From 5adcc6c7b48b7a213af91bc123a02ab87751228e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Feb 2013 19:06:14 +0100 Subject: [PATCH 0001/1103] Add support for NVCC5: most of the Core and part of LU are callable from CUDA code. Still a lot to do. --- Eigen/Core | 10 +++ Eigen/src/Core/Array.h | 19 ++++- Eigen/src/Core/Assign.h | 16 ++++ Eigen/src/Core/Block.h | 44 +++++++--- Eigen/src/Core/CommaInitializer.h | 6 ++ Eigen/src/Core/CwiseBinaryOp.h | 8 ++ Eigen/src/Core/CwiseUnaryOp.h | 9 ++ Eigen/src/Core/DenseBase.h | 49 ++++++++--- Eigen/src/Core/DenseCoeffsBase.h | 33 ++++++++ Eigen/src/Core/DenseStorage.h | 32 ++++---- Eigen/src/Core/EigenBase.h | 25 ++++-- Eigen/src/Core/Functors.h | 91 +++++++++++---------- Eigen/src/Core/GeneralProduct.h | 3 +- Eigen/src/Core/Map.h | 5 ++ Eigen/src/Core/MapBase.h | 21 +++-- Eigen/src/Core/Matrix.h | 22 ++++- Eigen/src/Core/MatrixBase.h | 30 +++++-- Eigen/src/Core/NoAlias.h | 7 ++ Eigen/src/Core/PlainObjectBase.h | 30 +++++++ Eigen/src/Core/Redux.h | 4 + Eigen/src/Core/ReturnByValue.h | 5 +- Eigen/src/Core/Stride.h | 13 ++- Eigen/src/Core/Transpose.h | 17 +++- Eigen/src/Core/VectorBlock.h | 2 + Eigen/src/Core/products/CoeffBasedProduct.h | 19 ++++- Eigen/src/Core/util/Macros.h | 2 +- Eigen/src/Core/util/XprHelper.h | 33 ++++---- Eigen/src/LU/FullPivLU.h | 2 + Eigen/src/LU/Inverse.h | 23 +++++- Eigen/src/LU/PartialPivLU.h | 5 ++ Eigen/src/plugins/ArrayCwiseUnaryOps.h | 14 ++++ Eigen/src/plugins/BlockMethods.h | 60 ++++++++++++++ Eigen/src/plugins/CommonCwiseBinaryOps.h | 1 + Eigen/src/plugins/CommonCwiseUnaryOps.h | 14 ++++ Eigen/src/plugins/MatrixCwiseBinaryOps.h | 8 ++ Eigen/src/plugins/MatrixCwiseUnaryOps.h | 5 ++ 36 files changed, 550 insertions(+), 137 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index d5b286e53..d8f62c825 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -49,6 +49,16 @@ #endif #endif +// Handle NVCC/CUDA +#ifdef __CUDACC__ + // Do not try to vectorize on CUDA! + #define EIGEN_DONT_VECTORIZE + // Do not try asserts on CUDA! + #define EIGEN_NO_DEBUG + // All functions callable from CUDA code must be qualified with __device__ + #define EIGEN_DEVICE_FUNC __host__ __device__ +#endif + #ifndef EIGEN_DONT_VECTORIZE #if defined (EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC) || defined(EIGEN_SSE2_ON_MSVC_2008_OR_LATER) diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h index 539e1d22b..707a9d7f2 100644 --- a/Eigen/src/Core/Array.h +++ b/Eigen/src/Core/Array.h @@ -69,6 +69,7 @@ class Array * the usage of 'using'. This should be done only for operator=. */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array& operator=(const EigenBase &other) { return Base::operator=(other); @@ -84,6 +85,7 @@ class Array * remain row-vectors and vectors remain vectors. */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array& operator=(const ArrayBase& other) { return Base::_set(other); @@ -92,6 +94,7 @@ class Array /** This is a special case of the templated operator=. Its purpose is to * prevent a default operator= from hiding the templated operator=. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array& operator=(const Array& other) { return Base::_set(other); @@ -107,6 +110,7 @@ class Array * * \sa resize(Index,Index) */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Array() : Base() { Base::_check_template_params(); @@ -116,6 +120,7 @@ class Array #ifndef EIGEN_PARSED_BY_DOXYGEN // FIXME is it still needed ?? /** \internal */ + EIGEN_DEVICE_FUNC Array(internal::constructor_without_unaligned_array_assert) : Base(internal::constructor_without_unaligned_array_assert()) { @@ -130,6 +135,7 @@ class Array * it is redundant to pass the dimension here, so it makes more sense to use the default * constructor Matrix() instead. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Array(Index dim) : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim) { @@ -142,6 +148,7 @@ class Array #ifndef EIGEN_PARSED_BY_DOXYGEN template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const T0& val0, const T1& val1) { Base::_check_template_params(); @@ -159,6 +166,7 @@ class Array #endif /** constructs an initialized 3D vector with given coefficients */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2) { Base::_check_template_params(); @@ -168,6 +176,7 @@ class Array m_storage.data()[2] = val2; } /** constructs an initialized 4D vector with given coefficients */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2, const Scalar& val3) { Base::_check_template_params(); @@ -178,10 +187,11 @@ class Array m_storage.data()[3] = val3; } - explicit Array(const Scalar *data); + EIGEN_DEVICE_FUNC explicit Array(const Scalar *data); /** Constructor copying the value of the expression \a other */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const ArrayBase& other) : Base(other.rows() * other.cols(), other.rows(), other.cols()) { @@ -189,6 +199,7 @@ class Array Base::_set_noalias(other); } /** Copy constructor */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const Array& other) : Base(other.rows() * other.cols(), other.rows(), other.cols()) { @@ -197,6 +208,7 @@ class Array } /** Copy constructor with in-place evaluation */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const ReturnByValue& other) { Base::_check_template_params(); @@ -206,6 +218,7 @@ class Array /** \sa MatrixBase::operator=(const EigenBase&) */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const EigenBase &other) : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols()) { @@ -221,8 +234,8 @@ class Array void swap(ArrayBase const & other) { this->_swap(other.derived()); } - inline Index innerStride() const { return 1; } - inline Index outerStride() const { return this->innerSize(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); } #ifdef EIGEN_ARRAY_PLUGIN #include EIGEN_ARRAY_PLUGIN diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index cd29a88f0..dc9b55fa4 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -139,6 +139,7 @@ struct assign_DefaultTraversal_CompleteUnrolling inner = Index % Derived1::InnerSizeAtCompileTime }; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) { dst.copyCoeffByOuterInner(outer, inner, src); @@ -149,12 +150,14 @@ struct assign_DefaultTraversal_CompleteUnrolling template struct assign_DefaultTraversal_CompleteUnrolling { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {} }; template struct assign_DefaultTraversal_InnerUnrolling { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, int outer) { dst.copyCoeffByOuterInner(outer, Index, src); @@ -165,6 +168,7 @@ struct assign_DefaultTraversal_InnerUnrolling template struct assign_DefaultTraversal_InnerUnrolling { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, int) {} }; @@ -175,6 +179,7 @@ struct assign_DefaultTraversal_InnerUnrolling template struct assign_LinearTraversal_CompleteUnrolling { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) { dst.copyCoeff(Index, src); @@ -185,6 +190,7 @@ struct assign_LinearTraversal_CompleteUnrolling template struct assign_LinearTraversal_CompleteUnrolling { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {} }; @@ -249,6 +255,7 @@ struct assign_impl; template struct assign_impl { + EIGEN_DEVICE_FUNC static inline void run(Derived1 &, const Derived2 &) { } }; @@ -256,6 +263,7 @@ template struct assign_impl { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { const Index innerSize = dst.innerSize(); @@ -269,6 +277,7 @@ struct assign_impl template struct assign_impl { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) { assign_DefaultTraversal_CompleteUnrolling @@ -280,6 +289,7 @@ template struct assign_impl { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) { const Index outerSize = dst.outerSize(); @@ -297,6 +307,7 @@ template struct assign_impl { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { const Index size = dst.size(); @@ -308,6 +319,7 @@ struct assign_impl template struct assign_impl { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) { assign_LinearTraversal_CompleteUnrolling @@ -519,18 +531,22 @@ struct assign_selector; template struct assign_selector { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); } }; template struct assign_selector { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.eval()); } }; template struct assign_selector { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); } }; template struct assign_selector { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); } }; diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 5f6426517..fbc2cf227 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -111,6 +111,7 @@ template class /** Column or Row constructor */ + EIGEN_DEVICE_FUNC inline Block(XprType& xpr, Index i) : Impl(xpr,i) { eigen_assert( (i>=0) && ( @@ -120,6 +121,7 @@ template class /** Fixed-size constructor */ + EIGEN_DEVICE_FUNC inline Block(XprType& xpr, Index a_startRow, Index a_startCol) : Impl(xpr, a_startRow, a_startCol) { @@ -130,6 +132,7 @@ template class /** Dynamic-size constructor */ + EIGEN_DEVICE_FUNC inline Block(XprType& xpr, Index a_startRow, Index a_startCol, Index blockRows, Index blockCols) @@ -153,8 +156,9 @@ class BlockImpl public: typedef Impl Base; EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) - inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {} - inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol) : Impl(xpr, a_startRow, a_startCol) {} + EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {} + EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol) : Impl(xpr, a_startRow, a_startCol) {} + EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index a_startRow, Index a_startCol, Index blockRows, Index blockCols) : Impl(xpr, a_startRow, a_startCol, blockRows, blockCols) {} }; @@ -176,6 +180,7 @@ template::type& nestedExpression() const + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& nestedExpression() const { return m_xpr; } - Index startRow() const + EIGEN_DEVICE_FUNC + Index startRow() const { return m_startRow.value(); } + EIGEN_DEVICE_FUNC Index startCol() const { return m_startCol.value(); @@ -322,6 +338,7 @@ class BlockImpl_dense /** Column or Row constructor */ + EIGEN_DEVICE_FUNC inline BlockImpl_dense(XprType& xpr, Index i) : Base(internal::const_cast_ptr(&xpr.coeffRef( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0, @@ -335,6 +352,7 @@ class BlockImpl_dense /** Fixed-size constructor */ + EIGEN_DEVICE_FUNC inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol) : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol))), m_xpr(xpr) { @@ -343,6 +361,7 @@ class BlockImpl_dense /** Dynamic-size constructor */ + EIGEN_DEVICE_FUNC inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) @@ -352,12 +371,14 @@ class BlockImpl_dense init(); } - const typename internal::remove_all::type& nestedExpression() const + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& nestedExpression() const { return m_xpr; } /** \sa MapBase::innerStride() */ + EIGEN_DEVICE_FUNC inline Index innerStride() const { return internal::traits::HasSameStorageOrderAsXprType @@ -366,6 +387,7 @@ class BlockImpl_dense } /** \sa MapBase::outerStride() */ + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_outerStride; @@ -379,6 +401,7 @@ class BlockImpl_dense #ifndef EIGEN_PARSED_BY_DOXYGEN /** \internal used by allowAligned() */ + EIGEN_DEVICE_FUNC inline BlockImpl_dense(XprType& xpr, const Scalar* data, Index blockRows, Index blockCols) : Base(data, blockRows, blockCols), m_xpr(xpr) { @@ -387,6 +410,7 @@ class BlockImpl_dense #endif protected: + EIGEN_DEVICE_FUNC void init() { m_outerStride = internal::traits::HasSameStorageOrderAsXprType diff --git a/Eigen/src/Core/CommaInitializer.h b/Eigen/src/Core/CommaInitializer.h index f20c1774c..1f801e2a0 100644 --- a/Eigen/src/Core/CommaInitializer.h +++ b/Eigen/src/Core/CommaInitializer.h @@ -30,6 +30,7 @@ struct CommaInitializer typedef typename XprType::Scalar Scalar; typedef typename XprType::Index Index; + EIGEN_DEVICE_FUNC inline CommaInitializer(XprType& xpr, const Scalar& s) : m_xpr(xpr), m_row(0), m_col(1), m_currentBlockRows(1) { @@ -37,6 +38,7 @@ struct CommaInitializer } template + EIGEN_DEVICE_FUNC inline CommaInitializer(XprType& xpr, const DenseBase& other) : m_xpr(xpr), m_row(0), m_col(other.cols()), m_currentBlockRows(other.rows()) { @@ -44,6 +46,7 @@ struct CommaInitializer } /* inserts a scalar value in the target matrix */ + EIGEN_DEVICE_FUNC CommaInitializer& operator,(const Scalar& s) { if (m_col==m_xpr.cols()) @@ -63,6 +66,7 @@ struct CommaInitializer /* inserts a matrix expression in the target matrix */ template + EIGEN_DEVICE_FUNC CommaInitializer& operator,(const DenseBase& other) { if(other.cols()==0 || other.rows()==0) @@ -88,6 +92,7 @@ struct CommaInitializer return *this; } + EIGEN_DEVICE_FUNC inline ~CommaInitializer() { eigen_assert((m_row+m_currentBlockRows) == m_xpr.rows() @@ -102,6 +107,7 @@ struct CommaInitializer * quaternion.fromRotationMatrix((Matrix3f() << axis0, axis1, axis2).finished()); * \endcode */ + EIGEN_DEVICE_FUNC inline XprType& finished() { return m_xpr; } XprType& m_xpr; // target expression diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index 686c2afa3..532b2b96e 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -122,6 +122,7 @@ class CwiseBinaryOp : internal::no_assignment_operator, typedef typename internal::remove_reference::type _LhsNested; typedef typename internal::remove_reference::type _RhsNested; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CwiseBinaryOp(const Lhs& aLhs, const Rhs& aRhs, const BinaryOp& func = BinaryOp()) : m_lhs(aLhs), m_rhs(aRhs), m_functor(func) { @@ -131,6 +132,7 @@ class CwiseBinaryOp : internal::no_assignment_operator, eigen_assert(aLhs.rows() == aRhs.rows() && aLhs.cols() == aRhs.cols()); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { // return the fixed size type if available to enable compile time optimizations if (internal::traits::type>::RowsAtCompileTime==Dynamic) @@ -138,6 +140,7 @@ class CwiseBinaryOp : internal::no_assignment_operator, else return m_lhs.rows(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { // return the fixed size type if available to enable compile time optimizations if (internal::traits::type>::ColsAtCompileTime==Dynamic) @@ -147,10 +150,13 @@ class CwiseBinaryOp : internal::no_assignment_operator, } /** \returns the left hand side nested expression */ + EIGEN_DEVICE_FUNC const _LhsNested& lhs() const { return m_lhs; } /** \returns the right hand side nested expression */ + EIGEN_DEVICE_FUNC const _RhsNested& rhs() const { return m_rhs; } /** \returns the functor representing the binary operation */ + EIGEN_DEVICE_FUNC const BinaryOp& functor() const { return m_functor; } protected: @@ -169,6 +175,7 @@ class CwiseBinaryOpImpl typedef typename internal::dense_xpr_base >::type Base; EIGEN_DENSE_PUBLIC_INTERFACE( Derived ) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const { return derived().functor()(derived().lhs().coeff(rowId, colId), @@ -182,6 +189,7 @@ class CwiseBinaryOpImpl derived().rhs().template packet(rowId, colId)); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const { return derived().functor()(derived().lhs().coeff(index), diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index f2de749f9..aa7df197f 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -64,20 +64,26 @@ class CwiseUnaryOp : internal::no_assignment_operator, typedef typename CwiseUnaryOpImpl::StorageKind>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryOp) + EIGEN_DEVICE_FUNC inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) : m_xpr(xpr), m_functor(func) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_xpr.cols(); } /** \returns the functor representing the unary operation */ + EIGEN_DEVICE_FUNC const UnaryOp& functor() const { return m_functor; } /** \returns the nested expression */ + EIGEN_DEVICE_FUNC const typename internal::remove_all::type& nestedExpression() const { return m_xpr; } /** \returns the nested expression */ + EIGEN_DEVICE_FUNC typename internal::remove_all::type& nestedExpression() { return m_xpr.const_cast_derived(); } @@ -98,6 +104,7 @@ class CwiseUnaryOpImpl typedef typename internal::dense_xpr_base >::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const { return derived().functor()(derived().nestedExpression().coeff(rowId, colId)); @@ -109,12 +116,14 @@ class CwiseUnaryOpImpl return derived().functor().packetOp(derived().nestedExpression().template packet(rowId, colId)); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const { return derived().functor()(derived().nestedExpression().coeff(index)); } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketScalar packet(Index index) const { return derived().functor().packetOp(derived().nestedExpression().template packet(index)); diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 62c73f1a9..12780354b 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -172,6 +172,7 @@ template class DenseBase /** \returns the number of nonzero coefficients which is in practice the number * of stored coefficients. */ + EIGEN_DEVICE_FUNC inline Index nonZeros() const { return size(); } /** \returns true if either the number of rows or the number of columns is equal to 1. * In other words, this function returns @@ -183,6 +184,7 @@ template class DenseBase * \note For a vector, this returns just 1. For a matrix (non-vector), this is the major dimension * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of columns for a * column-major matrix, and the number of rows for a row-major matrix. */ + EIGEN_DEVICE_FUNC Index outerSize() const { return IsVectorAtCompileTime ? 1 @@ -194,6 +196,7 @@ template class DenseBase * \note For a vector, this is just the size. For a matrix (non-vector), this is the minor dimension * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of rows for a * column-major matrix, and the number of columns for a row-major matrix. */ + EIGEN_DEVICE_FUNC Index innerSize() const { return IsVectorAtCompileTime ? this->size() @@ -204,6 +207,7 @@ template class DenseBase * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does * nothing else. */ + EIGEN_DEVICE_FUNC void resize(Index newSize) { EIGEN_ONLY_USED_FOR_DEBUG(newSize); @@ -214,6 +218,7 @@ template class DenseBase * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does * nothing else. */ + EIGEN_DEVICE_FUNC void resize(Index nbRows, Index nbCols) { EIGEN_ONLY_USED_FOR_DEBUG(nbRows); @@ -237,42 +242,54 @@ template class DenseBase /** Copies \a other into *this. \returns a reference to *this. */ template + EIGEN_DEVICE_FUNC Derived& operator=(const DenseBase& other); /** Special case of the template operator=, in order to prevent the compiler * from generating a default operator= (issue hit with g++ 4.1) */ + EIGEN_DEVICE_FUNC Derived& operator=(const DenseBase& other); template + EIGEN_DEVICE_FUNC Derived& operator=(const EigenBase &other); template + EIGEN_DEVICE_FUNC Derived& operator+=(const EigenBase &other); template + EIGEN_DEVICE_FUNC Derived& operator-=(const EigenBase &other); template + EIGEN_DEVICE_FUNC Derived& operator=(const ReturnByValue& func); #ifndef EIGEN_PARSED_BY_DOXYGEN /** Copies \a other into *this without evaluating other. \returns a reference to *this. */ template + EIGEN_DEVICE_FUNC Derived& lazyAssign(const DenseBase& other); #endif // not EIGEN_PARSED_BY_DOXYGEN + EIGEN_DEVICE_FUNC CommaInitializer operator<< (const Scalar& s); template const Flagged flagged() const; template + EIGEN_DEVICE_FUNC CommaInitializer operator<< (const DenseBase& other); + EIGEN_DEVICE_FUNC Eigen::Transpose transpose(); typedef const Transpose ConstTransposeReturnType; + EIGEN_DEVICE_FUNC ConstTransposeReturnType transpose() const; + EIGEN_DEVICE_FUNC void transposeInPlace(); #ifndef EIGEN_NO_DEBUG protected: @@ -346,6 +363,7 @@ template class DenseBase * Notice that in the case of a plain matrix or vector (not an expression) this function just returns * a const reference, in order to avoid a useless copy. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE EvalReturnType eval() const { // Even though MSVC does not honor strong inlining when the return type @@ -380,14 +398,14 @@ template class DenseBase template inline const typename internal::conditional,Derived&>::type forceAlignedAccessIf() const; template inline typename internal::conditional,Derived&>::type forceAlignedAccessIf(); - Scalar sum() const; - Scalar mean() const; - Scalar trace() const; + EIGEN_DEVICE_FUNC Scalar sum() const; + EIGEN_DEVICE_FUNC Scalar mean() const; + EIGEN_DEVICE_FUNC Scalar trace() const; - Scalar prod() const; + EIGEN_DEVICE_FUNC Scalar prod() const; - typename internal::traits::Scalar minCoeff() const; - typename internal::traits::Scalar maxCoeff() const; + EIGEN_DEVICE_FUNC typename internal::traits::Scalar minCoeff() const; + EIGEN_DEVICE_FUNC typename internal::traits::Scalar maxCoeff() const; template typename internal::traits::Scalar minCoeff(IndexType* row, IndexType* col) const; @@ -399,15 +417,18 @@ template class DenseBase typename internal::traits::Scalar maxCoeff(IndexType* index) const; template + EIGEN_DEVICE_FUNC typename internal::result_of::Scalar)>::type redux(const BinaryOp& func) const; template + EIGEN_DEVICE_FUNC void visit(Visitor& func) const; inline const WithFormat format(const IOFormat& fmt) const; /** \returns the unique coefficient of a 1x1 expression */ + EIGEN_DEVICE_FUNC CoeffReturnType value() const { EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) @@ -417,8 +438,8 @@ template class DenseBase /////////// Array module /////////// - bool all(void) const; - bool any(void) const; + bool all() const; + bool any() const; Index count() const; typedef VectorwiseOp RowwiseReturnType; @@ -480,14 +501,16 @@ template class DenseBase // disable the use of evalTo for dense objects with a nice compilation error - template inline void evalTo(Dest& ) const + template + EIGEN_DEVICE_FUNC + inline void evalTo(Dest& ) const { EIGEN_STATIC_ASSERT((internal::is_same::value),THE_EVAL_EVALTO_FUNCTION_SHOULD_NEVER_BE_CALLED_FOR_DENSE_OBJECTS); } protected: /** Default constructor. Do nothing. */ - DenseBase() + EIGEN_DEVICE_FUNC DenseBase() { /* Just checks for self-consistency of the flags. * Only do it when debugging Eigen, as this borders on paranoiac and could slow compilation down @@ -500,9 +523,9 @@ template class DenseBase } private: - explicit DenseBase(int); - DenseBase(int,int); - template explicit DenseBase(const DenseBase&); + EIGEN_DEVICE_FUNC explicit DenseBase(int); + EIGEN_DEVICE_FUNC DenseBase(int,int); + template EIGEN_DEVICE_FUNC explicit DenseBase(const DenseBase&); }; } // end namespace Eigen diff --git a/Eigen/src/Core/DenseCoeffsBase.h b/Eigen/src/Core/DenseCoeffsBase.h index 3c890f215..efabb5e67 100644 --- a/Eigen/src/Core/DenseCoeffsBase.h +++ b/Eigen/src/Core/DenseCoeffsBase.h @@ -61,6 +61,7 @@ class DenseCoeffsBase : public EigenBase using Base::size; using Base::derived; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner) const { return int(Derived::RowsAtCompileTime) == 1 ? 0 @@ -69,6 +70,7 @@ class DenseCoeffsBase : public EigenBase : inner; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner) const { return int(Derived::ColsAtCompileTime) == 1 ? 0 @@ -91,6 +93,7 @@ class DenseCoeffsBase : public EigenBase * * \sa operator()(Index,Index) const, coeffRef(Index,Index), coeff(Index) const */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const { eigen_internal_assert(row >= 0 && row < rows() @@ -98,6 +101,7 @@ class DenseCoeffsBase : public EigenBase return derived().coeff(row, col); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffByOuterInner(Index outer, Index inner) const { return coeff(rowIndexByOuterInner(outer, inner), @@ -108,6 +112,7 @@ class DenseCoeffsBase : public EigenBase * * \sa operator()(Index,Index), operator[](Index) */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType operator()(Index row, Index col) const { eigen_assert(row >= 0 && row < rows() @@ -130,6 +135,7 @@ class DenseCoeffsBase : public EigenBase * \sa operator[](Index) const, coeffRef(Index), coeff(Index,Index) const */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { @@ -146,6 +152,7 @@ class DenseCoeffsBase : public EigenBase * z() const, w() const */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType operator[](Index index) const { @@ -167,6 +174,7 @@ class DenseCoeffsBase : public EigenBase * z() const, w() const */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType operator()(Index index) const { @@ -176,21 +184,25 @@ class DenseCoeffsBase : public EigenBase /** equivalent to operator[](0). */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType x() const { return (*this)[0]; } /** equivalent to operator[](1). */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType y() const { return (*this)[1]; } /** equivalent to operator[](2). */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType z() const { return (*this)[2]; } /** equivalent to operator[](3). */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType w() const { return (*this)[3]; } @@ -311,6 +323,7 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && row < rows() @@ -318,6 +331,7 @@ class DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, const DenseBase& other) { eigen_internal_assert(row >= 0 && row < rows() @@ -489,6 +512,7 @@ class DenseCoeffsBase : public DenseCoeffsBase + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void copyCoeff(Index index, const DenseBase& other) { eigen_internal_assert(index >= 0 && index < size()); @@ -497,6 +521,7 @@ class DenseCoeffsBase : public DenseCoeffsBase + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void copyCoeffByOuterInner(Index outer, Index inner, const DenseBase& other) { const Index row = rowIndexByOuterInner(outer,inner); @@ -581,6 +606,7 @@ class DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase * * \sa outerStride(), rowStride(), colStride() */ + EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().innerStride(); @@ -662,6 +692,7 @@ class DenseCoeffsBase * * \sa innerStride(), rowStride(), colStride() */ + EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().outerStride(); @@ -677,6 +708,7 @@ class DenseCoeffsBase * * \sa innerStride(), outerStride(), colStride() */ + EIGEN_DEVICE_FUNC inline Index rowStride() const { return Derived::IsRowMajor ? outerStride() : innerStride(); @@ -686,6 +718,7 @@ class DenseCoeffsBase * * \sa innerStride(), outerStride(), rowStride() */ + EIGEN_DEVICE_FUNC inline Index colStride() const { return Derived::IsRowMajor ? innerStride() : outerStride(); diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 894dcf2c1..203944620 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -36,12 +36,14 @@ struct plain_array { T array[Size]; - plain_array() + EIGEN_DEVICE_FUNC + plain_array() { EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); } - plain_array(constructor_without_unaligned_array_assert) + EIGEN_DEVICE_FUNC + plain_array(constructor_without_unaligned_array_assert) { EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); } @@ -73,12 +75,14 @@ struct plain_array { EIGEN_USER_ALIGN16 T array[Size]; + EIGEN_DEVICE_FUNC plain_array() { EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(0xf); EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); } + EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) { EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); @@ -89,8 +93,8 @@ template struct plain_array { EIGEN_USER_ALIGN16 T array[1]; - plain_array() {} - plain_array(constructor_without_unaligned_array_assert) {} + EIGEN_DEVICE_FUNC plain_array() {} + EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {} }; } // end namespace internal @@ -114,17 +118,17 @@ template class DenseSt { internal::plain_array m_data; public: - inline explicit DenseStorage() {} - inline DenseStorage(internal::constructor_without_unaligned_array_assert) + EIGEN_DEVICE_FUNC inline explicit DenseStorage() {} + EIGEN_DEVICE_FUNC inline DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()) {} - inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} - inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); } - static inline DenseIndex rows(void) {return _Rows;} - static inline DenseIndex cols(void) {return _Cols;} - inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} - inline void resize(DenseIndex,DenseIndex,DenseIndex) {} - inline const T *data() const { return m_data.array; } - inline T *data() { return m_data.array; } + EIGEN_DEVICE_FUNC inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); } + EIGEN_DEVICE_FUNC static inline DenseIndex rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC static inline DenseIndex cols(void) {return _Cols;} + EIGEN_DEVICE_FUNC inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC inline void resize(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC inline const T *data() const { return m_data.array; } + EIGEN_DEVICE_FUNC inline T *data() { return m_data.array; } }; // null matrix diff --git a/Eigen/src/Core/EigenBase.h b/Eigen/src/Core/EigenBase.h index 2b8dd1b70..a25e823ab 100644 --- a/Eigen/src/Core/EigenBase.h +++ b/Eigen/src/Core/EigenBase.h @@ -31,29 +31,40 @@ template struct EigenBase typedef typename internal::traits::Index Index; /** \returns a reference to the derived object */ + EIGEN_DEVICE_FUNC Derived& derived() { return *static_cast(this); } /** \returns a const reference to the derived object */ + EIGEN_DEVICE_FUNC const Derived& derived() const { return *static_cast(this); } + EIGEN_DEVICE_FUNC inline Derived& const_cast_derived() const { return *static_cast(const_cast(this)); } + EIGEN_DEVICE_FUNC inline const Derived& const_derived() const { return *static_cast(this); } /** \returns the number of rows. \sa cols(), RowsAtCompileTime */ + EIGEN_DEVICE_FUNC inline Index rows() const { return derived().rows(); } /** \returns the number of columns. \sa rows(), ColsAtCompileTime*/ + EIGEN_DEVICE_FUNC inline Index cols() const { return derived().cols(); } /** \returns the number of coefficients, which is rows()*cols(). * \sa rows(), cols(), SizeAtCompileTime. */ + EIGEN_DEVICE_FUNC inline Index size() const { return rows() * cols(); } /** \internal Don't use it, but do the equivalent: \code dst = *this; \endcode */ - template inline void evalTo(Dest& dst) const + template + EIGEN_DEVICE_FUNC + inline void evalTo(Dest& dst) const { derived().evalTo(dst); } /** \internal Don't use it, but do the equivalent: \code dst += *this; \endcode */ - template inline void addTo(Dest& dst) const + template + EIGEN_DEVICE_FUNC + inline void addTo(Dest& dst) const { // This is the default implementation, // derived class can reimplement it in a more optimized way. @@ -63,7 +74,9 @@ template struct EigenBase } /** \internal Don't use it, but do the equivalent: \code dst -= *this; \endcode */ - template inline void subTo(Dest& dst) const + template + EIGEN_DEVICE_FUNC + inline void subTo(Dest& dst) const { // This is the default implementation, // derived class can reimplement it in a more optimized way. @@ -73,7 +86,8 @@ template struct EigenBase } /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheRight(*this); \endcode */ - template inline void applyThisOnTheRight(Dest& dst) const + template + EIGEN_DEVICE_FUNC inline void applyThisOnTheRight(Dest& dst) const { // This is the default implementation, // derived class can reimplement it in a more optimized way. @@ -81,7 +95,8 @@ template struct EigenBase } /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheLeft(*this); \endcode */ - template inline void applyThisOnTheLeft(Dest& dst) const + template + EIGEN_DEVICE_FUNC inline void applyThisOnTheLeft(Dest& dst) const { // This is the default implementation, // derived class can reimplement it in a more optimized way. diff --git a/Eigen/src/Core/Functors.h b/Eigen/src/Core/Functors.h index 2a6c3c003..6b6b36656 100644 --- a/Eigen/src/Core/Functors.h +++ b/Eigen/src/Core/Functors.h @@ -23,7 +23,7 @@ namespace internal { */ template struct scalar_sum_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op) - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::padd(a,b); } @@ -51,7 +51,7 @@ template struct scalar_product_op { }; typedef typename scalar_product_traits::ReturnType result_type; EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op) - EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pmul(a,b); } @@ -81,7 +81,7 @@ template struct scalar_conj_product_op { typedef typename scalar_product_traits::ReturnType result_type; EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op) - EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return conj_helper().pmul(a,b); } template @@ -103,7 +103,7 @@ struct functor_traits > { */ template struct scalar_min_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op) - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::min; return (min)(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::min; return (min)(a, b); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pmin(a,b); } @@ -126,7 +126,7 @@ struct functor_traits > { */ template struct scalar_max_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op) - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::max; return (max)(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::max; return (max)(a, b); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pmax(a,b); } @@ -150,7 +150,7 @@ struct functor_traits > { template struct scalar_hypot_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op) // typedef typename NumTraits::Real result_type; - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const { using std::max; using std::min; @@ -170,7 +170,7 @@ struct functor_traits > { */ template struct scalar_binary_pow_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_binary_pow_op) - inline Scalar operator() (const Scalar& a, const OtherScalar& b) const { return internal::pow(a, b); } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a, const OtherScalar& b) const { return internal::pow(a, b); } }; template struct functor_traits > { @@ -186,7 +186,7 @@ struct functor_traits > { */ template struct scalar_difference_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op) - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a - b; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a - b; } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::psub(a,b); } @@ -211,7 +211,7 @@ template struct scalar_quotient_op { }; typedef typename scalar_product_traits::ReturnType result_type; EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op) - EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pdiv(a,b); } @@ -233,7 +233,7 @@ struct functor_traits > { */ struct scalar_boolean_and_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op) - EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; } }; template<> struct functor_traits { enum { @@ -249,7 +249,7 @@ template<> struct functor_traits { */ struct scalar_boolean_or_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op) - EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; } }; template<> struct functor_traits { enum { @@ -267,7 +267,7 @@ template<> struct functor_traits { */ template struct scalar_opposite_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op) - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pnegate(a); } @@ -287,7 +287,7 @@ struct functor_traits > template struct scalar_abs_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op) typedef typename NumTraits::Real result_type; - EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using std::abs; return abs(a); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using std::abs; return abs(a); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pabs(a); } @@ -309,7 +309,7 @@ struct functor_traits > template struct scalar_abs2_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op) typedef typename NumTraits::Real result_type; - EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return internal::abs2(a); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return internal::abs2(a); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pmul(a,a); } @@ -325,7 +325,7 @@ struct functor_traits > */ template struct scalar_conjugate_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op) - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using internal::conj; return conj(a); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using internal::conj; return conj(a); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); } }; @@ -347,7 +347,7 @@ template struct scalar_cast_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) typedef NewType result_type; - EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast(a); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast(a); } }; template struct functor_traits > @@ -362,7 +362,7 @@ template struct scalar_real_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op) typedef typename NumTraits::Real result_type; - EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return internal::real(a); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return internal::real(a); } }; template struct functor_traits > @@ -377,7 +377,7 @@ template struct scalar_imag_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op) typedef typename NumTraits::Real result_type; - EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return internal::imag(a); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return internal::imag(a); } }; template struct functor_traits > @@ -392,7 +392,7 @@ template struct scalar_real_ref_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op) typedef typename NumTraits::Real result_type; - EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return internal::real_ref(*const_cast(&a)); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return internal::real_ref(*const_cast(&a)); } }; template struct functor_traits > @@ -407,7 +407,7 @@ template struct scalar_imag_ref_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op) typedef typename NumTraits::Real result_type; - EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return internal::imag_ref(*const_cast(&a)); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return internal::imag_ref(*const_cast(&a)); } }; template struct functor_traits > @@ -421,7 +421,7 @@ struct functor_traits > */ template struct scalar_exp_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op) - inline const Scalar operator() (const Scalar& a) const { using std::exp; return exp(a); } + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::exp; return exp(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::pexp(a); } }; @@ -437,7 +437,7 @@ struct functor_traits > */ template struct scalar_log_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op) - inline const Scalar operator() (const Scalar& a) const { using std::log; return log(a); } + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::log; return log(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::plog(a); } }; @@ -462,8 +462,11 @@ template struct scalar_multiple_op { typedef typename packet_traits::type Packet; // FIXME default copy constructors seems bugged with std::complex<> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple_op(const scalar_multiple_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple_op(const Scalar& other) : m_other(other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a * m_other; } EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pmul(a, pset1(m_other)); } @@ -478,7 +481,7 @@ struct scalar_multiple2_op { typedef typename scalar_product_traits::ReturnType result_type; EIGEN_STRONG_INLINE scalar_multiple2_op(const scalar_multiple2_op& other) : m_other(other.m_other) { } EIGEN_STRONG_INLINE scalar_multiple2_op(const Scalar2& other) : m_other(other) { } - EIGEN_STRONG_INLINE result_type operator() (const Scalar1& a) const { return a * m_other; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar1& a) const { return a * m_other; } typename add_const_on_value_type::Nested>::type m_other; }; template @@ -499,7 +502,7 @@ struct scalar_quotient1_op { // FIXME default copy constructors seems bugged with std::complex<> EIGEN_STRONG_INLINE scalar_quotient1_op(const scalar_quotient1_op& other) : m_other(other.m_other) { } EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {} - EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; } EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pdiv(a, pset1(m_other)); } typename add_const_on_value_type::Nested>::type m_other; @@ -516,7 +519,7 @@ struct scalar_constant_op { EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { } EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { } template - EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; } template EIGEN_STRONG_INLINE const Packet packetOp(Index, Index = 0) const { return internal::pset1(m_other); } const Scalar m_other; @@ -529,7 +532,7 @@ struct functor_traits > template struct scalar_identity_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op) template - EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); } }; template struct functor_traits > @@ -553,7 +556,7 @@ struct linspaced_op_impl m_base(padd(pset1(low),pmul(pset1(step),plset(-packet_traits::size)))) {} template - EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; } template EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); } @@ -576,7 +579,7 @@ struct linspaced_op_impl m_lowPacket(pset1(m_low)), m_stepPacket(pset1(m_step)), m_interPacket(plset(0)) {} template - EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; } template EIGEN_STRONG_INLINE const Packet packetOp(Index i) const @@ -603,12 +606,12 @@ template struct linspaced_op linspaced_op(Scalar low, Scalar high, int num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/(num_steps-1))) {} template - EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); } // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since // there row==0 and col is used for the actual iteration. template - EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { eigen_assert(col==0 || row==0); return impl(col + row); @@ -657,9 +660,9 @@ template struct scalar_add_op { typedef typename packet_traits::type Packet; // FIXME default copy constructors seems bugged with std::complex<> - inline scalar_add_op(const scalar_add_op& other) : m_other(other.m_other) { } - inline scalar_add_op(const Scalar& other) : m_other(other) { } - inline Scalar operator() (const Scalar& a) const { return a + m_other; } + EIGEN_DEVICE_FUNC inline scalar_add_op(const scalar_add_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC inline scalar_add_op(const Scalar& other) : m_other(other) { } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a + m_other; } inline const Packet packetOp(const Packet& a) const { return internal::padd(a, pset1(m_other)); } const Scalar m_other; @@ -674,7 +677,7 @@ struct functor_traits > */ template struct scalar_sqrt_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op) - inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return sqrt(a); } + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return sqrt(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); } }; @@ -692,7 +695,7 @@ struct functor_traits > */ template struct scalar_cos_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op) - inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::pcos(a); } }; @@ -711,7 +714,7 @@ struct functor_traits > */ template struct scalar_sin_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op) - inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); } + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::psin(a); } }; @@ -731,7 +734,7 @@ struct functor_traits > */ template struct scalar_tan_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op) - inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); } + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::ptan(a); } }; @@ -750,7 +753,7 @@ struct functor_traits > */ template struct scalar_acos_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op) - inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); } + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::pacos(a); } }; @@ -769,7 +772,7 @@ struct functor_traits > */ template struct scalar_asin_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op) - inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); } + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); } typedef typename packet_traits::type Packet; inline Packet packetOp(const Packet& a) const { return internal::pasin(a); } }; @@ -791,7 +794,7 @@ struct scalar_pow_op { // FIXME default copy constructors seems bugged with std::complex<> inline scalar_pow_op(const scalar_pow_op& other) : m_exponent(other.m_exponent) { } inline scalar_pow_op(const Scalar& exponent) : m_exponent(exponent) {} - inline Scalar operator() (const Scalar& a) const { return internal::pow(a, m_exponent); } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return internal::pow(a, m_exponent); } const Scalar m_exponent; }; template @@ -805,7 +808,7 @@ struct functor_traits > template struct scalar_inverse_mult_op { scalar_inverse_mult_op(const Scalar& other) : m_other(other) {} - inline Scalar operator() (const Scalar& a) const { return m_other / a; } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return m_other / a; } template inline const Packet packetOp(const Packet& a) const { return internal::pdiv(pset1(m_other),a); } @@ -819,7 +822,7 @@ struct scalar_inverse_mult_op { template struct scalar_inverse_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op) - inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; } template inline const Packet packetOp(const Packet& a) const { return internal::pdiv(pset1(Scalar(1)),a); } @@ -835,7 +838,7 @@ struct functor_traits > template struct scalar_square_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op) - inline Scalar operator() (const Scalar& a) const { return a*a; } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a; } template inline const Packet packetOp(const Packet& a) const { return internal::pmul(a,a); } @@ -851,7 +854,7 @@ struct functor_traits > template struct scalar_cube_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op) - inline Scalar operator() (const Scalar& a) const { return a*a*a; } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a*a; } template inline const Packet packetOp(const Packet& a) const { return internal::pmul(a,pmul(a,a)); } diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 9abc7b286..a070e618d 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -543,6 +543,7 @@ template<> struct gemv_selector * * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*() */ +#ifndef __CUDACC__ template template inline const typename ProductReturnType::Type @@ -572,7 +573,7 @@ MatrixBase::operator*(const MatrixBase &other) const #endif return typename ProductReturnType::Type(derived(), other.derived()); } - +#endif /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation. * * The returned product will behave like any other expressions: the coefficients of the product will be diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index f804c89d6..e054f2202 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -118,11 +118,13 @@ template class Ma inline PointerType cast_to_pointer_type(PointerArgType ptr) { return ptr; } #endif + EIGEN_DEVICE_FUNC inline Index innerStride() const { return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1; } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer() @@ -136,6 +138,7 @@ template class Ma * \param dataPtr pointer to the array to map * \param a_stride optional Stride object, passing the strides. */ + EIGEN_DEVICE_FUNC inline Map(PointerArgType dataPtr, const StrideType& a_stride = StrideType()) : Base(cast_to_pointer_type(dataPtr)), m_stride(a_stride) { @@ -148,6 +151,7 @@ template class Ma * \param a_size the size of the vector expression * \param a_stride optional Stride object, passing the strides. */ + EIGEN_DEVICE_FUNC inline Map(PointerArgType dataPtr, Index a_size, const StrideType& a_stride = StrideType()) : Base(cast_to_pointer_type(dataPtr), a_size), m_stride(a_stride) { @@ -161,6 +165,7 @@ template class Ma * \param nbCols the number of columns of the matrix expression * \param a_stride optional Stride object, passing the strides. */ + EIGEN_DEVICE_FUNC inline Map(PointerArgType dataPtr, Index nbRows, Index nbCols, const StrideType& a_stride = StrideType()) : Base(cast_to_pointer_type(dataPtr), nbRows, nbCols), m_stride(a_stride) { diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index 6876de588..8def7442d 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -76,8 +76,8 @@ template class MapBase typedef typename Base::CoeffReturnType CoeffReturnType; - inline Index rows() const { return m_rows.value(); } - inline Index cols() const { return m_cols.value(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_rows.value(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_cols.value(); } /** Returns a pointer to the first coefficient of the matrix or vector. * @@ -87,22 +87,26 @@ template class MapBase */ inline const Scalar* data() const { return m_data; } + EIGEN_DEVICE_FUNC inline const Scalar& coeff(Index rowId, Index colId) const { return m_data[colId * colStride() + rowId * rowStride()]; } + EIGEN_DEVICE_FUNC inline const Scalar& coeff(Index index) const { EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) return m_data[index * innerStride()]; } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index rowId, Index colId) const { return this->m_data[colId * colStride() + rowId * rowStride()]; } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index index) const { EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) @@ -123,12 +127,14 @@ template class MapBase return internal::ploadt(m_data + index * innerStride()); } + EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime) { EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) checkSanity(); } + EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index vecSize) : m_data(dataPtr), m_rows(RowsAtCompileTime == Dynamic ? vecSize : Index(RowsAtCompileTime)), @@ -140,6 +146,7 @@ template class MapBase checkSanity(); } + EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols) : m_data(dataPtr), m_rows(nbRows), m_cols(nbCols) { @@ -151,6 +158,7 @@ template class MapBase protected: + EIGEN_DEVICE_FUNC void checkSanity() const { EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(internal::traits::Flags&PacketAccessBit, @@ -198,11 +206,13 @@ template class MapBase inline const Scalar* data() const { return this->m_data; } inline ScalarWithConstIfNotLvalue* data() { return this->m_data; } // no const-cast here so non-const-correct code will give a compile error + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue& coeffRef(Index row, Index col) { return this->m_data[col * colStride() + row * rowStride()]; } + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue& coeffRef(Index index) { EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) @@ -224,10 +234,11 @@ template class MapBase (this->m_data + index * innerStride(), val); } - explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {} - inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {} - inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols) : Base(dataPtr, nbRows, nbCols) {} + EIGEN_DEVICE_FUNC explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {} + EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {} + EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index nbRows, Index nbCols) : Base(dataPtr, nbRows, nbCols) {} + EIGEN_DEVICE_FUNC Derived& operator=(const MapBase& other) { Base::Base::operator=(other); diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 99160b591..61af9d9a3 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -151,6 +151,7 @@ class Matrix * * \callgraph */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix& operator=(const Matrix& other) { return Base::_set(other); @@ -167,6 +168,7 @@ class Matrix * remain row-vectors and vectors remain vectors. */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix& operator=(const MatrixBase& other) { return Base::_set(other); @@ -179,12 +181,14 @@ class Matrix * \copydetails DenseBase::operator=(const EigenBase &other) */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix& operator=(const EigenBase &other) { return Base::operator=(other); } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix& operator=(const ReturnByValue& func) { return Base::operator=(func); @@ -200,6 +204,7 @@ class Matrix * * \sa resize(Index,Index) */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Matrix() : Base() { Base::_check_template_params(); @@ -207,6 +212,7 @@ class Matrix } // FIXME is it still needed + EIGEN_DEVICE_FUNC Matrix(internal::constructor_without_unaligned_array_assert) : Base(internal::constructor_without_unaligned_array_assert()) { Base::_check_template_params(); EIGEN_INITIALIZE_BY_ZERO_IF_THAT_OPTION_IS_ENABLED } @@ -217,6 +223,7 @@ class Matrix * it is redundant to pass the dimension here, so it makes more sense to use the default * constructor Matrix() instead. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit Matrix(Index dim) : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim) { @@ -229,6 +236,7 @@ class Matrix #ifndef EIGEN_PARSED_BY_DOXYGEN template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const T0& x, const T1& y) { Base::_check_template_params(); @@ -240,12 +248,14 @@ class Matrix * This is useful for dynamic-size matrices. For fixed-size matrices, * it is redundant to pass these parameters, so one should use the default constructor * Matrix() instead. */ + EIGEN_DEVICE_FUNC Matrix(Index rows, Index cols); /** \brief Constructs an initialized 2D vector with given coefficients */ Matrix(const Scalar& x, const Scalar& y); #endif /** \brief Constructs an initialized 3D vector with given coefficients */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z) { Base::_check_template_params(); @@ -255,6 +265,7 @@ class Matrix m_storage.data()[2] = z; } /** \brief Constructs an initialized 4D vector with given coefficients */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z, const Scalar& w) { Base::_check_template_params(); @@ -265,10 +276,12 @@ class Matrix m_storage.data()[3] = w; } + EIGEN_DEVICE_FUNC explicit Matrix(const Scalar *data); /** \brief Constructor copying the value of the expression \a other */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const MatrixBase& other) : Base(other.rows() * other.cols(), other.rows(), other.cols()) { @@ -281,6 +294,7 @@ class Matrix Base::_set_noalias(other); } /** \brief Copy constructor */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const Matrix& other) : Base(other.rows() * other.cols(), other.rows(), other.cols()) { @@ -289,6 +303,7 @@ class Matrix } /** \brief Copy constructor with in-place evaluation */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const ReturnByValue& other) { Base::_check_template_params(); @@ -300,6 +315,7 @@ class Matrix * \sa MatrixBase::operator=(const EigenBase&) */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const EigenBase &other) : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols()) { @@ -318,14 +334,16 @@ class Matrix void swap(MatrixBase const & other) { this->_swap(other.derived()); } - inline Index innerStride() const { return 1; } - inline Index outerStride() const { return this->innerSize(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); } /////////// Geometry module /////////// template + EIGEN_DEVICE_FUNC explicit Matrix(const RotationBase& r); template + EIGEN_DEVICE_FUNC Matrix& operator=(const RotationBase& r); #ifdef EIGEN2_SUPPORT diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index e12a6763e..c76192e46 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -145,22 +145,27 @@ template class MatrixBase /** Special case of the template operator=, in order to prevent the compiler * from generating a default operator= (issue hit with g++ 4.1) */ + EIGEN_DEVICE_FUNC Derived& operator=(const MatrixBase& other); // We cannot inherit here via Base::operator= since it is causing // trouble with MSVC. template + EIGEN_DEVICE_FUNC Derived& operator=(const DenseBase& other); template + EIGEN_DEVICE_FUNC Derived& operator=(const EigenBase& other); template + EIGEN_DEVICE_FUNC Derived& operator=(const ReturnByValue& other); #ifndef EIGEN_PARSED_BY_DOXYGEN template + EIGEN_DEVICE_FUNC Derived& lazyAssign(const ProductBase& other); template @@ -168,15 +173,26 @@ template class MatrixBase #endif // not EIGEN_PARSED_BY_DOXYGEN template + EIGEN_DEVICE_FUNC Derived& operator+=(const MatrixBase& other); template + EIGEN_DEVICE_FUNC Derived& operator-=(const MatrixBase& other); +#ifdef __CUDACC__ + template + EIGEN_DEVICE_FUNC + const typename LazyProductReturnType::Type + operator*(const MatrixBase &other) const + { return this->lazyProduct(other); } +#else template const typename ProductReturnType::Type operator*(const MatrixBase &other) const; +#endif template + EIGEN_DEVICE_FUNC const typename LazyProductReturnType::Type lazyProduct(const MatrixBase &other) const; @@ -194,6 +210,7 @@ template class MatrixBase operator*(const DiagonalBase &diagonal) const; template + EIGEN_DEVICE_FUNC typename internal::scalar_product_traits::Scalar,typename internal::traits::Scalar>::ReturnType dot(const MatrixBase& other) const; @@ -324,8 +341,8 @@ template class MatrixBase /////////// LU module /////////// - const FullPivLU fullPivLu() const; - const PartialPivLU partialPivLu() const; + EIGEN_DEVICE_FUNC const FullPivLU fullPivLu() const; + EIGEN_DEVICE_FUNC const PartialPivLU partialPivLu() const; #if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS const LU lu() const; @@ -346,6 +363,7 @@ template class MatrixBase } #endif + EIGEN_DEVICE_FUNC const internal::inverse_impl inverse() const; template void computeInverseAndDetWithCheck( @@ -495,12 +513,12 @@ template class MatrixBase #endif protected: - MatrixBase() : Base() {} + EIGEN_DEVICE_FUNC MatrixBase() : Base() {} private: - explicit MatrixBase(int); - MatrixBase(int,int); - template explicit MatrixBase(const MatrixBase&); + EIGEN_DEVICE_FUNC explicit MatrixBase(int); + EIGEN_DEVICE_FUNC MatrixBase(int,int); + template EIGEN_DEVICE_FUNC explicit MatrixBase(const MatrixBase&); protected: // mixing arrays and matrices is not legal template Derived& operator+=(const ArrayBase& ) diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index 0112c865b..9e371538a 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -37,11 +37,13 @@ class NoAlias /** Behaves like MatrixBase::lazyAssign(other) * \sa MatrixBase::lazyAssign() */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase& other) { return internal::assign_selector::run(m_expression,other.derived()); } /** \sa MatrixBase::operator+= */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase& other) { typedef SelfCwiseBinaryOp, ExpressionType, OtherDerived> SelfAdder; @@ -54,6 +56,7 @@ class NoAlias /** \sa MatrixBase::operator-= */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase& other) { typedef SelfCwiseBinaryOp, ExpressionType, OtherDerived> SelfAdder; @@ -66,10 +69,12 @@ class NoAlias #ifndef EIGEN_PARSED_BY_DOXYGEN template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator+=(const ProductBase& other) { other.derived().addTo(m_expression); return m_expression; } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator-=(const ProductBase& other) { other.derived().subTo(m_expression); return m_expression; } @@ -78,10 +83,12 @@ class NoAlias { return m_expression.derived() += CoeffBasedProduct(other.lhs(), other.rhs()); } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator-=(const CoeffBasedProduct& other) { return m_expression.derived() -= CoeffBasedProduct(other.lhs(), other.rhs()); } #endif + EIGEN_DEVICE_FUNC ExpressionType& expression() const { return m_expression; diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 5c94ef621..49a5518e3 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -23,6 +23,7 @@ namespace internal { template struct check_rows_cols_for_overflow { template + EIGEN_DEVICE_FUNC static EIGEN_ALWAYS_INLINE void run(Index, Index) { } @@ -30,6 +31,7 @@ template struct check_rows_cols_for_overflow { template<> struct check_rows_cols_for_overflow { template + EIGEN_DEVICE_FUNC static EIGEN_ALWAYS_INLINE void run(Index rows, Index cols) { // http://hg.mozilla.org/mozilla-central/file/6c8a909977d3/xpcom/ds/CheckedInt.h#l242 @@ -124,9 +126,12 @@ class PlainObjectBase : public internal::dense_xpr_base::type Base& base() { return *static_cast(this); } const Base& base() const { return *static_cast(this); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { return m_storage.rows(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_storage.cols(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(Index rowId, Index colId) const { if(Flags & RowMajorBit) @@ -135,11 +140,13 @@ class PlainObjectBase : public internal::dense_xpr_base::type return m_storage.data()[rowId + colId * m_storage.rows()]; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const { return m_storage.data()[index]; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index rowId, Index colId) { if(Flags & RowMajorBit) @@ -148,11 +155,13 @@ class PlainObjectBase : public internal::dense_xpr_base::type return m_storage.data()[rowId + colId * m_storage.rows()]; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) { return m_storage.data()[index]; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeffRef(Index rowId, Index colId) const { if(Flags & RowMajorBit) @@ -161,6 +170,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type return m_storage.data()[rowId + colId * m_storage.rows()]; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeffRef(Index index) const { return m_storage.data()[index]; @@ -224,6 +234,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * \sa resize(Index) for vectors, resize(NoChange_t, Index), resize(Index, NoChange_t) */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index nbRows, Index nbCols) { eigen_assert( EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,nbRows==RowsAtCompileTime) @@ -254,6 +265,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * \sa resize(Index,Index), resize(NoChange_t, Index), resize(Index, NoChange_t) */ + EIGEN_DEVICE_FUNC inline void resize(Index size) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(PlainObjectBase) @@ -278,6 +290,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * \sa resize(Index,Index) */ + EIGEN_DEVICE_FUNC inline void resize(NoChange_t, Index nbCols) { resize(rows(), nbCols); @@ -291,6 +304,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * \sa resize(Index,Index) */ + EIGEN_DEVICE_FUNC inline void resize(Index nbRows, NoChange_t) { resize(nbRows, cols()); @@ -304,6 +318,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * remain row-vectors and vectors remain vectors. */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resizeLike(const EigenBase& _other) { const OtherDerived& other = _other.derived(); @@ -393,6 +408,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type /** This is a special case of the templated operator=. Its purpose is to * prevent a default operator= from hiding the templated operator=. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const PlainObjectBase& other) { return _set(other); @@ -400,6 +416,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type /** \sa MatrixBase::lazyAssign() */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& lazyAssign(const DenseBase& other) { _resize_to_match(other); @@ -407,12 +424,14 @@ class PlainObjectBase : public internal::dense_xpr_base::type } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const ReturnByValue& func) { resize(func.rows(), func.cols()); return Base::operator=(func); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit PlainObjectBase() : m_storage() { // _check_template_params(); @@ -422,6 +441,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type #ifndef EIGEN_PARSED_BY_DOXYGEN // FIXME is it still needed ? /** \internal */ + EIGEN_DEVICE_FUNC PlainObjectBase(internal::constructor_without_unaligned_array_assert) : m_storage(internal::constructor_without_unaligned_array_assert()) { @@ -429,6 +449,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type } #endif + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PlainObjectBase(Index a_size, Index nbRows, Index nbCols) : m_storage(a_size, nbRows, nbCols) { @@ -439,6 +460,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type /** \copydoc MatrixBase::operator=(const EigenBase&) */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const EigenBase &other) { _resize_to_match(other); @@ -448,6 +470,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type /** \sa MatrixBase::operator=(const EigenBase&) */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase &other) : m_storage(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols()) { @@ -558,6 +581,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * remain row-vectors and vectors remain vectors. */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _resize_to_match(const EigenBase& other) { #ifdef EIGEN_NO_AUTOMATIC_RESIZING @@ -585,6 +609,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * \internal */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& _set(const DenseBase& other) { _set_selector(other.derived(), typename internal::conditional(int(OtherDerived::Flags) & EvalBeforeAssigningBit), internal::true_type, internal::false_type>::type()); @@ -592,9 +617,11 @@ class PlainObjectBase : public internal::dense_xpr_base::type } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::true_type&) { _set_noalias(other.eval()); } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::false_type&) { _set_noalias(other); } /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which @@ -603,6 +630,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * \sa operator=(const MatrixBase&), _set() */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& _set_noalias(const DenseBase& other) { // I don't think we need this resize call since the lazyAssign will anyways resize @@ -622,6 +650,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type resize(nbRows,nbCols); } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init2(const Scalar& val0, const Scalar& val1, typename internal::enable_if::type* = 0) { EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2) @@ -644,6 +673,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type public: #ifndef EIGEN_PARSED_BY_DOXYGEN + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void _check_template_params() { EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, (Options&RowMajor)==RowMajor) diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index b7ce7c658..12b3db584 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -82,6 +82,7 @@ struct redux_novec_unroller typedef typename Derived::Scalar Scalar; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) { return func(redux_novec_unroller::run(mat,func), @@ -99,6 +100,7 @@ struct redux_novec_unroller typedef typename Derived::Scalar Scalar; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&) { return mat.coeffByOuterInner(outer, inner); @@ -112,6 +114,7 @@ template struct redux_novec_unroller { typedef typename Derived::Scalar Scalar; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); } }; @@ -170,6 +173,7 @@ struct redux_impl { typedef typename Derived::Scalar Scalar; typedef typename Derived::Index Index; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func) { eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index 613912ffa..87c6d9194 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -57,10 +57,11 @@ template class ReturnByValue EIGEN_DENSE_PUBLIC_INTERFACE(ReturnByValue) template + EIGEN_DEVICE_FUNC inline void evalTo(Dest& dst) const { static_cast(this)->evalTo(dst); } - inline Index rows() const { return static_cast(this)->rows(); } - inline Index cols() const { return static_cast(this)->cols(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return static_cast(this)->rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return static_cast(this)->cols(); } #ifndef EIGEN_PARSED_BY_DOXYGEN #define Unusable YOU_ARE_TRYING_TO_ACCESS_A_SINGLE_COEFFICIENT_IN_A_SPECIAL_EXPRESSION_WHERE_THAT_IS_NOT_ALLOWED_BECAUSE_THAT_WOULD_BE_INEFFICIENT diff --git a/Eigen/src/Core/Stride.h b/Eigen/src/Core/Stride.h index 1e3f5fe9f..d3d454e4e 100644 --- a/Eigen/src/Core/Stride.h +++ b/Eigen/src/Core/Stride.h @@ -51,6 +51,7 @@ class Stride }; /** Default constructor, for use when strides are fixed at compile time */ + EIGEN_DEVICE_FUNC Stride() : m_outer(OuterStrideAtCompileTime), m_inner(InnerStrideAtCompileTime) { @@ -58,6 +59,7 @@ class Stride } /** Constructor allowing to pass the strides at runtime */ + EIGEN_DEVICE_FUNC Stride(Index outerStride, Index innerStride) : m_outer(outerStride), m_inner(innerStride) { @@ -65,13 +67,16 @@ class Stride } /** Copy constructor */ + EIGEN_DEVICE_FUNC Stride(const Stride& other) : m_outer(other.outer()), m_inner(other.inner()) {} /** \returns the outer stride */ + EIGEN_DEVICE_FUNC inline Index outer() const { return m_outer.value(); } /** \returns the inner stride */ + EIGEN_DEVICE_FUNC inline Index inner() const { return m_inner.value(); } protected: @@ -87,8 +92,8 @@ class InnerStride : public Stride<0, Value> typedef Stride<0, Value> Base; public: typedef DenseIndex Index; - InnerStride() : Base() {} - InnerStride(Index v) : Base(0, v) {} + EIGEN_DEVICE_FUNC InnerStride() : Base() {} + EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} }; /** \brief Convenience specialization of Stride to specify only an outer stride @@ -99,8 +104,8 @@ class OuterStride : public Stride typedef Stride Base; public: typedef DenseIndex Index; - OuterStride() : Base() {} - OuterStride(Index v) : Base(v,0) {} + EIGEN_DEVICE_FUNC OuterStride() : Base() {} + EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} }; } // end namespace Eigen diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 34944e055..b5e1468df 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -62,18 +62,21 @@ template class Transpose typedef typename TransposeImpl::StorageKind>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose) + EIGEN_DEVICE_FUNC inline Transpose(MatrixType& a_matrix) : m_matrix(a_matrix) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose) - inline Index rows() const { return m_matrix.cols(); } - inline Index cols() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.rows(); } /** \returns the nested expression */ + EIGEN_DEVICE_FUNC const typename internal::remove_all::type& nestedExpression() const { return m_matrix; } /** \returns the nested expression */ + EIGEN_DEVICE_FUNC typename internal::remove_all::type& nestedExpression() { return m_matrix.const_cast_derived(); } @@ -105,8 +108,8 @@ template class TransposeImpl typedef typename internal::TransposeImpl_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Transpose) - inline Index innerStride() const { return derived().nestedExpression().innerStride(); } - inline Index outerStride() const { return derived().nestedExpression().outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().nestedExpression().outerStride(); } typedef typename internal::conditional< internal::is_lvalue::value, @@ -117,33 +120,39 @@ template class TransposeImpl inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } inline const Scalar* data() const { return derived().nestedExpression().data(); } + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue& coeffRef(Index rowId, Index colId) { EIGEN_STATIC_ASSERT_LVALUE(MatrixType) return derived().nestedExpression().const_cast_derived().coeffRef(colId, rowId); } + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue& coeffRef(Index index) { EIGEN_STATIC_ASSERT_LVALUE(MatrixType) return derived().nestedExpression().const_cast_derived().coeffRef(index); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index rowId, Index colId) const { return derived().nestedExpression().coeffRef(colId, rowId); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index index) const { return derived().nestedExpression().coeffRef(index); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index rowId, Index colId) const { return derived().nestedExpression().coeff(colId, rowId); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index index) const { return derived().nestedExpression().coeff(index); diff --git a/Eigen/src/Core/VectorBlock.h b/Eigen/src/Core/VectorBlock.h index 1a7330f3c..216c568c4 100644 --- a/Eigen/src/Core/VectorBlock.h +++ b/Eigen/src/Core/VectorBlock.h @@ -72,6 +72,7 @@ template class VectorBlock /** Dynamic-size constructor */ + EIGEN_DEVICE_FUNC inline VectorBlock(VectorType& vector, Index start, Index size) : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start, @@ -82,6 +83,7 @@ template class VectorBlock /** Fixed-size constructor */ + EIGEN_DEVICE_FUNC inline VectorBlock(VectorType& vector, Index start) : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start) { diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index 403d25fa9..312a05c71 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -140,11 +140,13 @@ class CoeffBasedProduct public: + EIGEN_DEVICE_FUNC inline CoeffBasedProduct(const CoeffBasedProduct& other) : Base(), m_lhs(other.m_lhs), m_rhs(other.m_rhs) {} template + EIGEN_DEVICE_FUNC inline CoeffBasedProduct(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs) { @@ -157,9 +159,10 @@ class CoeffBasedProduct && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); } - EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const { Scalar res; @@ -170,6 +173,7 @@ class CoeffBasedProduct /* Allow index-based non-packet access. It is impossible though to allow index-based packed access, * which is why we don't set the LinearAccessBit. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const { Scalar res; @@ -191,22 +195,26 @@ class CoeffBasedProduct } // Implicit conversion to the nested type (trigger the evaluation of the product) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE operator const PlainObject& () const { m_result.lazyAssign(*this); return m_result; } - const _LhsNested& lhs() const { return m_lhs; } - const _RhsNested& rhs() const { return m_rhs; } + EIGEN_DEVICE_FUNC const _LhsNested& lhs() const { return m_lhs; } + EIGEN_DEVICE_FUNC const _RhsNested& rhs() const { return m_rhs; } + EIGEN_DEVICE_FUNC const Diagonal diagonal() const { return reinterpret_cast(*this); } template + EIGEN_DEVICE_FUNC const Diagonal diagonal() const { return reinterpret_cast(*this); } + EIGEN_DEVICE_FUNC const Diagonal diagonal(Index index) const { return reinterpret_cast(*this).diagonal(index); } @@ -239,6 +247,7 @@ template struct product_coeff_impl { typedef typename Lhs::Index Index; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res) { product_coeff_impl::run(row, col, lhs, rhs, res); @@ -250,6 +259,7 @@ template struct product_coeff_impl { typedef typename Lhs::Index Index; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res) { res = lhs.coeff(row, 0) * rhs.coeff(0, col); @@ -260,6 +270,7 @@ template struct product_coeff_impl { typedef typename Lhs::Index Index; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar& res) { eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix"); diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 933a34c9d..bf6a9293c 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -396,7 +396,7 @@ #define EIGEN_MAKE_CWISE_BINARY_OP(METHOD,FUNCTOR) \ template \ - EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const OtherDerived> \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const OtherDerived> \ (METHOD)(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const \ { \ return CwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); \ diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 3d1290cd2..f115d3779 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -16,8 +16,8 @@ // so currently we simply disable this optimization for gcc 4.3 #if (defined __GNUG__) && !((__GNUC__==4) && (__GNUC_MINOR__==3)) #define EIGEN_EMPTY_STRUCT_CTOR(X) \ - EIGEN_STRONG_INLINE X() {} \ - EIGEN_STRONG_INLINE X(const X& ) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X() {} \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X(const X& ) {} #else #define EIGEN_EMPTY_STRUCT_CTOR(X) #endif @@ -50,19 +50,19 @@ template class variable_if_dynamic { public: EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamic) - explicit variable_if_dynamic(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); assert(v == T(Value)); } - static T value() { return T(Value); } - void setValue(T) {} + EIGEN_DEVICE_FUNC explicit variable_if_dynamic(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); } + EIGEN_DEVICE_FUNC static T value() { return T(Value); } + EIGEN_DEVICE_FUNC void setValue(T) {} }; template class variable_if_dynamic { T m_value; - variable_if_dynamic() { assert(false); } + EIGEN_DEVICE_FUNC variable_if_dynamic() { eigen_assert(false); } public: - explicit variable_if_dynamic(T value) : m_value(value) {} - T value() const { return m_value; } - void setValue(T value) { m_value = value; } + EIGEN_DEVICE_FUNC explicit variable_if_dynamic(T value) : m_value(value) {} + EIGEN_DEVICE_FUNC T value() const { return m_value; } + EIGEN_DEVICE_FUNC void setValue(T value) { m_value = value; } }; /** \internal like variable_if_dynamic but for DynamicIndex @@ -71,19 +71,19 @@ template class variable_if_dynamicindex { public: EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamicindex) - explicit variable_if_dynamicindex(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); assert(v == T(Value)); } - static T value() { return T(Value); } - void setValue(T) {} + EIGEN_DEVICE_FUNC explicit variable_if_dynamicindex(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); } + EIGEN_DEVICE_FUNC static T value() { return T(Value); } + EIGEN_DEVICE_FUNC void setValue(T) {} }; template class variable_if_dynamicindex { T m_value; - variable_if_dynamicindex() { assert(false); } + EIGEN_DEVICE_FUNC variable_if_dynamicindex() { eigen_assert(false); } public: - explicit variable_if_dynamicindex(T value) : m_value(value) {} - T value() const { return m_value; } - void setValue(T value) { m_value = value; } + EIGEN_DEVICE_FUNC explicit variable_if_dynamicindex(T value) : m_value(value) {} + EIGEN_DEVICE_FUNC T value() const { return m_value; } + EIGEN_DEVICE_FUNC void setValue(T value) { m_value = value; } }; template struct functor_traits @@ -340,6 +340,7 @@ template::type> str }; template +EIGEN_DEVICE_FUNC T* const_cast_ptr(const T* ptr) { return const_cast(ptr); diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index 14a9c402d..bcd30be00 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -727,12 +727,14 @@ struct solve_retval, Rhs> * * \sa class FullPivLU */ +#ifndef __CUDACC__ template inline const FullPivLU::PlainObject> MatrixBase::fullPivLu() const { return FullPivLU(eval()); } +#endif } // end namespace Eigen diff --git a/Eigen/src/LU/Inverse.h b/Eigen/src/LU/Inverse.h index a5ae83bf4..57f9f686c 100644 --- a/Eigen/src/LU/Inverse.h +++ b/Eigen/src/LU/Inverse.h @@ -21,6 +21,7 @@ namespace internal { template struct compute_inverse { + EIGEN_DEVICE_FUNC static inline void run(const MatrixType& matrix, ResultType& result) { result = matrix.partialPivLu().inverse(); @@ -37,6 +38,7 @@ struct compute_inverse_and_det_with_check { /* nothing! general case not support template struct compute_inverse { + EIGEN_DEVICE_FUNC static inline void run(const MatrixType& matrix, ResultType& result) { typedef typename MatrixType::Scalar Scalar; @@ -47,6 +49,7 @@ struct compute_inverse template struct compute_inverse_and_det_with_check { + EIGEN_DEVICE_FUNC static inline void run( const MatrixType& matrix, const typename MatrixType::RealScalar& absDeterminantThreshold, @@ -67,6 +70,7 @@ struct compute_inverse_and_det_with_check ****************************/ template +EIGEN_DEVICE_FUNC inline void compute_inverse_size2_helper( const MatrixType& matrix, const typename ResultType::Scalar& invdet, ResultType& result) @@ -80,6 +84,7 @@ inline void compute_inverse_size2_helper( template struct compute_inverse { + EIGEN_DEVICE_FUNC static inline void run(const MatrixType& matrix, ResultType& result) { typedef typename ResultType::Scalar Scalar; @@ -91,6 +96,7 @@ struct compute_inverse template struct compute_inverse_and_det_with_check { + EIGEN_DEVICE_FUNC static inline void run( const MatrixType& matrix, const typename MatrixType::RealScalar& absDeterminantThreshold, @@ -114,6 +120,7 @@ struct compute_inverse_and_det_with_check ****************************/ template +EIGEN_DEVICE_FUNC inline typename MatrixType::Scalar cofactor_3x3(const MatrixType& m) { enum { @@ -127,6 +134,7 @@ inline typename MatrixType::Scalar cofactor_3x3(const MatrixType& m) } template +EIGEN_DEVICE_FUNC inline void compute_inverse_size3_helper( const MatrixType& matrix, const typename ResultType::Scalar& invdet, @@ -145,6 +153,7 @@ inline void compute_inverse_size3_helper( template struct compute_inverse { + EIGEN_DEVICE_FUNC static inline void run(const MatrixType& matrix, ResultType& result) { typedef typename ResultType::Scalar Scalar; @@ -161,6 +170,7 @@ struct compute_inverse template struct compute_inverse_and_det_with_check { + EIGEN_DEVICE_FUNC static inline void run( const MatrixType& matrix, const typename MatrixType::RealScalar& absDeterminantThreshold, @@ -188,6 +198,7 @@ struct compute_inverse_and_det_with_check ****************************/ template +EIGEN_DEVICE_FUNC inline const typename Derived::Scalar general_det3_helper (const MatrixBase& matrix, int i1, int i2, int i3, int j1, int j2, int j3) { @@ -196,6 +207,7 @@ inline const typename Derived::Scalar general_det3_helper } template +EIGEN_DEVICE_FUNC inline typename MatrixType::Scalar cofactor_4x4(const MatrixType& matrix) { enum { @@ -214,6 +226,7 @@ inline typename MatrixType::Scalar cofactor_4x4(const MatrixType& matrix) template struct compute_inverse_size4 { + EIGEN_DEVICE_FUNC static void run(const MatrixType& matrix, ResultType& result) { result.coeffRef(0,0) = cofactor_4x4(matrix); @@ -246,6 +259,7 @@ struct compute_inverse template struct compute_inverse_and_det_with_check { + EIGEN_DEVICE_FUNC static inline void run( const MatrixType& matrix, const typename MatrixType::RealScalar& absDeterminantThreshold, @@ -279,14 +293,17 @@ struct inverse_impl : public ReturnByValue > typedef typename remove_all::type MatrixTypeNestedCleaned; MatrixTypeNested m_matrix; + EIGEN_DEVICE_FUNC inverse_impl(const MatrixType& matrix) : m_matrix(matrix) {} - inline Index rows() const { return m_matrix.rows(); } - inline Index cols() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } - template inline void evalTo(Dest& dst) const + template + EIGEN_DEVICE_FUNC + inline void evalTo(Dest& dst) const { const int Size = EIGEN_PLAIN_ENUM_MIN(MatrixType::ColsAtCompileTime,Dest::ColsAtCompileTime); EIGEN_ONLY_USED_FOR_DEBUG(Size); diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index c9ff9dd5a..9cf1d61d8 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -469,12 +469,14 @@ struct solve_retval, Rhs> * * \sa class PartialPivLU */ +#ifndef __CUDACC__ template inline const PartialPivLU::PlainObject> MatrixBase::partialPivLu() const { return PartialPivLU(eval()); } +#endif #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS /** \lu_module @@ -485,6 +487,7 @@ MatrixBase::partialPivLu() const * * \sa class PartialPivLU */ +#ifndef __CUDACC__ template inline const PartialPivLU::PlainObject> MatrixBase::lu() const @@ -493,6 +496,8 @@ MatrixBase::lu() const } #endif +#endif + } // end namespace Eigen #endif // EIGEN_PARTIALLU_H diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index a59636790..afb0fb6e3 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -7,6 +7,7 @@ * * \sa abs2() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> abs() const { @@ -20,6 +21,7 @@ abs() const * * \sa abs(), square() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> abs2() const { @@ -33,6 +35,7 @@ abs2() const * * \sa pow(), log(), sin(), cos() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> exp() const { @@ -46,6 +49,7 @@ exp() const * * \sa exp() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> log() const { @@ -59,6 +63,7 @@ log() const * * \sa pow(), square() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> sqrt() const { @@ -72,6 +77,7 @@ sqrt() const * * \sa sin(), acos() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> cos() const { @@ -86,6 +92,7 @@ cos() const * * \sa cos(), asin() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> sin() const { @@ -99,6 +106,7 @@ sin() const * * \sa cos(), asin() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> acos() const { @@ -112,6 +120,7 @@ acos() const * * \sa sin(), acos() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> asin() const { @@ -125,6 +134,7 @@ asin() const * * \sa cos(), sin() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, Derived> tan() const { @@ -139,6 +149,7 @@ tan() const * * \sa exp(), log() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> pow(const Scalar& exponent) const { @@ -154,6 +165,7 @@ pow(const Scalar& exponent) const * * \sa operator/(), operator*() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> inverse() const { @@ -167,6 +179,7 @@ inverse() const * * \sa operator/(), operator*(), abs2() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> square() const { @@ -180,6 +193,7 @@ square() const * * \sa square(), pow() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> cube() const { diff --git a/Eigen/src/plugins/BlockMethods.h b/Eigen/src/plugins/BlockMethods.h index 19a491cf7..5ef373a81 100644 --- a/Eigen/src/plugins/BlockMethods.h +++ b/Eigen/src/plugins/BlockMethods.h @@ -53,12 +53,14 @@ template struct ConstFixedSegmentReturnType { typedef const VectorBloc * * \sa class Block, block(Index,Index) */ +EIGEN_DEVICE_FUNC inline Block block(Index startRow, Index startCol, Index blockRows, Index blockCols) { return Block(derived(), startRow, startCol, blockRows, blockCols); } /** This is the const version of block(Index,Index,Index,Index). */ +EIGEN_DEVICE_FUNC inline const Block block(Index startRow, Index startCol, Index blockRows, Index blockCols) const { return Block(derived(), startRow, startCol, blockRows, blockCols); @@ -77,12 +79,14 @@ inline const Block block(Index startRow, Index startCol, Index bl * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline Block topRightCorner(Index cRows, Index cCols) { return Block(derived(), 0, cols() - cCols, cRows, cCols); } /** This is the const version of topRightCorner(Index, Index).*/ +EIGEN_DEVICE_FUNC inline const Block topRightCorner(Index cRows, Index cCols) const { return Block(derived(), 0, cols() - cCols, cRows, cCols); @@ -98,6 +102,7 @@ inline const Block topRightCorner(Index cRows, Index cCols) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline Block topRightCorner() { return Block(derived(), 0, cols() - CCols); @@ -105,6 +110,7 @@ inline Block topRightCorner() /** This is the const version of topRightCorner().*/ template +EIGEN_DEVICE_FUNC inline const Block topRightCorner() const { return Block(derived(), 0, cols() - CCols); @@ -123,12 +129,14 @@ inline const Block topRightCorner() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline Block topLeftCorner(Index cRows, Index cCols) { return Block(derived(), 0, 0, cRows, cCols); } /** This is the const version of topLeftCorner(Index, Index).*/ +EIGEN_DEVICE_FUNC inline const Block topLeftCorner(Index cRows, Index cCols) const { return Block(derived(), 0, 0, cRows, cCols); @@ -144,6 +152,7 @@ inline const Block topLeftCorner(Index cRows, Index cCols) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline Block topLeftCorner() { return Block(derived(), 0, 0); @@ -151,6 +160,7 @@ inline Block topLeftCorner() /** This is the const version of topLeftCorner().*/ template +EIGEN_DEVICE_FUNC inline const Block topLeftCorner() const { return Block(derived(), 0, 0); @@ -168,12 +178,14 @@ inline const Block topLeftCorner() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline Block bottomRightCorner(Index cRows, Index cCols) { return Block(derived(), rows() - cRows, cols() - cCols, cRows, cCols); } /** This is the const version of bottomRightCorner(Index, Index).*/ +EIGEN_DEVICE_FUNC inline const Block bottomRightCorner(Index cRows, Index cCols) const { return Block(derived(), rows() - cRows, cols() - cCols, cRows, cCols); @@ -189,6 +201,7 @@ inline const Block bottomRightCorner(Index cRows, Index cCols) co * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline Block bottomRightCorner() { return Block(derived(), rows() - CRows, cols() - CCols); @@ -196,6 +209,7 @@ inline Block bottomRightCorner() /** This is the const version of bottomRightCorner().*/ template +EIGEN_DEVICE_FUNC inline const Block bottomRightCorner() const { return Block(derived(), rows() - CRows, cols() - CCols); @@ -213,12 +227,14 @@ inline const Block bottomRightCorner() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline Block bottomLeftCorner(Index cRows, Index cCols) { return Block(derived(), rows() - cRows, 0, cRows, cCols); } /** This is the const version of bottomLeftCorner(Index, Index).*/ +EIGEN_DEVICE_FUNC inline const Block bottomLeftCorner(Index cRows, Index cCols) const { return Block(derived(), rows() - cRows, 0, cRows, cCols); @@ -234,6 +250,7 @@ inline const Block bottomLeftCorner(Index cRows, Index cCols) con * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline Block bottomLeftCorner() { return Block(derived(), rows() - CRows, 0); @@ -241,6 +258,7 @@ inline Block bottomLeftCorner() /** This is the const version of bottomLeftCorner().*/ template +EIGEN_DEVICE_FUNC inline const Block bottomLeftCorner() const { return Block(derived(), rows() - CRows, 0); @@ -257,12 +275,14 @@ inline const Block bottomLeftCorner() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline RowsBlockXpr topRows(Index n) { return RowsBlockXpr(derived(), 0, 0, n, cols()); } /** This is the const version of topRows(Index).*/ +EIGEN_DEVICE_FUNC inline ConstRowsBlockXpr topRows(Index n) const { return ConstRowsBlockXpr(derived(), 0, 0, n, cols()); @@ -278,6 +298,7 @@ inline ConstRowsBlockXpr topRows(Index n) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline typename NRowsBlockXpr::Type topRows() { return typename NRowsBlockXpr::Type(derived(), 0, 0, N, cols()); @@ -285,6 +306,7 @@ inline typename NRowsBlockXpr::Type topRows() /** This is the const version of topRows().*/ template +EIGEN_DEVICE_FUNC inline typename ConstNRowsBlockXpr::Type topRows() const { return typename ConstNRowsBlockXpr::Type(derived(), 0, 0, N, cols()); @@ -301,12 +323,14 @@ inline typename ConstNRowsBlockXpr::Type topRows() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline RowsBlockXpr bottomRows(Index n) { return RowsBlockXpr(derived(), rows() - n, 0, n, cols()); } /** This is the const version of bottomRows(Index).*/ +EIGEN_DEVICE_FUNC inline ConstRowsBlockXpr bottomRows(Index n) const { return ConstRowsBlockXpr(derived(), rows() - n, 0, n, cols()); @@ -322,6 +346,7 @@ inline ConstRowsBlockXpr bottomRows(Index n) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline typename NRowsBlockXpr::Type bottomRows() { return typename NRowsBlockXpr::Type(derived(), rows() - N, 0, N, cols()); @@ -329,6 +354,7 @@ inline typename NRowsBlockXpr::Type bottomRows() /** This is the const version of bottomRows().*/ template +EIGEN_DEVICE_FUNC inline typename ConstNRowsBlockXpr::Type bottomRows() const { return typename ConstNRowsBlockXpr::Type(derived(), rows() - N, 0, N, cols()); @@ -346,12 +372,14 @@ inline typename ConstNRowsBlockXpr::Type bottomRows() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline RowsBlockXpr middleRows(Index startRow, Index numRows) { return RowsBlockXpr(derived(), startRow, 0, numRows, cols()); } /** This is the const version of middleRows(Index,Index).*/ +EIGEN_DEVICE_FUNC inline ConstRowsBlockXpr middleRows(Index startRow, Index numRows) const { return ConstRowsBlockXpr(derived(), startRow, 0, numRows, cols()); @@ -368,6 +396,7 @@ inline ConstRowsBlockXpr middleRows(Index startRow, Index numRows) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline typename NRowsBlockXpr::Type middleRows(Index startRow) { return typename NRowsBlockXpr::Type(derived(), startRow, 0, N, cols()); @@ -375,6 +404,7 @@ inline typename NRowsBlockXpr::Type middleRows(Index startRow) /** This is the const version of middleRows().*/ template +EIGEN_DEVICE_FUNC inline typename ConstNRowsBlockXpr::Type middleRows(Index startRow) const { return typename ConstNRowsBlockXpr::Type(derived(), startRow, 0, N, cols()); @@ -391,12 +421,14 @@ inline typename ConstNRowsBlockXpr::Type middleRows(Index startRow) const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline ColsBlockXpr leftCols(Index n) { return ColsBlockXpr(derived(), 0, 0, rows(), n); } /** This is the const version of leftCols(Index).*/ +EIGEN_DEVICE_FUNC inline ConstColsBlockXpr leftCols(Index n) const { return ConstColsBlockXpr(derived(), 0, 0, rows(), n); @@ -412,6 +444,7 @@ inline ConstColsBlockXpr leftCols(Index n) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline typename NColsBlockXpr::Type leftCols() { return typename NColsBlockXpr::Type(derived(), 0, 0, rows(), N); @@ -419,6 +452,7 @@ inline typename NColsBlockXpr::Type leftCols() /** This is the const version of leftCols().*/ template +EIGEN_DEVICE_FUNC inline typename ConstNColsBlockXpr::Type leftCols() const { return typename ConstNColsBlockXpr::Type(derived(), 0, 0, rows(), N); @@ -435,12 +469,14 @@ inline typename ConstNColsBlockXpr::Type leftCols() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline ColsBlockXpr rightCols(Index n) { return ColsBlockXpr(derived(), 0, cols() - n, rows(), n); } /** This is the const version of rightCols(Index).*/ +EIGEN_DEVICE_FUNC inline ConstColsBlockXpr rightCols(Index n) const { return ConstColsBlockXpr(derived(), 0, cols() - n, rows(), n); @@ -456,6 +492,7 @@ inline ConstColsBlockXpr rightCols(Index n) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline typename NColsBlockXpr::Type rightCols() { return typename NColsBlockXpr::Type(derived(), 0, cols() - N, rows(), N); @@ -463,6 +500,7 @@ inline typename NColsBlockXpr::Type rightCols() /** This is the const version of rightCols().*/ template +EIGEN_DEVICE_FUNC inline typename ConstNColsBlockXpr::Type rightCols() const { return typename ConstNColsBlockXpr::Type(derived(), 0, cols() - N, rows(), N); @@ -480,12 +518,14 @@ inline typename ConstNColsBlockXpr::Type rightCols() const * * \sa class Block, block(Index,Index,Index,Index) */ +EIGEN_DEVICE_FUNC inline ColsBlockXpr middleCols(Index startCol, Index numCols) { return ColsBlockXpr(derived(), 0, startCol, rows(), numCols); } /** This is the const version of middleCols(Index,Index).*/ +EIGEN_DEVICE_FUNC inline ConstColsBlockXpr middleCols(Index startCol, Index numCols) const { return ConstColsBlockXpr(derived(), 0, startCol, rows(), numCols); @@ -502,6 +542,7 @@ inline ConstColsBlockXpr middleCols(Index startCol, Index numCols) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline typename NColsBlockXpr::Type middleCols(Index startCol) { return typename NColsBlockXpr::Type(derived(), 0, startCol, rows(), N); @@ -509,6 +550,7 @@ inline typename NColsBlockXpr::Type middleCols(Index startCol) /** This is the const version of middleCols().*/ template +EIGEN_DEVICE_FUNC inline typename ConstNColsBlockXpr::Type middleCols(Index startCol) const { return typename ConstNColsBlockXpr::Type(derived(), 0, startCol, rows(), N); @@ -533,6 +575,7 @@ inline typename ConstNColsBlockXpr::Type middleCols(Index startCol) const * \sa class Block, block(Index,Index,Index,Index) */ template +EIGEN_DEVICE_FUNC inline Block block(Index startRow, Index startCol) { return Block(derived(), startRow, startCol); @@ -540,6 +583,7 @@ inline Block block(Index startRow, Index startCol /** This is the const version of block<>(Index, Index). */ template +EIGEN_DEVICE_FUNC inline const Block block(Index startRow, Index startCol) const { return Block(derived(), startRow, startCol); @@ -551,12 +595,14 @@ inline const Block block(Index startRow, In * Output: \verbinclude MatrixBase_col.out * * \sa row(), class Block */ +EIGEN_DEVICE_FUNC inline ColXpr col(Index i) { return ColXpr(derived(), i); } /** This is the const version of col(). */ +EIGEN_DEVICE_FUNC inline ConstColXpr col(Index i) const { return ConstColXpr(derived(), i); @@ -568,12 +614,14 @@ inline ConstColXpr col(Index i) const * Output: \verbinclude MatrixBase_row.out * * \sa col(), class Block */ +EIGEN_DEVICE_FUNC inline RowXpr row(Index i) { return RowXpr(derived(), i); } /** This is the const version of row(). */ +EIGEN_DEVICE_FUNC inline ConstRowXpr row(Index i) const { return ConstRowXpr(derived(), i); @@ -595,6 +643,7 @@ inline ConstRowXpr row(Index i) const * * \sa class Block, segment(Index) */ +EIGEN_DEVICE_FUNC inline SegmentReturnType segment(Index start, Index vecSize) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -603,6 +652,7 @@ inline SegmentReturnType segment(Index start, Index vecSize) /** This is the const version of segment(Index,Index).*/ +EIGEN_DEVICE_FUNC inline ConstSegmentReturnType segment(Index start, Index vecSize) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -624,6 +674,7 @@ inline ConstSegmentReturnType segment(Index start, Index vecSize) const * * \sa class Block, block(Index,Index) */ +EIGEN_DEVICE_FUNC inline SegmentReturnType head(Index vecSize) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -631,6 +682,7 @@ inline SegmentReturnType head(Index vecSize) } /** This is the const version of head(Index).*/ +EIGEN_DEVICE_FUNC inline ConstSegmentReturnType head(Index vecSize) const { @@ -653,6 +705,7 @@ inline ConstSegmentReturnType * * \sa class Block, block(Index,Index) */ +EIGEN_DEVICE_FUNC inline SegmentReturnType tail(Index vecSize) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -660,6 +713,7 @@ inline SegmentReturnType tail(Index vecSize) } /** This is the const version of tail(Index).*/ +EIGEN_DEVICE_FUNC inline ConstSegmentReturnType tail(Index vecSize) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -680,6 +734,7 @@ inline ConstSegmentReturnType tail(Index vecSize) const * \sa class Block */ template +EIGEN_DEVICE_FUNC inline typename FixedSegmentReturnType::Type segment(Index start) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -688,6 +743,7 @@ inline typename FixedSegmentReturnType::Type segment(Index start) /** This is the const version of segment(Index).*/ template +EIGEN_DEVICE_FUNC inline typename ConstFixedSegmentReturnType::Type segment(Index start) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -706,6 +762,7 @@ inline typename ConstFixedSegmentReturnType::Type segment(Index start) con * \sa class Block */ template +EIGEN_DEVICE_FUNC inline typename FixedSegmentReturnType::Type head() { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -714,6 +771,7 @@ inline typename FixedSegmentReturnType::Type head() /** This is the const version of head().*/ template +EIGEN_DEVICE_FUNC inline typename ConstFixedSegmentReturnType::Type head() const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -732,6 +790,7 @@ inline typename ConstFixedSegmentReturnType::Type head() const * \sa class Block */ template +EIGEN_DEVICE_FUNC inline typename FixedSegmentReturnType::Type tail() { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) @@ -740,6 +799,7 @@ inline typename FixedSegmentReturnType::Type tail() /** This is the const version of tail.*/ template +EIGEN_DEVICE_FUNC inline typename ConstFixedSegmentReturnType::Type tail() const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) diff --git a/Eigen/src/plugins/CommonCwiseBinaryOps.h b/Eigen/src/plugins/CommonCwiseBinaryOps.h index 688d22440..a8fa287c9 100644 --- a/Eigen/src/plugins/CommonCwiseBinaryOps.h +++ b/Eigen/src/plugins/CommonCwiseBinaryOps.h @@ -38,6 +38,7 @@ EIGEN_MAKE_CWISE_BINARY_OP(operator+,internal::scalar_sum_op) * \sa class CwiseBinaryOp, operator+(), operator-(), cwiseProduct() */ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp binaryExpr(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other, const CustomBinaryOp& func = CustomBinaryOp()) const { diff --git a/Eigen/src/plugins/CommonCwiseUnaryOps.h b/Eigen/src/plugins/CommonCwiseUnaryOps.h index 08e931aad..a17153e64 100644 --- a/Eigen/src/plugins/CommonCwiseUnaryOps.h +++ b/Eigen/src/plugins/CommonCwiseUnaryOps.h @@ -40,11 +40,13 @@ typedef CwiseUnaryView, Derived> NonConstIm /** \returns an expression of the opposite of \c *this */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp::Scalar>, const Derived> operator-() const { return derived(); } /** \returns an expression of \c *this scaled by the scalar factor \a scalar */ +EIGEN_DEVICE_FUNC inline const ScalarMultipleReturnType operator*(const Scalar& scalar) const { @@ -57,6 +59,7 @@ const ScalarMultipleReturnType operator*(const RealScalar& scalar) const; #endif /** \returns an expression of \c *this divided by the scalar value \a scalar */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp::Scalar>, const Derived> operator/(const Scalar& scalar) const { @@ -65,6 +68,7 @@ operator/(const Scalar& scalar) const } /** Overloaded for efficient real matrix times complex scalar value */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp >, const Derived> operator*(const std::complex& scalar) const { @@ -72,10 +76,12 @@ operator*(const std::complex& scalar) const (*static_cast(this), internal::scalar_multiple2_op >(scalar)); } +EIGEN_DEVICE_FUNC inline friend const ScalarMultipleReturnType operator*(const Scalar& scalar, const StorageBaseType& matrix) { return matrix*scalar; } +EIGEN_DEVICE_FUNC inline friend const CwiseUnaryOp >, const Derived> operator*(const std::complex& scalar, const StorageBaseType& matrix) { return matrix*scalar; } @@ -88,6 +94,7 @@ operator*(const std::complex& scalar, const StorageBaseType& matrix) * \sa class CwiseUnaryOp */ template +EIGEN_DEVICE_FUNC typename internal::cast_return_type::Scalar, NewType>, const Derived> >::type cast() const { @@ -97,6 +104,7 @@ cast() const /** \returns an expression of the complex conjugate of \c *this. * * \sa adjoint() */ +EIGEN_DEVICE_FUNC inline ConjugateReturnType conjugate() const { @@ -106,12 +114,14 @@ conjugate() const /** \returns a read-only expression of the real part of \c *this. * * \sa imag() */ +EIGEN_DEVICE_FUNC inline RealReturnType real() const { return derived(); } /** \returns an read-only expression of the imaginary part of \c *this. * * \sa real() */ +EIGEN_DEVICE_FUNC inline const ImagReturnType imag() const { return derived(); } @@ -135,6 +145,7 @@ imag() const { return derived(); } * \sa class CwiseUnaryOp, class CwiseBinaryOp */ template +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp unaryExpr(const CustomUnaryOp& func = CustomUnaryOp()) const { @@ -153,6 +164,7 @@ unaryExpr(const CustomUnaryOp& func = CustomUnaryOp()) const * \sa class CwiseUnaryOp, class CwiseBinaryOp */ template +EIGEN_DEVICE_FUNC inline const CwiseUnaryView unaryViewExpr(const CustomViewOp& func = CustomViewOp()) const { @@ -162,11 +174,13 @@ unaryViewExpr(const CustomViewOp& func = CustomViewOp()) const /** \returns a non const expression of the real part of \c *this. * * \sa imag() */ +EIGEN_DEVICE_FUNC inline NonConstRealReturnType real() { return derived(); } /** \returns a non const expression of the imaginary part of \c *this. * * \sa real() */ +EIGEN_DEVICE_FUNC inline NonConstImagReturnType imag() { return derived(); } diff --git a/Eigen/src/plugins/MatrixCwiseBinaryOps.h b/Eigen/src/plugins/MatrixCwiseBinaryOps.h index 3a737df7b..52e75a174 100644 --- a/Eigen/src/plugins/MatrixCwiseBinaryOps.h +++ b/Eigen/src/plugins/MatrixCwiseBinaryOps.h @@ -18,6 +18,7 @@ * \sa class CwiseBinaryOp, cwiseAbs2 */ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const EIGEN_CWISE_PRODUCT_RETURN_TYPE(Derived,OtherDerived) cwiseProduct(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -37,6 +38,7 @@ cwiseProduct(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const * \sa cwiseNotEqual(), isApprox(), isMuchSmallerThan() */ template +EIGEN_DEVICE_FUNC inline const CwiseBinaryOp, const Derived, const OtherDerived> cwiseEqual(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -56,6 +58,7 @@ cwiseEqual(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const * \sa cwiseEqual(), isApprox(), isMuchSmallerThan() */ template +EIGEN_DEVICE_FUNC inline const CwiseBinaryOp, const Derived, const OtherDerived> cwiseNotEqual(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -70,6 +73,7 @@ cwiseNotEqual(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const * \sa class CwiseBinaryOp, max() */ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const OtherDerived> cwiseMin(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -80,6 +84,7 @@ cwiseMin(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const * * \sa class CwiseBinaryOp, min() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const ConstantReturnType> cwiseMin(const Scalar &other) const { @@ -94,6 +99,7 @@ cwiseMin(const Scalar &other) const * \sa class CwiseBinaryOp, min() */ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const OtherDerived> cwiseMax(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -104,6 +110,7 @@ cwiseMax(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const * * \sa class CwiseBinaryOp, min() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const ConstantReturnType> cwiseMax(const Scalar &other) const { @@ -119,6 +126,7 @@ cwiseMax(const Scalar &other) const * \sa class CwiseBinaryOp, cwiseProduct(), cwiseInverse() */ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const OtherDerived> cwiseQuotient(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { diff --git a/Eigen/src/plugins/MatrixCwiseUnaryOps.h b/Eigen/src/plugins/MatrixCwiseUnaryOps.h index 0cf0640ba..1bb15f862 100644 --- a/Eigen/src/plugins/MatrixCwiseUnaryOps.h +++ b/Eigen/src/plugins/MatrixCwiseUnaryOps.h @@ -17,6 +17,7 @@ * * \sa cwiseAbs2() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> cwiseAbs() const { return derived(); } @@ -27,6 +28,7 @@ cwiseAbs() const { return derived(); } * * \sa cwiseAbs() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> cwiseAbs2() const { return derived(); } @@ -37,6 +39,7 @@ cwiseAbs2() const { return derived(); } * * \sa cwisePow(), cwiseSquare() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> cwiseSqrt() const { return derived(); } @@ -47,6 +50,7 @@ cwiseSqrt() const { return derived(); } * * \sa cwiseProduct() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> cwiseInverse() const { return derived(); } @@ -59,6 +63,7 @@ cwiseInverse() const { return derived(); } * * \sa cwiseEqual(const MatrixBase &) const */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp >, const Derived> cwiseEqual(const Scalar& s) const { From 968f7591f8b4a7d1457575bac22e75465340c3db Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 21 Feb 2013 12:51:58 +0100 Subject: [PATCH 0002/1103] Make it compile without nvcc --- Eigen/Core | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/Core b/Eigen/Core index d8f62c825..5d0a3e785 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -57,6 +57,8 @@ #define EIGEN_NO_DEBUG // All functions callable from CUDA code must be qualified with __device__ #define EIGEN_DEVICE_FUNC __host__ __device__ +#else + #define EIGEN_DEVICE_FUNC #endif #ifndef EIGEN_DONT_VECTORIZE From d93c1c113b9540e7a66d4b2c2dc3c1e99b5d26a1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 21 Feb 2013 19:05:23 +0100 Subject: [PATCH 0003/1103] NVCC: EIGEN_NO_DEBUG must be defined before including Macro.h --- Eigen/Core | 24 ++++++++++++------------ 1 file changed, 12 insertions(+), 12 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 5d0a3e785..dd8da38b5 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -14,6 +14,18 @@ // first thing Eigen does: stop the compiler from committing suicide #include "src/Core/util/DisableStupidWarnings.h" +// Handle NVCC/CUDA +#ifdef __CUDACC__ + // Do not try to vectorize on CUDA! + #define EIGEN_DONT_VECTORIZE + // Do not try asserts on CUDA! + #define EIGEN_NO_DEBUG + // All functions callable from CUDA code must be qualified with __device__ + #define EIGEN_DEVICE_FUNC __host__ __device__ +#else + #define EIGEN_DEVICE_FUNC +#endif + // then include this file where all our macros are defined. It's really important to do it first because // it's where we do all the alignment settings (platform detection and honoring the user's will if he // defined e.g. EIGEN_DONT_ALIGN) so it needs to be done before we do anything with vectorization. @@ -49,18 +61,6 @@ #endif #endif -// Handle NVCC/CUDA -#ifdef __CUDACC__ - // Do not try to vectorize on CUDA! - #define EIGEN_DONT_VECTORIZE - // Do not try asserts on CUDA! - #define EIGEN_NO_DEBUG - // All functions callable from CUDA code must be qualified with __device__ - #define EIGEN_DEVICE_FUNC __host__ __device__ -#else - #define EIGEN_DEVICE_FUNC -#endif - #ifndef EIGEN_DONT_VECTORIZE #if defined (EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC) || defined(EIGEN_SSE2_ON_MSVC_2008_OR_LATER) From 12439e12490b9fd79fe18aa30ce09648dcb644ba Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 5 Apr 2013 16:35:49 +0200 Subject: [PATCH 0004/1103] Port SelfCwiseBinaryOp and Dot.h to nvcc, fix portability issue with std::min/max --- Eigen/Core | 13 ++++++++++-- Eigen/src/Cholesky/LDLT.h | 2 +- Eigen/src/Core/CwiseNullaryOp.h | 6 ++++++ Eigen/src/Core/Dot.h | 6 ++++++ Eigen/src/Core/Functors.h | 20 ++++++++--------- Eigen/src/Core/Fuzzy.h | 2 +- Eigen/src/Core/GenericPacketMath.h | 4 ++-- Eigen/src/Core/Map.h | 1 + Eigen/src/Core/MapBase.h | 2 ++ Eigen/src/Core/MathFunctions.h | 33 +++++++++++++++++++++++++---- Eigen/src/Core/MatrixBase.h | 22 +++++++++---------- Eigen/src/Core/SelfCwiseBinaryOp.h | 21 +++++++++++++----- Eigen/src/Core/StableNorm.h | 6 +++--- Eigen/src/Eigenvalues/EigenSolver.h | 2 +- Eigen/src/Geometry/AngleAxis.h | 4 ++-- Eigen/src/Geometry/Quaternion.h | 2 +- Eigen/src/SVD/JacobiSVD.h | 2 +- 17 files changed, 104 insertions(+), 44 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index dd8da38b5..87f87cd70 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -18,12 +18,21 @@ #ifdef __CUDACC__ // Do not try to vectorize on CUDA! #define EIGEN_DONT_VECTORIZE - // Do not try asserts on CUDA! - #define EIGEN_NO_DEBUG + // All functions callable from CUDA code must be qualified with __device__ #define EIGEN_DEVICE_FUNC __host__ __device__ + #else #define EIGEN_DEVICE_FUNC + +#endif + +#if defined(__CUDA_ARCH__) + // Do not try asserts on CUDA! + #define EIGEN_NO_DEBUG + #define EIGEN_USING_STD_MATH(FUNC) +#else + #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC; #endif // then include this file where all our macros are defined. It's really important to do it first because diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 219a0e8b9..5dede8081 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -501,7 +501,7 @@ struct solve_retval, Rhs> // dst = D^-1 (L^-1 P b) // more precisely, use pseudo-inverse of D (see bug 241) using std::abs; - using std::max; + EIGEN_USING_STD_MATH(max); typedef typename LDLTType::MatrixType MatrixType; typedef typename LDLTType::Scalar Scalar; typedef typename LDLTType::RealScalar RealScalar; diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index a93bab2d0..14c78cb85 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -54,6 +54,7 @@ class CwiseNullaryOp : internal::no_assignment_operator, typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(CwiseNullaryOp) + EIGEN_DEVICE_FUNC CwiseNullaryOp(Index nbRows, Index nbCols, const NullaryOp& func = NullaryOp()) : m_rows(nbRows), m_cols(nbCols), m_functor(func) { @@ -63,9 +64,12 @@ class CwiseNullaryOp : internal::no_assignment_operator, && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == nbCols)); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { return m_rows.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_cols.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const { return m_functor(rowId, colId); @@ -77,6 +81,7 @@ class CwiseNullaryOp : internal::no_assignment_operator, return m_functor.packetOp(rowId, colId); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index index) const { return m_functor(index); @@ -89,6 +94,7 @@ class CwiseNullaryOp : internal::no_assignment_operator, } /** \returns the functor representing the nullary operation */ + EIGEN_DEVICE_FUNC const NullaryOp& functor() const { return m_functor; } protected: diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h index 9d513d699..6b058d3d9 100644 --- a/Eigen/src/Core/Dot.h +++ b/Eigen/src/Core/Dot.h @@ -59,6 +59,7 @@ struct dot_nocheck */ template template +EIGEN_DEVICE_FUNC typename internal::scalar_product_traits::Scalar,typename internal::traits::Scalar>::ReturnType MatrixBase::dot(const MatrixBase& other) const { @@ -151,6 +152,7 @@ MatrixBase::normalized() const * \sa norm(), normalized() */ template +EIGEN_DEVICE_FUNC inline void MatrixBase::normalize() { *this /= norm(); @@ -164,6 +166,7 @@ template struct lpNorm_selector { typedef typename NumTraits::Scalar>::Real RealScalar; + EIGEN_DEVICE_FUNC static inline RealScalar run(const MatrixBase& m) { return pow(m.cwiseAbs().array().pow(p).sum(), RealScalar(1)/p); @@ -173,6 +176,7 @@ struct lpNorm_selector template struct lpNorm_selector { + EIGEN_DEVICE_FUNC static inline typename NumTraits::Scalar>::Real run(const MatrixBase& m) { return m.cwiseAbs().sum(); @@ -182,6 +186,7 @@ struct lpNorm_selector template struct lpNorm_selector { + EIGEN_DEVICE_FUNC static inline typename NumTraits::Scalar>::Real run(const MatrixBase& m) { return m.norm(); @@ -191,6 +196,7 @@ struct lpNorm_selector template struct lpNorm_selector { + EIGEN_DEVICE_FUNC static inline typename NumTraits::Scalar>::Real run(const MatrixBase& m) { return m.cwiseAbs().maxCoeff(); diff --git a/Eigen/src/Core/Functors.h b/Eigen/src/Core/Functors.h index 6b6b36656..29cd739ce 100644 --- a/Eigen/src/Core/Functors.h +++ b/Eigen/src/Core/Functors.h @@ -103,7 +103,7 @@ struct functor_traits > { */ template struct scalar_min_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::min; return (min)(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(min); return (min)(a, b); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pmin(a,b); } @@ -126,7 +126,7 @@ struct functor_traits > { */ template struct scalar_max_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { using std::max; return (max)(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(max); return (max)(a, b); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pmax(a,b); } @@ -152,8 +152,8 @@ template struct scalar_hypot_op { // typedef typename NumTraits::Real result_type; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const { - using std::max; - using std::min; + EIGEN_USING_STD_MATH(max); + EIGEN_USING_STD_MATH(min); Scalar p = (max)(_x, _y); Scalar q = (min)(_x, _y); Scalar qp = q/p; @@ -479,8 +479,8 @@ struct functor_traits > template struct scalar_multiple2_op { typedef typename scalar_product_traits::ReturnType result_type; - EIGEN_STRONG_INLINE scalar_multiple2_op(const scalar_multiple2_op& other) : m_other(other.m_other) { } - EIGEN_STRONG_INLINE scalar_multiple2_op(const Scalar2& other) : m_other(other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple2_op(const scalar_multiple2_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple2_op(const Scalar2& other) : m_other(other) { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar1& a) const { return a * m_other; } typename add_const_on_value_type::Nested>::type m_other; }; @@ -500,8 +500,8 @@ template struct scalar_quotient1_op { typedef typename packet_traits::type Packet; // FIXME default copy constructors seems bugged with std::complex<> - EIGEN_STRONG_INLINE scalar_quotient1_op(const scalar_quotient1_op& other) : m_other(other.m_other) { } - EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_quotient1_op(const scalar_quotient1_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {} EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; } EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pdiv(a, pset1(m_other)); } @@ -516,8 +516,8 @@ struct functor_traits > template struct scalar_constant_op { typedef typename packet_traits::type Packet; - EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { } - EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; } template diff --git a/Eigen/src/Core/Fuzzy.h b/Eigen/src/Core/Fuzzy.h index 8fb9a01dd..9581c81bd 100644 --- a/Eigen/src/Core/Fuzzy.h +++ b/Eigen/src/Core/Fuzzy.h @@ -21,7 +21,7 @@ struct isApprox_selector { static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) { - using std::min; + EIGEN_USING_STD_MATH(min); typename internal::nested::type nested(x); typename internal::nested::type otherNested(y); return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum()); diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 51913fa5c..aee58f09b 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -121,12 +121,12 @@ pdiv(const Packet& a, /** \internal \returns the min of \a a and \a b (coeff-wise) */ template inline Packet pmin(const Packet& a, - const Packet& b) { using std::min; return (min)(a, b); } + const Packet& b) { EIGEN_USING_STD_MATH(min); return (min)(a, b); } /** \internal \returns the max of \a a and \a b (coeff-wise) */ template inline Packet pmax(const Packet& a, - const Packet& b) { using std::max; return (max)(a, b); } + const Packet& b) { EIGEN_USING_STD_MATH(max); return (max)(a, b); } /** \internal \returns the absolute value of \a a */ template inline Packet diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index e054f2202..8ea13cfb7 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -115,6 +115,7 @@ template class Ma inline PointerType cast_to_pointer_type(PointerArgType ptr) { return const_cast(ptr); } #else typedef PointerType PointerArgType; + EIGEN_DEVICE_FUNC inline PointerType cast_to_pointer_type(PointerArgType ptr) { return ptr; } #endif diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index 8def7442d..ffa1371c2 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -203,7 +203,9 @@ template class MapBase const Scalar >::type ScalarWithConstIfNotLvalue; + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return this->m_data; } + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return this->m_data; } // no const-cast here so non-const-correct code will give a compile error EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 4ce616746..b8d32ecec 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -63,6 +63,7 @@ template struct real_impl { typedef typename NumTraits::Real RealScalar; + EIGEN_DEVICE_FUNC static inline RealScalar run(const Scalar& x) { return x; @@ -72,6 +73,7 @@ struct real_impl template struct real_impl > { + EIGEN_DEVICE_FUNC static inline RealScalar run(const std::complex& x) { using std::real; @@ -86,6 +88,7 @@ struct real_retval }; template +EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(real, Scalar) real(const Scalar& x) { return EIGEN_MATHFUNC_IMPL(real, Scalar)::run(x); @@ -99,6 +102,7 @@ template struct imag_impl { typedef typename NumTraits::Real RealScalar; + EIGEN_DEVICE_FUNC static inline RealScalar run(const Scalar&) { return RealScalar(0); @@ -108,6 +112,7 @@ struct imag_impl template struct imag_impl > { + EIGEN_DEVICE_FUNC static inline RealScalar run(const std::complex& x) { using std::imag; @@ -122,6 +127,7 @@ struct imag_retval }; template +EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(imag, Scalar) imag(const Scalar& x) { return EIGEN_MATHFUNC_IMPL(imag, Scalar)::run(x); @@ -135,10 +141,12 @@ template struct real_ref_impl { typedef typename NumTraits::Real RealScalar; + EIGEN_DEVICE_FUNC static inline RealScalar& run(Scalar& x) { return reinterpret_cast(&x)[0]; } + EIGEN_DEVICE_FUNC static inline const RealScalar& run(const Scalar& x) { return reinterpret_cast(&x)[0]; @@ -152,12 +160,14 @@ struct real_ref_retval }; template +EIGEN_DEVICE_FUNC inline typename add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) >::type real_ref(const Scalar& x) { return real_ref_impl::run(x); } template +EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) real_ref(Scalar& x) { return EIGEN_MATHFUNC_IMPL(real_ref, Scalar)::run(x); @@ -171,10 +181,12 @@ template struct imag_ref_default_impl { typedef typename NumTraits::Real RealScalar; + EIGEN_DEVICE_FUNC static inline RealScalar& run(Scalar& x) { return reinterpret_cast(&x)[1]; } + EIGEN_DEVICE_FUNC static inline const RealScalar& run(const Scalar& x) { return reinterpret_cast(&x)[1]; @@ -184,10 +196,12 @@ struct imag_ref_default_impl template struct imag_ref_default_impl { + EIGEN_DEVICE_FUNC static inline Scalar run(Scalar&) { return Scalar(0); } + EIGEN_DEVICE_FUNC static inline const Scalar run(const Scalar&) { return Scalar(0); @@ -204,12 +218,14 @@ struct imag_ref_retval }; template +EIGEN_DEVICE_FUNC inline typename add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) >::type imag_ref(const Scalar& x) { return imag_ref_impl::run(x); } template +EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) imag_ref(Scalar& x) { return EIGEN_MATHFUNC_IMPL(imag_ref, Scalar)::run(x); @@ -222,6 +238,7 @@ inline EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) imag_ref(Scalar& x) template struct conj_impl { + EIGEN_DEVICE_FUNC static inline Scalar run(const Scalar& x) { return x; @@ -231,6 +248,7 @@ struct conj_impl template struct conj_impl > { + EIGEN_DEVICE_FUNC static inline std::complex run(const std::complex& x) { using std::conj; @@ -245,6 +263,7 @@ struct conj_retval }; template +EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(conj, Scalar) conj(const Scalar& x) { return EIGEN_MATHFUNC_IMPL(conj, Scalar)::run(x); @@ -258,6 +277,7 @@ template struct abs2_impl { typedef typename NumTraits::Real RealScalar; + EIGEN_DEVICE_FUNC static inline RealScalar run(const Scalar& x) { return x*x; @@ -267,6 +287,7 @@ struct abs2_impl template struct abs2_impl > { + EIGEN_DEVICE_FUNC static inline RealScalar run(const std::complex& x) { return real(x)*real(x) + imag(x)*imag(x); @@ -280,6 +301,7 @@ struct abs2_retval }; template +EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(abs2, Scalar) abs2(const Scalar& x) { return EIGEN_MATHFUNC_IMPL(abs2, Scalar)::run(x); @@ -293,6 +315,7 @@ template struct norm1_default_impl { typedef typename NumTraits::Real RealScalar; + EIGEN_DEVICE_FUNC static inline RealScalar run(const Scalar& x) { using std::abs; @@ -303,6 +326,7 @@ struct norm1_default_impl template struct norm1_default_impl { + EIGEN_DEVICE_FUNC static inline Scalar run(const Scalar& x) { using std::abs; @@ -320,6 +344,7 @@ struct norm1_retval }; template +EIGEN_DEVICE_FUNC inline EIGEN_MATHFUNC_RETVAL(norm1, Scalar) norm1(const Scalar& x) { return EIGEN_MATHFUNC_IMPL(norm1, Scalar)::run(x); @@ -335,8 +360,8 @@ struct hypot_impl typedef typename NumTraits::Real RealScalar; static inline RealScalar run(const Scalar& x, const Scalar& y) { - using std::max; - using std::min; + EIGEN_USING_STD_MATH(max); + EIGEN_USING_STD_MATH(min); using std::abs; RealScalar _x = abs(x); RealScalar _y = abs(y); @@ -631,7 +656,7 @@ struct scalar_fuzzy_default_impl } static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec) { - using std::min; + EIGEN_USING_STD_MATH(min); using std::abs; return abs(x - y) <= (min)(abs(x), abs(y)) * prec; } @@ -671,7 +696,7 @@ struct scalar_fuzzy_default_impl } static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec) { - using std::min; + EIGEN_USING_STD_MATH(min); return abs2(x - y) <= (min)(abs2(x), abs2(y)) * prec * prec; } }; diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index c76192e46..7b9ea8c0a 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -219,16 +219,16 @@ template class MatrixBase Scalar eigen2_dot(const MatrixBase& other) const; #endif - RealScalar squaredNorm() const; - RealScalar norm() const; + EIGEN_DEVICE_FUNC RealScalar squaredNorm() const; + EIGEN_DEVICE_FUNC RealScalar norm() const; RealScalar stableNorm() const; RealScalar blueNorm() const; RealScalar hypotNorm() const; - const PlainObject normalized() const; - void normalize(); + EIGEN_DEVICE_FUNC const PlainObject normalized() const; + EIGEN_DEVICE_FUNC void normalize(); - const AdjointReturnType adjoint() const; - void adjointInPlace(); + EIGEN_DEVICE_FUNC const AdjointReturnType adjoint() const; + EIGEN_DEVICE_FUNC void adjointInPlace(); typedef Diagonal DiagonalReturnType; DiagonalReturnType diagonal(); @@ -329,15 +329,15 @@ template class MatrixBase /////////// Array module /////////// - template RealScalar lpNorm() const; + template EIGEN_DEVICE_FUNC RealScalar lpNorm() const; - MatrixBase& matrix() { return *this; } - const MatrixBase& matrix() const { return *this; } + EIGEN_DEVICE_FUNC MatrixBase& matrix() { return *this; } + EIGEN_DEVICE_FUNC const MatrixBase& matrix() const { return *this; } /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix * \sa ArrayBase::matrix() */ - ArrayWrapper array() { return derived(); } - const ArrayWrapper array() const { return derived(); } + EIGEN_DEVICE_FUNC ArrayWrapper array() { return derived(); } + EIGEN_DEVICE_FUNC const ArrayWrapper array() const { return derived(); } /////////// LU module /////////// diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h index 0caf2bab1..40f9eb618 100644 --- a/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -52,21 +52,24 @@ template class SelfCwiseBinaryOp typedef typename internal::packet_traits::type Packet; + EIGEN_DEVICE_FUNC inline SelfCwiseBinaryOp(Lhs& xpr, const BinaryOp& func = BinaryOp()) : m_matrix(xpr), m_functor(func) {} - inline Index rows() const { return m_matrix.rows(); } - inline Index cols() const { return m_matrix.cols(); } - inline Index outerStride() const { return m_matrix.outerStride(); } - inline Index innerStride() const { return m_matrix.innerStride(); } - inline const Scalar* data() const { return m_matrix.data(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_matrix.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.innerStride(); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_matrix.data(); } // note that this function is needed by assign to correctly align loads/stores // TODO make Assign use .data() + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { EIGEN_STATIC_ASSERT_LVALUE(Lhs) return m_matrix.const_cast_derived().coeffRef(row, col); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index row, Index col) const { return m_matrix.coeffRef(row, col); @@ -74,17 +77,20 @@ template class SelfCwiseBinaryOp // note that this function is needed by assign to correctly align loads/stores // TODO make Assign use .data() + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { EIGEN_STATIC_ASSERT_LVALUE(Lhs) return m_matrix.const_cast_derived().coeffRef(index); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index index) const { return m_matrix.const_cast_derived().coeffRef(index); } template + EIGEN_DEVICE_FUNC void copyCoeff(Index row, Index col, const DenseBase& other) { OtherDerived& _other = other.const_cast_derived(); @@ -95,6 +101,7 @@ template class SelfCwiseBinaryOp } template + EIGEN_DEVICE_FUNC void copyCoeff(Index index, const DenseBase& other) { OtherDerived& _other = other.const_cast_derived(); @@ -125,6 +132,7 @@ template class SelfCwiseBinaryOp // reimplement lazyAssign to handle complex *= real // see CwiseBinaryOp ctor for details template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE SelfCwiseBinaryOp& lazyAssign(const DenseBase& rhs) { EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs,RhsDerived) @@ -144,17 +152,20 @@ template class SelfCwiseBinaryOp // overloaded to honor evaluation of special matrices // maybe another solution would be to not use SelfCwiseBinaryOp // at first... + EIGEN_DEVICE_FUNC SelfCwiseBinaryOp& operator=(const Rhs& _rhs) { typename internal::nested::type rhs(_rhs); return Base::operator=(rhs); } + EIGEN_DEVICE_FUNC Lhs& expression() const { return m_matrix; } + EIGEN_DEVICE_FUNC const BinaryOp& functor() const { return m_functor; diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h index 35626be21..1aefd91a8 100644 --- a/Eigen/src/Core/StableNorm.h +++ b/Eigen/src/Core/StableNorm.h @@ -36,8 +36,8 @@ blueNorm_impl(const EigenBase& _vec) typedef typename Derived::RealScalar RealScalar; typedef typename Derived::Index Index; using std::pow; - using std::min; - using std::max; + EIGEN_USING_STD_MATH(min); + EIGEN_USING_STD_MATH(max); using std::sqrt; using std::abs; const Derived& vec(_vec.derived()); @@ -141,7 +141,7 @@ template inline typename NumTraits::Scalar>::Real MatrixBase::stableNorm() const { - using std::min; + EIGEN_USING_STD_MATH(min); using std::sqrt; const Index blockSize = 4096; RealScalar scale(0); diff --git a/Eigen/src/Eigenvalues/EigenSolver.h b/Eigen/src/Eigenvalues/EigenSolver.h index 43d0ffa76..201ea620d 100644 --- a/Eigen/src/Eigenvalues/EigenSolver.h +++ b/Eigen/src/Eigenvalues/EigenSolver.h @@ -568,7 +568,7 @@ void EigenSolver::doComputeEigenvectors() } // Overflow control - using std::max; + EIGEN_USING_STD_MATH(max); Scalar t = (max)(abs(m_matT.coeff(i,n-1)),abs(m_matT.coeff(i,n))); if ((eps * t) * t > Scalar(1)) m_matT.block(i, n-1, size-i, 2) /= t; diff --git a/Eigen/src/Geometry/AngleAxis.h b/Eigen/src/Geometry/AngleAxis.h index 553d38c74..f424e6d7d 100644 --- a/Eigen/src/Geometry/AngleAxis.h +++ b/Eigen/src/Geometry/AngleAxis.h @@ -159,8 +159,8 @@ template AngleAxis& AngleAxis::operator=(const QuaternionBase& q) { using std::acos; - using std::min; - using std::max; + EIGEN_USING_STD_MATH(min); + EIGEN_USING_STD_MATH(max); using std::sqrt; Scalar n2 = q.vec().squaredNorm(); if (n2 < NumTraits::dummy_precision()*NumTraits::dummy_precision()) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index e7801abf9..45824fbeb 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -572,7 +572,7 @@ template template inline Derived& QuaternionBase::setFromTwoVectors(const MatrixBase& a, const MatrixBase& b) { - using std::max; + EIGEN_USING_STD_MATH(max); using std::sqrt; Vector3 v0 = a.normalized(); Vector3 v1 = b.normalized(); diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 5741a8ba1..1616fe560 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -766,7 +766,7 @@ JacobiSVD::compute(const MatrixType& matrix, unsig // if this 2x2 sub-matrix is not diagonal already... // notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't // keep us iterating forever. Similarly, small denormal numbers are considered zero. - using std::max; + EIGEN_USING_STD_MATH(max); RealScalar threshold = (max)(considerAsZero, precision * (max)(abs(m_workMatrix.coeff(p,p)), abs(m_workMatrix.coeff(q,q)))); if((max)(abs(m_workMatrix.coeff(p,q)),abs(m_workMatrix.coeff(q,p))) > threshold) From 4e2e615a7c2c719d2d708ab32840bad353322d8c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 19 Apr 2013 11:14:17 +0200 Subject: [PATCH 0005/1103] actually assertion are incompatible with nvcc even on host code --- Eigen/Core | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 87f87cd70..26a37cf42 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -16,6 +16,8 @@ // Handle NVCC/CUDA #ifdef __CUDACC__ + // Do not try asserts on CUDA! + #define EIGEN_NO_DEBUG // Do not try to vectorize on CUDA! #define EIGEN_DONT_VECTORIZE @@ -28,8 +30,6 @@ #endif #if defined(__CUDA_ARCH__) - // Do not try asserts on CUDA! - #define EIGEN_NO_DEBUG #define EIGEN_USING_STD_MATH(FUNC) #else #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC; From 8556ca3de5c78f1b5b23b848b78d97bf8a1adea1 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Wed, 15 May 2013 13:45:24 +0200 Subject: [PATCH 0006/1103] Adapted settings for the eol extension. --- .hgeol | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.hgeol b/.hgeol index 423676d31..5af3cc043 100644 --- a/.hgeol +++ b/.hgeol @@ -1,6 +1,8 @@ [patterns] +*.MINPACK = CRLF scripts/*.in = LF debug/msvc/*.dat = CRLF +debug/msvc/*.natvis = CRLF unsupported/test/mpreal/*.* = CRLF ** = native From 12e69ec89660166759ea817f35fe95fb6fea9cc1 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Wed, 15 May 2013 12:05:01 +0200 Subject: [PATCH 0007/1103] Added asserts to AngleAxis class which verify that the initial axis is normalized. --- Eigen/src/Eigen2Support/Geometry/AngleAxis.h | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Eigen2Support/Geometry/AngleAxis.h b/Eigen/src/Eigen2Support/Geometry/AngleAxis.h index af598a403..9f48c460b 100644 --- a/Eigen/src/Eigen2Support/Geometry/AngleAxis.h +++ b/Eigen/src/Eigen2Support/Geometry/AngleAxis.h @@ -67,12 +67,21 @@ public: /** Default constructor without initialization. */ AngleAxis() {} + /** Constructs and initialize the angle-axis rotation from an \a angle in radian * and an \a axis which must be normalized. */ template - inline AngleAxis(Scalar angle, const MatrixBase& axis) : m_axis(axis), m_angle(angle) {} + inline AngleAxis(Scalar angle, const MatrixBase& axis) : m_axis(axis), m_angle(angle) + { + using std::sqrt; + using std::abs; + // since we compare against 1, this is equal to computing the relative error + eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( dummy_precision() ) ); + } + /** Constructs and initialize the angle-axis rotation from a quaternion \a q. */ inline AngleAxis(const QuaternionType& q) { *this = q; } + /** Constructs and initialize the angle-axis rotation from a 3x3 rotation matrix. */ template inline explicit AngleAxis(const MatrixBase& m) { *this = m; } @@ -167,6 +176,11 @@ AngleAxis& AngleAxis::operator=(const QuaternionType& q) { m_angle = 2*std::acos(q.w()); m_axis = q.vec() / ei_sqrt(n2); + + using std::sqrt; + using std::abs; + // since we compare against 1, this is equal to computing the relative error + eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( dummy_precision() ) ); } return *this; } From 64054ee3964187853b941fbf4d5e3095c482bebb Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 5 Jun 2013 15:38:33 +0200 Subject: [PATCH 0008/1103] Add nvcc support for normalize, initializers, and fuzzy comparisons --- Eigen/src/Core/DenseBase.h | 85 ++++++++++++++++-------------- Eigen/src/Core/Dot.h | 3 +- Eigen/src/Core/Fuzzy.h | 6 +++ Eigen/src/Core/GenericPacketMath.h | 61 +++++++++++---------- Eigen/src/Core/MathFunctions.h | 18 ++++--- Eigen/src/Core/NumTraits.h | 5 ++ Eigen/src/Core/PlainObjectBase.h | 12 ++--- Eigen/src/Core/util/BlasUtil.h | 8 +-- 8 files changed, 113 insertions(+), 85 deletions(-) diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index ec646ce7d..6a6ba2954 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -299,65 +299,68 @@ template class DenseBase #endif - static const ConstantReturnType + EIGEN_DEVICE_FUNC static const ConstantReturnType Constant(Index rows, Index cols, const Scalar& value); - static const ConstantReturnType + EIGEN_DEVICE_FUNC static const ConstantReturnType Constant(Index size, const Scalar& value); - static const ConstantReturnType + EIGEN_DEVICE_FUNC static const ConstantReturnType Constant(const Scalar& value); - static const SequentialLinSpacedReturnType + EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high); - static const RandomAccessLinSpacedReturnType + EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType LinSpaced(Index size, const Scalar& low, const Scalar& high); - static const SequentialLinSpacedReturnType + EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType LinSpaced(Sequential_t, const Scalar& low, const Scalar& high); - static const RandomAccessLinSpacedReturnType + EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType LinSpaced(const Scalar& low, const Scalar& high); - template + template EIGEN_DEVICE_FUNC static const CwiseNullaryOp NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func); - template + template EIGEN_DEVICE_FUNC static const CwiseNullaryOp NullaryExpr(Index size, const CustomNullaryOp& func); - template + template EIGEN_DEVICE_FUNC static const CwiseNullaryOp NullaryExpr(const CustomNullaryOp& func); - static const ConstantReturnType Zero(Index rows, Index cols); - static const ConstantReturnType Zero(Index size); - static const ConstantReturnType Zero(); - static const ConstantReturnType Ones(Index rows, Index cols); - static const ConstantReturnType Ones(Index size); - static const ConstantReturnType Ones(); + EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index rows, Index cols); + EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index size); + EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(); + EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index rows, Index cols); + EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index size); + EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(); - void fill(const Scalar& value); - Derived& setConstant(const Scalar& value); - Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high); - Derived& setLinSpaced(const Scalar& low, const Scalar& high); - Derived& setZero(); - Derived& setOnes(); - Derived& setRandom(); + EIGEN_DEVICE_FUNC void fill(const Scalar& value); + EIGEN_DEVICE_FUNC Derived& setConstant(const Scalar& value); + EIGEN_DEVICE_FUNC Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high); + EIGEN_DEVICE_FUNC Derived& setLinSpaced(const Scalar& low, const Scalar& high); + EIGEN_DEVICE_FUNC Derived& setZero(); + EIGEN_DEVICE_FUNC Derived& setOnes(); + EIGEN_DEVICE_FUNC Derived& setRandom(); - template + template EIGEN_DEVICE_FUNC bool isApprox(const DenseBase& other, const RealScalar& prec = NumTraits::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isMuchSmallerThan(const RealScalar& other, const RealScalar& prec = NumTraits::dummy_precision()) const; - template + template EIGEN_DEVICE_FUNC bool isMuchSmallerThan(const DenseBase& other, const RealScalar& prec = NumTraits::dummy_precision()) const; - bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits::dummy_precision()) const; - bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits::dummy_precision()) const; - bool isZero(const RealScalar& prec = NumTraits::dummy_precision()) const; - bool isOnes(const RealScalar& prec = NumTraits::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isZero(const RealScalar& prec = NumTraits::dummy_precision()) const; + EIGEN_DEVICE_FUNC bool isOnes(const RealScalar& prec = NumTraits::dummy_precision()) const; inline bool hasNaN() const; inline bool isFinite() const; + EIGEN_DEVICE_FUNC inline Derived& operator*=(const Scalar& other); + EIGEN_DEVICE_FUNC inline Derived& operator/=(const Scalar& other); typedef typename internal::add_const_on_value_type::type>::type EvalReturnType; @@ -378,7 +381,7 @@ template class DenseBase /** swaps *this with the expression \a other. * */ - template + template EIGEN_DEVICE_FUNC void swap(const DenseBase& other, int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase) { @@ -388,18 +391,20 @@ template class DenseBase /** swaps *this with the matrix or array \a other. * */ - template + template EIGEN_DEVICE_FUNC void swap(PlainObjectBase& other) { SwapWrapper(derived()).lazyAssign(other.derived()); } - inline const NestByValue nestByValue() const; - inline const ForceAlignedAccess forceAlignedAccess() const; - inline ForceAlignedAccess forceAlignedAccess(); - template inline const typename internal::conditional,Derived&>::type forceAlignedAccessIf() const; - template inline typename internal::conditional,Derived&>::type forceAlignedAccessIf(); + EIGEN_DEVICE_FUNC inline const NestByValue nestByValue() const; + EIGEN_DEVICE_FUNC inline const ForceAlignedAccess forceAlignedAccess() const; + EIGEN_DEVICE_FUNC inline ForceAlignedAccess forceAlignedAccess(); + template EIGEN_DEVICE_FUNC + inline const typename internal::conditional,Derived&>::type forceAlignedAccessIf() const; + template EIGEN_DEVICE_FUNC + inline typename internal::conditional,Derived&>::type forceAlignedAccessIf(); EIGEN_DEVICE_FUNC Scalar sum() const; EIGEN_DEVICE_FUNC Scalar mean() const; @@ -410,13 +415,13 @@ template class DenseBase EIGEN_DEVICE_FUNC typename internal::traits::Scalar minCoeff() const; EIGEN_DEVICE_FUNC typename internal::traits::Scalar maxCoeff() const; - template + template EIGEN_DEVICE_FUNC typename internal::traits::Scalar minCoeff(IndexType* row, IndexType* col) const; - template + template EIGEN_DEVICE_FUNC typename internal::traits::Scalar maxCoeff(IndexType* row, IndexType* col) const; - template + template EIGEN_DEVICE_FUNC typename internal::traits::Scalar minCoeff(IndexType* index) const; - template + template EIGEN_DEVICE_FUNC typename internal::traits::Scalar maxCoeff(IndexType* index) const; template diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h index 6b058d3d9..dacd71f5f 100644 --- a/Eigen/src/Core/Dot.h +++ b/Eigen/src/Core/Dot.h @@ -29,6 +29,7 @@ template::Scalar,typename traits::Scalar>::ReturnType ResScalar; + EIGEN_DEVICE_FUNC static inline ResScalar run(const MatrixBase& a, const MatrixBase& b) { return a.template binaryExpr::Scalar,typename traits::Scalar> >(b).sum(); @@ -39,6 +40,7 @@ template struct dot_nocheck { typedef typename scalar_product_traits::Scalar,typename traits::Scalar>::ReturnType ResScalar; + EIGEN_DEVICE_FUNC static inline ResScalar run(const MatrixBase& a, const MatrixBase& b) { return a.transpose().template binaryExpr::Scalar,typename traits::Scalar> >(b).sum(); @@ -152,7 +154,6 @@ MatrixBase::normalized() const * \sa norm(), normalized() */ template -EIGEN_DEVICE_FUNC inline void MatrixBase::normalize() { *this /= norm(); diff --git a/Eigen/src/Core/Fuzzy.h b/Eigen/src/Core/Fuzzy.h index 9581c81bd..a7c91d302 100644 --- a/Eigen/src/Core/Fuzzy.h +++ b/Eigen/src/Core/Fuzzy.h @@ -19,6 +19,7 @@ namespace internal template::IsInteger> struct isApprox_selector { + EIGEN_DEVICE_FUNC static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) { EIGEN_USING_STD_MATH(min); @@ -31,6 +32,7 @@ struct isApprox_selector template struct isApprox_selector { + EIGEN_DEVICE_FUNC static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar&) { return x.matrix() == y.matrix(); @@ -40,6 +42,7 @@ struct isApprox_selector template::IsInteger> struct isMuchSmallerThan_object_selector { + EIGEN_DEVICE_FUNC static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) { return x.cwiseAbs2().sum() <= abs2(prec) * y.cwiseAbs2().sum(); @@ -49,6 +52,7 @@ struct isMuchSmallerThan_object_selector template struct isMuchSmallerThan_object_selector { + EIGEN_DEVICE_FUNC static bool run(const Derived& x, const OtherDerived&, const typename Derived::RealScalar&) { return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix(); @@ -58,6 +62,7 @@ struct isMuchSmallerThan_object_selector template::IsInteger> struct isMuchSmallerThan_scalar_selector { + EIGEN_DEVICE_FUNC static bool run(const Derived& x, const typename Derived::RealScalar& y, const typename Derived::RealScalar& prec) { return x.cwiseAbs2().sum() <= abs2(prec * y); @@ -67,6 +72,7 @@ struct isMuchSmallerThan_scalar_selector template struct isMuchSmallerThan_scalar_selector { + EIGEN_DEVICE_FUNC static bool run(const Derived& x, const typename Derived::RealScalar&, const typename Derived::RealScalar&) { return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix(); diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 967a37673..17b7ae87d 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -91,69 +91,69 @@ template struct packet_traits : default_packet_traits }; /** \internal \returns a + b (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet padd(const Packet& a, const Packet& b) { return a+b; } /** \internal \returns a - b (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet psub(const Packet& a, const Packet& b) { return a-b; } /** \internal \returns -a (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pnegate(const Packet& a) { return -a; } /** \internal \returns conj(a) (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pconj(const Packet& a) { return conj(a); } /** \internal \returns a * b (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pmul(const Packet& a, const Packet& b) { return a*b; } /** \internal \returns a / b (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pdiv(const Packet& a, const Packet& b) { return a/b; } /** \internal \returns the min of \a a and \a b (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pmin(const Packet& a, const Packet& b) { EIGEN_USING_STD_MATH(min); return (min)(a, b); } /** \internal \returns the max of \a a and \a b (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pmax(const Packet& a, const Packet& b) { EIGEN_USING_STD_MATH(max); return (max)(a, b); } /** \internal \returns the absolute value of \a a */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pabs(const Packet& a) { using std::abs; return abs(a); } /** \internal \returns the bitwise and of \a a and \a b */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pand(const Packet& a, const Packet& b) { return a & b; } /** \internal \returns the bitwise or of \a a and \a b */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet por(const Packet& a, const Packet& b) { return a | b; } /** \internal \returns the bitwise xor of \a a and \a b */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pxor(const Packet& a, const Packet& b) { return a ^ b; } /** \internal \returns the bitwise andnot of \a a and \a b */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pandnot(const Packet& a, const Packet& b) { return a & (!b); } /** \internal \returns a packet version of \a *from, from must be 16 bytes aligned */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pload(const typename unpacket_traits::type* from) { return *from; } /** \internal \returns a packet version of \a *from, (un-aligned load) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet ploadu(const typename unpacket_traits::type* from) { return *from; } /** \internal \returns a packet with elements of \a *from duplicated. @@ -161,11 +161,11 @@ ploadu(const typename unpacket_traits::type* from) { return *from; } * duplicated to form: {from[0],from[0],from[1],from[1],,from[2],from[2],,from[3],from[3]} * Currently, this function is only used for scalar * complex products. */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet ploaddup(const typename unpacket_traits::type* from) { return *from; } /** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pset1(const typename unpacket_traits::type& a) { return a; } /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */ @@ -173,11 +173,11 @@ template inline typename packet_traits::type plset(const Scalar& a) { return a; } /** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */ -template inline void pstore(Scalar* to, const Packet& from) +template EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from) { (*to) = from; } /** \internal copy the packet \a from to \a *to, (un-aligned store) */ -template inline void pstoreu(Scalar* to, const Packet& from) +template EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from) { (*to) = from; } /** \internal tries to do cache prefetching of \a addr */ @@ -189,36 +189,36 @@ __builtin_prefetch(addr); } /** \internal \returns the first element of a packet */ -template inline typename unpacket_traits::type pfirst(const Packet& a) +template EIGEN_DEVICE_FUNC inline typename unpacket_traits::type pfirst(const Packet& a) { return a; } /** \internal \returns a packet where the element i contains the sum of the packet of \a vec[i] */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet preduxp(const Packet* vecs) { return vecs[0]; } /** \internal \returns the sum of the elements of \a a*/ -template inline typename unpacket_traits::type predux(const Packet& a) +template EIGEN_DEVICE_FUNC inline typename unpacket_traits::type predux(const Packet& a) { return a; } /** \internal \returns the product of the elements of \a a*/ -template inline typename unpacket_traits::type predux_mul(const Packet& a) +template EIGEN_DEVICE_FUNC inline typename unpacket_traits::type predux_mul(const Packet& a) { return a; } /** \internal \returns the min of the elements of \a a*/ -template inline typename unpacket_traits::type predux_min(const Packet& a) +template EIGEN_DEVICE_FUNC inline typename unpacket_traits::type predux_min(const Packet& a) { return a; } /** \internal \returns the max of the elements of \a a*/ -template inline typename unpacket_traits::type predux_max(const Packet& a) +template EIGEN_DEVICE_FUNC inline typename unpacket_traits::type predux_max(const Packet& a) { return a; } /** \internal \returns the reversed elements of \a a*/ -template inline Packet preverse(const Packet& a) +template EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a) { return a; } /** \internal \returns \a a with real and imaginary part flipped (for complex type only) */ -template inline Packet pcplxflip(const Packet& a) +template EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a) { // FIXME: uncomment the following in case we drop the internal imag and real functions. // using std::imag; @@ -275,7 +275,7 @@ inline void pstore1(typename unpacket_traits::type* to, const typename u } /** \internal \returns a * b + c (coeff-wise) */ -template inline Packet +template EIGEN_DEVICE_FUNC inline Packet pmadd(const Packet& a, const Packet& b, const Packet& c) @@ -336,12 +336,17 @@ inline void palign(PacketType& first, const PacketType& second) * Fast complex products (GCC generates a function call which is very slow) ***************************************************************************/ +// Eigen+CUDA does not support complexes. +#ifndef __CUDACC__ + template<> inline std::complex pmul(const std::complex& a, const std::complex& b) { return std::complex(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); } template<> inline std::complex pmul(const std::complex& a, const std::complex& b) { return std::complex(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); } +#endif + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 9b5c94b07..9d3d7fbbb 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -649,18 +649,20 @@ template struct scalar_fuzzy_default_impl { typedef typename NumTraits::Real RealScalar; - template + template EIGEN_DEVICE_FUNC static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec) { using std::abs; return abs(x) <= abs(y) * prec; } + EIGEN_DEVICE_FUNC static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec) { EIGEN_USING_STD_MATH(min); using std::abs; return abs(x - y) <= (min)(abs(x), abs(y)) * prec; } + EIGEN_DEVICE_FUNC static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar& prec) { return x <= y || isApprox(x, y, prec); @@ -671,15 +673,17 @@ template struct scalar_fuzzy_default_impl { typedef typename NumTraits::Real RealScalar; - template + template EIGEN_DEVICE_FUNC static inline bool isMuchSmallerThan(const Scalar& x, const Scalar&, const RealScalar&) { return x == Scalar(0); } + EIGEN_DEVICE_FUNC static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar&) { return x == y; } + EIGEN_DEVICE_FUNC static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar&) { return x <= y; @@ -705,21 +709,21 @@ struct scalar_fuzzy_default_impl template struct scalar_fuzzy_impl : scalar_fuzzy_default_impl::IsComplex, NumTraits::IsInteger> {}; -template +template EIGEN_DEVICE_FUNC inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, typename NumTraits::Real precision = NumTraits::dummy_precision()) { return scalar_fuzzy_impl::template isMuchSmallerThan(x, y, precision); } -template +template EIGEN_DEVICE_FUNC inline bool isApprox(const Scalar& x, const Scalar& y, typename NumTraits::Real precision = NumTraits::dummy_precision()) { return scalar_fuzzy_impl::isApprox(x, y, precision); } -template +template EIGEN_DEVICE_FUNC inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, typename NumTraits::Real precision = NumTraits::dummy_precision()) { @@ -742,17 +746,19 @@ template<> struct scalar_fuzzy_impl { typedef bool RealScalar; - template + template EIGEN_DEVICE_FUNC static inline bool isMuchSmallerThan(const bool& x, const bool&, const bool&) { return !x; } + EIGEN_DEVICE_FUNC static inline bool isApprox(bool x, bool y, bool) { return x == y; } + EIGEN_DEVICE_FUNC static inline bool isApproxOrLessThan(const bool& x, const bool& y, const bool&) { return (!x) || y; diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h index bac9e50b8..eb5306bf2 100644 --- a/Eigen/src/Core/NumTraits.h +++ b/Eigen/src/Core/NumTraits.h @@ -69,12 +69,15 @@ template struct GenericNumTraits typedef T Nested; static inline Real epsilon() { return std::numeric_limits::epsilon(); } + EIGEN_DEVICE_FUNC static inline Real dummy_precision() { // make sure to override this for floating-point types return Real(0); } + EIGEN_DEVICE_FUNC static inline T highest() { return (std::numeric_limits::max)(); } + EIGEN_DEVICE_FUNC static inline T lowest() { return IsInteger ? (std::numeric_limits::min)() : (-(std::numeric_limits::max)()); } #ifdef EIGEN2_SUPPORT @@ -91,11 +94,13 @@ template struct NumTraits : GenericNumTraits template<> struct NumTraits : GenericNumTraits { + EIGEN_DEVICE_FUNC static inline float dummy_precision() { return 1e-5f; } }; template<> struct NumTraits : GenericNumTraits { + EIGEN_DEVICE_FUNC static inline double dummy_precision() { return 1e-12; } }; diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 4e159896e..037242595 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -558,16 +558,16 @@ class PlainObjectBase : public internal::dense_xpr_base::type //@} using Base::setConstant; - Derived& setConstant(Index size, const Scalar& value); - Derived& setConstant(Index rows, Index cols, const Scalar& value); + EIGEN_DEVICE_FUNC Derived& setConstant(Index size, const Scalar& value); + EIGEN_DEVICE_FUNC Derived& setConstant(Index rows, Index cols, const Scalar& value); using Base::setZero; - Derived& setZero(Index size); - Derived& setZero(Index rows, Index cols); + EIGEN_DEVICE_FUNC Derived& setZero(Index size); + EIGEN_DEVICE_FUNC Derived& setZero(Index rows, Index cols); using Base::setOnes; - Derived& setOnes(Index size); - Derived& setOnes(Index rows, Index cols); + EIGEN_DEVICE_FUNC Derived& setOnes(Index size); + EIGEN_DEVICE_FUNC Derived& setOnes(Index rows, Index cols); using Base::setRandom; Derived& setRandom(Index size); diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h index 91496651c..a4026376e 100644 --- a/Eigen/src/Core/util/BlasUtil.h +++ b/Eigen/src/Core/util/BlasUtil.h @@ -56,8 +56,8 @@ template<> struct conj_if { template struct conj_helper { - EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const { return internal::pmadd(x,y,c); } - EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const { return internal::pmul(x,y); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const { return internal::pmadd(x,y,c); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const { return internal::pmul(x,y); } }; template struct conj_helper, std::complex, false,true> @@ -109,11 +109,11 @@ template struct conj_helper struct get_factor { - static EIGEN_STRONG_INLINE To run(const From& x) { return x; } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE To run(const From& x) { return x; } }; template struct get_factor::Real> { - static EIGEN_STRONG_INLINE typename NumTraits::Real run(const Scalar& x) { return real(x); } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE typename NumTraits::Real run(const Scalar& x) { return real(x); } }; // Lightweight helper class to access matrix coefficients. From 231d4a6fdae342af5f2a482104539eafe4fc5cdb Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 25 Jun 2013 10:08:50 +0200 Subject: [PATCH 0009/1103] Workarounf nvcc not being able to find RowMajor when declaring a Matrix<...> inside another namespace. --- Eigen/src/Core/util/ForwardDeclarations.h | 16 ++++++++-------- Eigen/src/Core/util/Macros.h | 4 ++-- 2 files changed, 10 insertions(+), 10 deletions(-) diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index d6a814586..024a9ddc7 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -57,12 +57,12 @@ template class Array; diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 64348cd16..90fee79c5 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -87,9 +87,9 @@ #endif #ifdef EIGEN_DEFAULT_TO_ROW_MAJOR -#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION RowMajor +#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::RowMajor #else -#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ColMajor +#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::ColMajor #endif #ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE From 75b3391e3fb047d028730cdff77baeb276ee3f99 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Thu, 4 Jul 2013 18:37:46 +0800 Subject: [PATCH 0010/1103] Enable singular matrix power using unitary similarities. --- .../Eigen/src/MatrixFunctions/MatrixPower.h | 129 ++++++++++-------- 1 file changed, 72 insertions(+), 57 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index c32437281..18f6703b6 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -49,14 +49,14 @@ class MatrixPowerAtomic typedef typename MatrixType::RealScalar RealScalar; typedef std::complex ComplexScalar; typedef typename MatrixType::Index Index; - typedef Array ArrayType; + typedef Block ResultType; const MatrixType& m_A; RealScalar m_p; - void computePade(int degree, const MatrixType& IminusT, MatrixType& res) const; - void compute2x2(MatrixType& res, RealScalar p) const; - void computeBig(MatrixType& res) const; + void computePade(int degree, const MatrixType& IminusT, ResultType& res) const; + void compute2x2(ResultType& res, RealScalar p) const; + void computeBig(ResultType& res) const; static int getPadeDegree(float normIminusT); static int getPadeDegree(double normIminusT); static int getPadeDegree(long double normIminusT); @@ -65,7 +65,7 @@ class MatrixPowerAtomic public: MatrixPowerAtomic(const MatrixType& T, RealScalar p); - void compute(MatrixType& res) const; + void compute(ResultType& res) const; }; template @@ -74,9 +74,8 @@ MatrixPowerAtomic::MatrixPowerAtomic(const MatrixType& T, RealScalar { eigen_assert(T.rows() == T.cols()); } template -void MatrixPowerAtomic::compute(MatrixType& res) const +void MatrixPowerAtomic::compute(ResultType& res) const { - res.resizeLike(m_A); switch (m_A.rows()) { case 0: break; @@ -92,25 +91,24 @@ void MatrixPowerAtomic::compute(MatrixType& res) const } template -void MatrixPowerAtomic::computePade(int degree, const MatrixType& IminusT, MatrixType& res) const +void MatrixPowerAtomic::computePade(int degree, const MatrixType& IminusT, ResultType& res) const { - int i = degree<<1; - res = (m_p-degree) / ((i-1)<<1) * IminusT; + int i = 2*degree; + res = (m_p-degree) / (2*i-2) * IminusT; for (--i; i; --i) { res = (MatrixType::Identity(IminusT.rows(), IminusT.cols()) + res).template triangularView() - .solve((i==1 ? -m_p : i&1 ? (-m_p-(i>>1))/(i<<1) : (m_p-(i>>1))/((i-1)<<1)) * IminusT).eval(); + .solve((i==1 ? -m_p : i&1 ? (-m_p-i/2)/(2*i) : (m_p-i/2)/(2*i-2)) * IminusT).eval(); } res += MatrixType::Identity(IminusT.rows(), IminusT.cols()); } // This function assumes that res has the correct size (see bug 614) template -void MatrixPowerAtomic::compute2x2(MatrixType& res, RealScalar p) const +void MatrixPowerAtomic::compute2x2(ResultType& res, RealScalar p) const { using std::abs; using std::pow; - ArrayType logTdiag = m_A.diagonal().array().log(); res.coeffRef(0,0) = pow(m_A.coeff(0,0), p); for (Index i=1; i < m_A.cols(); ++i) { @@ -126,7 +124,7 @@ void MatrixPowerAtomic::compute2x2(MatrixType& res, RealScalar p) co } template -void MatrixPowerAtomic::computeBig(MatrixType& res) const +void MatrixPowerAtomic::computeBig(ResultType& res) const { const int digits = std::numeric_limits::digits; const RealScalar maxNormForPade = digits <= 24? 4.3386528e-1f: // sigle precision @@ -139,19 +137,6 @@ void MatrixPowerAtomic::computeBig(MatrixType& res) const int degree, degree2, numberOfSquareRoots = 0; bool hasExtraSquareRoot = false; - /* FIXME - * For singular T, norm(I - T) >= 1 but maxNormForPade < 1, leads to infinite - * loop. We should move 0 eigenvalues to bottom right corner. We need not - * worry about tiny values (e.g. 1e-300) because they will reach 1 if - * repetitively sqrt'ed. - * - * If the 0 eigenvalues are semisimple, they can form a 0 matrix at the - * bottom right corner. - * - * [ T A ]^p [ T^p (T^-1 T^p A) ] - * [ ] = [ ] - * [ 0 0 ] [ 0 0 ] - */ for (Index i=0; i < m_A.cols(); ++i) eigen_assert(m_A(i,i) != RealScalar(0)); @@ -275,12 +260,6 @@ template class MatrixPower { private: - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime, - MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, - MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime - }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; typedef typename MatrixType::Index Index; @@ -294,7 +273,11 @@ class MatrixPower * The class stores a reference to A, so it should not be changed * (or destroyed) before evaluation. */ - explicit MatrixPower(const MatrixType& A) : m_A(A), m_conditionNumber(0) + explicit MatrixPower(const MatrixType& A) : + m_A(A), + m_conditionNumber(0), + m_rank(A.cols()), + m_nulls(0) { eigen_assert(A.rows() == A.cols()); } /** @@ -322,15 +305,17 @@ class MatrixPower private: typedef std::complex ComplexScalar; - typedef Matrix ComplexMatrix; + typedef Matrix ComplexMatrix; typename MatrixType::Nested m_A; MatrixType m_tmp; ComplexMatrix m_T, m_U, m_fT; RealScalar m_conditionNumber; + Index m_rank, m_nulls; - RealScalar modfAndInit(RealScalar, RealScalar*); + RealScalar split(RealScalar, RealScalar*); + void initialize(); template void computeIntPower(ResultType&, RealScalar); @@ -362,37 +347,62 @@ void MatrixPower::compute(ResultType& res, RealScalar p) res(0,0) = std::pow(m_A.coeff(0,0), p); break; default: - RealScalar intpart, x = modfAndInit(p, &intpart); + RealScalar intpart, x = split(p, &intpart); computeIntPower(res, intpart); - computeFracPower(res, x); + if (x) computeFracPower(res, x); } } template -typename MatrixPower::RealScalar -MatrixPower::modfAndInit(RealScalar x, RealScalar* intpart) +typename MatrixPower::RealScalar MatrixPower::split(RealScalar x, RealScalar* intpart) { - typedef Array RealArray; - *intpart = std::floor(x); RealScalar res = x - *intpart; - if (!m_conditionNumber && res) { - const ComplexSchur schurOfA(m_A); - m_T = schurOfA.matrixT(); - m_U = schurOfA.matrixU(); - - const RealArray absTdiag = m_T.diagonal().array().abs(); - m_conditionNumber = absTdiag.maxCoeff() / absTdiag.minCoeff(); - } + if (!m_conditionNumber && res) + initialize(); - if (res>RealScalar(0.5) && res>(1-res)*std::pow(m_conditionNumber, res)) { + if (res > RealScalar(0.5) && res > (1-res) * std::pow(m_conditionNumber, res)) { --res; ++*intpart; } return res; } +template +void MatrixPower::initialize() +{ + const ComplexSchur schurOfA(m_A); + JacobiRotation rot; + ComplexScalar eigenvalue; + + m_fT.resizeLike(m_A); + m_T = schurOfA.matrixT(); + m_U = schurOfA.matrixU(); + m_conditionNumber = m_T.diagonal().array().abs().maxCoeff() / m_T.diagonal().array().abs().minCoeff(); + + for (Index i = cols()-1; i>=0; --i) { + if (m_rank <= 2) + return; + if (m_T.coeff(i,i) == RealScalar(0)) { + for (Index j=i+1; j < m_rank; ++j) { + eigenvalue = m_T.coeff(j,j); + rot.makeGivens(m_T.coeff(j-1,j), eigenvalue); + m_T.applyOnTheRight(j-1, j, rot); + m_T.applyOnTheLeft(j-1, j, rot.adjoint()); + m_T.coeffRef(j-1,j-1) = eigenvalue; + m_T.coeffRef(j,j) = RealScalar(0); + m_U.applyOnTheRight(j-1, j, rot); + } + --m_rank; + } + } + + m_nulls = rows() - m_rank; + if (m_nulls) + m_fT.bottomRows(m_nulls).fill(RealScalar(0)); +} + template template void MatrixPower::computeIntPower(ResultType& res, RealScalar p) @@ -415,12 +425,17 @@ template template void MatrixPower::computeFracPower(ResultType& res, RealScalar p) { - if (p) { - eigen_assert(m_conditionNumber); - MatrixPowerAtomic(m_T, p).compute(m_fT); - revertSchur(m_tmp, m_fT, m_U); - res = m_tmp * res; + Block blockTp(m_fT, 0, 0, m_rank, m_rank); + eigen_assert(m_conditionNumber); + eigen_assert(m_rank + m_nulls == rows()); + + MatrixPowerAtomic(m_T.topLeftCorner(m_rank, m_rank), p).compute(blockTp); + if (m_nulls) { + m_fT.topRightCorner(m_rank, m_nulls) = m_T.topLeftCorner(m_rank, m_rank).template triangularView() + .solve(blockTp * m_T.topRightCorner(m_rank, m_nulls)); } + revertSchur(m_tmp, m_fT, m_U); + res = m_tmp * res; } template From cce68d4e91a6b8f53bdea448144a300c60f0097b Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Thu, 4 Jul 2013 18:39:33 +0800 Subject: [PATCH 0011/1103] Remove unused inclusions. --- unsupported/Eigen/MatrixFunctions | 2 -- 1 file changed, 2 deletions(-) diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions index 0991817d5..df49fdafd 100644 --- a/unsupported/Eigen/MatrixFunctions +++ b/unsupported/Eigen/MatrixFunctions @@ -13,8 +13,6 @@ #include #include -#include -#include #include #include From 4e26057f66d62d65739557bee72f4b9ad211e495 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Fri, 5 Jul 2013 00:08:11 +0800 Subject: [PATCH 0012/1103] Remove unused declarations for MatrixPowerProduct. --- Eigen/src/Core/MatrixBase.h | 3 --- Eigen/src/Core/util/ForwardDeclarations.h | 1 - 2 files changed, 4 deletions(-) diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 198e51084..84b83ffdf 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -162,9 +162,6 @@ template class MatrixBase #ifndef EIGEN_PARSED_BY_DOXYGEN template Derived& lazyAssign(const ProductBase& other); - - template - Derived& lazyAssign(const MatrixPowerProduct& other); #endif // not EIGEN_PARSED_BY_DOXYGEN template diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index d6a814586..7dc87968c 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -271,7 +271,6 @@ template class MatrixFunctionReturnValue; template class MatrixSquareRootReturnValue; template class MatrixLogarithmReturnValue; template class MatrixPowerReturnValue; -template class MatrixPowerProduct; namespace internal { template From 04a9ad6e10a5b216c868d1f182380ac079be9837 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Fri, 5 Jul 2013 00:28:28 +0800 Subject: [PATCH 0013/1103] Let complex power fall back to "log, scale, exp". --- Eigen/src/Core/MatrixBase.h | 1 + Eigen/src/Core/util/ForwardDeclarations.h | 1 + .../Eigen/src/MatrixFunctions/MatrixPower.h | 32 +++++++++++++++++++ 3 files changed, 34 insertions(+) diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 84b83ffdf..0628ebd1f 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -455,6 +455,7 @@ template class MatrixBase const MatrixSquareRootReturnValue sqrt() const; const MatrixLogarithmReturnValue log() const; const MatrixPowerReturnValue pow(const RealScalar& p) const; + const MatrixComplexPowerReturnValue pow(const std::complex& p) const; #ifdef EIGEN2_SUPPORT template diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 7dc87968c..cbedb7da4 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -271,6 +271,7 @@ template class MatrixFunctionReturnValue; template class MatrixSquareRootReturnValue; template class MatrixLogarithmReturnValue; template class MatrixPowerReturnValue; +template class MatrixComplexPowerReturnValue; namespace internal { template diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index 18f6703b6..a49db1916 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -503,6 +503,30 @@ class MatrixPowerReturnValue : public ReturnByValue< MatrixPowerReturnValue +class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerReturnValue > +{ + public: + typedef typename Derived::PlainObject PlainObject; + typedef typename std::complex ComplexScalar; + typedef typename Derived::Index Index; + + MatrixComplexPowerReturnValue(const Derived& A, const ComplexScalar& p) : m_A(A), m_p(p) + { } + + template + inline void evalTo(ResultType& res) const + { res = (m_p * m_A.log()).exp(); } + + Index rows() const { return m_A.rows(); } + Index cols() const { return m_A.cols(); } + + private: + const Derived& m_A; + const ComplexScalar m_p; + MatrixComplexPowerReturnValue& operator=(const MatrixComplexPowerReturnValue&); +}; + namespace internal { template @@ -513,12 +537,20 @@ template struct traits< MatrixPowerReturnValue > { typedef typename Derived::PlainObject ReturnType; }; +template +struct traits< MatrixComplexPowerReturnValue > +{ typedef typename Derived::PlainObject ReturnType; }; + } template const MatrixPowerReturnValue MatrixBase::pow(const RealScalar& p) const { return MatrixPowerReturnValue(derived(), p); } +template +const MatrixComplexPowerReturnValue MatrixBase::pow(const std::complex& p) const +{ return MatrixComplexPowerReturnValue(derived(), p); } + } // namespace Eigen #endif // EIGEN_MATRIX_POWER From c273a6c37c74949ee62136b73290f0f9568bb5e1 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Fri, 5 Jul 2013 03:33:56 +0800 Subject: [PATCH 0014/1103] Avoid pow(Scalar, int) for C++11 conformance. --- .../src/MatrixFunctions/MatrixExponential.h | 44 ++++++++++++------- 1 file changed, 28 insertions(+), 16 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 6825a7882..f8bbd59c1 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -157,6 +157,24 @@ class MatrixExponential { /** \brief L1 norm of m_M. */ RealScalar m_l1norm; + + /** \brief Scaling operator. */ + struct ScalingOp; +}; + +template +struct MatrixExponential::ScalingOp +{ + ScalingOp(int squarings) : m_squarings(squarings) { } + + inline const RealScalar operator() (const RealScalar& x) const + { return std::ldexp(x, -m_squarings); } + + inline const ComplexScalar operator() (const ComplexScalar& x) const + { return ComplexScalar(std::ldexp(x.real(), -m_squarings), std::ldexp(x.imag(), -m_squarings)); } + + private: + int m_squarings; }; template @@ -283,17 +301,15 @@ EIGEN_STRONG_INLINE void MatrixExponential::pade17(const MatrixType template void MatrixExponential::computeUV(float) { - using std::frexp; - using std::pow; if (m_l1norm < 4.258730016922831e-001) { pade3(m_M); } else if (m_l1norm < 1.880152677804762e+000) { pade5(m_M); } else { const float maxnorm = 3.925724783138660f; - frexp(m_l1norm / maxnorm, &m_squarings); + std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = m_M / pow(Scalar(2), m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade7(A); } } @@ -301,8 +317,6 @@ void MatrixExponential::computeUV(float) template void MatrixExponential::computeUV(double) { - using std::frexp; - using std::pow; if (m_l1norm < 1.495585217958292e-002) { pade3(m_M); } else if (m_l1norm < 2.539398330063230e-001) { @@ -313,9 +327,9 @@ void MatrixExponential::computeUV(double) pade9(m_M); } else { const double maxnorm = 5.371920351148152; - frexp(m_l1norm / maxnorm, &m_squarings); + std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = m_M / pow(Scalar(2), m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade13(A); } } @@ -323,8 +337,6 @@ void MatrixExponential::computeUV(double) template void MatrixExponential::computeUV(long double) { - using std::frexp; - using std::pow; #if LDBL_MANT_DIG == 53 // double precision computeUV(double()); #elif LDBL_MANT_DIG <= 64 // extended precision @@ -338,9 +350,9 @@ void MatrixExponential::computeUV(long double) pade9(m_M); } else { const long double maxnorm = 4.0246098906697353063L; - frexp(m_l1norm / maxnorm, &m_squarings); + std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = m_M / pow(Scalar(2), m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade13(A); } #elif LDBL_MANT_DIG <= 106 // double-double @@ -356,9 +368,9 @@ void MatrixExponential::computeUV(long double) pade13(m_M); } else { const long double maxnorm = 3.2579440895405400856599663723517L; - frexp(m_l1norm / maxnorm, &m_squarings); + std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = m_M / pow(Scalar(2), m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade17(A); } #elif LDBL_MANT_DIG <= 112 // quadruple precison @@ -374,9 +386,9 @@ void MatrixExponential::computeUV(long double) pade13(m_M); } else { const long double maxnorm = 2.884233277829519311757165057717815L; - frexp(m_l1norm / maxnorm, &m_squarings); + std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = m_M / pow(Scalar(2), m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade17(A); } #else From 9e2b4eeac0dcf62f27e05bb0ae923d5c48ee98ed Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Fri, 5 Jul 2013 23:28:57 +0800 Subject: [PATCH 0015/1103] Const-correct the scaling functor. --- .../Eigen/src/MatrixFunctions/MatrixExponential.h | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index f8bbd59c1..ca4532357 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -309,7 +309,7 @@ void MatrixExponential::computeUV(float) const float maxnorm = 3.925724783138660f; std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade7(A); } } @@ -329,7 +329,7 @@ void MatrixExponential::computeUV(double) const double maxnorm = 5.371920351148152; std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade13(A); } } @@ -352,7 +352,7 @@ void MatrixExponential::computeUV(long double) const long double maxnorm = 4.0246098906697353063L; std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade13(A); } #elif LDBL_MANT_DIG <= 106 // double-double @@ -370,7 +370,7 @@ void MatrixExponential::computeUV(long double) const long double maxnorm = 3.2579440895405400856599663723517L; std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade17(A); } #elif LDBL_MANT_DIG <= 112 // quadruple precison @@ -388,7 +388,7 @@ void MatrixExponential::computeUV(long double) const long double maxnorm = 2.884233277829519311757165057717815L; std::frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); + MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade17(A); } #else From 55ec3cc6d50807ae99dfe603e15ff5f7d075ecb3 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 7 Jul 2013 19:34:13 +0800 Subject: [PATCH 0016/1103] Prevent copying with internal::noncopyable. --- .../Eigen/src/MatrixFunctions/MatrixExponential.h | 8 ++------ unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h | 10 +++------- .../Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h | 6 +----- .../Eigen/src/MatrixFunctions/MatrixLogarithm.h | 6 +----- unsupported/Eigen/src/MatrixFunctions/MatrixPower.h | 4 ++-- .../Eigen/src/MatrixFunctions/MatrixSquareRoot.h | 6 +++--- unsupported/Eigen/src/MatrixFunctions/StemFunction.h | 2 +- 7 files changed, 13 insertions(+), 29 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index ca4532357..e5d4e3ad2 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -21,8 +21,8 @@ namespace Eigen { * expected to be an instantiation of the Matrix class template. */ template -class MatrixExponential { - +class MatrixExponential : internal::noncopyable +{ public: /** \brief Constructor. @@ -43,10 +43,6 @@ class MatrixExponential { private: - // Prevent copying - MatrixExponential(const MatrixExponential&); - MatrixExponential& operator=(const MatrixExponential&); - /** \brief Compute the (3,3)-Padé approximant to the exponential. * * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index 7d426640c..a5476ac26 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -34,7 +34,7 @@ namespace Eigen { template ::Scalar>::IsComplex> -class MatrixFunction +class MatrixFunction : internal::noncopyable { public: @@ -65,7 +65,7 @@ class MatrixFunction * \brief Partial specialization of MatrixFunction for real matrices */ template -class MatrixFunction +class MatrixFunction : internal::noncopyable { private: @@ -111,8 +111,6 @@ class MatrixFunction private: typename internal::nested::type m_A; /**< \brief Reference to argument of matrix function. */ AtomicType& m_atomic; /**< \brief Class for computing matrix function of atomic blocks. */ - - MatrixFunction& operator=(const MatrixFunction&); }; @@ -120,7 +118,7 @@ class MatrixFunction * \brief Partial specialization of MatrixFunction for complex matrices */ template -class MatrixFunction +class MatrixFunction : internal::noncopyable { private: @@ -176,8 +174,6 @@ class MatrixFunction * separation constant is set to 0.1, a value taken from the * paper by Davies and Higham. */ static const RealScalar separation() { return static_cast(0.1); } - - MatrixFunction& operator=(const MatrixFunction&); }; /** \brief Constructor. diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h index efe332c48..d6ff5f1ce 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h @@ -21,7 +21,7 @@ namespace Eigen { * entries are close to each other. */ template -class MatrixFunctionAtomic +class MatrixFunctionAtomic : internal::noncopyable { public: @@ -44,10 +44,6 @@ class MatrixFunctionAtomic private: - // Prevent copying - MatrixFunctionAtomic(const MatrixFunctionAtomic&); - MatrixFunctionAtomic& operator=(const MatrixFunctionAtomic&); - void computeMu(); bool taylorConverged(Index s, const MatrixType& F, const MatrixType& Fincr, const MatrixType& P); diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index c744fc05f..586c034af 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -28,7 +28,7 @@ namespace Eigen { * \sa class MatrixFunctionAtomic, MatrixBase::log() */ template -class MatrixLogarithmAtomic +class MatrixLogarithmAtomic : internal::noncopyable { public: @@ -71,10 +71,6 @@ private: std::numeric_limits::digits<= 64? 8: // extended precision std::numeric_limits::digits<=106? 10: // double-double 11; // quadruple precision - - // Prevent copying - MatrixLogarithmAtomic(const MatrixLogarithmAtomic&); - MatrixLogarithmAtomic& operator=(const MatrixLogarithmAtomic&); }; /** \brief Compute logarithm of triangular matrix with clustered eigenvalues. */ diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index a49db1916..5d580ac9e 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -38,7 +38,7 @@ class MatrixPowerRetval : public ReturnByValue< MatrixPowerRetval > }; template -class MatrixPowerAtomic +class MatrixPowerAtomic : internal::noncopyable { private: enum { @@ -257,7 +257,7 @@ MatrixPowerAtomic::computeSuperDiag(RealScalar curr, RealScalar prev * Output: \verbinclude MatrixPower_optimal.out */ template -class MatrixPower +class MatrixPower : internal::noncopyable { private: typedef typename MatrixType::Scalar Scalar; diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h index b48ea9d46..f3bcef409 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h @@ -24,7 +24,7 @@ namespace Eigen { * \sa MatrixSquareRoot, MatrixSquareRootTriangular */ template -class MatrixSquareRootQuasiTriangular +class MatrixSquareRootQuasiTriangular : internal::noncopyable { public: @@ -253,7 +253,7 @@ void MatrixSquareRootQuasiTriangular * \sa MatrixSquareRoot, MatrixSquareRootQuasiTriangular */ template -class MatrixSquareRootTriangular +class MatrixSquareRootTriangular : internal::noncopyable { public: MatrixSquareRootTriangular(const MatrixType& A) @@ -370,7 +370,7 @@ class MatrixSquareRoot // ********** Partial specialization for complex matrices ********** template -class MatrixSquareRoot +class MatrixSquareRoot : internal::noncopyable { public: diff --git a/unsupported/Eigen/src/MatrixFunctions/StemFunction.h b/unsupported/Eigen/src/MatrixFunctions/StemFunction.h index 724e55c1d..0a815d834 100644 --- a/unsupported/Eigen/src/MatrixFunctions/StemFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/StemFunction.h @@ -16,7 +16,7 @@ namespace Eigen { * \brief Stem functions corresponding to standard mathematical functions. */ template -class StdStemFunctions +class StdStemFunctions : internal::noncopyable { public: From 00e30a5fc43bd3ff177c5e12fb53eaba8fb65ce8 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 7 Jul 2013 19:57:23 +0800 Subject: [PATCH 0017/1103] We need not prohibit assignment here. Thanks to changeset 3edd4681f2f04c1164cb3805f1ac37fbf9a618c0 . --- unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h | 2 -- unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h | 2 -- unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h | 2 -- unsupported/Eigen/src/MatrixFunctions/MatrixPower.h | 3 --- unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h | 2 -- 5 files changed, 11 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index e5d4e3ad2..221ec3115 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -435,8 +435,6 @@ template struct MatrixExponentialReturnValue protected: const Derived& m_src; - private: - MatrixExponentialReturnValue& operator=(const MatrixExponentialReturnValue&); }; namespace internal { diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index a5476ac26..ad54d8ed0 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -527,8 +527,6 @@ template class MatrixFunctionReturnValue private: typename internal::nested::type m_A; StemFunction *m_f; - - MatrixFunctionReturnValue& operator=(const MatrixFunctionReturnValue&); }; namespace internal { diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index 586c034af..32ec385e0 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -454,8 +454,6 @@ public: private: typename internal::nested::type m_A; - - MatrixLogarithmReturnValue& operator=(const MatrixLogarithmReturnValue&); }; namespace internal { diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index 5d580ac9e..e0a687978 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -34,7 +34,6 @@ class MatrixPowerRetval : public ReturnByValue< MatrixPowerRetval > private: MatrixPower& m_pow; const RealScalar m_p; - MatrixPowerRetval& operator=(const MatrixPowerRetval&); }; template @@ -500,7 +499,6 @@ class MatrixPowerReturnValue : public ReturnByValue< MatrixPowerReturnValue @@ -524,7 +522,6 @@ class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerRe private: const Derived& m_A; const ComplexScalar m_p; - MatrixComplexPowerReturnValue& operator=(const MatrixComplexPowerReturnValue&); }; namespace internal { diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h index f3bcef409..c405bfd61 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h @@ -442,8 +442,6 @@ template class MatrixSquareRootReturnValue protected: const Derived& m_src; - private: - MatrixSquareRootReturnValue& operator=(const MatrixSquareRootReturnValue&); }; namespace internal { From 04bd1e3fc003f4220e82a553c3496a4f8d509959 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Mon, 8 Jul 2013 16:49:27 +0800 Subject: [PATCH 0018/1103] Slightly optimize atanh2. --- Eigen/src/Core/MathFunctions.h | 52 ++++++++++++++++++++-------------- 1 file changed, 30 insertions(+), 22 deletions(-) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 5df2d8bca..d415bc719 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -332,37 +332,45 @@ inline NewType cast(const OldType& x) * Implementation of atanh2 * ****************************************************************************/ -template -struct atanh2_default_impl +template +struct atanh2_impl { - typedef Scalar retval; - typedef typename NumTraits::Real RealScalar; - static inline Scalar run(const Scalar& x, const Scalar& y) + static inline Scalar run(const Scalar& x, const Scalar& r) + { + EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar) + #if __cplusplus >= 201103L + using std::log1p; + return log1p(2 * x / (r - x)) / 2; + #else + using std::abs; + using std::log; + using std::sqrt; + Scalar z = x / r; + if (r == 0 || abs(z) > sqrt(NumTraits::epsilon())) + return log((r + x) / (r - x)) / 2; + else + return z + z*z*z / 3; + #endif + } +}; + +template +struct atanh2_impl > +{ + typedef std::complex Scalar; + static inline Scalar run(const Scalar& x, const Scalar& r) { - using std::abs; using std::log; + using std::norm; using std::sqrt; - Scalar z = x / y; - if (y == Scalar(0) || abs(z) > sqrt(NumTraits::epsilon())) - return RealScalar(0.5) * log((y + x) / (y - x)); + Scalar z = x / r; + if (r == Scalar(0) || norm(z) > NumTraits::epsilon()) + return RealScalar(0.5) * log((r + x) / (r - x)); else return z + z*z*z / RealScalar(3); } }; -template -struct atanh2_default_impl -{ - static inline Scalar run(const Scalar&, const Scalar&) - { - EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar) - return Scalar(0); - } -}; - -template -struct atanh2_impl : atanh2_default_impl::IsInteger> {}; - template struct atanh2_retval { From 25544dbec3429848226c9a567ccd7e82973c04e7 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 10 Jul 2013 02:36:34 +0800 Subject: [PATCH 0019/1103] Add assertion against undefined matrix power. --- unsupported/Eigen/src/MatrixFunctions/MatrixPower.h | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index e0a687978..e4f13c993 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -398,8 +398,11 @@ void MatrixPower::initialize() } m_nulls = rows() - m_rank; - if (m_nulls) + if (m_nulls) { + eigen_assert(m_T.bottomRightCorner(m_nulls, m_nulls).isZero() + && "Base of matrix power should be invertible or with a semisimple zero eigenvalue."); m_fT.bottomRows(m_nulls).fill(RealScalar(0)); + } } template From 159a3bed9e26274ccc8da07a08ea394285d05bd3 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 10 Jul 2013 02:43:10 +0800 Subject: [PATCH 0020/1103] Write doc for complex power of a matrix. --- unsupported/Eigen/MatrixFunctions | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions index df49fdafd..0bdd379d7 100644 --- a/unsupported/Eigen/MatrixFunctions +++ b/unsupported/Eigen/MatrixFunctions @@ -228,15 +228,16 @@ const MatrixPowerReturnValue MatrixBase::pow(RealScalar p) con \endcode \param[in] M base of the matrix power, should be a square matrix. -\param[in] p exponent of the matrix power, should be real. +\param[in] p exponent of the matrix power. The matrix power \f$ M^p \f$ is defined as \f$ \exp(p \log(M)) \f$, where exp denotes the matrix exponential, and log denotes the matrix logarithm. -The matrix \f$ M \f$ should meet the conditions to be an argument of -matrix logarithm. If \p p is not of the real scalar type of \p M, it -is casted into the real scalar type of \p M. +If \p p is complex, the scalar type of \p M should be the type of \p +p . \f$ M^p \f$ simply evaluates into \f$ \exp(p \log(M)) \f$. +Therefore, the matrix \f$ M \f$ should meet the conditions to be an +argument of matrix logarithm. This function computes the matrix power using the Schur-Padé algorithm as implemented by class MatrixPower. The exponent is split From c52cbd9de95553f2f8c7a9020903081db719573c Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 10 Jul 2013 02:44:38 +0800 Subject: [PATCH 0021/1103] Write doc for positive power of a matrix with a semisimple zero eigenvalue. --- unsupported/Eigen/MatrixFunctions | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions index 0bdd379d7..41dfab390 100644 --- a/unsupported/Eigen/MatrixFunctions +++ b/unsupported/Eigen/MatrixFunctions @@ -239,12 +239,29 @@ p . \f$ M^p \f$ simply evaluates into \f$ \exp(p \log(M)) \f$. Therefore, the matrix \f$ M \f$ should meet the conditions to be an argument of matrix logarithm. -This function computes the matrix power using the Schur-Padé +If \p p is real, it is casted into the real scalar type of \p M. Then +this function computes the matrix power using the Schur-Padé algorithm as implemented by class MatrixPower. The exponent is split into integral part and fractional part, where the fractional part is in the interval \f$ (-1, 1) \f$. The main diagonal and the first super-diagonal is directly computed. +If \p M is singular with a semisimple zero eigenvalue and \p p is +positive, the Schur factor \f$ T \f$ is reordered with Givens +rotations, i.e. + +\f[ T = \left[ \begin{array}{cc} + T_1 & T_2 \\ + 0 & 0 + \end{array} \right] \f] + +where \f$ T_1 \f$ is invertible. Then \f$ T^p \f$ is given by + +\f[ T^p = \left[ \begin{array}{cc} + T_1^p & T_1^{-1} T_1^p T_2 \\ + 0 & 0 + \end{array}. \right] \f] + Details of the algorithm can be found in: Nicholas J. Higham and Lijing Lin, "A Schur-Padé algorithm for fractional powers of a matrix," SIAM J. %Matrix Anal. Applic., From d204bb57d04e1d106c33790407ee4727e0f9816b Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 10 Jul 2013 02:48:17 +0800 Subject: [PATCH 0022/1103] Remove unused struct definition in test. --- unsupported/test/matrix_power.cpp | 27 --------------------------- 1 file changed, 27 deletions(-) diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index b9d513b45..38b16fba9 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -9,33 +9,6 @@ #include "matrix_functions.h" -template ::IsComplex> -struct generateTriangularMatrix; - -// for real matrices, make sure none of the eigenvalues are negative -template -struct generateTriangularMatrix -{ - static void run(MatrixType& result, typename MatrixType::Index size) - { - result.resize(size, size); - result.template triangularView() = MatrixType::Random(size, size); - for (typename MatrixType::Index i = 0; i < size; ++i) - result.coeffRef(i,i) = std::abs(result.coeff(i,i)); - } -}; - -// for complex matrices, any matrix is fine -template -struct generateTriangularMatrix -{ - static void run(MatrixType& result, typename MatrixType::Index size) - { - result.resize(size, size); - result.template triangularView() = MatrixType::Random(size, size); - } -}; - template void test2dRotation(double tol) { From 639d03d900e844171af0cc3b98dfa98eca79dad5 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 10 Jul 2013 02:53:15 +0800 Subject: [PATCH 0023/1103] These casts are unnecessary because isApprox already casts them. --- unsupported/test/matrix_power.cpp | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index 38b16fba9..8d53ec94f 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -26,7 +26,7 @@ void test2dRotation(double tol) C = Apow(std::ldexp(angle,1) / M_PI); std::cout << "test2dRotation: i = " << i << " error powerm = " << relerr(C,B) << '\n'; - VERIFY(C.isApprox(B, static_cast(tol))); + VERIFY(C.isApprox(B, tol)); } } @@ -48,7 +48,7 @@ void test2dHyperbolicRotation(double tol) C = Apow(angle); std::cout << "test2dHyperbolicRotation: i = " << i << " error powerm = " << relerr(C,B) << '\n'; - VERIFY(C.isApprox(B, static_cast(tol))); + VERIFY(C.isApprox(B, tol)); } } @@ -70,15 +70,15 @@ void testExponentLaws(const MatrixType& m, double tol) m4 = mpow(x+y); m5.noalias() = m2 * m3; - VERIFY(m4.isApprox(m5, static_cast(tol))); + VERIFY(m4.isApprox(m5, tol)); m4 = mpow(x*y); m5 = m2.pow(y); - VERIFY(m4.isApprox(m5, static_cast(tol))); + VERIFY(m4.isApprox(m5, tol)); m4 = (std::abs(x) * m1).pow(y); m5 = std::pow(std::abs(x), y) * m3; - VERIFY(m4.isApprox(m5, static_cast(tol))); + VERIFY(m4.isApprox(m5, tol)); } } From 5c95892b83f9c1b2fb7f5a236cfb965bf8e55c3f Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 10 Jul 2013 02:57:54 +0800 Subject: [PATCH 0024/1103] Test power of singular matrices. --- unsupported/test/matrix_power.cpp | 90 +++++++++++++++++++++++++++---- 1 file changed, 80 insertions(+), 10 deletions(-) diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index 8d53ec94f..a3055c81e 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -9,6 +9,36 @@ #include "matrix_functions.h" +// for complex matrices, any matrix is fine +template::Scalar>::IsComplex> +struct generateSingularMatrix +{ + static void run(MatrixType& result, typename MatrixType::Index size) + { + result = MatrixType::Random(size, size); + result.col(0).fill(0); + } +}; + +// for real matrices, make sure none of the eigenvalues are negative +template +struct generateSingularMatrix +{ + static void run(MatrixType& result, typename MatrixType::Index size) + { + MatrixType mat = MatrixType::Random(size, size); + mat.col(0).fill(0); + ComplexSchur schur(mat); + typename ComplexSchur::ComplexMatrixType T = schur.matrixT(); + + for (typename MatrixType::Index i = 0; i < size; ++i) { + if (T.coeff(i,i).imag() == 0 && T.coeff(i,i).real() < 0) + T.coeffRef(i,i) = -T.coeff(i,i); + } + result = (schur.matrixU() * (T.template triangularView() * schur.matrixU().adjoint())).real(); + } +}; + template void test2dRotation(double tol) { @@ -53,7 +83,7 @@ void test2dHyperbolicRotation(double tol) } template -void testExponentLaws(const MatrixType& m, double tol) +void testGeneral(const MatrixType& m, double tol) { typedef typename MatrixType::RealScalar RealScalar; MatrixType m1, m2, m3, m4, m5; @@ -82,6 +112,36 @@ void testExponentLaws(const MatrixType& m, double tol) } } +template +void testSingular(MatrixType m, double tol) +{ + typedef typename MatrixType::RealScalar RealScalar; + MatrixType m1, m2, m3, m4, m5; + RealScalar x, y; + + for (int i=0; i < g_repeat; ++i) { + generateTestMatrix::run(m1, m.rows()); + MatrixPower mpow(m1); + + x = internal::random(0, 1); + y = internal::random(0, 1); + m2 = mpow(x); + m3 = mpow(y); + + m4 = mpow(x+y); + m5.noalias() = m2 * m3; + VERIFY(m4.isApprox(m5, tol)); + + m4 = mpow(x*y); + m5 = m2.pow(y); + VERIFY(m4.isApprox(m5, tol)); + + m4 = (x * m1).pow(y); + m5 = std::pow(x, y) * m3; + VERIFY(m4.isApprox(m5, tol)); + } +} + typedef Matrix Matrix3dRowMajor; typedef Matrix MatrixXe; @@ -94,13 +154,23 @@ void test_matrix_power() CALL_SUBTEST_1(test2dHyperbolicRotation(1e-5)); CALL_SUBTEST_9(test2dHyperbolicRotation(1e-14)); - CALL_SUBTEST_2(testExponentLaws(Matrix2d(), 1e-13)); - CALL_SUBTEST_7(testExponentLaws(Matrix3dRowMajor(), 1e-13)); - CALL_SUBTEST_3(testExponentLaws(Matrix4cd(), 1e-13)); - CALL_SUBTEST_4(testExponentLaws(MatrixXd(8,8), 2e-12)); - CALL_SUBTEST_1(testExponentLaws(Matrix2f(), 1e-4)); - CALL_SUBTEST_5(testExponentLaws(Matrix3cf(), 1e-4)); - CALL_SUBTEST_8(testExponentLaws(Matrix4f(), 1e-4)); - CALL_SUBTEST_6(testExponentLaws(MatrixXf(2,2), 1e-3)); // see bug 614 - CALL_SUBTEST_9(testExponentLaws(MatrixXe(7,7), 1e-13)); + CALL_SUBTEST_2(testGeneral(Matrix2d(), 1e-13)); + CALL_SUBTEST_7(testGeneral(Matrix3dRowMajor(), 1e-13)); + CALL_SUBTEST_3(testGeneral(Matrix4cd(), 1e-13)); + CALL_SUBTEST_4(testGeneral(MatrixXd(8,8), 2e-12)); + CALL_SUBTEST_1(testGeneral(Matrix2f(), 1e-4)); + CALL_SUBTEST_5(testGeneral(Matrix3cf(), 1e-4)); + CALL_SUBTEST_8(testGeneral(Matrix4f(), 1e-4)); + CALL_SUBTEST_6(testGeneral(MatrixXf(2,2), 1e-3)); // see bug 614 + CALL_SUBTEST_9(testGeneral(MatrixXe(7,7), 1e-13)); + + CALL_SUBTEST_2(testSingular(Matrix2d(), 1e-13)); + CALL_SUBTEST_7(testSingular(Matrix3dRowMajor(), 1e-13)); + CALL_SUBTEST_3(testSingular(Matrix4cd(), 1e-13)); + CALL_SUBTEST_4(testSingular(MatrixXd(8,8), 2e-12)); + CALL_SUBTEST_1(testSingular(Matrix2f(), 1e-4)); + CALL_SUBTEST_5(testSingular(Matrix3cf(), 1e-4)); + CALL_SUBTEST_8(testSingular(Matrix4f(), 1e-4)); + CALL_SUBTEST_6(testSingular(MatrixXf(2,2), 1e-3)); + CALL_SUBTEST_9(testSingular(MatrixXe(7,7), 1e-13)); } From 4466875d544c7a279b84fabb3b67a73da6b15487 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 10 Jul 2013 02:59:16 +0800 Subject: [PATCH 0025/1103] The only(?) way to test complex matrix power. --- unsupported/test/matrix_power.cpp | 23 +++++++++++++++++++++++ 1 file changed, 23 insertions(+) diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index a3055c81e..b9b72b528 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -142,6 +142,19 @@ void testSingular(MatrixType m, double tol) } } +template +void testLogThenExp(MatrixType m, double tol) +{ + typedef typename MatrixType::Scalar Scalar; + Scalar x; + + for (int i=0; i < g_repeat; ++i) { + generateTestMatrix::run(m, m.rows()); + x = internal::random(); + VERIFY(m.pow(x).isApprox((x * m.log()).exp(), tol)); + } +} + typedef Matrix Matrix3dRowMajor; typedef Matrix MatrixXe; @@ -173,4 +186,14 @@ void test_matrix_power() CALL_SUBTEST_8(testSingular(Matrix4f(), 1e-4)); CALL_SUBTEST_6(testSingular(MatrixXf(2,2), 1e-3)); CALL_SUBTEST_9(testSingular(MatrixXe(7,7), 1e-13)); + + CALL_SUBTEST_2(testLogThenExp(Matrix2d(), 1e-13)); + CALL_SUBTEST_7(testLogThenExp(Matrix3dRowMajor(), 1e-13)); + CALL_SUBTEST_3(testLogThenExp(Matrix4cd(), 1e-13)); + CALL_SUBTEST_4(testLogThenExp(MatrixXd(8,8), 2e-12)); + CALL_SUBTEST_1(testLogThenExp(Matrix2f(), 1e-4)); + CALL_SUBTEST_5(testLogThenExp(Matrix3cf(), 1e-4)); + CALL_SUBTEST_8(testLogThenExp(Matrix4f(), 1e-4)); + CALL_SUBTEST_6(testLogThenExp(MatrixXf(2,2), 1e-3)); + CALL_SUBTEST_9(testLogThenExp(MatrixXe(7,7), 1e-13)); } From a992fa74ebb560d6c25a2debf8a1460eb326b8f8 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Thu, 11 Jul 2013 02:31:13 +0800 Subject: [PATCH 0026/1103] Make non-conversion unary constructors explicit. --- .../Eigen/src/MatrixFunctions/MatrixExponential.h | 4 ++-- .../Eigen/src/MatrixFunctions/MatrixLogarithm.h | 2 +- .../Eigen/src/MatrixFunctions/MatrixSquareRoot.h | 12 ++++++------ 3 files changed, 9 insertions(+), 9 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 221ec3115..750b0b989 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -32,7 +32,7 @@ class MatrixExponential : internal::noncopyable * * \param[in] M matrix whose exponential is to be computed. */ - MatrixExponential(const MatrixType &M); + explicit MatrixExponential(const MatrixType &M); /** \brief Computes the matrix exponential. * @@ -415,7 +415,7 @@ template struct MatrixExponentialReturnValue * \param[in] src %Matrix (expression) forming the argument of the * matrix exponential. */ - MatrixExponentialReturnValue(const Derived& src) : m_src(src) { } + explicit MatrixExponentialReturnValue(const Derived& src) : m_src(src) { } /** \brief Compute the matrix exponential. * diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index 32ec385e0..33cfadfb4 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -425,7 +425,7 @@ public: * * \param[in] A %Matrix (expression) forming the argument of the matrix logarithm. */ - MatrixLogarithmReturnValue(const Derived& A) : m_A(A) { } + explicit MatrixLogarithmReturnValue(const Derived& A) : m_A(A) { } /** \brief Compute the matrix logarithm. * diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h index c405bfd61..0cd39ebe4 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h @@ -36,7 +36,7 @@ class MatrixSquareRootQuasiTriangular : internal::noncopyable * The class stores a reference to \p A, so it should not be * changed (or destroyed) before compute() is called. */ - MatrixSquareRootQuasiTriangular(const MatrixType& A) + explicit MatrixSquareRootQuasiTriangular(const MatrixType& A) : m_A(A) { eigen_assert(A.rows() == A.cols()); @@ -256,7 +256,7 @@ template class MatrixSquareRootTriangular : internal::noncopyable { public: - MatrixSquareRootTriangular(const MatrixType& A) + explicit MatrixSquareRootTriangular(const MatrixType& A) : m_A(A) { eigen_assert(A.rows() == A.cols()); @@ -321,7 +321,7 @@ class MatrixSquareRoot * The class stores a reference to \p A, so it should not be * changed (or destroyed) before compute() is called. */ - MatrixSquareRoot(const MatrixType& A); + explicit MatrixSquareRoot(const MatrixType& A); /** \brief Compute the matrix square root * @@ -341,7 +341,7 @@ class MatrixSquareRoot { public: - MatrixSquareRoot(const MatrixType& A) + explicit MatrixSquareRoot(const MatrixType& A) : m_A(A) { eigen_assert(A.rows() == A.cols()); @@ -374,7 +374,7 @@ class MatrixSquareRoot : internal::noncopyable { public: - MatrixSquareRoot(const MatrixType& A) + explicit MatrixSquareRoot(const MatrixType& A) : m_A(A) { eigen_assert(A.rows() == A.cols()); @@ -422,7 +422,7 @@ template class MatrixSquareRootReturnValue * \param[in] src %Matrix (expression) forming the argument of the * matrix square root. */ - MatrixSquareRootReturnValue(const Derived& src) : m_src(src) { } + explicit MatrixSquareRootReturnValue(const Derived& src) : m_src(src) { } /** \brief Compute the matrix square root. * From 738d75d3eb187bf1b3169d0de4328c6a2a5ef500 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 13 Jul 2013 22:11:36 +0800 Subject: [PATCH 0027/1103] Document on the return type of MatrixPower::operator() --- .../Eigen/src/MatrixFunctions/MatrixPower.h | 42 ++++++++++++++++--- 1 file changed, 37 insertions(+), 5 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index e4f13c993..5ba6278ae 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -14,16 +14,48 @@ namespace Eigen { template class MatrixPower; +/** + * \ingroup MatrixFunctions_Module + * + * \brief Proxy for the matrix power of some matrix. + * + * \tparam MatrixType type of the base, a matrix. + * + * This class holds the arguments to the matrix power until it is + * assigned or evaluated for some other reason (so the argument + * should not be changed in the meantime). It is the return type of + * MatrixPower::operator() and related functions and most of the + * time this is the only way it is used. + */ +/* TODO This class is only used by MatrixPower, so it should be nested + * into MatrixPower, like MatrixPower::ReturnValue. However, my + * compiler complained about unused template parameter in the + * following declaration in namespace internal. + * + * template + * struct traits::ReturnValue>; + */ template -class MatrixPowerRetval : public ReturnByValue< MatrixPowerRetval > +class MatrixPowerParenthesesReturnValue : public ReturnByValue< MatrixPowerParenthesesReturnValue > { public: typedef typename MatrixType::RealScalar RealScalar; typedef typename MatrixType::Index Index; - MatrixPowerRetval(MatrixPower& pow, RealScalar p) : m_pow(pow), m_p(p) + /** + * \brief Constructor. + * + * \param[in] pow %MatrixPower storing the base. + * \param[in] p scalar, the exponent of the matrix power. + */ + MatrixPowerParenthesesReturnValue(MatrixPower& pow, RealScalar p) : m_pow(pow), m_p(p) { } + /** + * \brief Compute the matrix power. + * + * \param[out] result + */ template inline void evalTo(ResultType& res) const { m_pow.compute(res, m_p); } @@ -286,8 +318,8 @@ class MatrixPower : internal::noncopyable * \return The expression \f$ A^p \f$, where A is specified in the * constructor. */ - const MatrixPowerRetval operator()(RealScalar p) - { return MatrixPowerRetval(*this, p); } + const MatrixPowerParenthesesReturnValue operator()(RealScalar p) + { return MatrixPowerParenthesesReturnValue(*this, p); } /** * \brief Compute the matrix power. @@ -530,7 +562,7 @@ class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerRe namespace internal { template -struct traits< MatrixPowerRetval > +struct traits< MatrixPowerParenthesesReturnValue > { typedef typename MatrixPowerType::PlainObject ReturnType; }; template From 3c423ccfe215283c5a5159ac74a8a23603e3f9bd Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 13 Jul 2013 22:12:09 +0800 Subject: [PATCH 0028/1103] Document on complex matrix power. --- .../Eigen/src/MatrixFunctions/MatrixPower.h | 30 ++++++++++++++++++- 1 file changed, 29 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index 5ba6278ae..ffb344065 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -513,7 +513,7 @@ class MatrixPowerReturnValue : public ReturnByValue< MatrixPowerReturnValue class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerReturnValue > { @@ -544,9 +557,24 @@ class MatrixComplexPowerReturnValue : public ReturnByValue< MatrixComplexPowerRe typedef typename std::complex ComplexScalar; typedef typename Derived::Index Index; + /** + * \brief Constructor. + * + * \param[in] A %Matrix (expression), the base of the matrix power. + * \param[in] p complex scalar, the exponent of the matrix power. + */ MatrixComplexPowerReturnValue(const Derived& A, const ComplexScalar& p) : m_A(A), m_p(p) { } + /** + * \brief Compute the matrix power. + * + * Because \p p is complex, \f$ A^p \f$ is simply evaluated as \f$ + * \exp(p \log(A)) \f$. + * + * \param[out] result \f$ A^p \f$ where \p A and \p p are as in the + * constructor. + */ template inline void evalTo(ResultType& res) const { res = (m_p * m_A.log()).exp(); } From d5501d3a90c954a901fe95e654f4195744179831 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 13 Jul 2013 23:13:07 +0800 Subject: [PATCH 0029/1103] Document on MatrixPowerAtomic. --- .../Eigen/src/MatrixFunctions/MatrixPower.h | 38 ++++++++++++++++++- 1 file changed, 37 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index ffb344065..e1de7f606 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -68,6 +68,21 @@ class MatrixPowerParenthesesReturnValue : public ReturnByValue< MatrixPowerParen const RealScalar m_p; }; +/** + * \ingroup MatrixFunctions_Module + * + * \brief Class for computing matrix powers. + * + * \tparam MatrixType type of the base, expected to be an instantiation + * of the Matrix class template. + * + * This class is capable of computing triangular real/complex matrices + * raised to a power in the interval \f$ (-1, 1) \f$. + * + * \note Currently this class is only used by MatrixPower. One may + * insist that this be nested into MatrixPower. This class is here to + * faciliate future development of triangular matrix functions. + */ template class MatrixPowerAtomic : internal::noncopyable { @@ -95,14 +110,35 @@ class MatrixPowerAtomic : internal::noncopyable static RealScalar computeSuperDiag(RealScalar, RealScalar, RealScalar p); public: + /** + * \brief Constructor. + * + * \param[in] T the base of the matrix power. + * \param[in] p the exponent of the matrix power, should be in + * \f$ (-1, 1) \f$. + * + * The class stores a reference to T, so it should not be changed + * (or destroyed) before evaluation. Only the upper triangular + * part of T is read. + */ MatrixPowerAtomic(const MatrixType& T, RealScalar p); + + /** + * \brief Compute the matrix power. + * + * \param[out] res \f$ A^p \f$ where A and p are specified in the + * constructor. + */ void compute(ResultType& res) const; }; template MatrixPowerAtomic::MatrixPowerAtomic(const MatrixType& T, RealScalar p) : m_A(T), m_p(p) -{ eigen_assert(T.rows() == T.cols()); } +{ + eigen_assert(T.rows() == T.cols()); + eigen_assert(p > -1 && p < 1); +} template void MatrixPowerAtomic::compute(ResultType& res) const From eeb744dc8d8608df22d3e8c512f42fb507f402ae Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 14 Jul 2013 02:00:50 +0800 Subject: [PATCH 0030/1103] Add test3dRotation. --- unsupported/test/matrix_power.cpp | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index b9b72b528..8e33d6038 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -82,6 +82,20 @@ void test2dHyperbolicRotation(double tol) } } +template +void test3dRotation(double tol) +{ + Matrix v; + T angle; + + for (int i=0; i<=20; ++i) { + v = Matrix::Random(); + v.normalize(); + angle = pow(10, (i-10) / 5.); + VERIFY(AngleAxis(angle, v).matrix().isApprox(AngleAxis(1,v).matrix().pow(angle), tol)); + } +} + template void testGeneral(const MatrixType& m, double tol) { @@ -156,6 +170,7 @@ void testLogThenExp(MatrixType m, double tol) } typedef Matrix Matrix3dRowMajor; +typedef Matrix Matrix3e; typedef Matrix MatrixXe; void test_matrix_power() @@ -167,6 +182,10 @@ void test_matrix_power() CALL_SUBTEST_1(test2dHyperbolicRotation(1e-5)); CALL_SUBTEST_9(test2dHyperbolicRotation(1e-14)); + CALL_SUBTEST_10(test3dRotation(1e-13)); + CALL_SUBTEST_11(test3dRotation(1e-5)); + CALL_SUBTEST_12(test3dRotation(1e-13)); + CALL_SUBTEST_2(testGeneral(Matrix2d(), 1e-13)); CALL_SUBTEST_7(testGeneral(Matrix3dRowMajor(), 1e-13)); CALL_SUBTEST_3(testGeneral(Matrix4cd(), 1e-13)); @@ -176,6 +195,9 @@ void test_matrix_power() CALL_SUBTEST_8(testGeneral(Matrix4f(), 1e-4)); CALL_SUBTEST_6(testGeneral(MatrixXf(2,2), 1e-3)); // see bug 614 CALL_SUBTEST_9(testGeneral(MatrixXe(7,7), 1e-13)); + CALL_SUBTEST_10(testGeneral(Matrix3d(), 1e-13)); + CALL_SUBTEST_11(testGeneral(Matrix3f(), 1e-4)); + CALL_SUBTEST_12(testGeneral(Matrix3e(), 1e-13)); CALL_SUBTEST_2(testSingular(Matrix2d(), 1e-13)); CALL_SUBTEST_7(testSingular(Matrix3dRowMajor(), 1e-13)); @@ -186,6 +208,9 @@ void test_matrix_power() CALL_SUBTEST_8(testSingular(Matrix4f(), 1e-4)); CALL_SUBTEST_6(testSingular(MatrixXf(2,2), 1e-3)); CALL_SUBTEST_9(testSingular(MatrixXe(7,7), 1e-13)); + CALL_SUBTEST_10(testSingular(Matrix3d(), 1e-13)); + CALL_SUBTEST_11(testSingular(Matrix3f(), 1e-4)); + CALL_SUBTEST_12(testSingular(Matrix3e(), 1e-13)); CALL_SUBTEST_2(testLogThenExp(Matrix2d(), 1e-13)); CALL_SUBTEST_7(testLogThenExp(Matrix3dRowMajor(), 1e-13)); @@ -196,4 +221,7 @@ void test_matrix_power() CALL_SUBTEST_8(testLogThenExp(Matrix4f(), 1e-4)); CALL_SUBTEST_6(testLogThenExp(MatrixXf(2,2), 1e-3)); CALL_SUBTEST_9(testLogThenExp(MatrixXe(7,7), 1e-13)); + CALL_SUBTEST_10(testLogThenExp(Matrix3d(), 1e-13)); + CALL_SUBTEST_11(testLogThenExp(Matrix3f(), 1e-4)); + CALL_SUBTEST_12(testLogThenExp(Matrix3e(), 1e-13)); } From cbe92de2b57a7580e4e5a2e00ad8ea52dda707b8 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 14 Jul 2013 17:27:44 +0800 Subject: [PATCH 0031/1103] Fix typo in testSingular. --- unsupported/test/matrix_power.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index 8e33d6038..223af60bc 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -134,7 +134,7 @@ void testSingular(MatrixType m, double tol) RealScalar x, y; for (int i=0; i < g_repeat; ++i) { - generateTestMatrix::run(m1, m.rows()); + generateSingularMatrix::run(m1, m.rows()); MatrixPower mpow(m1); x = internal::random(0, 1); From b8f0364a1c56784fa666c783e21c4bd1b218b9b0 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Mon, 15 Jul 2013 00:10:17 +0800 Subject: [PATCH 0032/1103] Test singular matrix power with square roots. Exponent laws are too unstable. --- unsupported/test/matrix_power.cpp | 96 +++++++++++++++---------------- 1 file changed, 47 insertions(+), 49 deletions(-) diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index 223af60bc..3ee19fc56 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2012 Chen-Pang He +// Copyright (C) 2012, 2013 Chen-Pang He // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -9,36 +9,6 @@ #include "matrix_functions.h" -// for complex matrices, any matrix is fine -template::Scalar>::IsComplex> -struct generateSingularMatrix -{ - static void run(MatrixType& result, typename MatrixType::Index size) - { - result = MatrixType::Random(size, size); - result.col(0).fill(0); - } -}; - -// for real matrices, make sure none of the eigenvalues are negative -template -struct generateSingularMatrix -{ - static void run(MatrixType& result, typename MatrixType::Index size) - { - MatrixType mat = MatrixType::Random(size, size); - mat.col(0).fill(0); - ComplexSchur schur(mat); - typename ComplexSchur::ComplexMatrixType T = schur.matrixT(); - - for (typename MatrixType::Index i = 0; i < size; ++i) { - if (T.coeff(i,i).imag() == 0 && T.coeff(i,i).real() < 0) - T.coeffRef(i,i) = -T.coeff(i,i); - } - result = (schur.matrixU() * (T.template triangularView() * schur.matrixU().adjoint())).real(); - } -}; - template void test2dRotation(double tol) { @@ -126,33 +96,61 @@ void testGeneral(const MatrixType& m, double tol) } } +// For complex matrices, any matrix is fine. +template::Scalar>::IsComplex> +struct processTriangularMatrix +{ + static void run(MatrixType&, MatrixType&, const MatrixType&) + { } +}; + +// For real matrices, make sure none of the eigenvalues are negative. +template +struct processTriangularMatrix +{ + static void run(MatrixType& m, MatrixType& T, const MatrixType& U) + { + typedef typename MatrixType::Index Index; + const Index size = m.cols(); + + for (Index i=0; i < size; ++i) { + if (i == size - 1 || T.coeff(i+1,i) == 0) + T.coeffRef(i,i) = std::abs(T.coeff(i,i)); + else + ++i; + } + m = U * T * U.adjoint(); + } +}; + template void testSingular(MatrixType m, double tol) { - typedef typename MatrixType::RealScalar RealScalar; - MatrixType m1, m2, m3, m4, m5; - RealScalar x, y; + const int IsComplex = NumTraits::Scalar>::IsComplex; + typedef typename internal::conditional< IsComplex, MatrixSquareRootTriangular, + MatrixSquareRootQuasiTriangular >::type SquareRootType; + typedef typename internal::conditional, const MatrixType&>::type TriangularType; + typename internal::conditional< IsComplex, ComplexSchur, RealSchur >::type schur; + MatrixType T; for (int i=0; i < g_repeat; ++i) { - generateSingularMatrix::run(m1, m.rows()); - MatrixPower mpow(m1); + m.setRandom(); + m.col(0).fill(0); - x = internal::random(0, 1); - y = internal::random(0, 1); - m2 = mpow(x); - m3 = mpow(y); + schur.compute(m); + T = schur.matrixT(); + const MatrixType& U = schur.matrixU(); + processTriangularMatrix::run(m, T, U); + MatrixPower mpow(m); - m4 = mpow(x+y); - m5.noalias() = m2 * m3; - VERIFY(m4.isApprox(m5, tol)); + SquareRootType(T).compute(T); + VERIFY(mpow(0.5).isApprox(U * (TriangularType(T) * U.adjoint()), tol)); - m4 = mpow(x*y); - m5 = m2.pow(y); - VERIFY(m4.isApprox(m5, tol)); + SquareRootType(T).compute(T); + VERIFY(mpow(0.25).isApprox(U * (TriangularType(T) * U.adjoint()), tol)); - m4 = (x * m1).pow(y); - m5 = std::pow(x, y) * m3; - VERIFY(m4.isApprox(m5, tol)); + SquareRootType(T).compute(T); + VERIFY(mpow(0.125).isApprox(U * (TriangularType(T) * U.adjoint()), tol)); } } From 9be658f7015161989b7ecccd70fd050ce563cad9 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Mon, 15 Jul 2013 00:43:14 +0800 Subject: [PATCH 0033/1103] generateTestMatrix can use processTriangularMatrix --- unsupported/test/matrix_functions.h | 41 ++++++++++++++++++++++------- unsupported/test/matrix_power.cpp | 27 ------------------- 2 files changed, 31 insertions(+), 37 deletions(-) diff --git a/unsupported/test/matrix_functions.h b/unsupported/test/matrix_functions.h index 5817caef6..295da16b6 100644 --- a/unsupported/test/matrix_functions.h +++ b/unsupported/test/matrix_functions.h @@ -10,27 +10,48 @@ #include "main.h" #include +// For complex matrices, any matrix is fine. +template::Scalar>::IsComplex> +struct processTriangularMatrix +{ + static void run(MatrixType&, MatrixType&, const MatrixType&) + { } +}; + +// For real matrices, make sure none of the eigenvalues are negative. +template +struct processTriangularMatrix +{ + static void run(MatrixType& m, MatrixType& T, const MatrixType& U) + { + typedef typename MatrixType::Index Index; + const Index size = m.cols(); + + for (Index i=0; i < size; ++i) { + if (i == size - 1 || T.coeff(i+1,i) == 0) + T.coeffRef(i,i) = std::abs(T.coeff(i,i)); + else + ++i; + } + m = U * T * U.transpose(); + } +}; + template ::Scalar>::IsComplex> struct generateTestMatrix; -// for real matrices, make sure none of the eigenvalues are negative template struct generateTestMatrix { static void run(MatrixType& result, typename MatrixType::Index size) { - MatrixType mat = MatrixType::Random(size, size); - EigenSolver es(mat); - typename EigenSolver::EigenvalueType eivals = es.eigenvalues(); - for (typename MatrixType::Index i = 0; i < size; ++i) { - if (eivals(i).imag() == 0 && eivals(i).real() < 0) - eivals(i) = -eivals(i); - } - result = (es.eigenvectors() * eivals.asDiagonal() * es.eigenvectors().inverse()).real(); + result = MatrixType::Random(size, size); + RealSchur schur(result); + MatrixType T = schur.matrixT(); + processTriangularMatrix::run(result, T, schur.matrixU()); } }; -// for complex matrices, any matrix is fine template struct generateTestMatrix { diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index 3ee19fc56..849e4287b 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -96,33 +96,6 @@ void testGeneral(const MatrixType& m, double tol) } } -// For complex matrices, any matrix is fine. -template::Scalar>::IsComplex> -struct processTriangularMatrix -{ - static void run(MatrixType&, MatrixType&, const MatrixType&) - { } -}; - -// For real matrices, make sure none of the eigenvalues are negative. -template -struct processTriangularMatrix -{ - static void run(MatrixType& m, MatrixType& T, const MatrixType& U) - { - typedef typename MatrixType::Index Index; - const Index size = m.cols(); - - for (Index i=0; i < size; ++i) { - if (i == size - 1 || T.coeff(i+1,i) == 0) - T.coeffRef(i,i) = std::abs(T.coeff(i,i)); - else - ++i; - } - m = U * T * U.adjoint(); - } -}; - template void testSingular(MatrixType m, double tol) { From 4b780553e05e35cb4813cd57eaf12befb6062891 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Mon, 15 Jul 2013 09:10:17 +0800 Subject: [PATCH 0034/1103] Eliminate unnecessary copying for sparse Kronecker product. --- .../KroneckerProduct/KroneckerTensorProduct.h | 106 ++++++++++-------- unsupported/test/kronecker_product.cpp | 21 ++-- 2 files changed, 75 insertions(+), 52 deletions(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index 532896c3b..6ec8eb558 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -14,7 +14,42 @@ namespace Eigen { -template class SparseMatrix; +template +class KroneckerProductBase : public ReturnByValue +{ + private: + typedef typename internal::traits Traits; + typedef typename Traits::Lhs Lhs; + typedef typename Traits::Rhs Rhs; + typedef typename Traits::Scalar Scalar; + + protected: + typedef typename Traits::Index Index; + + public: + KroneckerProductBase(const Lhs& A, const Rhs& B) + : m_A(A), m_B(B) + {} + + inline Index rows() const { return m_A.rows() * m_B.rows(); } + inline Index cols() const { return m_A.cols() * m_B.cols(); } + + Scalar coeff(Index row, Index col) const + { + return m_A.coeff(row / m_B.rows(), col / m_B.cols()) * + m_B.coeff(row % m_B.rows(), col % m_B.cols()); + } + + Scalar coeff(Index i) const + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived); + return m_A.coeff(i / m_A.size()) * m_B.coeff(i % m_A.size()); + } + + protected: + typename Lhs::Nested m_A; + typename Rhs::Nested m_B; +}; /*! * \brief Kronecker tensor product helper class for dense matrices @@ -27,40 +62,21 @@ template class SparseMatrix; * \tparam Rhs Type of the rignt-hand side, a matrix expression. */ template -class KroneckerProduct : public ReturnByValue > +class KroneckerProduct : public KroneckerProductBase > { private: - typedef ReturnByValue Base; - typedef typename Base::Scalar Scalar; - typedef typename Base::Index Index; + typedef KroneckerProductBase Base; + using Base::m_A; + using Base::m_B; public: /*! \brief Constructor. */ KroneckerProduct(const Lhs& A, const Rhs& B) - : m_A(A), m_B(B) + : Base(A, B) {} /*! \brief Evaluate the Kronecker tensor product. */ template void evalTo(Dest& dst) const; - - inline Index rows() const { return m_A.rows() * m_B.rows(); } - inline Index cols() const { return m_A.cols() * m_B.cols(); } - - Scalar coeff(Index row, Index col) const - { - return m_A.coeff(row / m_B.rows(), col / m_B.cols()) * - m_B.coeff(row % m_B.rows(), col % m_B.cols()); - } - - Scalar coeff(Index i) const - { - EIGEN_STATIC_ASSERT_VECTOR_ONLY(KroneckerProduct); - return m_A.coeff(i / m_A.size()) * m_B.coeff(i % m_A.size()); - } - - private: - typename Lhs::Nested m_A; - typename Rhs::Nested m_B; }; /*! @@ -77,40 +93,28 @@ class KroneckerProduct : public ReturnByValue > * \tparam Rhs Type of the rignt-hand side, a matrix expression. */ template -class KroneckerProductSparse : public EigenBase > +class KroneckerProductSparse : public KroneckerProductBase > { private: - typedef typename internal::traits::Index Index; + typedef KroneckerProductBase Base; + using Base::m_A; + using Base::m_B; public: /*! \brief Constructor. */ KroneckerProductSparse(const Lhs& A, const Rhs& B) - : m_A(A), m_B(B) + : Base(A, B) {} /*! \brief Evaluate the Kronecker tensor product. */ template void evalTo(Dest& dst) const; - - inline Index rows() const { return m_A.rows() * m_B.rows(); } - inline Index cols() const { return m_A.cols() * m_B.cols(); } - - template - operator SparseMatrix() - { - SparseMatrix result; - evalTo(result.derived()); - return result; - } - - private: - typename Lhs::Nested m_A; - typename Rhs::Nested m_B; }; template template void KroneckerProduct::evalTo(Dest& dst) const { + typedef typename Base::Index Index; const int BlockRows = Rhs::RowsAtCompileTime, BlockCols = Rhs::ColsAtCompileTime; const Index Br = m_B.rows(), @@ -124,9 +128,10 @@ template template void KroneckerProductSparse::evalTo(Dest& dst) const { + typedef typename Base::Index Index; const Index Br = m_B.rows(), Bc = m_B.cols(); - dst.resize(rows(),cols()); + dst.resize(this->rows(), this->cols()); dst.resizeNonZeros(0); dst.reserve(m_A.nonZeros() * m_B.nonZeros()); @@ -155,6 +160,7 @@ struct traits > typedef typename remove_all<_Lhs>::type Lhs; typedef typename remove_all<_Rhs>::type Rhs; typedef typename scalar_product_traits::ReturnType Scalar; + typedef typename promote_index_type::type Index; enum { Rows = size_at_compile_time::RowsAtCompileTime, traits::RowsAtCompileTime>::ret, @@ -193,6 +199,8 @@ struct traits > | EvalBeforeNestingBit | EvalBeforeAssigningBit, CoeffReadCost = Dynamic }; + + typedef SparseMatrix ReturnType; }; } // end namespace internal @@ -228,6 +236,16 @@ KroneckerProduct kroneckerProduct(const MatrixBase& a, const MatrixBase< * Computes Kronecker tensor product of two matrices, at least one of * which is sparse * + * \warning If you want to replace a matrix by its Kronecker product + * with some matrix, do \b NOT do this: + * \code + * A = kroneckerProduct(A,B); // bug!!! caused by aliasing effect + * \endcode + * instead, use eval() to work around this: + * \code + * A = kroneckerProduct(A,B).eval(); + * \endcode + * * \param a Dense/sparse matrix a * \param b Dense/sparse matrix b * \return Kronecker tensor product of a and b, stored in a sparse diff --git a/unsupported/test/kronecker_product.cpp b/unsupported/test/kronecker_product.cpp index 8ddc6ec28..c68a07de8 100644 --- a/unsupported/test/kronecker_product.cpp +++ b/unsupported/test/kronecker_product.cpp @@ -107,31 +107,34 @@ void test_kronecker_product() SparseMatrix SM_row_a(SM_a), SM_row_b(SM_b); - // test kroneckerProduct(DM_block,DM,DM_fixedSize) + // test DM_fixedSize = kroneckerProduct(DM_block,DM) Matrix DM_fix_ab = kroneckerProduct(DM_a.topLeftCorner<2,3>(),DM_b); CALL_SUBTEST(check_kronecker_product(DM_fix_ab)); + CALL_SUBTEST(check_kronecker_product(kroneckerProduct(DM_a.topLeftCorner<2,3>(),DM_b))); for(int i=0;i(2,5) = kroneckerProduct(DM_a,DM_b); CALL_SUBTEST(check_kronecker_product(DM_block_ab.block<6,6>(2,5))); - // test kroneckerProduct(DM,DM,DM) + // test DM = kroneckerProduct(DM,DM) MatrixXd DM_ab = kroneckerProduct(DM_a,DM_b); CALL_SUBTEST(check_kronecker_product(DM_ab)); + CALL_SUBTEST(check_kronecker_product(kroneckerProduct(DM_a,DM_b))); - // test kroneckerProduct(SM,DM,SM) + // test SM = kroneckerProduct(SM,DM) SparseMatrix SM_ab = kroneckerProduct(SM_a,DM_b); CALL_SUBTEST(check_kronecker_product(SM_ab)); SparseMatrix SM_ab2 = kroneckerProduct(SM_a,DM_b); CALL_SUBTEST(check_kronecker_product(SM_ab2)); + CALL_SUBTEST(check_kronecker_product(kroneckerProduct(SM_a,DM_b))); - // test kroneckerProduct(DM,SM,SM) + // test SM = kroneckerProduct(DM,SM) SM_ab.setZero(); SM_ab.insert(0,0)=37.0; SM_ab = kroneckerProduct(DM_a,SM_b); @@ -140,8 +143,9 @@ void test_kronecker_product() SM_ab2.insert(0,0)=37.0; SM_ab2 = kroneckerProduct(DM_a,SM_b); CALL_SUBTEST(check_kronecker_product(SM_ab2)); + CALL_SUBTEST(check_kronecker_product(kroneckerProduct(DM_a,SM_b))); - // test kroneckerProduct(SM,SM,SM) + // test SM = kroneckerProduct(SM,SM) SM_ab.resize(2,33); SM_ab.insert(0,0)=37.0; SM_ab = kroneckerProduct(SM_a,SM_b); @@ -150,8 +154,9 @@ void test_kronecker_product() SM_ab2.insert(0,0)=37.0; SM_ab2 = kroneckerProduct(SM_a,SM_b); CALL_SUBTEST(check_kronecker_product(SM_ab2)); + CALL_SUBTEST(check_kronecker_product(kroneckerProduct(SM_a,SM_b))); - // test kroneckerProduct(SM,SM,SM) with sparse pattern + // test SM = kroneckerProduct(SM,SM) with sparse pattern SM_a.resize(4,5); SM_b.resize(3,2); SM_a.resizeNonZeros(0); @@ -169,7 +174,7 @@ void test_kronecker_product() SM_ab = kroneckerProduct(SM_a,SM_b); CALL_SUBTEST(check_sparse_kronecker_product(SM_ab)); - // test dimension of result of kroneckerProduct(DM,DM,DM) + // test dimension of result of DM = kroneckerProduct(DM,DM) MatrixXd DM_a2(2,1); MatrixXd DM_b2(5,4); MatrixXd DM_ab2 = kroneckerProduct(DM_a2,DM_b2); From 20e535e1429cdb2f2dace3e2e6915e33968aa198 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 17 Jul 2013 10:04:20 +0200 Subject: [PATCH 0035/1103] Bump default branch to 3.2.90 --- Eigen/src/Core/util/Macros.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 96737456a..6798a3e0f 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -12,8 +12,8 @@ #define EIGEN_MACROS_H #define EIGEN_WORLD_VERSION 3 -#define EIGEN_MAJOR_VERSION 1 -#define EIGEN_MINOR_VERSION 91 +#define EIGEN_MAJOR_VERSION 2 +#define EIGEN_MINOR_VERSION 90 #define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \ (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \ From 6fab4012a36ebbc307ef29b7a7d1a9cb606fa67d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 17 Jul 2013 21:13:45 +0200 Subject: [PATCH 0036/1103] Rename isFinite to hasNonFinite to avoid future naming collisions. --- Eigen/src/Core/BooleanRedux.h | 4 ++-- Eigen/src/Core/DenseBase.h | 2 +- test/special_numbers.cpp | 12 ++++++------ 3 files changed, 9 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h index f6afeb034..e9840257c 100644 --- a/Eigen/src/Core/BooleanRedux.h +++ b/Eigen/src/Core/BooleanRedux.h @@ -131,7 +131,7 @@ inline typename DenseBase::Index DenseBase::count() const /** \returns true is \c *this contains at least one Not A Number (NaN). * - * \sa isFinite() + * \sa hasNonFinite() */ template inline bool DenseBase::hasNaN() const @@ -144,7 +144,7 @@ inline bool DenseBase::hasNaN() const * \sa hasNaN() */ template -inline bool DenseBase::isFinite() const +inline bool DenseBase::hasNonFinite() const { return !((derived()-derived()).hasNaN()); } diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 9551b8384..44fd660d9 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -348,7 +348,7 @@ template class DenseBase bool isOnes(const RealScalar& prec = NumTraits::dummy_precision()) const; inline bool hasNaN() const; - inline bool isFinite() const; + inline bool hasNonFinite() const; inline Derived& operator*=(const Scalar& other); inline Derived& operator/=(const Scalar& other); diff --git a/test/special_numbers.cpp b/test/special_numbers.cpp index 910ddf21c..0d03b4c64 100644 --- a/test/special_numbers.cpp +++ b/test/special_numbers.cpp @@ -33,7 +33,7 @@ template void special_numbers() mboth = mnan + minf; VERIFY(!m1.hasNaN()); - VERIFY(m1.isFinite()); + VERIFY(m1.hasNonFinite()); VERIFY(mnan.hasNaN()); VERIFY((s1*mnan).hasNaN()); @@ -42,11 +42,11 @@ template void special_numbers() VERIFY(mboth.hasNaN()); VERIFY(mboth.array().hasNaN()); - VERIFY(!mnan.isFinite()); - VERIFY(!minf.isFinite()); - VERIFY(!(minf-mboth).isFinite()); - VERIFY(!mboth.isFinite()); - VERIFY(!mboth.array().isFinite()); + VERIFY(!mnan.hasNonFinite()); + VERIFY(!minf.hasNonFinite()); + VERIFY(!(minf-mboth).hasNonFinite()); + VERIFY(!mboth.hasNonFinite()); + VERIFY(!mboth.array().hasNonFinite()); } void test_special_numbers() From 736fe99fbfe4bf7d50062b8ce16982ccbd93efe3 Mon Sep 17 00:00:00 2001 From: Desire NUENTSA Date: Thu, 18 Jul 2013 10:32:31 +0200 Subject: [PATCH 0037/1103] Fix bug #326 : expose tridiagonal eigensolver to end-users through ComputeFromTridiagonal() --- .../src/Eigenvalues/SelfAdjointEigenSolver.h | 159 ++++++++++++------ test/eigensolver_selfadjoint.cpp | 9 + 2 files changed, 117 insertions(+), 51 deletions(-) diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index 3993046a8..4e06809c4 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -20,6 +20,8 @@ class GeneralizedSelfAdjointEigenSolver; namespace internal { template struct direct_selfadjoint_eigenvalues; +template +ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const typename MatrixType::Index maxIterations, bool computeEigenvectors, MatrixType& eivec); } /** \eigenvalues_module \ingroup Eigenvalues_Module @@ -98,6 +100,7 @@ template class SelfAdjointEigenSolver */ typedef typename internal::plain_col_type::type RealVectorType; typedef Tridiagonalization TridiagonalizationType; + typedef typename TridiagonalizationType::SubDiagonalType SubDiagonalType; /** \brief Default constructor for fixed-size matrices. * @@ -207,6 +210,19 @@ template class SelfAdjointEigenSolver */ SelfAdjointEigenSolver& computeDirect(const MatrixType& matrix, int options = ComputeEigenvectors); + /** + *\brief Computes the eigen decomposition from a tridiagonal symmetric matrix + * + * \param[in] diag The vector containing the diagonal of the matrix. + * \param[in] subdiag The subdiagonal of the matrix. + * \returns Reference to \c *this + * + * This function assumes that the matrix has been reduced to tridiagonal form. + * + * \sa compute(const MatrixType&, int) for more information + */ + SelfAdjointEigenSolver& computeFromTridiagonal(const RealVectorType& diag, const SubDiagonalType& subdiag , int options=ComputeEigenvectors); + /** \brief Returns the eigenvectors of given matrix. * * \returns A const reference to the matrix whose columns are the eigenvectors. @@ -415,58 +431,8 @@ SelfAdjointEigenSolver& SelfAdjointEigenSolver mat.template triangularView() /= scale; m_subdiag.resize(n-1); internal::tridiagonalization_inplace(mat, diag, m_subdiag, computeEigenvectors); - - Index end = n-1; - Index start = 0; - Index iter = 0; // total number of iterations - while (end>0) - { - for (Index i = start; i0 && m_subdiag[end-1]==0) - { - end--; - } - if (end<=0) - break; - - // if we spent too many iterations, we give up - iter++; - if(iter > m_maxIterations * n) break; - - start = end - 1; - while (start>0 && m_subdiag[start-1]!=0) - start--; - - internal::tridiagonal_qr_step(diag.data(), m_subdiag.data(), start, end, computeEigenvectors ? m_eivec.data() : (Scalar*)0, n); - } - - if (iter <= m_maxIterations * n) - m_info = Success; - else - m_info = NoConvergence; - - // Sort eigenvalues and corresponding vectors. - // TODO make the sort optional ? - // TODO use a better sort algorithm !! - if (m_info == Success) - { - for (Index i = 0; i < n-1; ++i) - { - Index k; - m_eivalues.segment(i,n-i).minCoeff(&k); - if (k > 0) - { - std::swap(m_eivalues[i], m_eivalues[k+i]); - if(computeEigenvectors) - m_eivec.col(i).swap(m_eivec.col(k+i)); - } - } - } + m_info = internal::computeFromTridiagonal_impl(diag, m_subdiag, m_maxIterations, computeEigenvectors, m_eivec); // scale back the eigen values m_eivalues *= scale; @@ -476,8 +442,99 @@ SelfAdjointEigenSolver& SelfAdjointEigenSolver return *this; } +template +SelfAdjointEigenSolver& SelfAdjointEigenSolver +::computeFromTridiagonal(const RealVectorType& diag, const SubDiagonalType& subdiag , int options) +{ + //TODO : Add an option to scale the values beforehand + bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; + + m_eivalues = diag; + m_subdiag = subdiag; + if (computeEigenvectors) + { + m_eivec.setIdentity(diag.size(), diag.size()); + } + m_info = computeFromTridiagonal_impl(m_eivalues, m_subdiag, m_maxIterations, computeEigenvectors, m_eivec); + + m_isInitialized = true; + m_eigenvectorsOk = computeEigenvectors; + return *this; +} namespace internal { +/** + * \internal + * \brief Compute the eigendecomposition from a tridiagonal matrix + * + * \param[in,out] diag : On input, the diagonal of the matrix, on output the eigenvalues + * \param[in] subdiag : The subdiagonal part of the matrix. + * \param[in,out] : On input, the maximum number of iterations, on output, the effective number of iterations. + * \param[out] eivec : The matrix to store the eigenvectors... if needed. allocated on input + * \returns \c Success or \c NoConvergence + */ +template +ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const typename MatrixType::Index maxIterations, bool computeEigenvectors, MatrixType& eivec) +{ + using std::abs; + + ComputationInfo info; + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + + Index n = diag.size(); + Index end = n-1; + Index start = 0; + Index iter = 0; // total number of iterations + + while (end>0) + { + for (Index i = start; i0 && subdiag[end-1]==0) + { + end--; + } + if (end<=0) + break; + + // if we spent too many iterations, we give up + iter++; + if(iter > maxIterations * n) break; + + start = end - 1; + while (start>0 && subdiag[start-1]!=0) + start--; + + internal::tridiagonal_qr_step(diag.data(), subdiag.data(), start, end, computeEigenvectors ? eivec.data() : (Scalar*)0, n); + } + if (iter <= maxIterations * n) + info = Success; + else + info = NoConvergence; + + // Sort eigenvalues and corresponding vectors. + // TODO make the sort optional ? + // TODO use a better sort algorithm !! + if (info == Success) + { + for (Index i = 0; i < n-1; ++i) + { + Index k; + diag.segment(i,n-i).minCoeff(&k); + if (k > 0) + { + std::swap(diag[i], diag[k+i]); + if(computeEigenvectors) + eivec.col(i).swap(eivec.col(k+i)); + } + } + } + return info; +} template struct direct_selfadjoint_eigenvalues { diff --git a/test/eigensolver_selfadjoint.cpp b/test/eigensolver_selfadjoint.cpp index 5c6ecd875..06a6a8654 100644 --- a/test/eigensolver_selfadjoint.cpp +++ b/test/eigensolver_selfadjoint.cpp @@ -99,6 +99,15 @@ template void selfadjointeigensolver(const MatrixType& m) // FIXME tridiag.matrixQ().adjoint() does not work VERIFY_IS_APPROX(MatrixType(symmA.template selfadjointView()), tridiag.matrixQ() * tridiag.matrixT().eval() * MatrixType(tridiag.matrixQ()).adjoint()); + // Test computation of eigenvalues from tridiagonal matrix + if(rows > 1) + { + SelfAdjointEigenSolver eiSymmTridiag; + eiSymmTridiag.computeFromTridiagonal(tridiag.matrixT().diagonal(), tridiag.matrixT().diagonal(-1), ComputeEigenvectors); + VERIFY_IS_APPROX(eiSymm.eigenvalues(), eiSymmTridiag.eigenvalues()); + VERIFY_IS_APPROX(tridiag.matrixT(), eiSymmTridiag.eigenvectors().real() * eiSymmTridiag.eigenvalues().asDiagonal() * eiSymmTridiag.eigenvectors().real().transpose()); + } + if (rows > 1) { // Test matrix with NaN From 4f0bd557a470d47d60f49374153135422bff5289 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 18 Jul 2013 11:27:04 +0200 Subject: [PATCH 0038/1103] Previous isFinite->hasNonFinite change was broken. After discussion let's rename it to allFinite --- Eigen/src/Core/BooleanRedux.h | 4 ++-- Eigen/src/Core/DenseBase.h | 2 +- test/special_numbers.cpp | 12 ++++++------ 3 files changed, 9 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h index e9840257c..6e37e031a 100644 --- a/Eigen/src/Core/BooleanRedux.h +++ b/Eigen/src/Core/BooleanRedux.h @@ -131,7 +131,7 @@ inline typename DenseBase::Index DenseBase::count() const /** \returns true is \c *this contains at least one Not A Number (NaN). * - * \sa hasNonFinite() + * \sa allFinite() */ template inline bool DenseBase::hasNaN() const @@ -144,7 +144,7 @@ inline bool DenseBase::hasNaN() const * \sa hasNaN() */ template -inline bool DenseBase::hasNonFinite() const +inline bool DenseBase::allFinite() const { return !((derived()-derived()).hasNaN()); } diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 44fd660d9..c5800f6c8 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -348,7 +348,7 @@ template class DenseBase bool isOnes(const RealScalar& prec = NumTraits::dummy_precision()) const; inline bool hasNaN() const; - inline bool hasNonFinite() const; + inline bool allFinite() const; inline Derived& operator*=(const Scalar& other); inline Derived& operator/=(const Scalar& other); diff --git a/test/special_numbers.cpp b/test/special_numbers.cpp index 0d03b4c64..2f1b704be 100644 --- a/test/special_numbers.cpp +++ b/test/special_numbers.cpp @@ -33,7 +33,7 @@ template void special_numbers() mboth = mnan + minf; VERIFY(!m1.hasNaN()); - VERIFY(m1.hasNonFinite()); + VERIFY(m1.allFinite()); VERIFY(mnan.hasNaN()); VERIFY((s1*mnan).hasNaN()); @@ -42,11 +42,11 @@ template void special_numbers() VERIFY(mboth.hasNaN()); VERIFY(mboth.array().hasNaN()); - VERIFY(!mnan.hasNonFinite()); - VERIFY(!minf.hasNonFinite()); - VERIFY(!(minf-mboth).hasNonFinite()); - VERIFY(!mboth.hasNonFinite()); - VERIFY(!mboth.array().hasNonFinite()); + VERIFY(!mnan.allFinite()); + VERIFY(!minf.allFinite()); + VERIFY(!(minf-mboth).allFinite()); + VERIFY(!mboth.allFinite()); + VERIFY(!mboth.array().allFinite()); } void test_special_numbers() From 660b905e129c92fd0e8271d2df06d11347f4f32f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 19 Jul 2013 11:46:54 +0200 Subject: [PATCH 0039/1103] Fix ICE with ICC 11 --- Eigen/src/SparseCore/SparseMatrixBase.h | 2 +- Eigen/src/SparseCore/SparseSelfAdjointView.h | 10 +++++----- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index 89ace19e5..706f699b8 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -105,7 +105,7 @@ template class SparseMatrixBase : public EigenBase >::type AdjointReturnType; - typedef SparseMatrix PlainObject; + typedef SparseMatrix PlainObject; #ifndef EIGEN_PARSED_BY_DOXYGEN diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 80e794411..0eda96bc4 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -75,22 +75,22 @@ template class SparseSelfAdjointView * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product. */ template - SparseSparseProduct::Flags&RowMajorBit) ? RowMajor : ColMajor),Index>, OtherDerived> + SparseSparseProduct operator*(const SparseMatrixBase& rhs) const { - return SparseSparseProduct::Flags&RowMajorBit) ? RowMajor : ColMajor, Index>, OtherDerived>(*this, rhs.derived()); + return SparseSparseProduct(*this, rhs.derived()); } - + /** \returns an expression of the matrix product between a sparse matrix \a lhs and a sparse self-adjoint matrix \a rhs. * * Note that there is no algorithmic advantage of performing such a product compared to a general sparse-sparse matrix product. * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product. */ template friend - SparseSparseProduct::Flags&RowMajorBit) ? RowMajor : ColMajor),Index> > + SparseSparseProduct operator*(const SparseMatrixBase& lhs, const SparseSelfAdjointView& rhs) { - return SparseSparseProduct< OtherDerived, SparseMatrix::Flags&RowMajorBit) ? RowMajor : ColMajor, Index> >(lhs.derived(), rhs.derived()); + return SparseSparseProduct(lhs.derived(), rhs); } /** Efficient sparse self-adjoint matrix times dense vector/matrix product */ From c587e63631b6c37e842033c2b0438f269865dd0a Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 20 Jul 2013 17:49:38 +0800 Subject: [PATCH 0040/1103] Simplify MatrixPower::split --- .../Eigen/src/MatrixFunctions/MatrixPower.h | 22 +++++++++---------- 1 file changed, 11 insertions(+), 11 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index e1de7f606..b43789bd0 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -381,7 +381,7 @@ class MatrixPower : internal::noncopyable RealScalar m_conditionNumber; Index m_rank, m_nulls; - RealScalar split(RealScalar, RealScalar*); + void split(RealScalar&, RealScalar&); void initialize(); template @@ -414,26 +414,26 @@ void MatrixPower::compute(ResultType& res, RealScalar p) res(0,0) = std::pow(m_A.coeff(0,0), p); break; default: - RealScalar intpart, x = split(p, &intpart); + RealScalar intpart; + split(p, intpart); computeIntPower(res, intpart); - if (x) computeFracPower(res, x); + if (p) computeFracPower(res, p); } } template -typename MatrixPower::RealScalar MatrixPower::split(RealScalar x, RealScalar* intpart) +void MatrixPower::split(RealScalar& p, RealScalar& intpart) { - *intpart = std::floor(x); - RealScalar res = x - *intpart; + intpart = std::floor(p); + p -= intpart; - if (!m_conditionNumber && res) + if (!m_conditionNumber && p) initialize(); - if (res > RealScalar(0.5) && res > (1-res) * std::pow(m_conditionNumber, res)) { - --res; - ++*intpart; + if (p > RealScalar(0.5) && p > (1-p) * std::pow(m_conditionNumber, p)) { + --p; + ++intpart; } - return res; } template From 2320073e4153b60405e5460eab397c93c51de7e9 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 20 Jul 2013 17:58:12 +0800 Subject: [PATCH 0041/1103] Comment on private members of MatrixPower. --- .../Eigen/src/MatrixFunctions/MatrixPower.h | 49 ++++++++++++++++--- 1 file changed, 42 insertions(+), 7 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index b43789bd0..be13095e5 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -375,20 +375,51 @@ class MatrixPower : internal::noncopyable typedef Matrix ComplexMatrix; + /** \brief Reference to the base of matrix power. */ typename MatrixType::Nested m_A; - MatrixType m_tmp; - ComplexMatrix m_T, m_U, m_fT; - RealScalar m_conditionNumber; - Index m_rank, m_nulls; - void split(RealScalar&, RealScalar&); + /** \brief Temporary storage for computing integral power. */ + MatrixType m_tmp; + + /** \brief Store the result of Schur decomposition. */ + ComplexMatrix m_T, m_U; + + /** \brief Store fractional power of m_T. */ + ComplexMatrix m_fT; + + /** + * \brief Condition number of m_A. + * + * It is initialized as 0 to avoid performing unnecessary Schur + * decomposition, which is the bottleneck. + */ + RealScalar m_conditionNumber; + + /** \brief Rank of m_A. */ + Index m_rank; + + /** \brief Rank deficiency of m_A. */ + Index m_nulls; + + /** + * \brief Split p into integral part and fractional part. + * + * \param[in] p The exponent. + * \param[out] p The fractional part ranging in \f$ (-1, 1) \f$. + * \param[out] intpart The integral part. + * + * Only if the fractional part is nonzero, it calls initialize(). + */ + void split(RealScalar& p, RealScalar& intpart); + + /** \brief Perform Schur decomposition for fractional power. */ void initialize(); template - void computeIntPower(ResultType&, RealScalar); + void computeIntPower(ResultType& res, RealScalar p); template - void computeFracPower(ResultType&, RealScalar); + void computeFracPower(ResultType& res, RealScalar p); template static void revertSchur( @@ -427,9 +458,12 @@ void MatrixPower::split(RealScalar& p, RealScalar& intpart) intpart = std::floor(p); p -= intpart; + // Perform Schur decomposition if it is not yet performed and the power is + // not an integer. if (!m_conditionNumber && p) initialize(); + // Choose the more stable of intpart = floor(p) and intpart = ceil(p). if (p > RealScalar(0.5) && p > (1-p) * std::pow(m_conditionNumber, p)) { --p; ++intpart; @@ -448,6 +482,7 @@ void MatrixPower::initialize() m_U = schurOfA.matrixU(); m_conditionNumber = m_T.diagonal().array().abs().maxCoeff() / m_T.diagonal().array().abs().minCoeff(); + // Move zero eigenvalues to the bottom right corner. for (Index i = cols()-1; i>=0; --i) { if (m_rank <= 2) return; From dda869051db084d22dcfc46b7732b04d77ff9b4b Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 20 Jul 2013 18:47:54 +0800 Subject: [PATCH 0042/1103] Optimize MatrixPower::computeIntPower --- .../Eigen/src/MatrixFunctions/MatrixPower.h | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index be13095e5..7124874f7 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -447,6 +447,8 @@ void MatrixPower::compute(ResultType& res, RealScalar p) default: RealScalar intpart; split(p, intpart); + + res = MatrixType::Identity(rows(), cols()); computeIntPower(res, intpart); if (p) computeFracPower(res, p); } @@ -514,15 +516,18 @@ void MatrixPower::computeIntPower(ResultType& res, RealScalar p) { RealScalar pp = std::abs(p); - if (p<0) m_tmp = m_A.inverse(); - else m_tmp = m_A; + if (p<0) + m_tmp = m_A.inverse(); + else + m_tmp = m_A; - res = MatrixType::Identity(rows(), cols()); - while (pp >= 1) { + while (true) { if (std::fmod(pp, 2) >= 1) res = m_tmp * res; - m_tmp *= m_tmp; pp /= 2; + if (pp < 1) + break; + m_tmp *= m_tmp; } } From ede27f5780880e5fb89662d11db5a0e3b5586766 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 20 Jul 2013 23:30:37 +0800 Subject: [PATCH 0043/1103] Apply argument-dependent lookup on user-defined types. (using std::) --- .../src/MatrixFunctions/MatrixExponential.h | 23 +++++++--- .../Eigen/src/MatrixFunctions/MatrixPower.h | 43 +++++++++++++------ 2 files changed, 46 insertions(+), 20 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 750b0b989..810f426f9 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -164,10 +164,16 @@ struct MatrixExponential::ScalingOp ScalingOp(int squarings) : m_squarings(squarings) { } inline const RealScalar operator() (const RealScalar& x) const - { return std::ldexp(x, -m_squarings); } + { + using std::ldexp; + return ldexp(x, -m_squarings); + } inline const ComplexScalar operator() (const ComplexScalar& x) const - { return ComplexScalar(std::ldexp(x.real(), -m_squarings), std::ldexp(x.imag(), -m_squarings)); } + { + using std::ldexp; + return ComplexScalar(ldexp(x.real(), -m_squarings), ldexp(x.imag(), -m_squarings)); + } private: int m_squarings; @@ -297,13 +303,14 @@ EIGEN_STRONG_INLINE void MatrixExponential::pade17(const MatrixType template void MatrixExponential::computeUV(float) { + using std::frexp; if (m_l1norm < 4.258730016922831e-001) { pade3(m_M); } else if (m_l1norm < 1.880152677804762e+000) { pade5(m_M); } else { const float maxnorm = 3.925724783138660f; - std::frexp(m_l1norm / maxnorm, &m_squarings); + frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade7(A); @@ -313,6 +320,7 @@ void MatrixExponential::computeUV(float) template void MatrixExponential::computeUV(double) { + using std::frexp; if (m_l1norm < 1.495585217958292e-002) { pade3(m_M); } else if (m_l1norm < 2.539398330063230e-001) { @@ -323,7 +331,7 @@ void MatrixExponential::computeUV(double) pade9(m_M); } else { const double maxnorm = 5.371920351148152; - std::frexp(m_l1norm / maxnorm, &m_squarings); + frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade13(A); @@ -333,6 +341,7 @@ void MatrixExponential::computeUV(double) template void MatrixExponential::computeUV(long double) { + using std::frexp; #if LDBL_MANT_DIG == 53 // double precision computeUV(double()); #elif LDBL_MANT_DIG <= 64 // extended precision @@ -346,7 +355,7 @@ void MatrixExponential::computeUV(long double) pade9(m_M); } else { const long double maxnorm = 4.0246098906697353063L; - std::frexp(m_l1norm / maxnorm, &m_squarings); + frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade13(A); @@ -364,7 +373,7 @@ void MatrixExponential::computeUV(long double) pade13(m_M); } else { const long double maxnorm = 3.2579440895405400856599663723517L; - std::frexp(m_l1norm / maxnorm, &m_squarings); + frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade17(A); @@ -382,7 +391,7 @@ void MatrixExponential::computeUV(long double) pade13(m_M); } else { const long double maxnorm = 2.884233277829519311757165057717815L; - std::frexp(m_l1norm / maxnorm, &m_squarings); + frexp(m_l1norm / maxnorm, &m_squarings); if (m_squarings < 0) m_squarings = 0; MatrixType A = CwiseUnaryOp(m_M, m_squarings); pade17(A); diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index 7124874f7..b419e245b 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -143,11 +143,12 @@ MatrixPowerAtomic::MatrixPowerAtomic(const MatrixType& T, RealScalar template void MatrixPowerAtomic::compute(ResultType& res) const { + using std::pow; switch (m_A.rows()) { case 0: break; case 1: - res(0,0) = std::pow(m_A(0,0), m_p); + res(0,0) = pow(m_A(0,0), m_p); break; case 2: compute2x2(res, m_p); @@ -162,6 +163,7 @@ void MatrixPowerAtomic::computePade(int degree, const MatrixType& Im { int i = 2*degree; res = (m_p-degree) / (2*i-2) * IminusT; + for (--i; i; --i) { res = (MatrixType::Identity(IminusT.rows(), IminusT.cols()) + res).template triangularView() .solve((i==1 ? -m_p : i&1 ? (-m_p-i/2)/(2*i) : (m_p-i/2)/(2*i-2)) * IminusT).eval(); @@ -175,7 +177,6 @@ void MatrixPowerAtomic::compute2x2(ResultType& res, RealScalar p) co { using std::abs; using std::pow; - res.coeffRef(0,0) = pow(m_A.coeff(0,0), p); for (Index i=1; i < m_A.cols(); ++i) { @@ -193,6 +194,7 @@ void MatrixPowerAtomic::compute2x2(ResultType& res, RealScalar p) co template void MatrixPowerAtomic::computeBig(ResultType& res) const { + using std::ldexp; const int digits = std::numeric_limits::digits; const RealScalar maxNormForPade = digits <= 24? 4.3386528e-1f: // sigle precision digits <= 53? 2.789358995219730e-1: // double precision @@ -224,7 +226,7 @@ void MatrixPowerAtomic::computeBig(ResultType& res) const computePade(degree, IminusT, res); for (; numberOfSquareRoots; --numberOfSquareRoots) { - compute2x2(res, std::ldexp(m_p, -numberOfSquareRoots)); + compute2x2(res, ldexp(m_p, -numberOfSquareRoots)); res = res.template triangularView() * res; } compute2x2(res, m_p); @@ -289,19 +291,28 @@ template inline typename MatrixPowerAtomic::ComplexScalar MatrixPowerAtomic::computeSuperDiag(const ComplexScalar& curr, const ComplexScalar& prev, RealScalar p) { - ComplexScalar logCurr = std::log(curr); - ComplexScalar logPrev = std::log(prev); - int unwindingNumber = std::ceil((numext::imag(logCurr - logPrev) - M_PI) / (2*M_PI)); + using std::ceil; + using std::exp; + using std::log; + using std::sinh; + + ComplexScalar logCurr = log(curr); + ComplexScalar logPrev = log(prev); + int unwindingNumber = ceil((numext::imag(logCurr - logPrev) - M_PI) / (2*M_PI)); ComplexScalar w = numext::atanh2(curr - prev, curr + prev) + ComplexScalar(0, M_PI*unwindingNumber); - return RealScalar(2) * std::exp(RealScalar(0.5) * p * (logCurr + logPrev)) * std::sinh(p * w) / (curr - prev); + return RealScalar(2) * exp(RealScalar(0.5) * p * (logCurr + logPrev)) * sinh(p * w) / (curr - prev); } template inline typename MatrixPowerAtomic::RealScalar MatrixPowerAtomic::computeSuperDiag(RealScalar curr, RealScalar prev, RealScalar p) { + using std::exp; + using std::log; + using std::sinh; + RealScalar w = numext::atanh2(curr - prev, curr + prev); - return 2 * std::exp(p * (std::log(curr) + std::log(prev)) / 2) * std::sinh(p * w) / (curr - prev); + return 2 * exp(p * (log(curr) + log(prev)) / 2) * sinh(p * w) / (curr - prev); } /** @@ -438,11 +449,12 @@ template template void MatrixPower::compute(ResultType& res, RealScalar p) { + using std::pow; switch (cols()) { case 0: break; case 1: - res(0,0) = std::pow(m_A.coeff(0,0), p); + res(0,0) = pow(m_A.coeff(0,0), p); break; default: RealScalar intpart; @@ -457,7 +469,10 @@ void MatrixPower::compute(ResultType& res, RealScalar p) template void MatrixPower::split(RealScalar& p, RealScalar& intpart) { - intpart = std::floor(p); + using std::floor; + using std::pow; + + intpart = floor(p); p -= intpart; // Perform Schur decomposition if it is not yet performed and the power is @@ -466,7 +481,7 @@ void MatrixPower::split(RealScalar& p, RealScalar& intpart) initialize(); // Choose the more stable of intpart = floor(p) and intpart = ceil(p). - if (p > RealScalar(0.5) && p > (1-p) * std::pow(m_conditionNumber, p)) { + if (p > RealScalar(0.5) && p > (1-p) * pow(m_conditionNumber, p)) { --p; ++intpart; } @@ -514,7 +529,9 @@ template template void MatrixPower::computeIntPower(ResultType& res, RealScalar p) { - RealScalar pp = std::abs(p); + using std::abs; + using std::fmod; + RealScalar pp = abs(p); if (p<0) m_tmp = m_A.inverse(); @@ -522,7 +539,7 @@ void MatrixPower::computeIntPower(ResultType& res, RealScalar p) m_tmp = m_A; while (true) { - if (std::fmod(pp, 2) >= 1) + if (fmod(pp, 2) >= 1) res = m_tmp * res; pp /= 2; if (pp < 1) From 3d94ed9fa0bc06771c39970bcae6705ca2524cad Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 21 Jul 2013 00:18:19 +0800 Subject: [PATCH 0044/1103] Document on MatrixExponential::ScalingOp --- .../src/MatrixFunctions/MatrixExponential.h | 20 ++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 810f426f9..30b36a179 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -126,6 +126,7 @@ class MatrixExponential : internal::noncopyable */ void computeUV(long double); + struct ScalingOp; typedef typename internal::traits::Scalar Scalar; typedef typename NumTraits::Real RealScalar; typedef typename std::complex ComplexScalar; @@ -153,22 +154,35 @@ class MatrixExponential : internal::noncopyable /** \brief L1 norm of m_M. */ RealScalar m_l1norm; - - /** \brief Scaling operator. */ - struct ScalingOp; }; +/** \brief Scaling operator. + * + * This struct is used by CwiseUnaryOp to scale a matrix by \f$ 2^{-s} \f$. + */ template struct MatrixExponential::ScalingOp { + /** \brief Constructor. + * + * \param[in] squarings The integer \f$ s \f$ in this document. + */ ScalingOp(int squarings) : m_squarings(squarings) { } + /** \brief Scale a matrix coefficient. + * + * \param[in,out] x The scalar to be scaled, becoming \f$ 2^{-s} x \f$. + */ inline const RealScalar operator() (const RealScalar& x) const { using std::ldexp; return ldexp(x, -m_squarings); } + /** \brief Scale a matrix coefficient. + * + * \param[in,out] x The scalar to be scaled, becoming \f$ 2^{-s} x \f$. + */ inline const ComplexScalar operator() (const ComplexScalar& x) const { using std::ldexp; From 1191949e8725106035d0f7276857217585a115ef Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 21 Jul 2013 00:19:46 +0800 Subject: [PATCH 0045/1103] Improve documentation on Kronecker product module. --- .../KroneckerProduct/KroneckerTensorProduct.h | 22 ++++++++++++++++++- 1 file changed, 21 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index 6ec8eb558..80be9fd7a 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -12,8 +12,15 @@ #ifndef KRONECKER_TENSOR_PRODUCT_H #define KRONECKER_TENSOR_PRODUCT_H -namespace Eigen { +namespace Eigen { +/*! + * \ingroup KroneckerProduct_Module + * + * \brief The base class of dense and sparse Kronecker product. + * + * \tparam Derived is the derived type. + */ template class KroneckerProductBase : public ReturnByValue { @@ -27,6 +34,7 @@ class KroneckerProductBase : public ReturnByValue typedef typename Traits::Index Index; public: + /*! \brief Constructor. */ KroneckerProductBase(const Lhs& A, const Rhs& B) : m_A(A), m_B(B) {} @@ -34,12 +42,20 @@ class KroneckerProductBase : public ReturnByValue inline Index rows() const { return m_A.rows() * m_B.rows(); } inline Index cols() const { return m_A.cols() * m_B.cols(); } + /*! + * This overrides ReturnByValue::coeff because this function is + * efficient enough. + */ Scalar coeff(Index row, Index col) const { return m_A.coeff(row / m_B.rows(), col / m_B.cols()) * m_B.coeff(row % m_B.rows(), col % m_B.cols()); } + /*! + * This overrides ReturnByValue::coeff because this function is + * efficient enough. + */ Scalar coeff(Index i) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived); @@ -52,6 +68,8 @@ class KroneckerProductBase : public ReturnByValue }; /*! + * \ingroup KroneckerProduct_Module + * * \brief Kronecker tensor product helper class for dense matrices * * This class is the return value of kroneckerProduct(MatrixBase, @@ -80,6 +98,8 @@ class KroneckerProduct : public KroneckerProductBase > }; /*! + * \ingroup KroneckerProduct_Module + * * \brief Kronecker tensor product helper class for sparse matrices * * If at least one of the operands is a sparse matrix expression, From 51573da3a4407288b900e8764e20de4fc09018d3 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 21 Jul 2013 01:00:36 +0800 Subject: [PATCH 0046/1103] Warn about power of a matrix with non-semisimple 0 eigenvalue. --- unsupported/Eigen/MatrixFunctions | 4 ++++ .../doc/examples/MatrixPower_failure.cpp | 20 +++++++++++++++++++ 2 files changed, 24 insertions(+) create mode 100644 unsupported/doc/examples/MatrixPower_failure.cpp diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions index 41dfab390..ff466a519 100644 --- a/unsupported/Eigen/MatrixFunctions +++ b/unsupported/Eigen/MatrixFunctions @@ -262,6 +262,10 @@ where \f$ T_1 \f$ is invertible. Then \f$ T^p \f$ is given by 0 & 0 \end{array}. \right] \f] +\warning Fractional power of a matrix with a non-semisimple zero +eigenvalue is not well-defined. We introduce an assertion failure +against inaccurate result, e.g. \include MatrixPower_failure.cpp + Details of the algorithm can be found in: Nicholas J. Higham and Lijing Lin, "A Schur-Padé algorithm for fractional powers of a matrix," SIAM J. %Matrix Anal. Applic., diff --git a/unsupported/doc/examples/MatrixPower_failure.cpp b/unsupported/doc/examples/MatrixPower_failure.cpp new file mode 100644 index 000000000..d20de78a0 --- /dev/null +++ b/unsupported/doc/examples/MatrixPower_failure.cpp @@ -0,0 +1,20 @@ +#include +#include + +int main() +{ + Eigen::Matrix4d A; + A << 0, 0, 2, 3, + 0, 0, 4, 5, + 0, 0, 6, 7, + 0, 0, 8, 9; + std::cout << A.pow(0.37) << std::endl; + + // The 1 makes eigenvalue 0 non-semisimple. + A.coeffRef(0, 1) = 1; + + // This fails if EIGEN_NO_DEBUG is undefined. + std::cout << A.pow(0.37) << std::endl; + + return 0; +} From c00f688c643038650f941f786227a76cd6fd2310 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 21 Jul 2013 05:40:56 +0800 Subject: [PATCH 0047/1103] Fix doc. (It is also used by computeFracPower) --- unsupported/Eigen/src/MatrixFunctions/MatrixPower.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index b419e245b..5548bd95c 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -389,7 +389,7 @@ class MatrixPower : internal::noncopyable /** \brief Reference to the base of matrix power. */ typename MatrixType::Nested m_A; - /** \brief Temporary storage for computing integral power. */ + /** \brief Temporary storage. */ MatrixType m_tmp; /** \brief Store the result of Schur decomposition. */ From 01190b3544cd0a674be6475185d5dd8e4b7890c5 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 21 Jul 2013 18:09:11 +0800 Subject: [PATCH 0048/1103] Directly code failing example, or it breaks `make doc`. --- unsupported/Eigen/MatrixFunctions | 23 ++++++++++++++++++- .../doc/examples/MatrixPower_failure.cpp | 20 ---------------- 2 files changed, 22 insertions(+), 21 deletions(-) delete mode 100644 unsupported/doc/examples/MatrixPower_failure.cpp diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions index ff466a519..0b12aaffb 100644 --- a/unsupported/Eigen/MatrixFunctions +++ b/unsupported/Eigen/MatrixFunctions @@ -264,7 +264,28 @@ where \f$ T_1 \f$ is invertible. Then \f$ T^p \f$ is given by \warning Fractional power of a matrix with a non-semisimple zero eigenvalue is not well-defined. We introduce an assertion failure -against inaccurate result, e.g. \include MatrixPower_failure.cpp +against inaccurate result, e.g. \code +#include +#include + +int main() +{ + Eigen::Matrix4d A; + A << 0, 0, 2, 3, + 0, 0, 4, 5, + 0, 0, 6, 7, + 0, 0, 8, 9; + std::cout << A.pow(0.37) << std::endl; + + // The 1 makes eigenvalue 0 non-semisimple. + A.coeffRef(0, 1) = 1; + + // This fails if EIGEN_NO_DEBUG is undefined. + std::cout << A.pow(0.37) << std::endl; + + return 0; +} +\endcode Details of the algorithm can be found in: Nicholas J. Higham and Lijing Lin, "A Schur-Padé algorithm for fractional powers of a diff --git a/unsupported/doc/examples/MatrixPower_failure.cpp b/unsupported/doc/examples/MatrixPower_failure.cpp deleted file mode 100644 index d20de78a0..000000000 --- a/unsupported/doc/examples/MatrixPower_failure.cpp +++ /dev/null @@ -1,20 +0,0 @@ -#include -#include - -int main() -{ - Eigen::Matrix4d A; - A << 0, 0, 2, 3, - 0, 0, 4, 5, - 0, 0, 6, 7, - 0, 0, 8, 9; - std::cout << A.pow(0.37) << std::endl; - - // The 1 makes eigenvalue 0 non-semisimple. - A.coeffRef(0, 1) = 1; - - // This fails if EIGEN_NO_DEBUG is undefined. - std::cout << A.pow(0.37) << std::endl; - - return 0; -} From 463343fb37cbd76e2527c560557a4d27a3d29c80 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sun, 21 Jul 2013 21:31:15 +0100 Subject: [PATCH 0049/1103] Clean-up of MatrixExponential: * put internal stuff in the internal namespace * replace member functions by free functions --- .../src/MatrixFunctions/MatrixExponential.h | 679 ++++++++---------- 1 file changed, 312 insertions(+), 367 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 30b36a179..54f07b27c 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -1,8 +1,8 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009, 2010 Jitse Niesen -// Copyright (C) 2011 Chen-Pang He +// Copyright (C) 2009, 2010, 2013 Jitse Niesen +// Copyright (C) 2011, 2013 Chen-Pang He // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -14,160 +14,21 @@ #include "StemFunction.h" namespace Eigen { - -/** \ingroup MatrixFunctions_Module - * \brief Class for computing the matrix exponential. - * \tparam MatrixType type of the argument of the exponential, - * expected to be an instantiation of the Matrix class template. - */ -template -class MatrixExponential : internal::noncopyable -{ - public: - - /** \brief Constructor. - * - * The class stores a reference to \p M, so it should not be - * changed (or destroyed) before compute() is called. - * - * \param[in] M matrix whose exponential is to be computed. - */ - explicit MatrixExponential(const MatrixType &M); - - /** \brief Computes the matrix exponential. - * - * \param[out] result the matrix exponential of \p M in the constructor. - */ - template - void compute(ResultType &result); - - private: - - /** \brief Compute the (3,3)-Padé approximant to the exponential. - * - * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé - * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. - * - * \param[in] A Argument of matrix exponential - */ - void pade3(const MatrixType &A); - - /** \brief Compute the (5,5)-Padé approximant to the exponential. - * - * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé - * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. - * - * \param[in] A Argument of matrix exponential - */ - void pade5(const MatrixType &A); - - /** \brief Compute the (7,7)-Padé approximant to the exponential. - * - * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé - * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. - * - * \param[in] A Argument of matrix exponential - */ - void pade7(const MatrixType &A); - - /** \brief Compute the (9,9)-Padé approximant to the exponential. - * - * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé - * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. - * - * \param[in] A Argument of matrix exponential - */ - void pade9(const MatrixType &A); - - /** \brief Compute the (13,13)-Padé approximant to the exponential. - * - * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé - * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. - * - * \param[in] A Argument of matrix exponential - */ - void pade13(const MatrixType &A); - - /** \brief Compute the (17,17)-Padé approximant to the exponential. - * - * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé - * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. - * - * This function activates only if your long double is double-double or quadruple. - * - * \param[in] A Argument of matrix exponential - */ - void pade17(const MatrixType &A); - - /** \brief Compute Padé approximant to the exponential. - * - * Computes \c m_U, \c m_V and \c m_squarings such that - * \f$ (V+U)(V-U)^{-1} \f$ is a Padé of - * \f$ \exp(2^{-\mbox{squarings}}M) \f$ around \f$ M = 0 \f$. The - * degree of the Padé approximant and the value of - * squarings are chosen such that the approximation error is no - * more than the round-off error. - * - * The argument of this function should correspond with the (real - * part of) the entries of \c m_M. It is used to select the - * correct implementation using overloading. - */ - void computeUV(double); - - /** \brief Compute Padé approximant to the exponential. - * - * \sa computeUV(double); - */ - void computeUV(float); - - /** \brief Compute Padé approximant to the exponential. - * - * \sa computeUV(double); - */ - void computeUV(long double); - - struct ScalingOp; - typedef typename internal::traits::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef typename std::complex ComplexScalar; - - /** \brief Reference to matrix whose exponential is to be computed. */ - typename internal::nested::type m_M; - - /** \brief Odd-degree terms in numerator of Padé approximant. */ - MatrixType m_U; - - /** \brief Even-degree terms in numerator of Padé approximant. */ - MatrixType m_V; - - /** \brief Used for temporary storage. */ - MatrixType m_tmp1; - - /** \brief Used for temporary storage. */ - MatrixType m_tmp2; - - /** \brief Identity matrix of the same size as \c m_M. */ - MatrixType m_Id; - - /** \brief Number of squarings required in the last step. */ - int m_squarings; - - /** \brief L1 norm of m_M. */ - RealScalar m_l1norm; -}; +namespace internal { /** \brief Scaling operator. * * This struct is used by CwiseUnaryOp to scale a matrix by \f$ 2^{-s} \f$. */ -template -struct MatrixExponential::ScalingOp +template +struct MatrixExponentialScalingOp { /** \brief Constructor. * * \param[in] squarings The integer \f$ s \f$ in this document. */ - ScalingOp(int squarings) : m_squarings(squarings) { } + MatrixExponentialScalingOp(int squarings) : m_squarings(squarings) { } + /** \brief Scale a matrix coefficient. * @@ -179,6 +40,8 @@ struct MatrixExponential::ScalingOp return ldexp(x, -m_squarings); } + typedef std::complex ComplexScalar; + /** \brief Scale a matrix coefficient. * * \param[in,out] x The scalar to be scaled, becoming \f$ 2^{-s} x \f$. @@ -193,228 +56,313 @@ struct MatrixExponential::ScalingOp int m_squarings; }; +/** \brief Compute the (3,3)-Padé approximant to the exponential. + * + * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé + * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. + */ template -MatrixExponential::MatrixExponential(const MatrixType &M) : - m_M(M), - m_U(M.rows(),M.cols()), - m_V(M.rows(),M.cols()), - m_tmp1(M.rows(),M.cols()), - m_tmp2(M.rows(),M.cols()), - m_Id(MatrixType::Identity(M.rows(), M.cols())), - m_squarings(0), - m_l1norm(M.cwiseAbs().colwise().sum().maxCoeff()) +void matrix_exp_pade3(const MatrixType &A, MatrixType &U, MatrixType &V) { - /* empty body */ + typedef typename NumTraits::Scalar>::Real RealScalar; + const RealScalar b[] = {120., 60., 12., 1.}; + const MatrixType A2 = A * A; + const MatrixType tmp = b[3] * A2 + b[1] * MatrixType::Identity(A.rows(), A.cols()); + U.noalias() = A * tmp; + V = b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols()); } +/** \brief Compute the (5,5)-Padé approximant to the exponential. + * + * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé + * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. + */ template -template -void MatrixExponential::compute(ResultType &result) +void matrix_exp_pade5(const MatrixType &A, MatrixType &U, MatrixType &V) +{ + typedef typename NumTraits::Scalar>::Real RealScalar; + const RealScalar b[] = {30240., 15120., 3360., 420., 30., 1.}; + const MatrixType A2 = A * A; + const MatrixType A4 = A2 * A2; + const MatrixType tmp = b[5] * A4 + b[3] * A2 + b[1] * MatrixType::Identity(A.rows(), A.cols()); + U.noalias() = A * tmp; + V = b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols()); +} + +/** \brief Compute the (7,7)-Padé approximant to the exponential. + * + * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé + * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. + */ +template +void matrix_exp_pade7(const MatrixType &A, MatrixType &U, MatrixType &V) +{ + typedef typename NumTraits::Scalar>::Real RealScalar; + const RealScalar b[] = {17297280., 8648640., 1995840., 277200., 25200., 1512., 56., 1.}; + const MatrixType A2 = A * A; + const MatrixType A4 = A2 * A2; + const MatrixType A6 = A4 * A2; + const MatrixType tmp = b[7] * A6 + b[5] * A4 + b[3] * A2 + + b[1] * MatrixType::Identity(A.rows(), A.cols()); + U.noalias() = A * tmp; + V = b[6] * A6 + b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols()); + +} + +/** \brief Compute the (9,9)-Padé approximant to the exponential. + * + * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé + * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. + */ +template +void matrix_exp_pade9(const MatrixType &A, MatrixType &U, MatrixType &V) +{ + typedef typename NumTraits::Scalar>::Real RealScalar; + const RealScalar b[] = {17643225600., 8821612800., 2075673600., 302702400., 30270240., + 2162160., 110880., 3960., 90., 1.}; + const MatrixType A2 = A * A; + const MatrixType A4 = A2 * A2; + const MatrixType A6 = A4 * A2; + const MatrixType A8 = A6 * A2; + const MatrixType tmp = b[9] * A8 + b[7] * A6 + b[5] * A4 + b[3] * A2 + + b[1] * MatrixType::Identity(A.rows(), A.cols()); + U.noalias() = A * tmp; + V = b[8] * A8 + b[6] * A6 + b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols()); +} + +/** \brief Compute the (13,13)-Padé approximant to the exponential. + * + * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé + * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. + */ +template +void matrix_exp_pade13(const MatrixType &A, MatrixType &U, MatrixType &V) +{ + typedef typename NumTraits::Scalar>::Real RealScalar; + const RealScalar b[] = {64764752532480000., 32382376266240000., 7771770303897600., + 1187353796428800., 129060195264000., 10559470521600., 670442572800., + 33522128640., 1323241920., 40840800., 960960., 16380., 182., 1.}; + const MatrixType A2 = A * A; + const MatrixType A4 = A2 * A2; + const MatrixType A6 = A4 * A2; + V = b[13] * A6 + b[11] * A4 + b[9] * A2; // used for temporary storage + MatrixType tmp = A6 * V; + tmp += b[7] * A6 + b[5] * A4 + b[3] * A2 + b[1] * MatrixType::Identity(A.rows(), A.cols()); + U.noalias() = A * tmp; + tmp = b[12] * A6 + b[10] * A4 + b[8] * A2; + V.noalias() = A6 * tmp; + V += b[6] * A6 + b[4] * A4 + b[2] * A2 + b[0] * MatrixType::Identity(A.rows(), A.cols()); +} + +/** \brief Compute the (17,17)-Padé approximant to the exponential. + * + * After exit, \f$ (V+U)(V-U)^{-1} \f$ is the Padé + * approximant of \f$ \exp(A) \f$ around \f$ A = 0 \f$. + * + * This function activates only if your long double is double-double or quadruple. + */ +#if LDBL_MANT_DIG > 64 +template +void matrix_exp_pade17(const MatrixType &A, MatrixType &U, MatrixType &V) +{ + typedef typename NumTraits::Scalar>::Real RealScalar; + const RealScalar b[] = {830034394580628357120000.L, 415017197290314178560000.L, + 100610229646136770560000.L, 15720348382208870400000.L, + 1774878043152614400000.L, 153822763739893248000.L, 10608466464820224000.L, + 595373117923584000.L, 27563570274240000.L, 1060137318240000.L, + 33924394183680.L, 899510451840.L, 19554575040.L, 341863200.L, 4651200.L, + 46512.L, 306.L, 1.L}; + const MatrixType A2 = A * A; + const MatrixType A4 = A2 * A2; + const MatrixType A6 = A4 * A2; + const MatrixType A8 = A4 * A4; + V = b[17] * m_tmp1 + b[15] * A6 + b[13] * A4 + b[11] * A2; // used for temporary storage + matrixType tmp = A8 * V; + tmp += b[9] * A8 + b[7] * A6 + b[5] * A4 + b[3] * A2 + + b[1] * MatrixType::Identity(A.rows(), A.cols()); + U.noalias() = A * tmp; + tmp = b[16] * A8 + b[14] * A6 + b[12] * A4 + b[10] * A2; + V.noalias() = tmp * A8; + V += b[8] * A8 + b[6] * A6 + b[4] * A4 + b[2] * A2 + + b[0] * MatrixType::Identity(A.rows(), A.cols()); +} +#endif + +template ::Scalar>::Real> +struct matrix_exp_computeUV +{ + /** \brief Compute Padé approximant to the exponential. + * + * Computes \c U, \c V and \c squarings such that \f$ (V+U)(V-U)^{-1} \f$ is a Padé + * approximant of \f$ \exp(2^{-\mbox{squarings}}M) \f$ around \f$ M = 0 \f$, where \f$ M \f$ + * denotes the matrix \c arg. The degree of the Padé approximant and the value of squarings + * are chosen such that the approximation error is no more than the round-off error. + */ + static void run(const MatrixType& arg, MatrixType& U, MatrixType& V, int& squarings); +}; + +template +struct matrix_exp_computeUV +{ + static void run(const MatrixType& arg, MatrixType& U, MatrixType& V, int& squarings) + { + using std::frexp; + using std::pow; + const float l1norm = arg.cwiseAbs().colwise().sum().maxCoeff(); + squarings = 0; + if (l1norm < 4.258730016922831e-001) { + matrix_exp_pade3(arg, U, V); + } else if (l1norm < 1.880152677804762e+000) { + matrix_exp_pade5(arg, U, V); + } else { + const float maxnorm = 3.925724783138660f; + frexp(l1norm / maxnorm, &squarings); + if (squarings < 0) squarings = 0; + MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp(squarings)); + matrix_exp_pade7(A, U, V); + } + } +}; + +template +struct matrix_exp_computeUV +{ + static void run(const MatrixType& arg, MatrixType& U, MatrixType& V, int& squarings) + { + using std::frexp; + using std::pow; + const double l1norm = arg.cwiseAbs().colwise().sum().maxCoeff(); + squarings = 0; + if (l1norm < 1.495585217958292e-002) { + matrix_exp_pade3(arg, U, V); + } else if (l1norm < 2.539398330063230e-001) { + matrix_exp_pade5(arg, U, V); + } else if (l1norm < 9.504178996162932e-001) { + matrix_exp_pade7(arg, U, V); + } else if (l1norm < 2.097847961257068e+000) { + matrix_exp_pade9(arg, U, V); + } else { + const double maxnorm = 5.371920351148152; + frexp(l1norm / maxnorm, &squarings); + if (squarings < 0) squarings = 0; + MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp(squarings)); + matrix_exp_pade13(A, U, V); + } + } +}; + +template +struct matrix_exp_computeUV +{ + static void run(const MatrixType& arg, MatrixType& U, MatrixType& V, int& squarings) + { +#if LDBL_MANT_DIG == 53 // double precision + + matrix_exp_computeUV(arg, U, V, squarings); + +#else + + using std::frexp; + using std::pow; + const long double l1norm = arg.cwiseAbs().colwise().sum().maxCoeff(); + squarings = 0; + +#if LDBL_MANT_DIG <= 64 // extended precision + + if (l1norm < 4.1968497232266989671e-003L) { + matrix_exp_pade3(arg, U, V); + } else if (l1norm < 1.1848116734693823091e-001L) { + matrix_exp_pade5(arg, U, V); + } else if (l1norm < 5.5170388480686700274e-001L) { + matrix_exp_pade7(arg, U, V); + } else if (l1norm < 1.3759868875587845383e+000L) { + matrix_exp_pade9(arg, U, V); + } else { + const long double maxnorm = 4.0246098906697353063L; + frexp(l1norm / maxnorm, &squarings); + if (squarings < 0) squarings = 0; + MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp(squarings)); + matrix_exp_pade13(A, U, V); + } + +#elif LDBL_MANT_DIG <= 106 // double-double + + if (l1norm < 3.2787892205607026992947488108213e-005L) { + matrix_exp_pade3(arg, U, V); + } else if (l1norm < 6.4467025060072760084130906076332e-003L) { + matrix_exp_pade5(arg, U, V); + } else if (l1norm < 6.8988028496595374751374122881143e-002L) { + matrix_exp_pade7(arg, U, V); + } else if (l1norm < 2.7339737518502231741495857201670e-001L) { + matrix_exp_pade9(arg, U, V); + } else if (l1norm < 1.3203382096514474905666448850278e+000L) { + matrix_exp_pade13(arg, U, V); + } else { + const long double maxnorm = 3.2579440895405400856599663723517L; + frexp(l1norm / maxnorm, &squarings); + if (squarings < 0) squarings = 0; + MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp(squarings)); + matrix_exp_pade17(A, U, V); + } + +#elif LDBL_MANT_DIG <= 112 // quadruple precison + + if (l1norm < 1.639394610288918690547467954466970e-005L) { + matrix_exp_pade3(arg, U, V); + } else if (l1norm < 4.253237712165275566025884344433009e-003L) { + matrix_exp_pade5(arg, U, V); + } else if (l1norm < 5.125804063165764409885122032933142e-002L) { + matrix_exp_pade7(arg, U, V); + } else if (l1norm < 2.170000765161155195453205651889853e-001L) { + matrix_exp_pade9(arg, U, V); + } else if (l1norm < 1.125358383453143065081397882891878e+000L) { + matrix_exp_pade13(arg, U, V); + } else { + frexp(l1norm / maxnorm, &squarings); + if (squarings < 0) squarings = 0; + MatrixType A = arg.unaryExpr(MatrixExponentialScalingOp(squarings)); + matrix_exp_pade17(A, U, V); + } + +#else + + // this case should be handled in compute() + eigen_assert(false && "Bug in MatrixExponential"); + +#endif +#endif // LDBL_MANT_DIG + } +}; + + +/* Computes the matrix exponential + * + * \param arg argument of matrix exponential (should be plain object) + * \param result variable in which result will be stored + */ +template +void matrix_exp_compute(const MatrixType& arg, ResultType &result) { #if LDBL_MANT_DIG > 112 // rarely happens - if(sizeof(RealScalar) > 14) { - result = m_M.matrixFunction(StdStemFunctions::exp); + typedef typename traits::Scalar Scalar; + typedef typename NumTraits::Real RealScalar; + typedef typename std::complex ComplexScalar; + if (sizeof(RealScalar) > 14) { + result = arg.matrixFunction(StdStemFunctions::exp); return; } #endif - computeUV(RealScalar()); - m_tmp1 = m_U + m_V; // numerator of Pade approximant - m_tmp2 = -m_U + m_V; // denominator of Pade approximant - result = m_tmp2.partialPivLu().solve(m_tmp1); - for (int i=0; i::run(arg, U, V, squarings); // Pade approximant is (U+V) / (-U+V) + MatrixType numer = U + V; + MatrixType denom = -U + V; + result = denom.partialPivLu().solve(numer); + for (int i=0; i -EIGEN_STRONG_INLINE void MatrixExponential::pade3(const MatrixType &A) -{ - const RealScalar b[] = {120., 60., 12., 1.}; - m_tmp1.noalias() = A * A; - m_tmp2 = b[3]*m_tmp1 + b[1]*m_Id; - m_U.noalias() = A * m_tmp2; - m_V = b[2]*m_tmp1 + b[0]*m_Id; -} - -template -EIGEN_STRONG_INLINE void MatrixExponential::pade5(const MatrixType &A) -{ - const RealScalar b[] = {30240., 15120., 3360., 420., 30., 1.}; - MatrixType A2 = A * A; - m_tmp1.noalias() = A2 * A2; - m_tmp2 = b[5]*m_tmp1 + b[3]*A2 + b[1]*m_Id; - m_U.noalias() = A * m_tmp2; - m_V = b[4]*m_tmp1 + b[2]*A2 + b[0]*m_Id; -} - -template -EIGEN_STRONG_INLINE void MatrixExponential::pade7(const MatrixType &A) -{ - const RealScalar b[] = {17297280., 8648640., 1995840., 277200., 25200., 1512., 56., 1.}; - MatrixType A2 = A * A; - MatrixType A4 = A2 * A2; - m_tmp1.noalias() = A4 * A2; - m_tmp2 = b[7]*m_tmp1 + b[5]*A4 + b[3]*A2 + b[1]*m_Id; - m_U.noalias() = A * m_tmp2; - m_V = b[6]*m_tmp1 + b[4]*A4 + b[2]*A2 + b[0]*m_Id; -} - -template -EIGEN_STRONG_INLINE void MatrixExponential::pade9(const MatrixType &A) -{ - const RealScalar b[] = {17643225600., 8821612800., 2075673600., 302702400., 30270240., - 2162160., 110880., 3960., 90., 1.}; - MatrixType A2 = A * A; - MatrixType A4 = A2 * A2; - MatrixType A6 = A4 * A2; - m_tmp1.noalias() = A6 * A2; - m_tmp2 = b[9]*m_tmp1 + b[7]*A6 + b[5]*A4 + b[3]*A2 + b[1]*m_Id; - m_U.noalias() = A * m_tmp2; - m_V = b[8]*m_tmp1 + b[6]*A6 + b[4]*A4 + b[2]*A2 + b[0]*m_Id; -} - -template -EIGEN_STRONG_INLINE void MatrixExponential::pade13(const MatrixType &A) -{ - const RealScalar b[] = {64764752532480000., 32382376266240000., 7771770303897600., - 1187353796428800., 129060195264000., 10559470521600., 670442572800., - 33522128640., 1323241920., 40840800., 960960., 16380., 182., 1.}; - MatrixType A2 = A * A; - MatrixType A4 = A2 * A2; - m_tmp1.noalias() = A4 * A2; - m_V = b[13]*m_tmp1 + b[11]*A4 + b[9]*A2; // used for temporary storage - m_tmp2.noalias() = m_tmp1 * m_V; - m_tmp2 += b[7]*m_tmp1 + b[5]*A4 + b[3]*A2 + b[1]*m_Id; - m_U.noalias() = A * m_tmp2; - m_tmp2 = b[12]*m_tmp1 + b[10]*A4 + b[8]*A2; - m_V.noalias() = m_tmp1 * m_tmp2; - m_V += b[6]*m_tmp1 + b[4]*A4 + b[2]*A2 + b[0]*m_Id; -} - -#if LDBL_MANT_DIG > 64 -template -EIGEN_STRONG_INLINE void MatrixExponential::pade17(const MatrixType &A) -{ - const RealScalar b[] = {830034394580628357120000.L, 415017197290314178560000.L, - 100610229646136770560000.L, 15720348382208870400000.L, - 1774878043152614400000.L, 153822763739893248000.L, 10608466464820224000.L, - 595373117923584000.L, 27563570274240000.L, 1060137318240000.L, - 33924394183680.L, 899510451840.L, 19554575040.L, 341863200.L, 4651200.L, - 46512.L, 306.L, 1.L}; - MatrixType A2 = A * A; - MatrixType A4 = A2 * A2; - MatrixType A6 = A4 * A2; - m_tmp1.noalias() = A4 * A4; - m_V = b[17]*m_tmp1 + b[15]*A6 + b[13]*A4 + b[11]*A2; // used for temporary storage - m_tmp2.noalias() = m_tmp1 * m_V; - m_tmp2 += b[9]*m_tmp1 + b[7]*A6 + b[5]*A4 + b[3]*A2 + b[1]*m_Id; - m_U.noalias() = A * m_tmp2; - m_tmp2 = b[16]*m_tmp1 + b[14]*A6 + b[12]*A4 + b[10]*A2; - m_V.noalias() = m_tmp1 * m_tmp2; - m_V += b[8]*m_tmp1 + b[6]*A6 + b[4]*A4 + b[2]*A2 + b[0]*m_Id; -} -#endif - -template -void MatrixExponential::computeUV(float) -{ - using std::frexp; - if (m_l1norm < 4.258730016922831e-001) { - pade3(m_M); - } else if (m_l1norm < 1.880152677804762e+000) { - pade5(m_M); - } else { - const float maxnorm = 3.925724783138660f; - frexp(m_l1norm / maxnorm, &m_squarings); - if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); - pade7(A); - } -} - -template -void MatrixExponential::computeUV(double) -{ - using std::frexp; - if (m_l1norm < 1.495585217958292e-002) { - pade3(m_M); - } else if (m_l1norm < 2.539398330063230e-001) { - pade5(m_M); - } else if (m_l1norm < 9.504178996162932e-001) { - pade7(m_M); - } else if (m_l1norm < 2.097847961257068e+000) { - pade9(m_M); - } else { - const double maxnorm = 5.371920351148152; - frexp(m_l1norm / maxnorm, &m_squarings); - if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); - pade13(A); - } -} - -template -void MatrixExponential::computeUV(long double) -{ - using std::frexp; -#if LDBL_MANT_DIG == 53 // double precision - computeUV(double()); -#elif LDBL_MANT_DIG <= 64 // extended precision - if (m_l1norm < 4.1968497232266989671e-003L) { - pade3(m_M); - } else if (m_l1norm < 1.1848116734693823091e-001L) { - pade5(m_M); - } else if (m_l1norm < 5.5170388480686700274e-001L) { - pade7(m_M); - } else if (m_l1norm < 1.3759868875587845383e+000L) { - pade9(m_M); - } else { - const long double maxnorm = 4.0246098906697353063L; - frexp(m_l1norm / maxnorm, &m_squarings); - if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); - pade13(A); - } -#elif LDBL_MANT_DIG <= 106 // double-double - if (m_l1norm < 3.2787892205607026992947488108213e-005L) { - pade3(m_M); - } else if (m_l1norm < 6.4467025060072760084130906076332e-003L) { - pade5(m_M); - } else if (m_l1norm < 6.8988028496595374751374122881143e-002L) { - pade7(m_M); - } else if (m_l1norm < 2.7339737518502231741495857201670e-001L) { - pade9(m_M); - } else if (m_l1norm < 1.3203382096514474905666448850278e+000L) { - pade13(m_M); - } else { - const long double maxnorm = 3.2579440895405400856599663723517L; - frexp(m_l1norm / maxnorm, &m_squarings); - if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); - pade17(A); - } -#elif LDBL_MANT_DIG <= 112 // quadruple precison - if (m_l1norm < 1.639394610288918690547467954466970e-005L) { - pade3(m_M); - } else if (m_l1norm < 4.253237712165275566025884344433009e-003L) { - pade5(m_M); - } else if (m_l1norm < 5.125804063165764409885122032933142e-002L) { - pade7(m_M); - } else if (m_l1norm < 2.170000765161155195453205651889853e-001L) { - pade9(m_M); - } else if (m_l1norm < 1.125358383453143065081397882891878e+000L) { - pade13(m_M); - } else { - const long double maxnorm = 2.884233277829519311757165057717815L; - frexp(m_l1norm / maxnorm, &m_squarings); - if (m_squarings < 0) m_squarings = 0; - MatrixType A = CwiseUnaryOp(m_M, m_squarings); - pade17(A); - } -#else - // this case should be handled in compute() - eigen_assert(false && "Bug in MatrixExponential"); -#endif // LDBL_MANT_DIG -} +} // end namespace Eigen::internal /** \ingroup MatrixFunctions_Module * @@ -422,11 +370,9 @@ void MatrixExponential::computeUV(long double) * * \tparam Derived Type of the argument to the matrix exponential. * - * This class holds the argument to the matrix exponential until it - * is assigned or evaluated for some other reason (so the argument - * should not be changed in the meantime). It is the return type of - * MatrixBase::exp() and most of the time this is the only way it is - * used. + * This class holds the argument to the matrix exponential until it is assigned or evaluated for + * some other reason (so the argument should not be changed in the meantime). It is the return type + * of MatrixBase::exp() and most of the time this is the only way it is used. */ template struct MatrixExponentialReturnValue : public ReturnByValue > @@ -435,22 +381,19 @@ template struct MatrixExponentialReturnValue public: /** \brief Constructor. * - * \param[in] src %Matrix (expression) forming the argument of the - * matrix exponential. + * \param src %Matrix (expression) forming the argument of the matrix exponential. */ - explicit MatrixExponentialReturnValue(const Derived& src) : m_src(src) { } + MatrixExponentialReturnValue(const Derived& src) : m_src(src) { } /** \brief Compute the matrix exponential. * - * \param[out] result the matrix exponential of \p src in the - * constructor. + * \param result the matrix exponential of \p src in the constructor. */ template inline void evalTo(ResultType& result) const { const typename Derived::PlainObject srcEvaluated = m_src.eval(); - MatrixExponential me(srcEvaluated); - me.compute(result); + internal::matrix_exp_compute(srcEvaluated, result); } Index rows() const { return m_src.rows(); } @@ -458,6 +401,8 @@ template struct MatrixExponentialReturnValue protected: const Derived& m_src; + private: + MatrixExponentialReturnValue& operator=(const MatrixExponentialReturnValue&); }; namespace internal { From 084dc63b4ccfcc9a83a12973505af74a8bc32839 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Mon, 22 Jul 2013 13:56:15 +0100 Subject: [PATCH 0050/1103] Clean-up of MatrixSquareRoot. --- .../src/MatrixFunctions/MatrixLogarithm.h | 2 +- .../Eigen/src/MatrixFunctions/MatrixPower.h | 2 +- .../src/MatrixFunctions/MatrixSquareRoot.h | 415 +++++++----------- unsupported/test/matrix_power.cpp | 8 +- 4 files changed, 164 insertions(+), 263 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index 33cfadfb4..4b1eb5a34 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -141,7 +141,7 @@ void MatrixLogarithmAtomic::computeBig(const MatrixType& A, MatrixTy break; ++numberOfExtraSquareRoots; } - MatrixSquareRootTriangular(T).compute(sqrtT); + matrix_sqrt_triangular(T, sqrtT); T = sqrtT.template triangularView(); ++numberOfSquareRoots; } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index 5548bd95c..ee665c18e 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -219,7 +219,7 @@ void MatrixPowerAtomic::computeBig(ResultType& res) const break; hasExtraSquareRoot = true; } - MatrixSquareRootTriangular(T).compute(sqrtT); + matrix_sqrt_triangular(T, sqrtT); T = sqrtT.template triangularView(); ++numberOfSquareRoots; } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h index 0cd39ebe4..0261d4aa9 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h @@ -12,133 +12,16 @@ namespace Eigen { -/** \ingroup MatrixFunctions_Module - * \brief Class for computing matrix square roots of upper quasi-triangular matrices. - * \tparam MatrixType type of the argument of the matrix square root, - * expected to be an instantiation of the Matrix class template. - * - * This class computes the square root of the upper quasi-triangular - * matrix stored in the upper Hessenberg part of the matrix passed to - * the constructor. - * - * \sa MatrixSquareRoot, MatrixSquareRootTriangular - */ -template -class MatrixSquareRootQuasiTriangular : internal::noncopyable -{ - public: - - /** \brief Constructor. - * - * \param[in] A upper quasi-triangular matrix whose square root - * is to be computed. - * - * The class stores a reference to \p A, so it should not be - * changed (or destroyed) before compute() is called. - */ - explicit MatrixSquareRootQuasiTriangular(const MatrixType& A) - : m_A(A) - { - eigen_assert(A.rows() == A.cols()); - } - - /** \brief Compute the matrix square root - * - * \param[out] result square root of \p A, as specified in the constructor. - * - * Only the upper Hessenberg part of \p result is updated, the - * rest is not touched. See MatrixBase::sqrt() for details on - * how this computation is implemented. - */ - template void compute(ResultType &result); - - private: - typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - - void computeDiagonalPartOfSqrt(MatrixType& sqrtT, const MatrixType& T); - void computeOffDiagonalPartOfSqrt(MatrixType& sqrtT, const MatrixType& T); - void compute2x2diagonalBlock(MatrixType& sqrtT, const MatrixType& T, typename MatrixType::Index i); - void compute1x1offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j); - void compute1x2offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j); - void compute2x1offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j); - void compute2x2offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j); - - template - static void solveAuxiliaryEquation(SmallMatrixType& X, const SmallMatrixType& A, - const SmallMatrixType& B, const SmallMatrixType& C); - - const MatrixType& m_A; -}; - -template -template -void MatrixSquareRootQuasiTriangular::compute(ResultType &result) -{ - result.resize(m_A.rows(), m_A.cols()); - computeDiagonalPartOfSqrt(result, m_A); - computeOffDiagonalPartOfSqrt(result, m_A); -} - -// pre: T is quasi-upper-triangular and sqrtT is a zero matrix of the same size -// post: the diagonal blocks of sqrtT are the square roots of the diagonal blocks of T -template -void MatrixSquareRootQuasiTriangular::computeDiagonalPartOfSqrt(MatrixType& sqrtT, - const MatrixType& T) -{ - using std::sqrt; - const Index size = m_A.rows(); - for (Index i = 0; i < size; i++) { - if (i == size - 1 || T.coeff(i+1, i) == 0) { - eigen_assert(T(i,i) >= 0); - sqrtT.coeffRef(i,i) = sqrt(T.coeff(i,i)); - } - else { - compute2x2diagonalBlock(sqrtT, T, i); - ++i; - } - } -} - -// pre: T is quasi-upper-triangular and diagonal blocks of sqrtT are square root of diagonal blocks of T. -// post: sqrtT is the square root of T. -template -void MatrixSquareRootQuasiTriangular::computeOffDiagonalPartOfSqrt(MatrixType& sqrtT, - const MatrixType& T) -{ - const Index size = m_A.rows(); - for (Index j = 1; j < size; j++) { - if (T.coeff(j, j-1) != 0) // if T(j-1:j, j-1:j) is a 2-by-2 block - continue; - for (Index i = j-1; i >= 0; i--) { - if (i > 0 && T.coeff(i, i-1) != 0) // if T(i-1:i, i-1:i) is a 2-by-2 block - continue; - bool iBlockIs2x2 = (i < size - 1) && (T.coeff(i+1, i) != 0); - bool jBlockIs2x2 = (j < size - 1) && (T.coeff(j+1, j) != 0); - if (iBlockIs2x2 && jBlockIs2x2) - compute2x2offDiagonalBlock(sqrtT, T, i, j); - else if (iBlockIs2x2 && !jBlockIs2x2) - compute2x1offDiagonalBlock(sqrtT, T, i, j); - else if (!iBlockIs2x2 && jBlockIs2x2) - compute1x2offDiagonalBlock(sqrtT, T, i, j); - else if (!iBlockIs2x2 && !jBlockIs2x2) - compute1x1offDiagonalBlock(sqrtT, T, i, j); - } - } -} +namespace internal { // pre: T.block(i,i,2,2) has complex conjugate eigenvalues // post: sqrtT.block(i,i,2,2) is square root of T.block(i,i,2,2) -template -void MatrixSquareRootQuasiTriangular - ::compute2x2diagonalBlock(MatrixType& sqrtT, const MatrixType& T, typename MatrixType::Index i) +template +void matrix_sqrt_quasi_triangular_2x2_diagonal_block(const MatrixType& T, typename MatrixType::Index i, ResultType& sqrtT) { // TODO: This case (2-by-2 blocks with complex conjugate eigenvalues) is probably hidden somewhere // in EigenSolver. If we expose it, we could call it directly from here. + typedef typename traits::Scalar Scalar; Matrix block = T.template block<2,2>(i,i); EigenSolver > es(block); sqrtT.template block<2,2>(i,i) @@ -148,21 +31,19 @@ void MatrixSquareRootQuasiTriangular // pre: block structure of T is such that (i,j) is a 1x1 block, // all blocks of sqrtT to left of and below (i,j) are correct // post: sqrtT(i,j) has the correct value -template -void MatrixSquareRootQuasiTriangular - ::compute1x1offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j) +template +void matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT) { + typedef typename traits::Scalar Scalar; Scalar tmp = (sqrtT.row(i).segment(i+1,j-i-1) * sqrtT.col(j).segment(i+1,j-i-1)).value(); sqrtT.coeffRef(i,j) = (T.coeff(i,j) - tmp) / (sqrtT.coeff(i,i) + sqrtT.coeff(j,j)); } // similar to compute1x1offDiagonalBlock() -template -void MatrixSquareRootQuasiTriangular - ::compute1x2offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j) +template +void matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT) { + typedef typename traits::Scalar Scalar; Matrix rhs = T.template block<1,2>(i,j); if (j-i > 1) rhs -= sqrtT.block(i, i+1, 1, j-i-1) * sqrtT.block(i+1, j, j-i-1, 2); @@ -172,11 +53,10 @@ void MatrixSquareRootQuasiTriangular } // similar to compute1x1offDiagonalBlock() -template -void MatrixSquareRootQuasiTriangular - ::compute2x1offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j) +template +void matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT) { + typedef typename traits::Scalar Scalar; Matrix rhs = T.template block<2,1>(i,j); if (j-i > 2) rhs -= sqrtT.block(i, i+2, 2, j-i-2) * sqrtT.block(i+2, j, j-i-2, 1); @@ -186,31 +66,25 @@ void MatrixSquareRootQuasiTriangular } // similar to compute1x1offDiagonalBlock() -template -void MatrixSquareRootQuasiTriangular - ::compute2x2offDiagonalBlock(MatrixType& sqrtT, const MatrixType& T, - typename MatrixType::Index i, typename MatrixType::Index j) +template +void matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(const MatrixType& T, typename MatrixType::Index i, typename MatrixType::Index j, ResultType& sqrtT) { + typedef typename traits::Scalar Scalar; Matrix A = sqrtT.template block<2,2>(i,i); Matrix B = sqrtT.template block<2,2>(j,j); Matrix C = T.template block<2,2>(i,j); if (j-i > 2) C -= sqrtT.block(i, i+2, 2, j-i-2) * sqrtT.block(i+2, j, j-i-2, 2); Matrix X; - solveAuxiliaryEquation(X, A, B, C); + matrix_sqrt_quasi_triangular_solve_auxiliary_equation(X, A, B, C); sqrtT.template block<2,2>(i,j) = X; } // solves the equation A X + X B = C where all matrices are 2-by-2 template -template -void MatrixSquareRootQuasiTriangular - ::solveAuxiliaryEquation(SmallMatrixType& X, const SmallMatrixType& A, - const SmallMatrixType& B, const SmallMatrixType& C) +void matrix_sqrt_quasi_triangular_solve_auxiliary_equation(MatrixType& X, const MatrixType& A, const MatrixType& B, const MatrixType& C) { - EIGEN_STATIC_ASSERT((internal::is_same >::value), - EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT); - + typedef typename traits::Scalar Scalar; Matrix coeffMatrix = Matrix::Zero(); coeffMatrix.coeffRef(0,0) = A.coeff(0,0) + B.coeff(0,0); coeffMatrix.coeffRef(1,1) = A.coeff(0,0) + B.coeff(1,1); @@ -241,164 +115,193 @@ void MatrixSquareRootQuasiTriangular } +// pre: T is quasi-upper-triangular and sqrtT is a zero matrix of the same size +// post: the diagonal blocks of sqrtT are the square roots of the diagonal blocks of T +template +void matrix_sqrt_quasi_triangular_diagonal(const MatrixType& T, ResultType& sqrtT) +{ + using std::sqrt; + typedef typename MatrixType::Index Index; + const Index size = T.rows(); + for (Index i = 0; i < size; i++) { + if (i == size - 1 || T.coeff(i+1, i) == 0) { + eigen_assert(T(i,i) >= 0); + sqrtT.coeffRef(i,i) = sqrt(T.coeff(i,i)); + } + else { + matrix_sqrt_quasi_triangular_2x2_diagonal_block(T, i, sqrtT); + ++i; + } + } +} + +// pre: T is quasi-upper-triangular and diagonal blocks of sqrtT are square root of diagonal blocks of T. +// post: sqrtT is the square root of T. +template +void matrix_sqrt_quasi_triangular_off_diagonal(const MatrixType& T, ResultType& sqrtT) +{ + typedef typename MatrixType::Index Index; + const Index size = T.rows(); + for (Index j = 1; j < size; j++) { + if (T.coeff(j, j-1) != 0) // if T(j-1:j, j-1:j) is a 2-by-2 block + continue; + for (Index i = j-1; i >= 0; i--) { + if (i > 0 && T.coeff(i, i-1) != 0) // if T(i-1:i, i-1:i) is a 2-by-2 block + continue; + bool iBlockIs2x2 = (i < size - 1) && (T.coeff(i+1, i) != 0); + bool jBlockIs2x2 = (j < size - 1) && (T.coeff(j+1, j) != 0); + if (iBlockIs2x2 && jBlockIs2x2) + matrix_sqrt_quasi_triangular_2x2_off_diagonal_block(T, i, j, sqrtT); + else if (iBlockIs2x2 && !jBlockIs2x2) + matrix_sqrt_quasi_triangular_2x1_off_diagonal_block(T, i, j, sqrtT); + else if (!iBlockIs2x2 && jBlockIs2x2) + matrix_sqrt_quasi_triangular_1x2_off_diagonal_block(T, i, j, sqrtT); + else if (!iBlockIs2x2 && !jBlockIs2x2) + matrix_sqrt_quasi_triangular_1x1_off_diagonal_block(T, i, j, sqrtT); + } + } +} + +} // end of namespace internal + /** \ingroup MatrixFunctions_Module - * \brief Class for computing matrix square roots of upper triangular matrices. - * \tparam MatrixType type of the argument of the matrix square root, - * expected to be an instantiation of the Matrix class template. + * \brief Compute matrix square root of quasi-triangular matrix. * - * This class computes the square root of the upper triangular matrix - * stored in the upper triangular part (including the diagonal) of - * the matrix passed to the constructor. + * \tparam MatrixType type of \p arg, the argument of matrix square root, + * expected to be an instantiation of the Matrix class template. + * \tparam ResultType type of \p result, where result is to be stored. + * \param[in] arg argument of matrix square root. + * \param[out] result matrix square root of upper Hessenberg part of \p arg. + * + * This function computes the square root of the upper quasi-triangular matrix stored in the upper + * Hessenberg part of \p arg. Only the upper Hessenberg part of \p result is updated, the rest is + * not touched. See MatrixBase::sqrt() for details on how this computation is implemented. * * \sa MatrixSquareRoot, MatrixSquareRootQuasiTriangular */ -template -class MatrixSquareRootTriangular : internal::noncopyable +template +void matrix_sqrt_quasi_triangular(const MatrixType &arg, ResultType &result) { - public: - explicit MatrixSquareRootTriangular(const MatrixType& A) - : m_A(A) - { - eigen_assert(A.rows() == A.cols()); - } + eigen_assert(arg.rows() == arg.cols()); + result.resize(arg.rows(), arg.cols()); + internal::matrix_sqrt_quasi_triangular_diagonal(arg, result); + internal::matrix_sqrt_quasi_triangular_off_diagonal(arg, result); +} - /** \brief Compute the matrix square root - * - * \param[out] result square root of \p A, as specified in the constructor. - * - * Only the upper triangular part (including the diagonal) of - * \p result is updated, the rest is not touched. See - * MatrixBase::sqrt() for details on how this computation is - * implemented. - */ - template void compute(ResultType &result); - private: - const MatrixType& m_A; -}; - -template -template -void MatrixSquareRootTriangular::compute(ResultType &result) +/** \ingroup MatrixFunctions_Module + * \brief Compute matrix square root of triangular matrix. + * + * \tparam MatrixType type of \p arg, the argument of matrix square root, + * expected to be an instantiation of the Matrix class template. + * \tparam ResultType type of \p result, where result is to be stored. + * \param[in] arg argument of matrix square root. + * \param[out] result matrix square root of upper triangular part of \p arg. + * + * Only the upper triangular part (including the diagonal) of \p result is updated, the rest is not + * touched. See MatrixBase::sqrt() for details on how this computation is implemented. + * + * \sa MatrixSquareRoot, MatrixSquareRootQuasiTriangular + */ +template +void matrix_sqrt_triangular(const MatrixType &arg, ResultType &result) { using std::sqrt; - - // Compute square root of m_A and store it in upper triangular part of result - // This uses that the square root of triangular matrices can be computed directly. - result.resize(m_A.rows(), m_A.cols()); typedef typename MatrixType::Index Index; - for (Index i = 0; i < m_A.rows(); i++) { - result.coeffRef(i,i) = sqrt(m_A.coeff(i,i)); - } - for (Index j = 1; j < m_A.cols(); j++) { - for (Index i = j-1; i >= 0; i--) { typedef typename MatrixType::Scalar Scalar; + + eigen_assert(arg.rows() == arg.cols()); + + // Compute square root of arg and store it in upper triangular part of result + // This uses that the square root of triangular matrices can be computed directly. + result.resize(arg.rows(), arg.cols()); + for (Index i = 0; i < arg.rows(); i++) { + result.coeffRef(i,i) = sqrt(arg.coeff(i,i)); + } + for (Index j = 1; j < arg.cols(); j++) { + for (Index i = j-1; i >= 0; i--) { // if i = j-1, then segment has length 0 so tmp = 0 Scalar tmp = (result.row(i).segment(i+1,j-i-1) * result.col(j).segment(i+1,j-i-1)).value(); // denominator may be zero if original matrix is singular - result.coeffRef(i,j) = (m_A.coeff(i,j) - tmp) / (result.coeff(i,i) + result.coeff(j,j)); + result.coeffRef(i,j) = (arg.coeff(i,j) - tmp) / (result.coeff(i,i) + result.coeff(j,j)); } } } +namespace internal { + /** \ingroup MatrixFunctions_Module - * \brief Class for computing matrix square roots of general matrices. + * \brief Helper struct for computing matrix square roots of general matrices. * \tparam MatrixType type of the argument of the matrix square root, * expected to be an instantiation of the Matrix class template. * * \sa MatrixSquareRootTriangular, MatrixSquareRootQuasiTriangular, MatrixBase::sqrt() */ template ::Scalar>::IsComplex> -class MatrixSquareRoot +struct matrix_sqrt_compute { - public: - - /** \brief Constructor. - * - * \param[in] A matrix whose square root is to be computed. - * - * The class stores a reference to \p A, so it should not be - * changed (or destroyed) before compute() is called. - */ - explicit MatrixSquareRoot(const MatrixType& A); - - /** \brief Compute the matrix square root - * - * \param[out] result square root of \p A, as specified in the constructor. - * - * See MatrixBase::sqrt() for details on how this computation is - * implemented. - */ - template void compute(ResultType &result); + /** \brief Compute the matrix square root + * + * \param[in] arg matrix whose square root is to be computed. + * \param[out] result square root of \p arg. + * + * See MatrixBase::sqrt() for details on how this computation is implemented. + */ + template static void run(const MatrixType &arg, ResultType &result); }; // ********** Partial specialization for real matrices ********** template -class MatrixSquareRoot +struct matrix_sqrt_compute { - public: + template + static void run(const MatrixType &arg, ResultType &result) + { + eigen_assert(arg.rows() == arg.cols()); - explicit MatrixSquareRoot(const MatrixType& A) - : m_A(A) - { - eigen_assert(A.rows() == A.cols()); - } - - template void compute(ResultType &result) - { - // Compute Schur decomposition of m_A - const RealSchur schurOfA(m_A); - const MatrixType& T = schurOfA.matrixT(); - const MatrixType& U = schurOfA.matrixU(); + // Compute Schur decomposition of arg + const RealSchur schurOfA(arg); + const MatrixType& T = schurOfA.matrixT(); + const MatrixType& U = schurOfA.matrixU(); - // Compute square root of T - MatrixType sqrtT = MatrixType::Zero(m_A.rows(), m_A.cols()); - MatrixSquareRootQuasiTriangular(T).compute(sqrtT); + // Compute square root of T + MatrixType sqrtT = MatrixType::Zero(arg.rows(), arg.cols()); + matrix_sqrt_quasi_triangular(T, sqrtT); - // Compute square root of m_A - result = U * sqrtT * U.adjoint(); - } - - private: - const MatrixType& m_A; + // Compute square root of arg + result = U * sqrtT * U.adjoint(); + } }; // ********** Partial specialization for complex matrices ********** template -class MatrixSquareRoot : internal::noncopyable +struct matrix_sqrt_compute { - public: + template + static void run(const MatrixType &arg, ResultType &result) + { + eigen_assert(arg.rows() == arg.cols()); - explicit MatrixSquareRoot(const MatrixType& A) - : m_A(A) - { - eigen_assert(A.rows() == A.cols()); - } - - template void compute(ResultType &result) - { - // Compute Schur decomposition of m_A - const ComplexSchur schurOfA(m_A); - const MatrixType& T = schurOfA.matrixT(); - const MatrixType& U = schurOfA.matrixU(); + // Compute Schur decomposition of arg + const ComplexSchur schurOfA(arg); + const MatrixType& T = schurOfA.matrixT(); + const MatrixType& U = schurOfA.matrixU(); - // Compute square root of T - MatrixType sqrtT; - MatrixSquareRootTriangular(T).compute(sqrtT); + // Compute square root of T + MatrixType sqrtT; + matrix_sqrt_triangular(T, sqrtT); - // Compute square root of m_A - result = U * (sqrtT.template triangularView() * U.adjoint()); - } - - private: - const MatrixType& m_A; + // Compute square root of arg + result = U * (sqrtT.template triangularView() * U.adjoint()); + } }; +} // end namespace internal /** \ingroup MatrixFunctions_Module * @@ -432,9 +335,9 @@ template class MatrixSquareRootReturnValue template inline void evalTo(ResultType& result) const { - const typename Derived::PlainObject srcEvaluated = m_src.eval(); - MatrixSquareRoot me(srcEvaluated); - me.compute(result); + typedef typename Derived::PlainObject PlainObject; + const PlainObject srcEvaluated = m_src.eval(); + internal::matrix_sqrt_compute::run(srcEvaluated, result); } Index rows() const { return m_src.rows(); } diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index 849e4287b..4c4cac509 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -100,8 +100,6 @@ template void testSingular(MatrixType m, double tol) { const int IsComplex = NumTraits::Scalar>::IsComplex; - typedef typename internal::conditional< IsComplex, MatrixSquareRootTriangular, - MatrixSquareRootQuasiTriangular >::type SquareRootType; typedef typename internal::conditional, const MatrixType&>::type TriangularType; typename internal::conditional< IsComplex, ComplexSchur, RealSchur >::type schur; MatrixType T; @@ -116,13 +114,13 @@ void testSingular(MatrixType m, double tol) processTriangularMatrix::run(m, T, U); MatrixPower mpow(m); - SquareRootType(T).compute(T); + T = T.sqrt(); VERIFY(mpow(0.5).isApprox(U * (TriangularType(T) * U.adjoint()), tol)); - SquareRootType(T).compute(T); + T = T.sqrt(); VERIFY(mpow(0.25).isApprox(U * (TriangularType(T) * U.adjoint()), tol)); - SquareRootType(T).compute(T); + T = T.sqrt(); VERIFY(mpow(0.125).isApprox(U * (TriangularType(T) * U.adjoint()), tol)); } } From a3a55357db7394281c872e911f13d69aba510aec Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Thu, 25 Jul 2013 15:08:53 +0100 Subject: [PATCH 0051/1103] Clean up MatrixFunction and MatrixLogarithm. --- .../src/MatrixFunctions/MatrixFunction.h | 684 +++++++++--------- .../MatrixFunctions/MatrixFunctionAtomic.h | 127 ---- .../src/MatrixFunctions/MatrixLogarithm.h | 466 +++++------- .../src/MatrixFunctions/MatrixSquareRoot.h | 2 +- 4 files changed, 517 insertions(+), 762 deletions(-) delete mode 100644 unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index ad54d8ed0..12e28793d 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009-2011 Jitse Niesen +// Copyright (C) 2009-2011, 2013 Jitse Niesen // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -11,394 +11,244 @@ #define EIGEN_MATRIX_FUNCTION #include "StemFunction.h" -#include "MatrixFunctionAtomic.h" namespace Eigen { +namespace internal { + +/** \brief Maximum distance allowed between eigenvalues to be considered "close". */ +static const float matrix_function_separation = 0.1; + /** \ingroup MatrixFunctions_Module - * \brief Class for computing matrix functions. - * \tparam MatrixType type of the argument of the matrix function, - * expected to be an instantiation of the Matrix class template. - * \tparam AtomicType type for computing matrix function of atomic blocks. - * \tparam IsComplex used internally to select correct specialization. + * \class MatrixFunctionAtomic + * \brief Helper class for computing matrix functions of atomic matrices. * - * This class implements the Schur-Parlett algorithm for computing matrix functions. The spectrum of the - * matrix is divided in clustered of eigenvalues that lies close together. This class delegates the - * computation of the matrix function on every block corresponding to these clusters to an object of type - * \p AtomicType and uses these results to compute the matrix function of the whole matrix. The class - * \p AtomicType should have a \p compute() member function for computing the matrix function of a block. - * - * \sa class MatrixFunctionAtomic, class MatrixLogarithmAtomic + * Here, an atomic matrix is a triangular matrix whose diagonal entries are close to each other. */ -template ::Scalar>::IsComplex> -class MatrixFunction : internal::noncopyable -{ - public: - - /** \brief Constructor. - * - * \param[in] A argument of matrix function, should be a square matrix. - * \param[in] atomic class for computing matrix function of atomic blocks. - * - * The class stores references to \p A and \p atomic, so they should not be - * changed (or destroyed) before compute() is called. - */ - MatrixFunction(const MatrixType& A, AtomicType& atomic); - - /** \brief Compute the matrix function. - * - * \param[out] result the function \p f applied to \p A, as - * specified in the constructor. - * - * See MatrixBase::matrixFunction() for details on how this computation - * is implemented. - */ - template - void compute(ResultType &result); -}; - - -/** \internal \ingroup MatrixFunctions_Module - * \brief Partial specialization of MatrixFunction for real matrices - */ -template -class MatrixFunction : internal::noncopyable -{ - private: - - typedef internal::traits Traits; - typedef typename Traits::Scalar Scalar; - static const int Rows = Traits::RowsAtCompileTime; - static const int Cols = Traits::ColsAtCompileTime; - static const int Options = MatrixType::Options; - static const int MaxRows = Traits::MaxRowsAtCompileTime; - static const int MaxCols = Traits::MaxColsAtCompileTime; - - typedef std::complex ComplexScalar; - typedef Matrix ComplexMatrix; - - public: - - /** \brief Constructor. - * - * \param[in] A argument of matrix function, should be a square matrix. - * \param[in] atomic class for computing matrix function of atomic blocks. - */ - MatrixFunction(const MatrixType& A, AtomicType& atomic) : m_A(A), m_atomic(atomic) { } - - /** \brief Compute the matrix function. - * - * \param[out] result the function \p f applied to \p A, as - * specified in the constructor. - * - * This function converts the real matrix \c A to a complex matrix, - * uses MatrixFunction and then converts the result back to - * a real matrix. - */ - template - void compute(ResultType& result) - { - ComplexMatrix CA = m_A.template cast(); - ComplexMatrix Cresult; - MatrixFunction mf(CA, m_atomic); - mf.compute(Cresult); - result = Cresult.real(); - } - - private: - typename internal::nested::type m_A; /**< \brief Reference to argument of matrix function. */ - AtomicType& m_atomic; /**< \brief Class for computing matrix function of atomic blocks. */ -}; - - -/** \internal \ingroup MatrixFunctions_Module - * \brief Partial specialization of MatrixFunction for complex matrices - */ -template -class MatrixFunction : internal::noncopyable +template +class MatrixFunctionAtomic { - private: + public: - typedef internal::traits Traits; typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; - static const int RowsAtCompileTime = Traits::RowsAtCompileTime; - static const int ColsAtCompileTime = Traits::ColsAtCompileTime; - static const int Options = MatrixType::Options; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - typedef Matrix IntVectorType; - typedef Matrix DynamicIntVectorType; - typedef std::list Cluster; - typedef std::list ListOfClusters; - typedef Matrix DynMatrixType; + typedef typename stem_function::type StemFunction; - public: + /** \brief Constructor + * \param[in] f matrix function to compute. + */ + MatrixFunctionAtomic(StemFunction f) : m_f(f) { } - MatrixFunction(const MatrixType& A, AtomicType& atomic); - template void compute(ResultType& result); + /** \brief Compute matrix function of atomic matrix + * \param[in] A argument of matrix function, should be upper triangular and atomic + * \returns f(A), the matrix function evaluated at the given matrix + */ + MatrixType compute(const MatrixType& A); private: - - void computeSchurDecomposition(); - void partitionEigenvalues(); - typename ListOfClusters::iterator findCluster(Scalar key); - void computeClusterSize(); - void computeBlockStart(); - void constructPermutation(); - void permuteSchur(); - void swapEntriesInSchur(Index index); - void computeBlockAtomic(); - Block block(MatrixType& A, Index i, Index j); - void computeOffDiagonal(); - DynMatrixType solveTriangularSylvester(const DynMatrixType& A, const DynMatrixType& B, const DynMatrixType& C); - - typename internal::nested::type m_A; /**< \brief Reference to argument of matrix function. */ - AtomicType& m_atomic; /**< \brief Class for computing matrix function of atomic blocks. */ - MatrixType m_T; /**< \brief Triangular part of Schur decomposition */ - MatrixType m_U; /**< \brief Unitary part of Schur decomposition */ - MatrixType m_fT; /**< \brief %Matrix function applied to #m_T */ - ListOfClusters m_clusters; /**< \brief Partition of eigenvalues into clusters of ei'vals "close" to each other */ - DynamicIntVectorType m_eivalToCluster; /**< \brief m_eivalToCluster[i] = j means i-th ei'val is in j-th cluster */ - DynamicIntVectorType m_clusterSize; /**< \brief Number of eigenvalues in each clusters */ - DynamicIntVectorType m_blockStart; /**< \brief Row index at which block corresponding to i-th cluster starts */ - IntVectorType m_permutation; /**< \brief Permutation which groups ei'vals in the same cluster together */ - - /** \brief Maximum distance allowed between eigenvalues to be considered "close". - * - * This is morally a \c static \c const \c Scalar, but only - * integers can be static constant class members in C++. The - * separation constant is set to 0.1, a value taken from the - * paper by Davies and Higham. */ - static const RealScalar separation() { return static_cast(0.1); } + StemFunction* m_f; }; -/** \brief Constructor. - * - * \param[in] A argument of matrix function, should be a square matrix. - * \param[in] atomic class for computing matrix function of atomic blocks. - */ -template -MatrixFunction::MatrixFunction(const MatrixType& A, AtomicType& atomic) - : m_A(A), m_atomic(atomic) +template +typename NumTraits::Real matrix_function_compute_mu(const MatrixType& A) { - /* empty body */ + typedef typename plain_col_type::type VectorType; + typename MatrixType::Index rows = A.rows(); + const MatrixType N = MatrixType::Identity(rows, rows) - A; + VectorType e = VectorType::Ones(rows); + N.template triangularView().solveInPlace(e); + return e.cwiseAbs().maxCoeff(); } -/** \brief Compute the matrix function. - * - * \param[out] result the function \p f applied to \p A, as - * specified in the constructor. +template +MatrixType MatrixFunctionAtomic::compute(const MatrixType& A) +{ + // TODO: Use that A is upper triangular + typedef typename NumTraits::Real RealScalar; + typedef typename MatrixType::Index Index; + Index rows = A.rows(); + Scalar avgEival = A.trace() / Scalar(RealScalar(rows)); + MatrixType Ashifted = A - avgEival * MatrixType::Identity(rows, rows); + RealScalar mu = matrix_function_compute_mu(Ashifted); + MatrixType F = m_f(avgEival, 0) * MatrixType::Identity(rows, rows); + MatrixType P = Ashifted; + MatrixType Fincr; + for (Index s = 1; s < 1.1 * rows + 10; s++) { // upper limit is fairly arbitrary + Fincr = m_f(avgEival, static_cast(s)) * P; + F += Fincr; + P = Scalar(RealScalar(1.0/(s + 1))) * P * Ashifted; + + // test whether Taylor series converged + const RealScalar F_norm = F.cwiseAbs().rowwise().sum().maxCoeff(); + const RealScalar Fincr_norm = Fincr.cwiseAbs().rowwise().sum().maxCoeff(); + if (Fincr_norm < NumTraits::epsilon() * F_norm) { + RealScalar delta = 0; + RealScalar rfactorial = 1; + for (Index r = 0; r < rows; r++) { + RealScalar mx = 0; + for (Index i = 0; i < rows; i++) + mx = (std::max)(mx, std::abs(m_f(Ashifted(i, i) + avgEival, static_cast(s+r)))); + if (r != 0) + rfactorial *= RealScalar(r); + delta = (std::max)(delta, mx / rfactorial); + } + const RealScalar P_norm = P.cwiseAbs().rowwise().sum().maxCoeff(); + if (mu * delta * P_norm < NumTraits::epsilon() * F_norm) // series converged + break; + } + } + return F; +} + +/** \brief Find cluster in \p clusters containing some value + * \param[in] key Value to find + * \returns Iterator to cluster containing \p key, or \c clusters.end() if no cluster in \p m_clusters + * contains \p key. */ -template -template -void MatrixFunction::compute(ResultType& result) +template +typename ListOfClusters::iterator matrix_function_find_cluster(Scalar key, ListOfClusters& clusters) { - computeSchurDecomposition(); - partitionEigenvalues(); - computeClusterSize(); - computeBlockStart(); - constructPermutation(); - permuteSchur(); - computeBlockAtomic(); - computeOffDiagonal(); - result = m_U * (m_fT.template triangularView() * m_U.adjoint()); -} - -/** \brief Store the Schur decomposition of #m_A in #m_T and #m_U */ -template -void MatrixFunction::computeSchurDecomposition() -{ - const ComplexSchur schurOfA(m_A); - m_T = schurOfA.matrixT(); - m_U = schurOfA.matrixU(); + typename std::list::iterator j; + for (typename ListOfClusters::iterator i = clusters.begin(); i != clusters.end(); ++i) { + j = std::find(i->begin(), i->end(), key); + if (j != i->end()) + return i; + } + return clusters.end(); } /** \brief Partition eigenvalues in clusters of ei'vals close to each other * - * This function computes #m_clusters. This is a partition of the - * eigenvalues of #m_T in clusters, such that - * # Any eigenvalue in a certain cluster is at most separation() away - * from another eigenvalue in the same cluster. - * # The distance between two eigenvalues in different clusters is - * more than separation(). - * The implementation follows Algorithm 4.1 in the paper of Davies - * and Higham. + * \param[in] eivals Eigenvalues + * \param[out] clusters Resulting partition of eigenvalues + * + * The partition satisfies the following two properties: + * # Any eigenvalue in a certain cluster is at most matrix_function_separation() away from another eigenvalue + * in the same cluster. + * # The distance between two eigenvalues in different clusters is more than matrix_function_separation(). + * The implementation follows Algorithm 4.1 in the paper of Davies and Higham. */ -template -void MatrixFunction::partitionEigenvalues() +template +void matrix_function_partition_eigenvalues(const EivalsType& eivals, std::list& clusters) { - using std::abs; - const Index rows = m_T.rows(); - VectorType diag = m_T.diagonal(); // contains eigenvalues of A - - for (Index i=0; i::iterator qi = matrix_function_find_cluster(eivals(i), clusters); + if (qi == clusters.end()) { Cluster l; - l.push_back(diag(i)); - m_clusters.push_back(l); - qi = m_clusters.end(); + l.push_back(eivals(i)); + clusters.push_back(l); + qi = clusters.end(); --qi; } // Look for other element to add to the set - for (Index j=i+1; jbegin(), qi->end(), diag(j)) == qi->end()) { - typename ListOfClusters::iterator qj = findCluster(diag(j)); - if (qj == m_clusters.end()) { - qi->push_back(diag(j)); + for (Index j=i+1; jbegin(), qi->end(), eivals(j)) == qi->end()) { + typename std::list::iterator qj = matrix_function_find_cluster(eivals(j), clusters); + if (qj == clusters.end()) { + qi->push_back(eivals(j)); } else { qi->insert(qi->end(), qj->begin(), qj->end()); - m_clusters.erase(qj); + clusters.erase(qj); } } } } } -/** \brief Find cluster in #m_clusters containing some value - * \param[in] key Value to find - * \returns Iterator to cluster containing \c key, or - * \c m_clusters.end() if no cluster in m_clusters contains \c key. - */ -template -typename MatrixFunction::ListOfClusters::iterator MatrixFunction::findCluster(Scalar key) +/** \brief Compute size of each cluster given a partitioning */ +template +void matrix_function_compute_cluster_size(const ListOfClusters& clusters, Matrix& clusterSize) { - typename Cluster::iterator j; - for (typename ListOfClusters::iterator i = m_clusters.begin(); i != m_clusters.end(); ++i) { - j = std::find(i->begin(), i->end(), key); - if (j != i->end()) - return i; + const Index numClusters = static_cast(clusters.size()); + clusterSize.setZero(numClusters); + Index clusterIndex = 0; + for (typename ListOfClusters::const_iterator cluster = clusters.begin(); cluster != clusters.end(); ++cluster) { + clusterSize[clusterIndex] = cluster->size(); + ++clusterIndex; } - return m_clusters.end(); } -/** \brief Compute #m_clusterSize and #m_eivalToCluster using #m_clusters */ -template -void MatrixFunction::computeClusterSize() +/** \brief Compute start of each block using clusterSize */ +template +void matrix_function_compute_block_start(const VectorType& clusterSize, VectorType& blockStart) { - const Index rows = m_T.rows(); - VectorType diag = m_T.diagonal(); - const Index numClusters = static_cast(m_clusters.size()); + blockStart.resize(clusterSize.rows()); + blockStart(0) = 0; + for (typename VectorType::Index i = 1; i < clusterSize.rows(); i++) { + blockStart(i) = blockStart(i-1) + clusterSize(i-1); + } +} - m_clusterSize.setZero(numClusters); - m_eivalToCluster.resize(rows); +/** \brief Compute mapping of eigenvalue indices to cluster indices */ +template +void matrix_function_compute_map(const EivalsType& eivals, const ListOfClusters& clusters, VectorType& eivalToCluster) +{ + typedef typename EivalsType::Index Index; + eivalToCluster.resize(eivals.rows()); Index clusterIndex = 0; - for (typename ListOfClusters::const_iterator cluster = m_clusters.begin(); cluster != m_clusters.end(); ++cluster) { - for (Index i = 0; i < diag.rows(); ++i) { - if (std::find(cluster->begin(), cluster->end(), diag(i)) != cluster->end()) { - ++m_clusterSize[clusterIndex]; - m_eivalToCluster[i] = clusterIndex; + for (typename ListOfClusters::const_iterator cluster = clusters.begin(); cluster != clusters.end(); ++cluster) { + for (Index i = 0; i < eivals.rows(); ++i) { + if (std::find(cluster->begin(), cluster->end(), eivals(i)) != cluster->end()) { + eivalToCluster[i] = clusterIndex; } } ++clusterIndex; } } -/** \brief Compute #m_blockStart using #m_clusterSize */ -template -void MatrixFunction::computeBlockStart() +/** \brief Compute permutation which groups ei'vals in same cluster together */ +template +void matrix_function_compute_permutation(const DynVectorType& blockStart, const DynVectorType& eivalToCluster, VectorType& permutation) { - m_blockStart.resize(m_clusterSize.rows()); - m_blockStart(0) = 0; - for (Index i = 1; i < m_clusterSize.rows(); i++) { - m_blockStart(i) = m_blockStart(i-1) + m_clusterSize(i-1); - } -} - -/** \brief Compute #m_permutation using #m_eivalToCluster and #m_blockStart */ -template -void MatrixFunction::constructPermutation() -{ - DynamicIntVectorType indexNextEntry = m_blockStart; - m_permutation.resize(m_T.rows()); - for (Index i = 0; i < m_T.rows(); i++) { - Index cluster = m_eivalToCluster[i]; - m_permutation[i] = indexNextEntry[cluster]; + typedef typename VectorType::Index Index; + DynVectorType indexNextEntry = blockStart; + permutation.resize(eivalToCluster.rows()); + for (Index i = 0; i < eivalToCluster.rows(); i++) { + Index cluster = eivalToCluster[i]; + permutation[i] = indexNextEntry[cluster]; ++indexNextEntry[cluster]; } } -/** \brief Permute Schur decomposition in #m_U and #m_T according to #m_permutation */ -template -void MatrixFunction::permuteSchur() +/** \brief Permute Schur decomposition in U and T according to permutation */ +template +void matrix_function_permute_schur(VectorType& permutation, MatrixType& U, MatrixType& T) { - IntVectorType p = m_permutation; - for (Index i = 0; i < p.rows() - 1; i++) { + typedef typename VectorType::Index Index; + for (Index i = 0; i < permutation.rows() - 1; i++) { Index j; - for (j = i; j < p.rows(); j++) { - if (p(j) == i) break; + for (j = i; j < permutation.rows(); j++) { + if (permutation(j) == i) break; } - eigen_assert(p(j) == i); + eigen_assert(permutation(j) == i); for (Index k = j-1; k >= i; k--) { - swapEntriesInSchur(k); - std::swap(p.coeffRef(k), p.coeffRef(k+1)); + JacobiRotation rotation; + rotation.makeGivens(T(k, k+1), T(k+1, k+1) - T(k, k)); + T.applyOnTheLeft(k, k+1, rotation.adjoint()); + T.applyOnTheRight(k, k+1, rotation); + U.applyOnTheRight(k, k+1, rotation); + std::swap(permutation.coeffRef(k), permutation.coeffRef(k+1)); } } } -/** \brief Swap rows \a index and \a index+1 in Schur decomposition in #m_U and #m_T */ -template -void MatrixFunction::swapEntriesInSchur(Index index) -{ - JacobiRotation rotation; - rotation.makeGivens(m_T(index, index+1), m_T(index+1, index+1) - m_T(index, index)); - m_T.applyOnTheLeft(index, index+1, rotation.adjoint()); - m_T.applyOnTheRight(index, index+1, rotation); - m_U.applyOnTheRight(index, index+1, rotation); -} - -/** \brief Compute block diagonal part of #m_fT. +/** \brief Compute block diagonal part of matrix function. * - * This routine computes the matrix function applied to the block diagonal part of #m_T, with the blocking - * given by #m_blockStart. The matrix function of each diagonal block is computed by #m_atomic. The - * off-diagonal parts of #m_fT are set to zero. + * This routine computes the matrix function applied to the block diagonal part of \p T (which should be + * upper triangular), with the blocking given by \p blockStart and \p clusterSize. The matrix function of + * each diagonal block is computed by \p atomic. The off-diagonal parts of \p fT are set to zero. */ -template -void MatrixFunction::computeBlockAtomic() +template +void matrix_function_compute_block_atomic(const MatrixType& T, AtomicType& atomic, const VectorType& blockStart, const VectorType& clusterSize, MatrixType& fT) { - m_fT.resize(m_T.rows(), m_T.cols()); - m_fT.setZero(); - for (Index i = 0; i < m_clusterSize.rows(); ++i) { - block(m_fT, i, i) = m_atomic.compute(block(m_T, i, i)); - } -} - -/** \brief Return block of matrix according to blocking given by #m_blockStart */ -template -Block MatrixFunction::block(MatrixType& A, Index i, Index j) -{ - return A.block(m_blockStart(i), m_blockStart(j), m_clusterSize(i), m_clusterSize(j)); -} - -/** \brief Compute part of #m_fT above block diagonal. - * - * This routine assumes that the block diagonal part of #m_fT (which - * equals the matrix function applied to #m_T) has already been computed and computes - * the part above the block diagonal. The part below the diagonal is - * zero, because #m_T is upper triangular. - */ -template -void MatrixFunction::computeOffDiagonal() -{ - for (Index diagIndex = 1; diagIndex < m_clusterSize.rows(); diagIndex++) { - for (Index blockIndex = 0; blockIndex < m_clusterSize.rows() - diagIndex; blockIndex++) { - // compute (blockIndex, blockIndex+diagIndex) block - DynMatrixType A = block(m_T, blockIndex, blockIndex); - DynMatrixType B = -block(m_T, blockIndex+diagIndex, blockIndex+diagIndex); - DynMatrixType C = block(m_fT, blockIndex, blockIndex) * block(m_T, blockIndex, blockIndex+diagIndex); - C -= block(m_T, blockIndex, blockIndex+diagIndex) * block(m_fT, blockIndex+diagIndex, blockIndex+diagIndex); - for (Index k = blockIndex + 1; k < blockIndex + diagIndex; k++) { - C += block(m_fT, blockIndex, k) * block(m_T, k, blockIndex+diagIndex); - C -= block(m_T, blockIndex, k) * block(m_fT, k, blockIndex+diagIndex); - } - block(m_fT, blockIndex, blockIndex+diagIndex) = solveTriangularSylvester(A, B, C); - } + fT.setZero(T.rows(), T.cols()); + for (typename VectorType::Index i = 0; i < clusterSize.rows(); ++i) { + fT.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i)) + = atomic.compute(T.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i))); } } @@ -410,8 +260,8 @@ void MatrixFunction::computeOffDiagonal() * * \returns the solution X. * - * If A is m-by-m and B is n-by-n, then both C and X are m-by-n. - * The (i,j)-th component of the Sylvester equation is + * If A is m-by-m and B is n-by-n, then both C and X are m-by-n. The (i,j)-th component of the Sylvester + * equation is * \f[ * \sum_{k=i}^m A_{ik} X_{kj} + \sum_{k=1}^j X_{ik} B_{kj} = C_{ij}. * \f] @@ -420,16 +270,12 @@ void MatrixFunction::computeOffDiagonal() * X_{ij} = \frac{1}{A_{ii} + B_{jj}} \Bigl( C_{ij} * - \sum_{k=i+1}^m A_{ik} X_{kj} - \sum_{k=1}^{j-1} X_{ik} B_{kj} \Bigr). * \f] - * It is assumed that A and B are such that the numerator is never - * zero (otherwise the Sylvester equation does not have a unique - * solution). In that case, these equations can be evaluated in the - * order \f$ i=m,\ldots,1 \f$ and \f$ j=1,\ldots,n \f$. + * It is assumed that A and B are such that the numerator is never zero (otherwise the Sylvester equation + * does not have a unique solution). In that case, these equations can be evaluated in the order + * \f$ i=m,\ldots,1 \f$ and \f$ j=1,\ldots,n \f$. */ -template -typename MatrixFunction::DynMatrixType MatrixFunction::solveTriangularSylvester( - const DynMatrixType& A, - const DynMatrixType& B, - const DynMatrixType& C) +template +MatrixType matrix_function_solve_triangular_sylvester(const MatrixType& A, const MatrixType& B, const MatrixType& C) { eigen_assert(A.rows() == A.cols()); eigen_assert(A.isUpperTriangular()); @@ -438,9 +284,12 @@ typename MatrixFunction::DynMatrixType MatrixFunction= 0; --i) { for (Index j = 0; j < n; ++j) { @@ -469,17 +318,164 @@ typename MatrixFunction::DynMatrixType MatrixFunction +void matrix_function_compute_above_diagonal(const MatrixType& T, const VectorType& blockStart, const VectorType& clusterSize, MatrixType& fT) +{ + typedef internal::traits Traits; + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::Index Index; + static const int RowsAtCompileTime = Traits::RowsAtCompileTime; + static const int ColsAtCompileTime = Traits::ColsAtCompileTime; + static const int Options = MatrixType::Options; + typedef Matrix DynMatrixType; + + for (Index k = 1; k < clusterSize.rows(); k++) { + for (Index i = 0; i < clusterSize.rows() - k; i++) { + // compute (i, i+k) block + DynMatrixType A = T.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i)); + DynMatrixType B = -T.block(blockStart(i+k), blockStart(i+k), clusterSize(i+k), clusterSize(i+k)); + DynMatrixType C = fT.block(blockStart(i), blockStart(i), clusterSize(i), clusterSize(i)) + * T.block(blockStart(i), blockStart(i+k), clusterSize(i), clusterSize(i+k)); + C -= T.block(blockStart(i), blockStart(i+k), clusterSize(i), clusterSize(i+k)) + * fT.block(blockStart(i+k), blockStart(i+k), clusterSize(i+k), clusterSize(i+k)); + for (Index m = i + 1; m < i + k; m++) { + C += fT.block(blockStart(i), blockStart(m), clusterSize(i), clusterSize(m)) + * T.block(blockStart(m), blockStart(i+k), clusterSize(m), clusterSize(i+k)); + C -= T.block(blockStart(i), blockStart(m), clusterSize(i), clusterSize(m)) + * fT.block(blockStart(m), blockStart(i+k), clusterSize(m), clusterSize(i+k)); + } + fT.block(blockStart(i), blockStart(i+k), clusterSize(i), clusterSize(i+k)) + = matrix_function_solve_triangular_sylvester(A, B, C); + } + } +} + +/** \ingroup MatrixFunctions_Module + * \brief Class for computing matrix functions. + * \tparam MatrixType type of the argument of the matrix function, + * expected to be an instantiation of the Matrix class template. + * \tparam AtomicType type for computing matrix function of atomic blocks. + * \tparam IsComplex used internally to select correct specialization. + * + * This class implements the Schur-Parlett algorithm for computing matrix functions. The spectrum of the + * matrix is divided in clustered of eigenvalues that lies close together. This class delegates the + * computation of the matrix function on every block corresponding to these clusters to an object of type + * \p AtomicType and uses these results to compute the matrix function of the whole matrix. The class + * \p AtomicType should have a \p compute() member function for computing the matrix function of a block. + * + * \sa class MatrixFunctionAtomic, class MatrixLogarithmAtomic + */ +template ::Scalar>::IsComplex> +struct matrix_function_compute +{ + /** \brief Compute the matrix function. + * + * \param[in] A argument of matrix function, should be a square matrix. + * \param[in] atomic class for computing matrix function of atomic blocks. + * \param[out] result the function \p f applied to \p A, as + * specified in the constructor. + * + * See MatrixBase::matrixFunction() for details on how this computation + * is implemented. + */ + template + static void run(const MatrixType& A, AtomicType& atomic, ResultType &result); +}; + +/** \internal \ingroup MatrixFunctions_Module + * \brief Partial specialization of MatrixFunction for real matrices + * + * This converts the real matrix to a complex matrix, compute the matrix function of that matrix, and then + * converts the result back to a real matrix. + */ +template +struct matrix_function_compute +{ + template + static void run(const MatrixType& A, AtomicType& atomic, ResultType &result) + { + typedef internal::traits Traits; + typedef typename Traits::Scalar Scalar; + static const int Rows = Traits::RowsAtCompileTime, Cols = Traits::ColsAtCompileTime; + static const int Options = MatrixType::Options; + static const int MaxRows = Traits::MaxRowsAtCompileTime, MaxCols = Traits::MaxColsAtCompileTime; + + typedef std::complex ComplexScalar; + typedef Matrix ComplexMatrix; + + ComplexMatrix CA = A.template cast(); + ComplexMatrix Cresult; + matrix_function_compute::run(CA, atomic, Cresult); + result = Cresult.real(); + } +}; + +/** \internal \ingroup MatrixFunctions_Module + * \brief Partial specialization of MatrixFunction for complex matrices + */ +template +struct matrix_function_compute +{ + template + static void run(const MatrixType& A, AtomicType& atomic, ResultType &result) + { + typedef internal::traits Traits; + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::Index Index; + + // compute Schur decomposition of A + const ComplexSchur schurOfA(A); + MatrixType T = schurOfA.matrixT(); + MatrixType U = schurOfA.matrixU(); + + // partition eigenvalues into clusters of ei'vals "close" to each other + std::list > clusters; + matrix_function_partition_eigenvalues(T.diagonal(), clusters); + + // compute size of each cluster + Matrix clusterSize; + matrix_function_compute_cluster_size(clusters, clusterSize); + + // blockStart[i] is row index at which block corresponding to i-th cluster starts + Matrix blockStart; + matrix_function_compute_block_start(clusterSize, blockStart); + + // compute map so that eivalToCluster[i] = j means that ei'val T(i,i) is in j-th cluster + Matrix eivalToCluster; + matrix_function_compute_map(T.diagonal(), clusters, eivalToCluster); + + // compute permutation which groups ei'vals in same cluster together + Matrix permutation; + matrix_function_compute_permutation(blockStart, eivalToCluster, permutation); + + // permute Schur decomposition + matrix_function_permute_schur(permutation, U, T); + + // compute result + MatrixType fT; // matrix function applied to T + matrix_function_compute_block_atomic(T, atomic, blockStart, clusterSize, fT); + matrix_function_compute_above_diagonal(T, blockStart, clusterSize, fT); + result = U * (fT.template triangularView() * U.adjoint()); + } +}; + +} // end of namespace internal + /** \ingroup MatrixFunctions_Module * * \brief Proxy for the matrix function of some matrix (expression). * * \tparam Derived Type of the argument to the matrix function. * - * This class holds the argument to the matrix function until it is - * assigned or evaluated for some other reason (so the argument - * should not be changed in the meantime). It is the return type of - * matrixBase::matrixFunction() and related functions and most of the - * time this is the only way it is used. + * This class holds the argument to the matrix function until it is assigned or evaluated for some other + * reason (so the argument should not be changed in the meantime). It is the return type of + * matrixBase::matrixFunction() and related functions and most of the time this is the only way it is used. */ template class MatrixFunctionReturnValue : public ReturnByValue > @@ -490,18 +486,16 @@ template class MatrixFunctionReturnValue typedef typename Derived::Index Index; typedef typename internal::stem_function::type StemFunction; - /** \brief Constructor. + /** \brief Constructor. * - * \param[in] A %Matrix (expression) forming the argument of the - * matrix function. + * \param[in] A %Matrix (expression) forming the argument of the matrix function. * \param[in] f Stem function for matrix function under consideration. */ MatrixFunctionReturnValue(const Derived& A, StemFunction f) : m_A(A), m_f(f) { } /** \brief Compute the matrix function. * - * \param[out] result \p f applied to \p A, where \p f and \p A - * are as in the constructor. + * \param[out] result \p f applied to \p A, where \p f and \p A are as in the constructor. */ template inline void evalTo(ResultType& result) const @@ -513,12 +507,12 @@ template class MatrixFunctionReturnValue static const int Options = PlainObject::Options; typedef std::complex::Real> ComplexScalar; typedef Matrix DynMatrixType; - typedef MatrixFunctionAtomic AtomicType; + + typedef internal::MatrixFunctionAtomic AtomicType; AtomicType atomic(m_f); const PlainObject Aevaluated = m_A.eval(); - MatrixFunction mf(Aevaluated, atomic); - mf.compute(result); + internal::matrix_function_compute::run(Aevaluated, atomic, result); } Index rows() const { return m_A.rows(); } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h deleted file mode 100644 index d6ff5f1ce..000000000 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunctionAtomic.h +++ /dev/null @@ -1,127 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2009 Jitse Niesen -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_MATRIX_FUNCTION_ATOMIC -#define EIGEN_MATRIX_FUNCTION_ATOMIC - -namespace Eigen { - -/** \ingroup MatrixFunctions_Module - * \class MatrixFunctionAtomic - * \brief Helper class for computing matrix functions of atomic matrices. - * - * \internal - * Here, an atomic matrix is a triangular matrix whose diagonal - * entries are close to each other. - */ -template -class MatrixFunctionAtomic : internal::noncopyable -{ - public: - - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; - typedef typename NumTraits::Real RealScalar; - typedef typename internal::stem_function::type StemFunction; - typedef Matrix VectorType; - - /** \brief Constructor - * \param[in] f matrix function to compute. - */ - MatrixFunctionAtomic(StemFunction f) : m_f(f) { } - - /** \brief Compute matrix function of atomic matrix - * \param[in] A argument of matrix function, should be upper triangular and atomic - * \returns f(A), the matrix function evaluated at the given matrix - */ - MatrixType compute(const MatrixType& A); - - private: - - void computeMu(); - bool taylorConverged(Index s, const MatrixType& F, const MatrixType& Fincr, const MatrixType& P); - - /** \brief Pointer to scalar function */ - StemFunction* m_f; - - /** \brief Size of matrix function */ - Index m_Arows; - - /** \brief Mean of eigenvalues */ - Scalar m_avgEival; - - /** \brief Argument shifted by mean of eigenvalues */ - MatrixType m_Ashifted; - - /** \brief Constant used to determine whether Taylor series has converged */ - RealScalar m_mu; -}; - -template -MatrixType MatrixFunctionAtomic::compute(const MatrixType& A) -{ - // TODO: Use that A is upper triangular - m_Arows = A.rows(); - m_avgEival = A.trace() / Scalar(RealScalar(m_Arows)); - m_Ashifted = A - m_avgEival * MatrixType::Identity(m_Arows, m_Arows); - computeMu(); - MatrixType F = m_f(m_avgEival, 0) * MatrixType::Identity(m_Arows, m_Arows); - MatrixType P = m_Ashifted; - MatrixType Fincr; - for (Index s = 1; s < 1.1 * m_Arows + 10; s++) { // upper limit is fairly arbitrary - Fincr = m_f(m_avgEival, static_cast(s)) * P; - F += Fincr; - P = Scalar(RealScalar(1.0/(s + 1))) * P * m_Ashifted; - if (taylorConverged(s, F, Fincr, P)) { - return F; - } - } - eigen_assert("Taylor series does not converge" && 0); - return F; -} - -/** \brief Compute \c m_mu. */ -template -void MatrixFunctionAtomic::computeMu() -{ - const MatrixType N = MatrixType::Identity(m_Arows, m_Arows) - m_Ashifted; - VectorType e = VectorType::Ones(m_Arows); - N.template triangularView().solveInPlace(e); - m_mu = e.cwiseAbs().maxCoeff(); -} - -/** \brief Determine whether Taylor series has converged */ -template -bool MatrixFunctionAtomic::taylorConverged(Index s, const MatrixType& F, - const MatrixType& Fincr, const MatrixType& P) -{ - const Index n = F.rows(); - const RealScalar F_norm = F.cwiseAbs().rowwise().sum().maxCoeff(); - const RealScalar Fincr_norm = Fincr.cwiseAbs().rowwise().sum().maxCoeff(); - if (Fincr_norm < NumTraits::epsilon() * F_norm) { - RealScalar delta = 0; - RealScalar rfactorial = 1; - for (Index r = 0; r < n; r++) { - RealScalar mx = 0; - for (Index i = 0; i < n; i++) - mx = (std::max)(mx, std::abs(m_f(m_Ashifted(i, i) + m_avgEival, static_cast(s+r)))); - if (r != 0) - rfactorial *= RealScalar(r); - delta = (std::max)(delta, mx / rfactorial); - } - const RealScalar P_norm = P.cwiseAbs().rowwise().sum().maxCoeff(); - if (m_mu * delta * P_norm < NumTraits::epsilon() * F_norm) - return true; - } - return false; -} - -} // end namespace Eigen - -#endif // EIGEN_MATRIX_FUNCTION_ATOMIC diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index 4b1eb5a34..6f84a31bd 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -17,81 +17,30 @@ namespace Eigen { -/** \ingroup MatrixFunctions_Module - * \class MatrixLogarithmAtomic - * \brief Helper class for computing matrix logarithm of atomic matrices. - * - * \internal - * Here, an atomic matrix is a triangular matrix whose diagonal - * entries are close to each other. - * - * \sa class MatrixFunctionAtomic, MatrixBase::log() - */ -template -class MatrixLogarithmAtomic : internal::noncopyable +namespace internal { + +template +struct matrix_log_min_pade_degree { -public: - - typedef typename MatrixType::Scalar Scalar; - // typedef typename MatrixType::Index Index; - typedef typename NumTraits::Real RealScalar; - // typedef typename internal::stem_function::type StemFunction; - // typedef Matrix VectorType; - - /** \brief Constructor. */ - MatrixLogarithmAtomic() { } - - /** \brief Compute matrix logarithm of atomic matrix - * \param[in] A argument of matrix logarithm, should be upper triangular and atomic - * \returns The logarithm of \p A. - */ - MatrixType compute(const MatrixType& A); - -private: - - void compute2x2(const MatrixType& A, MatrixType& result); - void computeBig(const MatrixType& A, MatrixType& result); - int getPadeDegree(float normTminusI); - int getPadeDegree(double normTminusI); - int getPadeDegree(long double normTminusI); - void computePade(MatrixType& result, const MatrixType& T, int degree); - void computePade3(MatrixType& result, const MatrixType& T); - void computePade4(MatrixType& result, const MatrixType& T); - void computePade5(MatrixType& result, const MatrixType& T); - void computePade6(MatrixType& result, const MatrixType& T); - void computePade7(MatrixType& result, const MatrixType& T); - void computePade8(MatrixType& result, const MatrixType& T); - void computePade9(MatrixType& result, const MatrixType& T); - void computePade10(MatrixType& result, const MatrixType& T); - void computePade11(MatrixType& result, const MatrixType& T); - - static const int minPadeDegree = 3; - static const int maxPadeDegree = std::numeric_limits::digits<= 24? 5: // single precision - std::numeric_limits::digits<= 53? 7: // double precision - std::numeric_limits::digits<= 64? 8: // extended precision - std::numeric_limits::digits<=106? 10: // double-double - 11; // quadruple precision + static const int value = 3; }; -/** \brief Compute logarithm of triangular matrix with clustered eigenvalues. */ -template -MatrixType MatrixLogarithmAtomic::compute(const MatrixType& A) +template +struct matrix_log_max_pade_degree { - using std::log; - MatrixType result(A.rows(), A.rows()); - if (A.rows() == 1) - result(0,0) = log(A(0,0)); - else if (A.rows() == 2) - compute2x2(A, result); - else - computeBig(A, result); - return result; -} + typedef typename NumTraits::Real RealScalar; + static const int value = std::numeric_limits::digits<= 24? 5: // single precision + std::numeric_limits::digits<= 53? 7: // double precision + std::numeric_limits::digits<= 64? 8: // extended precision + std::numeric_limits::digits<=106? 10: // double-double + 11; // quadruple precision +}; /** \brief Compute logarithm of 2x2 triangular matrix. */ template -void MatrixLogarithmAtomic::compute2x2(const MatrixType& A, MatrixType& result) +void matrix_log_compute_2x2(const MatrixType& A, MatrixType& result) { + typedef typename MatrixType::Scalar Scalar; using std::abs; using std::ceil; using std::imag; @@ -116,47 +65,14 @@ void MatrixLogarithmAtomic::compute2x2(const MatrixType& A, MatrixTy } } -/** \brief Compute logarithm of triangular matrices with size > 2. - * \details This uses a inverse scale-and-square algorithm. */ -template -void MatrixLogarithmAtomic::computeBig(const MatrixType& A, MatrixType& result) -{ - using std::pow; - int numberOfSquareRoots = 0; - int numberOfExtraSquareRoots = 0; - int degree; - MatrixType T = A, sqrtT; - const RealScalar maxNormForPade = maxPadeDegree<= 5? 5.3149729967117310e-1: // single precision - maxPadeDegree<= 7? 2.6429608311114350e-1: // double precision - maxPadeDegree<= 8? 2.32777776523703892094e-1L: // extended precision - maxPadeDegree<=10? 1.05026503471351080481093652651105e-1L: // double-double - 1.1880960220216759245467951592883642e-1L; // quadruple precision - - while (true) { - RealScalar normTminusI = (T - MatrixType::Identity(T.rows(), T.rows())).cwiseAbs().colwise().sum().maxCoeff(); - if (normTminusI < maxNormForPade) { - degree = getPadeDegree(normTminusI); - int degree2 = getPadeDegree(normTminusI / RealScalar(2)); - if ((degree - degree2 <= 1) || (numberOfExtraSquareRoots == 1)) - break; - ++numberOfExtraSquareRoots; - } - matrix_sqrt_triangular(T, sqrtT); - T = sqrtT.template triangularView(); - ++numberOfSquareRoots; - } - - computePade(result, T, degree); - result *= pow(RealScalar(2), numberOfSquareRoots); -} - /* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = float) */ -template -int MatrixLogarithmAtomic::getPadeDegree(float normTminusI) +int matrix_log_get_pade_degree(float normTminusI) { const float maxNormForPade[] = { 2.5111573934555054e-1 /* degree = 3 */ , 4.0535837411880493e-1, 5.3149729967117310e-1 }; - int degree = 3; + const int minPadeDegree = matrix_log_min_pade_degree::value; + const int maxPadeDegree = matrix_log_max_pade_degree::value; + int degree = minPadeDegree; for (; degree <= maxPadeDegree; ++degree) if (normTminusI <= maxNormForPade[degree - minPadeDegree]) break; @@ -164,12 +80,13 @@ int MatrixLogarithmAtomic::getPadeDegree(float normTminusI) } /* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = double) */ -template -int MatrixLogarithmAtomic::getPadeDegree(double normTminusI) +int matrix_log_get_pade_degree(double normTminusI) { const double maxNormForPade[] = { 1.6206284795015624e-2 /* degree = 3 */ , 5.3873532631381171e-2, 1.1352802267628681e-1, 1.8662860613541288e-1, 2.642960831111435e-1 }; - int degree = 3; + const int minPadeDegree = matrix_log_min_pade_degree::value; + const int maxPadeDegree = matrix_log_max_pade_degree::value; + int degree = minPadeDegree; for (; degree <= maxPadeDegree; ++degree) if (normTminusI <= maxNormForPade[degree - minPadeDegree]) break; @@ -177,8 +94,7 @@ int MatrixLogarithmAtomic::getPadeDegree(double normTminusI) } /* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = long double) */ -template -int MatrixLogarithmAtomic::getPadeDegree(long double normTminusI) +int matrix_log_get_pade_degree(long double normTminusI) { #if LDBL_MANT_DIG == 53 // double precision const long double maxNormForPade[] = { 1.6206284795015624e-2L /* degree = 3 */ , 5.3873532631381171e-2L, @@ -200,7 +116,9 @@ int MatrixLogarithmAtomic::getPadeDegree(long double normTminusI) 3.6688019729653446926585242192447447e-2L, 5.9290962294020186998954055264528393e-2L, 8.6998436081634343903250580992127677e-2L, 1.1880960220216759245467951592883642e-1L }; #endif - int degree = 3; + const int minPadeDegree = matrix_log_min_pade_degree::value; + const int maxPadeDegree = matrix_log_max_pade_degree::value; + int degree = minPadeDegree; for (; degree <= maxPadeDegree; ++degree) if (normTminusI <= maxNormForPade[degree - minPadeDegree]) break; @@ -209,197 +127,168 @@ int MatrixLogarithmAtomic::getPadeDegree(long double normTminusI) /* \brief Compute Pade approximation to matrix logarithm */ template -void MatrixLogarithmAtomic::computePade(MatrixType& result, const MatrixType& T, int degree) +void matrix_log_compute_pade(MatrixType& result, const MatrixType& T, int degree) { - switch (degree) { - case 3: computePade3(result, T); break; - case 4: computePade4(result, T); break; - case 5: computePade5(result, T); break; - case 6: computePade6(result, T); break; - case 7: computePade7(result, T); break; - case 8: computePade8(result, T); break; - case 9: computePade9(result, T); break; - case 10: computePade10(result, T); break; - case 11: computePade11(result, T); break; - default: assert(false); // should never happen + typedef typename NumTraits::Real RealScalar; + const int minPadeDegree = 3; + const int maxPadeDegree = 11; + assert(degree >= minPadeDegree && degree <= maxPadeDegree); + + const RealScalar nodes[][maxPadeDegree] = { + { 0.1127016653792583114820734600217600L, 0.5000000000000000000000000000000000L, // degree 3 + 0.8872983346207416885179265399782400L }, + { 0.0694318442029737123880267555535953L, 0.3300094782075718675986671204483777L, // degree 4 + 0.6699905217924281324013328795516223L, 0.9305681557970262876119732444464048L }, + { 0.0469100770306680036011865608503035L, 0.2307653449471584544818427896498956L, // degree 5 + 0.5000000000000000000000000000000000L, 0.7692346550528415455181572103501044L, + 0.9530899229693319963988134391496965L }, + { 0.0337652428984239860938492227530027L, 0.1693953067668677431693002024900473L, // degree 6 + 0.3806904069584015456847491391596440L, 0.6193095930415984543152508608403560L, + 0.8306046932331322568306997975099527L, 0.9662347571015760139061507772469973L }, + { 0.0254460438286207377369051579760744L, 0.1292344072003027800680676133596058L, // degree 7 + 0.2970774243113014165466967939615193L, 0.5000000000000000000000000000000000L, + 0.7029225756886985834533032060384807L, 0.8707655927996972199319323866403942L, + 0.9745539561713792622630948420239256L }, + { 0.0198550717512318841582195657152635L, 0.1016667612931866302042230317620848L, // degree 8 + 0.2372337950418355070911304754053768L, 0.4082826787521750975302619288199080L, + 0.5917173212478249024697380711800920L, 0.7627662049581644929088695245946232L, + 0.8983332387068133697957769682379152L, 0.9801449282487681158417804342847365L }, + { 0.0159198802461869550822118985481636L, 0.0819844463366821028502851059651326L, // degree 9 + 0.1933142836497048013456489803292629L, 0.3378732882980955354807309926783317L, + 0.5000000000000000000000000000000000L, 0.6621267117019044645192690073216683L, + 0.8066857163502951986543510196707371L, 0.9180155536633178971497148940348674L, + 0.9840801197538130449177881014518364L }, + { 0.0130467357414141399610179939577740L, 0.0674683166555077446339516557882535L, // degree 10 + 0.1602952158504877968828363174425632L, 0.2833023029353764046003670284171079L, + 0.4255628305091843945575869994351400L, 0.5744371694908156054424130005648600L, + 0.7166976970646235953996329715828921L, 0.8397047841495122031171636825574368L, + 0.9325316833444922553660483442117465L, 0.9869532642585858600389820060422260L }, + { 0.0108856709269715035980309994385713L, 0.0564687001159523504624211153480364L, // degree 11 + 0.1349239972129753379532918739844233L, 0.2404519353965940920371371652706952L, + 0.3652284220238275138342340072995692L, 0.5000000000000000000000000000000000L, + 0.6347715779761724861657659927004308L, 0.7595480646034059079628628347293048L, + 0.8650760027870246620467081260155767L, 0.9435312998840476495375788846519636L, + 0.9891143290730284964019690005614287L } }; + + const RealScalar weights[][maxPadeDegree] = { + { 0.2777777777777777777777777777777778L, 0.4444444444444444444444444444444444L, // degree 3 + 0.2777777777777777777777777777777778L }, + { 0.1739274225687269286865319746109997L, 0.3260725774312730713134680253890003L, // degree 4 + 0.3260725774312730713134680253890003L, 0.1739274225687269286865319746109997L }, + { 0.1184634425280945437571320203599587L, 0.2393143352496832340206457574178191L, // degree 5 + 0.2844444444444444444444444444444444L, 0.2393143352496832340206457574178191L, + 0.1184634425280945437571320203599587L }, + { 0.0856622461895851725201480710863665L, 0.1803807865240693037849167569188581L, // degree 6 + 0.2339569672863455236949351719947755L, 0.2339569672863455236949351719947755L, + 0.1803807865240693037849167569188581L, 0.0856622461895851725201480710863665L }, + { 0.0647424830844348466353057163395410L, 0.1398526957446383339507338857118898L, // degree 7 + 0.1909150252525594724751848877444876L, 0.2089795918367346938775510204081633L, + 0.1909150252525594724751848877444876L, 0.1398526957446383339507338857118898L, + 0.0647424830844348466353057163395410L }, + { 0.0506142681451881295762656771549811L, 0.1111905172266872352721779972131204L, // degree 8 + 0.1568533229389436436689811009933007L, 0.1813418916891809914825752246385978L, + 0.1813418916891809914825752246385978L, 0.1568533229389436436689811009933007L, + 0.1111905172266872352721779972131204L, 0.0506142681451881295762656771549811L }, + { 0.0406371941807872059859460790552618L, 0.0903240803474287020292360156214564L, // degree 9 + 0.1303053482014677311593714347093164L, 0.1561735385200014200343152032922218L, + 0.1651196775006298815822625346434870L, 0.1561735385200014200343152032922218L, + 0.1303053482014677311593714347093164L, 0.0903240803474287020292360156214564L, + 0.0406371941807872059859460790552618L }, + { 0.0333356721543440687967844049466659L, 0.0747256745752902965728881698288487L, // degree 10 + 0.1095431812579910219977674671140816L, 0.1346333596549981775456134607847347L, + 0.1477621123573764350869464973256692L, 0.1477621123573764350869464973256692L, + 0.1346333596549981775456134607847347L, 0.1095431812579910219977674671140816L, + 0.0747256745752902965728881698288487L, 0.0333356721543440687967844049466659L }, + { 0.0278342835580868332413768602212743L, 0.0627901847324523123173471496119701L, // degree 11 + 0.0931451054638671257130488207158280L, 0.1165968822959952399592618524215876L, + 0.1314022722551233310903444349452546L, 0.1364625433889503153572417641681711L, + 0.1314022722551233310903444349452546L, 0.1165968822959952399592618524215876L, + 0.0931451054638671257130488207158280L, 0.0627901847324523123173471496119701L, + 0.0278342835580868332413768602212743L } }; + + MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); + result.setZero(T.rows(), T.rows()); + for (int k = 0; k < degree; ++k) { + RealScalar weight = weights[degree-minPadeDegree][k]; + RealScalar node = nodes[degree-minPadeDegree][k]; + result += weight * (MatrixType::Identity(T.rows(), T.rows()) + node * TminusI) + .template triangularView().solve(TminusI); } } +/** \brief Compute logarithm of triangular matrices with size > 2. + * \details This uses a inverse scale-and-square algorithm. */ template -void MatrixLogarithmAtomic::computePade3(MatrixType& result, const MatrixType& T) +void matrix_log_compute_big(const MatrixType& A, MatrixType& result) { - const int degree = 3; - const RealScalar nodes[] = { 0.1127016653792583114820734600217600L, 0.5000000000000000000000000000000000L, - 0.8872983346207416885179265399782400L }; - const RealScalar weights[] = { 0.2777777777777777777777777777777778L, 0.4444444444444444444444444444444444L, - 0.2777777777777777777777777777777778L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); + typedef typename MatrixType::Scalar Scalar; + typedef typename NumTraits::Real RealScalar; + using std::pow; + + int numberOfSquareRoots = 0; + int numberOfExtraSquareRoots = 0; + int degree; + MatrixType T = A, sqrtT; + + int maxPadeDegree = matrix_log_max_pade_degree::value; + const RealScalar maxNormForPade = maxPadeDegree<= 5? 5.3149729967117310e-1: // single precision + maxPadeDegree<= 7? 2.6429608311114350e-1: // double precision + maxPadeDegree<= 8? 2.32777776523703892094e-1L: // extended precision + maxPadeDegree<=10? 1.05026503471351080481093652651105e-1L: // double-double + 1.1880960220216759245467951592883642e-1L; // quadruple precision + + while (true) { + RealScalar normTminusI = (T - MatrixType::Identity(T.rows(), T.rows())).cwiseAbs().colwise().sum().maxCoeff(); + if (normTminusI < maxNormForPade) { + degree = matrix_log_get_pade_degree(normTminusI); + int degree2 = matrix_log_get_pade_degree(normTminusI / RealScalar(2)); + if ((degree - degree2 <= 1) || (numberOfExtraSquareRoots == 1)) + break; + ++numberOfExtraSquareRoots; + } + matrix_sqrt_triangular(T, sqrtT); + T = sqrtT.template triangularView(); + ++numberOfSquareRoots; + } + + matrix_log_compute_pade(result, T, degree); + result *= pow(RealScalar(2), numberOfSquareRoots); } +/** \ingroup MatrixFunctions_Module + * \class MatrixLogarithmAtomic + * \brief Helper class for computing matrix logarithm of atomic matrices. + * + * Here, an atomic matrix is a triangular matrix whose diagonal entries are close to each other. + * + * \sa class MatrixFunctionAtomic, MatrixBase::log() + */ template -void MatrixLogarithmAtomic::computePade4(MatrixType& result, const MatrixType& T) +class MatrixLogarithmAtomic { - const int degree = 4; - const RealScalar nodes[] = { 0.0694318442029737123880267555535953L, 0.3300094782075718675986671204483777L, - 0.6699905217924281324013328795516223L, 0.9305681557970262876119732444464048L }; - const RealScalar weights[] = { 0.1739274225687269286865319746109997L, 0.3260725774312730713134680253890003L, - 0.3260725774312730713134680253890003L, 0.1739274225687269286865319746109997L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); -} +public: + /** \brief Compute matrix logarithm of atomic matrix + * \param[in] A argument of matrix logarithm, should be upper triangular and atomic + * \returns The logarithm of \p A. + */ + MatrixType compute(const MatrixType& A); +}; template -void MatrixLogarithmAtomic::computePade5(MatrixType& result, const MatrixType& T) +MatrixType MatrixLogarithmAtomic::compute(const MatrixType& A) { - const int degree = 5; - const RealScalar nodes[] = { 0.0469100770306680036011865608503035L, 0.2307653449471584544818427896498956L, - 0.5000000000000000000000000000000000L, 0.7692346550528415455181572103501044L, - 0.9530899229693319963988134391496965L }; - const RealScalar weights[] = { 0.1184634425280945437571320203599587L, 0.2393143352496832340206457574178191L, - 0.2844444444444444444444444444444444L, 0.2393143352496832340206457574178191L, - 0.1184634425280945437571320203599587L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); + using std::log; + MatrixType result(A.rows(), A.rows()); + if (A.rows() == 1) + result(0,0) = log(A(0,0)); + else if (A.rows() == 2) + matrix_log_compute_2x2(A, result); + else + matrix_log_compute_big(A, result); + return result; } -template -void MatrixLogarithmAtomic::computePade6(MatrixType& result, const MatrixType& T) -{ - const int degree = 6; - const RealScalar nodes[] = { 0.0337652428984239860938492227530027L, 0.1693953067668677431693002024900473L, - 0.3806904069584015456847491391596440L, 0.6193095930415984543152508608403560L, - 0.8306046932331322568306997975099527L, 0.9662347571015760139061507772469973L }; - const RealScalar weights[] = { 0.0856622461895851725201480710863665L, 0.1803807865240693037849167569188581L, - 0.2339569672863455236949351719947755L, 0.2339569672863455236949351719947755L, - 0.1803807865240693037849167569188581L, 0.0856622461895851725201480710863665L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); -} - -template -void MatrixLogarithmAtomic::computePade7(MatrixType& result, const MatrixType& T) -{ - const int degree = 7; - const RealScalar nodes[] = { 0.0254460438286207377369051579760744L, 0.1292344072003027800680676133596058L, - 0.2970774243113014165466967939615193L, 0.5000000000000000000000000000000000L, - 0.7029225756886985834533032060384807L, 0.8707655927996972199319323866403942L, - 0.9745539561713792622630948420239256L }; - const RealScalar weights[] = { 0.0647424830844348466353057163395410L, 0.1398526957446383339507338857118898L, - 0.1909150252525594724751848877444876L, 0.2089795918367346938775510204081633L, - 0.1909150252525594724751848877444876L, 0.1398526957446383339507338857118898L, - 0.0647424830844348466353057163395410L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); -} - -template -void MatrixLogarithmAtomic::computePade8(MatrixType& result, const MatrixType& T) -{ - const int degree = 8; - const RealScalar nodes[] = { 0.0198550717512318841582195657152635L, 0.1016667612931866302042230317620848L, - 0.2372337950418355070911304754053768L, 0.4082826787521750975302619288199080L, - 0.5917173212478249024697380711800920L, 0.7627662049581644929088695245946232L, - 0.8983332387068133697957769682379152L, 0.9801449282487681158417804342847365L }; - const RealScalar weights[] = { 0.0506142681451881295762656771549811L, 0.1111905172266872352721779972131204L, - 0.1568533229389436436689811009933007L, 0.1813418916891809914825752246385978L, - 0.1813418916891809914825752246385978L, 0.1568533229389436436689811009933007L, - 0.1111905172266872352721779972131204L, 0.0506142681451881295762656771549811L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); -} - -template -void MatrixLogarithmAtomic::computePade9(MatrixType& result, const MatrixType& T) -{ - const int degree = 9; - const RealScalar nodes[] = { 0.0159198802461869550822118985481636L, 0.0819844463366821028502851059651326L, - 0.1933142836497048013456489803292629L, 0.3378732882980955354807309926783317L, - 0.5000000000000000000000000000000000L, 0.6621267117019044645192690073216683L, - 0.8066857163502951986543510196707371L, 0.9180155536633178971497148940348674L, - 0.9840801197538130449177881014518364L }; - const RealScalar weights[] = { 0.0406371941807872059859460790552618L, 0.0903240803474287020292360156214564L, - 0.1303053482014677311593714347093164L, 0.1561735385200014200343152032922218L, - 0.1651196775006298815822625346434870L, 0.1561735385200014200343152032922218L, - 0.1303053482014677311593714347093164L, 0.0903240803474287020292360156214564L, - 0.0406371941807872059859460790552618L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); -} - -template -void MatrixLogarithmAtomic::computePade10(MatrixType& result, const MatrixType& T) -{ - const int degree = 10; - const RealScalar nodes[] = { 0.0130467357414141399610179939577740L, 0.0674683166555077446339516557882535L, - 0.1602952158504877968828363174425632L, 0.2833023029353764046003670284171079L, - 0.4255628305091843945575869994351400L, 0.5744371694908156054424130005648600L, - 0.7166976970646235953996329715828921L, 0.8397047841495122031171636825574368L, - 0.9325316833444922553660483442117465L, 0.9869532642585858600389820060422260L }; - const RealScalar weights[] = { 0.0333356721543440687967844049466659L, 0.0747256745752902965728881698288487L, - 0.1095431812579910219977674671140816L, 0.1346333596549981775456134607847347L, - 0.1477621123573764350869464973256692L, 0.1477621123573764350869464973256692L, - 0.1346333596549981775456134607847347L, 0.1095431812579910219977674671140816L, - 0.0747256745752902965728881698288487L, 0.0333356721543440687967844049466659L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); -} - -template -void MatrixLogarithmAtomic::computePade11(MatrixType& result, const MatrixType& T) -{ - const int degree = 11; - const RealScalar nodes[] = { 0.0108856709269715035980309994385713L, 0.0564687001159523504624211153480364L, - 0.1349239972129753379532918739844233L, 0.2404519353965940920371371652706952L, - 0.3652284220238275138342340072995692L, 0.5000000000000000000000000000000000L, - 0.6347715779761724861657659927004308L, 0.7595480646034059079628628347293048L, - 0.8650760027870246620467081260155767L, 0.9435312998840476495375788846519636L, - 0.9891143290730284964019690005614287L }; - const RealScalar weights[] = { 0.0278342835580868332413768602212743L, 0.0627901847324523123173471496119701L, - 0.0931451054638671257130488207158280L, 0.1165968822959952399592618524215876L, - 0.1314022722551233310903444349452546L, 0.1364625433889503153572417641681711L, - 0.1314022722551233310903444349452546L, 0.1165968822959952399592618524215876L, - 0.0931451054638671257130488207158280L, 0.0627901847324523123173471496119701L, - 0.0278342835580868332413768602212743L }; - eigen_assert(degree <= maxPadeDegree); - MatrixType TminusI = T - MatrixType::Identity(T.rows(), T.rows()); - result.setZero(T.rows(), T.rows()); - for (int k = 0; k < degree; ++k) - result += weights[k] * (MatrixType::Identity(T.rows(), T.rows()) + nodes[k] * TminusI) - .template triangularView().solve(TminusI); -} +} // end of namespace internal /** \ingroup MatrixFunctions_Module * @@ -441,12 +330,11 @@ public: static const int Options = PlainObject::Options; typedef std::complex::Real> ComplexScalar; typedef Matrix DynMatrixType; - typedef MatrixLogarithmAtomic AtomicType; + typedef internal::MatrixLogarithmAtomic AtomicType; AtomicType atomic; const PlainObject Aevaluated = m_A.eval(); - MatrixFunction mf(Aevaluated, atomic); - mf.compute(result); + internal::matrix_function_compute::run(Aevaluated, atomic, result); } Index rows() const { return m_A.rows(); } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h index 0261d4aa9..314b3f38e 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Jitse Niesen +// Copyright (C) 2011, 2013 Jitse Niesen // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed From 75edc7cc8bcc1bdf5a3defddc2f751062ec49ec1 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Fri, 26 Jul 2013 11:05:21 +0200 Subject: [PATCH 0052/1103] Fixed VC11 compilation. The typedefs Lhs/Rhs in the base class are now accessible by derived classes. --- .../src/KroneckerProduct/KroneckerTensorProduct.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index 80be9fd7a..0dccdb2bc 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -26,11 +26,11 @@ class KroneckerProductBase : public ReturnByValue { private: typedef typename internal::traits Traits; - typedef typename Traits::Lhs Lhs; - typedef typename Traits::Rhs Rhs; typedef typename Traits::Scalar Scalar; protected: + typedef typename Traits::Lhs Lhs; + typedef typename Traits::Rhs Rhs; typedef typename Traits::Index Index; public: @@ -79,8 +79,8 @@ class KroneckerProductBase : public ReturnByValue * \tparam Lhs Type of the left-hand side, a matrix expression. * \tparam Rhs Type of the rignt-hand side, a matrix expression. */ -template -class KroneckerProduct : public KroneckerProductBase > +template +class KroneckerProduct : public KroneckerProductBase > { private: typedef KroneckerProductBase Base; @@ -112,8 +112,8 @@ class KroneckerProduct : public KroneckerProductBase > * \tparam Lhs Type of the left-hand side, a matrix expression. * \tparam Rhs Type of the rignt-hand side, a matrix expression. */ -template -class KroneckerProductSparse : public KroneckerProductBase > +template +class KroneckerProductSparse : public KroneckerProductBase > { private: typedef KroneckerProductBase Base; From e43934d60f20d3f3884493d542bdcb0cb330c39a Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Fri, 26 Jul 2013 13:51:10 +0100 Subject: [PATCH 0053/1103] MatrixFunctions: Clean up StemFunction.h --- .../src/MatrixFunctions/MatrixFunction.h | 8 +- .../Eigen/src/MatrixFunctions/StemFunction.h | 170 ++++++++++-------- unsupported/test/matrix_function.cpp | 2 +- 3 files changed, 96 insertions(+), 84 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index 12e28793d..8a6b84c09 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -547,7 +547,7 @@ const MatrixFunctionReturnValue MatrixBase::sin() const { eigen_assert(rows() == cols()); typedef typename internal::stem_function::ComplexScalar ComplexScalar; - return MatrixFunctionReturnValue(derived(), StdStemFunctions::sin); + return MatrixFunctionReturnValue(derived(), internal::stem_function_sin); } template @@ -555,7 +555,7 @@ const MatrixFunctionReturnValue MatrixBase::cos() const { eigen_assert(rows() == cols()); typedef typename internal::stem_function::ComplexScalar ComplexScalar; - return MatrixFunctionReturnValue(derived(), StdStemFunctions::cos); + return MatrixFunctionReturnValue(derived(), internal::stem_function_cos); } template @@ -563,7 +563,7 @@ const MatrixFunctionReturnValue MatrixBase::sinh() const { eigen_assert(rows() == cols()); typedef typename internal::stem_function::ComplexScalar ComplexScalar; - return MatrixFunctionReturnValue(derived(), StdStemFunctions::sinh); + return MatrixFunctionReturnValue(derived(), internal::stem_function_sinh); } template @@ -571,7 +571,7 @@ const MatrixFunctionReturnValue MatrixBase::cosh() const { eigen_assert(rows() == cols()); typedef typename internal::stem_function::ComplexScalar ComplexScalar; - return MatrixFunctionReturnValue(derived(), StdStemFunctions::cosh); + return MatrixFunctionReturnValue(derived(), internal::stem_function_cosh); } } // end namespace Eigen diff --git a/unsupported/Eigen/src/MatrixFunctions/StemFunction.h b/unsupported/Eigen/src/MatrixFunctions/StemFunction.h index 0a815d834..cd45fe5b4 100644 --- a/unsupported/Eigen/src/MatrixFunctions/StemFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/StemFunction.h @@ -12,93 +12,105 @@ namespace Eigen { -/** \ingroup MatrixFunctions_Module - * \brief Stem functions corresponding to standard mathematical functions. - */ +namespace internal { + +/** \brief The exponential function (and its derivatives). */ template -class StdStemFunctions : internal::noncopyable +Scalar stem_function_exp(Scalar x, int) { - public: + using std::exp; + return exp(x); +} - /** \brief The exponential function (and its derivatives). */ - static Scalar exp(Scalar x, int) - { - return std::exp(x); - } +/** \brief Cosine (and its derivatives). */ +template +Scalar stem_function_cos(Scalar x, int n) +{ + using std::cos; + using std::sin; + Scalar res; - /** \brief Cosine (and its derivatives). */ - static Scalar cos(Scalar x, int n) - { - Scalar res; - switch (n % 4) { - case 0: - res = std::cos(x); - break; - case 1: - res = -std::sin(x); - break; - case 2: - res = -std::cos(x); - break; - case 3: - res = std::sin(x); - break; - } - return res; - } + switch (n % 4) { + case 0: + res = std::cos(x); + break; + case 1: + res = -std::sin(x); + break; + case 2: + res = -std::cos(x); + break; + case 3: + res = std::sin(x); + break; + } + return res; +} - /** \brief Sine (and its derivatives). */ - static Scalar sin(Scalar x, int n) - { - Scalar res; - switch (n % 4) { - case 0: - res = std::sin(x); - break; - case 1: - res = std::cos(x); - break; - case 2: - res = -std::sin(x); - break; - case 3: - res = -std::cos(x); - break; - } - return res; - } +/** \brief Sine (and its derivatives). */ +template +Scalar stem_function_sin(Scalar x, int n) +{ + using std::cos; + using std::sin; + Scalar res; - /** \brief Hyperbolic cosine (and its derivatives). */ - static Scalar cosh(Scalar x, int n) - { - Scalar res; - switch (n % 2) { - case 0: - res = std::cosh(x); - break; - case 1: - res = std::sinh(x); - break; - } - return res; - } + switch (n % 4) { + case 0: + res = std::sin(x); + break; + case 1: + res = std::cos(x); + break; + case 2: + res = -std::sin(x); + break; + case 3: + res = -std::cos(x); + break; + } + return res; +} + +/** \brief Hyperbolic cosine (and its derivatives). */ +template +Scalar stem_function_cosh(Scalar x, int n) +{ + using std::cosh; + using std::sinh; + Scalar res; + + switch (n % 2) { + case 0: + res = std::cosh(x); + break; + case 1: + res = std::sinh(x); + break; + } + return res; +} - /** \brief Hyperbolic sine (and its derivatives). */ - static Scalar sinh(Scalar x, int n) - { - Scalar res; - switch (n % 2) { - case 0: - res = std::sinh(x); - break; - case 1: - res = std::cosh(x); - break; - } - return res; - } +/** \brief Hyperbolic sine (and its derivatives). */ +template +Scalar stem_function_sinh(Scalar x, int n) +{ + using std::cosh; + using std::sinh; + Scalar res; + + switch (n % 2) { + case 0: + res = std::sinh(x); + break; + case 1: + res = std::cosh(x); + break; + } + return res; +} -}; // end of class StdStemFunctions +} // end namespace internal } // end namespace Eigen diff --git a/unsupported/test/matrix_function.cpp b/unsupported/test/matrix_function.cpp index 3c76cfb65..487d5a9b8 100644 --- a/unsupported/test/matrix_function.cpp +++ b/unsupported/test/matrix_function.cpp @@ -102,7 +102,7 @@ void testMatrixExponential(const MatrixType& A) typedef typename NumTraits::Real RealScalar; typedef std::complex ComplexScalar; - VERIFY_IS_APPROX(A.exp(), A.matrixFunction(StdStemFunctions::exp)); + VERIFY_IS_APPROX(A.exp(), A.matrixFunction(internal::stem_function_exp)); } template From 75dab1ce5e31d942edce1fabb28cc07685736dbd Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Fri, 26 Jul 2013 15:13:54 +0200 Subject: [PATCH 0054/1103] Fixed floating point warning. Fixed evaluation of matrix_exp_computeUV. --- unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h | 2 +- unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 54f07b27c..d67ded7ba 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -258,7 +258,7 @@ struct matrix_exp_computeUV { #if LDBL_MANT_DIG == 53 // double precision - matrix_exp_computeUV(arg, U, V, squarings); + matrix_exp_computeUV::run(arg, U, V, squarings); #else diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index 12e28793d..a948f2347 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -18,7 +18,7 @@ namespace Eigen { namespace internal { /** \brief Maximum distance allowed between eigenvalues to be considered "close". */ -static const float matrix_function_separation = 0.1; +static const float matrix_function_separation = 0.1f; /** \ingroup MatrixFunctions_Module * \class MatrixFunctionAtomic From 70131120abc723eb6e09c9066bee81a0b42335ae Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Fri, 26 Jul 2013 15:39:18 +0100 Subject: [PATCH 0055/1103] Fix bug in MatrixFunctions for matrices with multiple eigenvalues. Store indices, not eigenvalues, in clusters. Bug was introduced in changeset a3a55357db7394281c872e911f13d69aba510aec . --- Eigen/src/Core/util/XprHelper.h | 4 ++-- .../src/MatrixFunctions/MatrixFunction.h | 24 +++++++++---------- .../Eigen/src/MatrixFunctions/StemFunction.h | 2 +- 3 files changed, 15 insertions(+), 15 deletions(-) diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 3c4773054..0009ec049 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -307,9 +307,9 @@ struct transfer_constness * * Example. Suppose that a, b, and c are of type Matrix3d. The user forms the expression a*(b+c). * b+c is an expression "sum of matrices", which we will denote by S. In order to determine how to nest it, - * the Product expression uses: nested::ret, which turns out to be Matrix3d because the internal logic of + * the Product expression uses: nested::type, which turns out to be Matrix3d because the internal logic of * nested determined that in this case it was better to evaluate the expression b+c into a temporary. On the other hand, - * since a is of type Matrix3d, the Product expression nests it as nested::ret, which turns out to be + * since a is of type Matrix3d, the Product expression nests it as nested::type, which turns out to be * const Matrix3d&, because the internal logic of nested determined that since a was already a matrix, there was no point * in copying it into another matrix. */ diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index cf812542c..a5675228e 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -105,10 +105,10 @@ MatrixType MatrixFunctionAtomic::compute(const MatrixType& A) * \returns Iterator to cluster containing \p key, or \c clusters.end() if no cluster in \p m_clusters * contains \p key. */ -template -typename ListOfClusters::iterator matrix_function_find_cluster(Scalar key, ListOfClusters& clusters) +template +typename ListOfClusters::iterator matrix_function_find_cluster(Index key, ListOfClusters& clusters) { - typename std::list::iterator j; + typename std::list::iterator j; for (typename ListOfClusters::iterator i = clusters.begin(); i != clusters.end(); ++i) { j = std::find(i->begin(), i->end(), key); if (j != i->end()) @@ -133,11 +133,11 @@ void matrix_function_partition_eigenvalues(const EivalsType& eivals, std::list::iterator qi = matrix_function_find_cluster(eivals(i), clusters); + // Find cluster containing i-th ei'val, adding a new cluster if necessary + typename std::list::iterator qi = matrix_function_find_cluster(i, clusters); if (qi == clusters.end()) { Cluster l; - l.push_back(eivals(i)); + l.push_back(i); clusters.push_back(l); qi = clusters.end(); --qi; @@ -146,10 +146,10 @@ void matrix_function_partition_eigenvalues(const EivalsType& eivals, std::listbegin(), qi->end(), eivals(j)) == qi->end()) { - typename std::list::iterator qj = matrix_function_find_cluster(eivals(j), clusters); + && std::find(qi->begin(), qi->end(), j) == qi->end()) { + typename std::list::iterator qj = matrix_function_find_cluster(j, clusters); if (qj == clusters.end()) { - qi->push_back(eivals(j)); + qi->push_back(j); } else { qi->insert(qi->end(), qj->begin(), qj->end()); clusters.erase(qj); @@ -192,7 +192,7 @@ void matrix_function_compute_map(const EivalsType& eivals, const ListOfClusters& Index clusterIndex = 0; for (typename ListOfClusters::const_iterator cluster = clusters.begin(); cluster != clusters.end(); ++cluster) { for (Index i = 0; i < eivals.rows(); ++i) { - if (std::find(cluster->begin(), cluster->end(), eivals(i)) != cluster->end()) { + if (std::find(cluster->begin(), cluster->end(), i) != cluster->end()) { eivalToCluster[i] = clusterIndex; } } @@ -435,7 +435,7 @@ struct matrix_function_compute MatrixType U = schurOfA.matrixU(); // partition eigenvalues into clusters of ei'vals "close" to each other - std::list > clusters; + std::list > clusters; matrix_function_partition_eigenvalues(T.diagonal(), clusters); // compute size of each cluster @@ -446,7 +446,7 @@ struct matrix_function_compute Matrix blockStart; matrix_function_compute_block_start(clusterSize, blockStart); - // compute map so that eivalToCluster[i] = j means that ei'val T(i,i) is in j-th cluster + // compute map so that eivalToCluster[i] = j means that i-th ei'val is in j-th cluster Matrix eivalToCluster; matrix_function_compute_map(T.diagonal(), clusters, eivalToCluster); diff --git a/unsupported/Eigen/src/MatrixFunctions/StemFunction.h b/unsupported/Eigen/src/MatrixFunctions/StemFunction.h index cd45fe5b4..7604df903 100644 --- a/unsupported/Eigen/src/MatrixFunctions/StemFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/StemFunction.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2010 Jitse Niesen +// Copyright (C) 2010, 2013 Jitse Niesen // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed From dd27b5c4a85f6a7b67424ab0ad249e2c1fab88a4 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Sun, 28 Jul 2013 19:31:33 +0200 Subject: [PATCH 0056/1103] Fixed dummy_precision evaluation. --- Eigen/src/Eigen2Support/Geometry/AngleAxis.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Eigen2Support/Geometry/AngleAxis.h b/Eigen/src/Eigen2Support/Geometry/AngleAxis.h index 9f48c460b..a0b4ac44e 100644 --- a/Eigen/src/Eigen2Support/Geometry/AngleAxis.h +++ b/Eigen/src/Eigen2Support/Geometry/AngleAxis.h @@ -76,7 +76,7 @@ public: using std::sqrt; using std::abs; // since we compare against 1, this is equal to computing the relative error - eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( dummy_precision() ) ); + eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( NumTraits::dummy_precision() ) ); } /** Constructs and initialize the angle-axis rotation from a quaternion \a q. */ @@ -180,7 +180,7 @@ AngleAxis& AngleAxis::operator=(const QuaternionType& q) using std::sqrt; using std::abs; // since we compare against 1, this is equal to computing the relative error - eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( dummy_precision() ) ); + eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( NumTraits::dummy_precision() ) ); } return *this; } From 2437215221db4e67a158ab316dc2ab231ab2361d Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Sun, 28 Jul 2013 22:46:38 +0200 Subject: [PATCH 0057/1103] Fixed constness in Array- and MatrixWrapper. This also fixes the compilation on VC11. --- Eigen/src/Core/CoreEvaluators.h | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 272027c7b..3240ec6ed 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -1001,25 +1001,25 @@ protected: typename evaluator::nestedType m_argImpl; }; -template -struct evaluator_impl > - : evaluator_impl_wrapper_base > +template +struct evaluator_impl > + : evaluator_impl_wrapper_base > { - typedef MatrixWrapper XprType; + typedef MatrixWrapper XprType; evaluator_impl(const XprType& wrapper) - : evaluator_impl_wrapper_base >(wrapper.nestedExpression()) + : evaluator_impl_wrapper_base >(wrapper.nestedExpression()) { } }; -template -struct evaluator_impl > - : evaluator_impl_wrapper_base > +template +struct evaluator_impl > + : evaluator_impl_wrapper_base > { - typedef ArrayWrapper XprType; + typedef ArrayWrapper XprType; evaluator_impl(const XprType& wrapper) - : evaluator_impl_wrapper_base >(wrapper.nestedExpression()) + : evaluator_impl_wrapper_base >(wrapper.nestedExpression()) { } }; From 5f11db695b74db219e0d06682728c9eda90cac5c Mon Sep 17 00:00:00 2001 From: Sven Strothoff Date: Sun, 28 Jul 2013 23:59:37 +0200 Subject: [PATCH 0058/1103] bug #502: add bool intersects() methods to AlignedBox --- Eigen/src/Geometry/AlignedBox.h | 4 ++++ test/geo_alignedbox.cpp | 7 +++++++ 2 files changed, 11 insertions(+) diff --git a/Eigen/src/Geometry/AlignedBox.h b/Eigen/src/Geometry/AlignedBox.h index 8e186d57a..b6a2f0e24 100644 --- a/Eigen/src/Geometry/AlignedBox.h +++ b/Eigen/src/Geometry/AlignedBox.h @@ -185,6 +185,10 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim) inline bool contains(const AlignedBox& b) const { return (m_min.array()<=(b.min)().array()).all() && ((b.max)().array()<=m_max.array()).all(); } + /** \returns true if the box \a b is intersecting the box \c *this. */ + inline bool intersects(const AlignedBox& b) const + { return (m_min.array()<=(b.max)().array()).all() && ((b.min)().array()<=m_max.array()).all(); } + /** Extends \c *this such that it contains the point \a p and returns a reference to \c *this. */ template inline AlignedBox& extend(const MatrixBase& a_p) diff --git a/test/geo_alignedbox.cpp b/test/geo_alignedbox.cpp index 8e36adbe3..e2792ed18 100644 --- a/test/geo_alignedbox.cpp +++ b/test/geo_alignedbox.cpp @@ -54,6 +54,13 @@ template void alignedbox(const BoxType& _box) VERIFY(b2.contains(b1)); VERIFY_IS_APPROX(b2.clamp(b0), b0); + // intersection + BoxType box1(VectorType::Random(dim)); + box1.extend(VectorType::Random(dim)); + BoxType box2(VectorType::Random(dim)); + box2.extend(VectorType::Random(dim)); + + VERIFY(box1.intersects(box2) == !box1.intersection(box2).isEmpty()); // alignment -- make sure there is no memory alignment assertion BoxType *bp0 = new BoxType(dim); From 9ef3645cc748151565131fb946c400471fea8969 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Mon, 29 Jul 2013 12:08:50 +0200 Subject: [PATCH 0059/1103] Removed 'T' prefix from types and thus fixed compilation for GCC. --- .../Eigen/src/KroneckerProduct/KroneckerTensorProduct.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index 0dccdb2bc..a4516056d 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -79,8 +79,8 @@ class KroneckerProductBase : public ReturnByValue * \tparam Lhs Type of the left-hand side, a matrix expression. * \tparam Rhs Type of the rignt-hand side, a matrix expression. */ -template -class KroneckerProduct : public KroneckerProductBase > +template +class KroneckerProduct : public KroneckerProductBase > { private: typedef KroneckerProductBase Base; @@ -112,8 +112,8 @@ class KroneckerProduct : public KroneckerProductBase * \tparam Lhs Type of the left-hand side, a matrix expression. * \tparam Rhs Type of the rignt-hand side, a matrix expression. */ -template -class KroneckerProductSparse : public KroneckerProductBase > +template +class KroneckerProductSparse : public KroneckerProductBase > { private: typedef KroneckerProductBase Base; From acb82c7f1638389ac53b68c7db3a4fd8f6a92098 Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Mon, 29 Jul 2013 20:13:52 +0900 Subject: [PATCH 0060/1103] Quick fix in order to be custom-scalar friendly. --- Eigen/src/SparseCore/SparseMatrix.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index adceafe18..970c6be71 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -384,7 +384,7 @@ class SparseMatrix eigen_assert( (m_outerIndex[outer+1]-m_outerIndex[outer]==0 || m_data.index(m_data.size()-1) Date: Tue, 30 Jul 2013 08:05:10 +0200 Subject: [PATCH 0061/1103] Removed non-standard conforming (17.4.3.1.2/1) leading underscore. --- Eigen/src/SparseLU/SparseLU.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index dd9eab2c2..3abe998f2 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -219,9 +219,9 @@ class SparseLU : public internal::SparseLUImpl - bool _solve(const MatrixBase &B, MatrixBase &_X) const + bool _solve(const MatrixBase &B, MatrixBase &X_base) const { - Dest& X(_X.derived()); + Dest& X(X_base.derived()); eigen_assert(m_factorizationIsOk && "The matrix should be factorized first"); EIGEN_STATIC_ASSERT((Dest::Flags&RowMajorBit)==0, THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES); From 6126ad801f89cb88d46ed0dae0d8e9448652c506 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 31 Jul 2013 15:30:50 +0200 Subject: [PATCH 0062/1103] Extend support for nvcc to Array objects and wrappers --- Eigen/src/Core/ArrayBase.h | 12 +++++++++ Eigen/src/Core/ArrayWrapper.h | 33 +++++++++++++++++++++++++ Eigen/src/Core/CwiseBinaryOp.h | 1 + Eigen/src/Core/Diagonal.h | 18 ++++++++++++++ Eigen/src/Core/DiagonalMatrix.h | 33 +++++++++++++++++++++++++ Eigen/src/Core/MatrixBase.h | 20 ++++++++++++--- Eigen/src/plugins/ArrayCwiseBinaryOps.h | 11 +++++++++ Eigen/src/plugins/ArrayCwiseUnaryOps.h | 1 + 8 files changed, 126 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index 38852600d..9c6daf4fe 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -118,40 +118,52 @@ template class ArrayBase /** Special case of the template operator=, in order to prevent the compiler * from generating a default operator= (issue hit with g++ 4.1) */ + EIGEN_DEVICE_FUNC Derived& operator=(const ArrayBase& other) { return internal::assign_selector::run(derived(), other.derived()); } + EIGEN_DEVICE_FUNC Derived& operator+=(const Scalar& scalar) { return *this = derived() + scalar; } + EIGEN_DEVICE_FUNC Derived& operator-=(const Scalar& scalar) { return *this = derived() - scalar; } template + EIGEN_DEVICE_FUNC Derived& operator+=(const ArrayBase& other); template + EIGEN_DEVICE_FUNC Derived& operator-=(const ArrayBase& other); template + EIGEN_DEVICE_FUNC Derived& operator*=(const ArrayBase& other); template + EIGEN_DEVICE_FUNC Derived& operator/=(const ArrayBase& other); public: + EIGEN_DEVICE_FUNC ArrayBase& array() { return *this; } + EIGEN_DEVICE_FUNC const ArrayBase& array() const { return *this; } /** \returns an \link Eigen::MatrixBase Matrix \endlink expression of this array * \sa MatrixBase::array() */ + EIGEN_DEVICE_FUNC MatrixWrapper matrix() { return derived(); } + EIGEN_DEVICE_FUNC const MatrixWrapper matrix() const { return derived(); } // template // inline void evalTo(Dest& dst) const { dst = matrix(); } protected: + EIGEN_DEVICE_FUNC ArrayBase() : Base() {} private: diff --git a/Eigen/src/Core/ArrayWrapper.h b/Eigen/src/Core/ArrayWrapper.h index a791bc358..21830745b 100644 --- a/Eigen/src/Core/ArrayWrapper.h +++ b/Eigen/src/Core/ArrayWrapper.h @@ -48,41 +48,54 @@ class ArrayWrapper : public ArrayBase > typedef typename internal::nested::type NestedExpressionType; + EIGEN_DEVICE_FUNC inline ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {} + EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); } + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return m_expression.const_cast_derived().data(); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_expression.data(); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index rowId, Index colId) const { return m_expression.coeff(rowId, colId); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index rowId, Index colId) { return m_expression.const_cast_derived().coeffRef(rowId, colId); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index rowId, Index colId) const { return m_expression.const_cast_derived().coeffRef(rowId, colId); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index index) const { return m_expression.coeff(index); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { return m_expression.const_cast_derived().coeffRef(index); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index index) const { return m_expression.const_cast_derived().coeffRef(index); @@ -113,9 +126,11 @@ class ArrayWrapper : public ArrayBase > } template + EIGEN_DEVICE_FUNC inline void evalTo(Dest& dst) const { dst = m_expression; } const typename internal::remove_all::type& + EIGEN_DEVICE_FUNC nestedExpression() const { return m_expression; @@ -123,9 +138,11 @@ class ArrayWrapper : public ArrayBase > /** Forwards the resizing request to the nested expression * \sa DenseBase::resize(Index) */ + EIGEN_DEVICE_FUNC void resize(Index newSize) { m_expression.const_cast_derived().resize(newSize); } /** Forwards the resizing request to the nested expression * \sa DenseBase::resize(Index,Index)*/ + EIGEN_DEVICE_FUNC void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); } protected: @@ -168,41 +185,54 @@ class MatrixWrapper : public MatrixBase > typedef typename internal::nested::type NestedExpressionType; + EIGEN_DEVICE_FUNC inline MatrixWrapper(ExpressionType& a_matrix) : m_expression(a_matrix) {} + EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); } + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return m_expression.const_cast_derived().data(); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_expression.data(); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index rowId, Index colId) const { return m_expression.coeff(rowId, colId); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index rowId, Index colId) { return m_expression.const_cast_derived().coeffRef(rowId, colId); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index rowId, Index colId) const { return m_expression.derived().coeffRef(rowId, colId); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index index) const { return m_expression.coeff(index); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { return m_expression.const_cast_derived().coeffRef(index); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index index) const { return m_expression.const_cast_derived().coeffRef(index); @@ -232,6 +262,7 @@ class MatrixWrapper : public MatrixBase > m_expression.const_cast_derived().template writePacket(index, val); } + EIGEN_DEVICE_FUNC const typename internal::remove_all::type& nestedExpression() const { @@ -240,9 +271,11 @@ class MatrixWrapper : public MatrixBase > /** Forwards the resizing request to the nested expression * \sa DenseBase::resize(Index) */ + EIGEN_DEVICE_FUNC void resize(Index newSize) { m_expression.const_cast_derived().resize(newSize); } /** Forwards the resizing request to the nested expression * \sa DenseBase::resize(Index,Index)*/ + EIGEN_DEVICE_FUNC void resize(Index nbRows, Index nbCols) { m_expression.const_cast_derived().resize(nbRows,nbCols); } protected: diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index e603fb9d6..e20daacc8 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -235,3 +235,4 @@ MatrixBase::operator+=(const MatrixBase& other) } // end namespace Eigen #endif // EIGEN_CWISE_BINARY_OP_H + diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index aab8007b3..4436c6a69 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -70,20 +70,25 @@ template class Diagonal typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal) + EIGEN_DEVICE_FUNC inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal) + EIGEN_DEVICE_FUNC inline Index rows() const { return m_index.value()<0 ? (std::min)(m_matrix.cols(),m_matrix.rows()+m_index.value()) : (std::min)(m_matrix.rows(),m_matrix.cols()-m_index.value()); } + EIGEN_DEVICE_FUNC inline Index cols() const { return 1; } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.outerStride() + 1; } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return 0; @@ -95,47 +100,57 @@ template class Diagonal const Scalar >::type ScalarWithConstIfNotLvalue; + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return &(m_matrix.const_cast_derived().coeffRef(rowOffset(), colOffset())); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index) { EIGEN_STATIC_ASSERT_LVALUE(MatrixType) return m_matrix.const_cast_derived().coeffRef(row+rowOffset(), row+colOffset()); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index row, Index) const { return m_matrix.const_cast_derived().coeffRef(row+rowOffset(), row+colOffset()); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index row, Index) const { return m_matrix.coeff(row+rowOffset(), row+colOffset()); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index idx) { EIGEN_STATIC_ASSERT_LVALUE(MatrixType) return m_matrix.const_cast_derived().coeffRef(idx+rowOffset(), idx+colOffset()); } + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index idx) const { return m_matrix.const_cast_derived().coeffRef(idx+rowOffset(), idx+colOffset()); } + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index idx) const { return m_matrix.coeff(idx+rowOffset(), idx+colOffset()); } + EIGEN_DEVICE_FUNC const typename internal::remove_all::type& nestedExpression() const { return m_matrix; } + EIGEN_DEVICE_FUNC int index() const { return m_index.value(); @@ -147,8 +162,11 @@ template class Diagonal private: // some compilers may fail to optimize std::max etc in case of compile-time constants... + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index absDiagIndex() const { return m_index.value()>0 ? m_index.value() : -m_index.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; } // triger a compile time error is someone try to call packet template typename MatrixType::PacketReturnType packet(Index) const; diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index e6c220f41..f7ac22f8b 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -37,45 +37,59 @@ class DiagonalBase : public EigenBase typedef DenseMatrixType DenseType; typedef DiagonalMatrix PlainObject; + EIGEN_DEVICE_FUNC inline const Derived& derived() const { return *static_cast(this); } + EIGEN_DEVICE_FUNC inline Derived& derived() { return *static_cast(this); } + EIGEN_DEVICE_FUNC DenseMatrixType toDenseMatrix() const { return derived(); } template + EIGEN_DEVICE_FUNC void evalTo(MatrixBase &other) const; template + EIGEN_DEVICE_FUNC void addTo(MatrixBase &other) const { other.diagonal() += diagonal(); } template + EIGEN_DEVICE_FUNC void subTo(MatrixBase &other) const { other.diagonal() -= diagonal(); } + EIGEN_DEVICE_FUNC inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); } + EIGEN_DEVICE_FUNC inline DiagonalVectorType& diagonal() { return derived().diagonal(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return diagonal().size(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return diagonal().size(); } /** \returns the diagonal matrix product of \c *this by the matrix \a matrix. */ template + EIGEN_DEVICE_FUNC const DiagonalProduct operator*(const MatrixBase &matrix) const { return DiagonalProduct(matrix.derived(), derived()); } + EIGEN_DEVICE_FUNC inline const DiagonalWrapper, const DiagonalVectorType> > inverse() const { return diagonal().cwiseInverse(); } + EIGEN_DEVICE_FUNC inline const DiagonalWrapper, const DiagonalVectorType> > operator*(const Scalar& scalar) const { return diagonal() * scalar; } + EIGEN_DEVICE_FUNC friend inline const DiagonalWrapper, const DiagonalVectorType> > operator*(const Scalar& scalar, const DiagonalBase& other) { @@ -84,11 +98,13 @@ class DiagonalBase : public EigenBase #ifdef EIGEN2_SUPPORT template + EIGEN_DEVICE_FUNC bool isApprox(const DiagonalBase& other, typename NumTraits::Real precision = NumTraits::dummy_precision()) const { return diagonal().isApprox(other.diagonal(), precision); } template + EIGEN_DEVICE_FUNC bool isApprox(const MatrixBase& other, typename NumTraits::Real precision = NumTraits::dummy_precision()) const { return toDenseMatrix().isApprox(other, precision); @@ -151,24 +167,31 @@ class DiagonalMatrix public: /** const version of diagonal(). */ + EIGEN_DEVICE_FUNC inline const DiagonalVectorType& diagonal() const { return m_diagonal; } /** \returns a reference to the stored vector of diagonal coefficients. */ + EIGEN_DEVICE_FUNC inline DiagonalVectorType& diagonal() { return m_diagonal; } /** Default constructor without initialization */ + EIGEN_DEVICE_FUNC inline DiagonalMatrix() {} /** Constructs a diagonal matrix with given dimension */ + EIGEN_DEVICE_FUNC inline DiagonalMatrix(Index dim) : m_diagonal(dim) {} /** 2D constructor. */ + EIGEN_DEVICE_FUNC inline DiagonalMatrix(const Scalar& x, const Scalar& y) : m_diagonal(x,y) {} /** 3D constructor. */ + EIGEN_DEVICE_FUNC inline DiagonalMatrix(const Scalar& x, const Scalar& y, const Scalar& z) : m_diagonal(x,y,z) {} /** Copy constructor. */ template + EIGEN_DEVICE_FUNC inline DiagonalMatrix(const DiagonalBase& other) : m_diagonal(other.diagonal()) {} #ifndef EIGEN_PARSED_BY_DOXYGEN @@ -178,11 +201,13 @@ class DiagonalMatrix /** generic constructor from expression of the diagonal coefficients */ template + EIGEN_DEVICE_FUNC explicit inline DiagonalMatrix(const MatrixBase& other) : m_diagonal(other) {} /** Copy operator. */ template + EIGEN_DEVICE_FUNC DiagonalMatrix& operator=(const DiagonalBase& other) { m_diagonal = other.diagonal(); @@ -193,6 +218,7 @@ class DiagonalMatrix /** This is a special case of the templated operator=. Its purpose is to * prevent a default operator= from hiding the templated operator=. */ + EIGEN_DEVICE_FUNC DiagonalMatrix& operator=(const DiagonalMatrix& other) { m_diagonal = other.diagonal(); @@ -201,14 +227,19 @@ class DiagonalMatrix #endif /** Resizes to given size. */ + EIGEN_DEVICE_FUNC inline void resize(Index size) { m_diagonal.resize(size); } /** Sets all coefficients to zero. */ + EIGEN_DEVICE_FUNC inline void setZero() { m_diagonal.setZero(); } /** Resizes and sets all coefficients to zero. */ + EIGEN_DEVICE_FUNC inline void setZero(Index size) { m_diagonal.setZero(size); } /** Sets this matrix to be the identity matrix of the current size. */ + EIGEN_DEVICE_FUNC inline void setIdentity() { m_diagonal.setOnes(); } /** Sets this matrix to be the identity matrix of the given size. */ + EIGEN_DEVICE_FUNC inline void setIdentity(Index size) { m_diagonal.setOnes(size); } }; @@ -255,9 +286,11 @@ class DiagonalWrapper #endif /** Constructor from expression of diagonal coefficients to wrap. */ + EIGEN_DEVICE_FUNC inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {} /** \returns a const reference to the wrapped expression of diagonal coefficients. */ + EIGEN_DEVICE_FUNC const DiagonalVectorType& diagonal() const { return m_diagonal; } protected: diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index eed10de2d..8b8a868b7 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -98,6 +98,7 @@ template class MatrixBase /** \returns the size of the main diagonal, which is min(rows(),cols()). * \sa rows(), cols(), SizeAtCompileTime. */ + EIGEN_DEVICE_FUNC inline Index diagonalSize() const { return (std::min)(rows(),cols()); } /** \brief The plain matrix type corresponding to this expression. @@ -206,6 +207,7 @@ template class MatrixBase void applyOnTheRight(const EigenBase& other); template + EIGEN_DEVICE_FUNC const DiagonalProduct operator*(const DiagonalBase &diagonal) const; @@ -231,15 +233,23 @@ template class MatrixBase EIGEN_DEVICE_FUNC void adjointInPlace(); typedef Diagonal DiagonalReturnType; + EIGEN_DEVICE_FUNC DiagonalReturnType diagonal(); - typedef typename internal::add_const >::type ConstDiagonalReturnType; + + typedef typename internal::add_const >::type ConstDiagonalReturnType; + EIGEN_DEVICE_FUNC ConstDiagonalReturnType diagonal() const; template struct DiagonalIndexReturnType { typedef Diagonal Type; }; template struct ConstDiagonalIndexReturnType { typedef const Diagonal Type; }; - template typename DiagonalIndexReturnType::Type diagonal(); - template typename ConstDiagonalIndexReturnType::Type diagonal() const; + template + EIGEN_DEVICE_FUNC + typename DiagonalIndexReturnType::Type diagonal(); + + template + EIGEN_DEVICE_FUNC + typename ConstDiagonalIndexReturnType::Type diagonal() const; // Note: The "MatrixBase::" prefixes are added to help MSVC9 to match these declarations with the later implementations. // On the other hand they confuse MSVC8... @@ -247,7 +257,10 @@ template class MatrixBase typename MatrixBase::template DiagonalIndexReturnType::Type diagonal(Index index); typename MatrixBase::template ConstDiagonalIndexReturnType::Type diagonal(Index index) const; #else + EIGEN_DEVICE_FUNC typename DiagonalIndexReturnType::Type diagonal(Index index); + + EIGEN_DEVICE_FUNC typename ConstDiagonalIndexReturnType::Type diagonal(Index index) const; #endif @@ -285,6 +298,7 @@ template class MatrixBase static const BasisReturnType UnitZ(); static const BasisReturnType UnitW(); + EIGEN_DEVICE_FUNC const DiagonalWrapper asDiagonal() const; const PermutationWrapper asPermutation() const; diff --git a/Eigen/src/plugins/ArrayCwiseBinaryOps.h b/Eigen/src/plugins/ArrayCwiseBinaryOps.h index 5c8c476ee..2a77fd956 100644 --- a/Eigen/src/plugins/ArrayCwiseBinaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseBinaryOps.h @@ -3,6 +3,7 @@ * \sa MatrixBase::cwiseProduct */ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const EIGEN_CWISE_PRODUCT_RETURN_TYPE(Derived,OtherDerived) operator*(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -14,6 +15,7 @@ operator*(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const * \sa MatrixBase::cwiseQuotient */ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const OtherDerived> operator/(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -33,6 +35,7 @@ EIGEN_MAKE_CWISE_BINARY_OP(min,internal::scalar_min_op) * * \sa max() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const CwiseNullaryOp, PlainObject> > #ifdef EIGEN_PARSED_BY_DOXYGEN @@ -58,6 +61,7 @@ EIGEN_MAKE_CWISE_BINARY_OP(max,internal::scalar_max_op) * * \sa min() */ +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const CwiseNullaryOp, PlainObject> > #ifdef EIGEN_PARSED_BY_DOXYGEN @@ -143,12 +147,14 @@ EIGEN_MAKE_CWISE_BINARY_OP(operator!=,std::not_equal_to) * * \sa operator+=(), operator-() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> operator+(const Scalar& scalar) const { return CwiseUnaryOp, const Derived>(derived(), internal::scalar_add_op(scalar)); } +EIGEN_DEVICE_FUNC friend inline const CwiseUnaryOp, const Derived> operator+(const Scalar& scalar,const EIGEN_CURRENT_STORAGE_BASE_CLASS& other) { @@ -162,12 +168,14 @@ operator+(const Scalar& scalar,const EIGEN_CURRENT_STORAGE_BASE_CLASS& * * \sa operator+(), operator-=() */ +EIGEN_DEVICE_FUNC inline const CwiseUnaryOp, const Derived> operator-(const Scalar& scalar) const { return *this + (-scalar); } +EIGEN_DEVICE_FUNC friend inline const CwiseUnaryOp, const CwiseUnaryOp, const Derived> > operator-(const Scalar& scalar,const EIGEN_CURRENT_STORAGE_BASE_CLASS& other) { @@ -184,6 +192,7 @@ operator-(const Scalar& scalar,const EIGEN_CURRENT_STORAGE_BASE_CLASS& * \sa operator||(), select() */ template +EIGEN_DEVICE_FUNC inline const CwiseBinaryOp operator&&(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -202,6 +211,7 @@ operator&&(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const * \sa operator&&(), select() */ template +EIGEN_DEVICE_FUNC inline const CwiseBinaryOp operator||(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const { @@ -209,3 +219,4 @@ operator||(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL); return CwiseBinaryOp(derived(),other.derived()); } + diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index afb0fb6e3..71a7d4185 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -201,6 +201,7 @@ cube() const } #define EIGEN_MAKE_SCALAR_CWISE_UNARY_OP(METHOD_NAME,FUNCTOR) \ + EIGEN_DEVICE_FUNC \ inline const CwiseUnaryOp >, const Derived> \ METHOD_NAME(const Scalar& s) const { \ return CwiseUnaryOp >, const Derived> \ From 68168e9eae53b56a9b063d0fa5bd6f7f02cca1bf Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Wed, 31 Jul 2013 14:57:20 +0100 Subject: [PATCH 0063/1103] MatrixFunctions: replace eval() by nested. This eliminates an unnecessary copy in some situations, e.g. Map. --- .../src/MatrixFunctions/MatrixExponential.h | 7 ++----- .../src/MatrixFunctions/MatrixFunction.h | 17 ++++++++++------- .../src/MatrixFunctions/MatrixLogarithm.h | 19 +++++++++++-------- .../src/MatrixFunctions/MatrixSquareRoot.h | 10 ++++++---- 4 files changed, 29 insertions(+), 24 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index d67ded7ba..05df5a102 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -392,17 +392,14 @@ template struct MatrixExponentialReturnValue template inline void evalTo(ResultType& result) const { - const typename Derived::PlainObject srcEvaluated = m_src.eval(); - internal::matrix_exp_compute(srcEvaluated, result); + internal::matrix_exp_compute(m_src, result); } Index rows() const { return m_src.rows(); } Index cols() const { return m_src.cols(); } protected: - const Derived& m_src; - private: - MatrixExponentialReturnValue& operator=(const MatrixExponentialReturnValue&); + const typename internal::nested::type m_src; }; namespace internal { diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index a5675228e..f619d8ae5 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -481,11 +481,15 @@ template class MatrixFunctionReturnValue : public ReturnByValue > { public: - typedef typename Derived::Scalar Scalar; typedef typename Derived::Index Index; typedef typename internal::stem_function::type StemFunction; + protected: + typedef typename internal::nested::type DerivedNested; + + public: + /** \brief Constructor. * * \param[in] A %Matrix (expression) forming the argument of the matrix function. @@ -500,26 +504,25 @@ template class MatrixFunctionReturnValue template inline void evalTo(ResultType& result) const { - typedef typename Derived::PlainObject PlainObject; - typedef internal::traits Traits; + typedef typename internal::remove_all::type DerivedNestedClean; + typedef internal::traits Traits; static const int RowsAtCompileTime = Traits::RowsAtCompileTime; static const int ColsAtCompileTime = Traits::ColsAtCompileTime; - static const int Options = PlainObject::Options; + static const int Options = DerivedNestedClean::Options; typedef std::complex::Real> ComplexScalar; typedef Matrix DynMatrixType; typedef internal::MatrixFunctionAtomic AtomicType; AtomicType atomic(m_f); - const PlainObject Aevaluated = m_A.eval(); - internal::matrix_function_compute::run(Aevaluated, atomic, result); + internal::matrix_function_compute::run(m_A, atomic, result); } Index rows() const { return m_A.rows(); } Index cols() const { return m_A.cols(); } private: - typename internal::nested::type m_A; + const DerivedNested m_A; StemFunction *m_f; }; diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index 6f84a31bd..6248eeec7 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Jitse Niesen +// Copyright (C) 2011, 2013 Jitse Niesen // Copyright (C) 2011 Chen-Pang He // // This Source Code Form is subject to the terms of the Mozilla @@ -306,10 +306,14 @@ template class MatrixLogarithmReturnValue : public ReturnByValue > { public: - typedef typename Derived::Scalar Scalar; typedef typename Derived::Index Index; +protected: + typedef typename internal::nested::type DerivedNested; + +public: + /** \brief Constructor. * * \param[in] A %Matrix (expression) forming the argument of the matrix logarithm. @@ -323,25 +327,24 @@ public: template inline void evalTo(ResultType& result) const { - typedef typename Derived::PlainObject PlainObject; - typedef internal::traits Traits; + typedef typename internal::remove_all::type DerivedNestedClean; + typedef internal::traits Traits; static const int RowsAtCompileTime = Traits::RowsAtCompileTime; static const int ColsAtCompileTime = Traits::ColsAtCompileTime; - static const int Options = PlainObject::Options; + static const int Options = DerivedNestedClean::Options; typedef std::complex::Real> ComplexScalar; typedef Matrix DynMatrixType; typedef internal::MatrixLogarithmAtomic AtomicType; AtomicType atomic; - const PlainObject Aevaluated = m_A.eval(); - internal::matrix_function_compute::run(Aevaluated, atomic, result); + internal::matrix_function_compute::run(m_A, atomic, result); } Index rows() const { return m_A.rows(); } Index cols() const { return m_A.cols(); } private: - typename internal::nested::type m_A; + const DerivedNested m_A; }; namespace internal { diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h index 314b3f38e..8ca4f4864 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h @@ -318,7 +318,10 @@ struct matrix_sqrt_compute template class MatrixSquareRootReturnValue : public ReturnByValue > { + protected: typedef typename Derived::Index Index; + typedef typename internal::nested::type DerivedNested; + public: /** \brief Constructor. * @@ -335,16 +338,15 @@ template class MatrixSquareRootReturnValue template inline void evalTo(ResultType& result) const { - typedef typename Derived::PlainObject PlainObject; - const PlainObject srcEvaluated = m_src.eval(); - internal::matrix_sqrt_compute::run(srcEvaluated, result); + typedef typename internal::remove_all::type DerivedNestedClean; + internal::matrix_sqrt_compute::run(m_src, result); } Index rows() const { return m_src.rows(); } Index cols() const { return m_src.cols(); } protected: - const Derived& m_src; + const DerivedNested m_src; }; namespace internal { From 39491e3b75b8324cee480a7893b9dd2a19198b11 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Wed, 31 Jul 2013 16:16:08 +0200 Subject: [PATCH 0064/1103] Enable support for minimal rebuilds. --- cmake/EigenConfigureTesting.cmake | 28 ++++++++++++++++++---------- test/CMakeLists.txt | 25 +++++++++++++------------ 2 files changed, 31 insertions(+), 22 deletions(-) diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake index 228d29e97..7844bf4d3 100644 --- a/cmake/EigenConfigureTesting.cmake +++ b/cmake/EigenConfigureTesting.cmake @@ -24,31 +24,39 @@ set(CMAKE_MAKE_PROGRAM "@EIGEN_MAKECOMMAND_PLACEHOLDER@") # This call activates testing and generates the DartConfiguration.tcl include(CTest) -set(EIGEN_TEST_BUILD_FLAGS " " CACHE STRING "Options passed to the build command of unit tests") +set(EIGEN_TEST_BUILD_FLAGS "" CACHE STRING "Options passed to the build command of unit tests") # overwrite default DartConfiguration.tcl # The worarounds are different for each version of the MSVC IDE +set(EIGEN_TEST_TARGET buildtests) if(MSVC_IDE) if(CMAKE_MAKE_PROGRAM_SAVE MATCHES "devenv") # devenv - set(EIGEN_MAKECOMMAND_PLACEHOLDER "${CMAKE_MAKE_PROGRAM_SAVE} Eigen.sln /build \"Release\" /project buildtests ${EIGEN_TEST_BUILD_FLAGS} \n# ") + set(EIGEN_BUILD_COMMAND "${CMAKE_MAKE_PROGRAM_SAVE} Eigen.sln /build Release /project ${EIGEN_TEST_TARGET}") else() # msbuild - set(EIGEN_MAKECOMMAND_PLACEHOLDER "${CMAKE_MAKE_PROGRAM_SAVE} buildtests.vcxproj /p:Configuration=\${CTEST_CONFIGURATION_TYPE} ${EIGEN_TEST_BUILD_FLAGS}\n# ") + set(EIGEN_BUILD_COMMAND "${CMAKE_MAKE_PROGRAM_SAVE} ${EIGEN_TEST_TARGET}.vcxproj /p:Configuration=\${CTEST_CONFIGURATION_TYPE}") endif() + + # append the build flags if provided + if(NOT "${EIGEN_TEST_BUILD_FLAGS}" MATCHES "^[ \t]*$") + set(EIGEN_BUILD_COMMAND "${EIGEN_BUILD_COMMAND} ${EIGEN_TEST_BUILD_FLAGS}") + endif() + + # apply the dartconfig hack ... + set(EIGEN_MAKECOMMAND_PLACEHOLDER "${EIGEN_BUILD_COMMAND}\n#") else() # for make and nmake - set(EIGEN_MAKECOMMAND_PLACEHOLDER "${CMAKE_MAKE_PROGRAM_SAVE} buildtests ${EIGEN_TEST_BUILD_FLAGS}") + set(EIGEN_BUILD_COMMAND "${CMAKE_MAKE_PROGRAM_SAVE} ${EIGEN_TEST_TARGET} ${EIGEN_TEST_BUILD_FLAGS}") + set(EIGEN_MAKECOMMAND_PLACEHOLDER "${EIGEN_BUILD_COMMAND}") endif() -# copy ctest properties, which currently -# o raise the warning levels configure_file(${CMAKE_BINARY_DIR}/DartConfiguration.tcl ${CMAKE_BINARY_DIR}/DartConfiguration.tcl) # restore default CMAKE_MAKE_PROGRAM set(CMAKE_MAKE_PROGRAM ${CMAKE_MAKE_PROGRAM_SAVE}) -# un-set temporary variables so that it is like they never existed. -# CMake 2.6.3 introduces the more logical unset() syntax for this. -set(CMAKE_MAKE_PROGRAM_SAVE) -set(EIGEN_MAKECOMMAND_PLACEHOLDER) + +# un-set temporary variables so that it is like they never existed +unset(CMAKE_MAKE_PROGRAM_SAVE) +unset(EIGEN_MAKECOMMAND_PLACEHOLDER) configure_file(${CMAKE_SOURCE_DIR}/CTestCustom.cmake.in ${CMAKE_BINARY_DIR}/CTestCustom.cmake) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 580e2ef5e..19afe481e 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -1,16 +1,17 @@ - -# generate split test header file -message(STATUS ${CMAKE_CURRENT_BINARY_DIR}) -file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "") -foreach(i RANGE 1 999) - file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h - "#ifdef EIGEN_TEST_PART_${i}\n" - "#define CALL_SUBTEST_${i}(FUNC) CALL_SUBTEST(FUNC)\n" - "#else\n" - "#define CALL_SUBTEST_${i}(FUNC)\n" - "#endif\n\n" +# generate split test header file only if it does not yet exist +# in order to prevent a rebuild everytime cmake is configured +if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h) + file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h "") + foreach(i RANGE 1 999) + file(APPEND ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h + "#ifdef EIGEN_TEST_PART_${i}\n" + "#define CALL_SUBTEST_${i}(FUNC) CALL_SUBTEST(FUNC)\n" + "#else\n" + "#define CALL_SUBTEST_${i}(FUNC)\n" + "#endif\n\n" ) -endforeach() + endforeach() +endif() # configure blas/lapack (use Eigen's ones) set(BLAS_FOUND TRUE) From 55b57fcba6e56bea5c084cc756b50a447985e5c2 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 31 Jul 2013 16:56:31 +0200 Subject: [PATCH 0065/1103] Disable some shortcuts with nvcc --- Eigen/src/Cholesky/LDLT.h | 2 ++ Eigen/src/Cholesky/LLT.h | 4 +++- Eigen/src/QR/ColPivHouseholderQR.h | 2 ++ Eigen/src/QR/FullPivHouseholderQR.h | 2 ++ Eigen/src/QR/HouseholderQR.h | 2 ++ Eigen/src/SVD/JacobiSVD.h | 2 ++ 6 files changed, 13 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 05b300c56..4faedd257 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -575,6 +575,7 @@ MatrixType LDLT::reconstructedMatrix() const return res; } +#ifndef __CUDACC__ /** \cholesky_module * \returns the Cholesky decomposition with full pivoting without square root of \c *this */ @@ -594,6 +595,7 @@ MatrixBase::ldlt() const { return LDLT(derived()); } +#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index 2e6189f7d..2201c641e 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -465,6 +465,7 @@ MatrixType LLT::reconstructedMatrix() const return matrixL() * matrixL().adjoint().toDenseMatrix(); } +#ifndef __CUDACC__ /** \cholesky_module * \returns the LLT decomposition of \c *this */ @@ -484,7 +485,8 @@ SelfAdjointView::llt() const { return LLT(m_matrix); } - +#endif // __CUDACC__ + } // end namespace Eigen #endif // EIGEN_LLT_H diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index 8b01f8179..de9940cf5 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -563,6 +563,7 @@ typename ColPivHouseholderQR::HouseholderSequenceType ColPivHousehol return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate()).setLength(m_nonzero_pivots); } +#ifndef __CUDACC__ /** \return the column-pivoting Householder QR decomposition of \c *this. * * \sa class ColPivHouseholderQR @@ -573,6 +574,7 @@ MatrixBase::colPivHouseholderQr() const { return ColPivHouseholderQR(eval()); } +#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 0dd5ad347..a1de535a9 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -607,6 +607,7 @@ inline typename FullPivHouseholderQR::MatrixQReturnType FullPivHouse return MatrixQReturnType(m_qr, m_hCoeffs, m_rows_transpositions); } +#ifndef __CUDACC__ /** \return the full-pivoting Householder QR decomposition of \c *this. * * \sa class FullPivHouseholderQR @@ -617,6 +618,7 @@ MatrixBase::fullPivHouseholderQr() const { return FullPivHouseholderQR(eval()); } +#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index abc61bcbb..ad156396a 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -358,6 +358,7 @@ HouseholderQR& HouseholderQR::compute(const MatrixType& return *this; } +#ifndef __CUDACC__ /** \return the Householder QR decomposition of \c *this. * * \sa class HouseholderQR @@ -368,6 +369,7 @@ MatrixBase::householderQr() const { return HouseholderQR(eval()); } +#endif // __CUDACC__ } // end namespace Eigen diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 9fd9de669..7bedb02e9 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -860,6 +860,7 @@ struct solve_retval, Rhs> }; } // end namespace internal +#ifndef __CUDACC__ /** \svd_module * * \return the singular value decomposition of \c *this computed by two-sided @@ -873,6 +874,7 @@ MatrixBase::jacobiSvd(unsigned int computationOptions) const { return JacobiSVD(*this, computationOptions); } +#endif // __CUDACC__ } // end namespace Eigen From e41bc6cbbf9863c686e464d4479b4f562e5649c3 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Wed, 31 Jul 2013 17:21:32 +0200 Subject: [PATCH 0066/1103] Removed unused test files. --- test/eigen2/runtest.sh | 28 ----- test/eigen2/testsuite.cmake | 197 ------------------------------- test/runtest.sh | 20 ---- test/testsuite.cmake | 229 ------------------------------------ 4 files changed, 474 deletions(-) delete mode 100755 test/eigen2/runtest.sh delete mode 100644 test/eigen2/testsuite.cmake delete mode 100755 test/runtest.sh delete mode 100644 test/testsuite.cmake diff --git a/test/eigen2/runtest.sh b/test/eigen2/runtest.sh deleted file mode 100755 index bc693af13..000000000 --- a/test/eigen2/runtest.sh +++ /dev/null @@ -1,28 +0,0 @@ -#!/bin/bash - -black='\E[30m' -red='\E[31m' -green='\E[32m' -yellow='\E[33m' -blue='\E[34m' -magenta='\E[35m' -cyan='\E[36m' -white='\E[37m' - -if make test_$1 > /dev/null 2> .runtest.log ; then - if ! ./test_$1 r20 > /dev/null 2> .runtest.log ; then - echo -e $red Test $1 failed: $black - echo -e $blue - cat .runtest.log - echo -e $black - exit 1 - else - echo -e $green Test $1 passed$black - fi -else - echo -e $red Build of target $1 failed: $black - echo -e $blue - cat .runtest.log - echo -e $black - exit 1 -fi diff --git a/test/eigen2/testsuite.cmake b/test/eigen2/testsuite.cmake deleted file mode 100644 index 12b6bfa2e..000000000 --- a/test/eigen2/testsuite.cmake +++ /dev/null @@ -1,197 +0,0 @@ - -#################################################################### -# -# Usage: -# - create a new folder, let's call it cdash -# - in that folder, do: -# ctest -S path/to/eigen2/test/testsuite.cmake[,option1=value1[,option2=value2]] -# -# Options: -# - EIGEN_CXX: compiler, eg.: g++-4.2 -# default: default c++ compiler -# - EIGEN_SITE: eg, INRIA-Bdx_pc-gael, or the name of the contributor, etc. -# default: hostname -# - EIGEN_BUILD_STRING: a string which identify the system/compiler. It should be formed like that: -# --- -# with: -# = opensuse, debian, osx, windows, cygwin, freebsd, solaris, etc. -# = 11.1, XP, vista, leopard, etc. -# = i386, x86_64, ia64, powerpc, etc. -# = gcc-4.3.2, icc-11.0, MSVC-2008, etc. -# - EIGEN_EXPLICIT_VECTORIZATION: novec, SSE2, Altivec -# default: SSE2 for x86_64 systems, novec otherwise -# Its value is automatically appended to EIGEN_BUILD_STRING -# - EIGEN_CMAKE_DIR: path to cmake executable -# - EIGEN_MODE: dashboard model, can be Experimental, Nightly, or Continuous -# default: Nightly -# - EIGEN_WORK_DIR: directory used to download the source files and make the builds -# default: folder which contains this script -# - EIGEN_CMAKE_ARGS: additional arguments passed to cmake -# - CTEST_SOURCE_DIRECTORY: path to eigen's src (use a new and empty folder, not the one you are working on) -# default: /src -# - CTEST_BINARY_DIRECTORY: build directory -# default: /nightly- -# -# Here is an example running several compilers on a linux system: -# #!/bin/bash -# ARCH=`uname -m` -# SITE=`hostname` -# VERSION=opensuse-11.1 -# WORK_DIR=/home/gael/Coding/eigen2/cdash -# # get the last version of the script -# wget http://bitbucket.org/eigen/eigen/raw/tip/test/testsuite.cmake -o $WORK_DIR/testsuite.cmake -# COMMON="ctest -S $WORK_DIR/testsuite.cmake,EIGEN_WORK_DIR=$WORK_DIR,EIGEN_SITE=$SITE,EIGEN_MODE=$1,EIGEN_BUILD_STRING=$OS_VERSION-$ARCH" -# $COMMON-gcc-3.4.6,EIGEN_CXX=g++-3.4 -# $COMMON-gcc-4.0.1,EIGEN_CXX=g++-4.0.1 -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=novec -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=SSE2 -# $COMMON-icc-11.0,EIGEN_CXX=icpc -# -#################################################################### - -# process the arguments - -set(ARGLIST ${CTEST_SCRIPT_ARG}) -while(${ARGLIST} MATCHES ".+.*") - - # pick first - string(REGEX MATCH "([^,]*)(,.*)?" DUMMY ${ARGLIST}) - SET(TOP ${CMAKE_MATCH_1}) - - # remove first - string(REGEX MATCHALL "[^,]*,(.*)" DUMMY ${ARGLIST}) - SET(ARGLIST ${CMAKE_MATCH_1}) - - # decompose as a pair key=value - string(REGEX MATCH "([^=]*)(=.*)?" DUMMY ${TOP}) - SET(KEY ${CMAKE_MATCH_1}) - - string(REGEX MATCH "[^=]*=(.*)" DUMMY ${TOP}) - SET(VALUE ${CMAKE_MATCH_1}) - - # set the variable to the specified value - if(VALUE) - SET(${KEY} ${VALUE}) - else(VALUE) - SET(${KEY} ON) - endif(VALUE) - -endwhile(${ARGLIST} MATCHES ".+.*") - -#################################################################### -# Automatically set some user variables if they have not been defined manually -#################################################################### -cmake_minimum_required(VERSION 2.6 FATAL_ERROR) - -if(NOT EIGEN_SITE) - site_name(EIGEN_SITE) -endif(NOT EIGEN_SITE) - -if(NOT EIGEN_CMAKE_DIR) - SET(EIGEN_CMAKE_DIR "") -endif(NOT EIGEN_CMAKE_DIR) - -if(NOT EIGEN_BUILD_STRING) - - # let's try to find all information we need to make the build string ourself - - # OS - build_name(EIGEN_OS_VERSION) - - # arch - set(EIGEN_ARCH ${CMAKE_SYSTEM_PROCESSOR}) - if(WIN32) - set(EIGEN_ARCH $ENV{PROCESSOR_ARCHITECTURE}) - else(WIN32) - execute_process(COMMAND uname -m OUTPUT_VARIABLE EIGEN_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE) - endif(WIN32) - - set(EIGEN_BUILD_STRING ${EIGEN_OS_VERSION}${EIGEN_ARCH}-${EIGEN_CXX}) - -endif(NOT EIGEN_BUILD_STRING) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - set(EIGEN_BUILD_STRING ${EIGEN_BUILD_STRING}-${EIGEN_EXPLICIT_VECTORIZATION}) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(NOT EIGEN_WORK_DIR) - set(EIGEN_WORK_DIR ${CTEST_SCRIPT_DIRECTORY}) -endif(NOT EIGEN_WORK_DIR) - -if(NOT CTEST_SOURCE_DIRECTORY) - SET (CTEST_SOURCE_DIRECTORY "${EIGEN_WORK_DIR}/src") -endif(NOT CTEST_SOURCE_DIRECTORY) - -if(NOT CTEST_BINARY_DIRECTORY) - SET (CTEST_BINARY_DIRECTORY "${EIGEN_WORK_DIR}/nightly_${EIGEN_CXX}") -endif(NOT CTEST_BINARY_DIRECTORY) - -if(NOT EIGEN_MODE) - set(EIGEN_MODE Nightly) -endif(NOT EIGEN_MODE) - -## mandatory variables (the default should be ok in most cases): - -SET (CTEST_CVS_COMMAND "hg") -SET (CTEST_CVS_CHECKOUT "${CTEST_CVS_COMMAND} clone -r 2.0 http://bitbucket.org/eigen/eigen \"${CTEST_SOURCE_DIRECTORY}\"") - -# which ctest command to use for running the dashboard -SET (CTEST_COMMAND "${EIGEN_CMAKE_DIR}ctest -D ${EIGEN_MODE}") - -# what cmake command to use for configuring this dashboard -SET (CTEST_CMAKE_COMMAND "${EIGEN_CMAKE_DIR}cmake -DEIGEN_BUILD_TESTS=on ") - -#################################################################### -# The values in this section are optional you can either -# have them or leave them commented out -#################################################################### - -# this make sure we get consistent outputs -SET($ENV{LC_MESSAGES} "en_EN") - -# should ctest wipe the binary tree before running -SET(CTEST_START_WITH_EMPTY_BINARY_DIRECTORY TRUE) -SET(CTEST_BACKUP_AND_RESTORE TRUE) - -# this is the initial cache to use for the binary tree, be careful to escape -# any quotes inside of this string if you use it -if(WIN32 AND NOT UNIX) - #message(SEND_ERROR "win32") - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"NMake Makefiles\" -DCMAKE_MAKE_PROGRAM=nmake") - SET (CTEST_INITIAL_CACHE " - MAKECOMMAND:STRING=nmake -i - CMAKE_MAKE_PROGRAM:FILEPATH=nmake - CMAKE_GENERATOR:INTERNAL=NMake Makefiles - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") -else(WIN32 AND NOT UNIX) - SET (CTEST_INITIAL_CACHE " - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") -endif(WIN32 AND NOT UNIX) - -# set any extra environment variables to use during the execution of the script here: - -if(EIGEN_CXX) - set(CTEST_ENVIRONMENT "CXX=${EIGEN_CXX}") -endif(EIGEN_CXX) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - if(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE3) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES Altivec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_ALTIVEC=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES novec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_NO_EXPLICIT_VECTORIZATION=ON") - else(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - message(FATAL_ERROR "Invalid value for EIGEN_EXPLICIT_VECTORIZATION (${EIGEN_EXPLICIT_VECTORIZATION}), must be: novec, SSE2, SSE3, Altivec") - endif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(DEFINED EIGEN_CMAKE_ARGS) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} ${EIGEN_CMAKE_ARGS}") -endif(DEFINED EIGEN_CMAKE_ARGS) diff --git a/test/runtest.sh b/test/runtest.sh deleted file mode 100755 index 2be250819..000000000 --- a/test/runtest.sh +++ /dev/null @@ -1,20 +0,0 @@ -#!/bin/bash - -black='\E[30m' -red='\E[31m' -green='\E[32m' -yellow='\E[33m' -blue='\E[34m' -magenta='\E[35m' -cyan='\E[36m' -white='\E[37m' - -if ! ./$1 > /dev/null 2> .runtest.log ; then - echo -e $red Test $1 failed: $black - echo -e $blue - cat .runtest.log - echo -e $black - exit 1 -else -echo -e $green Test $1 passed$black -fi diff --git a/test/testsuite.cmake b/test/testsuite.cmake deleted file mode 100644 index 3bec56b3f..000000000 --- a/test/testsuite.cmake +++ /dev/null @@ -1,229 +0,0 @@ - -#################################################################### -# -# Usage: -# - create a new folder, let's call it cdash -# - in that folder, do: -# ctest -S path/to/eigen/test/testsuite.cmake[,option1=value1[,option2=value2]] -# -# Options: -# - EIGEN_CXX: compiler, eg.: g++-4.2 -# default: default c++ compiler -# - EIGEN_SITE: eg, INRIA-Bdx_pc-gael, or the name of the contributor, etc. -# default: hostname -# - EIGEN_BUILD_STRING: a string which identify the system/compiler. It should be formed like that: -# --- -# with: -# = opensuse, debian, osx, windows, cygwin, freebsd, solaris, etc. -# = 11.1, XP, vista, leopard, etc. -# = i386, x86_64, ia64, powerpc, etc. -# = gcc-4.3.2, icc-11.0, MSVC-2008, etc. -# - EIGEN_EXPLICIT_VECTORIZATION: novec, SSE2, Altivec -# default: SSE2 for x86_64 systems, novec otherwise -# Its value is automatically appended to EIGEN_BUILD_STRING -# - EIGEN_CMAKE_DIR: path to cmake executable -# - EIGEN_MODE: dashboard model, can be Experimental, Nightly, or Continuous -# default: Nightly -# - EIGEN_WORK_DIR: directory used to download the source files and make the builds -# default: folder which contains this script -# - EIGEN_CMAKE_ARGS: additional arguments passed to cmake -# - EIGEN_GENERATOR_TYPE: allows to overwrite the generator type -# default: nmake (windows -# See http://www.cmake.org/cmake/help/cmake2.6docs.html#section_Generators for a complete -# list of supported generators. -# - EIGEN_NO_UPDATE: allows to submit dash boards from local repositories -# This might be interesting in case you want to submit dashboards -# including local changes. -# - CTEST_SOURCE_DIRECTORY: path to eigen's src (use a new and empty folder, not the one you are working on) -# default: /src -# - CTEST_BINARY_DIRECTORY: build directory -# default: /nightly- -# -# Here is an example running several compilers on a linux system: -# #!/bin/bash -# ARCH=`uname -m` -# SITE=`hostname` -# VERSION=opensuse-11.1 -# WORK_DIR=/home/gael/Coding/eigen/cdash -# # get the last version of the script -# wget http://bitbucket.org/eigen/eigen/raw/tip/test/testsuite.cmake -o $WORK_DIR/testsuite.cmake -# COMMON="ctest -S $WORK_DIR/testsuite.cmake,EIGEN_WORK_DIR=$WORK_DIR,EIGEN_SITE=$SITE,EIGEN_MODE=$1,EIGEN_BUILD_STRING=$OS_VERSION-$ARCH" -# $COMMON-gcc-3.4.6,EIGEN_CXX=g++-3.4 -# $COMMON-gcc-4.0.1,EIGEN_CXX=g++-4.0.1 -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=novec -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=SSE2 -# $COMMON-icc-11.0,EIGEN_CXX=icpc -# -#################################################################### - -# process the arguments - -set(ARGLIST ${CTEST_SCRIPT_ARG}) -while(${ARGLIST} MATCHES ".+.*") - - # pick first - string(REGEX MATCH "([^,]*)(,.*)?" DUMMY ${ARGLIST}) - SET(TOP ${CMAKE_MATCH_1}) - - # remove first - string(REGEX MATCHALL "[^,]*,(.*)" DUMMY ${ARGLIST}) - SET(ARGLIST ${CMAKE_MATCH_1}) - - # decompose as a pair key=value - string(REGEX MATCH "([^=]*)(=.*)?" DUMMY ${TOP}) - SET(KEY ${CMAKE_MATCH_1}) - - string(REGEX MATCH "[^=]*=(.*)" DUMMY ${TOP}) - SET(VALUE ${CMAKE_MATCH_1}) - - # set the variable to the specified value - if(VALUE) - SET(${KEY} ${VALUE}) - else(VALUE) - SET(${KEY} ON) - endif(VALUE) - -endwhile(${ARGLIST} MATCHES ".+.*") - -#################################################################### -# Automatically set some user variables if they have not been defined manually -#################################################################### -cmake_minimum_required(VERSION 2.6 FATAL_ERROR) - -if(NOT EIGEN_SITE) - site_name(EIGEN_SITE) -endif(NOT EIGEN_SITE) - -if(NOT EIGEN_CMAKE_DIR) - SET(EIGEN_CMAKE_DIR "") -endif(NOT EIGEN_CMAKE_DIR) - -if(NOT EIGEN_BUILD_STRING) - - # let's try to find all information we need to make the build string ourself - - # OS - build_name(EIGEN_OS_VERSION) - - # arch - set(EIGEN_ARCH ${CMAKE_SYSTEM_PROCESSOR}) - if(WIN32) - set(EIGEN_ARCH $ENV{PROCESSOR_ARCHITECTURE}) - else(WIN32) - execute_process(COMMAND uname -m OUTPUT_VARIABLE EIGEN_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE) - endif(WIN32) - - set(EIGEN_BUILD_STRING ${EIGEN_OS_VERSION}${EIGEN_ARCH}-${EIGEN_CXX}) - -endif(NOT EIGEN_BUILD_STRING) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - set(EIGEN_BUILD_STRING ${EIGEN_BUILD_STRING}-${EIGEN_EXPLICIT_VECTORIZATION}) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(NOT EIGEN_WORK_DIR) - set(EIGEN_WORK_DIR ${CTEST_SCRIPT_DIRECTORY}) -endif(NOT EIGEN_WORK_DIR) - -if(NOT CTEST_SOURCE_DIRECTORY) - SET (CTEST_SOURCE_DIRECTORY "${EIGEN_WORK_DIR}/src") -endif(NOT CTEST_SOURCE_DIRECTORY) - -if(NOT CTEST_BINARY_DIRECTORY) - SET (CTEST_BINARY_DIRECTORY "${EIGEN_WORK_DIR}/nightly_${EIGEN_CXX}") -endif(NOT CTEST_BINARY_DIRECTORY) - -if(NOT EIGEN_MODE) - set(EIGEN_MODE Nightly) -endif(NOT EIGEN_MODE) - -## mandatory variables (the default should be ok in most cases): - -if(NOT EIGEN_NO_UPDATE) - SET (CTEST_CVS_COMMAND "hg") - SET (CTEST_CVS_CHECKOUT "${CTEST_CVS_COMMAND} clone http://bitbucket.org/eigen/eigen \"${CTEST_SOURCE_DIRECTORY}\"") - SET(CTEST_BACKUP_AND_RESTORE TRUE) # the backup is CVS related ... -endif(NOT EIGEN_NO_UPDATE) - -# which ctest command to use for running the dashboard -SET (CTEST_COMMAND "${EIGEN_CMAKE_DIR}ctest -D ${EIGEN_MODE} --no-compress-output") -if($ENV{EIGEN_CTEST_ARGS}) -SET (CTEST_COMMAND "${CTEST_COMMAND} $ENV{EIGEN_CTEST_ARGS}") -endif($ENV{EIGEN_CTEST_ARGS}) -# what cmake command to use for configuring this dashboard -SET (CTEST_CMAKE_COMMAND "${EIGEN_CMAKE_DIR}cmake -DEIGEN_LEAVE_TEST_IN_ALL_TARGET=ON") - -#################################################################### -# The values in this section are optional you can either -# have them or leave them commented out -#################################################################### - -# this make sure we get consistent outputs -SET($ENV{LC_MESSAGES} "en_EN") - -# should ctest wipe the binary tree before running -SET(CTEST_START_WITH_EMPTY_BINARY_DIRECTORY TRUE) - -# raise the warning/error limit -set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_WARNINGS "33331") -set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_ERRORS "33331") - -# this is the initial cache to use for the binary tree, be careful to escape -# any quotes inside of this string if you use it -if(WIN32 AND NOT UNIX) - #message(SEND_ERROR "win32") - if(EIGEN_GENERATOR_TYPE) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"${EIGEN_GENERATOR_TYPE}\"") - SET (CTEST_INITIAL_CACHE " - CMAKE_BUILD_TYPE:STRING=Release - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") - else(EIGEN_GENERATOR_TYPE) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"NMake Makefiles\" -DCMAKE_MAKE_PROGRAM=nmake") - SET (CTEST_INITIAL_CACHE " - MAKECOMMAND:STRING=nmake /i - CMAKE_MAKE_PROGRAM:FILEPATH=nmake - CMAKE_GENERATOR:INTERNAL=NMake Makefiles - CMAKE_BUILD_TYPE:STRING=Release - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") - endif(EIGEN_GENERATOR_TYPE) -else(WIN32 AND NOT UNIX) - SET (CTEST_INITIAL_CACHE " - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") -endif(WIN32 AND NOT UNIX) - -# set any extra environment variables to use during the execution of the script here: -# setting this variable on windows machines causes trouble ... - -if(EIGEN_CXX AND NOT WIN32) - set(CTEST_ENVIRONMENT "CXX=${EIGEN_CXX}") -endif(EIGEN_CXX AND NOT WIN32) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - if(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE3) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSSE3) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE4_1) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON -DEIGEN_TEST_SSE4_1=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE4_2) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON -DEIGEN_TEST_SSE4_1=ON -DEIGEN_TEST_SSE4_2=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES Altivec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_ALTIVEC=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES novec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_NO_EXPLICIT_VECTORIZATION=ON") - else(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - message(FATAL_ERROR "Invalid value for EIGEN_EXPLICIT_VECTORIZATION (${EIGEN_EXPLICIT_VECTORIZATION}), must be: novec, SSE2, SSE3, Altivec") - endif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(DEFINED EIGEN_CMAKE_ARGS) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} ${EIGEN_CMAKE_ARGS}") -endif(DEFINED EIGEN_CMAKE_ARGS) From 32d46dd9b8a095ba9720b92864a42d7d9c8d494d Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Wed, 31 Jul 2013 18:03:34 +0200 Subject: [PATCH 0067/1103] Backed out changeset: e41bc6cbbf9863c686e464d4479b4f562e5649c3 --- test/eigen2/runtest.sh | 28 +++++ test/eigen2/testsuite.cmake | 197 +++++++++++++++++++++++++++++++ test/runtest.sh | 20 ++++ test/testsuite.cmake | 229 ++++++++++++++++++++++++++++++++++++ 4 files changed, 474 insertions(+) create mode 100644 test/eigen2/runtest.sh create mode 100644 test/eigen2/testsuite.cmake create mode 100644 test/runtest.sh create mode 100644 test/testsuite.cmake diff --git a/test/eigen2/runtest.sh b/test/eigen2/runtest.sh new file mode 100644 index 000000000..bc693af13 --- /dev/null +++ b/test/eigen2/runtest.sh @@ -0,0 +1,28 @@ +#!/bin/bash + +black='\E[30m' +red='\E[31m' +green='\E[32m' +yellow='\E[33m' +blue='\E[34m' +magenta='\E[35m' +cyan='\E[36m' +white='\E[37m' + +if make test_$1 > /dev/null 2> .runtest.log ; then + if ! ./test_$1 r20 > /dev/null 2> .runtest.log ; then + echo -e $red Test $1 failed: $black + echo -e $blue + cat .runtest.log + echo -e $black + exit 1 + else + echo -e $green Test $1 passed$black + fi +else + echo -e $red Build of target $1 failed: $black + echo -e $blue + cat .runtest.log + echo -e $black + exit 1 +fi diff --git a/test/eigen2/testsuite.cmake b/test/eigen2/testsuite.cmake new file mode 100644 index 000000000..12b6bfa2e --- /dev/null +++ b/test/eigen2/testsuite.cmake @@ -0,0 +1,197 @@ + +#################################################################### +# +# Usage: +# - create a new folder, let's call it cdash +# - in that folder, do: +# ctest -S path/to/eigen2/test/testsuite.cmake[,option1=value1[,option2=value2]] +# +# Options: +# - EIGEN_CXX: compiler, eg.: g++-4.2 +# default: default c++ compiler +# - EIGEN_SITE: eg, INRIA-Bdx_pc-gael, or the name of the contributor, etc. +# default: hostname +# - EIGEN_BUILD_STRING: a string which identify the system/compiler. It should be formed like that: +# --- +# with: +# = opensuse, debian, osx, windows, cygwin, freebsd, solaris, etc. +# = 11.1, XP, vista, leopard, etc. +# = i386, x86_64, ia64, powerpc, etc. +# = gcc-4.3.2, icc-11.0, MSVC-2008, etc. +# - EIGEN_EXPLICIT_VECTORIZATION: novec, SSE2, Altivec +# default: SSE2 for x86_64 systems, novec otherwise +# Its value is automatically appended to EIGEN_BUILD_STRING +# - EIGEN_CMAKE_DIR: path to cmake executable +# - EIGEN_MODE: dashboard model, can be Experimental, Nightly, or Continuous +# default: Nightly +# - EIGEN_WORK_DIR: directory used to download the source files and make the builds +# default: folder which contains this script +# - EIGEN_CMAKE_ARGS: additional arguments passed to cmake +# - CTEST_SOURCE_DIRECTORY: path to eigen's src (use a new and empty folder, not the one you are working on) +# default: /src +# - CTEST_BINARY_DIRECTORY: build directory +# default: /nightly- +# +# Here is an example running several compilers on a linux system: +# #!/bin/bash +# ARCH=`uname -m` +# SITE=`hostname` +# VERSION=opensuse-11.1 +# WORK_DIR=/home/gael/Coding/eigen2/cdash +# # get the last version of the script +# wget http://bitbucket.org/eigen/eigen/raw/tip/test/testsuite.cmake -o $WORK_DIR/testsuite.cmake +# COMMON="ctest -S $WORK_DIR/testsuite.cmake,EIGEN_WORK_DIR=$WORK_DIR,EIGEN_SITE=$SITE,EIGEN_MODE=$1,EIGEN_BUILD_STRING=$OS_VERSION-$ARCH" +# $COMMON-gcc-3.4.6,EIGEN_CXX=g++-3.4 +# $COMMON-gcc-4.0.1,EIGEN_CXX=g++-4.0.1 +# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=novec +# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=SSE2 +# $COMMON-icc-11.0,EIGEN_CXX=icpc +# +#################################################################### + +# process the arguments + +set(ARGLIST ${CTEST_SCRIPT_ARG}) +while(${ARGLIST} MATCHES ".+.*") + + # pick first + string(REGEX MATCH "([^,]*)(,.*)?" DUMMY ${ARGLIST}) + SET(TOP ${CMAKE_MATCH_1}) + + # remove first + string(REGEX MATCHALL "[^,]*,(.*)" DUMMY ${ARGLIST}) + SET(ARGLIST ${CMAKE_MATCH_1}) + + # decompose as a pair key=value + string(REGEX MATCH "([^=]*)(=.*)?" DUMMY ${TOP}) + SET(KEY ${CMAKE_MATCH_1}) + + string(REGEX MATCH "[^=]*=(.*)" DUMMY ${TOP}) + SET(VALUE ${CMAKE_MATCH_1}) + + # set the variable to the specified value + if(VALUE) + SET(${KEY} ${VALUE}) + else(VALUE) + SET(${KEY} ON) + endif(VALUE) + +endwhile(${ARGLIST} MATCHES ".+.*") + +#################################################################### +# Automatically set some user variables if they have not been defined manually +#################################################################### +cmake_minimum_required(VERSION 2.6 FATAL_ERROR) + +if(NOT EIGEN_SITE) + site_name(EIGEN_SITE) +endif(NOT EIGEN_SITE) + +if(NOT EIGEN_CMAKE_DIR) + SET(EIGEN_CMAKE_DIR "") +endif(NOT EIGEN_CMAKE_DIR) + +if(NOT EIGEN_BUILD_STRING) + + # let's try to find all information we need to make the build string ourself + + # OS + build_name(EIGEN_OS_VERSION) + + # arch + set(EIGEN_ARCH ${CMAKE_SYSTEM_PROCESSOR}) + if(WIN32) + set(EIGEN_ARCH $ENV{PROCESSOR_ARCHITECTURE}) + else(WIN32) + execute_process(COMMAND uname -m OUTPUT_VARIABLE EIGEN_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE) + endif(WIN32) + + set(EIGEN_BUILD_STRING ${EIGEN_OS_VERSION}${EIGEN_ARCH}-${EIGEN_CXX}) + +endif(NOT EIGEN_BUILD_STRING) + +if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + set(EIGEN_BUILD_STRING ${EIGEN_BUILD_STRING}-${EIGEN_EXPLICIT_VECTORIZATION}) +endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + +if(NOT EIGEN_WORK_DIR) + set(EIGEN_WORK_DIR ${CTEST_SCRIPT_DIRECTORY}) +endif(NOT EIGEN_WORK_DIR) + +if(NOT CTEST_SOURCE_DIRECTORY) + SET (CTEST_SOURCE_DIRECTORY "${EIGEN_WORK_DIR}/src") +endif(NOT CTEST_SOURCE_DIRECTORY) + +if(NOT CTEST_BINARY_DIRECTORY) + SET (CTEST_BINARY_DIRECTORY "${EIGEN_WORK_DIR}/nightly_${EIGEN_CXX}") +endif(NOT CTEST_BINARY_DIRECTORY) + +if(NOT EIGEN_MODE) + set(EIGEN_MODE Nightly) +endif(NOT EIGEN_MODE) + +## mandatory variables (the default should be ok in most cases): + +SET (CTEST_CVS_COMMAND "hg") +SET (CTEST_CVS_CHECKOUT "${CTEST_CVS_COMMAND} clone -r 2.0 http://bitbucket.org/eigen/eigen \"${CTEST_SOURCE_DIRECTORY}\"") + +# which ctest command to use for running the dashboard +SET (CTEST_COMMAND "${EIGEN_CMAKE_DIR}ctest -D ${EIGEN_MODE}") + +# what cmake command to use for configuring this dashboard +SET (CTEST_CMAKE_COMMAND "${EIGEN_CMAKE_DIR}cmake -DEIGEN_BUILD_TESTS=on ") + +#################################################################### +# The values in this section are optional you can either +# have them or leave them commented out +#################################################################### + +# this make sure we get consistent outputs +SET($ENV{LC_MESSAGES} "en_EN") + +# should ctest wipe the binary tree before running +SET(CTEST_START_WITH_EMPTY_BINARY_DIRECTORY TRUE) +SET(CTEST_BACKUP_AND_RESTORE TRUE) + +# this is the initial cache to use for the binary tree, be careful to escape +# any quotes inside of this string if you use it +if(WIN32 AND NOT UNIX) + #message(SEND_ERROR "win32") + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"NMake Makefiles\" -DCMAKE_MAKE_PROGRAM=nmake") + SET (CTEST_INITIAL_CACHE " + MAKECOMMAND:STRING=nmake -i + CMAKE_MAKE_PROGRAM:FILEPATH=nmake + CMAKE_GENERATOR:INTERNAL=NMake Makefiles + BUILDNAME:STRING=${EIGEN_BUILD_STRING} + SITE:STRING=${EIGEN_SITE} + ") +else(WIN32 AND NOT UNIX) + SET (CTEST_INITIAL_CACHE " + BUILDNAME:STRING=${EIGEN_BUILD_STRING} + SITE:STRING=${EIGEN_SITE} + ") +endif(WIN32 AND NOT UNIX) + +# set any extra environment variables to use during the execution of the script here: + +if(EIGEN_CXX) + set(CTEST_ENVIRONMENT "CXX=${EIGEN_CXX}") +endif(EIGEN_CXX) + +if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + if(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE3) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES Altivec) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_ALTIVEC=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES novec) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_NO_EXPLICIT_VECTORIZATION=ON") + else(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) + message(FATAL_ERROR "Invalid value for EIGEN_EXPLICIT_VECTORIZATION (${EIGEN_EXPLICIT_VECTORIZATION}), must be: novec, SSE2, SSE3, Altivec") + endif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) +endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + +if(DEFINED EIGEN_CMAKE_ARGS) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} ${EIGEN_CMAKE_ARGS}") +endif(DEFINED EIGEN_CMAKE_ARGS) diff --git a/test/runtest.sh b/test/runtest.sh new file mode 100644 index 000000000..2be250819 --- /dev/null +++ b/test/runtest.sh @@ -0,0 +1,20 @@ +#!/bin/bash + +black='\E[30m' +red='\E[31m' +green='\E[32m' +yellow='\E[33m' +blue='\E[34m' +magenta='\E[35m' +cyan='\E[36m' +white='\E[37m' + +if ! ./$1 > /dev/null 2> .runtest.log ; then + echo -e $red Test $1 failed: $black + echo -e $blue + cat .runtest.log + echo -e $black + exit 1 +else +echo -e $green Test $1 passed$black +fi diff --git a/test/testsuite.cmake b/test/testsuite.cmake new file mode 100644 index 000000000..3bec56b3f --- /dev/null +++ b/test/testsuite.cmake @@ -0,0 +1,229 @@ + +#################################################################### +# +# Usage: +# - create a new folder, let's call it cdash +# - in that folder, do: +# ctest -S path/to/eigen/test/testsuite.cmake[,option1=value1[,option2=value2]] +# +# Options: +# - EIGEN_CXX: compiler, eg.: g++-4.2 +# default: default c++ compiler +# - EIGEN_SITE: eg, INRIA-Bdx_pc-gael, or the name of the contributor, etc. +# default: hostname +# - EIGEN_BUILD_STRING: a string which identify the system/compiler. It should be formed like that: +# --- +# with: +# = opensuse, debian, osx, windows, cygwin, freebsd, solaris, etc. +# = 11.1, XP, vista, leopard, etc. +# = i386, x86_64, ia64, powerpc, etc. +# = gcc-4.3.2, icc-11.0, MSVC-2008, etc. +# - EIGEN_EXPLICIT_VECTORIZATION: novec, SSE2, Altivec +# default: SSE2 for x86_64 systems, novec otherwise +# Its value is automatically appended to EIGEN_BUILD_STRING +# - EIGEN_CMAKE_DIR: path to cmake executable +# - EIGEN_MODE: dashboard model, can be Experimental, Nightly, or Continuous +# default: Nightly +# - EIGEN_WORK_DIR: directory used to download the source files and make the builds +# default: folder which contains this script +# - EIGEN_CMAKE_ARGS: additional arguments passed to cmake +# - EIGEN_GENERATOR_TYPE: allows to overwrite the generator type +# default: nmake (windows +# See http://www.cmake.org/cmake/help/cmake2.6docs.html#section_Generators for a complete +# list of supported generators. +# - EIGEN_NO_UPDATE: allows to submit dash boards from local repositories +# This might be interesting in case you want to submit dashboards +# including local changes. +# - CTEST_SOURCE_DIRECTORY: path to eigen's src (use a new and empty folder, not the one you are working on) +# default: /src +# - CTEST_BINARY_DIRECTORY: build directory +# default: /nightly- +# +# Here is an example running several compilers on a linux system: +# #!/bin/bash +# ARCH=`uname -m` +# SITE=`hostname` +# VERSION=opensuse-11.1 +# WORK_DIR=/home/gael/Coding/eigen/cdash +# # get the last version of the script +# wget http://bitbucket.org/eigen/eigen/raw/tip/test/testsuite.cmake -o $WORK_DIR/testsuite.cmake +# COMMON="ctest -S $WORK_DIR/testsuite.cmake,EIGEN_WORK_DIR=$WORK_DIR,EIGEN_SITE=$SITE,EIGEN_MODE=$1,EIGEN_BUILD_STRING=$OS_VERSION-$ARCH" +# $COMMON-gcc-3.4.6,EIGEN_CXX=g++-3.4 +# $COMMON-gcc-4.0.1,EIGEN_CXX=g++-4.0.1 +# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=novec +# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=SSE2 +# $COMMON-icc-11.0,EIGEN_CXX=icpc +# +#################################################################### + +# process the arguments + +set(ARGLIST ${CTEST_SCRIPT_ARG}) +while(${ARGLIST} MATCHES ".+.*") + + # pick first + string(REGEX MATCH "([^,]*)(,.*)?" DUMMY ${ARGLIST}) + SET(TOP ${CMAKE_MATCH_1}) + + # remove first + string(REGEX MATCHALL "[^,]*,(.*)" DUMMY ${ARGLIST}) + SET(ARGLIST ${CMAKE_MATCH_1}) + + # decompose as a pair key=value + string(REGEX MATCH "([^=]*)(=.*)?" DUMMY ${TOP}) + SET(KEY ${CMAKE_MATCH_1}) + + string(REGEX MATCH "[^=]*=(.*)" DUMMY ${TOP}) + SET(VALUE ${CMAKE_MATCH_1}) + + # set the variable to the specified value + if(VALUE) + SET(${KEY} ${VALUE}) + else(VALUE) + SET(${KEY} ON) + endif(VALUE) + +endwhile(${ARGLIST} MATCHES ".+.*") + +#################################################################### +# Automatically set some user variables if they have not been defined manually +#################################################################### +cmake_minimum_required(VERSION 2.6 FATAL_ERROR) + +if(NOT EIGEN_SITE) + site_name(EIGEN_SITE) +endif(NOT EIGEN_SITE) + +if(NOT EIGEN_CMAKE_DIR) + SET(EIGEN_CMAKE_DIR "") +endif(NOT EIGEN_CMAKE_DIR) + +if(NOT EIGEN_BUILD_STRING) + + # let's try to find all information we need to make the build string ourself + + # OS + build_name(EIGEN_OS_VERSION) + + # arch + set(EIGEN_ARCH ${CMAKE_SYSTEM_PROCESSOR}) + if(WIN32) + set(EIGEN_ARCH $ENV{PROCESSOR_ARCHITECTURE}) + else(WIN32) + execute_process(COMMAND uname -m OUTPUT_VARIABLE EIGEN_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE) + endif(WIN32) + + set(EIGEN_BUILD_STRING ${EIGEN_OS_VERSION}${EIGEN_ARCH}-${EIGEN_CXX}) + +endif(NOT EIGEN_BUILD_STRING) + +if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + set(EIGEN_BUILD_STRING ${EIGEN_BUILD_STRING}-${EIGEN_EXPLICIT_VECTORIZATION}) +endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + +if(NOT EIGEN_WORK_DIR) + set(EIGEN_WORK_DIR ${CTEST_SCRIPT_DIRECTORY}) +endif(NOT EIGEN_WORK_DIR) + +if(NOT CTEST_SOURCE_DIRECTORY) + SET (CTEST_SOURCE_DIRECTORY "${EIGEN_WORK_DIR}/src") +endif(NOT CTEST_SOURCE_DIRECTORY) + +if(NOT CTEST_BINARY_DIRECTORY) + SET (CTEST_BINARY_DIRECTORY "${EIGEN_WORK_DIR}/nightly_${EIGEN_CXX}") +endif(NOT CTEST_BINARY_DIRECTORY) + +if(NOT EIGEN_MODE) + set(EIGEN_MODE Nightly) +endif(NOT EIGEN_MODE) + +## mandatory variables (the default should be ok in most cases): + +if(NOT EIGEN_NO_UPDATE) + SET (CTEST_CVS_COMMAND "hg") + SET (CTEST_CVS_CHECKOUT "${CTEST_CVS_COMMAND} clone http://bitbucket.org/eigen/eigen \"${CTEST_SOURCE_DIRECTORY}\"") + SET(CTEST_BACKUP_AND_RESTORE TRUE) # the backup is CVS related ... +endif(NOT EIGEN_NO_UPDATE) + +# which ctest command to use for running the dashboard +SET (CTEST_COMMAND "${EIGEN_CMAKE_DIR}ctest -D ${EIGEN_MODE} --no-compress-output") +if($ENV{EIGEN_CTEST_ARGS}) +SET (CTEST_COMMAND "${CTEST_COMMAND} $ENV{EIGEN_CTEST_ARGS}") +endif($ENV{EIGEN_CTEST_ARGS}) +# what cmake command to use for configuring this dashboard +SET (CTEST_CMAKE_COMMAND "${EIGEN_CMAKE_DIR}cmake -DEIGEN_LEAVE_TEST_IN_ALL_TARGET=ON") + +#################################################################### +# The values in this section are optional you can either +# have them or leave them commented out +#################################################################### + +# this make sure we get consistent outputs +SET($ENV{LC_MESSAGES} "en_EN") + +# should ctest wipe the binary tree before running +SET(CTEST_START_WITH_EMPTY_BINARY_DIRECTORY TRUE) + +# raise the warning/error limit +set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_WARNINGS "33331") +set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_ERRORS "33331") + +# this is the initial cache to use for the binary tree, be careful to escape +# any quotes inside of this string if you use it +if(WIN32 AND NOT UNIX) + #message(SEND_ERROR "win32") + if(EIGEN_GENERATOR_TYPE) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"${EIGEN_GENERATOR_TYPE}\"") + SET (CTEST_INITIAL_CACHE " + CMAKE_BUILD_TYPE:STRING=Release + BUILDNAME:STRING=${EIGEN_BUILD_STRING} + SITE:STRING=${EIGEN_SITE} + ") + else(EIGEN_GENERATOR_TYPE) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"NMake Makefiles\" -DCMAKE_MAKE_PROGRAM=nmake") + SET (CTEST_INITIAL_CACHE " + MAKECOMMAND:STRING=nmake /i + CMAKE_MAKE_PROGRAM:FILEPATH=nmake + CMAKE_GENERATOR:INTERNAL=NMake Makefiles + CMAKE_BUILD_TYPE:STRING=Release + BUILDNAME:STRING=${EIGEN_BUILD_STRING} + SITE:STRING=${EIGEN_SITE} + ") + endif(EIGEN_GENERATOR_TYPE) +else(WIN32 AND NOT UNIX) + SET (CTEST_INITIAL_CACHE " + BUILDNAME:STRING=${EIGEN_BUILD_STRING} + SITE:STRING=${EIGEN_SITE} + ") +endif(WIN32 AND NOT UNIX) + +# set any extra environment variables to use during the execution of the script here: +# setting this variable on windows machines causes trouble ... + +if(EIGEN_CXX AND NOT WIN32) + set(CTEST_ENVIRONMENT "CXX=${EIGEN_CXX}") +endif(EIGEN_CXX AND NOT WIN32) + +if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + if(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE3) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSSE3) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE4_1) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON -DEIGEN_TEST_SSE4_1=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE4_2) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON -DEIGEN_TEST_SSE4_1=ON -DEIGEN_TEST_SSE4_2=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES Altivec) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_ALTIVEC=ON") + elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES novec) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_NO_EXPLICIT_VECTORIZATION=ON") + else(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) + message(FATAL_ERROR "Invalid value for EIGEN_EXPLICIT_VECTORIZATION (${EIGEN_EXPLICIT_VECTORIZATION}), must be: novec, SSE2, SSE3, Altivec") + endif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) +endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) + +if(DEFINED EIGEN_CMAKE_ARGS) + set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} ${EIGEN_CMAKE_ARGS}") +endif(DEFINED EIGEN_CMAKE_ARGS) From 8e6d0cba5ff1eaaf1307d158dffd75d9f5ca3f88 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Wed, 31 Jul 2013 18:20:58 +0200 Subject: [PATCH 0068/1103] Added a pattern which forces LF line endings for *.sh files. --- .hgeol | 1 + 1 file changed, 1 insertion(+) diff --git a/.hgeol b/.hgeol index 5af3cc043..5327df161 100644 --- a/.hgeol +++ b/.hgeol @@ -1,4 +1,5 @@ [patterns] +*.sh = LF *.MINPACK = CRLF scripts/*.in = LF debug/msvc/*.dat = CRLF From d0e543be26f1ba68a3b2b99ecdb16ffaa9d62acf Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 31 Jul 2013 23:11:43 +0200 Subject: [PATCH 0069/1103] Remove superfluous testing files (as changeset e41bc6cbbf9863c686e464d4479b4f562e5649c3 but more complete) --- cmake/EigenTesting.cmake | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index d9e22ab1a..ed683a4ef 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -68,11 +68,7 @@ macro(ei_add_test_internal testname testname_with_suffix) endif() endif() - if(EIGEN_BIN_BASH_EXISTS) - add_test(${testname_with_suffix} "${Eigen_SOURCE_DIR}/test/runtest.sh" "${testname_with_suffix}") - else() - add_test(${testname_with_suffix} "${targetname}") - endif() + add_test(${testname_with_suffix} "${targetname}") # Specify target and test labels accoirding to EIGEN_CURRENT_SUBPROJECT get_property(current_subproject GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT) From 222eedf5f345d2ca9f8a65bb71d6d88bf30dbc20 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Thu, 1 Aug 2013 12:14:03 +0200 Subject: [PATCH 0070/1103] Removed unused testing files. --- test/eigen2/runtest.sh | 28 ----- test/eigen2/testsuite.cmake | 197 ------------------------------- test/runtest.sh | 20 ---- test/testsuite.cmake | 229 ------------------------------------ 4 files changed, 474 deletions(-) delete mode 100644 test/eigen2/runtest.sh delete mode 100644 test/eigen2/testsuite.cmake delete mode 100644 test/runtest.sh delete mode 100644 test/testsuite.cmake diff --git a/test/eigen2/runtest.sh b/test/eigen2/runtest.sh deleted file mode 100644 index bc693af13..000000000 --- a/test/eigen2/runtest.sh +++ /dev/null @@ -1,28 +0,0 @@ -#!/bin/bash - -black='\E[30m' -red='\E[31m' -green='\E[32m' -yellow='\E[33m' -blue='\E[34m' -magenta='\E[35m' -cyan='\E[36m' -white='\E[37m' - -if make test_$1 > /dev/null 2> .runtest.log ; then - if ! ./test_$1 r20 > /dev/null 2> .runtest.log ; then - echo -e $red Test $1 failed: $black - echo -e $blue - cat .runtest.log - echo -e $black - exit 1 - else - echo -e $green Test $1 passed$black - fi -else - echo -e $red Build of target $1 failed: $black - echo -e $blue - cat .runtest.log - echo -e $black - exit 1 -fi diff --git a/test/eigen2/testsuite.cmake b/test/eigen2/testsuite.cmake deleted file mode 100644 index 12b6bfa2e..000000000 --- a/test/eigen2/testsuite.cmake +++ /dev/null @@ -1,197 +0,0 @@ - -#################################################################### -# -# Usage: -# - create a new folder, let's call it cdash -# - in that folder, do: -# ctest -S path/to/eigen2/test/testsuite.cmake[,option1=value1[,option2=value2]] -# -# Options: -# - EIGEN_CXX: compiler, eg.: g++-4.2 -# default: default c++ compiler -# - EIGEN_SITE: eg, INRIA-Bdx_pc-gael, or the name of the contributor, etc. -# default: hostname -# - EIGEN_BUILD_STRING: a string which identify the system/compiler. It should be formed like that: -# --- -# with: -# = opensuse, debian, osx, windows, cygwin, freebsd, solaris, etc. -# = 11.1, XP, vista, leopard, etc. -# = i386, x86_64, ia64, powerpc, etc. -# = gcc-4.3.2, icc-11.0, MSVC-2008, etc. -# - EIGEN_EXPLICIT_VECTORIZATION: novec, SSE2, Altivec -# default: SSE2 for x86_64 systems, novec otherwise -# Its value is automatically appended to EIGEN_BUILD_STRING -# - EIGEN_CMAKE_DIR: path to cmake executable -# - EIGEN_MODE: dashboard model, can be Experimental, Nightly, or Continuous -# default: Nightly -# - EIGEN_WORK_DIR: directory used to download the source files and make the builds -# default: folder which contains this script -# - EIGEN_CMAKE_ARGS: additional arguments passed to cmake -# - CTEST_SOURCE_DIRECTORY: path to eigen's src (use a new and empty folder, not the one you are working on) -# default: /src -# - CTEST_BINARY_DIRECTORY: build directory -# default: /nightly- -# -# Here is an example running several compilers on a linux system: -# #!/bin/bash -# ARCH=`uname -m` -# SITE=`hostname` -# VERSION=opensuse-11.1 -# WORK_DIR=/home/gael/Coding/eigen2/cdash -# # get the last version of the script -# wget http://bitbucket.org/eigen/eigen/raw/tip/test/testsuite.cmake -o $WORK_DIR/testsuite.cmake -# COMMON="ctest -S $WORK_DIR/testsuite.cmake,EIGEN_WORK_DIR=$WORK_DIR,EIGEN_SITE=$SITE,EIGEN_MODE=$1,EIGEN_BUILD_STRING=$OS_VERSION-$ARCH" -# $COMMON-gcc-3.4.6,EIGEN_CXX=g++-3.4 -# $COMMON-gcc-4.0.1,EIGEN_CXX=g++-4.0.1 -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=novec -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=SSE2 -# $COMMON-icc-11.0,EIGEN_CXX=icpc -# -#################################################################### - -# process the arguments - -set(ARGLIST ${CTEST_SCRIPT_ARG}) -while(${ARGLIST} MATCHES ".+.*") - - # pick first - string(REGEX MATCH "([^,]*)(,.*)?" DUMMY ${ARGLIST}) - SET(TOP ${CMAKE_MATCH_1}) - - # remove first - string(REGEX MATCHALL "[^,]*,(.*)" DUMMY ${ARGLIST}) - SET(ARGLIST ${CMAKE_MATCH_1}) - - # decompose as a pair key=value - string(REGEX MATCH "([^=]*)(=.*)?" DUMMY ${TOP}) - SET(KEY ${CMAKE_MATCH_1}) - - string(REGEX MATCH "[^=]*=(.*)" DUMMY ${TOP}) - SET(VALUE ${CMAKE_MATCH_1}) - - # set the variable to the specified value - if(VALUE) - SET(${KEY} ${VALUE}) - else(VALUE) - SET(${KEY} ON) - endif(VALUE) - -endwhile(${ARGLIST} MATCHES ".+.*") - -#################################################################### -# Automatically set some user variables if they have not been defined manually -#################################################################### -cmake_minimum_required(VERSION 2.6 FATAL_ERROR) - -if(NOT EIGEN_SITE) - site_name(EIGEN_SITE) -endif(NOT EIGEN_SITE) - -if(NOT EIGEN_CMAKE_DIR) - SET(EIGEN_CMAKE_DIR "") -endif(NOT EIGEN_CMAKE_DIR) - -if(NOT EIGEN_BUILD_STRING) - - # let's try to find all information we need to make the build string ourself - - # OS - build_name(EIGEN_OS_VERSION) - - # arch - set(EIGEN_ARCH ${CMAKE_SYSTEM_PROCESSOR}) - if(WIN32) - set(EIGEN_ARCH $ENV{PROCESSOR_ARCHITECTURE}) - else(WIN32) - execute_process(COMMAND uname -m OUTPUT_VARIABLE EIGEN_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE) - endif(WIN32) - - set(EIGEN_BUILD_STRING ${EIGEN_OS_VERSION}${EIGEN_ARCH}-${EIGEN_CXX}) - -endif(NOT EIGEN_BUILD_STRING) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - set(EIGEN_BUILD_STRING ${EIGEN_BUILD_STRING}-${EIGEN_EXPLICIT_VECTORIZATION}) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(NOT EIGEN_WORK_DIR) - set(EIGEN_WORK_DIR ${CTEST_SCRIPT_DIRECTORY}) -endif(NOT EIGEN_WORK_DIR) - -if(NOT CTEST_SOURCE_DIRECTORY) - SET (CTEST_SOURCE_DIRECTORY "${EIGEN_WORK_DIR}/src") -endif(NOT CTEST_SOURCE_DIRECTORY) - -if(NOT CTEST_BINARY_DIRECTORY) - SET (CTEST_BINARY_DIRECTORY "${EIGEN_WORK_DIR}/nightly_${EIGEN_CXX}") -endif(NOT CTEST_BINARY_DIRECTORY) - -if(NOT EIGEN_MODE) - set(EIGEN_MODE Nightly) -endif(NOT EIGEN_MODE) - -## mandatory variables (the default should be ok in most cases): - -SET (CTEST_CVS_COMMAND "hg") -SET (CTEST_CVS_CHECKOUT "${CTEST_CVS_COMMAND} clone -r 2.0 http://bitbucket.org/eigen/eigen \"${CTEST_SOURCE_DIRECTORY}\"") - -# which ctest command to use for running the dashboard -SET (CTEST_COMMAND "${EIGEN_CMAKE_DIR}ctest -D ${EIGEN_MODE}") - -# what cmake command to use for configuring this dashboard -SET (CTEST_CMAKE_COMMAND "${EIGEN_CMAKE_DIR}cmake -DEIGEN_BUILD_TESTS=on ") - -#################################################################### -# The values in this section are optional you can either -# have them or leave them commented out -#################################################################### - -# this make sure we get consistent outputs -SET($ENV{LC_MESSAGES} "en_EN") - -# should ctest wipe the binary tree before running -SET(CTEST_START_WITH_EMPTY_BINARY_DIRECTORY TRUE) -SET(CTEST_BACKUP_AND_RESTORE TRUE) - -# this is the initial cache to use for the binary tree, be careful to escape -# any quotes inside of this string if you use it -if(WIN32 AND NOT UNIX) - #message(SEND_ERROR "win32") - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"NMake Makefiles\" -DCMAKE_MAKE_PROGRAM=nmake") - SET (CTEST_INITIAL_CACHE " - MAKECOMMAND:STRING=nmake -i - CMAKE_MAKE_PROGRAM:FILEPATH=nmake - CMAKE_GENERATOR:INTERNAL=NMake Makefiles - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") -else(WIN32 AND NOT UNIX) - SET (CTEST_INITIAL_CACHE " - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") -endif(WIN32 AND NOT UNIX) - -# set any extra environment variables to use during the execution of the script here: - -if(EIGEN_CXX) - set(CTEST_ENVIRONMENT "CXX=${EIGEN_CXX}") -endif(EIGEN_CXX) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - if(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE3) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES Altivec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_ALTIVEC=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES novec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_NO_EXPLICIT_VECTORIZATION=ON") - else(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - message(FATAL_ERROR "Invalid value for EIGEN_EXPLICIT_VECTORIZATION (${EIGEN_EXPLICIT_VECTORIZATION}), must be: novec, SSE2, SSE3, Altivec") - endif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(DEFINED EIGEN_CMAKE_ARGS) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} ${EIGEN_CMAKE_ARGS}") -endif(DEFINED EIGEN_CMAKE_ARGS) diff --git a/test/runtest.sh b/test/runtest.sh deleted file mode 100644 index 2be250819..000000000 --- a/test/runtest.sh +++ /dev/null @@ -1,20 +0,0 @@ -#!/bin/bash - -black='\E[30m' -red='\E[31m' -green='\E[32m' -yellow='\E[33m' -blue='\E[34m' -magenta='\E[35m' -cyan='\E[36m' -white='\E[37m' - -if ! ./$1 > /dev/null 2> .runtest.log ; then - echo -e $red Test $1 failed: $black - echo -e $blue - cat .runtest.log - echo -e $black - exit 1 -else -echo -e $green Test $1 passed$black -fi diff --git a/test/testsuite.cmake b/test/testsuite.cmake deleted file mode 100644 index 3bec56b3f..000000000 --- a/test/testsuite.cmake +++ /dev/null @@ -1,229 +0,0 @@ - -#################################################################### -# -# Usage: -# - create a new folder, let's call it cdash -# - in that folder, do: -# ctest -S path/to/eigen/test/testsuite.cmake[,option1=value1[,option2=value2]] -# -# Options: -# - EIGEN_CXX: compiler, eg.: g++-4.2 -# default: default c++ compiler -# - EIGEN_SITE: eg, INRIA-Bdx_pc-gael, or the name of the contributor, etc. -# default: hostname -# - EIGEN_BUILD_STRING: a string which identify the system/compiler. It should be formed like that: -# --- -# with: -# = opensuse, debian, osx, windows, cygwin, freebsd, solaris, etc. -# = 11.1, XP, vista, leopard, etc. -# = i386, x86_64, ia64, powerpc, etc. -# = gcc-4.3.2, icc-11.0, MSVC-2008, etc. -# - EIGEN_EXPLICIT_VECTORIZATION: novec, SSE2, Altivec -# default: SSE2 for x86_64 systems, novec otherwise -# Its value is automatically appended to EIGEN_BUILD_STRING -# - EIGEN_CMAKE_DIR: path to cmake executable -# - EIGEN_MODE: dashboard model, can be Experimental, Nightly, or Continuous -# default: Nightly -# - EIGEN_WORK_DIR: directory used to download the source files and make the builds -# default: folder which contains this script -# - EIGEN_CMAKE_ARGS: additional arguments passed to cmake -# - EIGEN_GENERATOR_TYPE: allows to overwrite the generator type -# default: nmake (windows -# See http://www.cmake.org/cmake/help/cmake2.6docs.html#section_Generators for a complete -# list of supported generators. -# - EIGEN_NO_UPDATE: allows to submit dash boards from local repositories -# This might be interesting in case you want to submit dashboards -# including local changes. -# - CTEST_SOURCE_DIRECTORY: path to eigen's src (use a new and empty folder, not the one you are working on) -# default: /src -# - CTEST_BINARY_DIRECTORY: build directory -# default: /nightly- -# -# Here is an example running several compilers on a linux system: -# #!/bin/bash -# ARCH=`uname -m` -# SITE=`hostname` -# VERSION=opensuse-11.1 -# WORK_DIR=/home/gael/Coding/eigen/cdash -# # get the last version of the script -# wget http://bitbucket.org/eigen/eigen/raw/tip/test/testsuite.cmake -o $WORK_DIR/testsuite.cmake -# COMMON="ctest -S $WORK_DIR/testsuite.cmake,EIGEN_WORK_DIR=$WORK_DIR,EIGEN_SITE=$SITE,EIGEN_MODE=$1,EIGEN_BUILD_STRING=$OS_VERSION-$ARCH" -# $COMMON-gcc-3.4.6,EIGEN_CXX=g++-3.4 -# $COMMON-gcc-4.0.1,EIGEN_CXX=g++-4.0.1 -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=novec -# $COMMON-gcc-4.3.2,EIGEN_CXX=g++-4.3,EIGEN_EXPLICIT_VECTORIZATION=SSE2 -# $COMMON-icc-11.0,EIGEN_CXX=icpc -# -#################################################################### - -# process the arguments - -set(ARGLIST ${CTEST_SCRIPT_ARG}) -while(${ARGLIST} MATCHES ".+.*") - - # pick first - string(REGEX MATCH "([^,]*)(,.*)?" DUMMY ${ARGLIST}) - SET(TOP ${CMAKE_MATCH_1}) - - # remove first - string(REGEX MATCHALL "[^,]*,(.*)" DUMMY ${ARGLIST}) - SET(ARGLIST ${CMAKE_MATCH_1}) - - # decompose as a pair key=value - string(REGEX MATCH "([^=]*)(=.*)?" DUMMY ${TOP}) - SET(KEY ${CMAKE_MATCH_1}) - - string(REGEX MATCH "[^=]*=(.*)" DUMMY ${TOP}) - SET(VALUE ${CMAKE_MATCH_1}) - - # set the variable to the specified value - if(VALUE) - SET(${KEY} ${VALUE}) - else(VALUE) - SET(${KEY} ON) - endif(VALUE) - -endwhile(${ARGLIST} MATCHES ".+.*") - -#################################################################### -# Automatically set some user variables if they have not been defined manually -#################################################################### -cmake_minimum_required(VERSION 2.6 FATAL_ERROR) - -if(NOT EIGEN_SITE) - site_name(EIGEN_SITE) -endif(NOT EIGEN_SITE) - -if(NOT EIGEN_CMAKE_DIR) - SET(EIGEN_CMAKE_DIR "") -endif(NOT EIGEN_CMAKE_DIR) - -if(NOT EIGEN_BUILD_STRING) - - # let's try to find all information we need to make the build string ourself - - # OS - build_name(EIGEN_OS_VERSION) - - # arch - set(EIGEN_ARCH ${CMAKE_SYSTEM_PROCESSOR}) - if(WIN32) - set(EIGEN_ARCH $ENV{PROCESSOR_ARCHITECTURE}) - else(WIN32) - execute_process(COMMAND uname -m OUTPUT_VARIABLE EIGEN_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE) - endif(WIN32) - - set(EIGEN_BUILD_STRING ${EIGEN_OS_VERSION}${EIGEN_ARCH}-${EIGEN_CXX}) - -endif(NOT EIGEN_BUILD_STRING) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - set(EIGEN_BUILD_STRING ${EIGEN_BUILD_STRING}-${EIGEN_EXPLICIT_VECTORIZATION}) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(NOT EIGEN_WORK_DIR) - set(EIGEN_WORK_DIR ${CTEST_SCRIPT_DIRECTORY}) -endif(NOT EIGEN_WORK_DIR) - -if(NOT CTEST_SOURCE_DIRECTORY) - SET (CTEST_SOURCE_DIRECTORY "${EIGEN_WORK_DIR}/src") -endif(NOT CTEST_SOURCE_DIRECTORY) - -if(NOT CTEST_BINARY_DIRECTORY) - SET (CTEST_BINARY_DIRECTORY "${EIGEN_WORK_DIR}/nightly_${EIGEN_CXX}") -endif(NOT CTEST_BINARY_DIRECTORY) - -if(NOT EIGEN_MODE) - set(EIGEN_MODE Nightly) -endif(NOT EIGEN_MODE) - -## mandatory variables (the default should be ok in most cases): - -if(NOT EIGEN_NO_UPDATE) - SET (CTEST_CVS_COMMAND "hg") - SET (CTEST_CVS_CHECKOUT "${CTEST_CVS_COMMAND} clone http://bitbucket.org/eigen/eigen \"${CTEST_SOURCE_DIRECTORY}\"") - SET(CTEST_BACKUP_AND_RESTORE TRUE) # the backup is CVS related ... -endif(NOT EIGEN_NO_UPDATE) - -# which ctest command to use for running the dashboard -SET (CTEST_COMMAND "${EIGEN_CMAKE_DIR}ctest -D ${EIGEN_MODE} --no-compress-output") -if($ENV{EIGEN_CTEST_ARGS}) -SET (CTEST_COMMAND "${CTEST_COMMAND} $ENV{EIGEN_CTEST_ARGS}") -endif($ENV{EIGEN_CTEST_ARGS}) -# what cmake command to use for configuring this dashboard -SET (CTEST_CMAKE_COMMAND "${EIGEN_CMAKE_DIR}cmake -DEIGEN_LEAVE_TEST_IN_ALL_TARGET=ON") - -#################################################################### -# The values in this section are optional you can either -# have them or leave them commented out -#################################################################### - -# this make sure we get consistent outputs -SET($ENV{LC_MESSAGES} "en_EN") - -# should ctest wipe the binary tree before running -SET(CTEST_START_WITH_EMPTY_BINARY_DIRECTORY TRUE) - -# raise the warning/error limit -set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_WARNINGS "33331") -set(CTEST_CUSTOM_MAXIMUM_NUMBER_OF_ERRORS "33331") - -# this is the initial cache to use for the binary tree, be careful to escape -# any quotes inside of this string if you use it -if(WIN32 AND NOT UNIX) - #message(SEND_ERROR "win32") - if(EIGEN_GENERATOR_TYPE) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"${EIGEN_GENERATOR_TYPE}\"") - SET (CTEST_INITIAL_CACHE " - CMAKE_BUILD_TYPE:STRING=Release - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") - else(EIGEN_GENERATOR_TYPE) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -G \"NMake Makefiles\" -DCMAKE_MAKE_PROGRAM=nmake") - SET (CTEST_INITIAL_CACHE " - MAKECOMMAND:STRING=nmake /i - CMAKE_MAKE_PROGRAM:FILEPATH=nmake - CMAKE_GENERATOR:INTERNAL=NMake Makefiles - CMAKE_BUILD_TYPE:STRING=Release - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") - endif(EIGEN_GENERATOR_TYPE) -else(WIN32 AND NOT UNIX) - SET (CTEST_INITIAL_CACHE " - BUILDNAME:STRING=${EIGEN_BUILD_STRING} - SITE:STRING=${EIGEN_SITE} - ") -endif(WIN32 AND NOT UNIX) - -# set any extra environment variables to use during the execution of the script here: -# setting this variable on windows machines causes trouble ... - -if(EIGEN_CXX AND NOT WIN32) - set(CTEST_ENVIRONMENT "CXX=${EIGEN_CXX}") -endif(EIGEN_CXX AND NOT WIN32) - -if(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - if(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE3) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSSE3) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE4_1) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON -DEIGEN_TEST_SSE4_1=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE4_2) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_SSE2=ON -DEIGEN_TEST_SSE3=ON -DEIGEN_TEST_SSSE3=ON -DEIGEN_TEST_SSE4_1=ON -DEIGEN_TEST_SSE4_2=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES Altivec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_ALTIVEC=ON") - elseif(EIGEN_EXPLICIT_VECTORIZATION MATCHES novec) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} -DEIGEN_TEST_NO_EXPLICIT_VECTORIZATION=ON") - else(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) - message(FATAL_ERROR "Invalid value for EIGEN_EXPLICIT_VECTORIZATION (${EIGEN_EXPLICIT_VECTORIZATION}), must be: novec, SSE2, SSE3, Altivec") - endif(EIGEN_EXPLICIT_VECTORIZATION MATCHES SSE2) -endif(DEFINED EIGEN_EXPLICIT_VECTORIZATION) - -if(DEFINED EIGEN_CMAKE_ARGS) - set(CTEST_CMAKE_COMMAND "${CTEST_CMAKE_COMMAND} ${EIGEN_CMAKE_ARGS}") -endif(DEFINED EIGEN_CMAKE_ARGS) From ddf775363147fc7ee778b42c21b642f085193f55 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 1 Aug 2013 16:26:57 +0200 Subject: [PATCH 0071/1103] Add nvcc support for small eigenvalues decompositions and workaround lack of support for std::swap and std::numeric_limits --- Eigen/Core | 2 +- Eigen/src/Core/Assign.h | 8 ++ Eigen/src/Core/CwiseNullaryOp.h | 2 + Eigen/src/Core/DenseBase.h | 6 +- Eigen/src/Core/Matrix.h | 1 + Eigen/src/Core/MatrixBase.h | 40 +++++++--- Eigen/src/Core/NumTraits.h | 12 ++- Eigen/src/Core/PlainObjectBase.h | 11 +++ Eigen/src/Core/SelfAdjointView.h | 24 ++++++ Eigen/src/Core/Swap.h | 14 ++++ Eigen/src/Core/TriangularMatrix.h | 76 ++++++++++++++++++- Eigen/src/Core/util/Meta.h | 24 ++++++ .../src/Eigenvalues/SelfAdjointEigenSolver.h | 40 ++++++++-- Eigen/src/Geometry/OrthoMethods.h | 3 + 14 files changed, 234 insertions(+), 29 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index d15795aed..1aaac5ae5 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -30,7 +30,7 @@ #endif #if defined(__CUDA_ARCH__) - #define EIGEN_USING_STD_MATH(FUNC) + #define EIGEN_USING_STD_MATH(FUNC) using ::FUNC; #else #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC; #endif diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index b5a5a9fdb..906adcf82 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -532,6 +532,7 @@ struct assign_selector { EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); } template + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; } }; template @@ -544,6 +545,7 @@ struct assign_selector { EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); } template + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { Transpose dstTrans(dst); other.evalTo(dstTrans); return dst; } }; template @@ -556,18 +558,21 @@ struct assign_selector { template template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::operator=(const DenseBase& other) { return internal::assign_selector::run(derived(), other.derived()); } template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase::operator=(const DenseBase& other) { return internal::assign_selector::run(derived(), other.derived()); } template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const MatrixBase& other) { return internal::assign_selector::run(derived(), other.derived()); @@ -575,6 +580,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const MatrixBase& ot template template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const DenseBase& other) { return internal::assign_selector::run(derived(), other.derived()); @@ -582,6 +588,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const DenseBase template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const EigenBase& other) { return internal::assign_selector::evalTo(derived(), other.derived()); @@ -589,6 +596,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const EigenBase template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const ReturnByValue& other) { return internal::assign_selector::evalTo(derived(), other.derived()); diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index 14c78cb85..1d4ee50a8 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -746,6 +746,7 @@ namespace internal { template=16)> struct setIdentity_impl { + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& m) { return m = Derived::Identity(m.rows(), m.cols()); @@ -756,6 +757,7 @@ template struct setIdentity_impl { typedef typename Derived::Index Index; + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& m) { m.setZero(); diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 097717075..889e2eafb 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -391,7 +391,8 @@ template class DenseBase /** swaps *this with the expression \a other. * */ - template EIGEN_DEVICE_FUNC + template + EIGEN_DEVICE_FUNC void swap(const DenseBase& other, int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase) { @@ -401,7 +402,8 @@ template class DenseBase /** swaps *this with the matrix or array \a other. * */ - template EIGEN_DEVICE_FUNC + template + EIGEN_DEVICE_FUNC void swap(PlainObjectBase& other) { SwapWrapper(derived()).lazyAssign(other.derived()); diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 8c3780ca2..b80d97c8c 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -331,6 +331,7 @@ class Matrix * of same type it is enough to swap the data pointers. */ template + EIGEN_DEVICE_FUNC void swap(MatrixBase const & other) { this->_swap(other.derived()); } diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 8b8a868b7..ff4526d75 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -278,31 +278,41 @@ template class MatrixBase template struct TriangularViewReturnType { typedef TriangularView Type; }; template struct ConstTriangularViewReturnType { typedef const TriangularView Type; }; - template typename TriangularViewReturnType::Type triangularView(); - template typename ConstTriangularViewReturnType::Type triangularView() const; + template + EIGEN_DEVICE_FUNC + typename TriangularViewReturnType::Type triangularView(); + template + EIGEN_DEVICE_FUNC + typename ConstTriangularViewReturnType::Type triangularView() const; template struct SelfAdjointViewReturnType { typedef SelfAdjointView Type; }; template struct ConstSelfAdjointViewReturnType { typedef const SelfAdjointView Type; }; - template typename SelfAdjointViewReturnType::Type selfadjointView(); - template typename ConstSelfAdjointViewReturnType::Type selfadjointView() const; + template + EIGEN_DEVICE_FUNC + typename SelfAdjointViewReturnType::Type selfadjointView(); + template + EIGEN_DEVICE_FUNC + typename ConstSelfAdjointViewReturnType::Type selfadjointView() const; const SparseView sparseView(const Scalar& m_reference = Scalar(0), const typename NumTraits::Real& m_epsilon = NumTraits::dummy_precision()) const; - static const IdentityReturnType Identity(); - static const IdentityReturnType Identity(Index rows, Index cols); - static const BasisReturnType Unit(Index size, Index i); - static const BasisReturnType Unit(Index i); - static const BasisReturnType UnitX(); - static const BasisReturnType UnitY(); - static const BasisReturnType UnitZ(); - static const BasisReturnType UnitW(); + EIGEN_DEVICE_FUNC static const IdentityReturnType Identity(); + EIGEN_DEVICE_FUNC static const IdentityReturnType Identity(Index rows, Index cols); + EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index size, Index i); + EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index i); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitX(); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitY(); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitZ(); + EIGEN_DEVICE_FUNC static const BasisReturnType UnitW(); EIGEN_DEVICE_FUNC const DiagonalWrapper asDiagonal() const; const PermutationWrapper asPermutation() const; + EIGEN_DEVICE_FUNC Derived& setIdentity(); + EIGEN_DEVICE_FUNC Derived& setIdentity(Index rows, Index cols); bool isIdentity(const RealScalar& prec = NumTraits::dummy_precision()) const; @@ -430,11 +440,17 @@ template class MatrixBase }; #endif // EIGEN_PARSED_BY_DOXYGEN template + EIGEN_DEVICE_FUNC typename cross_product_return_type::type cross(const MatrixBase& other) const; + template + EIGEN_DEVICE_FUNC PlainObject cross3(const MatrixBase& other) const; + + EIGEN_DEVICE_FUNC PlainObject unitOrthogonal(void) const; + Matrix eulerAngles(Index a0, Index a1, Index a2) const; #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h index eb5306bf2..2b6633c9c 100644 --- a/Eigen/src/Core/NumTraits.h +++ b/Eigen/src/Core/NumTraits.h @@ -68,16 +68,22 @@ template struct GenericNumTraits >::type NonInteger; typedef T Nested; - static inline Real epsilon() { return std::numeric_limits::epsilon(); } + EIGEN_DEVICE_FUNC + static inline Real epsilon() + { + #if defined(__CUDA_ARCH__) + return internal::device::numeric_limits::epsilon(); + #else + return std::numeric_limits::epsilon(); + #endif + } EIGEN_DEVICE_FUNC static inline Real dummy_precision() { // make sure to override this for floating-point types return Real(0); } - EIGEN_DEVICE_FUNC static inline T highest() { return (std::numeric_limits::max)(); } - EIGEN_DEVICE_FUNC static inline T lowest() { return IsInteger ? (std::numeric_limits::min)() : (-(std::numeric_limits::max)()); } #ifdef EIGEN2_SUPPORT diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 39d422f78..618e00e02 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -128,7 +128,9 @@ class PlainObjectBase : public internal::dense_xpr_base::type enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits::Flags & AlignedBit) != 0 }; EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) + EIGEN_DEVICE_FUNC Base& base() { return *static_cast(this); } + EIGEN_DEVICE_FUNC const Base& base() const { return *static_cast(this); } EIGEN_DEVICE_FUNC @@ -351,6 +353,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * Matrices are resized relative to the top-left element. In case values need to be * appended to the matrix they will be uninitialized. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, Index nbCols) { internal::conservative_resize_like_impl::run(*this, nbRows, nbCols); @@ -363,6 +366,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * In case the matrix is growing, new rows will be uninitialized. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void conservativeResize(Index nbRows, NoChange_t) { // Note: see the comment in conservativeResize(Index,Index) @@ -376,6 +380,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * In case the matrix is growing, new columns will be uninitialized. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index nbCols) { // Note: see the comment in conservativeResize(Index,Index) @@ -390,6 +395,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * When values are appended, they will be uninitialized. */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void conservativeResize(Index size) { internal::conservative_resize_like_impl::run(*this, size); @@ -405,6 +411,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * appended to the matrix they will copied from \c other. */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void conservativeResizeLike(const DenseBase& other) { internal::conservative_resize_like_impl::run(*this, other); @@ -647,6 +654,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init2(Index nbRows, Index nbCols, typename internal::enable_if::type* = 0) { EIGEN_STATIC_ASSERT(bool(NumTraits::IsInteger) && @@ -670,6 +678,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type * data pointers. */ template + EIGEN_DEVICE_FUNC void _swap(DenseBase const & other) { enum { SwapPointers = internal::is_same::value && Base::SizeAtCompileTime==Dynamic }; @@ -790,6 +799,7 @@ struct conservative_resize_like_impl template struct matrix_swap_impl { + EIGEN_DEVICE_FUNC static inline void run(MatrixTypeA& a, MatrixTypeB& b) { a.base().swap(b); @@ -799,6 +809,7 @@ struct matrix_swap_impl template struct matrix_swap_impl { + EIGEN_DEVICE_FUNC static inline void run(MatrixTypeA& a, MatrixTypeB& b) { static_cast(a).m_storage.swap(static_cast(b).m_storage); diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 6fa7cd15e..8231e3f5c 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -69,17 +69,23 @@ template class SelfAdjointView }; typedef typename MatrixType::PlainObject PlainObject; + EIGEN_DEVICE_FUNC inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix) {} + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_matrix.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.innerStride(); } /** \sa MatrixBase::coeff() * \warning the coordinates must fit into the referenced triangular part */ + EIGEN_DEVICE_FUNC inline Scalar coeff(Index row, Index col) const { Base::check_coordinates_internal(row, col); @@ -89,6 +95,7 @@ template class SelfAdjointView /** \sa MatrixBase::coeffRef() * \warning the coordinates must fit into the referenced triangular part */ + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { Base::check_coordinates_internal(row, col); @@ -96,13 +103,17 @@ template class SelfAdjointView } /** \internal */ + EIGEN_DEVICE_FUNC const MatrixTypeNestedCleaned& _expression() const { return m_matrix; } + EIGEN_DEVICE_FUNC const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; } + EIGEN_DEVICE_FUNC MatrixTypeNestedCleaned& nestedExpression() { return *const_cast(&m_matrix); } /** Efficient self-adjoint matrix times vector/matrix product */ template + EIGEN_DEVICE_FUNC SelfadjointProductMatrix operator*(const MatrixBase& rhs) const { @@ -113,6 +124,7 @@ template class SelfAdjointView /** Efficient vector/matrix times self-adjoint matrix product */ template friend + EIGEN_DEVICE_FUNC SelfadjointProductMatrix operator*(const MatrixBase& lhs, const SelfAdjointView& rhs) { @@ -132,6 +144,7 @@ template class SelfAdjointView * \sa rankUpdate(const MatrixBase&, Scalar) */ template + EIGEN_DEVICE_FUNC SelfAdjointView& rankUpdate(const MatrixBase& u, const MatrixBase& v, const Scalar& alpha = Scalar(1)); /** Perform a symmetric rank K update of the selfadjoint matrix \c *this: @@ -145,6 +158,7 @@ template class SelfAdjointView * \sa rankUpdate(const MatrixBase&, const MatrixBase&, Scalar) */ template + EIGEN_DEVICE_FUNC SelfAdjointView& rankUpdate(const MatrixBase& u, const Scalar& alpha = Scalar(1)); /////////// Cholesky module /////////// @@ -159,11 +173,14 @@ template class SelfAdjointView /** Return type of eigenvalues() */ typedef Matrix::ColsAtCompileTime, 1> EigenvaluesReturnType; + EIGEN_DEVICE_FUNC EigenvaluesReturnType eigenvalues() const; + EIGEN_DEVICE_FUNC RealScalar operatorNorm() const; #ifdef EIGEN2_SUPPORT template + EIGEN_DEVICE_FUNC SelfAdjointView& operator=(const MatrixBase& other) { enum { @@ -174,6 +191,7 @@ template class SelfAdjointView return *this; } template + EIGEN_DEVICE_FUNC SelfAdjointView& operator=(const TriangularView& other) { enum { @@ -209,6 +227,7 @@ struct triangular_assignment_selector::run(dst, src); @@ -223,6 +242,7 @@ struct triangular_assignment_selector struct triangular_assignment_selector { + EIGEN_DEVICE_FUNC static inline void run(Derived1 &, const Derived2 &) {} }; @@ -234,6 +254,7 @@ struct triangular_assignment_selector::run(dst, src); @@ -248,6 +269,7 @@ struct triangular_assignment_selector struct triangular_assignment_selector { + EIGEN_DEVICE_FUNC static inline void run(Derived1 &, const Derived2 &) {} }; @@ -255,6 +277,7 @@ template struct triangular_assignment_selector { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { for(Index j = 0; j < dst.cols(); ++j) @@ -272,6 +295,7 @@ struct triangular_assignment_selector struct triangular_assignment_selector { + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { typedef typename Derived1::Index Index; diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index bf58bd599..d602fba65 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -33,11 +33,16 @@ template class SwapWrapper EIGEN_DENSE_PUBLIC_INTERFACE(SwapWrapper) typedef typename internal::packet_traits::type Packet; + EIGEN_DEVICE_FUNC inline SwapWrapper(ExpressionType& xpr) : m_expression(xpr) {} + EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); } typedef typename internal::conditional< @@ -46,30 +51,37 @@ template class SwapWrapper const Scalar >::type ScalarWithConstIfNotLvalue; + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_expression.data(); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index rowId, Index colId) { return m_expression.const_cast_derived().coeffRef(rowId, colId); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { return m_expression.const_cast_derived().coeffRef(index); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index rowId, Index colId) const { return m_expression.coeffRef(rowId, colId); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) const { return m_expression.coeffRef(index); } template + EIGEN_DEVICE_FUNC void copyCoeff(Index rowId, Index colId, const DenseBase& other) { OtherDerived& _other = other.const_cast_derived(); @@ -81,6 +93,7 @@ template class SwapWrapper } template + EIGEN_DEVICE_FUNC void copyCoeff(Index index, const DenseBase& other) { OtherDerived& _other = other.const_cast_derived(); @@ -115,6 +128,7 @@ template class SwapWrapper _other.template writePacket(index, tmp); } + EIGEN_DEVICE_FUNC ExpressionType& expression() const { return m_expression; } protected: diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index fba07365f..1d6e34650 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -44,29 +44,39 @@ template class TriangularBase : public EigenBase typedef typename internal::traits::DenseMatrixType DenseMatrixType; typedef DenseMatrixType DenseType; + EIGEN_DEVICE_FUNC inline TriangularBase() { eigen_assert(!((Mode&UnitDiag) && (Mode&ZeroDiag))); } + EIGEN_DEVICE_FUNC inline Index rows() const { return derived().rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return derived().cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().innerStride(); } + EIGEN_DEVICE_FUNC inline Scalar coeff(Index row, Index col) const { return derived().coeff(row,col); } + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { return derived().coeffRef(row,col); } /** \see MatrixBase::copyCoeff(row,col) */ template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, Other& other) { derived().coeffRef(row, col) = other.coeff(row, col); } + EIGEN_DEVICE_FUNC inline Scalar operator()(Index row, Index col) const { check_coordinates(row, col); return coeff(row,col); } + EIGEN_DEVICE_FUNC inline Scalar& operator()(Index row, Index col) { check_coordinates(row, col); @@ -74,15 +84,20 @@ template class TriangularBase : public EigenBase } #ifndef EIGEN_PARSED_BY_DOXYGEN + EIGEN_DEVICE_FUNC inline const Derived& derived() const { return *static_cast(this); } + EIGEN_DEVICE_FUNC inline Derived& derived() { return *static_cast(this); } #endif // not EIGEN_PARSED_BY_DOXYGEN template + EIGEN_DEVICE_FUNC void evalTo(MatrixBase &other) const; template + EIGEN_DEVICE_FUNC void evalToLazy(MatrixBase &other) const; + EIGEN_DEVICE_FUNC DenseMatrixType toDenseMatrix() const { DenseMatrixType res(rows(), cols()); @@ -189,36 +204,52 @@ template class TriangularView | (Mode & (ZeroDiag)) }; + EIGEN_DEVICE_FUNC inline TriangularView(const MatrixType& matrix) : m_matrix(matrix) {} + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_matrix.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.innerStride(); } - /** \sa MatrixBase::operator+=() */ - template TriangularView& operator+=(const DenseBase& other) { return *this = m_matrix + other.derived(); } + /** \sa MatrixBase::operator+=() */ + template + EIGEN_DEVICE_FUNC + TriangularView& operator+=(const DenseBase& other) { return *this = m_matrix + other.derived(); } /** \sa MatrixBase::operator-=() */ - template TriangularView& operator-=(const DenseBase& other) { return *this = m_matrix - other.derived(); } + template + EIGEN_DEVICE_FUNC + TriangularView& operator-=(const DenseBase& other) { return *this = m_matrix - other.derived(); } /** \sa MatrixBase::operator*=() */ + EIGEN_DEVICE_FUNC TriangularView& operator*=(const typename internal::traits::Scalar& other) { return *this = m_matrix * other; } /** \sa MatrixBase::operator/=() */ + EIGEN_DEVICE_FUNC TriangularView& operator/=(const typename internal::traits::Scalar& other) { return *this = m_matrix / other; } /** \sa MatrixBase::fill() */ + EIGEN_DEVICE_FUNC void fill(const Scalar& value) { setConstant(value); } /** \sa MatrixBase::setConstant() */ + EIGEN_DEVICE_FUNC TriangularView& setConstant(const Scalar& value) { return *this = MatrixType::Constant(rows(), cols(), value); } /** \sa MatrixBase::setZero() */ + EIGEN_DEVICE_FUNC TriangularView& setZero() { return setConstant(Scalar(0)); } /** \sa MatrixBase::setOnes() */ + EIGEN_DEVICE_FUNC TriangularView& setOnes() { return setConstant(Scalar(1)); } /** \sa MatrixBase::coeff() * \warning the coordinates must fit into the referenced triangular part */ + EIGEN_DEVICE_FUNC inline Scalar coeff(Index row, Index col) const { Base::check_coordinates_internal(row, col); @@ -228,49 +259,62 @@ template class TriangularView /** \sa MatrixBase::coeffRef() * \warning the coordinates must fit into the referenced triangular part */ + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { Base::check_coordinates_internal(row, col); return m_matrix.const_cast_derived().coeffRef(row, col); } + EIGEN_DEVICE_FUNC const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; } + EIGEN_DEVICE_FUNC MatrixTypeNestedCleaned& nestedExpression() { return *const_cast(&m_matrix); } /** Assigns a triangular matrix to a triangular part of a dense matrix */ template + EIGEN_DEVICE_FUNC TriangularView& operator=(const TriangularBase& other); template + EIGEN_DEVICE_FUNC TriangularView& operator=(const MatrixBase& other); + EIGEN_DEVICE_FUNC TriangularView& operator=(const TriangularView& other) { return *this = other.nestedExpression(); } template + EIGEN_DEVICE_FUNC void lazyAssign(const TriangularBase& other); template + EIGEN_DEVICE_FUNC void lazyAssign(const MatrixBase& other); /** \sa MatrixBase::conjugate() */ + EIGEN_DEVICE_FUNC inline TriangularView conjugate() { return m_matrix.conjugate(); } /** \sa MatrixBase::conjugate() const */ + EIGEN_DEVICE_FUNC inline const TriangularView conjugate() const { return m_matrix.conjugate(); } /** \sa MatrixBase::adjoint() const */ + EIGEN_DEVICE_FUNC inline const TriangularView adjoint() const { return m_matrix.adjoint(); } /** \sa MatrixBase::transpose() */ + EIGEN_DEVICE_FUNC inline TriangularView,TransposeMode> transpose() { EIGEN_STATIC_ASSERT_LVALUE(MatrixType) return m_matrix.const_cast_derived().transpose(); } /** \sa MatrixBase::transpose() const */ + EIGEN_DEVICE_FUNC inline const TriangularView,TransposeMode> transpose() const { return m_matrix.transpose(); @@ -278,6 +322,7 @@ template class TriangularView /** Efficient triangular matrix times vector/matrix product */ template + EIGEN_DEVICE_FUNC TriangularProduct operator*(const MatrixBase& rhs) const { @@ -288,6 +333,7 @@ template class TriangularView /** Efficient vector/matrix times triangular matrix product */ template friend + EIGEN_DEVICE_FUNC TriangularProduct operator*(const MatrixBase& lhs, const TriangularView& rhs) { @@ -326,26 +372,32 @@ template class TriangularView #endif // EIGEN2_SUPPORT template + EIGEN_DEVICE_FUNC inline const internal::triangular_solve_retval solve(const MatrixBase& other) const; template + EIGEN_DEVICE_FUNC void solveInPlace(const MatrixBase& other) const; template + EIGEN_DEVICE_FUNC inline const internal::triangular_solve_retval solve(const MatrixBase& other) const { return solve(other); } template + EIGEN_DEVICE_FUNC void solveInPlace(const MatrixBase& other) const { return solveInPlace(other); } + EIGEN_DEVICE_FUNC const SelfAdjointView selfadjointView() const { EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR); return SelfAdjointView(m_matrix); } + EIGEN_DEVICE_FUNC SelfAdjointView selfadjointView() { EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR); @@ -353,18 +405,21 @@ template class TriangularView } template + EIGEN_DEVICE_FUNC void swap(TriangularBase const & other) { TriangularView,Mode>(const_cast(m_matrix)).lazyAssign(other.derived()); } template + EIGEN_DEVICE_FUNC void swap(MatrixBase const & other) { SwapWrapper swaper(const_cast(m_matrix)); TriangularView,Mode>(swaper).lazyAssign(other.derived()); } + EIGEN_DEVICE_FUNC Scalar determinant() const { if (Mode & UnitDiag) @@ -377,6 +432,7 @@ template class TriangularView // TODO simplify the following: template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& operator=(const ProductBase& other) { setZero(); @@ -384,12 +440,14 @@ template class TriangularView } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& operator+=(const ProductBase& other) { return assignProduct(other,1); } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& operator-=(const ProductBase& other) { return assignProduct(other,-1); @@ -397,6 +455,7 @@ template class TriangularView template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& operator=(const ScaledProduct& other) { setZero(); @@ -404,12 +463,14 @@ template class TriangularView } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& operator+=(const ScaledProduct& other) { return assignProduct(other,other.alpha()); } template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& operator-=(const ScaledProduct& other) { return assignProduct(other,-other.alpha()); @@ -418,6 +479,7 @@ template class TriangularView protected: template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& assignProduct(const ProductBase& prod, const Scalar& alpha); MatrixTypeNested m_matrix; @@ -439,6 +501,7 @@ struct triangular_assignment_selector typedef typename Derived1::Scalar Scalar; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { triangular_assignment_selector::run(dst, src); @@ -467,6 +530,7 @@ struct triangular_assignment_selector template struct triangular_assignment_selector { + EIGEN_DEVICE_FUNC static inline void run(Derived1 &, const Derived2 &) {} }; @@ -475,6 +539,7 @@ struct triangular_assignment_selector struct triangular_assignment_selector { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { for(Index j = 0; j < dst.cols(); ++j) @@ -512,6 +578,7 @@ struct triangular_assignment_selector struct triangular_assignment_selector { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { for(Index j = 0; j < dst.cols(); ++j) @@ -548,6 +616,7 @@ template struct triangular_assignment_selector { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { for(Index j = 0; j < dst.cols(); ++j) @@ -568,6 +637,7 @@ template struct triangular_assignment_selector { typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC static inline void run(Derived1 &dst, const Derived2 &src) { for(Index j = 0; j < dst.cols(); ++j) diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h index 71d587108..aea168b46 100644 --- a/Eigen/src/Core/util/Meta.h +++ b/Eigen/src/Core/util/Meta.h @@ -88,7 +88,31 @@ template struct enable_if; template struct enable_if { typedef T type; }; +#if defined(__CUDA_ARCH__) +template EIGEN_DEVICE_FUNC void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; } +namespace device { +template struct numeric_limits +{ + EIGEN_DEVICE_FUNC + static T epsilon() { return 0; } +}; +template<> struct numeric_limits +{ + EIGEN_DEVICE_FUNC + static float epsilon() { return __FLT_EPSILON__; } +}; +template<> struct numeric_limits +{ + EIGEN_DEVICE_FUNC + static double epsilon() { return __DBL_EPSILON__; } +}; + +} + +#else +template EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); } +#endif /** \internal * A base class do disable default copy ctor and copy assignement operator. diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index 3993046a8..23a334b5c 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -109,6 +109,7 @@ template class SelfAdjointEigenSolver * Example: \include SelfAdjointEigenSolver_SelfAdjointEigenSolver.cpp * Output: \verbinclude SelfAdjointEigenSolver_SelfAdjointEigenSolver.out */ + EIGEN_DEVICE_FUNC SelfAdjointEigenSolver() : m_eivec(), m_eivalues(), @@ -128,6 +129,7 @@ template class SelfAdjointEigenSolver * * \sa compute() for an example */ + EIGEN_DEVICE_FUNC SelfAdjointEigenSolver(Index size) : m_eivec(size, size), m_eivalues(size), @@ -150,6 +152,7 @@ template class SelfAdjointEigenSolver * * \sa compute(const MatrixType&, int) */ + EIGEN_DEVICE_FUNC SelfAdjointEigenSolver(const MatrixType& matrix, int options = ComputeEigenvectors) : m_eivec(matrix.rows(), matrix.cols()), m_eivalues(matrix.cols()), @@ -189,6 +192,7 @@ template class SelfAdjointEigenSolver * * \sa SelfAdjointEigenSolver(const MatrixType&, int) */ + EIGEN_DEVICE_FUNC SelfAdjointEigenSolver& compute(const MatrixType& matrix, int options = ComputeEigenvectors); /** \brief Computes eigendecomposition of given matrix using a direct algorithm @@ -205,6 +209,7 @@ template class SelfAdjointEigenSolver * * \sa compute(const MatrixType&, int options) */ + EIGEN_DEVICE_FUNC SelfAdjointEigenSolver& computeDirect(const MatrixType& matrix, int options = ComputeEigenvectors); /** \brief Returns the eigenvectors of given matrix. @@ -225,6 +230,7 @@ template class SelfAdjointEigenSolver * * \sa eigenvalues() */ + EIGEN_DEVICE_FUNC const MatrixType& eigenvectors() const { eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); @@ -247,6 +253,7 @@ template class SelfAdjointEigenSolver * * \sa eigenvectors(), MatrixBase::eigenvalues() */ + EIGEN_DEVICE_FUNC const RealVectorType& eigenvalues() const { eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); @@ -271,6 +278,7 @@ template class SelfAdjointEigenSolver * \sa operatorInverseSqrt(), * \ref MatrixFunctions_Module "MatrixFunctions Module" */ + EIGEN_DEVICE_FUNC MatrixType operatorSqrt() const { eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); @@ -296,6 +304,7 @@ template class SelfAdjointEigenSolver * \sa operatorSqrt(), MatrixBase::inverse(), * \ref MatrixFunctions_Module "MatrixFunctions Module" */ + EIGEN_DEVICE_FUNC MatrixType operatorInverseSqrt() const { eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); @@ -307,6 +316,7 @@ template class SelfAdjointEigenSolver * * \returns \c Success if computation was succesful, \c NoConvergence otherwise. */ + EIGEN_DEVICE_FUNC ComputationInfo info() const { eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized."); @@ -321,6 +331,7 @@ template class SelfAdjointEigenSolver static const int m_maxIterations = 30; #ifdef EIGEN2_SUPPORT + EIGEN_DEVICE_FUNC SelfAdjointEigenSolver(const MatrixType& matrix, bool computeEigenvectors) : m_eivec(matrix.rows(), matrix.cols()), m_eivalues(matrix.cols()), @@ -330,6 +341,7 @@ template class SelfAdjointEigenSolver compute(matrix, computeEigenvectors); } + EIGEN_DEVICE_FUNC SelfAdjointEigenSolver(const MatrixType& matA, const MatrixType& matB, bool computeEigenvectors = true) : m_eivec(matA.cols(), matA.cols()), m_eivalues(matA.cols()), @@ -339,11 +351,13 @@ template class SelfAdjointEigenSolver static_cast*>(this)->compute(matA, matB, computeEigenvectors ? ComputeEigenvectors : EigenvaluesOnly); } + EIGEN_DEVICE_FUNC void compute(const MatrixType& matrix, bool computeEigenvectors) { compute(matrix, computeEigenvectors ? ComputeEigenvectors : EigenvaluesOnly); } + EIGEN_DEVICE_FUNC void compute(const MatrixType& matA, const MatrixType& matB, bool computeEigenvectors = true) { compute(matA, matB, computeEigenvectors ? ComputeEigenvectors : EigenvaluesOnly); @@ -377,10 +391,12 @@ template class SelfAdjointEigenSolver */ namespace internal { template +EIGEN_DEVICE_FUNC static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n); } template +EIGEN_DEVICE_FUNC SelfAdjointEigenSolver& SelfAdjointEigenSolver ::compute(const MatrixType& matrix, int options) { @@ -481,6 +497,7 @@ namespace internal { template struct direct_selfadjoint_eigenvalues { + EIGEN_DEVICE_FUNC static inline void run(SolverType& eig, const typename SolverType::MatrixType& A, int options) { eig.compute(A,options); } }; @@ -491,12 +508,13 @@ template struct direct_selfadjoint_eigenvalues struct direct_selfadjoint_eigenvalues= roots(1)) - std::swap(roots(0),roots(1)); + internal::swap(roots(0),roots(1)); if (roots(1) >= roots(2)) { - std::swap(roots(1),roots(2)); + internal::swap(roots(1),roots(2)); if (roots(0) >= roots(1)) - std::swap(roots(0),roots(1)); + internal::swap(roots(0),roots(1)); } } + EIGEN_DEVICE_FUNC static inline void run(SolverType& solver, const MatrixType& mat, int options) { using std::sqrt; @@ -660,12 +679,14 @@ template struct direct_selfadjoint_eigenvalues struct direct_selfadjoint_eigenvalues +template +struct direct_selfadjoint_eigenvalues { typedef typename SolverType::MatrixType MatrixType; typedef typename SolverType::RealVectorType VectorType; typedef typename SolverType::Scalar Scalar; + EIGEN_DEVICE_FUNC static inline void computeRoots(const MatrixType& m, VectorType& roots) { using std::sqrt; @@ -675,6 +696,7 @@ template struct direct_selfadjoint_eigenvalues struct direct_selfadjoint_eigenvalues +EIGEN_DEVICE_FUNC SelfAdjointEigenSolver& SelfAdjointEigenSolver ::computeDirect(const MatrixType& matrix, int options) { @@ -737,6 +760,7 @@ SelfAdjointEigenSolver& SelfAdjointEigenSolver namespace internal { template +EIGEN_DEVICE_FUNC static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n) { using std::abs; diff --git a/Eigen/src/Geometry/OrthoMethods.h b/Eigen/src/Geometry/OrthoMethods.h index 556bc8160..26be3ee5b 100644 --- a/Eigen/src/Geometry/OrthoMethods.h +++ b/Eigen/src/Geometry/OrthoMethods.h @@ -132,6 +132,7 @@ struct unitOrthogonal_selector typedef typename NumTraits::Real RealScalar; typedef typename Derived::Index Index; typedef Matrix Vector2; + EIGEN_DEVICE_FUNC static inline VectorType run(const Derived& src) { VectorType perp = VectorType::Zero(src.size()); @@ -154,6 +155,7 @@ struct unitOrthogonal_selector typedef typename plain_matrix_type::type VectorType; typedef typename traits::Scalar Scalar; typedef typename NumTraits::Real RealScalar; + EIGEN_DEVICE_FUNC static inline VectorType run(const Derived& src) { VectorType perp; @@ -192,6 +194,7 @@ template struct unitOrthogonal_selector { typedef typename plain_matrix_type::type VectorType; + EIGEN_DEVICE_FUNC static inline VectorType run(const Derived& src) { return VectorType(-numext::conj(src.y()), numext::conj(src.x())).normalized(); } }; From cf884a9815c51c27f9172465afb32f3d1d899953 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Thu, 1 Aug 2013 16:38:05 +0200 Subject: [PATCH 0072/1103] Added build name support for VC11 and its service packs. --- cmake/EigenDetermineVSServicePack.cmake | 16 ++++++++++++++-- cmake/EigenTesting.cmake | 18 +++++++++--------- 2 files changed, 23 insertions(+), 11 deletions(-) diff --git a/cmake/EigenDetermineVSServicePack.cmake b/cmake/EigenDetermineVSServicePack.cmake index 8e5546a85..a1a7348f6 100644 --- a/cmake/EigenDetermineVSServicePack.cmake +++ b/cmake/EigenDetermineVSServicePack.cmake @@ -4,7 +4,6 @@ include(CMakeDetermineVSServicePack) # _DetermineVSServicePack_FastCheckVersionWithCompiler which lead to errors on some systems. function(EigenDetermineVSServicePack _pack) if(NOT DETERMINED_VS_SERVICE_PACK OR NOT ${_pack}) - if(NOT DETERMINED_VS_SERVICE_PACK) _DetermineVSServicePack_CheckVersionWithTryCompile(DETERMINED_VS_SERVICE_PACK _cl_version) if(NOT DETERMINED_VS_SERVICE_PACK) @@ -13,10 +12,23 @@ function(EigenDetermineVSServicePack _pack) endif() if(DETERMINED_VS_SERVICE_PACK) - if(_cl_version) # Call helper function to determine VS version _DetermineVSServicePackFromCompiler(_sp "${_cl_version}") + + # temporary fix, until CMake catches up + if (NOT _sp) + if(${_cl_version} VERSION_EQUAL "17.00.50727.1") + set(_sp "vc110") + elseif(${_cl_version} VERSION_EQUAL "17.00.51106.1") + set(_sp "vc110sp1") + elseif(${_cl_version} VERSION_EQUAL "17.00.60315.1") + set(_sp "vc110sp2") + elseif(${_cl_version} VERSION_EQUAL "17.00.60610.1") + set(_sp "vc110sp3") + endif() + endif() + if(_sp) set(${_pack} ${_sp} CACHE INTERNAL "The Visual Studio Release with Service Pack") diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index ed683a4ef..73a62fd73 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -306,17 +306,17 @@ macro(ei_set_sitename) endmacro(ei_set_sitename) macro(ei_get_compilerver VAR) - if(MSVC) - # on windows system, we use a modified CMake script - include(EigenDetermineVSServicePack) - EigenDetermineVSServicePack( my_service_pack ) + if(MSVC) + # on windows system, we use a modified CMake script + include(EigenDetermineVSServicePack) + EigenDetermineVSServicePack( my_service_pack ) - if( my_service_pack ) - set(${VAR} ${my_service_pack}) + if( my_service_pack ) + set(${VAR} ${my_service_pack}) + else() + set(${VAR} "na") + endif() else() - set(${VAR} "na") - endif() - else() # on all other system we rely on ${CMAKE_CXX_COMPILER} # supporting a "--version" flag From e90229a429c74b302368e788a36d8bbef4d1e617 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 2 Aug 2013 00:36:06 +0200 Subject: [PATCH 0073/1103] reduce cancellation probablity --- test/product_trmm.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/test/product_trmm.cpp b/test/product_trmm.cpp index 506a1aeb9..d715b9a36 100644 --- a/test/product_trmm.cpp +++ b/test/product_trmm.cpp @@ -51,6 +51,7 @@ void trmm(int rows=internal::random(1,EIGEN_TEST_MAX_SIZE), ge_xs_save = ge_xs; VERIFY_IS_APPROX( (ge_xs_save + s1*triTr.conjugate() * (s2*ge_left.adjoint())).eval(), ge_xs.noalias() += (s1*mat.adjoint()).template triangularView() * (s2*ge_left.adjoint()) ); + ge_sx.setRandom(); ge_sx_save = ge_sx; VERIFY_IS_APPROX( ge_sx_save - (ge_right.adjoint() * (-s1 * triTr).conjugate()).eval(), ge_sx.noalias() -= (ge_right.adjoint() * (-s1 * mat).adjoint().template triangularView()).eval()); From 8ea7413a64aceb888af7106e42f6b8b3c48b3f61 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 2 Aug 2013 11:05:00 +0200 Subject: [PATCH 0074/1103] Fix compilation and warning of PARDISO --- Eigen/src/PardisoSupport/PardisoSupport.h | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/Eigen/src/PardisoSupport/PardisoSupport.h b/Eigen/src/PardisoSupport/PardisoSupport.h index 1c48f0df7..e459ea96e 100644 --- a/Eigen/src/PardisoSupport/PardisoSupport.h +++ b/Eigen/src/PardisoSupport/PardisoSupport.h @@ -219,7 +219,7 @@ class PardisoImpl void pardisoInit(int type) { m_type = type; - bool symmetric = abs(m_type) < 10; + bool symmetric = std::abs(m_type) < 10; m_iparm[0] = 1; // No solver default m_iparm[1] = 3; // use Metis for the ordering m_iparm[2] = 1; // Numbers of processors, value of OMP_NUM_THREADS @@ -278,7 +278,7 @@ class PardisoImpl Index m_size; private: - PardisoImpl(PardisoImpl &) {} + PardisoImpl(const PardisoImpl&); }; template @@ -436,7 +436,7 @@ class PardisoLU : public PardisoImpl< PardisoLU > } private: - PardisoLU(PardisoLU& ) {} + PardisoLU(const PardisoLU&); }; /** \ingroup PardisoSupport_Module @@ -495,7 +495,7 @@ class PardisoLLT : public PardisoImpl< PardisoLLT > } private: - PardisoLLT(PardisoLLT& ) {} + PardisoLLT(const PardisoLLT&); }; /** \ingroup PardisoSupport_Module @@ -554,7 +554,7 @@ class PardisoLDLT : public PardisoImpl< PardisoLDLT > } private: - PardisoLDLT(PardisoLDLT& ) {} + PardisoLDLT(const PardisoLDLT&); }; namespace internal { From e3058dd88ba93b8848e2e0fa84f8d1f510c985bc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 2 Aug 2013 11:09:02 +0200 Subject: [PATCH 0075/1103] Make Pardiso solvers non copyabe --- Eigen/src/PardisoSupport/PardisoSupport.h | 13 +------------ 1 file changed, 1 insertion(+), 12 deletions(-) diff --git a/Eigen/src/PardisoSupport/PardisoSupport.h b/Eigen/src/PardisoSupport/PardisoSupport.h index e459ea96e..b6571069e 100644 --- a/Eigen/src/PardisoSupport/PardisoSupport.h +++ b/Eigen/src/PardisoSupport/PardisoSupport.h @@ -96,7 +96,7 @@ namespace internal } template -class PardisoImpl +class PardisoImpl : internal::noncopyable { typedef internal::pardiso_traits Traits; public: @@ -277,8 +277,6 @@ class PardisoImpl mutable IntColVectorType m_perm; Index m_size; - private: - PardisoImpl(const PardisoImpl&); }; template @@ -434,9 +432,6 @@ class PardisoLU : public PardisoImpl< PardisoLU > { m_matrix = matrix; } - - private: - PardisoLU(const PardisoLU&); }; /** \ingroup PardisoSupport_Module @@ -493,9 +488,6 @@ class PardisoLLT : public PardisoImpl< PardisoLLT > m_matrix.resize(matrix.rows(), matrix.cols()); m_matrix.template selfadjointView() = matrix.template selfadjointView().twistedBy(p_null); } - - private: - PardisoLLT(const PardisoLLT&); }; /** \ingroup PardisoSupport_Module @@ -552,9 +544,6 @@ class PardisoLDLT : public PardisoImpl< PardisoLDLT > m_matrix.resize(matrix.rows(), matrix.cols()); m_matrix.template selfadjointView() = matrix.template selfadjointView().twistedBy(p_null); } - - private: - PardisoLDLT(const PardisoLDLT&); }; namespace internal { From b72a686830b95afd8941d3c4020eb936628b6398 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 2 Aug 2013 11:11:21 +0200 Subject: [PATCH 0076/1103] Fix bug #635: add isCompressed to MappedSparseMatrix for compatibility --- Eigen/src/SparseCore/MappedSparseMatrix.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/SparseCore/MappedSparseMatrix.h b/Eigen/src/SparseCore/MappedSparseMatrix.h index 93cd4832d..ab1a266a9 100644 --- a/Eigen/src/SparseCore/MappedSparseMatrix.h +++ b/Eigen/src/SparseCore/MappedSparseMatrix.h @@ -50,6 +50,8 @@ class MappedSparseMatrix inline Index cols() const { return IsRowMajor ? m_innerSize : m_outerSize; } inline Index innerSize() const { return m_innerSize; } inline Index outerSize() const { return m_outerSize; } + + bool isCompressed() const { return true; } //---------------------------------------- // direct access interface From 7c99b38b7c34c5a031e347b81277e67ab9cfdd93 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Fri, 2 Aug 2013 19:59:43 +0200 Subject: [PATCH 0077/1103] Added move support for Matrix and Array. Added EIGEN_HAVE_RVALUE_REFERENCES define. Added move unit tests. Removed superfluous 'inline' declarations in DenseStorage. --- Eigen/src/Core/Array.h | 15 ++ Eigen/src/Core/DenseStorage.h | 226 ++++++++++++++++++++----------- Eigen/src/Core/Matrix.h | 15 ++ Eigen/src/Core/PlainObjectBase.h | 14 ++ Eigen/src/Core/util/Macros.h | 13 ++ test/CMakeLists.txt | 1 + test/rvalue_types.cpp | 73 ++++++++++ 7 files changed, 277 insertions(+), 80 deletions(-) create mode 100644 test/rvalue_types.cpp diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h index 497efff66..3d0350834 100644 --- a/Eigen/src/Core/Array.h +++ b/Eigen/src/Core/Array.h @@ -124,6 +124,21 @@ class Array } #endif +#ifdef EIGEN_HAVE_RVALUE_REFERENCES + Array(Array&& other) + : Base(std::move(other)) + { + Base::_check_template_params(); + if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic) + Base::_set_noalias(other); + } + Array& operator=(Array&& other) + { + other.swap(*this); + return *this; + } +#endif + /** Constructs a vector or row-vector with given dimension. \only_for_vectors * * Note that this is only useful for dynamic-size vectors. For fixed-size vectors, diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 3e7f9c1b7..ae86b4d11 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -114,33 +114,37 @@ template class DenseSt { internal::plain_array m_data; public: - inline DenseStorage() {} - inline DenseStorage(internal::constructor_without_unaligned_array_assert) + DenseStorage() {} + DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()) {} - inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} - inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); } - static inline DenseIndex rows(void) {return _Rows;} - static inline DenseIndex cols(void) {return _Cols;} - inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} - inline void resize(DenseIndex,DenseIndex,DenseIndex) {} - inline const T *data() const { return m_data.array; } - inline T *data() { return m_data.array; } + DenseStorage(const DenseStorage& other) : m_data(other.m_data) {} + DenseStorage& operator=(DenseStorage other) { other.swap(*this); return *this; } + DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} + void swap(DenseStorage& other) { std::swap(m_data,other.m_data); } + static DenseIndex rows(void) {return _Rows;} + static DenseIndex cols(void) {return _Cols;} + void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} + void resize(DenseIndex,DenseIndex,DenseIndex) {} + const T *data() const { return m_data.array; } + T *data() { return m_data.array; } }; // null matrix template class DenseStorage { public: - inline DenseStorage() {} - inline DenseStorage(internal::constructor_without_unaligned_array_assert) {} - inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} - inline void swap(DenseStorage& ) {} - static inline DenseIndex rows(void) {return _Rows;} - static inline DenseIndex cols(void) {return _Cols;} - inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} - inline void resize(DenseIndex,DenseIndex,DenseIndex) {} - inline const T *data() const { return 0; } - inline T *data() { return 0; } + DenseStorage() {} + DenseStorage(internal::constructor_without_unaligned_array_assert) {} + DenseStorage(const DenseStorage&) {} + DenseStorage& operator=(const DenseStorage&) { return *this; } + DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} + void swap(DenseStorage& ) {} + static DenseIndex rows(void) {return _Rows;} + static DenseIndex cols(void) {return _Cols;} + void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} + void resize(DenseIndex,DenseIndex,DenseIndex) {} + const T *data() const { return 0; } + T *data() { return 0; } }; // more specializations for null matrices; these are necessary to resolve ambiguities @@ -160,18 +164,20 @@ template class DenseStorage class DenseStorage m_data; DenseIndex m_rows; public: - inline DenseStorage() : m_rows(0) {} - inline DenseStorage(internal::constructor_without_unaligned_array_assert) + DenseStorage() : m_rows(0) {} + DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {} - inline DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex) : m_rows(nbRows) {} - inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } - inline DenseIndex rows(void) const {return m_rows;} - inline DenseIndex cols(void) const {return _Cols;} - inline void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; } - inline void resize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; } - inline const T *data() const { return m_data.array; } - inline T *data() { return m_data.array; } + DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {} + DenseStorage& operator=(DenseStorage other) { other.swap(*this); return *this; } + DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex) : m_rows(nbRows) {} + void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } + DenseIndex rows(void) const {return m_rows;} + DenseIndex cols(void) const {return _Cols;} + void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; } + void resize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; } + const T *data() const { return m_data.array; } + T *data() { return m_data.array; } }; // dynamic-size matrix with fixed-size storage and fixed height @@ -199,17 +207,19 @@ template class DenseStorage m_data; DenseIndex m_cols; public: - inline DenseStorage() : m_cols(0) {} - inline DenseStorage(internal::constructor_without_unaligned_array_assert) + DenseStorage() : m_cols(0) {} + DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {} - inline DenseStorage(DenseIndex, DenseIndex, DenseIndex nbCols) : m_cols(nbCols) {} - inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } - inline DenseIndex rows(void) const {return _Rows;} - inline DenseIndex cols(void) const {return m_cols;} - inline void conservativeResize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; } - inline void resize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; } - inline const T *data() const { return m_data.array; } - inline T *data() { return m_data.array; } + DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {} + DenseStorage& operator=(DenseStorage other) { other.swap(*this); return *this; } + DenseStorage(DenseIndex, DenseIndex, DenseIndex nbCols) : m_cols(nbCols) {} + void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } + DenseIndex rows(void) const {return _Rows;} + DenseIndex cols(void) const {return m_cols;} + void conservativeResize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; } + void resize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; } + const T *data() const { return m_data.array; } + T *data() { return m_data.array; } }; // purely dynamic matrix. @@ -219,18 +229,35 @@ template class DenseStorage(size)), m_rows(nbRows), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } - inline ~DenseStorage() { internal::conditional_aligned_delete_auto(m_data, m_rows*m_cols); } - inline void swap(DenseStorage& other) +#ifdef EIGEN_HAVE_RVALUE_REFERENCES + DenseStorage(DenseStorage&& other) + : m_data(std::move(other.m_data)) + , m_rows(std::move(other.m_rows)) + , m_cols(std::move(other.m_cols)) + { + other.m_data = nullptr; + } + DenseStorage& operator=(DenseStorage&& other) + { + using std::swap; + swap(m_data, other.m_data); + swap(m_rows, other.m_rows); + swap(m_cols, other.m_cols); + return *this; + } +#endif + ~DenseStorage() { internal::conditional_aligned_delete_auto(m_data, m_rows*m_cols); } + void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); } - inline DenseIndex rows(void) const {return m_rows;} - inline DenseIndex cols(void) const {return m_cols;} - inline void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols) + DenseIndex rows(void) const {return m_rows;} + DenseIndex cols(void) const {return m_cols;} + void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, m_rows*m_cols); m_rows = nbRows; @@ -250,8 +277,11 @@ template class DenseStorage class DenseStorage(size)), m_cols(nbCols) + DenseStorage() : m_data(0), m_cols(0) {} + DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {} + DenseStorage(DenseIndex size, DenseIndex, DenseIndex nbCols) : m_data(internal::conditional_aligned_new_auto(size)), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } - inline ~DenseStorage() { internal::conditional_aligned_delete_auto(m_data, _Rows*m_cols); } - inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } - static inline DenseIndex rows(void) {return _Rows;} - inline DenseIndex cols(void) const {return m_cols;} - inline void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols) +#ifdef EIGEN_HAVE_RVALUE_REFERENCES + DenseStorage(DenseStorage&& other) + : m_data(std::move(other.m_data)) + , m_cols(std::move(other.m_cols)) + { + other.m_data = nullptr; + } + DenseStorage& operator=(DenseStorage&& other) + { + using std::swap; + swap(m_data, other.m_data); + swap(m_cols, other.m_cols); + return *this; + } +#endif + ~DenseStorage() { internal::conditional_aligned_delete_auto(m_data, _Rows*m_cols); } + void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } + static DenseIndex rows(void) {return _Rows;} + DenseIndex cols(void) const {return m_cols;} + void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, _Rows*m_cols); m_cols = nbCols; @@ -286,8 +331,11 @@ template class DenseStorage class DenseStorage(size)), m_rows(nbRows) + DenseStorage() : m_data(0), m_rows(0) {} + DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {} + DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex) : m_data(internal::conditional_aligned_new_auto(size)), m_rows(nbRows) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } - inline ~DenseStorage() { internal::conditional_aligned_delete_auto(m_data, _Cols*m_rows); } - inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } - inline DenseIndex rows(void) const {return m_rows;} - static inline DenseIndex cols(void) {return _Cols;} - inline void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex) +#ifdef EIGEN_HAVE_RVALUE_REFERENCES + DenseStorage(DenseStorage&& other) + : m_data(std::move(other.m_data)) + , m_rows(std::move(other.m_rows)) + { + other.m_data = nullptr; + } + DenseStorage& operator=(DenseStorage&& other) + { + using std::swap; + swap(m_data, other.m_data); + swap(m_rows, other.m_rows); + return *this; + } +#endif + ~DenseStorage() { internal::conditional_aligned_delete_auto(m_data, _Cols*m_rows); } + void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } + DenseIndex rows(void) const {return m_rows;} + static DenseIndex cols(void) {return _Cols;} + void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, m_rows*_Cols); m_rows = nbRows; @@ -322,8 +385,11 @@ template class DenseStorage::type } #endif +#ifdef EIGEN_HAVE_RVALUE_REFERENCES + PlainObjectBase(PlainObjectBase&& other) + : m_storage( std::move(other.m_storage) ) + { + } + + PlainObjectBase& operator=(PlainObjectBase&& other) + { + using std::swap; + swap(m_storage, other.m_storage); + return *this; + } +#endif + EIGEN_STRONG_INLINE PlainObjectBase(Index a_size, Index nbRows, Index nbCols) : m_storage(a_size, nbRows, nbCols) { diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 6798a3e0f..2a1ef296e 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -96,6 +96,19 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t #endif +// A Clang feature extension to determine compiler features. +// We use it to determine 'cxx_rvalue_references' +#ifndef __has_feature +# define __has_feature(x) 0 +#endif + +// Do we support r-value references? +#if (__has_feature(cxx_rvalue_references) || \ + defined(__GXX_EXPERIMENTAL_CXX0X__) || \ + (defined(_MSC_VER) && _MSC_VER >= 1600)) + #define EIGEN_HAVE_RVALUE_REFERENCES +#endif + /** Allows to disable some optimizations which might affect the accuracy of the result. * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them. * They currently include: diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 19afe481e..84bf65ed4 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -224,6 +224,7 @@ ei_add_test(sizeoverflow) ei_add_test(prec_inverse_4x4) ei_add_test(vectorwiseop) ei_add_test(special_numbers) +ei_add_test(rvalue_types) ei_add_test(simplicial_cholesky) ei_add_test(conjugate_gradient) diff --git a/test/rvalue_types.cpp b/test/rvalue_types.cpp new file mode 100644 index 000000000..5a581d0df --- /dev/null +++ b/test/rvalue_types.cpp @@ -0,0 +1,73 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Hauke Heibel +// +// Eigen is free software; you can redistribute it and/or +// modify it under the terms of the GNU Lesser General Public +// License as published by the Free Software Foundation; either +// version 3 of the License, or (at your option) any later version. +// +// Alternatively, you can redistribute it and/or +// modify it under the terms of the GNU General Public License as +// published by the Free Software Foundation; either version 2 of +// the License, or (at your option) any later version. +// +// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY +// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS +// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the +// GNU General Public License for more details. +// +// You should have received a copy of the GNU Lesser General Public +// License and a copy of the GNU General Public License along with +// Eigen. If not, see . + +#include "main.h" + +#include + +template +void rvalue_copyassign(const MatrixType& m) +{ +#ifdef EIGEN_HAVE_RVALUE_REFERENCES + typedef typename internal::traits::Scalar Scalar; + + // create a temporary which we are about to destroy by moving + MatrixType tmp = m; + long src_address = reinterpret_cast(tmp.data()); + + // move the temporary to n + MatrixType n = std::move(tmp); + long dst_address = reinterpret_cast(n.data()); + + if (MatrixType::RowsAtCompileTime==Dynamic|| MatrixType::ColsAtCompileTime==Dynamic) + { + // verify that we actually moved the guts + VERIFY_IS_EQUAL(src_address, dst_address); + } + + // verify that the content did not change + Scalar abs_diff = (m-n).array().abs().sum(); + VERIFY_IS_EQUAL(abs_diff, Scalar(0)); +#endif +} + +void test_rvalue_types() +{ + CALL_SUBTEST_1(rvalue_copyassign( MatrixXf::Random(50,50).eval() )); + CALL_SUBTEST_1(rvalue_copyassign( ArrayXXf::Random(50,50).eval() )); + + CALL_SUBTEST_1(rvalue_copyassign( Matrix::Random(50).eval() )); + CALL_SUBTEST_1(rvalue_copyassign( Array::Random(50).eval() )); + + CALL_SUBTEST_1(rvalue_copyassign( Matrix::Random(50).eval() )); + CALL_SUBTEST_1(rvalue_copyassign( Array::Random(50).eval() )); + + CALL_SUBTEST_2(rvalue_copyassign( Array::Random().eval() )); + CALL_SUBTEST_2(rvalue_copyassign( Array::Random().eval() )); + CALL_SUBTEST_2(rvalue_copyassign( Array::Random().eval() )); + + CALL_SUBTEST_2(rvalue_copyassign( Array::Random().eval() )); + CALL_SUBTEST_2(rvalue_copyassign( Array::Random().eval() )); + CALL_SUBTEST_2(rvalue_copyassign( Array::Random().eval() )); +} From 51b361b3bb40d547ecf39e4c2e90a806c56c9751 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Fri, 2 Aug 2013 21:07:39 +0200 Subject: [PATCH 0078/1103] Ensure that (potentially aligned) stack objects are passed by reference. --- Eigen/src/Core/DenseStorage.h | 37 +++++++++++++++++++++++++++---- unsupported/test/matrix_power.cpp | 4 ++-- 2 files changed, 35 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index ae86b4d11..67507db28 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -118,7 +118,11 @@ template class DenseSt DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()) {} DenseStorage(const DenseStorage& other) : m_data(other.m_data) {} - DenseStorage& operator=(DenseStorage other) { other.swap(*this); return *this; } + DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) m_data = other.m_data; + return *this; + } DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} void swap(DenseStorage& other) { std::swap(m_data,other.m_data); } static DenseIndex rows(void) {return _Rows;} @@ -168,7 +172,16 @@ template class DenseStorage class DenseStorage class DenseStorage -void testSingular(MatrixType m, double tol) +void testSingular(const MatrixType& m, double tol) { const int IsComplex = NumTraits::Scalar>::IsComplex; typedef typename internal::conditional, const MatrixType&>::type TriangularType; @@ -126,7 +126,7 @@ void testSingular(MatrixType m, double tol) } template -void testLogThenExp(MatrixType m, double tol) +void testLogThenExp(const MatrixType& m, double tol) { typedef typename MatrixType::Scalar Scalar; Scalar x; From 8f4d93a4b7b15fdd240f4c3c283128cead6eae16 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Fri, 2 Aug 2013 22:40:36 +0200 Subject: [PATCH 0079/1103] Fix compilation. The Matrix is required to be mutable but it also needs to be a reference and temporaries do not bind to non-const references - thus we need a hack and cast away the constness. --- .../Eigen/src/MatrixFunctions/MatrixFunction.h | 1 - unsupported/test/matrix_power.cpp | 12 ++++++++++-- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index f619d8ae5..a35c11be5 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -426,7 +426,6 @@ struct matrix_function_compute static void run(const MatrixType& A, AtomicType& atomic, ResultType &result) { typedef internal::traits Traits; - typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::Index Index; // compute Schur decomposition of A diff --git a/unsupported/test/matrix_power.cpp b/unsupported/test/matrix_power.cpp index 58644f84b..baf183d12 100644 --- a/unsupported/test/matrix_power.cpp +++ b/unsupported/test/matrix_power.cpp @@ -97,8 +97,12 @@ void testGeneral(const MatrixType& m, double tol) } template -void testSingular(const MatrixType& m, double tol) +void testSingular(const MatrixType& m_const, double tol) { + // we need to pass by reference in order to prevent errors with + // MSVC for aligned data types ... + MatrixType& m = const_cast(m_const); + const int IsComplex = NumTraits::Scalar>::IsComplex; typedef typename internal::conditional, const MatrixType&>::type TriangularType; typename internal::conditional< IsComplex, ComplexSchur, RealSchur >::type schur; @@ -126,8 +130,12 @@ void testSingular(const MatrixType& m, double tol) } template -void testLogThenExp(const MatrixType& m, double tol) +void testLogThenExp(const MatrixType& m_const, double tol) { + // we need to pass by reference in order to prevent errors with + // MSVC for aligned data types ... + MatrixType& m = const_cast(m_const); + typedef typename MatrixType::Scalar Scalar; Scalar x; From 3444f06f684ca10df70b29af812e8f0c572a709a Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Fri, 2 Aug 2013 22:54:01 +0200 Subject: [PATCH 0080/1103] Removed a warning when rvalue references are not unsed. --- test/rvalue_types.cpp | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-) diff --git a/test/rvalue_types.cpp b/test/rvalue_types.cpp index 5a581d0df..0e32f111e 100644 --- a/test/rvalue_types.cpp +++ b/test/rvalue_types.cpp @@ -26,10 +26,11 @@ #include +#ifdef EIGEN_HAVE_RVALUE_REFERENCES template void rvalue_copyassign(const MatrixType& m) -{ -#ifdef EIGEN_HAVE_RVALUE_REFERENCES +{ + typedef typename internal::traits::Scalar Scalar; // create a temporary which we are about to destroy by moving @@ -49,8 +50,11 @@ void rvalue_copyassign(const MatrixType& m) // verify that the content did not change Scalar abs_diff = (m-n).array().abs().sum(); VERIFY_IS_EQUAL(abs_diff, Scalar(0)); -#endif } +#else +template +void rvalue_copyassign(const MatrixType&) {} +#endif void test_rvalue_types() { From 8fdffdd573e6d36b04cde595460887dd4e1a2340 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Fri, 2 Aug 2013 22:33:12 +0100 Subject: [PATCH 0081/1103] Move inheritance from Eigen example in stand-alone file. Also fix a small mistake (Vector3d instead of VectorXd). --- doc/CustomizingEigen.dox | 53 ++++++------------- doc/TemplateKeyword.dox | 3 +- doc/examples/CustomizingEigen_Inheritance.cpp | 30 +++++++++++ 3 files changed, 47 insertions(+), 39 deletions(-) create mode 100644 doc/examples/CustomizingEigen_Inheritance.cpp diff --git a/doc/CustomizingEigen.dox b/doc/CustomizingEigen.dox index 5a0890ea9..0850863aa 100644 --- a/doc/CustomizingEigen.dox +++ b/doc/CustomizingEigen.dox @@ -72,55 +72,32 @@ Then one can the following declaration in the config.h or whatever prerequisites \section InheritingFromMatrix Inheriting from Matrix -Before inheriting from Matrix, be really, i mean REALLY sure that using +Before inheriting from Matrix, be really, I mean REALLY, sure that using EIGEN_MATRIX_PLUGIN is not what you really want (see previous section). If you just need to add few members to Matrix, this is the way to go. -An example of when you actually need to inherit Matrix, is when you have -several layers of heritage such as MyVerySpecificVector1,MyVerySpecificVector1 -> MyVector1 -> Matrix and. -MyVerySpecificVector3,MyVerySpecificVector4 -> MyVector2 -> Matrix. +An example of when you actually need to inherit Matrix, is when you +have several layers of heritage such as +MyVerySpecificVector1, MyVerySpecificVector2 -> MyVector1 -> Matrix and +MyVerySpecificVector3, MyVerySpecificVector4 -> MyVector2 -> Matrix. In order for your object to work within the %Eigen framework, you need to define a few members in your inherited class. -Here is a minimalistic example:\n -\code -class MyVectorType : public Eigen::VectorXd -{ -public: - MyVectorType(void):Eigen::VectorXd() {} +Here is a minimalistic example: - typedef Eigen::VectorXd Base; +\include CustomizingEigen_Inheritance.cpp - // This constructor allows you to construct MyVectorType from Eigen expressions - template - MyVectorType(const Eigen::MatrixBase& other) - : Eigen::Vector3d(other) - { } - - // This method allows you to assign Eigen expressions to MyVectorType - template - MyVectorType & operator= (const Eigen::MatrixBase & other) - { - this->Base::operator=(other); - return *this; - } -}; -\endcode +Output: \verbinclude CustomizingEigen_Inheritance.out This is the kind of error you can get if you don't provide those methods -\code -error: no match for ‘operator=’ in ‘delta = -(((Eigen::MatrixBase, 10000, 1, 2, 10000, -1> >*)(& delta)) + 8u)->Eigen::MatrixBase::cwise [with Derived = -Eigen::Matrix, 10000, 1, 2, 10000, -1>]().Eigen::Cwise::operator* [with OtherDerived = -Eigen::Matrix, 10000, 1, 2, 10000, 1>, ExpressionType = -Eigen::Matrix, 10000, 1, 2, 10000, 1>](((const -Eigen::MatrixBase, 10000, 1, 2, 10000, 1> ->&)(((const Eigen::MatrixBase, 10000, 1, ->2, 10000, 1> >*)((const spectral1d*)where)) + 8u)))’ -\endcode +\verbatim +error: no match for ‘operator=’ in ‘v = Eigen::operator*( +const Eigen::MatrixBase >::Scalar&, +const Eigen::MatrixBase >::StorageBaseType&) +(((const Eigen::MatrixBase >::StorageBaseType&) +((const Eigen::MatrixBase >::StorageBaseType*)(& v))))’ +\endverbatim \anchor user_defined_scalars \section CustomScalarType Using custom scalar types diff --git a/doc/TemplateKeyword.dox b/doc/TemplateKeyword.dox index f4e4f237e..c9944ae05 100644 --- a/doc/TemplateKeyword.dox +++ b/doc/TemplateKeyword.dox @@ -5,7 +5,8 @@ namespace Eigen { There are two uses for the \c template and \c typename keywords in C++. One of them is fairly well known amongst programmers: to define templates. The other use is more obscure: to specify that an expression refers to a template function or a type. This regularly trips up programmers that use the %Eigen library, often -leading to error messages from the compiler that are difficult to understand. +leading to error messages from the compiler that are difficult to understand, such as "expected expression" or +"no match for operator<". \eigenAutoToc diff --git a/doc/examples/CustomizingEigen_Inheritance.cpp b/doc/examples/CustomizingEigen_Inheritance.cpp new file mode 100644 index 000000000..48df64ee3 --- /dev/null +++ b/doc/examples/CustomizingEigen_Inheritance.cpp @@ -0,0 +1,30 @@ +#include +#include + +class MyVectorType : public Eigen::VectorXd +{ +public: + MyVectorType(void):Eigen::VectorXd() {} + + // This constructor allows you to construct MyVectorType from Eigen expressions + template + MyVectorType(const Eigen::MatrixBase& other) + : Eigen::VectorXd(other) + { } + + // This method allows you to assign Eigen expressions to MyVectorType + template + MyVectorType& operator=(const Eigen::MatrixBase & other) + { + this->Eigen::VectorXd::operator=(other); + return *this; + } +}; + +int main() +{ + MyVectorType v = MyVectorType::Ones(4); + v(2) += 10; + v = 2 * v; + std::cout << v.transpose() << std::endl; +} From 8710440951860e7dbf4382916244e5fed498d756 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Sat, 3 Aug 2013 10:09:31 +0200 Subject: [PATCH 0082/1103] Removed errornous swap for stack storage. --- Eigen/src/Core/DenseStorage.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 67507db28..8a68316df 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -180,7 +180,7 @@ template class DenseStorage Date: Tue, 6 Aug 2013 08:03:39 +0100 Subject: [PATCH 0083/1103] Remove LinearLeastSquares.dox , which should not have been added. Accidentally included in changeset e37ff98bbb21f2ee44c6d912002ddf2cdf05ccda . --- doc/LinearLeastSquares.dox | 27 --------------------------- 1 file changed, 27 deletions(-) delete mode 100644 doc/LinearLeastSquares.dox diff --git a/doc/LinearLeastSquares.dox b/doc/LinearLeastSquares.dox deleted file mode 100644 index ab21a87ae..000000000 --- a/doc/LinearLeastSquares.dox +++ /dev/null @@ -1,27 +0,0 @@ -namespace Eigen { - -/** \eigenManualPage LinearLeastSquares Solving linear least squares problems - -lede - -\eigenAutoToc - -\section LinearLeastSquaresCopied Copied - -The best way to do least squares solving is with a SVD decomposition. Eigen provides one as the JacobiSVD class, and its solve() -is doing least-squares solving. - -Here is an example: - - - - - - -
Example:Output:
\include TutorialLinAlgSVDSolve.cpp \verbinclude TutorialLinAlgSVDSolve.out
- -For more information, including faster but less reliable methods, read our page concentrating on \ref LinearLeastSquares "linear least squares problems". - -*/ - -} From 616f9cc593fb47e39e0a47b6a749f15d66a5c734 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Tue, 6 Aug 2013 09:49:44 +0100 Subject: [PATCH 0084/1103] doc: Explain type of result for VectorwiseOp member functions. Prompted by a question on the forum. --- Eigen/src/Core/VectorwiseOp.h | 11 ++++++++++- doc/snippets/PartialRedux_count.cpp | 4 +++- 2 files changed, 13 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index 511564875..328b6179b 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -301,6 +301,7 @@ template class VectorwiseOp /** \returns a row (or column) vector expression of the squared norm * of each column (or row) of the referenced expression. + * This is a vector with real entries, even if the original matrix has complex entries. * * Example: \include PartialRedux_squaredNorm.cpp * Output: \verbinclude PartialRedux_squaredNorm.out @@ -311,6 +312,7 @@ template class VectorwiseOp /** \returns a row (or column) vector expression of the norm * of each column (or row) of the referenced expression. + * This is a vector with real entries, even if the original matrix has complex entries. * * Example: \include PartialRedux_norm.cpp * Output: \verbinclude PartialRedux_norm.out @@ -322,7 +324,8 @@ template class VectorwiseOp /** \returns a row (or column) vector expression of the norm * of each column (or row) of the referenced expression, using - * blue's algorithm. + * Blue's algorithm. + * This is a vector with real entries, even if the original matrix has complex entries. * * \sa DenseBase::blueNorm() */ const typename ReturnType::Type blueNorm() const @@ -332,6 +335,7 @@ template class VectorwiseOp /** \returns a row (or column) vector expression of the norm * of each column (or row) of the referenced expression, avoiding * underflow and overflow. + * This is a vector with real entries, even if the original matrix has complex entries. * * \sa DenseBase::stableNorm() */ const typename ReturnType::Type stableNorm() const @@ -341,6 +345,7 @@ template class VectorwiseOp /** \returns a row (or column) vector expression of the norm * of each column (or row) of the referenced expression, avoiding * underflow and overflow using a concatenation of hypot() calls. + * This is a vector with real entries, even if the original matrix has complex entries. * * \sa DenseBase::hypotNorm() */ const typename ReturnType::Type hypotNorm() const @@ -365,6 +370,7 @@ template class VectorwiseOp /** \returns a row (or column) vector expression representing * whether \b all coefficients of each respective column (or row) are \c true. + * This expression can be assigned to a vector with entries of type \c bool. * * \sa DenseBase::all() */ const typename ReturnType::Type all() const @@ -372,6 +378,7 @@ template class VectorwiseOp /** \returns a row (or column) vector expression representing * whether \b at \b least one coefficient of each respective column (or row) is \c true. + * This expression can be assigned to a vector with entries of type \c bool. * * \sa DenseBase::any() */ const typename ReturnType::Type any() const @@ -379,6 +386,8 @@ template class VectorwiseOp /** \returns a row (or column) vector expression representing * the number of \c true coefficients of each respective column (or row). + * This expression can be assigned to a vector whose entries have the same type as is used to + * index entries of the original matrix; for dense matrices, this is \c std::ptrdiff_t . * * Example: \include PartialRedux_count.cpp * Output: \verbinclude PartialRedux_count.out diff --git a/doc/snippets/PartialRedux_count.cpp b/doc/snippets/PartialRedux_count.cpp index c7b3097e4..1c3b3a28f 100644 --- a/doc/snippets/PartialRedux_count.cpp +++ b/doc/snippets/PartialRedux_count.cpp @@ -1,3 +1,5 @@ Matrix3d m = Matrix3d::Random(); cout << "Here is the matrix m:" << endl << m << endl; -cout << "Here is the count of elements larger or equal than 0.5 of each row:" << endl << (m.array() >= 0.5).rowwise().count() << endl; +Matrix res = (m.array() >= 0.5).rowwise().count(); +cout << "Here is the count of elements larger or equal than 0.5 of each row:" << endl; +cout << res << endl; From 306ce33e1c08716005554228dd17138f11a778c0 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Wed, 7 Aug 2013 16:34:34 +0100 Subject: [PATCH 0085/1103] BDCSVD: Streamline compute() and copyUV() --- unsupported/Eigen/src/SVD/BDCSVD.h | 61 ++++++++++-------------------- 1 file changed, 20 insertions(+), 41 deletions(-) diff --git a/unsupported/Eigen/src/SVD/BDCSVD.h b/unsupported/Eigen/src/SVD/BDCSVD.h index 11d4882e4..dab419a47 100644 --- a/unsupported/Eigen/src/SVD/BDCSVD.h +++ b/unsupported/Eigen/src/SVD/BDCSVD.h @@ -139,7 +139,7 @@ public: void setSwitchSize(int s) { - eigen_assert(s>3 && "BDCSVD the size of the algo switch has to be greater than 4"); + eigen_assert(s>3 && "BDCSVD the size of the algo switch has to be greater than 3"); algoswap = s; } @@ -201,7 +201,7 @@ private: void deflation43(Index firstCol, Index shift, Index i, Index size); void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size); void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift); - void copyUV(MatrixXr naiveU, MatrixXr naiveV, MatrixX householderU, MatrixX houseHolderV); + void copyUV(MatrixX householderU, MatrixX houseHolderV); protected: MatrixXr m_naiveU, m_naiveV; @@ -282,15 +282,8 @@ BDCSVD::compute(const MatrixType& matrix, unsigned int computationOp internal::UpperBidiagonalization bid(copy); //**** step 2 Divide - // this is ugly and has to be redone (care of complex cast) - MatrixXr temp; - temp = bid.bidiagonal().toDenseMatrix().transpose(); - m_computed.setZero(); - for (int i=0; im_diagSize - 1; i++) { - m_computed(i, i) = temp(i, i); - m_computed(i + 1, i) = temp(i + 1, i); - } - m_computed(this->m_diagSize - 1, this->m_diagSize - 1) = temp(this->m_diagSize - 1, this->m_diagSize - 1); + m_computed.topRows(this->m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose(); + m_computed.template bottomRows<1>().setZero(); divide(0, this->m_diagSize - 1, 0, 0, 0); //**** step 3 copy @@ -307,45 +300,31 @@ BDCSVD::compute(const MatrixType& matrix, unsigned int computationOp break; } } - copyUV(m_naiveV, m_naiveU, bid.householderU(), bid.householderV()); + copyUV(bid.householderU(), bid.householderV()); this->m_isInitialized = true; return *this; }// end compute +// TODO: this function should accept householder sequences to save converting them to matrix template -void BDCSVD::copyUV(MatrixXr naiveU, MatrixXr naiveV, MatrixX householderU, MatrixX householderV){ +void BDCSVD::copyUV(MatrixX householderU, MatrixX householderV){ + // Note exchange of U and V: m_matrixU is set from m_naiveV and vice versa if (this->computeU()){ - MatrixX temp = MatrixX::Zero(naiveU.rows(), naiveU.cols()); - temp.real() = naiveU; - if (this->m_computeThinU){ - this->m_matrixU = MatrixX::Identity(householderU.cols(), this->m_nonzeroSingularValues ); - this->m_matrixU.block(0, 0, this->m_diagSize, this->m_nonzeroSingularValues) = - temp.block(0, 0, this->m_diagSize, this->m_nonzeroSingularValues); - this->m_matrixU = householderU * this->m_matrixU ; - } - else - { - this->m_matrixU = MatrixX::Identity(householderU.cols(), householderU.cols()); - this->m_matrixU.block(0, 0, this->m_diagSize, this->m_diagSize) = temp.block(0, 0, this->m_diagSize, this->m_diagSize); - this->m_matrixU = householderU * this->m_matrixU ; - } + Index Ucols = this->m_computeThinU ? this->m_nonzeroSingularValues : householderU.cols(); + this->m_matrixU = MatrixX::Identity(householderU.cols(), Ucols); + Index blockCols = this->m_computeThinU ? this->m_nonzeroSingularValues : this->m_diagSize; + this->m_matrixU.block(0, 0, this->m_diagSize, blockCols) = + m_naiveV.template cast().block(0, 0, this->m_diagSize, blockCols); + this->m_matrixU = householderU * this->m_matrixU; } if (this->computeV()){ - MatrixX temp = MatrixX::Zero(naiveV.rows(), naiveV.cols()); - temp.real() = naiveV; - if (this->m_computeThinV){ - this->m_matrixV = MatrixX::Identity(householderV.cols(),this->m_nonzeroSingularValues ); - this->m_matrixV.block(0, 0, this->m_nonzeroSingularValues, this->m_nonzeroSingularValues) = - temp.block(0, 0, this->m_nonzeroSingularValues, this->m_nonzeroSingularValues); - this->m_matrixV = householderV * this->m_matrixV ; - } - else - { - this->m_matrixV = MatrixX::Identity(householderV.cols(), householderV.cols()); - this->m_matrixV.block(0, 0, this->m_diagSize, this->m_diagSize) = temp.block(0, 0, this->m_diagSize, this->m_diagSize); - this->m_matrixV = householderV * this->m_matrixV; - } + Index Vcols = this->m_computeThinV ? this->m_nonzeroSingularValues : householderV.cols(); + this->m_matrixV = MatrixX::Identity(householderV.cols(), Vcols); + Index blockCols = this->m_computeThinV ? this->m_nonzeroSingularValues : this->m_diagSize; + this->m_matrixV.block(0, 0, this->m_diagSize, blockCols) = + m_naiveU.template cast().block(0, 0, this->m_diagSize, blockCols); + this->m_matrixV = householderV * this->m_matrixV; } } From 097a105603e61eb32856e9b772e5da155b25ba15 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Sat, 10 Aug 2013 19:10:23 +0200 Subject: [PATCH 0086/1103] Disabled std::log1p on Cygwin. --- Eigen/src/Core/MathFunctions.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 454f9ce52..fc1af1a5d 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -338,7 +338,7 @@ struct atanh2_impl static inline Scalar run(const Scalar& x, const Scalar& r) { EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar) - #if __cplusplus >= 201103L + #if (__cplusplus >= 201103L) && !defined(__CYGWIN__) using std::log1p; return log1p(2 * x / (r - x)) / 2; #else From 8a89ba927564d9e6209495da7729d8e93f381dec Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Sat, 10 Aug 2013 19:11:03 +0200 Subject: [PATCH 0087/1103] Added alternative C++11 detection. --- Eigen/src/Core/util/Macros.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 2a1ef296e..7e5289cdd 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -104,8 +104,9 @@ // Do we support r-value references? #if (__has_feature(cxx_rvalue_references) || \ + (defined(__cplusplus) && __cplusplus >= 201103L) || \ defined(__GXX_EXPERIMENTAL_CXX0X__) || \ - (defined(_MSC_VER) && _MSC_VER >= 1600)) + (defined(_MSC_VER) && _MSC_VER >= 1600)) #define EIGEN_HAVE_RVALUE_REFERENCES #endif From e4acd6e2fd32bbef0eaf82aef4b0442c95573ead Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Sat, 10 Aug 2013 19:13:46 +0200 Subject: [PATCH 0088/1103] Added copy constructor and assignment to DenseStorage. Required by the standard even when its not used but elided. Added a test for DenseStorage copying and assignment. --- Eigen/src/Core/DenseStorage.h | 57 ++++++++++++++++++---- test/CMakeLists.txt | 1 + test/dense_storage.cpp | 91 +++++++++++++++++++++++++++++++++++ 3 files changed, 139 insertions(+), 10 deletions(-) create mode 100644 test/dense_storage.cpp diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 8a68316df..d464904d1 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -3,7 +3,7 @@ // // Copyright (C) 2008 Gael Guennebaud // Copyright (C) 2006-2009 Benoit Jacob -// Copyright (C) 2010 Hauke Heibel +// Copyright (C) 2010-2013 Hauke Heibel // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -264,6 +264,22 @@ template class DenseStorage(size)), m_rows(nbRows), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } + DenseStorage(const DenseStorage& other) + : m_data(internal::conditional_aligned_new_auto(other.m_rows*other.m_cols)) + , m_rows(other.m_rows) + , m_cols(other.m_cols) + { + internal::smart_copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data); + } + DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + DenseStorage tmp(other); + this->swap(tmp); + } + return *this; + } #ifdef EIGEN_HAVE_RVALUE_REFERENCES DenseStorage(DenseStorage&& other) : m_data(std::move(other.m_data)) @@ -308,9 +324,6 @@ template class DenseStorage class DenseStorage(size)), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } + DenseStorage(const DenseStorage& other) + : m_data(internal::conditional_aligned_new_auto(_Rows*other.m_cols)) + , m_cols(other.m_cols) + { + internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data); + } + DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + DenseStorage tmp(other); + this->swap(tmp); + } + return *this; + } #ifdef EIGEN_HAVE_RVALUE_REFERENCES DenseStorage(DenseStorage&& other) : m_data(std::move(other.m_data)) @@ -362,9 +390,6 @@ template class DenseStorage class DenseStorage(size)), m_rows(nbRows) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } + DenseStorage(const DenseStorage& other) + : m_data(internal::conditional_aligned_new_auto(other.m_rows*_Cols)) + , m_rows(other.m_rows) + { + internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data); + } + DenseStorage& operator=(const DenseStorage& other) + { + if (this != &other) + { + DenseStorage tmp(other); + this->swap(tmp); + } + return *this; + } #ifdef EIGEN_HAVE_RVALUE_REFERENCES DenseStorage(DenseStorage&& other) : m_data(std::move(other.m_data)) @@ -416,9 +456,6 @@ template class DenseStorage +// +// Eigen is free software; you can redistribute it and/or +// modify it under the terms of the GNU Lesser General Public +// License as published by the Free Software Foundation; either +// version 3 of the License, or (at your option) any later version. +// +// Alternatively, you can redistribute it and/or +// modify it under the terms of the GNU General Public License as +// published by the Free Software Foundation; either version 2 of +// the License, or (at your option) any later version. +// +// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY +// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS +// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the +// GNU General Public License for more details. +// +// You should have received a copy of the GNU Lesser General Public +// License and a copy of the GNU General Public License along with +// Eigen. If not, see . + +#include "main.h" + +#include + +template +void dense_storage_copy() +{ + static const int Size = ((Rows==Dynamic || Cols==Dynamic) ? Dynamic : Rows*Cols); + typedef DenseStorage DenseStorageType; + + const int rows = (Rows==Dynamic) ? 4 : Rows; + const int cols = (Cols==Dynamic) ? 3 : Cols; + const int size = rows*cols; + DenseStorageType reference(size, rows, cols); + T* raw_reference = reference.data(); + for (int i=0; i(i); + + DenseStorageType copied_reference(reference); + const T* raw_copied_reference = copied_reference.data(); + for (int i=0; i +void dense_storage_assignment() +{ + static const int Size = ((Rows==Dynamic || Cols==Dynamic) ? Dynamic : Rows*Cols); + typedef DenseStorage DenseStorageType; + + const int rows = (Rows==Dynamic) ? 4 : Rows; + const int cols = (Cols==Dynamic) ? 3 : Cols; + const int size = rows*cols; + DenseStorageType reference(size, rows, cols); + T* raw_reference = reference.data(); + for (int i=0; i(i); + + DenseStorageType copied_reference; + copied_reference = reference; + const T* raw_copied_reference = copied_reference.data(); + for (int i=0; i(); + dense_storage_copy(); + dense_storage_copy(); + dense_storage_copy(); + + dense_storage_copy(); + dense_storage_copy(); + dense_storage_copy(); + dense_storage_copy(); + + dense_storage_assignment(); + dense_storage_assignment(); + dense_storage_assignment(); + dense_storage_assignment(); + + dense_storage_assignment(); + dense_storage_assignment(); + dense_storage_assignment(); + dense_storage_assignment(); +} From c13e9bbabf885abad49aafb77535192262355448 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sun, 11 Aug 2013 10:17:23 +0100 Subject: [PATCH 0089/1103] QuickReference.dox: std::tan(array) --> tan(array), same for other functions. --- doc/QuickReference.dox | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/doc/QuickReference.dox b/doc/QuickReference.dox index 5e0168649..a4be0f68a 100644 --- a/doc/QuickReference.dox +++ b/doc/QuickReference.dox @@ -405,19 +405,19 @@ array1 == array2 array1 != array2 array1 == scalar array1 != scalar array1.min(array2) array1.max(array2) array1.abs2() -array1.abs() std::abs(array1) -array1.sqrt() std::sqrt(array1) -array1.log() std::log(array1) -array1.exp() std::exp(array1) -array1.pow(exponent) std::pow(array1,exponent) +array1.abs() abs(array1) +array1.sqrt() sqrt(array1) +array1.log() log(array1) +array1.exp() exp(array1) +array1.pow(exponent) pow(array1,exponent) array1.square() array1.cube() array1.inverse() -array1.sin() std::sin(array1) -array1.cos() std::cos(array1) -array1.tan() std::tan(array1) -array1.asin() std::asin(array1) -array1.acos() std::acos(array1) +array1.sin() sin(array1) +array1.cos() cos(array1) +array1.tan() tan(array1) +array1.asin() asin(array1) +array1.acos() acos(array1) \endcode From 6719e56b5bfe9ae4badc9a6e894c5824f663d62e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 11 Aug 2013 17:52:43 +0200 Subject: [PATCH 0090/1103] Ref<> objects must be nested by reference because they potentially store a temporary object --- Eigen/src/Core/Ref.h | 5 +++-- test/ref.cpp | 4 ++-- unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h | 2 +- unsupported/Eigen/src/MatrixFunctions/MatrixPower.h | 2 +- 4 files changed, 7 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index aba795bdb..00d9e6d2b 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -94,7 +94,8 @@ struct traits > typedef _PlainObjectType PlainObjectType; typedef _StrideType StrideType; enum { - Options = _Options + Options = _Options, + Flags = traits >::Flags | NestByRefBit }; template struct match { @@ -111,7 +112,7 @@ struct traits > }; typedef typename internal::conditional::type type; }; - + }; template diff --git a/test/ref.cpp b/test/ref.cpp index 65b4f5a3e..f639d900b 100644 --- a/test/ref.cpp +++ b/test/ref.cpp @@ -14,9 +14,9 @@ static int nb_temporaries; -inline void on_temporary_creation(int size) { +inline void on_temporary_creation(int) { // here's a great place to set a breakpoint when debugging failures in this test! - if(size!=0) nb_temporaries++; + nb_temporaries++; } diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h index 8d42e69b9..763ed414a 100644 --- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h +++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h @@ -604,7 +604,7 @@ atan2(const AutoDiffScalar& a, const AutoDiffScalar& b) Scalar tmp4 = tmp3/(tmp2+tmp3); if (tmp4!=0) - ret.derivatives() = (a.derivatives() * b.value() - a.value() * b.derivatives()) * (tmp2+tmp3); + ret.derivatives() = (a.value() * b.derivatives() - a.derivatives() * b.value()) * (tmp2+tmp3); return ret; } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index ee665c18e..b91d8632a 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -299,7 +299,7 @@ MatrixPowerAtomic::computeSuperDiag(const ComplexScalar& curr, const ComplexScalar logCurr = log(curr); ComplexScalar logPrev = log(prev); int unwindingNumber = ceil((numext::imag(logCurr - logPrev) - M_PI) / (2*M_PI)); - ComplexScalar w = numext::atanh2(curr - prev, curr + prev) + ComplexScalar(0, M_PI*unwindingNumber); + ComplexScalar w = numext::atanh2(curr - prev, curr + prev) + ComplexScalar(RealScalar(0), RealScalar(M_PI*unwindingNumber)); return RealScalar(2) * exp(RealScalar(0.5) * p * (logCurr + logPrev)) * sinh(p * w) / (curr - prev); } From bffdc491b304cb207ed63cc421f39e597f8a6a9d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 11 Aug 2013 19:21:43 +0200 Subject: [PATCH 0091/1103] Fix cost evaluation of partial reduxions -> improve performance of vectorwise/replicate expressions involving partial reduxions --- Eigen/src/Core/VectorwiseOp.h | 5 +++-- test/vectorwiseop.cpp | 10 ++++++++++ 2 files changed, 13 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index 328b6179b..f25ddca17 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -50,7 +50,7 @@ struct traits > MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime, Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits, Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0), - TraversalSize = Direction==Vertical ? RowsAtCompileTime : ColsAtCompileTime + TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime : MatrixType::ColsAtCompileTime }; #if EIGEN_GNUC_AT_LEAST(3,4) typedef typename MemberOp::template Cost CostOpType; @@ -58,7 +58,8 @@ struct traits > typedef typename MemberOp::template Cost CostOpType; #endif enum { - CoeffReadCost = TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value) + CoeffReadCost = TraversalSize==Dynamic ? Dynamic + : TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value) }; }; } diff --git a/test/vectorwiseop.cpp b/test/vectorwiseop.cpp index 9d60b0388..6cd1acdda 100644 --- a/test/vectorwiseop.cpp +++ b/test/vectorwiseop.cpp @@ -101,6 +101,16 @@ template void vectorwiseop_array(const ArrayType& m) VERIFY_RAISES_ASSERT(m2.rowwise() /= rowvec.transpose()); VERIFY_RAISES_ASSERT(m1.rowwise() / rowvec.transpose()); + + m2 = m1; + // yes, there might be an aliasing issue there but ".rowwise() /=" + // is suppposed to evaluate " m2.colwise().sum()" into to temporary to avoid + // evaluating the reducions multiple times + if(ArrayType::RowsAtCompileTime>2 || ArrayType::RowsAtCompileTime==Dynamic) + { + m2.rowwise() /= m2.colwise().sum(); + VERIFY_IS_APPROX(m2, m1.rowwise() / m1.colwise().sum()); + } } template void vectorwiseop_matrix(const MatrixType& m) From 916d29e58f678d265eca84115148f7bd97f588a7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 11 Aug 2013 19:26:41 +0200 Subject: [PATCH 0092/1103] Backout parts of changeset 6719e56b5bfe9ae4badc9a6e894c5824f663d62e (these changes were not intended to be commited) --- unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h | 2 +- unsupported/Eigen/src/MatrixFunctions/MatrixPower.h | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h index 763ed414a..8d42e69b9 100644 --- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h +++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h @@ -604,7 +604,7 @@ atan2(const AutoDiffScalar& a, const AutoDiffScalar& b) Scalar tmp4 = tmp3/(tmp2+tmp3); if (tmp4!=0) - ret.derivatives() = (a.value() * b.derivatives() - a.derivatives() * b.value()) * (tmp2+tmp3); + ret.derivatives() = (a.derivatives() * b.value() - a.value() * b.derivatives()) * (tmp2+tmp3); return ret; } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h index b91d8632a..ee665c18e 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixPower.h @@ -299,7 +299,7 @@ MatrixPowerAtomic::computeSuperDiag(const ComplexScalar& curr, const ComplexScalar logCurr = log(curr); ComplexScalar logPrev = log(prev); int unwindingNumber = ceil((numext::imag(logCurr - logPrev) - M_PI) / (2*M_PI)); - ComplexScalar w = numext::atanh2(curr - prev, curr + prev) + ComplexScalar(RealScalar(0), RealScalar(M_PI*unwindingNumber)); + ComplexScalar w = numext::atanh2(curr - prev, curr + prev) + ComplexScalar(0, M_PI*unwindingNumber); return RealScalar(2) * exp(RealScalar(0.5) * p * (logCurr + logPrev)) * sinh(p * w) / (curr - prev); } From 6f5f488a80307adc6299839c4d35fb1a82b5fe37 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Mon, 12 Aug 2013 07:39:24 +0200 Subject: [PATCH 0093/1103] Switched to MPL2 license. --- test/dense_storage.cpp | 23 ++++------------------- test/rvalue_types.cpp | 23 ++++------------------- 2 files changed, 8 insertions(+), 38 deletions(-) diff --git a/test/dense_storage.cpp b/test/dense_storage.cpp index e9c3e5b2d..e63712b1a 100644 --- a/test/dense_storage.cpp +++ b/test/dense_storage.cpp @@ -1,26 +1,11 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2010 Hauke Heibel +// Copyright (C) 2013 Hauke Heibel // -// Eigen is free software; you can redistribute it and/or -// modify it under the terms of the GNU Lesser General Public -// License as published by the Free Software Foundation; either -// version 3 of the License, or (at your option) any later version. -// -// Alternatively, you can redistribute it and/or -// modify it under the terms of the GNU General Public License as -// published by the Free Software Foundation; either version 2 of -// the License, or (at your option) any later version. -// -// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY -// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS -// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the -// GNU General Public License for more details. -// -// You should have received a copy of the GNU Lesser General Public -// License and a copy of the GNU General Public License along with -// Eigen. If not, see . +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. #include "main.h" diff --git a/test/rvalue_types.cpp b/test/rvalue_types.cpp index 0e32f111e..3eebfc61b 100644 --- a/test/rvalue_types.cpp +++ b/test/rvalue_types.cpp @@ -1,26 +1,11 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2010 Hauke Heibel +// Copyright (C) 2013 Hauke Heibel // -// Eigen is free software; you can redistribute it and/or -// modify it under the terms of the GNU Lesser General Public -// License as published by the Free Software Foundation; either -// version 3 of the License, or (at your option) any later version. -// -// Alternatively, you can redistribute it and/or -// modify it under the terms of the GNU General Public License as -// published by the Free Software Foundation; either version 2 of -// the License, or (at your option) any later version. -// -// Eigen is distributed in the hope that it will be useful, but WITHOUT ANY -// WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS -// FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the -// GNU General Public License for more details. -// -// You should have received a copy of the GNU Lesser General Public -// License and a copy of the GNU General Public License along with -// Eigen. If not, see . +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. #include "main.h" From 956251b738a4d955fb4322c2bc5dc5170d9b8367 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 12 Aug 2013 13:37:47 +0200 Subject: [PATCH 0094/1103] bug #638: fix typos in sparse tutorial --- doc/TutorialSparse.dox | 2 +- doc/special_examples/CMakeLists.txt | 6 +++--- doc/special_examples/Tutorial_sparse_example_details.cpp | 4 ++-- 3 files changed, 6 insertions(+), 6 deletions(-) diff --git a/doc/TutorialSparse.dox b/doc/TutorialSparse.dox index dbfb4a9eb..835c59354 100644 --- a/doc/TutorialSparse.dox +++ b/doc/TutorialSparse.dox @@ -83,7 +83,7 @@ There is no notion of compressed/uncompressed mode for a SparseVector. \section TutorialSparseExample First example -Before describing each individual class, let's start with the following typical example: solving the Lapace equation \f$ \nabla u = 0 \f$ on a regular 2D grid using a finite difference scheme and Dirichlet boundary conditions. +Before describing each individual class, let's start with the following typical example: solving the Laplace equation \f$ \nabla u = 0 \f$ on a regular 2D grid using a finite difference scheme and Dirichlet boundary conditions. Such problem can be mathematically expressed as a linear problem of the form \f$ Ax=b \f$ where \f$ x \f$ is the vector of \c m unknowns (in our case, the values of the pixels), \f$ b \f$ is the right hand side vector resulting from the boundary conditions, and \f$ A \f$ is an \f$ m \times m \f$ matrix containing only a few non-zero elements resulting from the discretization of the Laplacian operator. diff --git a/doc/special_examples/CMakeLists.txt b/doc/special_examples/CMakeLists.txt index 138a2f6ef..0c9b3c3ba 100644 --- a/doc/special_examples/CMakeLists.txt +++ b/doc/special_examples/CMakeLists.txt @@ -10,12 +10,12 @@ endif(NOT EIGEN_TEST_NOQT) if(QT4_FOUND) add_executable(Tutorial_sparse_example Tutorial_sparse_example.cpp Tutorial_sparse_example_details.cpp) target_link_libraries(Tutorial_sparse_example ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO} ${QT_QTCORE_LIBRARY} ${QT_QTGUI_LIBRARY}) - + add_custom_command( TARGET Tutorial_sparse_example POST_BUILD - COMMAND Tutorial_sparse_example - ARGS ${CMAKE_CURRENT_BINARY_DIR}/../html/Tutorial_sparse_example.jpeg + COMMAND Tutorial_sparse_example ARGS ${CMAKE_CURRENT_BINARY_DIR}/../html/Tutorial_sparse_example.jpeg ) + add_dependencies(all_examples Tutorial_sparse_example) endif(QT4_FOUND) diff --git a/doc/special_examples/Tutorial_sparse_example_details.cpp b/doc/special_examples/Tutorial_sparse_example_details.cpp index 8c3020b63..7d820b44a 100644 --- a/doc/special_examples/Tutorial_sparse_example_details.cpp +++ b/doc/special_examples/Tutorial_sparse_example_details.cpp @@ -11,8 +11,8 @@ void insertCoefficient(int id, int i, int j, double w, std::vector& coeffs, int n = boundary.size(); int id1 = i+j*n; - if(i==-1 || i==n) b(id) -= w * boundary(j); // constrained coeffcieint - else if(j==-1 || j==n) b(id) -= w * boundary(i); // constrained coeffcieint + if(i==-1 || i==n) b(id) -= w * boundary(j); // constrained coefficient + else if(j==-1 || j==n) b(id) -= w * boundary(i); // constrained coefficient else coeffs.push_back(T(id,id1,w)); // unknown coefficient } From ace2ed7b87e9568993bd04ff6615b83e44c86369 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 12 Aug 2013 13:38:25 +0200 Subject: [PATCH 0095/1103] Fix broken link on transforming normals --- doc/TutorialGeometry.dox | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/doc/TutorialGeometry.dox b/doc/TutorialGeometry.dox index 81aeec978..372a275de 100644 --- a/doc/TutorialGeometry.dox +++ b/doc/TutorialGeometry.dox @@ -127,7 +127,7 @@ VectorNf vec1, vec2; vec2 = t.linear() * vec1;\endcode From 1d89554f1b6bb0f6f9e16282cf0b3cd6fdccfb65 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 13 Aug 2013 14:54:09 +0200 Subject: [PATCH 0096/1103] Deprecate boolean sum operator (bug #426) --- Eigen/src/Core/Functors.h | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/Eigen/src/Core/Functors.h b/Eigen/src/Core/Functors.h index 04fb21732..080338617 100644 --- a/Eigen/src/Core/Functors.h +++ b/Eigen/src/Core/Functors.h @@ -39,6 +39,13 @@ struct functor_traits > { }; }; + +template<> struct scalar_sum_op : scalar_sum_op { + EIGEN_DEPRECATED + scalar_sum_op() {} +}; + + /** \internal * \brief Template functor to compute the product of two scalars * From e0dbc2913a4f176876ea56f247409ff340c5f1d3 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 16 Aug 2013 16:43:02 +0200 Subject: [PATCH 0097/1103] Documentation of deprecated struct. Closing bug #426. --- Eigen/src/Core/Functors.h | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Functors.h b/Eigen/src/Core/Functors.h index 080338617..69d95ea30 100644 --- a/Eigen/src/Core/Functors.h +++ b/Eigen/src/Core/Functors.h @@ -19,7 +19,7 @@ namespace internal { /** \internal * \brief Template functor to compute the sum of two scalars * - * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, MatrixBase::sum() + * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, DenseBase::sum() */ template struct scalar_sum_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op) @@ -39,7 +39,11 @@ struct functor_traits > { }; }; - +/** \internal + * \brief Template specialization to deprecate the summation of boolean expressions. + * This is required to solve Bug 426. + * \sa DenseBase::count(), DenseBase::any(), ArrayBase::cast(), MatrixBase::cast() + */ template<> struct scalar_sum_op : scalar_sum_op { EIGEN_DEPRECATED scalar_sum_op() {} From ebd6a7a46ce6887b6a7a0d0c998d914910d5025c Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Mon, 2 Sep 2013 19:09:39 +0900 Subject: [PATCH 0098/1103] Added support for custom-scalars --- Eigen/src/SparseQR/SparseQR.h | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 07c46e4b9..c42dfc556 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -205,7 +205,7 @@ class SparseQR /** \brief Reports whether previous computation was successful. * - * \returns \c Success if computation was succesful, + * \returns \c Success if computation was successful, * \c NumericalIssue if the QR factorization reports a numerical problem * \c InvalidInput if the input matrix is invalid * @@ -256,7 +256,7 @@ class SparseQR /** \brief Preprocessing step of a QR factorization * * In this step, the fill-reducing permutation is computed and applied to the columns of A - * and the column elimination tree is computed as well. Only the sparcity pattern of \a mat is exploited. + * and the column elimination tree is computed as well. Only the sparsity pattern of \a mat is exploited. * * \note In this step it is assumed that there is no empty row in the matrix \a mat. */ @@ -292,7 +292,7 @@ void SparseQR::analyzePattern(const MatrixType& mat) /** \brief Performs the numerical QR factorization of the input matrix * * The function SparseQR::analyzePattern(const MatrixType&) must have been called beforehand with - * a matrix having the same sparcity pattern than \a mat. + * a matrix having the same sparsity pattern than \a mat. * * \param mat The sparse column-major matrix */ @@ -445,7 +445,8 @@ void SparseQR::factorize(const MatrixType& mat) } else { - beta = std::sqrt(numext::abs2(c0) + sqrNorm); + using std::sqrt; + beta = sqrt(numext::abs2(c0) + sqrNorm); if(numext::real(c0) >= RealScalar(0)) beta = -beta; tval(Qidx(0)) = 1; From d4dd6aaed2c70b5e32541e96b4864b90dc07c614 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 19 Aug 2013 16:02:27 +0200 Subject: [PATCH 0099/1103] Fix bug #642: add vectorization of sqrt for doubles, and make sqrt really safe if EIGEN_FAST_MATH is disabled --- Eigen/src/Core/arch/SSE/MathFunctions.h | 11 +++++++++++ Eigen/src/Core/arch/SSE/PacketMath.h | 3 ++- doc/PreprocessorDirectives.dox | 6 +++--- 3 files changed, 16 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index 3376a984e..7a0aee658 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -442,8 +442,11 @@ Packet4f pcos(const Packet4f& _x) return _mm_xor_ps(y, sign_bit); } +#if EIGEN_FAST_MATH + // This is based on Quake3's fast inverse square root. // For detail see here: http://www.beyond3d.com/content/articles/8/ +// It lacks 1 (or 2 bits in some rare cases) of precision, and does not handle negative, +inf, or denormalized numbers correctly. template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f psqrt(const Packet4f& _x) { @@ -457,6 +460,14 @@ Packet4f psqrt(const Packet4f& _x) return pmul(_x,x); } +#else + +template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet4f& x) { return _mm_sqrt_ps(x); } + +#endif + +template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet2d& x) { return _mm_sqrt_pd(x); } + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index e256f4bac..f85d2e06e 100644 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -83,7 +83,8 @@ template<> struct packet_traits : default_packet_traits size=2, HasDiv = 1, - HasExp = 1 + HasExp = 1, + HasSqrt = 1 }; }; template<> struct packet_traits : default_packet_traits diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox index eedd5524a..981083e96 100644 --- a/doc/PreprocessorDirectives.dox +++ b/doc/PreprocessorDirectives.dox @@ -64,9 +64,9 @@ run time. However, these assertions do cost time and can thus be turned off. \c EIGEN_DONT_ALIGN is defined. - \b EIGEN_DONT_VECTORIZE - disables explicit vectorization when defined. Not defined by default, unless alignment is disabled by %Eigen's platform test or the user defining \c EIGEN_DONT_ALIGN. - - \b EIGEN_FAST_MATH - enables some optimizations which might affect the accuracy of the result. The only - optimization this currently includes is single precision sin() and cos() in the present of SSE - vectorization. Defined by default. + - \b EIGEN_FAST_MATH - enables some optimizations which might affect the accuracy of the result. This currently + enables the SSE vectorization of sin() and cos(), and speedups sqrt() for single precision. Defined to 1 by default. + Define it to 0 to disable. - \b EIGEN_UNROLLING_LIMIT - defines the size of a loop to enable meta unrolling. Set it to zero to disable unrolling. The size of a loop here is expressed in %Eigen's own notion of "number of FLOPS", it does not correspond to the number of iterations or the number of instructions. The default is value 100. From c47010e3d221396cde2aad6b3250fa698a930e28 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 19 Aug 2013 16:10:00 +0200 Subject: [PATCH 0100/1103] typo --- Eigen/src/Core/arch/SSE/MathFunctions.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index 7a0aee658..99cbd0d95 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -466,7 +466,7 @@ template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet4f& x) { ret #endif -template<> EIGEN_STRONG_INLINE Packet4f psqrt(const Packet2d& x) { return _mm_sqrt_pd(x); } +template<> EIGEN_STRONG_INLINE Packet2d psqrt(const Packet2d& x) { return _mm_sqrt_pd(x); } } // end namespace internal From 127d7f2071c9be2a62f881d53d0b696f43053453 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 19 Aug 2013 16:34:09 +0200 Subject: [PATCH 0101/1103] Fix bug #643: enable vectorization of compound assignement for fixed size objects --- Eigen/src/Core/SelfCwiseBinaryOp.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h index 22f3047b4..96012eaf8 100644 --- a/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -35,7 +35,7 @@ struct traits > enum { // Note that it is still a good idea to preserve the DirectAccessBit // so that assign can correctly align the data. - Flags = traits >::Flags | (Lhs::Flags&DirectAccessBit) | (Lhs::Flags&LvalueBit), + Flags = traits >::Flags | (Lhs::Flags&AlignedBit) | (Lhs::Flags&DirectAccessBit) | (Lhs::Flags&LvalueBit), OuterStrideAtCompileTime = Lhs::OuterStrideAtCompileTime, InnerStrideAtCompileTime = Lhs::InnerStrideAtCompileTime }; From 2b15e001068f548b852d58472b9b29f1a7bf1a2c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 19 Aug 2013 16:40:50 +0200 Subject: [PATCH 0102/1103] Make ArrayBase operator+=(scalar) and -=(scalar) use SelfCwiseBinaryOp optimization --- Eigen/Core | 2 +- Eigen/src/Core/ArrayBase.h | 6 ++---- Eigen/src/Core/SelfCwiseBinaryOp.h | 18 ++++++++++++++++++ 3 files changed, 21 insertions(+), 5 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 97aa581e2..a508b97f6 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -284,6 +284,7 @@ using std::ptrdiff_t; #include "src/Core/Assign.h" #endif +#include "src/Core/ArrayBase.h" #include "src/Core/util/BlasUtil.h" #include "src/Core/DenseStorage.h" #include "src/Core/NestByValue.h" @@ -347,7 +348,6 @@ using std::ptrdiff_t; #include "src/Core/Random.h" #include "src/Core/Replicate.h" #include "src/Core/Reverse.h" -#include "src/Core/ArrayBase.h" #include "src/Core/ArrayWrapper.h" #ifdef EIGEN_ENABLE_EVALUATORS diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index 38852600d..b7c4a1c71 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -123,10 +123,8 @@ template class ArrayBase return internal::assign_selector::run(derived(), other.derived()); } - Derived& operator+=(const Scalar& scalar) - { return *this = derived() + scalar; } - Derived& operator-=(const Scalar& scalar) - { return *this = derived() - scalar; } + Derived& operator+=(const Scalar& scalar); + Derived& operator-=(const Scalar& scalar); template Derived& operator+=(const ArrayBase& other); diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h index 96012eaf8..3d2deff98 100644 --- a/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -177,6 +177,24 @@ inline Derived& DenseBase::operator*=(const Scalar& other) return derived(); } +template +inline Derived& ArrayBase::operator+=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + SelfCwiseBinaryOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived()); + tmp = PlainObject::Constant(rows(),cols(),other); + return derived(); +} + +template +inline Derived& ArrayBase::operator-=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + SelfCwiseBinaryOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived()); + tmp = PlainObject::Constant(rows(),cols(),other); + return derived(); +} + template inline Derived& DenseBase::operator/=(const Scalar& other) { From d908ccc01cbb3483d3271589a16edde4358420fd Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Tue, 20 Aug 2013 15:00:28 +0900 Subject: [PATCH 0103/1103] Added support for custom scalars --- Eigen/src/SparseLU/SparseLU.h | 6 ++++-- Eigen/src/SparseLU/SparseLU_pivotL.h | 6 ++++-- 2 files changed, 8 insertions(+), 4 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 3abe998f2..aae82a1bc 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -267,7 +267,8 @@ class SparseLU : public internal::SparseLUImpl::pivotL(const Index jcol, const RealScalar& dia RealScalar rtemp; Index isub, icol, itemp, k; for (isub = nsupc; isub < nsupr; ++isub) { - rtemp = std::abs(lu_col_ptr[isub]); + using std::abs; + rtemp = abs(lu_col_ptr[isub]); if (rtemp > pivmax) { pivmax = rtemp; pivptr = isub; @@ -101,7 +102,8 @@ Index SparseLUImpl::pivotL(const Index jcol, const RealScalar& dia if (diag >= 0 ) { // Diagonal element exists - rtemp = std::abs(lu_col_ptr[diag]); + using std::abs; + rtemp = abs(lu_col_ptr[diag]); if (rtemp != 0.0 && rtemp >= thresh) pivptr = diag; } pivrow = lsub_ptr[pivptr]; From e4ffb7729ad41d58cb1a4b284601440b8f52bfb7 Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Tue, 20 Aug 2013 16:06:13 +0900 Subject: [PATCH 0104/1103] Removed unnecessary parentheses --- Eigen/src/SparseLU/SparseLU.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index aae82a1bc..f2b907326 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -268,7 +268,7 @@ class SparseLU : public internal::SparseLUImpl Date: Tue, 20 Aug 2013 11:52:48 +0200 Subject: [PATCH 0105/1103] Make FullPivHouseholderQR::solve returns the least-square solution instead of aborting if no exact solution exist --- Eigen/src/QR/FullPivHouseholderQR.h | 16 +++------------- 1 file changed, 3 insertions(+), 13 deletions(-) diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 0dd5ad347..611467367 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -126,11 +126,12 @@ template class FullPivHouseholderQR } /** This method finds a solution x to the equation Ax=b, where A is the matrix of which - * *this is the QR decomposition, if any exists. + * \c *this is the QR decomposition. * * \param b the right-hand-side of the equation to solve. * - * \returns a solution. + * \returns the exact or least-square solution if the rank is greater or equal to the number of columns of A, + * and an arbitrary solution otherwise. * * \note The case where b is a matrix is not yet implemented. Also, this * code is space inefficient. @@ -516,17 +517,6 @@ struct solve_retval, Rhs> dec().hCoeffs().coeff(k), &temp.coeffRef(0)); } - if(!dec().isSurjective()) - { - // is c is in the image of R ? - RealScalar biggest_in_upper_part_of_c = c.topRows( dec().rank() ).cwiseAbs().maxCoeff(); - RealScalar biggest_in_lower_part_of_c = c.bottomRows(rows-dec().rank()).cwiseAbs().maxCoeff(); - // FIXME brain dead - const RealScalar m_precision = NumTraits::epsilon() * (std::min)(rows,cols); - // this internal:: prefix is needed by at least gcc 3.4 and ICC - if(!internal::isMuchSmallerThan(biggest_in_lower_part_of_c, biggest_in_upper_part_of_c, m_precision)) - return; - } dec().matrixQR() .topLeftCorner(dec().rank(), dec().rank()) .template triangularView() From 7bca2910c7922021ea447739eb0596f15790c891 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 20 Aug 2013 13:59:33 +0200 Subject: [PATCH 0106/1103] Make the static assertions on maximal fixed size object use EIGEN_STACK_ALLOCATION_LIMIT, and raise its default value to 128KB --- Eigen/src/Core/DenseStorage.h | 8 ++++---- Eigen/src/Core/util/Macros.h | 3 ++- doc/PreprocessorDirectives.dox | 5 ++++- 3 files changed, 10 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index d464904d1..0ea2bb71f 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -38,12 +38,12 @@ struct plain_array plain_array() { - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); } plain_array(constructor_without_unaligned_array_assert) { - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); } }; @@ -76,12 +76,12 @@ struct plain_array plain_array() { EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(0xf); - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); } plain_array(constructor_without_unaligned_array_assert) { - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= 128 * 128 * 8, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); } }; diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 7e5289cdd..95f9eb7d1 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -298,7 +298,8 @@ #endif #ifndef EIGEN_STACK_ALLOCATION_LIMIT -#define EIGEN_STACK_ALLOCATION_LIMIT 20000 +// 131072 == 128 KB +#define EIGEN_STACK_ALLOCATION_LIMIT 131072 #endif #ifndef EIGEN_DEFAULT_IO_FORMAT diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox index 981083e96..eb42ee7fa 100644 --- a/doc/PreprocessorDirectives.dox +++ b/doc/PreprocessorDirectives.dox @@ -69,7 +69,10 @@ run time. However, these assertions do cost time and can thus be turned off. Define it to 0 to disable. - \b EIGEN_UNROLLING_LIMIT - defines the size of a loop to enable meta unrolling. Set it to zero to disable unrolling. The size of a loop here is expressed in %Eigen's own notion of "number of FLOPS", it does not - correspond to the number of iterations or the number of instructions. The default is value 100. + correspond to the number of iterations or the number of instructions. The default is value 100. + - \b EIGEN_STACK_ALLOCATION_LIMIT - defines the maximum bytes for a buffer to be allocated on the stack. For internal + temporary buffers, dynamic memory allocation is employed as a fall back. For fixed-size matrices or arrays, exceeding + this threshold raises a compile time assertion. Default is 128 KB. \section TopicPreprocessorDirectivesPlugins Plugins From c06e373bebd872be84d0524f10c208cce072090a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 20 Aug 2013 14:12:42 +0200 Subject: [PATCH 0107/1103] Fix compilation with non-msvc compilers. --- Eigen/src/SparseLU/SparseLU.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index f2b907326..44239e57a 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -255,6 +255,7 @@ class SparseLU : public internal::SparseLUImplcols(); ++j) @@ -295,7 +298,6 @@ class SparseLU : public internal::SparseLUImpl Date: Tue, 20 Aug 2013 14:13:41 +0200 Subject: [PATCH 0108/1103] Fix indentation --- Eigen/src/SparseLU/SparseLU.h | 84 +++++++++++++++++------------------ 1 file changed, 42 insertions(+), 42 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 44239e57a..71fe844ac 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -253,7 +253,7 @@ class SparseLU : public internal::SparseLUImplcols(); ++j) - { - for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it) - { - if(it.row() < j) continue; - if(it.row() == j) - { - det += log(abs(it.value())); - break; - } - } - } - return det; - } + /** \returns the natural log of the absolute value of the determinant of the matrix + * of which **this is the QR decomposition + * + * \note This method is useful to work around the risk of overflow/underflow that's + * inherent to the determinant computation. + * + * \sa absDeterminant(), signDeterminant() + */ + Scalar logAbsDeterminant() const + { + using std::log; + using std::abs; - /** \returns A number representing the sign of the determinant - * - * \sa absDeterminant(), logAbsDeterminant() - */ - Scalar signDeterminant() - { - eigen_assert(m_factorizationIsOk && "The matrix should be factorized first."); - return Scalar(m_detPermR); - } + eigen_assert(m_factorizationIsOk && "The matrix should be factorized first."); + Scalar det = Scalar(0.); + for (Index j = 0; j < this->cols(); ++j) + { + for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it) + { + if(it.row() < j) continue; + if(it.row() == j) + { + det += log(abs(it.value())); + break; + } + } + } + return det; + } + + /** \returns A number representing the sign of the determinant + * + * \sa absDeterminant(), logAbsDeterminant() + */ + Scalar signDeterminant() + { + eigen_assert(m_factorizationIsOk && "The matrix should be factorized first."); + return Scalar(m_detPermR); + } protected: // Functions From 403be74861071daa7a81372abf2debe71cf1aed6 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Tue, 20 Aug 2013 14:10:55 +0100 Subject: [PATCH 0109/1103] BDCSVD: Compute SVD of combined problem directly. First step at implementing final stage in BDCSVD algorithm. Uses bisection method to solve nonlinear equation. Still lots of room for optimization. --- unsupported/Eigen/src/SVD/BDCSVD.h | 177 +++++++++++++++++++++++++---- 1 file changed, 156 insertions(+), 21 deletions(-) diff --git a/unsupported/Eigen/src/SVD/BDCSVD.h b/unsupported/Eigen/src/SVD/BDCSVD.h index dab419a47..3c41e4e46 100644 --- a/unsupported/Eigen/src/SVD/BDCSVD.h +++ b/unsupported/Eigen/src/SVD/BDCSVD.h @@ -10,6 +10,7 @@ // Copyright (C) 2013 Nicolas Carre // Copyright (C) 2013 Jean Ceccato // Copyright (C) 2013 Pierre Zoppitelli +// Copyright (C) 2013 Jitse Niesen // // Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -20,7 +21,7 @@ #define EPSILON 0.0000000000000001 -#define ALGOSWAP 32 +#define ALGOSWAP 16 namespace Eigen { /** \ingroup SVD_Module @@ -198,6 +199,7 @@ private: void allocate(Index rows, Index cols, unsigned int computationOptions); void divide (Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift); + void computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V); void deflation43(Index firstCol, Index shift, Index i, Index size); void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size); void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift); @@ -292,6 +294,7 @@ BDCSVD::compute(const MatrixType& matrix, unsigned int computationOp this->m_singularValues.coeffRef(i) = a; if (a == 0){ this->m_nonzeroSingularValues = i; + this->m_singularValues.tail(this->m_diagSize - i - 1).setZero(); break; } else if (i == this->m_diagSize - 1) @@ -455,27 +458,159 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) << betaK * f.transpose().real(); - // the line below do the deflation of the matrix for the third part of the algorithm - // Here the deflation is commented because the third part of the algorithm is not implemented - // the third part of the algorithm is a fast SVD on the matrix m_computed which works thanks to the deflation - + // Second part: try to deflate singular values in combined matrix deflation(firstCol, lastCol, k, firstRowW, firstColW, shift); - // Third part of the algorithm, since the real third part of the algorithm is not implemeted we use a JacobiSVD - JacobiSVD res= JacobiSVD(m_computed.block(firstCol + shift, firstCol +shift, n + 1, n), - ComputeFullU | (ComputeFullV * compV)) ; - if (compU) m_naiveU.block(firstCol, firstCol, n + 1, n + 1) *= res.matrixU(); - else m_naiveU.block(0, firstCol, 2, n + 1) *= res.matrixU(); - - if (compV) m_naiveV.block(firstRowW, firstColW, n, n) *= res.matrixV(); - m_computed.block(firstCol + shift, firstCol + shift, n, n) << MatrixXr::Zero(n, n); - for (int i=0; i +void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V) +{ + using std::abs; + Block block = m_computed.block(firstCol, firstCol, n, n); + + // TODO Get rid of these copies (?) + Array col0 = m_computed.block(firstCol, firstCol, n, 1); + Array diag = m_computed.block(firstCol, firstCol, n, n).diagonal(); + diag(0) = 0; + + // compute singular values and vectors (in decreasing order) + singVals.resize(n); + U.resize(n+1, n+1); + if (compV) V.resize(n, n); + + if (col0.hasNaN() || diag.hasNaN()) return; + + Array shifts(n), mus(n); + for (Index k = 0; k < n; ++k) { + if (col0(k) == 0) { + // entry is deflated, so singular value is on diagonal + singVals(k) = diag(k); + mus(k) = 0; + shifts(k) = diag(k); + continue; + } + + // otherwise, use bisection to find singular value + RealScalar left = diag(k); + RealScalar right = (k != n-1) ? diag(k+1) : (diag(n-1) + col0.matrix().norm()); + + // first decide whether it's closer to the left end or the right end + RealScalar mid = left + (right-left) / 2; + RealScalar fMid = 1 + (col0.square() / ((diag + mid) * (diag - mid))).sum(); + + RealScalar shift; + if (k == 0 || fMid > 0) shift = left; + else shift = right; + + // measure everything relative to shifted + Array diagShifted = diag - shift; + RealScalar leftShifted, rightShifted; + if (shift == left) { + leftShifted = 1e-30; + if (k == 0) rightShifted = right - left; + else rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe + } else { + leftShifted = -(right - left) * 0.6; + rightShifted = -1e-30; + } + + RealScalar fLeft = 1 + (col0.square() / ((diagShifted - leftShifted) * (diag + shift + leftShifted))).sum(); + RealScalar fRight = 1 + (col0.square() / ((diagShifted - rightShifted) * (diag + shift + rightShifted))).sum(); + assert(fLeft * fRight < 0); + + while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (std::max)(abs(leftShifted), abs(rightShifted))) { + RealScalar midShifted = (leftShifted + rightShifted) / 2; + RealScalar fMid = 1 + (col0.square() / ((diagShifted - midShifted) * (diag + shift + midShifted))).sum(); + if (fLeft * fMid < 0) { + rightShifted = midShifted; + fRight = fMid; + } else { + leftShifted = midShifted; + fLeft = fMid; + } + } + + singVals[k] = shift + (leftShifted + rightShifted) / 2; + shifts[k] = shift; + mus[k] = (leftShifted + rightShifted) / 2; + + // perturb singular value slightly if it equals diagonal entry to avoid division by zero later + // (deflation is supposed to avoid this from happening) + if (singVals[k] == left) singVals[k] *= 1 + NumTraits::epsilon(); + if (singVals[k] == right) singVals[k] *= 1 - NumTraits::epsilon(); + } + + // zhat is perturbation of col0 for which singular vectors can be computed stably (see Section 3.1) + Array zhat(n); + for (Index k = 0; k < n; ++k) { + if (col0(k) == 0) + zhat(k) = 0; + else { + // see equation (3.6) + using std::sqrt; + RealScalar tmp = + sqrt( + (singVals(n-1) + diag(k)) * (mus(n-1) + (shifts(n-1) - diag(k))) + * ( + ((singVals.head(k).array() + diag(k)) * (mus.head(k) + (shifts.head(k) - diag(k)))) + / ((diag.head(k).array() + diag(k)) * (diag.head(k).array() - diag(k))) + ).prod() + * ( + ((singVals.segment(k, n-k-1).array() + diag(k)) * (mus.segment(k, n-k-1) + (shifts.segment(k, n-k-1) - diag(k)))) + / ((diag.tail(n-k-1) + diag(k)) * (diag.tail(n-k-1) - diag(k))) + ).prod() + ); + if (col0(k) > 0) zhat(k) = tmp; + else zhat(k) = -tmp; + } + } + + MatrixXr Mhat = MatrixXr::Zero(n,n); + Mhat.diagonal() = diag; + Mhat.col(0) = zhat; + + // compute singular vectors + for (Index k = 0; k < n; ++k) { + if (zhat(k) == 0) { + U.col(k) = VectorType::Unit(n+1, k); + if (compV) V.col(k) = VectorType::Unit(n, k); + } else { + U.col(k).head(n) = zhat / (((diag - shifts(k)) - mus(k)) * (diag + singVals[k])); + U(n,k) = 0; + U.col(k).normalize(); + + if (compV) { + V.col(k).tail(n-1) = (diag * zhat / (((diag - shifts(k)) - mus(k)) * (diag + singVals[k]))).tail(n-1); + V(0,k) = -1; + V.col(k).normalize(); + } + } + } + U.col(n) = VectorType::Unit(n+1, n); + + // Reverse order so that singular values in increased order + singVals.reverseInPlace(); + U.leftCols(n) = U.leftCols(n).rowwise().reverse().eval(); + if (compV) V = V.rowwise().reverse().eval(); +} + // page 12_13 // i >= 1, di almost null and zi non null. @@ -551,15 +686,16 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi }// end deflation 44 - +// acts on block from (firstCol+shift, firstCol+shift) to (lastCol+shift, lastCol+shift) [inclusive] template void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift){ //condition 4.1 - RealScalar EPS = EPSILON * (std::max(m_computed(firstCol + shift + 1, firstCol + shift + 1), m_computed(firstCol + k, firstCol + k))); + RealScalar EPS = NumTraits::epsilon() * 10 * (std::max(m_computed(firstCol + shift + 1, firstCol + shift + 1), m_computed(firstCol + k, firstCol + k))); const Index length = lastCol + 1 - firstCol; if (m_computed(firstCol + shift, firstCol + shift) < EPS){ m_computed(firstCol + shift, firstCol + shift) = EPS; } + //condition 4.2 for (Index i=firstCol + shift + 1;i<=lastCol + shift;i++){ if (std::abs(m_computed(i, firstCol + shift)) < EPS){ @@ -672,7 +808,6 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index delete [] permutation; delete [] realInd; delete [] realCol; - }//end deflation From d9381598bc8f4bb81f90125d347d4329f57e5aff Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 21 Aug 2013 14:29:00 +0200 Subject: [PATCH 0110/1103] Allows EIGEN_STACK_ALLOCATION_LIMIT to be 0 for no limit --- Eigen/src/Core/DenseStorage.h | 16 ++++++++++++---- doc/PreprocessorDirectives.dox | 2 +- 2 files changed, 13 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 0ea2bb71f..2549bd84e 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -24,6 +24,14 @@ namespace internal { struct constructor_without_unaligned_array_assert {}; +template void check_static_allocation_size() +{ + // if EIGEN_STACK_ALLOCATION_LIMIT is defined to 0, then no limit + #if EIGEN_STACK_ALLOCATION_LIMIT + EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + #endif +} + /** \internal * Static array. If the MatrixOrArrayOptions require auto-alignment, the array will be automatically aligned: * to 16 bytes boundary if the total size is a multiple of 16 bytes. @@ -38,12 +46,12 @@ struct plain_array plain_array() { - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + check_static_allocation_size(); } plain_array(constructor_without_unaligned_array_assert) { - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + check_static_allocation_size(); } }; @@ -76,12 +84,12 @@ struct plain_array plain_array() { EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(0xf); - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + check_static_allocation_size(); } plain_array(constructor_without_unaligned_array_assert) { - EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG); + check_static_allocation_size(); } }; diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox index eb42ee7fa..6e40f919f 100644 --- a/doc/PreprocessorDirectives.dox +++ b/doc/PreprocessorDirectives.dox @@ -72,7 +72,7 @@ run time. However, these assertions do cost time and can thus be turned off. correspond to the number of iterations or the number of instructions. The default is value 100. - \b EIGEN_STACK_ALLOCATION_LIMIT - defines the maximum bytes for a buffer to be allocated on the stack. For internal temporary buffers, dynamic memory allocation is employed as a fall back. For fixed-size matrices or arrays, exceeding - this threshold raises a compile time assertion. Default is 128 KB. + this threshold raises a compile time assertion. Use 0 to set no limit. Default is 128 KB. \section TopicPreprocessorDirectivesPlugins Plugins From 4ecfdc4716f0e58b0b872d36ba0698b0d852715a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 21 Aug 2013 14:29:53 +0200 Subject: [PATCH 0111/1103] Add explanations of the logic behind the matrix-vector products --- Eigen/src/Core/products/GeneralMatrixVector.h | 28 +++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h index c1cb78498..15dc59b98 100644 --- a/Eigen/src/Core/products/GeneralMatrixVector.h +++ b/Eigen/src/Core/products/GeneralMatrixVector.h @@ -26,6 +26,34 @@ namespace internal { * |real |cplx |real | alpha is converted to a cplx when calling the run function, no vectorization * |cplx |real |cplx | invalid, the caller has to do tmp: = A * B; C += alpha*tmp * |cplx |real |real | optimal case, vectorization possible via real-cplx mul + * + * Accesses to the matrix coefficients follow the following logic: + * + * - if all columns have the same alignment then + * - if the columns have the same alignment as the result vector, then easy! (-> AllAligned case) + * - otherwise perform unaligned loads only (-> NoneAligned case) + * - otherwise + * - if even columns have the same alignment then + * // odd columns are guaranteed to have the same alignment too + * - if even or odd columns have the same alignment as the result, then + * // for a register size of 2 scalars, this is guarantee to be the case (e.g., SSE with double) + * - perform half aligned and half unaligned loads (-> EvenAligned case) + * - otherwise perform unaligned loads only (-> NoneAligned case) + * - otherwise, if the register size is 4 scalars (e.g., SSE with float) then + * - one over 4 consecutive columns is guaranteed to be aligned with the result vector, + * perform simple aligned loads for this column and aligned loads plus re-alignment for the other. (-> FirstAligned case) + * // this re-alignment is done by the palign function implemented for SSE in Eigen/src/Core/arch/SSE/PacketMath.h + * - otherwise, + * // if we get here, this means the register size is greater than 4 (e.g., AVX with floats), + * // we currently fall back to the NoneAligned case + * + * The same reasoning apply for the transposed case. + * + * The last case (PacketSize>4) could probably be improved by generalizing the FirstAligned case, but since we do not support AVX yet... + * One might also wonder why in the EvenAligned case we perform unaligned loads instead of using the aligned-loads plus re-alignment + * strategy as in the FirstAligned case. The reason is that we observed that unaligned loads on a 8 byte boundary are not too slow + * compared to unaligned loads on a 4 byte boundary. + * */ template struct general_matrix_vector_product From 1b8394f71f613b60525976a8240e30bc3ebdd4dc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 21 Aug 2013 15:28:53 +0200 Subject: [PATCH 0112/1103] Fix compilation with ICC/MSVC combo --- Eigen/src/SparseCore/SparseBlock.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 0b3e193db..b6a2d232c 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -66,6 +66,8 @@ public: typename XprType::Nested m_matrix; Index m_outerStart; const internal::variable_if_dynamic m_outerSize; + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) }; From d1c48f16061a60d92ab6f4df1e877f20c17d78d0 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Wed, 21 Aug 2013 14:34:48 +0100 Subject: [PATCH 0113/1103] BDCSVD: Use HouseholderSeq directly. --- unsupported/Eigen/src/SVD/BDCSVD.h | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/src/SVD/BDCSVD.h b/unsupported/Eigen/src/SVD/BDCSVD.h index 3c41e4e46..d5f0a3f21 100644 --- a/unsupported/Eigen/src/SVD/BDCSVD.h +++ b/unsupported/Eigen/src/SVD/BDCSVD.h @@ -203,7 +203,8 @@ private: void deflation43(Index firstCol, Index shift, Index i, Index size); void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size); void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift); - void copyUV(MatrixX householderU, MatrixX houseHolderV); + void copyUV(const typename internal::UpperBidiagonalization::HouseholderUSequenceType& householderU, + const typename internal::UpperBidiagonalization::HouseholderVSequenceType& householderV); protected: MatrixXr m_naiveU, m_naiveV; @@ -281,7 +282,7 @@ BDCSVD::compute(const MatrixType& matrix, unsigned int computationOp if (isTranspose) copy = matrix.adjoint(); else copy = matrix; - internal::UpperBidiagonalization bid(copy); + internal::UpperBidiagonalization bid(copy); //**** step 2 Divide m_computed.topRows(this->m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose(); @@ -309,9 +310,10 @@ BDCSVD::compute(const MatrixType& matrix, unsigned int computationOp }// end compute -// TODO: this function should accept householder sequences to save converting them to matrix template -void BDCSVD::copyUV(MatrixX householderU, MatrixX householderV){ +void BDCSVD::copyUV(const typename internal::UpperBidiagonalization::HouseholderUSequenceType& householderU, + const typename internal::UpperBidiagonalization::HouseholderVSequenceType& householderV) +{ // Note exchange of U and V: m_matrixU is set from m_naiveV and vice versa if (this->computeU()){ Index Ucols = this->m_computeThinU ? this->m_nonzeroSingularValues : householderU.cols(); From a147500dee85092fbfbf755123553f412a218ec2 Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Sun, 25 Aug 2013 18:00:28 +0900 Subject: [PATCH 0114/1103] Added smart_memmove with support of non-POD scalars (e.g. needed in SparseBlock.h). --- Eigen/src/Core/util/Memory.h | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 451535a0c..3a357e079 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -6,6 +6,7 @@ // Copyright (C) 2009 Kenneth Riddile // Copyright (C) 2010 Hauke Heibel // Copyright (C) 2010 Thomas Capricelli +// Copyright (C) 2013 Pavel Holoborodko // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -514,6 +515,34 @@ template struct smart_copy_helper { { std::copy(start, end, target); } }; +// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise. +template struct smart_memmove_helper; + +template void smart_memmove(const T* start, const T* end, T* target) +{ + smart_memmove_helper::RequireInitialization>::run(start, end, target); +} + +template struct smart_memmove_helper { + static inline void run(const T* start, const T* end, T* target) + { std::memmove(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); } +}; + +template struct smart_memmove_helper { + static inline void run(const T* start, const T* end, T* target) + { + if (uintptr_t(target) < uintptr_t(start)) + { + std::copy(start, end, target); + } + else + { + std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T); + std::copy_backward(start, end, target + count); + } + } +}; + /***************************************************************************** *** Implementation of runtime stack allocation (falling back to malloc) *** From 1472f4bc61192e4ffd06c6baf419747d7cc4c90b Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Sun, 25 Aug 2013 18:02:07 +0900 Subject: [PATCH 0115/1103] Fixed bug #647 by using smart_copy instead of bitwise memcpy. --- Eigen/src/SparseCore/CompressedStorage.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index 3321fab4a..10d516a66 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -51,8 +51,8 @@ class CompressedStorage CompressedStorage& operator=(const CompressedStorage& other) { resize(other.size()); - memcpy(m_values, other.m_values, m_size * sizeof(Scalar)); - memcpy(m_indices, other.m_indices, m_size * sizeof(Index)); + internal::smart_copy(other.m_values, other.m_values + m_size, m_values); + internal::smart_copy(other.m_indices, other.m_indices + m_size, m_indices); return *this; } From e6462c2ce3e2d871892bae13f795d56afc7ec109 Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Sun, 25 Aug 2013 18:03:49 +0900 Subject: [PATCH 0116/1103] Switched to smart_copy to support non-trivial scalar types --- Eigen/src/SparseCore/SparseMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 970c6be71..4dad50562 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -711,7 +711,7 @@ class SparseMatrix initAssignment(other); if(other.isCompressed()) { - memcpy(m_outerIndex, other.m_outerIndex, (m_outerSize+1)*sizeof(Index)); + internal::smart_copy(other.m_outerIndex, other.m_outerIndex + m_outerSize + 1, m_outerIndex); m_data = other.m_data; } else From 41321e436606fa9399cf9b89586b04c2ba95c23e Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Sun, 25 Aug 2013 18:12:15 +0900 Subject: [PATCH 0117/1103] Replaced memcpy & memmove to smart_* alternatives for non-POD scalar types --- Eigen/src/SparseCore/SparseBlock.h | 29 +++++++++++++++-------------- 1 file changed, 15 insertions(+), 14 deletions(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index b6a2d232c..efd44284f 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -66,13 +66,14 @@ public: typename XprType::Nested m_matrix; Index m_outerStart; const internal::variable_if_dynamic m_outerSize; - + + public: EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) }; /*************************************************************************** -* specialisation for SparseMatrix +* specialization for SparseMatrix ***************************************************************************/ template @@ -125,7 +126,7 @@ public: { typedef typename internal::remove_all::type _NestedMatrixType; _NestedMatrixType& matrix = const_cast<_NestedMatrixType&>(m_matrix);; - // This assignement is slow if this vector set is not empty + // This assignment is slow if this vector set is not empty // and/or it is not at the end of the nonzeros of the underlying matrix. // 1 - eval to a temporary to avoid transposition and/or aliasing issues @@ -134,7 +135,7 @@ public: // 2 - let's check whether there is enough allocated memory Index nnz = tmp.nonZeros(); Index start = m_outerStart==0 ? 0 : matrix.outerIndexPtr()[m_outerStart]; // starting position of the current block - Index end = m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]; // ending posiiton of the current block + Index end = m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]; // ending position of the current block Index block_size = end - start; // available room in the current block Index tail_size = m_matrix.outerIndexPtr()[m_matrix.outerSize()] - end; @@ -147,14 +148,14 @@ public: // realloc manually to reduce copies typename SparseMatrixType::Storage newdata(m_matrix.data().allocatedSize() - block_size + nnz); - std::memcpy(&newdata.value(0), &m_matrix.data().value(0), start*sizeof(Scalar)); - std::memcpy(&newdata.index(0), &m_matrix.data().index(0), start*sizeof(Index)); + internal::smart_copy(&m_matrix.data().value(0), &m_matrix.data().value(0) + start, &newdata.value(0)); + internal::smart_copy(&m_matrix.data().index(0), &m_matrix.data().index(0) + start, &newdata.index(0)); - std::memcpy(&newdata.value(start), &tmp.data().value(0), nnz*sizeof(Scalar)); - std::memcpy(&newdata.index(start), &tmp.data().index(0), nnz*sizeof(Index)); + internal::smart_copy(&tmp.data().value(0), &tmp.data().value(0) + nnz, &newdata.value(start)); + internal::smart_copy(&tmp.data().index(0), &tmp.data().index(0) + nnz, &newdata.index(start)); - std::memcpy(&newdata.value(start+nnz), &matrix.data().value(end), tail_size*sizeof(Scalar)); - std::memcpy(&newdata.index(start+nnz), &matrix.data().index(end), tail_size*sizeof(Index)); + internal::smart_copy(&matrix.data().value(end), &matrix.data().value(end) + tail_size, &newdata.value(start+nnz)); + internal::smart_copy(&matrix.data().index(end), &matrix.data().index(end) + tail_size, &newdata.index(start+nnz)); newdata.resize(m_matrix.outerIndexPtr()[m_matrix.outerSize()] - block_size + nnz); @@ -165,11 +166,11 @@ public: // no need to realloc, simply copy the tail at its respective position and insert tmp matrix.data().resize(start + nnz + tail_size); - std::memmove(&matrix.data().value(start+nnz), &matrix.data().value(end), tail_size*sizeof(Scalar)); - std::memmove(&matrix.data().index(start+nnz), &matrix.data().index(end), tail_size*sizeof(Index)); + internal::smart_memmove(&matrix.data().value(end), &matrix.data().value(end) + tail_size, &matrix.data().value(start + nnz)); + internal::smart_memmove(&matrix.data().index(end), &matrix.data().index(end) + tail_size, &matrix.data().index(start + nnz)); - std::memcpy(&matrix.data().value(start), &tmp.data().value(0), nnz*sizeof(Scalar)); - std::memcpy(&matrix.data().index(start), &tmp.data().index(0), nnz*sizeof(Index)); + internal::smart_copy(&tmp.data().value(0), &tmp.data().value(0) + nnz, &matrix.value(start)); + internal::smart_copy(&tmp.data().index(0), &tmp.data().index(0) + nnz, &matrix.index(start)); } // update innerNonZeros From d2c4f4ab21bd39485cb2acf3658aac1e2628d031 Mon Sep 17 00:00:00 2001 From: Pavel Holoborodko Date: Mon, 26 Aug 2013 00:22:18 +0900 Subject: [PATCH 0118/1103] Updated mpfr::mpreal. Move semantic support, RVO, other new features --- unsupported/test/mpreal/mpreal.h | 199 ++++++++++++++++++++----------- 1 file changed, 127 insertions(+), 72 deletions(-) diff --git a/unsupported/test/mpreal/mpreal.h b/unsupported/test/mpreal/mpreal.h index ef0a6a9f0..75bda7db7 100644 --- a/unsupported/test/mpreal/mpreal.h +++ b/unsupported/test/mpreal/mpreal.h @@ -5,7 +5,7 @@ Project homepage: http://www.holoborodko.com/pavel/mpfr Contact e-mail: pavel@holoborodko.com - Copyright (c) 2008-2012 Pavel Holoborodko + Copyright (c) 2008-2013 Pavel Holoborodko Contributors: Dmitriy Gubanov, Konstantin Holoborodko, Brian Gladman, @@ -65,6 +65,7 @@ #include #include #include +#include #include // Options @@ -82,6 +83,20 @@ #define IsInf(x) std::isinf(x) // GNU C/C++ (and/or other compilers), just hope for C99 conformance #endif +// Detect support for r-value references (move semantic). Borrowed from Eigen. +// A Clang feature extension to determine compiler features. +// We use it to determine 'cxx_rvalue_references' +#ifndef __has_feature +# define __has_feature(x) 0 +#endif + +#if (__has_feature(cxx_rvalue_references) || \ + defined(__GXX_EXPERIMENTAL_CXX0X__) || \ + (defined(_MSC_VER) && _MSC_VER >= 1600)) +#define MPREAL_HAVE_MOVE_SUPPORT +#endif + +// Detect available 64-bit capabilities #if defined(MPREAL_HAVE_INT64_SUPPORT) #define MPFR_USE_INTMAX_T // Should be defined before mpfr.h @@ -111,7 +126,7 @@ #endif #if defined(MPREAL_HAVE_MSVC_DEBUGVIEW) && defined(_MSC_VER) && defined(_DEBUG) -#define MPREAL_MSVC_DEBUGVIEW_CODE DebugView = toString(); + #define MPREAL_MSVC_DEBUGVIEW_CODE DebugView = toString(); #define MPREAL_MSVC_DEBUGVIEW_DATA std::string DebugView; #else #define MPREAL_MSVC_DEBUGVIEW_CODE @@ -149,7 +164,6 @@ public: // Constructors && type conversions mpreal(); mpreal(const mpreal& u); - mpreal(const mpfr_t u); mpreal(const mpf_t u); mpreal(const mpz_t u, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t mode = mpreal::get_default_rnd()); mpreal(const mpq_t u, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t mode = mpreal::get_default_rnd()); @@ -159,6 +173,10 @@ public: mpreal(const unsigned int u, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t mode = mpreal::get_default_rnd()); mpreal(const long int u, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t mode = mpreal::get_default_rnd()); mpreal(const int u, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t mode = mpreal::get_default_rnd()); + + // Construct mpreal from mpfr_t structure. + // shared = true allows to avoid deep copy, so that mpreal and 'u' shared same data & pointers. + mpreal(const mpfr_t u, bool shared = false); #if defined (MPREAL_HAVE_INT64_SUPPORT) mpreal(const uint64_t u, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t mode = mpreal::get_default_rnd()); @@ -170,6 +188,11 @@ public: ~mpreal(); +#ifdef MPREAL_HAVE_MOVE_SUPPORT + mpreal& operator=(mpreal&& v); + mpreal(mpreal&& u); +#endif + // Operations // = // +, -, *, /, ++, --, <<, >> @@ -415,6 +438,8 @@ public: friend const mpreal digamma (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); friend const mpreal ai (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); friend const mpreal urandom (gmp_randstate_t& state, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); // use gmp_randinit_default() to init state, gmp_randclear() to clear + friend const mpreal grandom (gmp_randstate_t& state, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); // use gmp_randinit_default() to init state, gmp_randclear() to clear + friend const mpreal grandom (unsigned int seed = 0); #endif // Uniformly distributed random number generation in [0,1] using @@ -422,7 +447,7 @@ public: // Use parameter to setup seed, e.g.: random((unsigned)time(NULL)) // Check urandom() for more precise control. friend const mpreal random(unsigned int seed = 0); - + // Exponent and mantissa manipulation friend const mpreal frexp(const mpreal& v, mp_exp_t* exp); friend const mpreal ldexp(const mpreal& v, mp_exp_t exp); @@ -531,12 +556,15 @@ private: // Human friendly Debug Preview in Visual Studio. // Put one of these lines: // - // mpfr::mpreal= ; Show value only + // mpfr::mpreal= ; Show value only // mpfr::mpreal=, bits ; Show value & precision // // at the beginning of // [Visual Studio Installation Folder]\Common7\Packages\Debugger\autoexp.dat MPREAL_MSVC_DEBUGVIEW_DATA + + // "Smart" resources deallocation. Checks if instance initialized before deletion. + void clear(::mpfr_ptr); }; ////////////////////////////////////////////////////////////////////////// @@ -565,10 +593,36 @@ inline mpreal::mpreal(const mpreal& u) MPREAL_MSVC_DEBUGVIEW_CODE; } -inline mpreal::mpreal(const mpfr_t u) +#ifdef MPREAL_HAVE_MOVE_SUPPORT +inline mpreal::mpreal(mpreal&& other) { - mpfr_init2(mp,mpfr_get_prec(u)); - mpfr_set(mp,u,mpreal::get_default_rnd()); + mpfr_ptr()->_mpfr_d = 0; // make sure "other" holds no pinter to actual data + + mpfr_swap(mpfr_ptr(), other.mpfr_ptr()); + + MPREAL_MSVC_DEBUGVIEW_CODE; +} + +inline mpreal& mpreal::operator=(mpreal&& other) +{ + mpfr_swap(mpfr_ptr(), other.mpfr_ptr()); + + MPREAL_MSVC_DEBUGVIEW_CODE; + return *this; +} +#endif + +inline mpreal::mpreal(const mpfr_t u, bool shared) +{ + if(shared) + { + std::memcpy(mp, u, sizeof(mpfr_t)); + } + else + { + mpfr_init2(mp, mpfr_get_prec(u)); + mpfr_set (mp, u, mpreal::get_default_rnd()); + } MPREAL_MSVC_DEBUGVIEW_CODE; } @@ -688,9 +742,15 @@ inline mpreal::mpreal(const std::string& s, mp_prec_t prec, int base, mp_rnd_t m MPREAL_MSVC_DEBUGVIEW_CODE; } +inline void mpreal::clear(::mpfr_ptr x) +{ + // Use undocumented field in mpfr_t structure to check if it was initialized + if(0 != x->_mpfr_d) mpfr_clear(x); +} + inline mpreal::~mpreal() { - mpfr_clear(mp); + clear(mpfr_ptr()); } // internal namespace needed for template magic @@ -864,8 +924,8 @@ inline mpreal& mpreal::operator=(const mpreal& v) mp_prec_t vp = mpfr_get_prec(v.mp); if(tp != vp){ - mpfr_clear(mp); - mpfr_init2(mp, vp); + clear(mpfr_ptr()); + mpfr_init2(mpfr_ptr(), vp); } mpfr_set(mp, v.mp, mpreal::get_default_rnd()); @@ -974,7 +1034,7 @@ inline mpreal& mpreal::operator=(const char* s) MPREAL_MSVC_DEBUGVIEW_CODE; } - mpfr_clear(t); + clear(t); return *this; } @@ -997,7 +1057,7 @@ inline mpreal& mpreal::operator=(const std::string& s) MPREAL_MSVC_DEBUGVIEW_CODE; } - mpfr_clear(t); + clear(t); return *this; } @@ -1202,7 +1262,7 @@ inline const mpreal mpreal::operator-()const inline const mpreal operator-(const mpreal& a, const mpreal& b) { - mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpreal c(0, std::max(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); mpfr_sub(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); return c; } @@ -1319,7 +1379,7 @@ inline mpreal& mpreal::operator*=(const int v) inline const mpreal operator*(const mpreal& a, const mpreal& b) { - mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpreal c(0, std::max(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); mpfr_mul(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); return c; } @@ -1395,7 +1455,7 @@ inline mpreal& mpreal::operator/=(const int v) inline const mpreal operator/(const mpreal& a, const mpreal& b) { - mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpreal c(0, std::max(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); mpfr_div(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); return c; } @@ -1922,17 +1982,17 @@ inline mpreal maxval(mp_prec_t prec) inline bool isEqualUlps(const mpreal& a, const mpreal& b, int maxUlps) { - return abs(a - b) <= machine_epsilon((max)(abs(a), abs(b))) * maxUlps; + return abs(a - b) <= machine_epsilon((max)(abs(a), abs(b))) * maxUlps; } inline bool isEqualFuzzy(const mpreal& a, const mpreal& b, const mpreal& eps) { - return abs(a - b) <= (min)(abs(a), abs(b)) * eps; + return abs(a - b) <= eps; } inline bool isEqualFuzzy(const mpreal& a, const mpreal& b) { - return isEqualFuzzy(a, b, machine_epsilon((min)(abs(a), abs(b)))); + return isEqualFuzzy(a, b, machine_epsilon((max)(1, (min)(abs(a), abs(b))))); } inline const mpreal modf(const mpreal& v, mpreal& n) @@ -2048,12 +2108,12 @@ inline const mpreal dim(const mpreal& a, const mpreal& b, mp_rnd_t r) inline int cmpabs(const mpreal& a,const mpreal& b) { - return mpfr_cmpabs(a.mp,b.mp); + return mpfr_cmpabs(a.mpfr_ptr(), b.mpfr_srcptr()); } inline int sin_cos(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode) { - return mpfr_sin_cos(s.mp,c.mp,v.mp,rnd_mode); + return mpfr_sin_cos(s.mpfr_ptr(), c.mpfr_ptr(), v.mpfr_srcptr(), rnd_mode); } inline const mpreal sqrt (const long double v, mp_rnd_t rnd_mode) { return sqrt(mpreal(v),rnd_mode); } @@ -2074,9 +2134,9 @@ inline const mpreal tan (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FU inline const mpreal sec (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(sec ); } inline const mpreal csc (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(csc ); } inline const mpreal cot (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(cot ); } -inline const mpreal acos (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(acos); } -inline const mpreal asin (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(asin); } -inline const mpreal atan (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(atan); } +inline const mpreal acos (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(acos ); } +inline const mpreal asin (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(asin ); } +inline const mpreal atan (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(atan ); } inline const mpreal acot (const mpreal& v, mp_rnd_t r) { return atan (1/v, r); } inline const mpreal asec (const mpreal& v, mp_rnd_t r) { return acos (1/v, r); } @@ -2110,68 +2170,36 @@ inline const mpreal bessely1(const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_ inline const mpreal atan2 (const mpreal& y, const mpreal& x, mp_rnd_t rnd_mode) { - mpreal a; - mp_prec_t yp, xp; - - yp = y.get_prec(); - xp = x.get_prec(); - - a.set_prec(yp>xp?yp:xp); - - mpfr_atan2(a.mp, y.mp, x.mp, rnd_mode); - + mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); + mpfr_atan2(a.mpfr_ptr(), y.mpfr_srcptr(), x.mpfr_srcptr(), rnd_mode); return a; } inline const mpreal hypot (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) { - mpreal a; - mp_prec_t yp, xp; - - yp = y.get_prec(); - xp = x.get_prec(); - - a.set_prec(yp>xp?yp:xp); - - mpfr_hypot(a.mp, x.mp, y.mp, rnd_mode); - + mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); + mpfr_hypot(a.mpfr_ptr(), x.mpfr_srcptr(), y.mpfr_srcptr(), rnd_mode); return a; } inline const mpreal remainder (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) { - mpreal a; - mp_prec_t yp, xp; - - yp = y.get_prec(); - xp = x.get_prec(); - - a.set_prec(yp>xp?yp:xp); - - mpfr_remainder(a.mp, x.mp, y.mp, rnd_mode); - + mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); + mpfr_remainder(a.mpfr_ptr(), x.mpfr_srcptr(), y.mpfr_srcptr(), rnd_mode); return a; } inline const mpreal remquo (long* q, const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) { - mpreal a; - mp_prec_t yp, xp; - - yp = y.get_prec(); - xp = x.get_prec(); - - a.set_prec(yp>xp?yp:xp); - - mpfr_remquo(a.mp,q, x.mp, y.mp, rnd_mode); - + mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); + mpfr_remquo(a.mpfr_ptr(),q, x.mpfr_srcptr(), y.mpfr_srcptr(), rnd_mode); return a; } inline const mpreal fac_ui (unsigned long int v, mp_prec_t prec, mp_rnd_t rnd_mode) { mpreal x(0, prec); - mpfr_fac_ui(x.mp,v,rnd_mode); + mpfr_fac_ui(x.mpfr_ptr(),v,rnd_mode); return x; } @@ -2181,7 +2209,7 @@ inline const mpreal lgamma (const mpreal& v, int *signp, mp_rnd_t rnd_mode) mpreal x(v); int tsignp; - if(signp) mpfr_lgamma(x.mp,signp,v.mp,rnd_mode); + if(signp) mpfr_lgamma(x.mp, signp,v.mp,rnd_mode); else mpfr_lgamma(x.mp,&tsignp,v.mp,rnd_mode); return x; @@ -2365,7 +2393,7 @@ inline const mpreal const_catalan (mp_prec_t p, mp_rnd_t r) return x; } -inline const mpreal const_infinity (int sign, mp_prec_t p, mp_rnd_t /*r*/) +inline const mpreal const_infinity (int sign, mp_prec_t p, mp_rnd_t r) { mpreal x(0, p); mpfr_set_inf(x.mpfr_ptr(), sign); @@ -2465,6 +2493,14 @@ inline const mpreal urandom (gmp_randstate_t& state, mp_rnd_t rnd_mode) mpfr_urandom(x.mp,state,rnd_mode); return x; } + +inline const mpreal grandom (gmp_randstate_t& state, mp_rnd_t rnd_mode) +{ + mpreal x; + mpfr_grandom(x.mp, NULL, state, rnd_mode); + return x; +} + #endif #if (MPFR_VERSION <= MPFR_VERSION_NUM(2,4,2)) @@ -2504,6 +2540,25 @@ inline const mpreal random(unsigned int seed) } +#if (MPFR_VERSION >= MPFR_VERSION_NUM(3,0,0)) +inline const mpreal grandom(unsigned int seed) +{ + static gmp_randstate_t state; + static bool isFirstTime = true; + + if(isFirstTime) + { + gmp_randinit_default(state); + gmp_randseed_ui(state,0); + isFirstTime = false; + } + + if(seed != 0) gmp_randseed_ui(state,seed); + + return mpfr::grandom(state); +} +#endif + ////////////////////////////////////////////////////////////////////////// // Set/Get global properties inline void mpreal::set_default_prec(mp_prec_t prec) @@ -2859,10 +2914,10 @@ namespace std switch (r) { - case MPFR_RNDN: return round_to_nearest; - case MPFR_RNDZ: return round_toward_zero; - case MPFR_RNDU: return round_toward_infinity; - case MPFR_RNDD: return round_toward_neg_infinity; + case GMP_RNDN: return round_to_nearest; + case GMP_RNDZ: return round_toward_zero; + case GMP_RNDU: return round_toward_infinity; + case GMP_RNDD: return round_toward_neg_infinity; default: return round_indeterminate; } } @@ -2881,7 +2936,7 @@ namespace std { mp_rnd_t r = mpfr::mpreal::get_default_rnd(); - if(r == MPFR_RNDN) return mpfr::mpreal(0.5, precision); + if(r == GMP_RNDN) return mpfr::mpreal(0.5, precision); else return mpfr::mpreal(1.0, precision); } From 5864e3fbd5d8fa4750b4fbf3363813a0aa63ff82 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 27 Aug 2013 07:23:31 +0200 Subject: [PATCH 0119/1103] Implement a blocked upper-bidiagonalization algorithm. The computeUnblocked function is currently for benchmarking purpose. --- Eigen/src/SVD/UpperBidiagonalization.h | 313 ++++++++++++++++++++++--- 1 file changed, 283 insertions(+), 30 deletions(-) diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 587de37a5..2a55d550d 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -31,7 +31,7 @@ template class UpperBidiagonalization typedef typename MatrixType::Index Index; typedef Matrix RowVectorType; typedef Matrix ColVectorType; - typedef BandMatrix BidiagonalType; + typedef BandMatrix BidiagonalType; typedef Matrix DiagVectorType; typedef Matrix SuperDiagVectorType; typedef HouseholderSequence< @@ -61,6 +61,7 @@ template class UpperBidiagonalization } UpperBidiagonalization& compute(const MatrixType& matrix); + UpperBidiagonalization& computeUnblocked(const MatrixType& matrix); const MatrixType& householder() const { return m_householder; } const BidiagonalType& bidiagonal() const { return m_bidiagonal; } @@ -85,45 +86,297 @@ template class UpperBidiagonalization bool m_isInitialized; }; -template -UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::compute(const _MatrixType& matrix) -{ - Index rows = matrix.rows(); - Index cols = matrix.cols(); - - eigen_assert(rows >= cols && "UpperBidiagonalization is only for matrices satisfying rows>=cols."); - - m_householder = matrix; - ColVectorType temp(rows); +/** \internal + * Helper routine for the block bidiagonal reduction. + * This function reduces to bidiagonal form the left \c rows x \a blockSize vertical + * and \a blockSize x \c cols horizontal panels of the \a A. The bottom-right block + * is left unchanged, and the Arices X and Y. + */ +template +void upperbidiagonalization_inplace_helper(MatrixType& A, + typename MatrixType::RealScalar *diagonal, + typename MatrixType::RealScalar *upper_diagonal, + typename MatrixType::Index bs, + Ref > X, + Ref > Y) +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + typedef Ref > SubColumnType; + typedef Ref, 0, InnerStride<> > SubRowType; + typedef Ref > SubMatType; + + Index brows = A.rows(); + Index bcols = A.cols(); + + Scalar tau_u, tau_u_prev(0), tau_v; + + for(Index k = 0; k < bs; ++k) + { + Index remainingRows = brows - k; + Index remainingSizeInBlock = bs - k - 1; + Index remainingCols = bcols - k - 1; + + SubMatType X_k1( X.block(k,0, remainingRows,k) ); + SubMatType V_k1( A.block(k,0, remainingRows,k) ); + + // 1 - update the k-th column of A + SubColumnType v_k = A.col(k).tail(remainingRows); + v_k -= V_k1 * Y.row(k).head(k).adjoint(); + if(k) v_k -= X_k1 * A.col(k).head(k); + + // 2 - construct left Householder transform in-place + v_k.makeHouseholderInPlace(tau_v, diagonal[k]); + + if(k+10) A.coeffRef(k-1,k) = tau_u_prev; + tau_u_prev = tau_u; + } + else + A.coeffRef(k-1,k) = tau_u_prev; + + A.coeffRef(k,k) = tau_v; + } + + if(bsbs && brows>bs) + { + SubMatType A22( A.bottomRightCorner(brows-bs,bcols-bs) ); + SubMatType A21( A.block(bs,0, brows-bs,bs) ); + SubMatType A12( A.block(0,bs, bs, bcols-bs) ); + Scalar tmp = A12(bs-1,0); + A12(bs-1,0) = 1; + A22.noalias() -= A21 * Y.topLeftCorner(bcols,bs).bottomRows(bcols-bs).adjoint(); + A22.noalias() -= X.topLeftCorner(brows,bs).bottomRows(brows-bs) * A12; + A12(bs-1,0) = tmp; + } +} + +/** \internal + * + * Implementation of a block-bidiagonal reduction. + * It is based on the following paper: + * The Design of a Parallel Dense Linear Algebra Software Library: Reduction to Hessenberg, Tridiagonal, and Bidiagonal Form. + * by Jaeyoung Choi, Jack J. Dongarra, David W. Walker. (1995) + * section 3.3 + */ +template +void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagonal, + typename MatrixType::Index maxBlockSize=32, + typename MatrixType::Scalar* tempData = 0) +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + typedef Block BlockType; + + Index rows = A.rows(); + Index cols = A.cols(); + Index size = (std::min)(rows, cols); + + typedef Matrix TempType; + TempType tempVector; + if(tempData==0) + { + tempVector.resize(cols); + tempData = tempVector.data(); + } + + Matrix X(rows,maxBlockSize), Y(cols, maxBlockSize); + X.setOnes(); + Y.setOnes(); + + Index blockSize = (std::min)(maxBlockSize,size); + + + + Index k = 0; + for (k = 0; k < size; k += blockSize) + { + Index bs = (std::min)(size-k,blockSize); // actual size of the block + Index tcols = cols - k - bs; // trailing columns + Index trows = rows - k - bs; // trailing rows + Index brows = rows - k; // rows of the block + Index bcols = cols - k; // columns of the block + + // partition the matrix A: + // + // | A00 A01 A02 | + // | | + // A = | A10 A11 A12 | + // | | + // | A20 A21 A22 | + // + // where A11 is a bs x bs diagonal block, + // and performs the bidiagonalization of A11, A21, A12, without updating A22. + // + // A22 will be updated in a second stage using matrices X and Y and level 3 operations: + // A22 -= V*Y^T - X*U^T + // where V and U contains the left and right Householder vectors + // + // Finally, the algorithm continue on the updated A22. + + + + // Let: + // | A11 A12 | + // B = | | + // | A21 A22 | + BlockType B = A.block(k,k,brows,bcols); + upperbidiagonalization_inplace_helper(B, + &(bidiagonal.template diagonal<0>().coeffRef(k)), + &(bidiagonal.template diagonal<1>().coeffRef(k)), + bs, + X.topLeftCorner(brows,bs), + Y.topLeftCorner(bcols,bs) + ); + } +} + +// Standard upper bidiagonalization without fancy optimizations +// This version should be faster for small matrix size +template +void upperbidiagonalization_inplace_unblocked(MatrixType& mat, BidiagType& bidiagonal, + typename MatrixType::Scalar* tempData = 0) +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + + Index rows = mat.rows(); + Index cols = mat.cols(); + Index size = (std::min)(rows, cols); + + typedef Matrix TempType; + TempType tempVector; + if(tempData==0) + { + tempVector.resize(rows); + tempData = tempVector.data(); + } for (Index k = 0; /* breaks at k==cols-1 below */ ; ++k) { Index remainingRows = rows - k; Index remainingCols = cols - k - 1; - // construct left householder transform in-place in m_householder - m_householder.col(k).tail(remainingRows) - .makeHouseholderInPlace(m_householder.coeffRef(k,k), - m_bidiagonal.template diagonal<0>().coeffRef(k)); - // apply householder transform to remaining part of m_householder on the left - m_householder.bottomRightCorner(remainingRows, remainingCols) - .applyHouseholderOnTheLeft(m_householder.col(k).tail(remainingRows-1), - m_householder.coeff(k,k), - temp.data()); + // construct left householder transform in-place in A + mat.col(k).tail(remainingRows) + .makeHouseholderInPlace(mat.coeffRef(k,k), bidiagonal.template diagonal<0>().coeffRef(k)); + // apply householder transform to remaining part of A on the left + mat.bottomRightCorner(remainingRows, remainingCols) + .applyHouseholderOnTheLeft(mat.col(k).tail(remainingRows-1), mat.coeff(k,k), tempData); if(k == cols-1) break; - - // construct right householder transform in-place in m_householder - m_householder.row(k).tail(remainingCols) - .makeHouseholderInPlace(m_householder.coeffRef(k,k+1), - m_bidiagonal.template diagonal<1>().coeffRef(k)); - // apply householder transform to remaining part of m_householder on the left - m_householder.bottomRightCorner(remainingRows-1, remainingCols) - .applyHouseholderOnTheRight(m_householder.row(k).tail(remainingCols-1).transpose(), - m_householder.coeff(k,k+1), - temp.data()); + + // construct right householder transform in-place in mat + mat.row(k).tail(remainingCols) + .makeHouseholderInPlace(mat.coeffRef(k,k+1), bidiagonal.template diagonal<1>().coeffRef(k)); + // apply householder transform to remaining part of mat on the left + mat.bottomRightCorner(remainingRows-1, remainingCols) + .applyHouseholderOnTheRight(mat.row(k).tail(remainingCols-1).transpose(), mat.coeff(k,k+1), tempData); } +} + +template +UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::computeUnblocked(const _MatrixType& matrix) +{ + Index rows = matrix.rows(); + Index cols = matrix.cols(); + + eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols."); + + m_householder = matrix; + + ColVectorType temp(rows); + + upperbidiagonalization_inplace_unblocked(m_householder, m_bidiagonal, temp.data()); + + MatrixType A = matrix; + BidiagonalType B(cols,cols); + upperbidiagonalization_inplace_blocked(A, B, 8); + + std::cout << (m_householder-A)/*.maxCoeff()*/ << "\n\n"; + std::cout << m_householder << "\n\n" + << m_bidiagonal.template diagonal<0>() << "\n" + << m_bidiagonal.template diagonal<1>() << "\n\n"; + std::cout << A << "\n\n" + << B.template diagonal<0>() << "\n" + << B.template diagonal<1>() << "\n\n"; + + m_isInitialized = true; + return *this; +} + +template +UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::compute(const _MatrixType& matrix) +{ + Index rows = matrix.rows(); + Index cols = matrix.cols(); + + eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols."); + + m_householder = matrix; + upperbidiagonalization_inplace_blocked(m_householder, m_bidiagonal); + m_isInitialized = true; return *this; } From 94a7a1ec00d2f26fc301a1e81a1d7d3aa07b57e9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 27 Aug 2013 13:47:15 +0200 Subject: [PATCH 0120/1103] Use unblocked version if the matrix is too small, plus some cleaning. --- Eigen/src/SVD/UpperBidiagonalization.h | 215 ++++++++++++------------- 1 file changed, 106 insertions(+), 109 deletions(-) diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 2a55d550d..6cac3af92 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -86,15 +86,71 @@ template class UpperBidiagonalization bool m_isInitialized; }; +// Standard upper bidiagonalization without fancy optimizations +// This version should be faster for small matrix size +template +void upperbidiagonalization_inplace_unblocked(MatrixType& mat, + typename MatrixType::RealScalar *diagonal, + typename MatrixType::RealScalar *upper_diagonal, + typename MatrixType::Scalar* tempData = 0) +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + + Index rows = mat.rows(); + Index cols = mat.cols(); + Index size = (std::min)(rows, cols); + + typedef Matrix TempType; + TempType tempVector; + if(tempData==0) + { + tempVector.resize(rows); + tempData = tempVector.data(); + } + + for (Index k = 0; /* breaks at k==cols-1 below */ ; ++k) + { + Index remainingRows = rows - k; + Index remainingCols = cols - k - 1; + + // construct left householder transform in-place in A + mat.col(k).tail(remainingRows) + .makeHouseholderInPlace(mat.coeffRef(k,k), diagonal[k]); + // apply householder transform to remaining part of A on the left + mat.bottomRightCorner(remainingRows, remainingCols) + .applyHouseholderOnTheLeft(mat.col(k).tail(remainingRows-1), mat.coeff(k,k), tempData); + + if(k == cols-1) break; + + // construct right householder transform in-place in mat + mat.row(k).tail(remainingCols) + .makeHouseholderInPlace(mat.coeffRef(k,k+1), upper_diagonal[k]); + // apply householder transform to remaining part of mat on the left + mat.bottomRightCorner(remainingRows-1, remainingCols) + .applyHouseholderOnTheRight(mat.row(k).tail(remainingCols-1).transpose(), mat.coeff(k,k+1), tempData); + } +} /** \internal - * Helper routine for the block bidiagonal reduction. - * This function reduces to bidiagonal form the left \c rows x \a blockSize vertical - * and \a blockSize x \c cols horizontal panels of the \a A. The bottom-right block - * is left unchanged, and the Arices X and Y. + * Helper routine for the block reduction to upper bidiagonal form. + * + * Let's partition the matrix A: + * + * | A00 A01 | + * A = | | + * | A10 A11 | + * + * This function reduces to bidiagonal form the left \c rows x \a blockSize vertical panel [A00/A10] + * and the \a blockSize x \c cols horizontal panel [A00 A01] of the matrix \a A. The bottom-right block A11 + * is updated using matrix-matrix products: + * A22 -= V * Y^T - X * U^T + * where V and U contains the left and right Householder vectors. U and V are stored in A10, and A01 + * respectively, and the update matrices X and Y are computed during the reduction. + * */ template -void upperbidiagonalization_inplace_helper(MatrixType& A, +void upperbidiagonalization_blocked_helper(MatrixType& A, typename MatrixType::RealScalar *diagonal, typename MatrixType::RealScalar *upper_diagonal, typename MatrixType::Index bs, @@ -145,10 +201,10 @@ void upperbidiagonalization_inplace_helper(MatrixType& A, // let's use the begining of column k of Y as a temporary vector SubColumnType tmp( Y.col(k).head(k) ); y_k.noalias() = A.block(k,k+1, remainingRows,remainingCols).adjoint() * v_k; // bottleneck - tmp.noalias() = V_k1.adjoint() * v_k; + tmp.noalias() = V_k1.adjoint() * v_k; y_k.noalias() -= Y_k.leftCols(k) * tmp; - tmp.noalias() = X_k1.adjoint() * v_k; - y_k.noalias() -= U_k1.adjoint() * tmp; + tmp.noalias() = X_k1.adjoint() * v_k; + y_k.noalias() -= U_k1.adjoint() * tmp; y_k *= numext::conj(tau_v); } @@ -201,14 +257,14 @@ void upperbidiagonalization_inplace_helper(MatrixType& A, // update A22 if(bcols>bs && brows>bs) { - SubMatType A22( A.bottomRightCorner(brows-bs,bcols-bs) ); - SubMatType A21( A.block(bs,0, brows-bs,bs) ); - SubMatType A12( A.block(0,bs, bs, bcols-bs) ); - Scalar tmp = A12(bs-1,0); - A12(bs-1,0) = 1; - A22.noalias() -= A21 * Y.topLeftCorner(bcols,bs).bottomRows(bcols-bs).adjoint(); - A22.noalias() -= X.topLeftCorner(brows,bs).bottomRows(brows-bs) * A12; - A12(bs-1,0) = tmp; + SubMatType A11( A.bottomRightCorner(brows-bs,bcols-bs) ); + SubMatType A10( A.block(bs,0, brows-bs,bs) ); + SubMatType A01( A.block(0,bs, bs,bcols-bs) ); + Scalar tmp = A01(bs-1,0); + A01(bs-1,0) = 1; + A11.noalias() -= A10 * Y.topLeftCorner(bcols,bs).bottomRows(bcols-bs).adjoint(); + A11.noalias() -= X.topLeftCorner(brows,bs).bottomRows(brows-bs) * A01; + A01(bs-1,0) = tmp; } } @@ -223,7 +279,7 @@ void upperbidiagonalization_inplace_helper(MatrixType& A, template void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagonal, typename MatrixType::Index maxBlockSize=32, - typename MatrixType::Scalar* tempData = 0) + typename MatrixType::Scalar* /*tempData*/ = 0) { typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; @@ -233,28 +289,14 @@ void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagona Index cols = A.cols(); Index size = (std::min)(rows, cols); - typedef Matrix TempType; - TempType tempVector; - if(tempData==0) - { - tempVector.resize(cols); - tempData = tempVector.data(); - } - - Matrix X(rows,maxBlockSize), Y(cols, maxBlockSize); - X.setOnes(); - Y.setOnes(); - + Matrix X(rows,maxBlockSize); + Matrix Y(cols,maxBlockSize); Index blockSize = (std::min)(maxBlockSize,size); - - Index k = 0; - for (k = 0; k < size; k += blockSize) + for(k = 0; k < size; k += blockSize) { Index bs = (std::min)(size-k,blockSize); // actual size of the block - Index tcols = cols - k - bs; // trailing columns - Index trows = rows - k - bs; // trailing rows Index brows = rows - k; // rows of the block Index bcols = cols - k; // columns of the block @@ -267,72 +309,36 @@ void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagona // | A20 A21 A22 | // // where A11 is a bs x bs diagonal block, - // and performs the bidiagonalization of A11, A21, A12, without updating A22. - // - // A22 will be updated in a second stage using matrices X and Y and level 3 operations: - // A22 -= V*Y^T - X*U^T - // where V and U contains the left and right Householder vectors - // - // Finally, the algorithm continue on the updated A22. - - - - // Let: + // and let: // | A11 A12 | // B = | | // | A21 A22 | + BlockType B = A.block(k,k,brows,bcols); - upperbidiagonalization_inplace_helper(B, - &(bidiagonal.template diagonal<0>().coeffRef(k)), - &(bidiagonal.template diagonal<1>().coeffRef(k)), - bs, - X.topLeftCorner(brows,bs), - Y.topLeftCorner(bcols,bs) - ); - } -} - -// Standard upper bidiagonalization without fancy optimizations -// This version should be faster for small matrix size -template -void upperbidiagonalization_inplace_unblocked(MatrixType& mat, BidiagType& bidiagonal, - typename MatrixType::Scalar* tempData = 0) -{ - typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - - Index rows = mat.rows(); - Index cols = mat.cols(); - Index size = (std::min)(rows, cols); - - typedef Matrix TempType; - TempType tempVector; - if(tempData==0) - { - tempVector.resize(rows); - tempData = tempVector.data(); - } - - for (Index k = 0; /* breaks at k==cols-1 below */ ; ++k) - { - Index remainingRows = rows - k; - Index remainingCols = cols - k - 1; - - // construct left householder transform in-place in A - mat.col(k).tail(remainingRows) - .makeHouseholderInPlace(mat.coeffRef(k,k), bidiagonal.template diagonal<0>().coeffRef(k)); - // apply householder transform to remaining part of A on the left - mat.bottomRightCorner(remainingRows, remainingCols) - .applyHouseholderOnTheLeft(mat.col(k).tail(remainingRows-1), mat.coeff(k,k), tempData); - - if(k == cols-1) break; - - // construct right householder transform in-place in mat - mat.row(k).tail(remainingCols) - .makeHouseholderInPlace(mat.coeffRef(k,k+1), bidiagonal.template diagonal<1>().coeffRef(k)); - // apply householder transform to remaining part of mat on the left - mat.bottomRightCorner(remainingRows-1, remainingCols) - .applyHouseholderOnTheRight(mat.row(k).tail(remainingCols-1).transpose(), mat.coeff(k,k+1), tempData); + + // This stage performs the bidiagonalization of A11, A21, A12, and updating of A22. + // Finally, the algorithm continue on the updated A22. + // + // However, if B is too small, or A22 empty, then let's use an unblocked strategy + if(k+bs==cols || bcols<48) // somewhat arbitrary threshold + { + upperbidiagonalization_inplace_unblocked(B, + &(bidiagonal.template diagonal<0>().coeffRef(k)), + &(bidiagonal.template diagonal<1>().coeffRef(k)), + X.data() + ); + break; // We're done + } + else + { + upperbidiagonalization_blocked_helper( B, + &(bidiagonal.template diagonal<0>().coeffRef(k)), + &(bidiagonal.template diagonal<1>().coeffRef(k)), + bs, + X.topLeftCorner(brows,bs), + Y.topLeftCorner(bcols,bs) + ); + } } } @@ -348,19 +354,10 @@ UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::comput ColVectorType temp(rows); - upperbidiagonalization_inplace_unblocked(m_householder, m_bidiagonal, temp.data()); - - MatrixType A = matrix; - BidiagonalType B(cols,cols); - upperbidiagonalization_inplace_blocked(A, B, 8); - - std::cout << (m_householder-A)/*.maxCoeff()*/ << "\n\n"; - std::cout << m_householder << "\n\n" - << m_bidiagonal.template diagonal<0>() << "\n" - << m_bidiagonal.template diagonal<1>() << "\n\n"; - std::cout << A << "\n\n" - << B.template diagonal<0>() << "\n" - << B.template diagonal<1>() << "\n\n"; + upperbidiagonalization_inplace_unblocked(m_householder, + &(m_bidiagonal.template diagonal<0>().coeffRef(0)), + &(m_bidiagonal.template diagonal<1>().coeffRef(0)), + temp.data()); m_isInitialized = true; return *this; From 69c057ccb17ec77fcfc4b72795bdf2494b1313a8 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Tue, 27 Aug 2013 14:54:57 +0200 Subject: [PATCH 0121/1103] Fixed InnerPanel definition in the Transformation class. Added some inital documentation on InnerPanel. --- Eigen/src/Core/Block.h | 3 +++ Eigen/src/Geometry/Transform.h | 4 ++-- 2 files changed, 5 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 358b3188b..3efdcfee3 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -21,6 +21,9 @@ namespace Eigen { * \param XprType the type of the expression in which we are taking a block * \param BlockRows the number of rows of the block we are taking at compile time (optional) * \param BlockCols the number of columns of the block we are taking at compile time (optional) + * \param InnerPanel is true, if the block maps to a set of rows of a row major matrix or + * to set of columns of a column major matrix (optional). The parameter allows to determine + * at compile time whether aligned access is possible on the block expression. * * This class represents an expression of either a fixed-size or dynamic-size block. It is the return * type of DenseBase::block(Index,Index,Index,Index) and DenseBase::block(Index,Index) and diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 887e718d6..419f90148 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -208,9 +208,9 @@ public: /** type of a vector */ typedef Matrix VectorType; /** type of a read/write reference to the translation part of the rotation */ - typedef Block TranslationPart; + typedef Block::Flags & RowMajorBit)> TranslationPart; /** type of a read reference to the translation part of the rotation */ - typedef const Block ConstTranslationPart; + typedef const Block::Flags & RowMajorBit)> ConstTranslationPart; /** corresponding translation type */ typedef Translation TranslationType; From 86daf2f75c08b2236fc3d657292aeee600f26a12 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Tue, 27 Aug 2013 15:41:18 +0200 Subject: [PATCH 0122/1103] Added missing inline statements in order to prevent linker errors. --- unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index 6248eeec7..d46ccc145 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -66,7 +66,7 @@ void matrix_log_compute_2x2(const MatrixType& A, MatrixType& result) } /* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = float) */ -int matrix_log_get_pade_degree(float normTminusI) +inline int matrix_log_get_pade_degree(float normTminusI) { const float maxNormForPade[] = { 2.5111573934555054e-1 /* degree = 3 */ , 4.0535837411880493e-1, 5.3149729967117310e-1 }; @@ -80,7 +80,7 @@ int matrix_log_get_pade_degree(float normTminusI) } /* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = double) */ -int matrix_log_get_pade_degree(double normTminusI) +inline int matrix_log_get_pade_degree(double normTminusI) { const double maxNormForPade[] = { 1.6206284795015624e-2 /* degree = 3 */ , 5.3873532631381171e-2, 1.1352802267628681e-1, 1.8662860613541288e-1, 2.642960831111435e-1 }; @@ -94,7 +94,7 @@ int matrix_log_get_pade_degree(double normTminusI) } /* \brief Get suitable degree for Pade approximation. (specialized for RealScalar = long double) */ -int matrix_log_get_pade_degree(long double normTminusI) +inline int matrix_log_get_pade_degree(long double normTminusI) { #if LDBL_MANT_DIG == 53 // double precision const long double maxNormForPade[] = { 1.6206284795015624e-2L /* degree = 3 */ , 5.3873532631381171e-2L, From 16cbd3d72dd94f9152d9f340de73500bd8593c4a Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Tue, 27 Aug 2013 15:30:11 +0100 Subject: [PATCH 0123/1103] BDCSVD: Use rational interpolation to solve secular equation. Algorithm is rather ad-hoc and falls back on bisection if required. --- unsupported/Eigen/src/SVD/BDCSVD.h | 173 ++++++++++++++++++++--------- 1 file changed, 123 insertions(+), 50 deletions(-) diff --git a/unsupported/Eigen/src/SVD/BDCSVD.h b/unsupported/Eigen/src/SVD/BDCSVD.h index d5f0a3f21..80006fd60 100644 --- a/unsupported/Eigen/src/SVD/BDCSVD.h +++ b/unsupported/Eigen/src/SVD/BDCSVD.h @@ -70,6 +70,7 @@ public: typedef Matrix MatrixX; typedef Matrix MatrixXr; typedef Matrix VectorType; + typedef Array ArrayXr; /** \brief Default Constructor. * @@ -78,7 +79,7 @@ public: */ BDCSVD() : SVDBase<_MatrixType>::SVDBase(), - algoswap(ALGOSWAP) + algoswap(ALGOSWAP), m_numIters(0) {} @@ -90,7 +91,7 @@ public: */ BDCSVD(Index rows, Index cols, unsigned int computationOptions = 0) : SVDBase<_MatrixType>::SVDBase(), - algoswap(ALGOSWAP) + algoswap(ALGOSWAP), m_numIters(0) { allocate(rows, cols, computationOptions); } @@ -107,7 +108,7 @@ public: */ BDCSVD(const MatrixType& matrix, unsigned int computationOptions = 0) : SVDBase<_MatrixType>::SVDBase(), - algoswap(ALGOSWAP) + algoswap(ALGOSWAP), m_numIters(0) { compute(matrix, computationOptions); } @@ -197,9 +198,14 @@ public: private: void allocate(Index rows, Index cols, unsigned int computationOptions); - void divide (Index firstCol, Index lastCol, Index firstRowW, - Index firstColW, Index shift); + void divide(Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift); void computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V); + void computeSingVals(const ArrayXr& col0, const ArrayXr& diag, VectorType& singVals, + ArrayXr& shifts, ArrayXr& mus); + void perturbCol0(const ArrayXr& col0, const ArrayXr& diag, const VectorType& singVals, + const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat); + void computeSingVecs(const ArrayXr& zhat, const ArrayXr& diag, const VectorType& singVals, + const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V); void deflation43(Index firstCol, Index shift, Index i, Index size); void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size); void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift); @@ -212,7 +218,9 @@ protected: Index nRec; int algoswap; bool isTranspose, compU, compV; - + +public: + int m_numIters; }; //end class BDCSVD @@ -484,12 +492,9 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, template void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V) { - using std::abs; - Block block = m_computed.block(firstCol, firstCol, n, n); - // TODO Get rid of these copies (?) - Array col0 = m_computed.block(firstCol, firstCol, n, 1); - Array diag = m_computed.block(firstCol, firstCol, n, n).diagonal(); + ArrayXr col0 = m_computed.block(firstCol, firstCol, n, 1); + ArrayXr diag = m_computed.block(firstCol, firstCol, n, n).diagonal(); diag(0) = 0; // compute singular values and vectors (in decreasing order) @@ -499,7 +504,25 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec if (col0.hasNaN() || diag.hasNaN()) return; - Array shifts(n), mus(n); + ArrayXr shifts(n), mus(n), zhat(n); + computeSingVals(col0, diag, singVals, shifts, mus); + perturbCol0(col0, diag, singVals, shifts, mus, zhat); + computeSingVecs(zhat, diag, singVals, shifts, mus, U, V); + + // Reverse order so that singular values in increased order + singVals.reverseInPlace(); + U.leftCols(n) = U.leftCols(n).rowwise().reverse().eval(); + if (compV) V = V.rowwise().reverse().eval(); +} + +template +void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& diag, + VectorType& singVals, ArrayXr& shifts, ArrayXr& mus) +{ + using std::abs; + using std::swap; + + Index n = col0.size(); for (Index k = 0; k < n; ++k) { if (col0(k) == 0) { // entry is deflated, so singular value is on diagonal @@ -509,7 +532,7 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec continue; } - // otherwise, use bisection to find singular value + // otherwise, use secular equation to find singular value RealScalar left = diag(k); RealScalar right = (k != n-1) ? diag(k+1) : (diag(n-1) + col0.matrix().norm()); @@ -518,49 +541,98 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec RealScalar fMid = 1 + (col0.square() / ((diag + mid) * (diag - mid))).sum(); RealScalar shift; - if (k == 0 || fMid > 0) shift = left; + if (k == n-1 || fMid > 0) shift = left; else shift = right; - - // measure everything relative to shifted - Array diagShifted = diag - shift; - RealScalar leftShifted, rightShifted; + + // measure everything relative to shift + ArrayXr diagShifted = diag - shift; + + // initial guess + RealScalar muPrev, muCur; if (shift == left) { - leftShifted = 1e-30; - if (k == 0) rightShifted = right - left; - else rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe + muPrev = (right - left) * 0.1; + if (k == n-1) muCur = right - left; + else muCur = (right - left) * 0.5; } else { - leftShifted = -(right - left) * 0.6; - rightShifted = -1e-30; + muPrev = -(right - left) * 0.1; + muCur = -(right - left) * 0.5; } - RealScalar fLeft = 1 + (col0.square() / ((diagShifted - leftShifted) * (diag + shift + leftShifted))).sum(); - RealScalar fRight = 1 + (col0.square() / ((diagShifted - rightShifted) * (diag + shift + rightShifted))).sum(); - assert(fLeft * fRight < 0); - - while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (std::max)(abs(leftShifted), abs(rightShifted))) { - RealScalar midShifted = (leftShifted + rightShifted) / 2; - RealScalar fMid = 1 + (col0.square() / ((diagShifted - midShifted) * (diag + shift + midShifted))).sum(); - if (fLeft * fMid < 0) { - rightShifted = midShifted; - fRight = fMid; + RealScalar fPrev = 1 + (col0.square() / ((diagShifted - muPrev) * (diag + shift + muPrev))).sum(); + RealScalar fCur = 1 + (col0.square() / ((diagShifted - muCur) * (diag + shift + muCur))).sum(); + if (abs(fPrev) < abs(fCur)) { + swap(fPrev, fCur); + swap(muPrev, muCur); + } + + // rational interpolation: fit a function of the form a / mu + b through the two previous + // iterates and use its zero to compute the next iterate + bool useBisection = false; + while (abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (std::max)(abs(muCur), abs(muPrev)) && fCur != fPrev && !useBisection) { + ++m_numIters; + + RealScalar a = (fCur - fPrev) / (1/muCur - 1/muPrev); + RealScalar b = fCur - a / muCur; + + muPrev = muCur; + fPrev = fCur; + muCur = -a / b; + fCur = 1 + (col0.square() / ((diagShifted - muCur) * (diag + shift + muCur))).sum(); + + if (shift == left && (muCur < 0 || muCur > right - left)) useBisection = true; + if (shift == right && (muCur < -(right - left) || muCur > 0)) useBisection = true; + } + + // fall back on bisection method if rational interpolation did not work + if (useBisection) { + RealScalar leftShifted, rightShifted; + if (shift == left) { + leftShifted = 1e-30; + if (k == 0) rightShifted = right - left; + else rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe } else { - leftShifted = midShifted; - fLeft = fMid; + leftShifted = -(right - left) * 0.6; + rightShifted = -1e-30; } + + RealScalar fLeft = 1 + (col0.square() / ((diagShifted - leftShifted) * (diag + shift + leftShifted))).sum(); + RealScalar fRight = 1 + (col0.square() / ((diagShifted - rightShifted) * (diag + shift + rightShifted))).sum(); + assert(fLeft * fRight < 0); + + while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (std::max)(abs(leftShifted), abs(rightShifted))) { + RealScalar midShifted = (leftShifted + rightShifted) / 2; + RealScalar fMid = 1 + (col0.square() / ((diagShifted - midShifted) * (diag + shift + midShifted))).sum(); + if (fLeft * fMid < 0) { + rightShifted = midShifted; + fRight = fMid; + } else { + leftShifted = midShifted; + fLeft = fMid; + } + } + + muCur = (leftShifted + rightShifted) / 2; } - singVals[k] = shift + (leftShifted + rightShifted) / 2; + singVals[k] = shift + muCur; shifts[k] = shift; - mus[k] = (leftShifted + rightShifted) / 2; + mus[k] = muCur; // perturb singular value slightly if it equals diagonal entry to avoid division by zero later // (deflation is supposed to avoid this from happening) if (singVals[k] == left) singVals[k] *= 1 + NumTraits::epsilon(); if (singVals[k] == right) singVals[k] *= 1 - NumTraits::epsilon(); } +} - // zhat is perturbation of col0 for which singular vectors can be computed stably (see Section 3.1) - Array zhat(n); + +// zhat is perturbation of col0 for which singular vectors can be computed stably (see Section 3.1) +template +void BDCSVD::perturbCol0 + (const ArrayXr& col0, const ArrayXr& diag, const VectorType& singVals, + const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat) +{ + Index n = col0.size(); for (Index k = 0; k < n; ++k) { if (col0(k) == 0) zhat(k) = 0; @@ -583,12 +655,15 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec else zhat(k) = -tmp; } } +} - MatrixXr Mhat = MatrixXr::Zero(n,n); - Mhat.diagonal() = diag; - Mhat.col(0) = zhat; - - // compute singular vectors +// compute singular vectors +template +void BDCSVD::computeSingVecs + (const ArrayXr& zhat, const ArrayXr& diag, const VectorType& singVals, + const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V) +{ + Index n = zhat.size(); for (Index k = 0; k < n; ++k) { if (zhat(k) == 0) { U.col(k) = VectorType::Unit(n+1, k); @@ -606,11 +681,6 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec } } U.col(n) = VectorType::Unit(n+1, n); - - // Reverse order so that singular values in increased order - singVals.reverseInPlace(); - U.leftCols(n) = U.leftCols(n).rowwise().reverse().eval(); - if (compV) V = V.rowwise().reverse().eval(); } @@ -692,8 +762,11 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi template void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift){ //condition 4.1 - RealScalar EPS = NumTraits::epsilon() * 10 * (std::max(m_computed(firstCol + shift + 1, firstCol + shift + 1), m_computed(firstCol + k, firstCol + k))); + using std::sqrt; const Index length = lastCol + 1 - firstCol; + RealScalar norm1 = m_computed.block(firstCol+shift, firstCol+shift, length, 1).squaredNorm(); + RealScalar norm2 = m_computed.block(firstCol+shift, firstCol+shift, length, length).diagonal().squaredNorm(); + RealScalar EPS = 10 * NumTraits::epsilon() * sqrt(norm1 + norm2); if (m_computed(firstCol + shift, firstCol + shift) < EPS){ m_computed(firstCol + shift, firstCol + shift) = EPS; } From eda2f8948a4948ad9225f9baa9f1a737d69976ba Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 3 Sep 2013 21:42:59 +0200 Subject: [PATCH 0124/1103] Another compilation fix with ICC/MSVC combo --- Eigen/src/SparseCore/SparseBlock.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index b6a2d232c..16a20a574 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -393,6 +393,8 @@ public: protected: friend class InnerIterator; friend class ReverseInnerIterator; + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) typename XprType::Nested m_matrix; const internal::variable_if_dynamic m_startRow; From 7fa007e8bfa8aa90731ad5d228e998d71c97b791 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 7 Sep 2013 00:00:13 +0200 Subject: [PATCH 0125/1103] Fix sparse block --- Eigen/src/SparseCore/SparseBlock.h | 804 ++++++++++++++--------------- 1 file changed, 402 insertions(+), 402 deletions(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index a0ba7643b..38c7f1ff6 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -1,410 +1,410 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008-2009 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_SPARSE_BLOCK_H -#define EIGEN_SPARSE_BLOCK_H - -namespace Eigen { - -template -class BlockImpl - : public SparseMatrixBase > -{ - typedef typename internal::remove_all::type _MatrixTypeNested; - typedef Block BlockType; -public: - enum { IsRowMajor = internal::traits::IsRowMajor }; -protected: - enum { OuterSize = IsRowMajor ? BlockRows : BlockCols }; -public: - EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) - - class InnerIterator: public XprType::InnerIterator - { - typedef typename BlockImpl::Index Index; - public: - inline InnerIterator(const BlockType& xpr, Index outer) - : XprType::InnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) - {} - inline Index row() const { return IsRowMajor ? m_outer : this->index(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; - class ReverseInnerIterator: public XprType::ReverseInnerIterator - { - typedef typename BlockImpl::Index Index; - public: - inline ReverseInnerIterator(const BlockType& xpr, Index outer) - : XprType::ReverseInnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) - {} - inline Index row() const { return IsRowMajor ? m_outer : this->index(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; - - inline BlockImpl(const XprType& xpr, int i) - : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) - {} - - inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) - : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) - {} - - EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } - - protected: - - typename XprType::Nested m_matrix; - Index m_outerStart; - const internal::variable_if_dynamic m_outerSize; - - public: - EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) -}; - - -/*************************************************************************** -* specialization for SparseMatrix -***************************************************************************/ - -template -class BlockImpl,BlockRows,BlockCols,true,Sparse> - : public SparseMatrixBase,BlockRows,BlockCols,true> > -{ - typedef SparseMatrix<_Scalar, _Options, _Index> SparseMatrixType; - typedef typename internal::remove_all::type _MatrixTypeNested; - typedef Block BlockType; -public: - enum { IsRowMajor = internal::traits::IsRowMajor }; - EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) -protected: - enum { OuterSize = IsRowMajor ? BlockRows : BlockCols }; -public: - - class InnerIterator: public SparseMatrixType::InnerIterator - { - public: - inline InnerIterator(const BlockType& xpr, Index outer) - : SparseMatrixType::InnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) - {} - inline Index row() const { return IsRowMajor ? m_outer : this->index(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; - class ReverseInnerIterator: public SparseMatrixType::ReverseInnerIterator - { - public: - inline ReverseInnerIterator(const BlockType& xpr, Index outer) - : SparseMatrixType::ReverseInnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) - {} - inline Index row() const { return IsRowMajor ? m_outer : this->index(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; - - inline BlockImpl(const SparseMatrixType& xpr, int i) - : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) - {} - - inline BlockImpl(const SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) - : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) - {} - - template - inline BlockType& operator=(const SparseMatrixBase& other) - { - typedef typename internal::remove_all::type _NestedMatrixType; - _NestedMatrixType& matrix = const_cast<_NestedMatrixType&>(m_matrix);; - // This assignment is slow if this vector set is not empty - // and/or it is not at the end of the nonzeros of the underlying matrix. - - // 1 - eval to a temporary to avoid transposition and/or aliasing issues - SparseMatrix tmp(other); - - // 2 - let's check whether there is enough allocated memory - Index nnz = tmp.nonZeros(); - Index start = m_outerStart==0 ? 0 : matrix.outerIndexPtr()[m_outerStart]; // starting position of the current block - Index end = m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]; // ending position of the current block - Index block_size = end - start; // available room in the current block - Index tail_size = m_matrix.outerIndexPtr()[m_matrix.outerSize()] - end; - - Index free_size = m_matrix.isCompressed() - ? Index(matrix.data().allocatedSize()) + block_size - : block_size; - - if(nnz>free_size) - { - // realloc manually to reduce copies - typename SparseMatrixType::Storage newdata(m_matrix.data().allocatedSize() - block_size + nnz); - +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SPARSE_BLOCK_H +#define EIGEN_SPARSE_BLOCK_H + +namespace Eigen { + +template +class BlockImpl + : public SparseMatrixBase > +{ + typedef typename internal::remove_all::type _MatrixTypeNested; + typedef Block BlockType; +public: + enum { IsRowMajor = internal::traits::IsRowMajor }; +protected: + enum { OuterSize = IsRowMajor ? BlockRows : BlockCols }; +public: + EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) + + class InnerIterator: public XprType::InnerIterator + { + typedef typename BlockImpl::Index Index; + public: + inline InnerIterator(const BlockType& xpr, Index outer) + : XprType::InnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) + {} + inline Index row() const { return IsRowMajor ? m_outer : this->index(); } + inline Index col() const { return IsRowMajor ? this->index() : m_outer; } + protected: + Index m_outer; + }; + class ReverseInnerIterator: public XprType::ReverseInnerIterator + { + typedef typename BlockImpl::Index Index; + public: + inline ReverseInnerIterator(const BlockType& xpr, Index outer) + : XprType::ReverseInnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) + {} + inline Index row() const { return IsRowMajor ? m_outer : this->index(); } + inline Index col() const { return IsRowMajor ? this->index() : m_outer; } + protected: + Index m_outer; + }; + + inline BlockImpl(const XprType& xpr, int i) + : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) + {} + + inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) + : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) + {} + + EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } + EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } + + protected: + + typename XprType::Nested m_matrix; + Index m_outerStart; + const internal::variable_if_dynamic m_outerSize; + + public: + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) +}; + + +/*************************************************************************** +* specialization for SparseMatrix +***************************************************************************/ + +template +class BlockImpl,BlockRows,BlockCols,true,Sparse> + : public SparseMatrixBase,BlockRows,BlockCols,true> > +{ + typedef SparseMatrix<_Scalar, _Options, _Index> SparseMatrixType; + typedef typename internal::remove_all::type _MatrixTypeNested; + typedef Block BlockType; +public: + enum { IsRowMajor = internal::traits::IsRowMajor }; + EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) +protected: + enum { OuterSize = IsRowMajor ? BlockRows : BlockCols }; +public: + + class InnerIterator: public SparseMatrixType::InnerIterator + { + public: + inline InnerIterator(const BlockType& xpr, Index outer) + : SparseMatrixType::InnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) + {} + inline Index row() const { return IsRowMajor ? m_outer : this->index(); } + inline Index col() const { return IsRowMajor ? this->index() : m_outer; } + protected: + Index m_outer; + }; + class ReverseInnerIterator: public SparseMatrixType::ReverseInnerIterator + { + public: + inline ReverseInnerIterator(const BlockType& xpr, Index outer) + : SparseMatrixType::ReverseInnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) + {} + inline Index row() const { return IsRowMajor ? m_outer : this->index(); } + inline Index col() const { return IsRowMajor ? this->index() : m_outer; } + protected: + Index m_outer; + }; + + inline BlockImpl(const SparseMatrixType& xpr, int i) + : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) + {} + + inline BlockImpl(const SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) + : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) + {} + + template + inline BlockType& operator=(const SparseMatrixBase& other) + { + typedef typename internal::remove_all::type _NestedMatrixType; + _NestedMatrixType& matrix = const_cast<_NestedMatrixType&>(m_matrix);; + // This assignment is slow if this vector set is not empty + // and/or it is not at the end of the nonzeros of the underlying matrix. + + // 1 - eval to a temporary to avoid transposition and/or aliasing issues + SparseMatrix tmp(other); + + // 2 - let's check whether there is enough allocated memory + Index nnz = tmp.nonZeros(); + Index start = m_outerStart==0 ? 0 : matrix.outerIndexPtr()[m_outerStart]; // starting position of the current block + Index end = m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]; // ending position of the current block + Index block_size = end - start; // available room in the current block + Index tail_size = m_matrix.outerIndexPtr()[m_matrix.outerSize()] - end; + + Index free_size = m_matrix.isCompressed() + ? Index(matrix.data().allocatedSize()) + block_size + : block_size; + + if(nnz>free_size) + { + // realloc manually to reduce copies + typename SparseMatrixType::Storage newdata(m_matrix.data().allocatedSize() - block_size + nnz); + internal::smart_copy(&m_matrix.data().value(0), &m_matrix.data().value(0) + start, &newdata.value(0)); internal::smart_copy(&m_matrix.data().index(0), &m_matrix.data().index(0) + start, &newdata.index(0)); - + internal::smart_copy(&tmp.data().value(0), &tmp.data().value(0) + nnz, &newdata.value(start)); internal::smart_copy(&tmp.data().index(0), &tmp.data().index(0) + nnz, &newdata.index(start)); - + internal::smart_copy(&matrix.data().value(end), &matrix.data().value(end) + tail_size, &newdata.value(start+nnz)); internal::smart_copy(&matrix.data().index(end), &matrix.data().index(end) + tail_size, &newdata.index(start+nnz)); - - newdata.resize(m_matrix.outerIndexPtr()[m_matrix.outerSize()] - block_size + nnz); - - matrix.data().swap(newdata); - } - else - { - // no need to realloc, simply copy the tail at its respective position and insert tmp - matrix.data().resize(start + nnz + tail_size); - + + newdata.resize(m_matrix.outerIndexPtr()[m_matrix.outerSize()] - block_size + nnz); + + matrix.data().swap(newdata); + } + else + { + // no need to realloc, simply copy the tail at its respective position and insert tmp + matrix.data().resize(start + nnz + tail_size); + internal::smart_memmove(&matrix.data().value(end), &matrix.data().value(end) + tail_size, &matrix.data().value(start + nnz)); internal::smart_memmove(&matrix.data().index(end), &matrix.data().index(end) + tail_size, &matrix.data().index(start + nnz)); - - internal::smart_copy(&tmp.data().value(0), &tmp.data().value(0) + nnz, &matrix.value(start)); - internal::smart_copy(&tmp.data().index(0), &tmp.data().index(0) + nnz, &matrix.index(start)); - } - - // update innerNonZeros - if(!m_matrix.isCompressed()) - for(Index j=0; j(other); - } - - inline const Scalar* valuePtr() const - { return m_matrix.valuePtr() + m_matrix.outerIndexPtr()[m_outerStart]; } - inline Scalar* valuePtr() - { return m_matrix.const_cast_derived().valuePtr() + m_matrix.outerIndexPtr()[m_outerStart]; } - - inline const Index* innerIndexPtr() const - { return m_matrix.innerIndexPtr() + m_matrix.outerIndexPtr()[m_outerStart]; } - inline Index* innerIndexPtr() - { return m_matrix.const_cast_derived().innerIndexPtr() + m_matrix.outerIndexPtr()[m_outerStart]; } - - inline const Index* outerIndexPtr() const - { return m_matrix.outerIndexPtr() + m_outerStart; } - inline Index* outerIndexPtr() - { return m_matrix.const_cast_derived().outerIndexPtr() + m_outerStart; } - - Index nonZeros() const - { - if(m_matrix.isCompressed()) - return std::size_t(m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]) - - std::size_t(m_matrix.outerIndexPtr()[m_outerStart]); - else if(m_outerSize.value()==0) - return 0; - else - return Map >(m_matrix.innerNonZeroPtr()+m_outerStart, m_outerSize.value()).sum(); - } - - const Scalar& lastCoeff() const - { - EIGEN_STATIC_ASSERT_VECTOR_ONLY(BlockImpl); - eigen_assert(nonZeros()>0); - if(m_matrix.isCompressed()) - return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart+1]-1]; - else - return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart]+m_matrix.innerNonZeroPtr()[m_outerStart]-1]; - } - - EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } - - protected: - - typename SparseMatrixType::Nested m_matrix; - Index m_outerStart; - const internal::variable_if_dynamic m_outerSize; - -}; - -//---------- - -/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this - * is col-major (resp. row-major). - */ -template -typename SparseMatrixBase::InnerVectorReturnType SparseMatrixBase::innerVector(Index outer) -{ return InnerVectorReturnType(derived(), outer); } - -/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this - * is col-major (resp. row-major). Read-only. - */ -template -const typename SparseMatrixBase::ConstInnerVectorReturnType SparseMatrixBase::innerVector(Index outer) const -{ return ConstInnerVectorReturnType(derived(), outer); } - -/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this - * is col-major (resp. row-major). - */ -template -Block SparseMatrixBase::innerVectors(Index outerStart, Index outerSize) -{ - return Block(derived(), - IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart, - IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize); - -} - -/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this - * is col-major (resp. row-major). Read-only. - */ -template -const Block SparseMatrixBase::innerVectors(Index outerStart, Index outerSize) const -{ - return Block(derived(), - IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart, - IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize); - -} - -/** Generic implementation of sparse Block expression. - * Real-only. - */ -template -class BlockImpl - : public SparseMatrixBase >, internal::no_assignment_operator -{ - typedef typename internal::remove_all::type _MatrixTypeNested; - typedef Block BlockType; -public: - enum { IsRowMajor = internal::traits::IsRowMajor }; - EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) - - /** Column or Row constructor - */ - inline BlockImpl(const XprType& xpr, int i) - : m_matrix(xpr), - m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0), - m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0), - m_blockRows(xpr.rows()), - m_blockCols(xpr.cols()) - {} - - /** Dynamic-size constructor - */ - inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) - : m_matrix(xpr), m_startRow(startRow), m_startCol(startCol), m_blockRows(blockRows), m_blockCols(blockCols) - {} - - inline int rows() const { return m_blockRows.value(); } - inline int cols() const { return m_blockCols.value(); } - - inline Scalar& coeffRef(int row, int col) - { - return m_matrix.const_cast_derived() - .coeffRef(row + m_startRow.value(), col + m_startCol.value()); - } - - inline const Scalar coeff(int row, int col) const - { - return m_matrix.coeff(row + m_startRow.value(), col + m_startCol.value()); - } - - inline Scalar& coeffRef(int index) - { - return m_matrix.const_cast_derived() - .coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), - m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); - } - - inline const Scalar coeff(int index) const - { - return m_matrix - .coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), - m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); - } - - inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } - - class InnerIterator : public _MatrixTypeNested::InnerIterator - { - typedef typename _MatrixTypeNested::InnerIterator Base; - const BlockType& m_block; - Index m_end; - public: - - EIGEN_STRONG_INLINE InnerIterator(const BlockType& block, Index outer) - : Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())), - m_block(block), - m_end(IsRowMajor ? block.m_startCol.value()+block.m_blockCols.value() : block.m_startRow.value()+block.m_blockRows.value()) - { - while( (Base::operator bool()) && (Base::index() < (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value())) ) - Base::operator++(); - } - - inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } - inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } - inline Index row() const { return Base::row() - m_block.m_startRow.value(); } - inline Index col() const { return Base::col() - m_block.m_startCol.value(); } - - inline operator bool() const { return Base::operator bool() && Base::index() < m_end; } - }; - class ReverseInnerIterator : public _MatrixTypeNested::ReverseInnerIterator - { - typedef typename _MatrixTypeNested::ReverseInnerIterator Base; - const BlockType& m_block; - Index m_begin; - public: - - EIGEN_STRONG_INLINE ReverseInnerIterator(const BlockType& block, Index outer) - : Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())), - m_block(block), - m_begin(IsRowMajor ? block.m_startCol.value() : block.m_startRow.value()) - { - while( (Base::operator bool()) && (Base::index() >= (IsRowMajor ? m_block.m_startCol.value()+block.m_blockCols.value() : m_block.m_startRow.value()+block.m_blockRows.value())) ) - Base::operator--(); - } - - inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } - inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } - inline Index row() const { return Base::row() - m_block.m_startRow.value(); } - inline Index col() const { return Base::col() - m_block.m_startCol.value(); } - - inline operator bool() const { return Base::operator bool() && Base::index() >= m_begin; } - }; - protected: - friend class InnerIterator; - friend class ReverseInnerIterator; - - EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) - - typename XprType::Nested m_matrix; - const internal::variable_if_dynamic m_startRow; - const internal::variable_if_dynamic m_startCol; - const internal::variable_if_dynamic m_blockRows; - const internal::variable_if_dynamic m_blockCols; - -}; - -} // end namespace Eigen - -#endif // EIGEN_SPARSE_BLOCK_H + + internal::smart_copy(&tmp.data().value(0), &tmp.data().value(0) + nnz, &matrix.data().value(start)); + internal::smart_copy(&tmp.data().index(0), &tmp.data().index(0) + nnz, &matrix.data().index(start)); + } + + // update innerNonZeros + if(!m_matrix.isCompressed()) + for(Index j=0; j(other); + } + + inline const Scalar* valuePtr() const + { return m_matrix.valuePtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + inline Scalar* valuePtr() + { return m_matrix.const_cast_derived().valuePtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + + inline const Index* innerIndexPtr() const + { return m_matrix.innerIndexPtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + inline Index* innerIndexPtr() + { return m_matrix.const_cast_derived().innerIndexPtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + + inline const Index* outerIndexPtr() const + { return m_matrix.outerIndexPtr() + m_outerStart; } + inline Index* outerIndexPtr() + { return m_matrix.const_cast_derived().outerIndexPtr() + m_outerStart; } + + Index nonZeros() const + { + if(m_matrix.isCompressed()) + return std::size_t(m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]) + - std::size_t(m_matrix.outerIndexPtr()[m_outerStart]); + else if(m_outerSize.value()==0) + return 0; + else + return Map >(m_matrix.innerNonZeroPtr()+m_outerStart, m_outerSize.value()).sum(); + } + + const Scalar& lastCoeff() const + { + EIGEN_STATIC_ASSERT_VECTOR_ONLY(BlockImpl); + eigen_assert(nonZeros()>0); + if(m_matrix.isCompressed()) + return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart+1]-1]; + else + return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart]+m_matrix.innerNonZeroPtr()[m_outerStart]-1]; + } + + EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } + EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } + + protected: + + typename SparseMatrixType::Nested m_matrix; + Index m_outerStart; + const internal::variable_if_dynamic m_outerSize; + +}; + +//---------- + +/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this + * is col-major (resp. row-major). + */ +template +typename SparseMatrixBase::InnerVectorReturnType SparseMatrixBase::innerVector(Index outer) +{ return InnerVectorReturnType(derived(), outer); } + +/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this + * is col-major (resp. row-major). Read-only. + */ +template +const typename SparseMatrixBase::ConstInnerVectorReturnType SparseMatrixBase::innerVector(Index outer) const +{ return ConstInnerVectorReturnType(derived(), outer); } + +/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this + * is col-major (resp. row-major). + */ +template +Block SparseMatrixBase::innerVectors(Index outerStart, Index outerSize) +{ + return Block(derived(), + IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart, + IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize); + +} + +/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this + * is col-major (resp. row-major). Read-only. + */ +template +const Block SparseMatrixBase::innerVectors(Index outerStart, Index outerSize) const +{ + return Block(derived(), + IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart, + IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize); + +} + +/** Generic implementation of sparse Block expression. + * Real-only. + */ +template +class BlockImpl + : public SparseMatrixBase >, internal::no_assignment_operator +{ + typedef typename internal::remove_all::type _MatrixTypeNested; + typedef Block BlockType; +public: + enum { IsRowMajor = internal::traits::IsRowMajor }; + EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) + + /** Column or Row constructor + */ + inline BlockImpl(const XprType& xpr, int i) + : m_matrix(xpr), + m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0), + m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0), + m_blockRows(xpr.rows()), + m_blockCols(xpr.cols()) + {} + + /** Dynamic-size constructor + */ + inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) + : m_matrix(xpr), m_startRow(startRow), m_startCol(startCol), m_blockRows(blockRows), m_blockCols(blockCols) + {} + + inline int rows() const { return m_blockRows.value(); } + inline int cols() const { return m_blockCols.value(); } + + inline Scalar& coeffRef(int row, int col) + { + return m_matrix.const_cast_derived() + .coeffRef(row + m_startRow.value(), col + m_startCol.value()); + } + + inline const Scalar coeff(int row, int col) const + { + return m_matrix.coeff(row + m_startRow.value(), col + m_startCol.value()); + } + + inline Scalar& coeffRef(int index) + { + return m_matrix.const_cast_derived() + .coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), + m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); + } + + inline const Scalar coeff(int index) const + { + return m_matrix + .coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), + m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); + } + + inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } + + class InnerIterator : public _MatrixTypeNested::InnerIterator + { + typedef typename _MatrixTypeNested::InnerIterator Base; + const BlockType& m_block; + Index m_end; + public: + + EIGEN_STRONG_INLINE InnerIterator(const BlockType& block, Index outer) + : Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())), + m_block(block), + m_end(IsRowMajor ? block.m_startCol.value()+block.m_blockCols.value() : block.m_startRow.value()+block.m_blockRows.value()) + { + while( (Base::operator bool()) && (Base::index() < (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value())) ) + Base::operator++(); + } + + inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } + inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } + inline Index row() const { return Base::row() - m_block.m_startRow.value(); } + inline Index col() const { return Base::col() - m_block.m_startCol.value(); } + + inline operator bool() const { return Base::operator bool() && Base::index() < m_end; } + }; + class ReverseInnerIterator : public _MatrixTypeNested::ReverseInnerIterator + { + typedef typename _MatrixTypeNested::ReverseInnerIterator Base; + const BlockType& m_block; + Index m_begin; + public: + + EIGEN_STRONG_INLINE ReverseInnerIterator(const BlockType& block, Index outer) + : Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())), + m_block(block), + m_begin(IsRowMajor ? block.m_startCol.value() : block.m_startRow.value()) + { + while( (Base::operator bool()) && (Base::index() >= (IsRowMajor ? m_block.m_startCol.value()+block.m_blockCols.value() : m_block.m_startRow.value()+block.m_blockRows.value())) ) + Base::operator--(); + } + + inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } + inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } + inline Index row() const { return Base::row() - m_block.m_startRow.value(); } + inline Index col() const { return Base::col() - m_block.m_startCol.value(); } + + inline operator bool() const { return Base::operator bool() && Base::index() >= m_begin; } + }; + protected: + friend class InnerIterator; + friend class ReverseInnerIterator; + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) + + typename XprType::Nested m_matrix; + const internal::variable_if_dynamic m_startRow; + const internal::variable_if_dynamic m_startCol; + const internal::variable_if_dynamic m_blockRows; + const internal::variable_if_dynamic m_blockCols; + +}; + +} // end namespace Eigen + +#endif // EIGEN_SPARSE_BLOCK_H From 07417bd03f501efd276302123eb77b2f5a1a67b7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 7 Sep 2013 00:01:04 +0200 Subject: [PATCH 0126/1103] Fix bug #654: allow implicit transposition in Array to Matrix and Matrix to Array constructors --- Eigen/src/Core/Array.h | 2 +- Eigen/src/Core/Matrix.h | 2 +- test/array_for_matrix.cpp | 7 +++++++ 3 files changed, 9 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h index 3d0350834..0b9c38c82 100644 --- a/Eigen/src/Core/Array.h +++ b/Eigen/src/Core/Array.h @@ -225,7 +225,7 @@ class Array : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols()) { Base::_check_template_params(); - Base::resize(other.rows(), other.cols()); + Base::_resize_to_match(other); *this = other; } diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 8dba1bf36..02be142d8 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -319,7 +319,7 @@ class Matrix : Base(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols()) { Base::_check_template_params(); - Base::resize(other.rows(), other.cols()); + Base::_resize_to_match(other); // FIXME/CHECK: isn't *this = other.derived() more efficient. it allows to // go for pure _set() implementations, right? *this = other; diff --git a/test/array_for_matrix.cpp b/test/array_for_matrix.cpp index 1250c9ff5..67e7a2a44 100644 --- a/test/array_for_matrix.cpp +++ b/test/array_for_matrix.cpp @@ -202,6 +202,12 @@ template void resize(const MatrixTraits& t) VERIFY(a1.size()==cols); } +void regression_bug_654() +{ + ArrayXf a = RowVectorXf(3); + VectorXf v = Array(3); +} + void test_array_for_matrix() { for(int i = 0; i < g_repeat; i++) { @@ -239,4 +245,5 @@ void test_array_for_matrix() CALL_SUBTEST_5( resize(MatrixXf(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); CALL_SUBTEST_6( resize(MatrixXi(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); } + CALL_SUBTEST_6( regression_bug_654() ); } From 4612a1cd870b4a72b3849608fcdce9a18dc80a65 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 10 Sep 2013 16:13:59 +0200 Subject: [PATCH 0127/1103] Fix ploaddup and lin-spaced with AltiVec. --- Eigen/src/Core/arch/AltiVec/PacketMath.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) mode change 100644 => 100755 Eigen/src/Core/arch/AltiVec/PacketMath.h diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h old mode 100644 new mode 100755 index e4089962d..45d1954f7 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -56,8 +56,8 @@ typedef __vector unsigned char Packet16uc; #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride)) // Define global static constants: -static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 }; -static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 }; +static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; +static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; static Packet16uc p16uc_REVERSE = {12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; @@ -286,15 +286,15 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) template<> EIGEN_STRONG_INLINE Packet4f ploaddup(const float* from) { Packet4f p; - if((ptrdiff_t(&from) % 16) == 0) p = pload(from); - else p = ploadu(from); + if((ptrdiff_t(from) % 16) == 0) p = pload(from); + else p = ploadu(from); return vec_perm(p, p, p16uc_DUPLICATE); } template<> EIGEN_STRONG_INLINE Packet4i ploaddup(const int* from) { Packet4i p; - if((ptrdiff_t(&from) % 16) == 0) p = pload(from); - else p = ploadu(from); + if((ptrdiff_t(from) % 16) == 0) p = pload(from); + else p = ploadu(from); return vec_perm(p, p, p16uc_DUPLICATE); } From a1f056cf2a0318d7a8d4a4bf4c5019c09a8fdbd6 Mon Sep 17 00:00:00 2001 From: Martinho Fernandes Date: Tue, 10 Sep 2013 17:08:04 +0200 Subject: [PATCH 0128/1103] Fix bug #503 C++11 support on simple allocators comes for free. `aligned_allocator` does not need to add any `construct` overloads to work with C++11 compilers. --- Eigen/src/Core/util/Memory.h | 9 --------- 1 file changed, 9 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 3a357e079..fab90f0da 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -758,15 +758,6 @@ public: ::new( p ) T( value ); } - // Support for c++11 -#if (__cplusplus >= 201103L) - template - void construct(pointer p, Args&&... args) - { - ::new(p) T(std::forward(args)...); - } -#endif - void destroy( pointer p ) { p->~T(); From 1b4623e7132c0959925386801e9e342ba5565f19 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 12 Sep 2013 22:16:35 +0200 Subject: [PATCH 0129/1103] Fix elimination tree and SparseQR with rows=m; pp(col) = col; cset = col; root(cset) = col; @@ -105,6 +106,7 @@ int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowEl Index i = col; if(it) i = it.index(); if (i == col) found_diag = true; + row = firstRowElt(i); if (row >= col) continue; rset = internal::etree_find(row, pp); // Find the name of the set containing row diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index c42dfc556..50c4c1145 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -161,8 +161,9 @@ class SparseQR b = y; // Solve with the triangular matrix R + y.resize((std::max)(cols(),Index(y.rows())),y.cols()); y.topRows(rank) = this->matrixR().topLeftCorner(rank, rank).template triangularView().solve(b.topRows(rank)); - y.bottomRows(y.size()-rank).setZero(); + y.bottomRows(y.rows()-rank).setZero(); // Apply the column permutation if (m_perm_c.size()) dest.topRows(cols()) = colsPermutation() * y.topRows(cols()); @@ -246,7 +247,7 @@ class SparseQR Index m_nonzeropivots; // Number of non zero pivots found IndexVector m_etree; // Column elimination tree IndexVector m_firstRowElt; // First element in each row - bool m_isQSorted; // whether Q is sorted or not + bool m_isQSorted; // whether Q is sorted or not template friend struct SparseQR_QProduct; template friend struct SparseQRMatrixQReturnType; @@ -338,7 +339,7 @@ void SparseQR::factorize(const MatrixType& mat) Index nonzeroCol = 0; // Record the number of valid pivots // Left looking rank-revealing QR factorization: compute a column of R and Q at a time - for (Index col = 0; col < n; ++col) + for (Index col = 0; col < (std::min)(n,m); ++col) { mark.setConstant(-1); m_R.startVec(col); @@ -346,7 +347,7 @@ void SparseQR::factorize(const MatrixType& mat) mark(nonzeroCol) = col; Qidx(0) = nonzeroCol; nzcolR = 0; nzcolQ = 1; - found_diag = false; + found_diag = col>=m; tval.setZero(); // Symbolic factorization: find the nonzero locations of the column k of the factors R and Q, i.e., From bd21c82a942c69a46a40b7cdde9e1e9023501314 Mon Sep 17 00:00:00 2001 From: Desire NUENTSA Date: Fri, 20 Sep 2013 18:49:32 +0200 Subject: [PATCH 0130/1103] Fix assert bug in sparseQR --- Eigen/src/SparseQR/SparseQR.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 50c4c1145..e9b2a6bec 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -580,7 +580,7 @@ struct SparseQR_QProduct : ReturnByValue Date: Fri, 20 Sep 2013 18:54:17 +0200 Subject: [PATCH 0131/1103] Add a block sparse matrix class. tests to be added --- Eigen/src/SparseCore/SparseDenseProduct.h | 26 +- .../Eigen/src/SparseExtra/BlockSparseMatrix.h | 1079 +++++++++++++++++ 2 files changed, 1097 insertions(+), 8 deletions(-) create mode 100644 unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 30975c29c..6132a4880 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -139,12 +139,13 @@ struct traits > }; template struct sparse_time_dense_product_impl; template -struct sparse_time_dense_product_impl +struct sparse_time_dense_product_impl { typedef typename internal::remove_all::type Lhs; typedef typename internal::remove_all::type Rhs; @@ -167,21 +168,30 @@ struct sparse_time_dense_product_impl -struct sparse_time_dense_product_impl +template +struct scalar_product_traits > +{ + enum { + Defined = 1 + }; + typedef typename CwiseUnaryOp, T2>::PlainObject ReturnType; +}; +template +struct sparse_time_dense_product_impl { typedef typename internal::remove_all::type Lhs; typedef typename internal::remove_all::type Rhs; typedef typename internal::remove_all::type Res; typedef typename Lhs::InnerIterator LhsInnerIterator; typedef typename Lhs::Index Index; - static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) + static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) { for(Index c=0; c::ReturnType rhs_j(alpha * rhs.coeff(j,c)); for(LhsInnerIterator it(lhs,j); it ;++it) res.coeffRef(it.index(),c) += it.value() * rhs_j; } @@ -190,7 +200,7 @@ struct sparse_time_dense_product_impl -struct sparse_time_dense_product_impl +struct sparse_time_dense_product_impl { typedef typename internal::remove_all::type Lhs; typedef typename internal::remove_all::type Rhs; @@ -209,7 +219,7 @@ struct sparse_time_dense_product_impl -struct sparse_time_dense_product_impl +struct sparse_time_dense_product_impl { typedef typename internal::remove_all::type Lhs; typedef typename internal::remove_all::type Rhs; @@ -230,7 +240,7 @@ struct sparse_time_dense_product_impl inline void sparse_time_dense_product(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) { - sparse_time_dense_product_impl::run(lhs, rhs, res, alpha); + sparse_time_dense_product_impl::run(lhs, rhs, res, alpha); } } // end namespace internal diff --git a/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h b/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h new file mode 100644 index 000000000..6d845961e --- /dev/null +++ b/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h @@ -0,0 +1,1079 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2013 Desire Nuentsa +// Copyright (C) 2013 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SPARSEBLOCKMATRIX_H +#define EIGEN_SPARSEBLOCKMATRIX_H + +namespace Eigen { +/** \ingroup SparseCore_Module + * + * \class BlockSparseMatrix + * + * \brief A versatile sparse matrix representation where each element is a block + * + * This class provides routines to manipulate block sparse matrices stored in a + * BSR-like representation. There are two main types : + * + * 1. All blocks have the same number of rows and columns, called block size + * in the following. In this case, if this block size is known at compile time, + * it can be given as a template parameter like + * \code + * BlockSparseMatrix bmat(b_rows, b_cols); + * \endcode + * Here, bmat is a b_rows x b_cols block sparse matrix + * where each coefficient is a 3x3 dense matrix. + * If the block size is fixed but will be given at runtime, + * \code + * BlockSparseMatrix bmat(b_rows, b_cols); + * bmat.setBlockSize(block_size); + * \endcode + * + * 2. The second case is for variable-block sparse matrices. + * Here each block has its own dimensions. The only restriction is that all the blocks + * in a row (resp. a column) should have the same number of rows (resp. of columns). + * It is thus required in this case to describe the layout of the matrix by calling + * setBlockLayout(rowBlocks, colBlocks). + * + * In any of the previous case, the matrix can be filled by calling setFromTriplets(). + * A regular sparse matrix can be converted to a block sparse matrix and vice versa. + * It is obviously required to describe the block layout beforehand by calling either + * setBlockSize() for fixed-size blocks or setBlockLayout for variable-size blocks. + * + * \tparam _Scalar The Scalar type + * \tparam _BlockAtCompileTime The block layout option. It takes the following values + * Dynamic : block size known at runtime + * a numeric number : fixed-size block known at compile time + */ +template class BlockSparseMatrix; + +template class BlockSparseMatrixView; + +namespace internal { +template +struct traits > +{ + typedef _Scalar Scalar; + typedef _Index Index; + typedef Sparse StorageKind; // FIXME Where is it used ?? + typedef MatrixXpr XprKind; + enum { + RowsAtCompileTime = Dynamic, + ColsAtCompileTime = Dynamic, + MaxRowsAtCompileTime = Dynamic, + MaxColsAtCompileTime = Dynamic, + BlockSize = _BlockAtCompileTime, + Flags = _Options | NestByRefBit | LvalueBit, + CoeffReadCost = NumTraits::ReadCost, + SupportedAccessPatterns = InnerRandomAccessPattern + }; +}; +template +struct traits > +{ + typedef Ref > Scalar; + typedef Ref > RealScalar; + +}; + +// Function object to sort a triplet list +template +struct TripletComp +{ + typedef typename Iterator::value_type Triplet; + bool operator()(const Triplet& a, const Triplet& b) + { if(IsColMajor) + return ((a.col() == b.col() && a.row() < b.row()) || (a.col() < b.col())); + else + return ((a.row() == b.row() && a.col() < b.col()) || (a.row() < b.row())); + } +}; +} // end namespace internal + + +/* Proxy to view the block sparse matrix as a regular sparse matrix */ +template +class BlockSparseMatrixView : public SparseMatrixBase +{ + public: + typedef Ref Scalar; + typedef Ref RealScalar; + typedef typename BlockSparseMatrixT::Index Index; + typedef BlockSparseMatrixT Nested; + enum { + Flags = BlockSparseMatrixT::Options, + Options = BlockSparseMatrixT::Options, + RowsAtCompileTime = BlockSparseMatrixT::RowsAtCompileTime, + ColsAtCompileTime = BlockSparseMatrixT::ColsAtCompileTime, + MaxColsAtCompileTime = BlockSparseMatrixT::MaxColsAtCompileTime, + MaxRowsAtCompileTime = BlockSparseMatrixT::MaxRowsAtCompileTime + }; + public: + BlockSparseMatrixView(const BlockSparseMatrixT& spblockmat) + : m_spblockmat(spblockmat) + {} + + Index outerSize() const + { + return (Flags&RowMajorBit) == 1 ? this->rows() : this->cols(); + } + Index cols() const + { + return m_spblockmat.blockCols(); + } + Index rows() const + { + return m_spblockmat.blockRows(); + } + Scalar coeff(Index row, Index col) + { + return m_spblockmat.coeff(row, col); + } + Scalar coeffRef(Index row, Index col) + { + return m_spblockmat.coeffRef(row, col); + } + // Wrapper to iterate over all blocks + class InnerIterator : public BlockSparseMatrixT::BlockInnerIterator + { + public: + InnerIterator(const BlockSparseMatrixView& mat, Index outer) + : BlockSparseMatrixT::BlockInnerIterator(mat.m_spblockmat, outer) + {} + + }; + + protected: + const BlockSparseMatrixT& m_spblockmat; +}; + +// Proxy to view a regular vector as a block vector +template +class BlockVectorView +{ + public: + enum { + BlockSize = BlockSparseMatrixT::BlockSize, + ColsAtCompileTime = VectorType::ColsAtCompileTime, + RowsAtCompileTime = VectorType::RowsAtCompileTime, + Flags = VectorType::Flags + }; + typedef Ref >Scalar; + typedef typename BlockSparseMatrixT::Index Index; + public: + BlockVectorView(const BlockSparseMatrixT& spblockmat, const VectorType& vec) + : m_spblockmat(spblockmat),m_vec(vec) + { } + inline Index cols() const + { + return m_vec.cols(); + } + inline Index size() const + { + return m_spblockmat.blockRows(); + } + inline Scalar coeff(Index bi) const + { + Index startRow = m_spblockmat.blockRowsIndex(bi); + Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow; + return m_vec.middleRows(startRow, rowSize); + } + inline Scalar coeff(Index bi, Index j) const + { + Index startRow = m_spblockmat.blockRowsIndex(bi); + Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow; + return m_vec.block(startRow, j, rowSize, 1); + } + protected: + const BlockSparseMatrixT& m_spblockmat; + const VectorType& m_vec; +}; + +template class BlockVectorReturn; + + +// Proxy to view a regular vector as a block vector +template +class BlockVectorReturn +{ + public: + enum { + ColsAtCompileTime = VectorType::ColsAtCompileTime, + RowsAtCompileTime = VectorType::RowsAtCompileTime, + Flags = VectorType::Flags + }; + typedef Ref > Scalar; + typedef typename BlockSparseMatrixT::Index Index; + public: + BlockVectorReturn(const BlockSparseMatrixT& spblockmat, VectorType& vec) + : m_spblockmat(spblockmat),m_vec(vec) + { } + inline Index size() const + { + return m_spblockmat.blockRows(); + } + inline Scalar coeffRef(Index bi) + { + Index startRow = m_spblockmat.blockRowsIndex(bi); + Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow; + return m_vec.middleRows(startRow, rowSize); + } + inline Scalar coeffRef(Index bi, Index j) + { + Index startRow = m_spblockmat.blockRowsIndex(bi); + Index rowSize = m_spblockmat.blockRowsIndex(bi+1) - startRow; + return m_vec.block(startRow, j, rowSize, 1); + } + + protected: + const BlockSparseMatrixT& m_spblockmat; + VectorType& m_vec; +}; + +// Block version of the sparse dense product +template +class BlockSparseTimeDenseProduct; + +namespace internal { + +template +struct traits > +{ + typedef Dense StorageKind; + typedef MatrixXpr XprKind; + typedef typename BlockSparseMatrixT::Scalar Scalar; + typedef typename BlockSparseMatrixT::Index Index; + enum { + RowsAtCompileTime = Dynamic, + ColsAtCompileTime = Dynamic, + MaxRowsAtCompileTime = Dynamic, + MaxColsAtCompileTime = Dynamic, + Flags = 0, + CoeffReadCost = internal::traits::CoeffReadCost + }; +}; +} // end namespace internal + +template +class BlockSparseTimeDenseProduct + : public ProductBase, Lhs, Rhs> +{ + public: + EIGEN_PRODUCT_PUBLIC_INTERFACE(BlockSparseTimeDenseProduct) + + BlockSparseTimeDenseProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) + {} + + template void scaleAndAddTo(Dest& dest, const typename Rhs::Scalar& alpha) const + { + BlockVectorReturn tmpDest(m_lhs, dest); + internal::sparse_time_dense_product( BlockSparseMatrixView(m_lhs), BlockVectorView(m_lhs, m_rhs), tmpDest, alpha); + } + + private: + BlockSparseTimeDenseProduct& operator=(const BlockSparseTimeDenseProduct&); +}; + +template +class BlockSparseMatrix : public SparseMatrixBase > +{ + public: + typedef _Scalar Scalar; + typedef typename NumTraits::Real RealScalar; + typedef _Index Index; + typedef typename internal::nested >::type Nested; + + enum { + Options = _Options, + Flags = Options, + BlockSize=_BlockAtCompileTime, + RowsAtCompileTime = Dynamic, + ColsAtCompileTime = Dynamic, + MaxRowsAtCompileTime = Dynamic, + MaxColsAtCompileTime = Dynamic, + IsVectorAtCompileTime = 0, + IsColMajor = Flags&RowMajorBit ? 0 : 1 + }; + typedef Matrix BlockScalar; + typedef Matrix BlockRealScalar; + typedef typename internal::conditional<_BlockAtCompileTime==Dynamic, Scalar, BlockScalar>::type BlockScalarReturnType; + typedef BlockSparseMatrix PlainObject; + public: + // Default constructor + BlockSparseMatrix() + : m_innerBSize(0),m_outerBSize(0),m_innerOffset(0),m_outerOffset(0), + m_nonzerosblocks(0),m_values(0),m_blockPtr(0),m_indices(0), + m_outerIndex(0),m_blockSize(BlockSize) + { } + + + /** + * \brief Construct and resize + * + */ + BlockSparseMatrix(Index brow, Index bcol) + : m_innerBSize(IsColMajor ? brow : bcol), + m_outerBSize(IsColMajor ? bcol : brow), + m_innerOffset(0),m_outerOffset(0),m_nonzerosblocks(0), + m_values(0),m_blockPtr(0),m_indices(0), + m_outerIndex(0),m_blockSize(BlockSize) + { } + + /** + * \brief Copy-constructor + */ + BlockSparseMatrix(const BlockSparseMatrix& other) + : m_innerBSize(other.m_innerBSize),m_outerBSize(other.m_outerBSize), + m_nonzerosblocks(other.m_nonzerosblocks),m_nonzeros(other.m_nonzeros), + m_blockPtr(0),m_blockSize(other.m_blockSize) + { + // should we allow copying between variable-size blocks and fixed-size blocks ?? + eigen_assert(m_blockSize == BlockSize && " CAN NOT COPY BETWEEN FIXED-SIZE AND VARIABLE-SIZE BLOCKS"); + + std::copy(other.m_innerOffset, other.m_innerOffset+m_innerBSize+1, m_innerOffset); + std::copy(other.m_outerOffset, other.m_outerOffset+m_outerBSize+1, m_outerOffset); + std::copy(other.m_values, other.m_values+m_nonzeros, m_values); + + if(m_blockSize != Dynamic) + std::copy(other.m_blockPtr, other.m_blockPtr+m_nonzerosblocks, m_blockPtr); + + std::copy(other.m_indices, other.m_indices+m_nonzerosblocks, m_indices); + std::copy(other.m_outerIndex, other.m_outerIndex+m_outerBSize, m_outerIndex); + } + + friend void swap(BlockSparseMatrix& first, BlockSparseMatrix& second) + { + std::swap(first.m_innerBSize, second.m_innerBSize); + std::swap(first.m_outerBSize, second.m_outerBSize); + std::swap(first.m_innerOffset, second.m_innerOffset); + std::swap(first.m_outerOffset, second.m_outerOffset); + std::swap(first.m_nonzerosblocks, second.m_nonzerosblocks); + std::swap(first.m_nonzeros, second.m_nonzeros); + std::swap(first.m_values, second.m_values); + std::swap(first.m_blockPtr, second.m_blockPtr); + std::swap(first.m_indices, second.m_indices); + std::swap(first.m_outerIndex, second.m_outerIndex); + std::swap(first.m_BlockSize, second.m_blockSize); + } + + BlockSparseMatrix& operator=(BlockSparseMatrix other) + { + //Copy-and-swap paradigm ... avoid leaked data if thrown + swap(*this, other); + return *this; + } + + // Destructor + ~BlockSparseMatrix() + { + delete[] m_outerIndex; + delete[] m_innerOffset; + delete[] m_outerOffset; + delete[] m_indices; + delete[] m_blockPtr; + delete[] m_values; + } + + + /** + * \brief Constructor from a sparse matrix + * + */ + template + inline BlockSparseMatrix(const MatrixType& spmat) : m_blockSize(BlockSize) + { + EIGEN_STATIC_ASSERT((m_blockSize != Dynamic), THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE); + + *this = spmat; + } + + /** + * \brief Assignment from a sparse matrix with the same storage order + * + * Convert from a sparse matrix to block sparse matrix. + * \warning Before calling this function, tt is necessary to call + * either setBlockLayout() (matrices with variable-size blocks) + * or setBlockSize() (for fixed-size blocks). + */ + template + inline BlockSparseMatrix& operator=(const MatrixType& spmat) + { + eigen_assert((m_innerBSize != 0 && m_outerBSize != 0) + && "Trying to assign to a zero-size matrix, call resize() first"); + eigen_assert(((MatrixType::Options&RowMajorBit) != IsColMajor) && "Wrong storage order"); + typedef SparseMatrix MatrixPatternType; + MatrixPatternType blockPattern(blockRows(), blockCols()); + m_nonzeros = 0; + + // First, compute the number of nonzero blocks and their locations + for(Index bj = 0; bj < m_outerBSize; ++bj) + { + // Browse each outer block and compute the structure + std::vector nzblocksFlag(m_innerBSize,false); // Record the existing blocks + blockPattern.startVec(bj); + for(Index j = blockOuterIndex(bj); j < blockOuterIndex(bj+1); ++j) + { + typename MatrixType::InnerIterator it_spmat(spmat, j); + for(; it_spmat; ++it_spmat) + { + Index bi = innerToBlock(it_spmat.index()); // Index of the current nonzero block + if(!nzblocksFlag[bi]) + { + // Save the index of this nonzero block + nzblocksFlag[bi] = true; + blockPattern.insertBackByOuterInnerUnordered(bj, bi) = true; + // Compute the total number of nonzeros (including explicit zeros in blocks) + m_nonzeros += blockOuterSize(bj) * blockInnerSize(bi); + } + } + } // end current outer block + } + blockPattern.finalize(); + + // Allocate the internal arrays + setBlockStructure(blockPattern); + + for(Index nz = 0; nz < m_nonzeros; ++nz) m_values[nz] = Scalar(0); + for(Index bj = 0; bj < m_outerBSize; ++bj) + { + // Now copy the values + for(Index j = blockOuterIndex(bj); j < blockOuterIndex(bj+1); ++j) + { + // Browse the outer block column by column (for column-major matrices) + typename MatrixType::InnerIterator it_spmat(spmat, j); + for(; it_spmat; ++it_spmat) + { + Index idx = 0; // Position of this block in the column block + Index bi = innerToBlock(it_spmat.index()); // Index of the current nonzero block + // Go to the inner block where this element belongs to + while(bi > m_indices[m_outerIndex[bj]+idx]) ++idx; // Not expensive for ordered blocks + Index idxVal;// Get the right position in the array of values for this element + if(m_blockSize == Dynamic) + { + // Offset from all blocks before ... + idxVal = m_blockPtr[m_outerIndex[bj]+idx]; + // ... and offset inside the block + idxVal += (j - blockOuterIndex(bj)) * blockOuterSize(bj) + it_spmat.index() - m_innerOffset[bi]; + } + else + { + // All blocks before + idxVal = (m_outerIndex[bj] + idx) * m_blockSize * m_blockSize; + // inside the block + idxVal += (j - blockOuterIndex(bj)) * m_blockSize + (it_spmat.index()%m_blockSize); + } + // Insert the value + m_values[idxVal] = it_spmat.value(); + } // end of this column + } // end of this block + } // end of this outer block + + return *this; + } + + /** + * \brief Set the nonzero block pattern of the matrix + * + * Given a sparse matrix describing the nonzero block pattern, + * this function prepares the internal pointers for values. + * After calling this function, any *nonzero* block (bi, bj) can be set + * with a simple call to coeffRef(bi,bj). + * + * + * \warning Before calling this function, tt is necessary to call + * either setBlockLayout() (matrices with variable-size blocks) + * or setBlockSize() (for fixed-size blocks). + * + * \param blockPattern Sparse matrix of boolean elements describing the block structure + * + * \sa setBlockLayout() \sa setBlockSize() + */ + template + void setBlockStructure(const MatrixType& blockPattern) + { + resize(blockPattern.rows(), blockPattern.cols()); + reserve(blockPattern.nonZeros()); + + // Browse the block pattern and set up the various pointers + m_outerIndex[0] = 0; + if(m_blockSize == Dynamic) m_blockPtr[0] = 0; + for(Index nz = 0; nz < m_nonzeros; ++nz) m_values[nz] = Scalar(0); + for(Index bj = 0; bj < m_outerBSize; ++bj) + { + //Browse each outer block + + //First, copy and save the indices of nonzero blocks + //FIXME : find a way to avoid this ... + std::vector nzBlockIdx; + typename MatrixType::InnerIterator it(blockPattern, bj); + for(; it; ++it) + { + nzBlockIdx.push_back(it.index()); + } + std::sort(nzBlockIdx.begin(), nzBlockIdx.end()); + + // Now, fill block indices and (eventually) pointers to blocks + for(Index idx = 0; idx < nzBlockIdx.size(); ++idx) + { + Index offset = m_outerIndex[bj]+idx; // offset in m_indices + m_indices[offset] = nzBlockIdx[idx]; + if(m_blockSize == Dynamic) + m_blockPtr[offset] = m_blockPtr[offset-1] + blockInnerSize(nzBlockIdx[idx]) * blockOuterSize(bj); + // There is no blockPtr for fixed-size blocks... not needed !??? + } + // Save the pointer to the next outer block + m_outerIndex[bj+1] = m_outerIndex[bj] + nzBlockIdx.size(); + } + } + + /** + * \brief Set the number of rows and columns blocks + */ + inline void resize(Index brow, Index bcol) + { + m_innerBSize = IsColMajor ? brow : bcol; + m_outerBSize = IsColMajor ? bcol : brow; + } + + /** + * \brief set the block size at runtime for fixed-size block layout + * + * Call this only for fixed-size blocks + */ + inline void setBlockSize(Index blockSize) + { + m_blockSize = blockSize; + } + + /** + * \brief Set the row and column block layouts, + * + * This function set the size of each row and column block. + * So this function should be used only for blocks with variable size. + * \param rowBlocks : Number of rows per row block + * \param colBlocks : Number of columns per column block + * \sa resize(), setBlockSize() + */ + inline void setBlockLayout(const VectorXi& rowBlocks, const VectorXi& colBlocks) + { + const VectorXi& innerBlocks = IsColMajor ? rowBlocks : colBlocks; + const VectorXi& outerBlocks = IsColMajor ? colBlocks : rowBlocks; + eigen_assert(m_innerBSize == innerBlocks.size() && "CHECK THE NUMBER OF ROW OR COLUMN BLOCKS"); + eigen_assert(m_outerBSize == outerBlocks.size() && "CHECK THE NUMBER OF ROW OR COLUMN BLOCKS"); + m_outerBSize = outerBlocks.size(); + // starting index of blocks... cumulative sums + m_innerOffset = new Index[m_innerBSize+1]; + m_outerOffset = new Index[m_outerBSize+1]; + m_innerOffset[0] = 0; + m_outerOffset[0] = 0; + std::partial_sum(&innerBlocks[0], &innerBlocks[m_innerBSize-1]+1, &m_innerOffset[1]); + std::partial_sum(&outerBlocks[0], &outerBlocks[m_outerBSize-1]+1, &m_outerOffset[1]); + + // Compute the total number of nonzeros + m_nonzeros = 0; + for(Index bj = 0; bj < m_outerBSize; ++bj) + for(Index bi = 0; bi < m_innerBSize; ++bi) + m_nonzeros += outerBlocks[bj] * innerBlocks[bi]; + + } + + /** + * \brief Allocate the internal array of pointers to blocks and their inner indices + * + * \note For fixed-size blocks, call setBlockSize() to set the block. + * And For variable-size blocks, call setBlockLayout() before using this function + * + * \param nonzerosblocks Number of nonzero blocks. The total number of nonzeros is + * is computed in setBlockLayout() for variable-size blocks + * \sa setBlockSize() + */ + inline void reserve(const Index nonzerosblocks) + { + eigen_assert((m_innerBSize != 0 && m_outerBSize != 0) && + "TRYING TO RESERVE ZERO-SIZE MATRICES, CALL resize() first"); + + //FIXME Should free if already allocated + m_outerIndex = new Index[m_outerBSize+1]; + + m_nonzerosblocks = nonzerosblocks; + if(m_blockSize != Dynamic) + { + m_nonzeros = nonzerosblocks * (m_blockSize * m_blockSize); + m_blockPtr = 0; + } + else + { + // m_nonzeros is already computed in setBlockLayout() + m_blockPtr = new Index[m_nonzerosblocks+1]; + } + m_indices = new Index[m_nonzerosblocks+1]; + m_values = new Scalar[m_nonzeros]; + } + + + /** + * \brief Fill values in a matrix from a triplet list. + * + * Each triplet item has a block stored in an Eigen dense matrix. + * The InputIterator class should provide the functions row(), col() and value() + * + * \note For fixed-size blocks, call setBlockSize() before this function. + * + * FIXME Do not accept duplicates + */ + template + void setFromTriplets(const InputIterator& begin, const InputIterator& end) + { + eigen_assert((m_innerBSize!=0 && m_outerBSize !=0) && "ZERO BLOCKS, PLEASE CALL resize() before"); + + /* First, sort the triplet list + * FIXME This can be unnecessarily expensive since only the inner indices have to be sorted + * The best approach is like in SparseMatrix::setFromTriplets() + */ + internal::TripletComp tripletcomp; + std::sort(begin, end, tripletcomp); + + /* Count the number of rows and column blocks, + * and the number of nonzero blocks per outer dimension + */ + VectorXi rowBlocks(m_innerBSize); // Size of each block row + VectorXi colBlocks(m_outerBSize); // Size of each block column + rowBlocks.setZero(); colBlocks.setZero(); + VectorXi nzblock_outer(m_outerBSize); // Number of nz blocks per outer vector + VectorXi nz_outer(m_outerBSize); // Number of nz per outer vector...for variable-size blocks + nzblock_outer.setZero(); + nz_outer.setZero(); + for(InputIterator it(begin); it !=end; ++it) + { + eigen_assert(it->row() >= 0 && it->row() < this->blockRows() && it->col() >= 0 && it->col() < this->blockCols()); + eigen_assert((it->value().rows() == it->value().cols() && (it->value().rows() == m_blockSize)) + || (m_blockSize == Dynamic)); + + if(m_blockSize == Dynamic) + { + eigen_assert((rowBlocks[it->row()] == 0 || rowBlocks[it->row()] == it->value().rows()) && + "NON CORRESPONDING SIZES FOR ROW BLOCKS"); + eigen_assert((colBlocks[it->col()] == 0 || colBlocks[it->col()] == it->value().cols()) && + "NON CORRESPONDING SIZES FOR COLUMN BLOCKS"); + rowBlocks[it->row()] =it->value().rows(); + colBlocks[it->col()] = it->value().cols(); + } + nz_outer(IsColMajor ? it->col() : it->row()) += it->value().rows() * it->value().cols(); + nzblock_outer(IsColMajor ? it->col() : it->row())++; + } + // Allocate member arrays + if(m_blockSize == Dynamic) setBlockLayout(rowBlocks, colBlocks); + Index nzblocks = nzblock_outer.sum(); + reserve(nzblocks); + + // Temporary markers + VectorXi block_id(m_outerBSize); // To be used as a block marker during insertion + + // Setup outer index pointers and markers + m_outerIndex[0] = 0; + if (m_blockSize == Dynamic) m_blockPtr[0] = 0; + for(Index bj = 0; bj < m_outerBSize; ++bj) + { + m_outerIndex[bj+1] = m_outerIndex[bj] + nzblock_outer(bj); + block_id(bj) = m_outerIndex[bj]; + if(m_blockSize==Dynamic) + { + m_blockPtr[m_outerIndex[bj+1]] = m_blockPtr[m_outerIndex[bj]] + nz_outer(bj); + } + } + + // Fill the matrix + for(InputIterator it(begin); it!=end; ++it) + { + Index outer = IsColMajor ? it->col() : it->row(); + Index inner = IsColMajor ? it->row() : it->col(); + m_indices[block_id(outer)] = inner; + Index block_size = it->value().rows()*it->value().cols(); + Index nz_marker = blockPtr(block_id[outer]); + memcpy(&(m_values[nz_marker]), it->value().data(), block_size * sizeof(Scalar)); + if(m_blockSize == Dynamic) + { + m_blockPtr[block_id(outer)+1] = m_blockPtr[block_id(outer)] + block_size; + } + block_id(outer)++; + } + + // An alternative when the outer indices are sorted...no need to use an array of markers +// for(Index bcol = 0; bcol < m_outerBSize; ++bcol) +// { +// Index id = 0, id_nz = 0, id_nzblock = 0; +// for(InputIterator it(begin); it!=end; ++it) +// { +// while (idvalue().rows()*it->value().cols(); +// m_blockPtr[id_nzblock+1] = m_blockPtr[id_nzblock] + block_size; +// id_nzblock++; +// memcpy(&(m_values[id_nz]),it->value().data(), block_size*sizeof(Scalar)); +// id_nz += block_size; +// } +// while(id < m_outerBSize-1) // Empty columns at the end +// { +// id++; +// m_outerIndex[id+1]=m_outerIndex[id]; +// } +// } + } + + + /** + * \returns the number of rows + */ + inline Index rows() const + { +// return blockRows(); + return (IsColMajor ? innerSize() : outerSize()); + } + + /** + * \returns the number of cols + */ + inline Index cols() const + { +// return blockCols(); + return (IsColMajor ? outerSize() : innerSize()); + } + + inline Index innerSize() const + { + if(m_blockSize == Dynamic) return m_innerOffset[m_innerBSize]; + else return (m_innerBSize * m_blockSize) ; + } + + inline Index outerSize() const + { + if(m_blockSize == Dynamic) return m_outerOffset[m_outerBSize]; + else return (m_outerBSize * m_blockSize) ; + } + /** \returns the number of rows grouped by blocks */ + inline Index blockRows() const + { + return (IsColMajor ? m_innerBSize : m_outerBSize); + } + /** \returns the number of columns grouped by blocks */ + inline Index blockCols() const + { + return (IsColMajor ? m_outerBSize : m_innerBSize); + } + + inline Index outerBlocks() const { return m_outerBSize; } + inline Index innerBlocks() const { return m_innerBSize; } + + /** \returns the block index where outer belongs to */ + inline Index outerToBlock(Index outer) const + { + eigen_assert(outer < outerSize() && "OUTER INDEX OUT OF BOUNDS"); + + if(m_blockSize != Dynamic) + return (outer / m_blockSize); // Integer division + + Index b_outer = 0; + while(m_outerOffset[b_outer] <= outer) ++b_outer; + return b_outer - 1; + } + /** \returns the block index where inner belongs to */ + inline Index innerToBlock(Index inner) const + { + eigen_assert(inner < innerSize() && "OUTER INDEX OUT OF BOUNDS"); + + if(m_blockSize != Dynamic) + return (inner / m_blockSize); // Integer division + + Index b_inner = 0; + while(m_innerOffset[b_inner] <= inner) ++b_inner; + return b_inner - 1; + } + + /** + *\returns a reference to the (i,j) block as an Eigen Dense Matrix + */ + Ref coeffRef(Index brow, Index bcol) + { + eigen_assert(brow < blockRows() && "BLOCK ROW INDEX OUT OF BOUNDS"); + eigen_assert(bcol < blockCols() && "BLOCK nzblocksFlagCOLUMN OUT OF BOUNDS"); + + Index rsize = IsColMajor ? blockInnerSize(brow): blockOuterSize(bcol); + Index csize = IsColMajor ? blockOuterSize(bcol) : blockInnerSize(brow); + Index inner = IsColMajor ? brow : bcol; + Index outer = IsColMajor ? bcol : brow; + Index offset = m_outerIndex[outer]; + while(offset < m_outerIndex[outer+1] && m_indices[offset] != inner) + offset++; + if(m_indices[offset] == inner) + { + return Map(&(m_values[blockPtr(offset)]), rsize, csize); + } + else + { + //FIXME the block does not exist, Insert it !!!!!!!!! + eigen_assert("DYNAMIC INSERTION IS NOT YET SUPPORTED"); + } + } + + /** + * \returns the value of the (i,j) block as an Eigen Dense Matrix + */ + Map coeff(Index brow, Index bcol) const + { + eigen_assert(brow < blockRows() && "BLOCK ROW INDEX OUT OF BOUNDS"); + eigen_assert(bcol < blockCols() && "BLOCK COLUMN OUT OF BOUNDS"); + + Index rsize = IsColMajor ? blockInnerSize(brow): blockOuterSize(bcol); + Index csize = IsColMajor ? blockOuterSize(bcol) : blockInnerSize(brow); + Index inner = IsColMajor ? brow : bcol; + Index outer = IsColMajor ? bcol : brow; + Index offset = m_outerIndex[outer]; + while(offset < m_outerIndex[outer+1] && m_indices[offset] != inner) offset++; + if(m_indices[offset] == inner) + { + return Map (&(m_values[blockPtr(offset)]), rsize, csize); + } + else +// return BlockScalar::Zero(rsize, csize); + eigen_assert("NOT YET SUPPORTED"); + } + + // Block Matrix times vector product + template + BlockSparseTimeDenseProduct operator*(const VecType& lhs) const + { + return BlockSparseTimeDenseProduct(*this, lhs); + } + + /** \returns the number of nonzero blocks */ + inline Index nonZerosBlocks() const { return m_nonzerosblocks; } + /** \returns the total number of nonzero elements, including eventual explicit zeros in blocks */ + inline Index nonZeros() const { return m_nonzeros; } + + inline BlockScalarReturnType *valuePtr() {return static_cast(m_values);} +// inline Scalar *valuePtr(){ return m_values; } + inline Index *innerIndexPtr() {return m_indices; } + inline const Index *innerIndexPtr() const {return m_indices; } + inline Index *outerIndexPtr() {return m_outerIndex; } + inline const Index* outerIndexPtr() const {return m_outerIndex; } + + /** \brief for compatibility purposes with the SparseMatrix class */ + inline bool isCompressed() const {return true;} + /** + * \returns the starting index of the bi row block + */ + inline Index blockRowsIndex(Index bi) const + { + return IsColMajor ? blockInnerIndex(bi) : blockOuterIndex(bi); + } + + /** + * \returns the starting index of the bj col block + */ + inline Index blockColsIndex(Index bj) const + { + return IsColMajor ? blockOuterIndex(bj) : blockInnerIndex(bj); + } + + inline Index blockOuterIndex(Index bj) const + { + return (m_blockSize == Dynamic) ? m_outerOffset[bj] : (bj * m_blockSize); + } + inline Index blockInnerIndex(Index bi) const + { + return (m_blockSize == Dynamic) ? m_innerOffset[bi] : (bi * m_blockSize); + } + + // Not needed ??? + inline Index blockInnerSize(Index bi) const + { + return (m_blockSize == Dynamic) ? (m_innerOffset[bi+1] - m_innerOffset[bi]) : m_blockSize; + } + inline Index blockOuterSize(Index bj) const + { + return (m_blockSize == Dynamic) ? (m_outerOffset[bj+1]- m_outerOffset[bj]) : m_blockSize; + } + + /** + * \brief Browse the matrix by outer index + */ + class InnerIterator; // Browse column by column + + /** + * \brief Browse the matrix by block outer index + */ + class BlockInnerIterator; // Browse block by block + + friend std::ostream & operator << (std::ostream & s, const BlockSparseMatrix& m) + { + for (Index j = 0; j < m.outerBlocks(); ++j) + { + BlockInnerIterator itb(m, j); + for(; itb; ++itb) + { + s << "("< in the array of values + */ + Index blockPtr(Index id) const + { + if(m_blockSize == Dynamic) return m_blockPtr[id]; + else return id * m_blockSize * m_blockSize; + //return blockDynIdx(id, typename internal::conditional<(BlockSize==Dynamic), internal::true_type, internal::false_type>::type()); + } + + + protected: +// inline Index blockDynIdx(Index id, internal::true_type) const +// { +// return m_blockPtr[id]; +// } +// inline Index blockDynIdx(Index id, internal::false_type) const +// { +// return id * BlockSize * BlockSize; +// } + + // To be implemented + // Insert a block at a particular location... need to make a room for that + Map insert(Index brow, Index bcol); + + Index m_innerBSize; // Number of block rows + Index m_outerBSize; // Number of block columns + Index *m_innerOffset; // Starting index of each inner block (size m_innerBSize+1) + Index *m_outerOffset; // Starting index of each outer block (size m_outerBSize+1) + Index m_nonzerosblocks; // Total nonzeros blocks (lower than m_innerBSize x m_outerBSize) + Index m_nonzeros; // Total nonzeros elements + Scalar *m_values; //Values stored block column after block column (size m_nonzeros) + Index *m_blockPtr; // Pointer to the beginning of each block in m_values, size m_nonzeroblocks ... null for fixed-size blocks + Index *m_indices; //Inner block indices, size m_nonzerosblocks ... OK + Index *m_outerIndex; // Starting pointer of each block column in m_indices (size m_outerBSize)... OK + Index m_blockSize; // Size of a block for fixed-size blocks, otherwise -1 +}; + +template +class BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _Index>::BlockInnerIterator +{ + public: + + enum{ + Flags = _Options + }; + + BlockInnerIterator(const BlockSparseMatrix& mat, const Index outer) + : m_mat(mat),m_outer(outer), + m_id(mat.m_outerIndex[outer]), + m_end(mat.m_outerIndex[outer+1]) + { + } + + inline BlockInnerIterator& operator++() {m_id++; return *this; } + + inline const Map value() const + { + return Map(&(m_mat.m_values[m_mat.blockPtr(m_id)]), + rows(),cols()); + } + inline Map valueRef() + { + return Map(&(m_mat.m_values[m_mat.blockPtr(m_id)]), + rows(),cols()); + } + // Block inner index + inline Index index() const {return m_mat.m_indices[m_id]; } + inline Index outer() const { return m_outer; } + // block row index + inline Index row() const {return index(); } + // block column index + inline Index col() const {return outer(); } + // FIXME Number of rows in the current block + inline Index rows() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_innerOffset[index()+1] - m_mat.m_innerOffset[index()]) : m_mat.m_blockSize; } + // Number of columns in the current block ... + inline Index cols() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_outerOffset[m_outer+1]-m_mat.m_outerOffset[m_outer]) : m_mat.m_blockSize;} + inline operator bool() const { return (m_id < m_end); } + + protected: + const BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, Index>& m_mat; + const Index m_outer; + Index m_id; + Index m_end; +}; + +template +class BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _Index>::InnerIterator +{ + public: + InnerIterator(const BlockSparseMatrix& mat, Index outer) + : m_mat(mat),m_outerB(mat.outerToBlock(outer)),m_outer(outer), + itb(mat, mat.outerToBlock(outer)), + m_offset(outer - mat.blockOuterIndex(m_outerB)) + { + if (itb) + { + m_id = m_mat.blockInnerIndex(itb.index()); + m_start = m_id; + m_end = m_mat.blockInnerIndex(itb.index()+1); + } + } + inline InnerIterator& operator++() + { + m_id++; + if (m_id >= m_end) + { + ++itb; + if (itb) + { + m_id = m_mat.blockInnerIndex(itb.index()); + m_start = m_id; + m_end = m_mat.blockInnerIndex(itb.index()+1); + } + } + return *this; + } + inline const Scalar& value() const + { + return itb.value().coeff(m_id - m_start, m_offset); + } + inline Scalar& valueRef() + { + return itb.valueRef().coeff(m_id - m_start, m_offset); + } + inline Index index() const { return m_id; } + inline Index outer() const {return m_outer; } + inline Index col() const {return outer(); } + inline Index row() const { return index();} + inline operator bool() const + { + return itb; + } + protected: + const BlockSparseMatrix& m_mat; + const Index m_outer; + const Index m_outerB; + BlockInnerIterator itb; // Iterator through the blocks + const Index m_offset; // Position of this column in the block + Index m_start; // starting inner index of this block + Index m_id; // current inner index in the block + Index m_end; // starting inner index of the next block + +}; +} // end namespace Eigen + +#endif // EIGEN_SPARSEBLOCKMATRIX_H From 00dc45d0f912407366b20e132df76cffbbf22449 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 20 Sep 2013 23:28:10 +0200 Subject: [PATCH 0132/1103] Reduce explicit zeros when applying SparseQR's matrix Q --- Eigen/src/SparseQR/SparseQR.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index e9b2a6bec..267c48bc3 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -573,6 +573,7 @@ struct SparseQR_QProduct : ReturnByValue Date: Tue, 24 Sep 2013 15:56:56 +0200 Subject: [PATCH 0133/1103] Fix leaked memory for successive calls to SPQR --- Eigen/src/SPQRSupport/SuiteSparseQRSupport.h | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h index aa41f434c..e1a320446 100644 --- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h @@ -64,7 +64,8 @@ class SPQR typedef PermutationMatrix PermutationType; public: SPQR() - : m_ordering(SPQR_ORDERING_DEFAULT), + : m_isInitialized(false), + m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) { @@ -72,7 +73,8 @@ class SPQR } SPQR(const _MatrixType& matrix) - : m_ordering(SPQR_ORDERING_DEFAULT), + : m_isInitialized(false), + m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) { @@ -82,16 +84,22 @@ class SPQR ~SPQR() { - // Calls SuiteSparseQR_free() + SPQR_free(); + cholmod_l_finish(&m_cc); + } + void SPQR_free() + { cholmod_l_free_sparse(&m_H, &m_cc); cholmod_l_free_sparse(&m_cR, &m_cc); cholmod_l_free_dense(&m_HTau, &m_cc); std::free(m_E); std::free(m_HPinv); - cholmod_l_finish(&m_cc); } + void compute(const _MatrixType& matrix) { + if(m_isInitialized) SPQR_free(); + MatrixType mat(matrix); cholmod_sparse A; A = viewAsCholmod(mat); From 54e576c88af67aa2c8a66d95a104a4ffc5fd1985 Mon Sep 17 00:00:00 2001 From: Desire NUENTSA Date: Thu, 26 Sep 2013 15:00:22 +0200 Subject: [PATCH 0134/1103] Fix SPQR Solve() when assigning to a Map object --- Eigen/src/SPQRSupport/SuiteSparseQRSupport.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h index e1a320446..a2cc2a9e2 100644 --- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h @@ -147,7 +147,7 @@ class SPQR eigen_assert(b.cols()==1 && "This method is for vectors only"); //Compute Q^T * b - Dest y; + typename Dest::PlainObject y; y = matrixQ().transpose() * b; // Solves with the triangular matrix R Index rk = this->rank(); From 446320b226c152ced3e42dbff08049a3f82ffee7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Oct 2013 22:37:10 +0200 Subject: [PATCH 0135/1103] Fix dot*w to return 0 for empty vectors (BLAS interface) --- blas/level1_cplx_impl.h | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-) diff --git a/blas/level1_cplx_impl.h b/blas/level1_cplx_impl.h index 283b9f827..ffe192481 100644 --- a/blas/level1_cplx_impl.h +++ b/blas/level1_cplx_impl.h @@ -39,13 +39,16 @@ RealScalar EIGEN_CAT(EIGEN_CAT(REAL_SCALAR_SUFFIX,SCALAR_SUFFIX),asum_)(int *n, // computes a dot product of a conjugated vector with another vector. int EIGEN_BLAS_FUNC(dotcw)(int *n, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar* pres) { -// std::cerr << "_dotc " << *n << " " << *incx << " " << *incy << "\n"; + Scalar* res = reinterpret_cast(pres); - if(*n<=0) return 0; + if(*n<=0) + { + *res = Scalar(0); + return 0; + } Scalar* x = reinterpret_cast(px); Scalar* y = reinterpret_cast(py); - Scalar* res = reinterpret_cast(pres); if(*incx==1 && *incy==1) *res = (vector(x,*n).dot(vector(y,*n))); else if(*incx>0 && *incy>0) *res = (vector(x,*n,*incx).dot(vector(y,*n,*incy))); @@ -58,13 +61,16 @@ int EIGEN_BLAS_FUNC(dotcw)(int *n, RealScalar *px, int *incx, RealScalar *py, in // computes a vector-vector dot product without complex conjugation. int EIGEN_BLAS_FUNC(dotuw)(int *n, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar* pres) { -// std::cerr << "_dotu " << *n << " " << *incx << " " << *incy << "\n"; + Scalar* res = reinterpret_cast(pres); - if(*n<=0) return 0; + if(*n<=0) + { + *res = Scalar(0); + return 0; + } Scalar* x = reinterpret_cast(px); Scalar* y = reinterpret_cast(py); - Scalar* res = reinterpret_cast(pres); if(*incx==1 && *incy==1) *res = (vector(x,*n).cwiseProduct(vector(y,*n))).sum(); else if(*incx>0 && *incy>0) *res = (vector(x,*n,*incx).cwiseProduct(vector(y,*n,*incy))).sum(); From 40f1548b32671ed27b3a7d02ea64c2291a34443f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 2 Oct 2013 23:31:59 +0200 Subject: [PATCH 0136/1103] Sparse is stable now, so Eigen/Eigen should include Sparse --- Eigen/Eigen | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/Eigen b/Eigen/Eigen index 19b40ea4e..654c8dc63 100644 --- a/Eigen/Eigen +++ b/Eigen/Eigen @@ -1,2 +1,2 @@ #include "Dense" -//#include "Sparse" +#include "Sparse" From 3736e00ae719d745fbde5012d8fe9047e85828c3 Mon Sep 17 00:00:00 2001 From: vanhoucke Date: Fri, 4 Oct 2013 00:21:03 +0000 Subject: [PATCH 0137/1103] Silence unused variable warning. --- Eigen/src/Eigen2Support/SVD.h | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/Eigen/src/Eigen2Support/SVD.h b/Eigen/src/Eigen2Support/SVD.h index 077d26d54..3d03d2288 100644 --- a/Eigen/src/Eigen2Support/SVD.h +++ b/Eigen/src/Eigen2Support/SVD.h @@ -512,8 +512,7 @@ template template bool SVD::solve(const MatrixBase &b, ResultType* result) const { - const int rows = m_matU.rows(); - ei_assert(b.rows() == rows); + ei_assert(b.rows() == m_matU.rows()); Scalar maxVal = m_sigma.cwise().abs().maxCoeff(); for (int j=0; j Date: Wed, 9 Oct 2013 10:25:50 +0200 Subject: [PATCH 0138/1103] Add cmake config files --- CMakeLists.txt | 28 ++++++++++++++++++++++++++++ cmake/Eigen3Config.cmake.in | 28 ++++++++++++++++++++++++++++ cmake/UseEigen3.cmake | 6 ++++++ 3 files changed, 62 insertions(+) create mode 100644 cmake/Eigen3Config.cmake.in create mode 100644 cmake/UseEigen3.cmake diff --git a/CMakeLists.txt b/CMakeLists.txt index ad0269ea6..4135339b1 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -418,3 +418,31 @@ else() endif() message(STATUS "") + +set ( EIGEN_CONFIG_CMAKE_PATH + lib${LIB_SUFFIX}/cmake/eigen3 + CACHE PATH "The directory where the CMake files are installed" + ) +if ( NOT IS_ABSOLUTE EIGEN_CONFIG_CMAKE_PATH ) + set ( EIGEN_CONFIG_CMAKE_PATH ${CMAKE_INSTALL_PREFIX}/${EIGEN_CONFIG_CMAKE_PATH} ) +endif () + +set ( EIGEN_USE_FILE ${EIGEN_CONFIG_CMAKE_PATH}/UseEigen3.cmake ) +set ( EIGEN_VERSION_STRING ${EIGEN_VERSION_NUMBER} ) +set ( EIGEN_VERSION_MAJOR ${EIGEN_WORLD_VERSION} ) +set ( EIGEN_VERSION_MINOR ${EIGEN_MAJOR_VERSION} ) +set ( EIGEN_VERSION_PATCH ${EIGEN_MINOR_VERSION} ) +set ( EIGEN_DEFINITIONS "") +set ( EIGEN_INCLUDE_DIR ${INCLUDE_INSTALL_DIR} ) +set ( EIGEN_INCLUDE_DIRS ${EIGEN_INCLUDE_DIR} ) +set ( EIGEN_ROOT_DIR ${CMAKE_INSTALL_PREFIX} ) + +configure_file ( ${CMAKE_SOURCE_DIR}/cmake/Eigen3Config.cmake.in + ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake + @ONLY ESCAPE_QUOTES + ) + +install ( FILES ${CMAKE_SOURCE_DIR}/cmake/UseEigen3.cmake + ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake + DESTINATION ${EIGEN_CONFIG_CMAKE_PATH} + ) diff --git a/cmake/Eigen3Config.cmake.in b/cmake/Eigen3Config.cmake.in new file mode 100644 index 000000000..257c595ed --- /dev/null +++ b/cmake/Eigen3Config.cmake.in @@ -0,0 +1,28 @@ +# -*- cmake -*- +# +# Eigen3Config.cmake(.in) + +# Use the following variables to compile and link against Eigen: +# EIGEN_FOUND - True if Eigen was found on your system +# EIGEN_USE_FILE - The file making Eigen usable +# EIGEN_DEFINITIONS - Definitions needed to build with Eigen +# EIGEN_INCLUDE_DIR - Directory where signature_of_eigen3_matrix_library can be found +# EIGEN_INCLUDE_DIRS - List of directories of Eigen and it's dependencies +# EIGEN_ROOT_DIR - The base directory of Eigen +# EIGEN_VERSION_STRING - A human-readable string containing the version +# EIGEN_VERSION_MAJOR - The major version of Eigen +# EIGEN_VERSION_MINOR - The minor version of Eigen +# EIGEN_VERSION_PATCH - The patch version of Eigen + +set ( EIGEN_FOUND 1 ) +set ( EIGEN_USE_FILE "@EIGEN_USE_FILE@" ) + +set ( EIGEN_DEFINITIONS "@EIGEN_DEFINITIONS@" ) +set ( EIGEN_INCLUDE_DIR "@EIGEN_INCLUDE_DIR@" ) +set ( EIGEN_INCLUDE_DIRS "@EIGEN_INCLUDE_DIRS@" ) +set ( EIGEN_ROOT_DIR "@EIGEN_ROOT_DIR@" ) + +set ( EIGEN_VERSION_STRING "@EIGEN_VERSION_STRING@" ) +set ( EIGEN_VERSION_MAJOR "@EIGEN_VERSION_MAJOR@" ) +set ( EIGEN_VERSION_MINOR "@EIGEN_VERSION_MINOR@" ) +set ( EIGEN_VERSION_PATCH "@EIGEN_VERSION_PATCH@" ) diff --git a/cmake/UseEigen3.cmake b/cmake/UseEigen3.cmake new file mode 100644 index 000000000..a38bac82d --- /dev/null +++ b/cmake/UseEigen3.cmake @@ -0,0 +1,6 @@ +# -*- cmake -*- +# +# UseEigen3.cmake + +add_definitions ( ${EIGEN3_DEFINITIONS} ) +include_directories ( ${EIGEN3_INCLUDE_DIRS} ) From 6bef527f9dbf470bc78655a8311c20c1508db7ec Mon Sep 17 00:00:00 2001 From: Thomas Capricelli Date: Fri, 11 Oct 2013 20:46:18 +0200 Subject: [PATCH 0139/1103] uniformize piwik code among branches --- doc/eigendoxy_footer.html.in | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/doc/eigendoxy_footer.html.in b/doc/eigendoxy_footer.html.in index e62af8ed5..878244a19 100644 --- a/doc/eigendoxy_footer.html.in +++ b/doc/eigendoxy_footer.html.in @@ -17,8 +17,7 @@ $generatedby   - - ---> From 0bce534c8f48e36223b302b21855fe10fc9012db Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 15 Oct 2013 19:09:09 +0200 Subject: [PATCH 0140/1103] Fix bug #679 --- Eigen/src/plugins/MatrixCwiseBinaryOps.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/plugins/MatrixCwiseBinaryOps.h b/Eigen/src/plugins/MatrixCwiseBinaryOps.h index 3a737df7b..7f62149e0 100644 --- a/Eigen/src/plugins/MatrixCwiseBinaryOps.h +++ b/Eigen/src/plugins/MatrixCwiseBinaryOps.h @@ -83,7 +83,7 @@ cwiseMin(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const ConstantReturnType> cwiseMin(const Scalar &other) const { - return cwiseMin(Derived::PlainObject::Constant(rows(), cols(), other)); + return cwiseMin(Derived::Constant(rows(), cols(), other)); } /** \returns an expression of the coefficient-wise max of *this and \a other @@ -107,7 +107,7 @@ cwiseMax(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const EIGEN_STRONG_INLINE const CwiseBinaryOp, const Derived, const ConstantReturnType> cwiseMax(const Scalar &other) const { - return cwiseMax(Derived::PlainObject::Constant(rows(), cols(), other)); + return cwiseMax(Derived::Constant(rows(), cols(), other)); } From 2c0303c89eb4c1b979c6dcd2379c9d3f98ba4c99 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 15 Oct 2013 23:51:01 +0200 Subject: [PATCH 0141/1103] bug #679: add respective unit test --- test/array_for_matrix.cpp | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/test/array_for_matrix.cpp b/test/array_for_matrix.cpp index 67e7a2a44..9a50f99ab 100644 --- a/test/array_for_matrix.cpp +++ b/test/array_for_matrix.cpp @@ -163,10 +163,14 @@ template void cwise_min_max(const MatrixType& m) // min/max with scalar input VERIFY_IS_APPROX(MatrixType::Constant(rows,cols, minM1), m1.cwiseMin( minM1)); - VERIFY_IS_APPROX(m1, m1.cwiseMin( maxM1)); + VERIFY_IS_APPROX(m1, m1.cwiseMin(maxM1)); + VERIFY_IS_APPROX(-m1, (-m1).cwiseMin(-minM1)); + VERIFY_IS_APPROX(-m1.array(), ((-m1).array().min)( -minM1)); VERIFY_IS_APPROX(MatrixType::Constant(rows,cols, maxM1), m1.cwiseMax( maxM1)); - VERIFY_IS_APPROX(m1, m1.cwiseMax( minM1)); + VERIFY_IS_APPROX(m1, m1.cwiseMax(minM1)); + VERIFY_IS_APPROX(-m1, (-m1).cwiseMax(-maxM1)); + VERIFY_IS_APPROX(-m1.array(), ((-m1).array().max)(-maxM1)); VERIFY_IS_APPROX(MatrixType::Constant(rows,cols, minM1).array(), (m1.array().min)( minM1)); VERIFY_IS_APPROX(m1.array(), (m1.array().min)( maxM1)); From b433fb28576af729251594dbccd659607be44280 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 16 Oct 2013 12:07:33 +0200 Subject: [PATCH 0142/1103] Allow .conservativeResize(rows,cols) on vectors --- Eigen/src/Core/PlainObjectBase.h | 16 ++++++++++++---- test/conservative_resize.cpp | 31 ++++++++++++++++++++++++------- 2 files changed, 36 insertions(+), 11 deletions(-) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index e9aa3a88c..1a4c19233 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -47,7 +47,10 @@ template<> struct check_rows_cols_for_overflow { } }; -template struct conservative_resize_like_impl; +template +struct conservative_resize_like_impl; template struct matrix_swap_impl; @@ -682,8 +685,10 @@ private: enum { ThisConstantIsPrivateInPlainObjectBase }; }; +namespace internal { + template -struct internal::conservative_resize_like_impl +struct conservative_resize_like_impl { typedef typename Derived::Index Index; static void run(DenseBase& _this, Index rows, Index cols) @@ -743,11 +748,14 @@ struct internal::conservative_resize_like_impl } }; -namespace internal { - +// Here, the specialization for vectors inherits from the general matrix case +// to allow calling .conservativeResize(rows,cols) on vectors. template struct conservative_resize_like_impl + : conservative_resize_like_impl { + using conservative_resize_like_impl::run; + typedef typename Derived::Index Index; static void run(DenseBase& _this, Index size) { diff --git a/test/conservative_resize.cpp b/test/conservative_resize.cpp index 2d1ab3f03..498421b4c 100644 --- a/test/conservative_resize.cpp +++ b/test/conservative_resize.cpp @@ -60,34 +60,51 @@ void run_matrix_tests() template void run_vector_tests() { - typedef Matrix MatrixType; + typedef Matrix VectorType; - MatrixType m, n; + VectorType m, n; // boundary cases ... - m = n = MatrixType::Random(50); + m = n = VectorType::Random(50); m.conservativeResize(1); VERIFY_IS_APPROX(m, n.segment(0,1)); - m = n = MatrixType::Random(50); + m = n = VectorType::Random(50); m.conservativeResize(50); VERIFY_IS_APPROX(m, n.segment(0,50)); + + m = n = VectorType::Random(50); + m.conservativeResize(m.rows(),1); + VERIFY_IS_APPROX(m, n.segment(0,1)); + + m = n = VectorType::Random(50); + m.conservativeResize(m.rows(),50); + VERIFY_IS_APPROX(m, n.segment(0,50)); // random shrinking ... for (int i=0; i<50; ++i) { const int size = internal::random(1,50); - m = n = MatrixType::Random(50); + m = n = VectorType::Random(50); m.conservativeResize(size); VERIFY_IS_APPROX(m, n.segment(0,size)); + + m = n = VectorType::Random(50); + m.conservativeResize(m.rows(), size); + VERIFY_IS_APPROX(m, n.segment(0,size)); } // random growing with zeroing ... for (int i=0; i<50; ++i) { const int size = internal::random(50,100); - m = n = MatrixType::Random(50); - m.conservativeResizeLike(MatrixType::Zero(size)); + m = n = VectorType::Random(50); + m.conservativeResizeLike(VectorType::Zero(size)); + VERIFY_IS_APPROX(m.segment(0,50), n); + VERIFY( size<=50 || m.segment(50,size-50).sum() == Scalar(0) ); + + m = n = VectorType::Random(50); + m.conservativeResizeLike(Matrix::Zero(1,size)); VERIFY_IS_APPROX(m.segment(0,50), n); VERIFY( size<=50 || m.segment(50,size-50).sum() == Scalar(0) ); } From 4a4284351365554708b4801d408456f0b360c507 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 16 Oct 2013 13:08:09 +0200 Subject: [PATCH 0143/1103] Make index type of Triplet default to SparseMatrix::Index as suggested by Kolja Brix. Fixes bug #665. --- Eigen/src/SparseCore/SparseUtil.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 064a40707..05023858b 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -143,7 +143,7 @@ template struct plain_matrix_type * * \sa SparseMatrix::setFromTriplets() */ -template +template::Index > class Triplet { public: From b61facb08b1e6c49724f27e15d72b490c6c07d8f Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 16 Oct 2013 13:10:15 +0200 Subject: [PATCH 0144/1103] Use != instead of < to check for emptiness of iterator range (fixes bug #664) --- Eigen/src/SparseCore/SparseMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 4dad50562..f2e234eee 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -941,7 +941,7 @@ void set_from_triplets(const InputIterator& begin, const InputIterator& end, Spa typedef typename SparseMatrixType::Scalar Scalar; SparseMatrix trMat(mat.rows(),mat.cols()); - if(begin Date: Wed, 16 Oct 2013 15:25:39 +0200 Subject: [PATCH 0145/1103] Fix bug #674: typo in documentation example for BiCGSTAB. They are now proper snippet files. --- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 25 ++------------------- doc/snippets/BiCGSTAB_simple.cpp | 11 +++++++++ doc/snippets/BiCGSTAB_step_by_step.cpp | 14 ++++++++++++ doc/snippets/compile_snippet.cpp.in | 2 +- 4 files changed, 28 insertions(+), 24 deletions(-) create mode 100644 doc/snippets/BiCGSTAB_simple.cpp create mode 100644 doc/snippets/BiCGSTAB_step_by_step.cpp diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index 6fc6ab852..7a46b51fa 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -136,34 +136,13 @@ struct traits > * and NumTraits::epsilon() for the tolerance. * * This class can be used as the direct solver classes. Here is a typical usage example: - * \code - * int n = 10000; - * VectorXd x(n), b(n); - * SparseMatrix A(n,n); - * // fill A and b - * BiCGSTAB > solver; - * solver(A); - * x = solver.solve(b); - * std::cout << "#iterations: " << solver.iterations() << std::endl; - * std::cout << "estimated error: " << solver.error() << std::endl; - * // update b, and solve again - * x = solver.solve(b); - * \endcode + * \include BiCGSTAB_simple.cpp * * By default the iterations start with x=0 as an initial guess of the solution. * One can control the start using the solveWithGuess() method. Here is a step by * step execution example starting with a random guess and printing the evolution * of the estimated error: - * * \code - * x = VectorXd::Random(n); - * solver.setMaxIterations(1); - * int i = 0; - * do { - * x = solver.solveWithGuess(b,x); - * std::cout << i << " : " << solver.error() << std::endl; - * ++i; - * } while (solver.info()!=Success && i<100); - * \endcode + * \include BiCGSTAB_step_by_step.cpp * Note that such a step by step excution is slightly slower. * * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner diff --git a/doc/snippets/BiCGSTAB_simple.cpp b/doc/snippets/BiCGSTAB_simple.cpp new file mode 100644 index 000000000..5520f4f1f --- /dev/null +++ b/doc/snippets/BiCGSTAB_simple.cpp @@ -0,0 +1,11 @@ + int n = 10000; + VectorXd x(n), b(n); + SparseMatrix A(n,n); + /* ... fill A and b ... */ + BiCGSTAB > solver; + solver.compute(A); + x = solver.solve(b); + std::cout << "#iterations: " << solver.iterations() << std::endl; + std::cout << "estimated error: " << solver.error() << std::endl; + /* ... update b ... */ + x = solver.solve(b); // solve again \ No newline at end of file diff --git a/doc/snippets/BiCGSTAB_step_by_step.cpp b/doc/snippets/BiCGSTAB_step_by_step.cpp new file mode 100644 index 000000000..06147bb81 --- /dev/null +++ b/doc/snippets/BiCGSTAB_step_by_step.cpp @@ -0,0 +1,14 @@ + int n = 10000; + VectorXd x(n), b(n); + SparseMatrix A(n,n); + /* ... fill A and b ... */ + BiCGSTAB > solver(A); + // start from a random solution + x = VectorXd::Random(n); + solver.setMaxIterations(1); + int i = 0; + do { + x = solver.solveWithGuess(b,x); + std::cout << i << " : " << solver.error() << std::endl; + ++i; + } while (solver.info()!=Success && i<100); \ No newline at end of file diff --git a/doc/snippets/compile_snippet.cpp.in b/doc/snippets/compile_snippet.cpp.in index 894cd526c..82ae89162 100644 --- a/doc/snippets/compile_snippet.cpp.in +++ b/doc/snippets/compile_snippet.cpp.in @@ -1,4 +1,4 @@ -#include +#include #include using namespace Eigen; From 3390db099a086c4cc5c9b90cb0ae339c4ba93832 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 17 Oct 2013 00:03:00 +0200 Subject: [PATCH 0146/1103] Fixes bug #681 Also fixed some spelling issues in the documentation --- Eigen/src/CholmodSupport/CholmodSupport.h | 23 +++++++++++++---------- 1 file changed, 13 insertions(+), 10 deletions(-) diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h index 783324b0b..c449960de 100644 --- a/Eigen/src/CholmodSupport/CholmodSupport.h +++ b/Eigen/src/CholmodSupport/CholmodSupport.h @@ -58,10 +58,12 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat) res.p = mat.outerIndexPtr(); res.i = mat.innerIndexPtr(); res.x = mat.valuePtr(); + res.z = 0; res.sorted = 1; if(mat.isCompressed()) { res.packed = 1; + res.nz = 0; } else { @@ -170,6 +172,7 @@ class CholmodBase : internal::noncopyable CholmodBase() : m_cholmodFactor(0), m_info(Success), m_isInitialized(false) { + m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0); cholmod_start(&m_cholmod); } @@ -241,7 +244,7 @@ class CholmodBase : internal::noncopyable return internal::sparse_solve_retval(*this, b.derived()); } - /** Performs a symbolic decomposition on the sparcity of \a matrix. + /** Performs a symbolic decomposition on the sparsity pattern of \a matrix. * * This function is particularly useful when solving for several problems having the same structure. * @@ -265,7 +268,7 @@ class CholmodBase : internal::noncopyable /** Performs a numeric decomposition of \a matrix * - * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed. + * The given matrix must have the same sparsity pattern as the matrix on which the symbolic decomposition has been performed. * * \sa analyzePattern() */ @@ -302,7 +305,7 @@ class CholmodBase : internal::noncopyable { this->m_info = NumericalIssue; } - // TODO optimize this copy by swapping when possible (be carreful with alignment, etc.) + // TODO optimize this copy by swapping when possible (be careful with alignment, etc.) dest = Matrix::Map(reinterpret_cast(x_cd->x),b.rows(),b.cols()); cholmod_free_dense(&x_cd, &m_cholmod); } @@ -323,7 +326,7 @@ class CholmodBase : internal::noncopyable { this->m_info = NumericalIssue; } - // TODO optimize this copy by swapping when possible (be carreful with alignment, etc.) + // TODO optimize this copy by swapping when possible (be careful with alignment, etc.) dest = viewAsEigen(*x_cs); cholmod_free_sparse(&x_cs, &m_cholmod); } @@ -365,8 +368,8 @@ class CholmodBase : internal::noncopyable * * This class allows to solve for A.X = B sparse linear problems via a simplicial LL^T Cholesky factorization * using the Cholmod library. - * This simplicial variant is equivalent to Eigen's built-in SimplicialLLT class. Thefore, it has little practical interest. - * The sparse matrix A must be selfajoint and positive definite. The vectors or matrices + * This simplicial variant is equivalent to Eigen's built-in SimplicialLLT class. Therefore, it has little practical interest. + * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices * X and B can be either dense or sparse. * * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<> @@ -412,8 +415,8 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl * * This class allows to solve for A.X = B sparse linear problems via a simplicial LDL^T Cholesky factorization * using the Cholmod library. - * This simplicial variant is equivalent to Eigen's built-in SimplicialLDLT class. Thefore, it has little practical interest. - * The sparse matrix A must be selfajoint and positive definite. The vectors or matrices + * This simplicial variant is equivalent to Eigen's built-in SimplicialLDLT class. Therefore, it has little practical interest. + * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices * X and B can be either dense or sparse. * * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<> @@ -458,7 +461,7 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp * This class allows to solve for A.X = B sparse linear problems via a supernodal LL^T Cholesky factorization * using the Cholmod library. * This supernodal variant performs best on dense enough problems, e.g., 3D FEM, or very high order 2D FEM. - * The sparse matrix A must be selfajoint and positive definite. The vectors or matrices + * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices * X and B can be either dense or sparse. * * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<> @@ -501,7 +504,7 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper * \brief A general Cholesky factorization and solver based on Cholmod * * This class allows to solve for A.X = B sparse linear problems via a LL^T or LDL^T Cholesky factorization - * using the Cholmod library. The sparse matrix A must be selfajoint and positive definite. The vectors or matrices + * using the Cholmod library. The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices * X and B can be either dense or sparse. * * This variant permits to change the underlying Cholesky method at runtime. From 4d7dfafbe74f8fd05811ca75e095a23c773e2a2b Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 17 Oct 2013 13:49:56 +0200 Subject: [PATCH 0147/1103] Don't add rowSpacer if columns are not to be aligned --- Eigen/src/Core/IO.h | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/IO.h b/Eigen/src/Core/IO.h index c8d5f6379..7e8cb8c0b 100644 --- a/Eigen/src/Core/IO.h +++ b/Eigen/src/Core/IO.h @@ -49,7 +49,7 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& */ struct IOFormat { - /** Default contructor, see class IOFormat for the meaning of the parameters */ + /** Default constructor, see class IOFormat for the meaning of the parameters */ IOFormat(int _precision = StreamPrecision, int _flags = 0, const std::string& _coeffSeparator = " ", const std::string& _rowSeparator = "\n", const std::string& _rowPrefix="", const std::string& _rowSuffix="", @@ -57,6 +57,10 @@ struct IOFormat : matPrefix(_matPrefix), matSuffix(_matSuffix), rowPrefix(_rowPrefix), rowSuffix(_rowSuffix), rowSeparator(_rowSeparator), rowSpacer(""), coeffSeparator(_coeffSeparator), precision(_precision), flags(_flags) { + // TODO check if rowPrefix, rowSuffix or rowSeparator contains a newline + // don't add rowSpacer if columns are not to be aligned + if((fmt.flags & DontAlignCols)) + return; int i = int(matSuffix.length())-1; while (i>=0 && matSuffix[i]!='\n') { From ff075def5cfb7a638cfd523f5313d2a2458f8721 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 17 Oct 2013 14:02:00 +0200 Subject: [PATCH 0148/1103] Copy and paste mistake in last commit --- Eigen/src/Core/IO.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/IO.h b/Eigen/src/Core/IO.h index 7e8cb8c0b..13830861f 100644 --- a/Eigen/src/Core/IO.h +++ b/Eigen/src/Core/IO.h @@ -59,7 +59,7 @@ struct IOFormat { // TODO check if rowPrefix, rowSuffix or rowSeparator contains a newline // don't add rowSpacer if columns are not to be aligned - if((fmt.flags & DontAlignCols)) + if((flags & DontAlignCols)) return; int i = int(matSuffix.length())-1; while (i>=0 && matSuffix[i]!='\n') From ad9dc05663c75b271e2ddf8bd8d039eaa92313ef Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 17 Oct 2013 14:14:06 +0200 Subject: [PATCH 0149/1103] consider all columns for aligned output (fixes bug #616) --- Eigen/src/Core/IO.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/IO.h b/Eigen/src/Core/IO.h index 13830861f..6f41bdc18 100644 --- a/Eigen/src/Core/IO.h +++ b/Eigen/src/Core/IO.h @@ -193,7 +193,7 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& if(align_cols) { // compute the largest width - for(Index j = 1; j < m.cols(); ++j) + for(Index j = 0; j < m.cols(); ++j) for(Index i = 0; i < m.rows(); ++i) { std::stringstream sstr; From 3d2a3bc755f34dd1df330aa2439ef8bacc78fe93 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 17 Oct 2013 14:30:09 +0200 Subject: [PATCH 0150/1103] Copy all format flags (not only precision) from actual output stream when calculating the maximal width --- Eigen/src/Core/IO.h | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/IO.h b/Eigen/src/Core/IO.h index 6f41bdc18..b96802753 100644 --- a/Eigen/src/Core/IO.h +++ b/Eigen/src/Core/IO.h @@ -189,6 +189,9 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& explicit_precision = fmt.precision; } + std::streamsize old_precision = 0; + if(explicit_precision) old_precision = s.precision(explicit_precision); + bool align_cols = !(fmt.flags & DontAlignCols); if(align_cols) { @@ -197,13 +200,11 @@ std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& for(Index i = 0; i < m.rows(); ++i) { std::stringstream sstr; - if(explicit_precision) sstr.precision(explicit_precision); + sstr.copyfmt(s); sstr << m.coeff(i,j); width = std::max(width, Index(sstr.str().length())); } } - std::streamsize old_precision = 0; - if(explicit_precision) old_precision = s.precision(explicit_precision); s << fmt.matPrefix; for(Index i = 0; i < m.rows(); ++i) { From 36052c4911ff4a3034af260b56c7e8c7bdda017b Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 17 Oct 2013 15:37:29 +0200 Subject: [PATCH 0151/1103] Added comparisons scalar to array (previously only the array to scalar was possible) (Fixes bug #147) Extended the unit test for that --- Eigen/src/plugins/ArrayCwiseUnaryOps.h | 5 +++++ test/array.cpp | 14 ++++++++++---- 2 files changed, 15 insertions(+), 4 deletions(-) diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index a59636790..b58d8667c 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -191,6 +191,11 @@ cube() const METHOD_NAME(const Scalar& s) const { \ return CwiseUnaryOp >, const Derived> \ (derived(), std::bind2nd(FUNCTOR(), s)); \ + } \ + friend inline const CwiseUnaryOp >, const Derived> \ + METHOD_NAME(const Scalar& s, const Derived& d) { \ + return CwiseUnaryOp >, const Derived> \ + (d, std::bind1st(FUNCTOR(), s)); \ } EIGEN_MAKE_SCALAR_CWISE_UNARY_OP(operator==, std::equal_to) diff --git a/test/array.cpp b/test/array.cpp index 8960e49f8..f1deda7e3 100644 --- a/test/array.cpp +++ b/test/array.cpp @@ -110,11 +110,17 @@ template void comparisons(const ArrayType& m) VERIFY(! (m1 > m3).all() ); } - // comparisons to scalar + // comparisons array to scalar VERIFY( (m1 != (m1(r,c)+1) ).any() ); - VERIFY( (m1 > (m1(r,c)-1) ).any() ); - VERIFY( (m1 < (m1(r,c)+1) ).any() ); - VERIFY( (m1 == m1(r,c) ).any() ); + VERIFY( (m1 > (m1(r,c)-1) ).any() ); + VERIFY( (m1 < (m1(r,c)+1) ).any() ); + VERIFY( (m1 == m1(r,c) ).any() ); + + // comparisons scalar to array + VERIFY( ( (m1(r,c)+1) != m1).any() ); + VERIFY( ( (m1(r,c)-1) < m1).any() ); + VERIFY( ( (m1(r,c)+1) > m1).any() ); + VERIFY( ( m1(r,c) == m1).any() ); // test Select VERIFY_IS_APPROX( (m1 Date: Fri, 18 Oct 2013 12:56:15 +0200 Subject: [PATCH 0152/1103] simplify/uniformize eigen_gen_docs --- scripts/eigen_gen_docs | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/scripts/eigen_gen_docs b/scripts/eigen_gen_docs index 9b71cd8e0..787dcb325 100644 --- a/scripts/eigen_gen_docs +++ b/scripts/eigen_gen_docs @@ -4,6 +4,7 @@ # You should call this script with USER set as you want, else some default # will be used USER=${USER:-'orzel'} +UPLOAD_DIR=dox-devel #ulimit -v 1024000 @@ -14,10 +15,10 @@ mkdir build -p #step 2 : upload # (the '/' at the end of path is very important, see rsync documentation) -rsync -az --no-p --delete build/doc/html/ $USER@ssh.tuxfamily.org:eigen/eigen.tuxfamily.org-web/htdocs/dox-devel/ || { echo "upload failed"; exit 1; } +rsync -az --no-p --delete build/doc/html/ $USER@ssh.tuxfamily.org:eigen/eigen.tuxfamily.org-web/htdocs/$UPLOAD_DIR/ || { echo "upload failed"; exit 1; } #step 3 : fix the perm -ssh $USER@ssh.tuxfamily.org 'chmod -R g+w /home/eigen/eigen.tuxfamily.org-web/htdocs/dox-devel' || { echo "perm failed"; exit 1; } +ssh $USER@ssh.tuxfamily.org "chmod -R g+w /home/eigen/eigen.tuxfamily.org-web/htdocs/$UPLOAD_DIR" || { echo "perm failed"; exit 1; } echo "Uploaded successfully" From a0e8577b49693df630b9b35b0fee2c89a9ec12ad Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Oct 2013 14:56:36 +0200 Subject: [PATCH 0153/1103] Fix bug #684: optimize vectorization of array-scalar and scalar-array --- Eigen/src/Core/Functors.h | 41 +++++++++++++++++++++++-- Eigen/src/plugins/ArrayCwiseBinaryOps.h | 8 ++--- 2 files changed, 43 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/Functors.h b/Eigen/src/Core/Functors.h index 69d95ea30..d0e05fcf2 100644 --- a/Eigen/src/Core/Functors.h +++ b/Eigen/src/Core/Functors.h @@ -22,6 +22,7 @@ namespace internal { * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, DenseBase::sum() */ template struct scalar_sum_op { +// typedef Scalar result_type; EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op) EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; } template @@ -45,8 +46,8 @@ struct functor_traits > { * \sa DenseBase::count(), DenseBase::any(), ArrayBase::cast(), MatrixBase::cast() */ template<> struct scalar_sum_op : scalar_sum_op { - EIGEN_DEPRECATED - scalar_sum_op() {} + EIGEN_DEPRECATED + scalar_sum_op() {} }; @@ -689,6 +690,42 @@ template struct functor_traits > { enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; +/** \internal + * \brief Template functor to subtract a fixed scalar to another one + * \sa class CwiseUnaryOp, Array::operator-, struct scalar_add_op, struct scalar_rsub_op + */ +template +struct scalar_sub_op { + typedef typename packet_traits::type Packet; + inline scalar_sub_op(const scalar_sub_op& other) : m_other(other.m_other) { } + inline scalar_sub_op(const Scalar& other) : m_other(other) { } + inline Scalar operator() (const Scalar& a) const { return a - m_other; } + inline const Packet packetOp(const Packet& a) const + { return internal::psub(a, pset1(m_other)); } + const Scalar m_other; +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; + +/** \internal + * \brief Template functor to subtract a scalar to fixed another one + * \sa class CwiseUnaryOp, Array::operator-, struct scalar_add_op, struct scalar_sub_op + */ +template +struct scalar_rsub_op { + typedef typename packet_traits::type Packet; + inline scalar_rsub_op(const scalar_rsub_op& other) : m_other(other.m_other) { } + inline scalar_rsub_op(const Scalar& other) : m_other(other) { } + inline Scalar operator() (const Scalar& a) const { return m_other - a; } + inline const Packet packetOp(const Packet& a) const + { return internal::psub(pset1(m_other), a); } + const Scalar m_other; +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; + /** \internal * \brief Template functor to compute the square root of a scalar * \sa class CwiseUnaryOp, Cwise::sqrt() diff --git a/Eigen/src/plugins/ArrayCwiseBinaryOps.h b/Eigen/src/plugins/ArrayCwiseBinaryOps.h index 5c8c476ee..65d198749 100644 --- a/Eigen/src/plugins/ArrayCwiseBinaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseBinaryOps.h @@ -162,16 +162,16 @@ operator+(const Scalar& scalar,const EIGEN_CURRENT_STORAGE_BASE_CLASS& * * \sa operator+(), operator-=() */ -inline const CwiseUnaryOp, const Derived> +inline const CwiseUnaryOp, const Derived> operator-(const Scalar& scalar) const { - return *this + (-scalar); + return CwiseUnaryOp, const Derived>(derived(), internal::scalar_sub_op(scalar));; } -friend inline const CwiseUnaryOp, const CwiseUnaryOp, const Derived> > +friend inline const CwiseUnaryOp, const Derived> operator-(const Scalar& scalar,const EIGEN_CURRENT_STORAGE_BASE_CLASS& other) { - return (-other) + scalar; + return CwiseUnaryOp, const Derived>(other.derived(), internal::scalar_rsub_op(scalar));; } /** \returns an expression of the coefficient-wise && operator of *this and \a other From 9f3f42d66a150d10ca5c0a037eed88f8c9e55336 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 26 Oct 2013 13:59:02 +0200 Subject: [PATCH 0154/1103] fix a few "dead stores" warnings --- Eigen/src/Core/arch/SSE/MathFunctions.h | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index 99cbd0d95..09f74c651 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -126,7 +126,7 @@ Packet4f pexp(const Packet4f& _x) _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); - Packet4f tmp = _mm_setzero_ps(), fx; + Packet4f tmp, fx; Packet4i emm0; // clamp x @@ -138,6 +138,7 @@ Packet4f pexp(const Packet4f& _x) #ifdef EIGEN_VECTORIZE_SSE4_1 fx = _mm_floor_ps(fx); #else + tmp = _mm_setzero_ps(); emm0 = _mm_cvttps_epi32(fx); tmp = _mm_cvtepi32_ps(emm0); /* if greater, substract 1 */ @@ -195,7 +196,7 @@ Packet2d pexp(const Packet2d& _x) _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6); static const __m128i p4i_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0); - Packet2d tmp = _mm_setzero_pd(), fx; + Packet2d tmp, fx; Packet4i emm0; // clamp x @@ -206,6 +207,7 @@ Packet2d pexp(const Packet2d& _x) #ifdef EIGEN_VECTORIZE_SSE4_1 fx = _mm_floor_pd(fx); #else + tmp = _mm_setzero_pd(); emm0 = _mm_cvttpd_epi32(fx); tmp = _mm_cvtepi32_pd(emm0); /* if greater, substract 1 */ @@ -279,7 +281,7 @@ Packet4f psin(const Packet4f& _x) _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI - Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y; + Packet4f xmm1, xmm2, xmm3, sign_bit, y; Packet4i emm0, emm2; sign_bit = x; @@ -378,7 +380,7 @@ Packet4f pcos(const Packet4f& _x) _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827E-002f); _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI - Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, y; + Packet4f xmm1, xmm2, xmm3, y; Packet4i emm0, emm2; x = pabs(x); From 285112fc559ad8c31528c3d34d890f53d016052a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 28 Oct 2013 15:56:30 +0100 Subject: [PATCH 0155/1103] Fix bug #688: make it clearer that CG is for both dense and sparse matrices. --- Eigen/src/IterativeLinearSolvers/ConjugateGradient.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index a74a8155e..3ce517940 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -107,12 +107,12 @@ struct traits > } /** \ingroup IterativeLinearSolvers_Module - * \brief A conjugate gradient solver for sparse self-adjoint problems + * \brief A conjugate gradient solver for sparse (or dense) self-adjoint problems * - * This class allows to solve for A.x = b sparse linear problems using a conjugate gradient algorithm. - * The sparse matrix A must be selfadjoint. The vectors x and b can be either dense or sparse. + * This class allows to solve for A.x = b linear problems using an iterative conjugate gradient algorithm. + * The matrix A must be selfadjoint. The matrix A and the vectors x and b can be either dense or sparse. * - * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix. + * \tparam _MatrixType the type of the matrix A, can be a dense or a sparse matrix. * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower * or Upper. Default is Lower. * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner From 2e606394b15a87301d299eb2d6aeb68e2b4d6b15 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 28 Oct 2013 17:16:03 +0100 Subject: [PATCH 0156/1103] Fix bug #685: document the range of Random and setRandom --- Eigen/src/Core/Random.h | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/Eigen/src/Core/Random.h b/Eigen/src/Core/Random.h index 480fea408..a0bd3039e 100644 --- a/Eigen/src/Core/Random.h +++ b/Eigen/src/Core/Random.h @@ -28,12 +28,16 @@ struct functor_traits > /** \returns a random matrix expression * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * * The parameters \a rows and \a cols are the number of rows and of columns of * the returned matrix. Must be compatible with this MatrixBase type. * * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, * it is redundant to pass \a rows and \a cols as arguments, so Random() should be used * instead. + * * * Example: \include MatrixBase_random_int_int.cpp * Output: \verbinclude MatrixBase_random_int_int.out @@ -52,6 +56,9 @@ DenseBase::Random(Index rows, Index cols) } /** \returns a random vector expression + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. * * The parameter \a size is the size of the returned vector. * Must be compatible with this MatrixBase type. @@ -80,6 +87,9 @@ DenseBase::Random(Index size) /** \returns a fixed-size random matrix or vector expression * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you * need to use the variants taking size arguments. * @@ -101,6 +111,9 @@ DenseBase::Random() /** Sets all coefficients in this expression to random values. * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * * Example: \include MatrixBase_setRandom.cpp * Output: \verbinclude MatrixBase_setRandom.out * @@ -114,6 +127,9 @@ inline Derived& DenseBase::setRandom() /** Resizes to the given \a newSize, and sets all coefficients in this expression to random values. * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. + * * \only_for_vectors * * Example: \include Matrix_setRandom_int.cpp @@ -130,6 +146,9 @@ PlainObjectBase::setRandom(Index newSize) } /** Resizes to the given size, and sets all coefficients in this expression to random values. + * + * Numbers are uniformly spread through their whole definition range for integer types, + * and in the [-1:1] range for floating point scalar types. * * \param nbRows the new number of rows * \param nbCols the new number of columns From d3e84b747add85220119e951fff837ad9b25b8c8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 28 Oct 2013 17:44:07 +0100 Subject: [PATCH 0157/1103] Clarify the meaning of AlignedBit (bug #359) --- Eigen/src/Core/util/Constants.h | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 14b9624e1..05107fdfe 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -143,7 +143,14 @@ const unsigned int DirectAccessBit = 0x40; /** \ingroup flags * - * means the first coefficient packet is guaranteed to be aligned */ + * means the first coefficient packet is guaranteed to be aligned. + * An expression cannot has the AlignedBit without the PacketAccessBit flag. + * In other words, this means we are allow to perform an aligned packet access to the first element regardless + * of the expression kind: + * \code + * expression.packet(0); + * \endcode + */ const unsigned int AlignedBit = 0x80; const unsigned int NestByRefBit = 0x100; From 11fbbc51faa17babefa439423bfa6b69c14fc9e8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 28 Oct 2013 17:48:32 +0100 Subject: [PATCH 0158/1103] Fix bug #359: fix AlignedBit flag of CoeffBasedProduct thus enabling the vectorization of more matrix products --- Eigen/src/Core/products/CoeffBasedProduct.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index c06a0df1c..357e04cb0 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -85,7 +85,8 @@ struct traits > Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit) | (EvalToRowMajor ? RowMajorBit : 0) | NestingFlags - | (LhsFlags & RhsFlags & AlignedBit) + | (CanVectorizeLhs ? (LhsFlags & AlignedBit) : 0) + | (CanVectorizeRhs ? (RhsFlags & AlignedBit) : 0) // TODO enable vectorization for mixed types | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0), @@ -305,7 +306,6 @@ struct product_coeff_impl::run(row, col, lhs, rhs, pres); - product_coeff_impl::run(row, col, lhs, rhs, res); res = predux(pres); } }; From 90b5d303db61b6ea399a1ea43c3830b8ebba55f6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 29 Oct 2013 11:26:52 +0100 Subject: [PATCH 0159/1103] Fix bug #672: use exceptions in SuperLU if they are enabled only --- Eigen/src/SparseLU/SparseLU_Memory.h | 34 ++++++++++++++++------------ 1 file changed, 20 insertions(+), 14 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU_Memory.h b/Eigen/src/SparseLU/SparseLU_Memory.h index a5158025c..767519754 100644 --- a/Eigen/src/SparseLU/SparseLU_Memory.h +++ b/Eigen/src/SparseLU/SparseLU_Memory.h @@ -77,16 +77,23 @@ Index SparseLUImpl::expand(VectorType& vec, Index& length, Index old_vec = vec.segment(0,nbElts); //Allocate or expand the current vector - try +#ifdef EIGEN_EXCEPTIONS + try +#endif { vec.resize(new_len); } +#ifdef EIGEN_EXCEPTIONS catch(std::bad_alloc& ) +#else + if(!vec.size()) +#endif { - if ( !num_expansions ) + if (!num_expansions) { // First time to allocate from LUMemInit() - throw; // Pass the exception to LUMemInit() which has a try... catch block + // Let LUMemInit() deals with it. + return 0; } if (keep_prev) { @@ -101,11 +108,17 @@ Index SparseLUImpl::expand(VectorType& vec, Index& length, Index { alpha = (alpha + 1)/2; new_len = Index(alpha * length); +#ifdef EIGEN_EXCEPTIONS try +#endif { vec.resize(new_len); } +#ifdef EIGEN_EXCEPTIONS catch(std::bad_alloc& ) +#else + if (!vec.size()) +#endif { tries += 1; if ( tries > 10) return new_len; @@ -166,14 +179,10 @@ Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lw // Reserve memory for L/U factors do { - try - { - expand(glu.lusup, glu.nzlumax, 0, 0, num_expansions); - expand(glu.ucol,glu.nzumax, 0, 0, num_expansions); - expand(glu.lsub,glu.nzlmax, 0, 0, num_expansions); - expand(glu.usub,glu.nzumax, 0, 1, num_expansions); - } - catch(std::bad_alloc& ) + if( (!expand(glu.lusup, glu.nzlumax, 0, 0, num_expansions)) + || (!expand(glu.ucol, glu.nzumax, 0, 0, num_expansions)) + || (!expand (glu.lsub, glu.nzlmax, 0, 0, num_expansions)) + || (!expand (glu.usub, glu.nzumax, 0, 1, num_expansions)) ) { //Reduce the estimated size and retry glu.nzlumax /= 2; @@ -181,10 +190,7 @@ Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lw glu.nzlmax /= 2; if (glu.nzlumax < annz ) return glu.nzlumax; } - } while (!glu.lusup.size() || !glu.ucol.size() || !glu.lsub.size() || !glu.usub.size()); - - ++num_expansions; return 0; From 7fae9b358d68b43330e2b9889c413e262a9a04cf Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 29 Oct 2013 12:42:46 +0100 Subject: [PATCH 0160/1103] Use aligned loads in Matrix-Vector product where possible. Fixes bug #689 --- Eigen/src/Core/products/GeneralMatrixVector.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h index 15dc59b98..c8697c913 100644 --- a/Eigen/src/Core/products/GeneralMatrixVector.h +++ b/Eigen/src/Core/products/GeneralMatrixVector.h @@ -293,7 +293,7 @@ EIGEN_DONT_INLINE void general_matrix_vector_product(&lhs0[i]), ptmp0, pload(&res[i]))); + pstore(&res[i], pcj.pmadd(pload(&lhs0[i]), ptmp0, pload(&res[i]))); else for (Index i = alignedStart;i(&lhs0[i]), ptmp0, pload(&res[i]))); From 5974685866b5deb1a84680e9e9689d4e8b7d034e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 29 Oct 2013 17:43:21 +0100 Subject: [PATCH 0161/1103] Fix parenthesis min/max issue in mpreal --- unsupported/test/mpreal/mpreal.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/unsupported/test/mpreal/mpreal.h b/unsupported/test/mpreal/mpreal.h index 75bda7db7..70d37ab71 100644 --- a/unsupported/test/mpreal/mpreal.h +++ b/unsupported/test/mpreal/mpreal.h @@ -1262,9 +1262,9 @@ inline const mpreal mpreal::operator-()const inline const mpreal operator-(const mpreal& a, const mpreal& b) { - mpreal c(0, std::max(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); - mpfr_sub(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); - return c; + mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpfr_sub(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); + return c; } inline const mpreal operator-(const double b, const mpreal& a) @@ -1379,7 +1379,7 @@ inline mpreal& mpreal::operator*=(const int v) inline const mpreal operator*(const mpreal& a, const mpreal& b) { - mpreal c(0, std::max(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); mpfr_mul(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); return c; } @@ -1455,7 +1455,7 @@ inline mpreal& mpreal::operator/=(const int v) inline const mpreal operator/(const mpreal& a, const mpreal& b) { - mpreal c(0, std::max(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); mpfr_div(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); return c; } @@ -2393,7 +2393,7 @@ inline const mpreal const_catalan (mp_prec_t p, mp_rnd_t r) return x; } -inline const mpreal const_infinity (int sign, mp_prec_t p, mp_rnd_t r) +inline const mpreal const_infinity (int sign, mp_prec_t p, mp_rnd_t /*r*/) { mpreal x(0, p); mpfr_set_inf(x.mpfr_ptr(), sign); From 58c0a6f0fd6d86703b121961861d59f422366338 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 29 Oct 2013 17:51:19 +0100 Subject: [PATCH 0162/1103] Fix unused variable warnings --- Eigen/src/SVD/UpperBidiagonalization.h | 2 -- 1 file changed, 2 deletions(-) diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 6cac3af92..40067682c 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -99,7 +99,6 @@ void upperbidiagonalization_inplace_unblocked(MatrixType& mat, Index rows = mat.rows(); Index cols = mat.cols(); - Index size = (std::min)(rows, cols); typedef Matrix TempType; TempType tempVector; @@ -171,7 +170,6 @@ void upperbidiagonalization_blocked_helper(MatrixType& A, for(Index k = 0; k < bs; ++k) { Index remainingRows = brows - k; - Index remainingSizeInBlock = bs - k - 1; Index remainingCols = bcols - k - 1; SubMatType X_k1( X.block(k,0, remainingRows,k) ); From 2702788da71303afa71e53dfa6542a560115425a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 29 Oct 2013 18:02:18 +0100 Subject: [PATCH 0163/1103] Fix bug #678: vectors of row and columns transpositions were not properly resized in FullPivQR --- Eigen/src/QR/FullPivHouseholderQR.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 611467367..8273eaa3b 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -418,8 +418,8 @@ FullPivHouseholderQR& FullPivHouseholderQR::compute(cons m_precision = NumTraits::epsilon() * size; - m_rows_transpositions.resize(matrix.rows()); - m_cols_transpositions.resize(matrix.cols()); + m_rows_transpositions.resize(size); + m_cols_transpositions.resize(size); Index number_of_transpositions = 0; RealScalar biggest(0); From 6dc0e59b1e320660acd9a3d930473170b743454a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 31 Oct 2013 13:52:43 +0100 Subject: [PATCH 0164/1103] Fix bug #677: compilation issue on arm64 which does not have the PLD instruction --- Eigen/src/Core/arch/NEON/PacketMath.h | 20 +++++++++++++++----- 1 file changed, 15 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 163bac215..05e891df2 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -48,9 +48,19 @@ typedef uint32x4_t Packet4ui; #define EIGEN_INIT_NEON_PACKET2(X, Y) {X, Y} #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {X, Y, Z, W} #endif - -#ifndef __pld -#define __pld(x) asm volatile ( " pld [%[addr]]\n" :: [addr] "r" (x) : "cc" ); + + +// arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function +// which available on LLVM and GCC (at least) +#if (defined(__has_builtin) && __has_builtin(__builtin_prefetch)) || defined(__GNUC__) + #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR); +#elif defined __pld + #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR) +#elif !defined(__aarch64__) + #define EIGEN_ARM_PREFETCH(ADDR) asm volatile ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); +#else + // by default no explicit prefetching + #define EIGEN_ARM_PREFETCH(ADDR) #endif template<> struct packet_traits : default_packet_traits @@ -209,8 +219,8 @@ template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& f template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); } -template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { __pld(addr); } -template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { __pld(addr); } +template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { EIGEN_ARM_PREFETCH(addr); } +template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { EIGEN_ARM_PREFETCH(addr); } // FIXME only store the 2 first elements ? template<> EIGEN_STRONG_INLINE float pfirst(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; } From 8f496cd3a3e1ce555da7391e72f5f51f3b4d1ec4 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Nov 2013 18:17:55 +0100 Subject: [PATCH 0165/1103] Fix changeset 2702788da71303afa71e53dfa6542a560115425a for fixed size matrices --- Eigen/src/QR/FullPivHouseholderQR.h | 17 +++++++++-------- 1 file changed, 9 insertions(+), 8 deletions(-) diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 8273eaa3b..a0050859c 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -63,9 +63,10 @@ template class FullPivHouseholderQR typedef typename MatrixType::Index Index; typedef internal::FullPivHouseholderQRMatrixQReturnType MatrixQReturnType; typedef typename internal::plain_diag_type::type HCoeffsType; - typedef Matrix IntRowVectorType; + typedef Matrix IntDiagSizeVectorType; typedef PermutationMatrix PermutationType; - typedef typename internal::plain_col_type::type IntColVectorType; typedef typename internal::plain_row_type::type RowVectorType; typedef typename internal::plain_col_type::type ColVectorType; @@ -173,7 +174,7 @@ template class FullPivHouseholderQR } /** \returns a const reference to the vector of indices representing the rows transpositions */ - const IntColVectorType& rowsTranspositions() const + const IntDiagSizeVectorType& rowsTranspositions() const { eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); return m_rows_transpositions; @@ -369,8 +370,8 @@ template class FullPivHouseholderQR protected: MatrixType m_qr; HCoeffsType m_hCoeffs; - IntColVectorType m_rows_transpositions; - IntRowVectorType m_cols_transpositions; + IntDiagSizeVectorType m_rows_transpositions; + IntDiagSizeVectorType m_cols_transpositions; PermutationType m_cols_permutation; RowVectorType m_temp; bool m_isInitialized, m_usePrescribedThreshold; @@ -538,14 +539,14 @@ template struct FullPivHouseholderQRMatrixQReturnType { public: typedef typename MatrixType::Index Index; - typedef typename internal::plain_col_type::type IntColVectorType; + typedef typename FullPivHouseholderQR::IntDiagSizeVectorType IntDiagSizeVectorType; typedef typename internal::plain_diag_type::type HCoeffsType; typedef Matrix WorkVectorType; FullPivHouseholderQRMatrixQReturnType(const MatrixType& qr, const HCoeffsType& hCoeffs, - const IntColVectorType& rowsTranspositions) + const IntDiagSizeVectorType& rowsTranspositions) : m_qr(qr), m_hCoeffs(hCoeffs), m_rowsTranspositions(rowsTranspositions) @@ -585,7 +586,7 @@ public: protected: typename MatrixType::Nested m_qr; typename HCoeffsType::Nested m_hCoeffs; - typename IntColVectorType::Nested m_rowsTranspositions; + typename IntDiagSizeVectorType::Nested m_rowsTranspositions; }; } // end namespace internal From bbd49d194adc34e1a0e25cdfb5d488f2e0551ce0 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Nov 2013 18:21:46 +0100 Subject: [PATCH 0166/1103] Add a rank method with threshold control to JacobiSVD, and make solve uses it to return the minimal norm solution for rank-deficient problems --- Eigen/src/QR/ColPivHouseholderQR.h | 3 +- Eigen/src/SVD/JacobiSVD.h | 71 +++++++++++++++++++++++++++++- 2 files changed, 71 insertions(+), 3 deletions(-) diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index 8b01f8179..905bc1935 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -76,7 +76,8 @@ template class ColPivHouseholderQR m_colsTranspositions(), m_temp(), m_colSqNorms(), - m_isInitialized(false) {} + m_isInitialized(false), + m_usePrescribedThreshold(false) {} /** \brief Default Constructor with memory preallocation * diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 4786768ff..fcf404587 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -528,6 +528,7 @@ template class JacobiSVD JacobiSVD() : m_isInitialized(false), m_isAllocated(false), + m_usePrescribedThreshold(false), m_computationOptions(0), m_rows(-1), m_cols(-1) {} @@ -542,6 +543,7 @@ template class JacobiSVD JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0) : m_isInitialized(false), m_isAllocated(false), + m_usePrescribedThreshold(false), m_computationOptions(0), m_rows(-1), m_cols(-1) { @@ -561,6 +563,7 @@ template class JacobiSVD JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0) : m_isInitialized(false), m_isAllocated(false), + m_usePrescribedThreshold(false), m_computationOptions(0), m_rows(-1), m_cols(-1) { @@ -662,6 +665,69 @@ template class JacobiSVD eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); return m_nonzeroSingularValues; } + + /** \returns the rank of the matrix of which \c *this is the SVD. + * + * \note This method has to determine which singular values should be considered nonzero. + * For that, it uses the threshold value that you can control by calling + * setThreshold(const RealScalar&). + */ + inline Index rank() const + { + using std::abs; + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + if(m_singularValues.size()==0) return 0; + RealScalar premultiplied_threshold = m_singularValues.coeff(0) * threshold(); + Index i = m_nonzeroSingularValues-1; + while(i>=0 && m_singularValues.coeff(i) < premultiplied_threshold) --i; + return i+1; + } + + /** Allows to prescribe a threshold to be used by certain methods, such as rank() and solve(), + * which need to determine when singular values are to be considered nonzero. + * This is not used for the SVD decomposition itself. + * + * When it needs to get the threshold value, Eigen calls threshold(). + * The default is \c NumTraits::epsilon() + * + * \param threshold The new value to use as the threshold. + * + * A singular value will be considered nonzero if its value is strictly greater than + * \f$ \vert singular value \vert \leqslant threshold \times \vert max singular value \vert \f$. + * + * If you want to come back to the default behavior, call setThreshold(Default_t) + */ + JacobiSVD& setThreshold(const RealScalar& threshold) + { + m_usePrescribedThreshold = true; + m_prescribedThreshold = threshold; + return *this; + } + + /** Allows to come back to the default behavior, letting Eigen use its default formula for + * determining the threshold. + * + * You should pass the special object Eigen::Default as parameter here. + * \code svd.setThreshold(Eigen::Default); \endcode + * + * See the documentation of setThreshold(const RealScalar&). + */ + JacobiSVD& setThreshold(Default_t) + { + m_usePrescribedThreshold = false; + return *this; + } + + /** Returns the threshold that will be used by certain methods such as rank(). + * + * See the documentation of setThreshold(const RealScalar&). + */ + RealScalar threshold() const + { + eigen_assert(m_isInitialized || m_usePrescribedThreshold); + return m_usePrescribedThreshold ? m_prescribedThreshold + : NumTraits::epsilon(); + } inline Index rows() const { return m_rows; } inline Index cols() const { return m_cols; } @@ -674,11 +740,12 @@ template class JacobiSVD MatrixVType m_matrixV; SingularValuesType m_singularValues; WorkMatrixType m_workMatrix; - bool m_isInitialized, m_isAllocated; + bool m_isInitialized, m_isAllocated, m_usePrescribedThreshold; bool m_computeFullU, m_computeThinU; bool m_computeFullV, m_computeThinV; unsigned int m_computationOptions; Index m_nonzeroSingularValues, m_rows, m_cols, m_diagSize; + RealScalar m_prescribedThreshold; template friend struct internal::svd_precondition_2x2_block_to_be_real; @@ -851,7 +918,7 @@ struct solve_retval, Rhs> // So A^{-1} = V S^{-1} U^* Matrix tmp; - Index nonzeroSingVals = dec().nonzeroSingularValues(); + Index nonzeroSingVals = dec().rank(); tmp.noalias() = dec().matrixU().leftCols(nonzeroSingVals).adjoint() * rhs(); tmp = dec().singularValues().head(nonzeroSingVals).asDiagonal().inverse() * tmp; From 19521c83b8d5c9b0567ceadebe7ad4b1119dd7cc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 2 Nov 2013 12:10:48 +0100 Subject: [PATCH 0167/1103] bug #677: fix usage of pld instrinsics for ccomplexes --- Eigen/src/Core/arch/NEON/Complex.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index f183d31de..8d9255eef 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -110,7 +110,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { __pld((float *)addr); } +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((float *)addr); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cf& a) { From 019dcfc21dc74f0c735d72918731b5349a62a26a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 3 Nov 2013 13:18:56 +0100 Subject: [PATCH 0168/1103] JacobiSVD: move from Lapack to Matlab strategy for the default threshold --- Eigen/src/SVD/JacobiSVD.h | 12 ++++----- test/jacobisvd.cpp | 53 +++++++++++++++++++++++++++++++++++++++ 2 files changed, 59 insertions(+), 6 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index fcf404587..98a3ba22a 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -530,7 +530,7 @@ template class JacobiSVD m_isAllocated(false), m_usePrescribedThreshold(false), m_computationOptions(0), - m_rows(-1), m_cols(-1) + m_rows(-1), m_cols(-1), m_diagSize(0) {} @@ -726,7 +726,7 @@ template class JacobiSVD { eigen_assert(m_isInitialized || m_usePrescribedThreshold); return m_usePrescribedThreshold ? m_prescribedThreshold - : NumTraits::epsilon(); + : (std::max)(1,m_diagSize)*NumTraits::epsilon(); } inline Index rows() const { return m_rows; } @@ -918,11 +918,11 @@ struct solve_retval, Rhs> // So A^{-1} = V S^{-1} U^* Matrix tmp; - Index nonzeroSingVals = dec().rank(); + Index rank = dec().rank(); - tmp.noalias() = dec().matrixU().leftCols(nonzeroSingVals).adjoint() * rhs(); - tmp = dec().singularValues().head(nonzeroSingVals).asDiagonal().inverse() * tmp; - dst = dec().matrixV().leftCols(nonzeroSingVals) * tmp; + tmp.noalias() = dec().matrixU().leftCols(rank).adjoint() * rhs(); + tmp = dec().singularValues().head(rank).asDiagonal().inverse() * tmp; + dst = dec().matrixV().leftCols(rank) * tmp; } }; } // end namespace internal diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index 76157c30f..e378de477 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -84,6 +84,59 @@ void jacobisvd_solve(const MatrixType& m, unsigned int computationOptions) SolutionType x = svd.solve(rhs); // evaluate normal equation which works also for least-squares solutions VERIFY_IS_APPROX(m.adjoint()*m*x,m.adjoint()*rhs); + + // check minimal norm solutions + { + // generate a full-rank m x n problem with m MatrixType2; + typedef Matrix RhsType2; + typedef Matrix MatrixType2T; + Index rank = RankAtCompileTime2==Dynamic ? internal::random(1,cols) : Index(RankAtCompileTime2); + MatrixType2 m2(rank,cols); + int guard = 0; + do { + m2.setRandom(); + } while(m2.jacobiSvd().setThreshold(test_precision()).rank()!=rank && (++guard)<10); + VERIFY(guard<10); + RhsType2 rhs2 = RhsType2::Random(rank); + // use QR to find a reference minimal norm solution + HouseholderQR qr(m2.adjoint()); + Matrix tmp = qr.matrixQR().topLeftCorner(rank,rank).template triangularView().adjoint().solve(rhs2); + tmp.conservativeResize(cols); + tmp.tail(cols-rank).setZero(); + SolutionType x21 = qr.householderQ() * tmp; + // now check with SVD + JacobiSVD svd2(m2, computationOptions); + SolutionType x22 = svd2.solve(rhs2); + VERIFY_IS_APPROX(m2*x21, rhs2); + VERIFY_IS_APPROX(m2*x22, rhs2); + VERIFY_IS_APPROX(x21, x22); + + // Now check with a rank deficient matrix + typedef Matrix MatrixType3; + typedef Matrix RhsType3; + Index rows3 = RowsAtCompileTime3==Dynamic ? internal::random(rank+1,2*cols) : Index(RowsAtCompileTime3); + Matrix C = Matrix::Random(rows3,rank); + MatrixType3 m3 = C * m2; + RhsType3 rhs3 = C * rhs2; + JacobiSVD svd3(m3, computationOptions); + SolutionType x3 = svd3.solve(rhs3); + if(svd3.rank()!=rank) { + std::cout << m3 << "\n\n"; + std::cout << svd3.singularValues().transpose() << "\n"; + std::cout << svd3.rank() << " == " << rank << "\n"; + std::cout << x21.norm() << " == " << x3.norm() << "\n"; + } +// VERIFY_IS_APPROX(m3*x3, rhs3); + VERIFY_IS_APPROX(m3*x21, rhs3); + VERIFY_IS_APPROX(m2*x3, rhs2); + + VERIFY_IS_APPROX(x21, x3); + } } template From ad1dc50b5779f4e6f87eb9e5865ce6fcfcc5a828 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 3 Nov 2013 13:19:55 +0100 Subject: [PATCH 0169/1103] Check for minimal norm solutions --- test/jacobisvd.cpp | 8 +------- 1 file changed, 1 insertion(+), 7 deletions(-) diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index e378de477..a0d8562c1 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -125,13 +125,7 @@ void jacobisvd_solve(const MatrixType& m, unsigned int computationOptions) RhsType3 rhs3 = C * rhs2; JacobiSVD svd3(m3, computationOptions); SolutionType x3 = svd3.solve(rhs3); - if(svd3.rank()!=rank) { - std::cout << m3 << "\n\n"; - std::cout << svd3.singularValues().transpose() << "\n"; - std::cout << svd3.rank() << " == " << rank << "\n"; - std::cout << x21.norm() << " == " << x3.norm() << "\n"; - } -// VERIFY_IS_APPROX(m3*x3, rhs3); + VERIFY_IS_APPROX(m3*x3, rhs3); VERIFY_IS_APPROX(m3*x21, rhs3); VERIFY_IS_APPROX(m2*x3, rhs2); From a236e15048014290b02472c18e8fb21d2a6ed99c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 4 Nov 2013 23:58:18 +0100 Subject: [PATCH 0170/1103] JacobiSVD: fix a 0/0 issue for complexes --- Eigen/src/SVD/JacobiSVD.h | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 98a3ba22a..747a4f958 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -380,7 +380,10 @@ struct svd_precondition_2x2_block_to_be_real z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q); work_matrix.row(p) *= z; if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z); - z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q); + if(work_matrix.coeff(q,q)!=Scalar(0)) + z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q); + else + z = Scalar(0); work_matrix.row(q) *= z; if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z); } From 7c9cdd6030a95f0027d5723a758e2290d0b61deb Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 5 Nov 2013 00:12:14 +0100 Subject: [PATCH 0171/1103] SparseLU: fix estimated non-zeros in U --- Eigen/src/SparseLU/SparseLU_Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseLU/SparseLU_Memory.h b/Eigen/src/SparseLU/SparseLU_Memory.h index 767519754..e1e7de043 100644 --- a/Eigen/src/SparseLU/SparseLU_Memory.h +++ b/Eigen/src/SparseLU/SparseLU_Memory.h @@ -153,7 +153,7 @@ Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lw { Index& num_expansions = glu.num_expansions; //No memory expansions so far num_expansions = 0; - glu.nzumax = glu.nzlumax = (std::max)(fillratio * annz, m*n); // estimated number of nonzeros in U + glu.nzumax = glu.nzlumax = (std::min)(fillratio * annz, m*n); // estimated number of nonzeros in U glu.nzlmax = (std::max)(Index(4), fillratio) * annz / 4; // estimated nnz in L factor // Return the estimated size to the user if necessary From 87aee5fda1d42f5e6fdbce3c5c91f28e291147cd Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 5 Nov 2013 15:40:58 +0100 Subject: [PATCH 0172/1103] Allow calling attributes of dynamic size objects from device --- Eigen/src/Core/DenseStorage.h | 88 ++++++++++++++++++----------------- Eigen/src/Core/util/Memory.h | 55 +++++++++++----------- 2 files changed, 73 insertions(+), 70 deletions(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index af14832cf..b9dd75ade 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -24,7 +24,9 @@ namespace internal { struct constructor_without_unaligned_array_assert {}; -template void check_static_allocation_size() +template +EIGEN_DEVICE_FUNC +void check_static_allocation_size() { // if EIGEN_STACK_ALLOCATION_LIMIT is defined to 0, then no limit #if EIGEN_STACK_ALLOCATION_LIMIT @@ -152,18 +154,18 @@ template class DenseSt template class DenseStorage { public: - DenseStorage() {} - DenseStorage(internal::constructor_without_unaligned_array_assert) {} - DenseStorage(const DenseStorage&) {} - DenseStorage& operator=(const DenseStorage&) { return *this; } - DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} - void swap(DenseStorage& ) {} - static DenseIndex rows(void) {return _Rows;} - static DenseIndex cols(void) {return _Cols;} - void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} - void resize(DenseIndex,DenseIndex,DenseIndex) {} - const T *data() const { return 0; } - T *data() { return 0; } + EIGEN_DEVICE_FUNC DenseStorage() {} + EIGEN_DEVICE_FUNC DenseStorage(internal::constructor_without_unaligned_array_assert) {} + EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage&) {} + EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage&) { return *this; } + EIGEN_DEVICE_FUNC DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC void swap(DenseStorage& ) {} + EIGEN_DEVICE_FUNC static DenseIndex rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC static DenseIndex cols(void) {return _Cols;} + EIGEN_DEVICE_FUNC void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC void resize(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC const T *data() const { return 0; } + EIGEN_DEVICE_FUNC T *data() { return 0; } }; // more specializations for null matrices; these are necessary to resolve ambiguities @@ -183,7 +185,7 @@ template class DenseStorage class DenseStorage class DenseStorage m_data; DenseIndex m_rows; public: - DenseStorage() : m_rows(0) {} + EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0) {} DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {} DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {} @@ -229,12 +231,12 @@ template class DenseStorage class DenseStorage m_data; DenseIndex m_cols; public: - DenseStorage() : m_cols(0) {} + EIGEN_DEVICE_FUNC DenseStorage() : m_cols(0) {} DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {} DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {} @@ -258,12 +260,12 @@ template class DenseStorage class DenseStorage class DenseStorage(m_data, m_rows*m_cols); } void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); } - DenseIndex rows(void) const {return m_rows;} - DenseIndex cols(void) const {return m_cols;} + EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return m_rows;} + EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return m_cols;} void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, m_rows*m_cols); @@ -337,8 +339,8 @@ template class DenseStorage class DenseStorage(size)), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } @@ -383,8 +385,8 @@ template class DenseStorage(m_data, _Rows*m_cols); } void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } - static DenseIndex rows(void) {return _Rows;} - DenseIndex cols(void) const {return m_cols;} + EIGEN_DEVICE_FUNC static DenseIndex rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return m_cols;} void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, _Rows*m_cols); @@ -403,8 +405,8 @@ template class DenseStorage class DenseStorage(size)), m_rows(nbRows) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } @@ -449,8 +451,8 @@ template class DenseStorage(m_data, _Cols*m_rows); } void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } - DenseIndex rows(void) const {return m_rows;} - static DenseIndex cols(void) {return _Cols;} + EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return m_rows;} + EIGEN_DEVICE_FUNC static DenseIndex cols(void) {return _Cols;} void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, m_rows*_Cols); @@ -469,8 +471,8 @@ template class DenseStorage struct smart_copy_helper { { std::copy(start, end, target); } }; -// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise. -template struct smart_memmove_helper; - -template void smart_memmove(const T* start, const T* end, T* target) -{ - smart_memmove_helper::RequireInitialization>::run(start, end, target); -} - -template struct smart_memmove_helper { - static inline void run(const T* start, const T* end, T* target) - { std::memmove(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); } -}; - -template struct smart_memmove_helper { - static inline void run(const T* start, const T* end, T* target) - { - if (uintptr_t(target) < uintptr_t(start)) - { - std::copy(start, end, target); - } - else - { - std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T); - std::copy_backward(start, end, target + count); - } - } -}; +// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise. +template struct smart_memmove_helper; + +template void smart_memmove(const T* start, const T* end, T* target) +{ + smart_memmove_helper::RequireInitialization>::run(start, end, target); +} + +template struct smart_memmove_helper { + static inline void run(const T* start, const T* end, T* target) + { std::memmove(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); } +}; + +template struct smart_memmove_helper { + static inline void run(const T* start, const T* end, T* target) + { + if (uintptr_t(target) < uintptr_t(start)) + { + std::copy(start, end, target); + } + else + { + std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T); + std::copy_backward(start, end, target + count); + } + } +}; /***************************************************************************** From 4f572e4c14445158bd9e58c2ba651528847053d6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 5 Nov 2013 15:41:45 +0100 Subject: [PATCH 0173/1103] Add minimalistic unit tests for NVCC support --- Eigen/Core | 9 ++- cmake/EigenTesting.cmake | 25 +++++++-- test/CMakeLists.txt | 22 ++++++++ test/cuda_basic.cu | 116 +++++++++++++++++++++++++++++++++++++++ test/cuda_common.h | 98 +++++++++++++++++++++++++++++++++ test/main.h | 8 ++- 6 files changed, 272 insertions(+), 6 deletions(-) create mode 100644 test/cuda_basic.cu create mode 100644 test/cuda_common.h diff --git a/Eigen/Core b/Eigen/Core index d0f0adbe4..4c9c3d297 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -17,7 +17,14 @@ // Handle NVCC/CUDA #ifdef __CUDACC__ // Do not try asserts on CUDA! + #ifndef EIGEN_NO_DEBUG #define EIGEN_NO_DEBUG + #endif + + #ifdef EIGEN_INTERNAL_DEBUGGING + #undef EIGEN_INTERNAL_DEBUGGING + #endif + // Do not try to vectorize on CUDA! #define EIGEN_DONT_VECTORIZE @@ -190,7 +197,7 @@ #include #endif -#if defined(_CPPUNWIND) || defined(__EXCEPTIONS) +#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) #define EIGEN_EXCEPTIONS #endif diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index 73a62fd73..fdc166bf8 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -11,9 +11,20 @@ endmacro(ei_add_property) #internal. See documentation of ei_add_test for details. macro(ei_add_test_internal testname testname_with_suffix) set(targetname ${testname_with_suffix}) - - set(filename ${testname}.cpp) - add_executable(${targetname} ${filename}) + + if(EIGEN_ADD_TEST_FILENAME_EXTENSION) + set(filename ${testname}.${EIGEN_ADD_TEST_FILENAME_EXTENSION}) + else() + set(filename ${testname}.cpp) + endif() + + if(EIGEN_ADD_TEST_FILENAME_EXTENSION STREQUAL cu) + cuda_add_executable(${targetname} ${filename}) + else() + add_executable(${targetname} ${filename}) + endif() + + if (targetname MATCHES "^eigen2_") add_dependencies(eigen2_buildtests ${targetname}) else() @@ -127,7 +138,13 @@ macro(ei_add_test testname) set(EIGEN_TESTS_LIST "${EIGEN_TESTS_LIST}${testname}\n") set_property(GLOBAL PROPERTY EIGEN_TESTS_LIST "${EIGEN_TESTS_LIST}") - file(READ "${testname}.cpp" test_source) + if(EIGEN_ADD_TEST_FILENAME_EXTENSION) + set(filename ${testname}.${EIGEN_ADD_TEST_FILENAME_EXTENSION}) + else() + set(filename ${testname}.cpp) + endif() + + file(READ "${filename}" test_source) set(parts 0) string(REGEX MATCHALL "CALL_SUBTEST_[0-9]+|EIGEN_TEST_PART_[0-9]+|EIGEN_SUFFIXES(;[0-9]+)+" occurences "${test_source}") diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 547d7505c..5b9e92f01 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -286,3 +286,25 @@ option(EIGEN_TEST_EIGEN2 "Run whole Eigen2 test suite against EIGEN2_SUPPORT" OF if(EIGEN_TEST_EIGEN2) add_subdirectory(eigen2) endif() + + +# NVCC unit tests +option(EIGEN_TEST_NVCC "Enable NVCC support in unit tests" OFF) +if(EIGEN_TEST_NVCC) + +find_package(CUDA) +if(CUDA_FOUND) + + set(CUDA_PROPAGATE_HOST_FLAGS OFF) + set(CUDA_HOST_COMPILER ${CMAKE_CXX_COMPILER}) + cuda_include_directories(${CMAKE_CURRENT_BINARY_DIR}) + set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu") + + ei_add_test(cuda_basic) + + unset(EIGEN_ADD_TEST_FILENAME_EXTENSION) + +endif(CUDA_FOUND) + +endif(EIGEN_TEST_NVCC) + diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu new file mode 100644 index 000000000..a062947a8 --- /dev/null +++ b/test/cuda_basic.cu @@ -0,0 +1,116 @@ + + +#define EIGEN_TEST_NO_LONGDOUBLE +#define EIGEN_TEST_NO_COMPLEX +#define EIGEN_TEST_FUNC cuda_basic +#include "main.h" +#include "cuda_common.h" + +#include + +// struct Foo{ +// EIGEN_DEVICE_FUNC +// void operator()(int i, const float* mats, float* vecs) const { +// using namespace Eigen; +// // Matrix3f M(data); +// // Vector3f x(data+9); +// // Map(data+9) = M.inverse() * x; +// Matrix3f M(mats+i/16); +// Vector3f x(vecs+i*3); +// // using std::min; +// // using std::sqrt; +// Map(vecs+i*3) << x.minCoeff(), 1, 2;// / x.dot(x);//(M.inverse() * x) / x.x(); +// //x = x*2 + x.y() * x + x * x.maxCoeff() - x / x.sum(); +// } +// }; + +template +struct coeff_wise { + EIGEN_DEVICE_FUNC + void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const + { + using namespace Eigen; + T x1(in+i); + T x2(in+i+1); + T x3(in+i+2); + Map res(out+i*T::MaxSizeAtCompileTime); + + res.array() += (in[0] * x1 + x2).array() * x3.array(); + } +}; + +template +struct redux { + EIGEN_DEVICE_FUNC + void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const + { + using namespace Eigen; + int N = 6; + T x1(in+i); + out[i*N+0] = x1.minCoeff(); + out[i*N+1] = x1.maxCoeff(); + out[i*N+2] = x1.sum(); + out[i*N+3] = x1.prod(); +// out[i*N+4] = x1.colwise().sum().maxCoeff(); +// out[i*N+5] = x1.rowwise().maxCoeff().sum(); + } +}; + +template +struct prod { + EIGEN_DEVICE_FUNC + void operator()(int i, const typename T1::Scalar* in, typename T1::Scalar* out) const + { + using namespace Eigen; + typedef Matrix T3; + T1 x1(in+i); + T2 x2(in+i+1); + Map res(out+i*T3::MaxSizeAtCompileTime); + res += in[i] * x1 * x2; + } +}; + + +template +struct eigenvalues { + EIGEN_DEVICE_FUNC + void operator()(int i, const typename T::Scalar* in, typename T::Scalar* out) const + { + using namespace Eigen; + typedef Matrix Vec; + T M(in+i); + Map res(out+i*Vec::MaxSizeAtCompileTime); + T A = M*M.adjoint(); + SelfAdjointEigenSolver eig; + eig.computeDirect(A); + res = A.eigenvalues(); + } +}; + + +void test_cuda_basic() +{ + ei_test_init_cuda(); + + int nthreads = 100; + Eigen::VectorXf in, out; + + #ifndef __CUDA_ARCH__ + int data_size = nthreads * 16; + in.setRandom(data_size); + out.setRandom(data_size); + #endif + + CALL_SUBTEST( run_and_compare_to_cuda(coeff_wise(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(coeff_wise(), nthreads, in, out) ); + + CALL_SUBTEST( run_and_compare_to_cuda(redux(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(redux(), nthreads, in, out) ); + + CALL_SUBTEST( run_and_compare_to_cuda(prod(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(prod(), nthreads, in, out) ); + +// CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); +// CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); + +} diff --git a/test/cuda_common.h b/test/cuda_common.h new file mode 100644 index 000000000..0dd870b9c --- /dev/null +++ b/test/cuda_common.h @@ -0,0 +1,98 @@ + +#ifndef EIGEN_TEST_CUDA_COMMON_H +#define EIGEN_TEST_CUDA_COMMON_H + +#include +#include +#include +#include + +#ifndef __CUDACC__ +dim3 threadIdx, blockDim, blockIdx; +#endif + +template +void run_on_cpu(const Kernel& ker, int n, const Input& in, Output& out) +{ + for(int i=0; i +__global__ +void run_on_cuda_meta_kernel(const Kernel ker, int n, const Input* in, Output* out) +{ + int i = threadIdx.x + blockIdx.x*blockDim.x; + if(i +void run_on_cuda(const Kernel& ker, int n, const Input& in, Output& out) +{ + typename Input::Scalar* d_in; + typename Output::Scalar* d_out; + std::ptrdiff_t in_bytes = in.size() * sizeof(typename Input::Scalar); + std::ptrdiff_t out_bytes = out.size() * sizeof(typename Output::Scalar); + + cudaMalloc((void**)(&d_in), in_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_in, in.data(), in_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_out, out.data(), out_bytes, cudaMemcpyHostToDevice); + + // Simple and non-optimal 1D mapping assuming n is not too large + // That's only for unit testing! + dim3 Blocks(128); + dim3 Grids( (n+int(Blocks.x)-1)/int(Blocks.x) ); + + cudaThreadSynchronize(); + run_on_cuda_meta_kernel<<>>(ker, n, d_in, d_out); + cudaThreadSynchronize(); + + // check inputs have not been modified + cudaMemcpy(const_cast(in.data()), d_in, in_bytes, cudaMemcpyDeviceToHost); + cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost); + + cudaFree(d_in); + cudaFree(d_out); +} + + +template +void run_and_compare_to_cuda(const Kernel& ker, int n, const Input& in, Output& out) +{ + Input in_ref, in_cuda; + Output out_ref, out_cuda; + #ifndef __CUDA_ARCH__ + in_ref = in_cuda = in; + out_ref = out_cuda = out; + #endif + run_on_cpu (ker, n, in_ref, out_ref); + run_on_cuda(ker, n, in_cuda, out_cuda); + #ifndef __CUDA_ARCH__ + VERIFY_IS_APPROX(in_ref, in_cuda); + VERIFY_IS_APPROX(out_ref, out_cuda); + #endif +} + + +void ei_test_init_cuda() +{ + int device = 0; + cudaDeviceProp deviceProp; + cudaGetDeviceProperties(&deviceProp, device); + std::cout << "CUDA device info:\n"; + std::cout << " capability: " << deviceProp.major << "." << deviceProp.minor << "\n"; + std::cout << " multiProcessorCount: " << deviceProp.multiProcessorCount << "\n"; + std::cout << " maxThreadsPerMultiProcessor: " << deviceProp.maxThreadsPerMultiProcessor << "\n"; + std::cout << " regsPerBlock: " << deviceProp.regsPerBlock << "\n"; + std::cout << " concurrentKernels: " << deviceProp.concurrentKernels << "\n"; + std::cout << " clockRate: " << deviceProp.clockRate << "\n"; + std::cout << " computeMode: " << deviceProp.computeMode << "\n"; +} + +#endif // EIGEN_TEST_CUDA_COMMON_H diff --git a/test/main.h b/test/main.h index 14f0d2f78..c889fa6c5 100644 --- a/test/main.h +++ b/test/main.h @@ -138,7 +138,7 @@ namespace Eigen Eigen::internal::push_assert = false; \ } - #else // EIGEN_DEBUG_ASSERTS + #elif !defined(__CUDACC__) // EIGEN_DEBUG_ASSERTS // see bug 89. The copy_bool here is working around a bug in gcc <= 4.3 #define eigen_assert(a) \ if( (!Eigen::internal::copy_bool(a)) && (!no_more_assert) )\ @@ -162,7 +162,9 @@ namespace Eigen #endif // EIGEN_DEBUG_ASSERTS + #if !defined(__CUDACC__) #define EIGEN_USE_CUSTOM_ASSERT + #endif #else // EIGEN_NO_ASSERTION_CHECKING @@ -238,6 +240,7 @@ inline bool test_isMuchSmallerThan(const double& a, const double& b) inline bool test_isApproxOrLessThan(const double& a, const double& b) { return internal::isApproxOrLessThan(a, b, test_precision()); } +#ifndef EIGEN_TEST_NO_COMPLEX inline bool test_isApprox(const std::complex& a, const std::complex& b) { return internal::isApprox(a, b, test_precision >()); } inline bool test_isMuchSmallerThan(const std::complex& a, const std::complex& b) @@ -247,7 +250,9 @@ inline bool test_isApprox(const std::complex& a, const std::complex >()); } inline bool test_isMuchSmallerThan(const std::complex& a, const std::complex& b) { return internal::isMuchSmallerThan(a, b, test_precision >()); } +#endif +#ifndef EIGEN_TEST_NO_LONGDOUBLE inline bool test_isApprox(const long double& a, const long double& b) { bool ret = internal::isApprox(a, b, test_precision()); @@ -261,6 +266,7 @@ inline bool test_isMuchSmallerThan(const long double& a, const long double& b) { return internal::isMuchSmallerThan(a, b, test_precision()); } inline bool test_isApproxOrLessThan(const long double& a, const long double& b) { return internal::isApproxOrLessThan(a, b, test_precision()); } +#endif // EIGEN_TEST_NO_LONGDOUBLE template inline bool test_isApprox(const Type1& a, const Type2& b) From 03de5c24102678c5c7ad78a006fd72a247b276ef Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 6 Nov 2013 10:36:10 +0100 Subject: [PATCH 0174/1103] Split the huge Functors.h file --- Eigen/Core | 6 +- Eigen/src/Core/Functors.h | 1044 --------------------- Eigen/src/Core/functors/BinaryFunctors.h | 448 +++++++++ Eigen/src/Core/functors/NullaryFunctors.h | 158 ++++ Eigen/src/Core/functors/StlFunctors.h | 129 +++ Eigen/src/Core/functors/UnaryFunctors.h | 376 ++++++++ 6 files changed, 1116 insertions(+), 1045 deletions(-) delete mode 100644 Eigen/src/Core/Functors.h create mode 100644 Eigen/src/Core/functors/BinaryFunctors.h create mode 100644 Eigen/src/Core/functors/NullaryFunctors.h create mode 100644 Eigen/src/Core/functors/StlFunctors.h create mode 100644 Eigen/src/Core/functors/UnaryFunctors.h diff --git a/Eigen/Core b/Eigen/Core index 4c9c3d297..bf2d3a908 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -301,7 +301,11 @@ using std::ptrdiff_t; #include "src/Core/arch/Default/Settings.h" -#include "src/Core/Functors.h" +#include "src/Core/functors/BinaryFunctors.h" +#include "src/Core/functors/UnaryFunctors.h" +#include "src/Core/functors/NullaryFunctors.h" +#include "src/Core/functors/StlFunctors.h" + #include "src/Core/DenseCoeffsBase.h" #include "src/Core/DenseBase.h" #include "src/Core/MatrixBase.h" diff --git a/Eigen/src/Core/Functors.h b/Eigen/src/Core/Functors.h deleted file mode 100644 index 3d43b528f..000000000 --- a/Eigen/src/Core/Functors.h +++ /dev/null @@ -1,1044 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008-2010 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_FUNCTORS_H -#define EIGEN_FUNCTORS_H - -namespace Eigen { - -namespace internal { - -// associative functors: - -/** \internal - * \brief Template functor to compute the sum of two scalars - * - * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, DenseBase::sum() - */ -template struct scalar_sum_op { -// typedef Scalar result_type; - EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const - { return internal::padd(a,b); } - template - EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const - { return internal::predux(a); } -}; -template -struct functor_traits > { - enum { - Cost = NumTraits::AddCost, - PacketAccess = packet_traits::HasAdd - }; -}; - -/** \internal - * \brief Template specialization to deprecate the summation of boolean expressions. - * This is required to solve Bug 426. - * \sa DenseBase::count(), DenseBase::any(), ArrayBase::cast(), MatrixBase::cast() - */ -template<> struct scalar_sum_op : scalar_sum_op { - EIGEN_DEPRECATED - scalar_sum_op() {} -}; - - -/** \internal - * \brief Template functor to compute the product of two scalars - * - * \sa class CwiseBinaryOp, Cwise::operator*(), class VectorwiseOp, MatrixBase::redux() - */ -template struct scalar_product_op { - enum { - // TODO vectorize mixed product - Vectorizable = is_same::value && packet_traits::HasMul && packet_traits::HasMul - }; - typedef typename scalar_product_traits::ReturnType result_type; - EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const - { return internal::pmul(a,b); } - template - EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const - { return internal::predux_mul(a); } -}; -template -struct functor_traits > { - enum { - Cost = (NumTraits::MulCost + NumTraits::MulCost)/2, // rough estimate! - PacketAccess = scalar_product_op::Vectorizable - }; -}; - -/** \internal - * \brief Template functor to compute the conjugate product of two scalars - * - * This is a short cut for conj(x) * y which is needed for optimization purpose; in Eigen2 support mode, this becomes x * conj(y) - */ -template struct scalar_conj_product_op { - - enum { - Conj = NumTraits::IsComplex - }; - - typedef typename scalar_product_traits::ReturnType result_type; - - EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const - { return conj_helper().pmul(a,b); } - - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const - { return conj_helper().pmul(a,b); } -}; -template -struct functor_traits > { - enum { - Cost = NumTraits::MulCost, - PacketAccess = internal::is_same::value && packet_traits::HasMul - }; -}; - -/** \internal - * \brief Template functor to compute the min of two scalars - * - * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff() - */ -template struct scalar_min_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(min); return (min)(a, b); } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const - { return internal::pmin(a,b); } - template - EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const - { return internal::predux_min(a); } -}; -template -struct functor_traits > { - enum { - Cost = NumTraits::AddCost, - PacketAccess = packet_traits::HasMin - }; -}; - -/** \internal - * \brief Template functor to compute the max of two scalars - * - * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff() - */ -template struct scalar_max_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(max); return (max)(a, b); } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const - { return internal::pmax(a,b); } - template - EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const - { return internal::predux_max(a); } -}; -template -struct functor_traits > { - enum { - Cost = NumTraits::AddCost, - PacketAccess = packet_traits::HasMax - }; -}; - -/** \internal - * \brief Template functor to compute the hypot of two scalars - * - * \sa MatrixBase::stableNorm(), class Redux - */ -template struct scalar_hypot_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op) -// typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const - { - EIGEN_USING_STD_MATH(max); - EIGEN_USING_STD_MATH(min); - using std::sqrt; - Scalar p = (max)(_x, _y); - Scalar q = (min)(_x, _y); - Scalar qp = q/p; - return p * sqrt(Scalar(1) + qp*qp); - } -}; -template -struct functor_traits > { - enum { Cost = 5 * NumTraits::MulCost, PacketAccess=0 }; -}; - -/** \internal - * \brief Template functor to compute the pow of two scalars - */ -template struct scalar_binary_pow_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_binary_pow_op) - EIGEN_DEVICE_FUNC - inline Scalar operator() (const Scalar& a, const OtherScalar& b) const { return numext::pow(a, b); } -}; -template -struct functor_traits > { - enum { Cost = 5 * NumTraits::MulCost, PacketAccess = false }; -}; - -// other binary functors: - -/** \internal - * \brief Template functor to compute the difference of two scalars - * - * \sa class CwiseBinaryOp, MatrixBase::operator- - */ -template struct scalar_difference_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a - b; } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const - { return internal::psub(a,b); } -}; -template -struct functor_traits > { - enum { - Cost = NumTraits::AddCost, - PacketAccess = packet_traits::HasSub - }; -}; - -/** \internal - * \brief Template functor to compute the quotient of two scalars - * - * \sa class CwiseBinaryOp, Cwise::operator/() - */ -template struct scalar_quotient_op { - enum { - // TODO vectorize mixed product - Vectorizable = is_same::value && packet_traits::HasDiv && packet_traits::HasDiv - }; - typedef typename scalar_product_traits::ReturnType result_type; - EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const - { return internal::pdiv(a,b); } -}; -template -struct functor_traits > { - enum { - Cost = (NumTraits::MulCost + NumTraits::MulCost), // rough estimate! - PacketAccess = scalar_quotient_op::Vectorizable - }; -}; - - - -/** \internal - * \brief Template functor to compute the and of two booleans - * - * \sa class CwiseBinaryOp, ArrayBase::operator&& - */ -struct scalar_boolean_and_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; } -}; -template<> struct functor_traits { - enum { - Cost = NumTraits::AddCost, - PacketAccess = false - }; -}; - -/** \internal - * \brief Template functor to compute the or of two booleans - * - * \sa class CwiseBinaryOp, ArrayBase::operator|| - */ -struct scalar_boolean_or_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; } -}; -template<> struct functor_traits { - enum { - Cost = NumTraits::AddCost, - PacketAccess = false - }; -}; - -// unary functors: - -/** \internal - * \brief Template functor to compute the opposite of a scalar - * - * \sa class CwiseUnaryOp, MatrixBase::operator- - */ -template struct scalar_opposite_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const - { return internal::pnegate(a); } -}; -template -struct functor_traits > -{ enum { - Cost = NumTraits::AddCost, - PacketAccess = packet_traits::HasNegate }; -}; - -/** \internal - * \brief Template functor to compute the absolute value of a scalar - * - * \sa class CwiseUnaryOp, Cwise::abs - */ -template struct scalar_abs_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op) - typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using std::abs; return abs(a); } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const - { return internal::pabs(a); } -}; -template -struct functor_traits > -{ - enum { - Cost = NumTraits::AddCost, - PacketAccess = packet_traits::HasAbs - }; -}; - -/** \internal - * \brief Template functor to compute the squared absolute value of a scalar - * - * \sa class CwiseUnaryOp, Cwise::abs2 - */ -template struct scalar_abs2_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op) - typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const - { return internal::pmul(a,a); } -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasAbs2 }; }; - -/** \internal - * \brief Template functor to compute the conjugate of a complex value - * - * \sa class CwiseUnaryOp, MatrixBase::conjugate() - */ -template struct scalar_conjugate_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op) - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); } - template - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); } -}; -template -struct functor_traits > -{ - enum { - Cost = NumTraits::IsComplex ? NumTraits::AddCost : 0, - PacketAccess = packet_traits::HasConj - }; -}; - -/** \internal - * \brief Template functor to cast a scalar to another type - * - * \sa class CwiseUnaryOp, MatrixBase::cast() - */ -template -struct scalar_cast_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) - typedef NewType result_type; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast(a); } -}; -template -struct functor_traits > -{ enum { Cost = is_same::value ? 0 : NumTraits::AddCost, PacketAccess = false }; }; - -/** \internal - * \brief Template functor to extract the real part of a complex - * - * \sa class CwiseUnaryOp, MatrixBase::real() - */ -template -struct scalar_real_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op) - typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::real(a); } -}; -template -struct functor_traits > -{ enum { Cost = 0, PacketAccess = false }; }; - -/** \internal - * \brief Template functor to extract the imaginary part of a complex - * - * \sa class CwiseUnaryOp, MatrixBase::imag() - */ -template -struct scalar_imag_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op) - typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::imag(a); } -}; -template -struct functor_traits > -{ enum { Cost = 0, PacketAccess = false }; }; - -/** \internal - * \brief Template functor to extract the real part of a complex as a reference - * - * \sa class CwiseUnaryOp, MatrixBase::real() - */ -template -struct scalar_real_ref_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op) - typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::real_ref(*const_cast(&a)); } -}; -template -struct functor_traits > -{ enum { Cost = 0, PacketAccess = false }; }; - -/** \internal - * \brief Template functor to extract the imaginary part of a complex as a reference - * - * \sa class CwiseUnaryOp, MatrixBase::imag() - */ -template -struct scalar_imag_ref_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op) - typedef typename NumTraits::Real result_type; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::imag_ref(*const_cast(&a)); } -}; -template -struct functor_traits > -{ enum { Cost = 0, PacketAccess = false }; }; - -/** \internal - * - * \brief Template functor to compute the exponential of a scalar - * - * \sa class CwiseUnaryOp, Cwise::exp() - */ -template struct scalar_exp_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op) - EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::exp; return exp(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::pexp(a); } -}; -template -struct functor_traits > -{ enum { Cost = 5 * NumTraits::MulCost, PacketAccess = packet_traits::HasExp }; }; - -/** \internal - * - * \brief Template functor to compute the logarithm of a scalar - * - * \sa class CwiseUnaryOp, Cwise::log() - */ -template struct scalar_log_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op) - EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::log; return log(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::plog(a); } -}; -template -struct functor_traits > -{ enum { Cost = 5 * NumTraits::MulCost, PacketAccess = packet_traits::HasLog }; }; - -/** \internal - * \brief Template functor to multiply a scalar by a fixed other one - * - * \sa class CwiseUnaryOp, MatrixBase::operator*, MatrixBase::operator/ - */ -/* NOTE why doing the pset1() in packetOp *is* an optimization ? - * indeed it seems better to declare m_other as a Packet and do the pset1() once - * in the constructor. However, in practice: - * - GCC does not like m_other as a Packet and generate a load every time it needs it - * - on the other hand GCC is able to moves the pset1() outside the loop :) - * - simpler code ;) - * (ICC and gcc 4.4 seems to perform well in both cases, the issue is visible with y = a*x + b*y) - */ -template -struct scalar_multiple_op { - typedef typename packet_traits::type Packet; - // FIXME default copy constructors seems bugged with std::complex<> - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE scalar_multiple_op(const scalar_multiple_op& other) : m_other(other.m_other) { } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE scalar_multiple_op(const Scalar& other) : m_other(other) { } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a * m_other; } - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const - { return internal::pmul(a, pset1(m_other)); } - typename add_const_on_value_type::Nested>::type m_other; -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasMul }; }; - -template -struct scalar_multiple2_op { - typedef typename scalar_product_traits::ReturnType result_type; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple2_op(const scalar_multiple2_op& other) : m_other(other.m_other) { } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple2_op(const Scalar2& other) : m_other(other) { } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar1& a) const { return a * m_other; } - typename add_const_on_value_type::Nested>::type m_other; -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::MulCost, PacketAccess = false }; }; - -/** \internal - * \brief Template functor to divide a scalar by a fixed other one - * - * This functor is used to implement the quotient of a matrix by - * a scalar where the scalar type is not necessarily a floating point type. - * - * \sa class CwiseUnaryOp, MatrixBase::operator/ - */ -template -struct scalar_quotient1_op { - typedef typename packet_traits::type Packet; - // FIXME default copy constructors seems bugged with std::complex<> - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_quotient1_op(const scalar_quotient1_op& other) : m_other(other.m_other) { } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {} - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; } - EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const - { return internal::pdiv(a, pset1(m_other)); } - typename add_const_on_value_type::Nested>::type m_other; -}; -template -struct functor_traits > -{ enum { Cost = 2 * NumTraits::MulCost, PacketAccess = packet_traits::HasDiv }; }; - -// nullary functors - -template -struct scalar_constant_op { - typedef typename packet_traits::type Packet; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { } - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; } - template - EIGEN_STRONG_INLINE const Packet packetOp(Index, Index = 0) const { return internal::pset1(m_other); } - const Scalar m_other; -}; -template -struct functor_traits > -// FIXME replace this packet test by a safe one -{ enum { Cost = 1, PacketAccess = packet_traits::Vectorizable, IsRepeatable = true }; }; - -template struct scalar_identity_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op) - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); } -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::AddCost, PacketAccess = false, IsRepeatable = true }; }; - -template struct linspaced_op_impl; - -// linear access for packet ops: -// 1) initialization -// base = [low, ..., low] + ([step, ..., step] * [-size, ..., 0]) -// 2) each step (where size is 1 for coeff access or PacketSize for packet access) -// base += [size*step, ..., size*step] -// -// TODO: Perhaps it's better to initialize lazily (so not in the constructor but in packetOp) -// in order to avoid the padd() in operator() ? -template -struct linspaced_op_impl -{ - typedef typename packet_traits::type Packet; - - linspaced_op_impl(const Scalar& low, const Scalar& step) : - m_low(low), m_step(step), - m_packetStep(pset1(packet_traits::size*step)), - m_base(padd(pset1(low), pmul(pset1(step),plset(-packet_traits::size)))) {} - - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const - { - m_base = padd(m_base, pset1(m_step)); - return m_low+Scalar(i)*m_step; - } - - template - EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); } - - const Scalar m_low; - const Scalar m_step; - const Packet m_packetStep; - mutable Packet m_base; -}; - -// random access for packet ops: -// 1) each step -// [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) ) -template -struct linspaced_op_impl -{ - typedef typename packet_traits::type Packet; - - linspaced_op_impl(const Scalar& low, const Scalar& step) : - m_low(low), m_step(step), - m_lowPacket(pset1(m_low)), m_stepPacket(pset1(m_step)), m_interPacket(plset(0)) {} - - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; } - - template - EIGEN_STRONG_INLINE const Packet packetOp(Index i) const - { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1(i),m_interPacket))); } - - const Scalar m_low; - const Scalar m_step; - const Packet m_lowPacket; - const Packet m_stepPacket; - const Packet m_interPacket; -}; - -// ----- Linspace functor ---------------------------------------------------------------- - -// Forward declaration (we default to random access which does not really give -// us a speed gain when using packet access but it allows to use the functor in -// nested expressions). -template struct linspaced_op; -template struct functor_traits< linspaced_op > -{ enum { Cost = 1, PacketAccess = packet_traits::HasSetLinear, IsRepeatable = true }; }; -template struct linspaced_op -{ - typedef typename packet_traits::type Packet; - linspaced_op(const Scalar& low, const Scalar& high, DenseIndex num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/(num_steps-1))) {} - - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); } - - // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since - // there row==0 and col is used for the actual iteration. - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const - { - eigen_assert(col==0 || row==0); - return impl(col + row); - } - - template - EIGEN_STRONG_INLINE const Packet packetOp(Index i) const { return impl.packetOp(i); } - - // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since - // there row==0 and col is used for the actual iteration. - template - EIGEN_STRONG_INLINE const Packet packetOp(Index row, Index col) const - { - eigen_assert(col==0 || row==0); - return impl.packetOp(col + row); - } - - // This proxy object handles the actual required temporaries, the different - // implementations (random vs. sequential access) as well as the - // correct piping to size 2/4 packet operations. - const linspaced_op_impl impl; -}; - -// all functors allow linear access, except scalar_identity_op. So we fix here a quick meta -// to indicate whether a functor allows linear access, just always answering 'yes' except for -// scalar_identity_op. -// FIXME move this to functor_traits adding a functor_default -template struct functor_has_linear_access { enum { ret = 1 }; }; -template struct functor_has_linear_access > { enum { ret = 0 }; }; - -// In Eigen, any binary op (Product, CwiseBinaryOp) require the Lhs and Rhs to have the same scalar type, except for multiplication -// where the mixing of different types is handled by scalar_product_traits -// In particular, real * complex is allowed. -// FIXME move this to functor_traits adding a functor_default -template struct functor_is_product_like { enum { ret = 0 }; }; -template struct functor_is_product_like > { enum { ret = 1 }; }; -template struct functor_is_product_like > { enum { ret = 1 }; }; -template struct functor_is_product_like > { enum { ret = 1 }; }; - - -/** \internal - * \brief Template functor to add a scalar to a fixed other one - * \sa class CwiseUnaryOp, Array::operator+ - */ -/* If you wonder why doing the pset1() in packetOp() is an optimization check scalar_multiple_op */ -template -struct scalar_add_op { - typedef typename packet_traits::type Packet; - // FIXME default copy constructors seems bugged with std::complex<> - EIGEN_DEVICE_FUNC inline scalar_add_op(const scalar_add_op& other) : m_other(other.m_other) { } - EIGEN_DEVICE_FUNC inline scalar_add_op(const Scalar& other) : m_other(other) { } - EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a + m_other; } - inline const Packet packetOp(const Packet& a) const - { return internal::padd(a, pset1(m_other)); } - const Scalar m_other; -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; - -/** \internal - * \brief Template functor to subtract a fixed scalar to another one - * \sa class CwiseUnaryOp, Array::operator-, struct scalar_add_op, struct scalar_rsub_op - */ -template -struct scalar_sub_op { - typedef typename packet_traits::type Packet; - inline scalar_sub_op(const scalar_sub_op& other) : m_other(other.m_other) { } - inline scalar_sub_op(const Scalar& other) : m_other(other) { } - inline Scalar operator() (const Scalar& a) const { return a - m_other; } - inline const Packet packetOp(const Packet& a) const - { return internal::psub(a, pset1(m_other)); } - const Scalar m_other; -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; - -/** \internal - * \brief Template functor to subtract a scalar to fixed another one - * \sa class CwiseUnaryOp, Array::operator-, struct scalar_add_op, struct scalar_sub_op - */ -template -struct scalar_rsub_op { - typedef typename packet_traits::type Packet; - inline scalar_rsub_op(const scalar_rsub_op& other) : m_other(other.m_other) { } - inline scalar_rsub_op(const Scalar& other) : m_other(other) { } - inline Scalar operator() (const Scalar& a) const { return m_other - a; } - inline const Packet packetOp(const Packet& a) const - { return internal::psub(pset1(m_other), a); } - const Scalar m_other; -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; - -/** \internal - * \brief Template functor to compute the square root of a scalar - * \sa class CwiseUnaryOp, Cwise::sqrt() - */ -template struct scalar_sqrt_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op) - EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return sqrt(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); } -}; -template -struct functor_traits > -{ enum { - Cost = 5 * NumTraits::MulCost, - PacketAccess = packet_traits::HasSqrt - }; -}; - -/** \internal - * \brief Template functor to compute the cosine of a scalar - * \sa class CwiseUnaryOp, ArrayBase::cos() - */ -template struct scalar_cos_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op) - EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::pcos(a); } -}; -template -struct functor_traits > -{ - enum { - Cost = 5 * NumTraits::MulCost, - PacketAccess = packet_traits::HasCos - }; -}; - -/** \internal - * \brief Template functor to compute the sine of a scalar - * \sa class CwiseUnaryOp, ArrayBase::sin() - */ -template struct scalar_sin_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op) - EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::psin(a); } -}; -template -struct functor_traits > -{ - enum { - Cost = 5 * NumTraits::MulCost, - PacketAccess = packet_traits::HasSin - }; -}; - - -/** \internal - * \brief Template functor to compute the tan of a scalar - * \sa class CwiseUnaryOp, ArrayBase::tan() - */ -template struct scalar_tan_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op) - EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::ptan(a); } -}; -template -struct functor_traits > -{ - enum { - Cost = 5 * NumTraits::MulCost, - PacketAccess = packet_traits::HasTan - }; -}; - -/** \internal - * \brief Template functor to compute the arc cosine of a scalar - * \sa class CwiseUnaryOp, ArrayBase::acos() - */ -template struct scalar_acos_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op) - EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::pacos(a); } -}; -template -struct functor_traits > -{ - enum { - Cost = 5 * NumTraits::MulCost, - PacketAccess = packet_traits::HasACos - }; -}; - -/** \internal - * \brief Template functor to compute the arc sine of a scalar - * \sa class CwiseUnaryOp, ArrayBase::asin() - */ -template struct scalar_asin_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op) - EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); } - typedef typename packet_traits::type Packet; - inline Packet packetOp(const Packet& a) const { return internal::pasin(a); } -}; -template -struct functor_traits > -{ - enum { - Cost = 5 * NumTraits::MulCost, - PacketAccess = packet_traits::HasASin - }; -}; - -/** \internal - * \brief Template functor to raise a scalar to a power - * \sa class CwiseUnaryOp, Cwise::pow - */ -template -struct scalar_pow_op { - // FIXME default copy constructors seems bugged with std::complex<> - inline scalar_pow_op(const scalar_pow_op& other) : m_exponent(other.m_exponent) { } - inline scalar_pow_op(const Scalar& exponent) : m_exponent(exponent) {} - EIGEN_DEVICE_FUNC - inline Scalar operator() (const Scalar& a) const { return numext::pow(a, m_exponent); } - const Scalar m_exponent; -}; -template -struct functor_traits > -{ enum { Cost = 5 * NumTraits::MulCost, PacketAccess = false }; }; - -/** \internal - * \brief Template functor to compute the quotient between a scalar and array entries. - * \sa class CwiseUnaryOp, Cwise::inverse() - */ -template -struct scalar_inverse_mult_op { - scalar_inverse_mult_op(const Scalar& other) : m_other(other) {} - EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return m_other / a; } - template - inline const Packet packetOp(const Packet& a) const - { return internal::pdiv(pset1(m_other),a); } - Scalar m_other; -}; - -/** \internal - * \brief Template functor to compute the inverse of a scalar - * \sa class CwiseUnaryOp, Cwise::inverse() - */ -template -struct scalar_inverse_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op) - EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; } - template - inline const Packet packetOp(const Packet& a) const - { return internal::pdiv(pset1(Scalar(1)),a); } -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasDiv }; }; - -/** \internal - * \brief Template functor to compute the square of a scalar - * \sa class CwiseUnaryOp, Cwise::square() - */ -template -struct scalar_square_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op) - EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a; } - template - inline const Packet packetOp(const Packet& a) const - { return internal::pmul(a,a); } -}; -template -struct functor_traits > -{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasMul }; }; - -/** \internal - * \brief Template functor to compute the cube of a scalar - * \sa class CwiseUnaryOp, Cwise::cube() - */ -template -struct scalar_cube_op { - EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op) - EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a*a; } - template - inline const Packet packetOp(const Packet& a) const - { return internal::pmul(a,pmul(a,a)); } -}; -template -struct functor_traits > -{ enum { Cost = 2*NumTraits::MulCost, PacketAccess = packet_traits::HasMul }; }; - -// default functor traits for STL functors: - -template -struct functor_traits > -{ enum { Cost = NumTraits::MulCost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = NumTraits::MulCost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = NumTraits::AddCost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = NumTraits::AddCost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = NumTraits::AddCost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = functor_traits::Cost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = functor_traits::Cost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1 + functor_traits::Cost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 1 + functor_traits::Cost, PacketAccess = false }; }; - -#ifdef EIGEN_STDEXT_SUPPORT - -template -struct functor_traits > -{ enum { Cost = 0, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = 0, PacketAccess = false }; }; - -template -struct functor_traits > > -{ enum { Cost = 0, PacketAccess = false }; }; - -template -struct functor_traits > > -{ enum { Cost = 0, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = functor_traits::Cost + functor_traits::Cost, PacketAccess = false }; }; - -template -struct functor_traits > -{ enum { Cost = functor_traits::Cost + functor_traits::Cost + functor_traits::Cost, PacketAccess = false }; }; - -#endif // EIGEN_STDEXT_SUPPORT - -// allow to add new functors and specializations of functor_traits from outside Eigen. -// this macro is really needed because functor_traits must be specialized after it is declared but before it is used... -#ifdef EIGEN_FUNCTORS_PLUGIN -#include EIGEN_FUNCTORS_PLUGIN -#endif - -} // end namespace internal - -} // end namespace Eigen - -#endif // EIGEN_FUNCTORS_H diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h new file mode 100644 index 000000000..ba094f5d1 --- /dev/null +++ b/Eigen/src/Core/functors/BinaryFunctors.h @@ -0,0 +1,448 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_BINARY_FUNCTORS_H +#define EIGEN_BINARY_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +//---------- associative binary functors ---------- + +/** \internal + * \brief Template functor to compute the sum of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, DenseBase::sum() + */ +template struct scalar_sum_op { +// typedef Scalar result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a + b; } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::padd(a,b); } + template + EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const + { return internal::predux(a); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::AddCost, + PacketAccess = packet_traits::HasAdd + }; +}; + +/** \internal + * \brief Template specialization to deprecate the summation of boolean expressions. + * This is required to solve Bug 426. + * \sa DenseBase::count(), DenseBase::any(), ArrayBase::cast(), MatrixBase::cast() + */ +template<> struct scalar_sum_op : scalar_sum_op { + EIGEN_DEPRECATED + scalar_sum_op() {} +}; + + +/** \internal + * \brief Template functor to compute the product of two scalars + * + * \sa class CwiseBinaryOp, Cwise::operator*(), class VectorwiseOp, MatrixBase::redux() + */ +template struct scalar_product_op { + enum { + // TODO vectorize mixed product + Vectorizable = is_same::value && packet_traits::HasMul && packet_traits::HasMul + }; + typedef typename scalar_product_traits::ReturnType result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pmul(a,b); } + template + EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const + { return internal::predux_mul(a); } +}; +template +struct functor_traits > { + enum { + Cost = (NumTraits::MulCost + NumTraits::MulCost)/2, // rough estimate! + PacketAccess = scalar_product_op::Vectorizable + }; +}; + +/** \internal + * \brief Template functor to compute the conjugate product of two scalars + * + * This is a short cut for conj(x) * y which is needed for optimization purpose; in Eigen2 support mode, this becomes x * conj(y) + */ +template struct scalar_conj_product_op { + + enum { + Conj = NumTraits::IsComplex + }; + + typedef typename scalar_product_traits::ReturnType result_type; + + EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const + { return conj_helper().pmul(a,b); } + + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return conj_helper().pmul(a,b); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::MulCost, + PacketAccess = internal::is_same::value && packet_traits::HasMul + }; +}; + +/** \internal + * \brief Template functor to compute the min of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff() + */ +template struct scalar_min_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(min); return (min)(a, b); } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pmin(a,b); } + template + EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const + { return internal::predux_min(a); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::AddCost, + PacketAccess = packet_traits::HasMin + }; +}; + +/** \internal + * \brief Template functor to compute the max of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff() + */ +template struct scalar_max_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(max); return (max)(a, b); } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pmax(a,b); } + template + EIGEN_STRONG_INLINE const Scalar predux(const Packet& a) const + { return internal::predux_max(a); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::AddCost, + PacketAccess = packet_traits::HasMax + }; +}; + +/** \internal + * \brief Template functor to compute the hypot of two scalars + * + * \sa MatrixBase::stableNorm(), class Redux + */ +template struct scalar_hypot_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op) +// typedef typename NumTraits::Real result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const + { + EIGEN_USING_STD_MATH(max); + EIGEN_USING_STD_MATH(min); + using std::sqrt; + Scalar p = (max)(_x, _y); + Scalar q = (min)(_x, _y); + Scalar qp = q/p; + return p * sqrt(Scalar(1) + qp*qp); + } +}; +template +struct functor_traits > { + enum { Cost = 5 * NumTraits::MulCost, PacketAccess=0 }; +}; + +/** \internal + * \brief Template functor to compute the pow of two scalars + */ +template struct scalar_binary_pow_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_binary_pow_op) + EIGEN_DEVICE_FUNC + inline Scalar operator() (const Scalar& a, const OtherScalar& b) const { return numext::pow(a, b); } +}; +template +struct functor_traits > { + enum { Cost = 5 * NumTraits::MulCost, PacketAccess = false }; +}; + + + +//---------- non associative binary functors ---------- + +/** \internal + * \brief Template functor to compute the difference of two scalars + * + * \sa class CwiseBinaryOp, MatrixBase::operator- + */ +template struct scalar_difference_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return a - b; } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::psub(a,b); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::AddCost, + PacketAccess = packet_traits::HasSub + }; +}; + +/** \internal + * \brief Template functor to compute the quotient of two scalars + * + * \sa class CwiseBinaryOp, Cwise::operator/() + */ +template struct scalar_quotient_op { + enum { + // TODO vectorize mixed product + Vectorizable = is_same::value && packet_traits::HasDiv && packet_traits::HasDiv + }; + typedef typename scalar_product_traits::ReturnType result_type; + EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const + { return internal::pdiv(a,b); } +}; +template +struct functor_traits > { + enum { + Cost = (NumTraits::MulCost + NumTraits::MulCost), // rough estimate! + PacketAccess = scalar_quotient_op::Vectorizable + }; +}; + + + +/** \internal + * \brief Template functor to compute the and of two booleans + * + * \sa class CwiseBinaryOp, ArrayBase::operator&& + */ +struct scalar_boolean_and_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; } +}; +template<> struct functor_traits { + enum { + Cost = NumTraits::AddCost, + PacketAccess = false + }; +}; + +/** \internal + * \brief Template functor to compute the or of two booleans + * + * \sa class CwiseBinaryOp, ArrayBase::operator|| + */ +struct scalar_boolean_or_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; } +}; +template<> struct functor_traits { + enum { + Cost = NumTraits::AddCost, + PacketAccess = false + }; +}; + + + +//---------- binary functors bound to a constant, thus appearing as a unary functor ---------- + +/** \internal + * \brief Template functor to multiply a scalar by a fixed other one + * + * \sa class CwiseUnaryOp, MatrixBase::operator*, MatrixBase::operator/ + */ +/* NOTE why doing the pset1() in packetOp *is* an optimization ? + * indeed it seems better to declare m_other as a Packet and do the pset1() once + * in the constructor. However, in practice: + * - GCC does not like m_other as a Packet and generate a load every time it needs it + * - on the other hand GCC is able to moves the pset1() outside the loop :) + * - simpler code ;) + * (ICC and gcc 4.4 seems to perform well in both cases, the issue is visible with y = a*x + b*y) + */ +template +struct scalar_multiple_op { + typedef typename packet_traits::type Packet; + // FIXME default copy constructors seems bugged with std::complex<> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE scalar_multiple_op(const scalar_multiple_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE scalar_multiple_op(const Scalar& other) : m_other(other) { } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a * m_other; } + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pmul(a, pset1(m_other)); } + typename add_const_on_value_type::Nested>::type m_other; +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasMul }; }; + +template +struct scalar_multiple2_op { + typedef typename scalar_product_traits::ReturnType result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple2_op(const scalar_multiple2_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_multiple2_op(const Scalar2& other) : m_other(other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar1& a) const { return a * m_other; } + typename add_const_on_value_type::Nested>::type m_other; +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::MulCost, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to divide a scalar by a fixed other one + * + * This functor is used to implement the quotient of a matrix by + * a scalar where the scalar type is not necessarily a floating point type. + * + * \sa class CwiseUnaryOp, MatrixBase::operator/ + */ +template +struct scalar_quotient1_op { + typedef typename packet_traits::type Packet; + // FIXME default copy constructors seems bugged with std::complex<> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_quotient1_op(const scalar_quotient1_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_quotient1_op(const Scalar& other) : m_other(other) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator() (const Scalar& a) const { return a / m_other; } + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pdiv(a, pset1(m_other)); } + typename add_const_on_value_type::Nested>::type m_other; +}; +template +struct functor_traits > +{ enum { Cost = 2 * NumTraits::MulCost, PacketAccess = packet_traits::HasDiv }; }; + +// In Eigen, any binary op (Product, CwiseBinaryOp) require the Lhs and Rhs to have the same scalar type, except for multiplication +// where the mixing of different types is handled by scalar_product_traits +// In particular, real * complex is allowed. +// FIXME move this to functor_traits adding a functor_default +template struct functor_is_product_like { enum { ret = 0 }; }; +template struct functor_is_product_like > { enum { ret = 1 }; }; +template struct functor_is_product_like > { enum { ret = 1 }; }; +template struct functor_is_product_like > { enum { ret = 1 }; }; + + +/** \internal + * \brief Template functor to add a scalar to a fixed other one + * \sa class CwiseUnaryOp, Array::operator+ + */ +/* If you wonder why doing the pset1() in packetOp() is an optimization check scalar_multiple_op */ +template +struct scalar_add_op { + typedef typename packet_traits::type Packet; + // FIXME default copy constructors seems bugged with std::complex<> + EIGEN_DEVICE_FUNC inline scalar_add_op(const scalar_add_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC inline scalar_add_op(const Scalar& other) : m_other(other) { } + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a + m_other; } + inline const Packet packetOp(const Packet& a) const + { return internal::padd(a, pset1(m_other)); } + const Scalar m_other; +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; + +/** \internal + * \brief Template functor to subtract a fixed scalar to another one + * \sa class CwiseUnaryOp, Array::operator-, struct scalar_add_op, struct scalar_rsub_op + */ +template +struct scalar_sub_op { + typedef typename packet_traits::type Packet; + inline scalar_sub_op(const scalar_sub_op& other) : m_other(other.m_other) { } + inline scalar_sub_op(const Scalar& other) : m_other(other) { } + inline Scalar operator() (const Scalar& a) const { return a - m_other; } + inline const Packet packetOp(const Packet& a) const + { return internal::psub(a, pset1(m_other)); } + const Scalar m_other; +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; + +/** \internal + * \brief Template functor to subtract a scalar to fixed another one + * \sa class CwiseUnaryOp, Array::operator-, struct scalar_add_op, struct scalar_sub_op + */ +template +struct scalar_rsub_op { + typedef typename packet_traits::type Packet; + inline scalar_rsub_op(const scalar_rsub_op& other) : m_other(other.m_other) { } + inline scalar_rsub_op(const Scalar& other) : m_other(other) { } + inline Scalar operator() (const Scalar& a) const { return m_other - a; } + inline const Packet packetOp(const Packet& a) const + { return internal::psub(pset1(m_other), a); } + const Scalar m_other; +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = packet_traits::HasAdd }; }; + +/** \internal + * \brief Template functor to raise a scalar to a power + * \sa class CwiseUnaryOp, Cwise::pow + */ +template +struct scalar_pow_op { + // FIXME default copy constructors seems bugged with std::complex<> + inline scalar_pow_op(const scalar_pow_op& other) : m_exponent(other.m_exponent) { } + inline scalar_pow_op(const Scalar& exponent) : m_exponent(exponent) {} + EIGEN_DEVICE_FUNC + inline Scalar operator() (const Scalar& a) const { return numext::pow(a, m_exponent); } + const Scalar m_exponent; +}; +template +struct functor_traits > +{ enum { Cost = 5 * NumTraits::MulCost, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to compute the quotient between a scalar and array entries. + * \sa class CwiseUnaryOp, Cwise::inverse() + */ +template +struct scalar_inverse_mult_op { + scalar_inverse_mult_op(const Scalar& other) : m_other(other) {} + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return m_other / a; } + template + inline const Packet packetOp(const Packet& a) const + { return internal::pdiv(pset1(m_other),a); } + Scalar m_other; +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_BINARY_FUNCTORS_H diff --git a/Eigen/src/Core/functors/NullaryFunctors.h b/Eigen/src/Core/functors/NullaryFunctors.h new file mode 100644 index 000000000..950acd93b --- /dev/null +++ b/Eigen/src/Core/functors/NullaryFunctors.h @@ -0,0 +1,158 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_NULLARY_FUNCTORS_H +#define EIGEN_NULLARY_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +template +struct scalar_constant_op { + typedef typename packet_traits::type Packet; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index, Index = 0) const { return m_other; } + template + EIGEN_STRONG_INLINE const Packet packetOp(Index, Index = 0) const { return internal::pset1(m_other); } + const Scalar m_other; +}; +template +struct functor_traits > +// FIXME replace this packet test by a safe one +{ enum { Cost = 1, PacketAccess = packet_traits::Vectorizable, IsRepeatable = true }; }; + +template struct scalar_identity_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op) + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const { return row==col ? Scalar(1) : Scalar(0); } +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = false, IsRepeatable = true }; }; + +template struct linspaced_op_impl; + +// linear access for packet ops: +// 1) initialization +// base = [low, ..., low] + ([step, ..., step] * [-size, ..., 0]) +// 2) each step (where size is 1 for coeff access or PacketSize for packet access) +// base += [size*step, ..., size*step] +// +// TODO: Perhaps it's better to initialize lazily (so not in the constructor but in packetOp) +// in order to avoid the padd() in operator() ? +template +struct linspaced_op_impl +{ + typedef typename packet_traits::type Packet; + + linspaced_op_impl(const Scalar& low, const Scalar& step) : + m_low(low), m_step(step), + m_packetStep(pset1(packet_traits::size*step)), + m_base(padd(pset1(low), pmul(pset1(step),plset(-packet_traits::size)))) {} + + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const + { + m_base = padd(m_base, pset1(m_step)); + return m_low+Scalar(i)*m_step; + } + + template + EIGEN_STRONG_INLINE const Packet packetOp(Index) const { return m_base = padd(m_base,m_packetStep); } + + const Scalar m_low; + const Scalar m_step; + const Packet m_packetStep; + mutable Packet m_base; +}; + +// random access for packet ops: +// 1) each step +// [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) ) +template +struct linspaced_op_impl +{ + typedef typename packet_traits::type Packet; + + linspaced_op_impl(const Scalar& low, const Scalar& step) : + m_low(low), m_step(step), + m_lowPacket(pset1(m_low)), m_stepPacket(pset1(m_step)), m_interPacket(plset(0)) {} + + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return m_low+i*m_step; } + + template + EIGEN_STRONG_INLINE const Packet packetOp(Index i) const + { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1(i),m_interPacket))); } + + const Scalar m_low; + const Scalar m_step; + const Packet m_lowPacket; + const Packet m_stepPacket; + const Packet m_interPacket; +}; + +// ----- Linspace functor ---------------------------------------------------------------- + +// Forward declaration (we default to random access which does not really give +// us a speed gain when using packet access but it allows to use the functor in +// nested expressions). +template struct linspaced_op; +template struct functor_traits< linspaced_op > +{ enum { Cost = 1, PacketAccess = packet_traits::HasSetLinear, IsRepeatable = true }; }; +template struct linspaced_op +{ + typedef typename packet_traits::type Packet; + linspaced_op(const Scalar& low, const Scalar& high, DenseIndex num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/(num_steps-1))) {} + + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); } + + // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since + // there row==0 and col is used for the actual iteration. + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index row, Index col) const + { + eigen_assert(col==0 || row==0); + return impl(col + row); + } + + template + EIGEN_STRONG_INLINE const Packet packetOp(Index i) const { return impl.packetOp(i); } + + // We need this function when assigning e.g. a RowVectorXd to a MatrixXd since + // there row==0 and col is used for the actual iteration. + template + EIGEN_STRONG_INLINE const Packet packetOp(Index row, Index col) const + { + eigen_assert(col==0 || row==0); + return impl.packetOp(col + row); + } + + // This proxy object handles the actual required temporaries, the different + // implementations (random vs. sequential access) as well as the + // correct piping to size 2/4 packet operations. + const linspaced_op_impl impl; +}; + +// all functors allow linear access, except scalar_identity_op. So we fix here a quick meta +// to indicate whether a functor allows linear access, just always answering 'yes' except for +// scalar_identity_op. +// FIXME move this to functor_traits adding a functor_default +template struct functor_has_linear_access { enum { ret = 1 }; }; +template struct functor_has_linear_access > { enum { ret = 0 }; }; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_NULLARY_FUNCTORS_H diff --git a/Eigen/src/Core/functors/StlFunctors.h b/Eigen/src/Core/functors/StlFunctors.h new file mode 100644 index 000000000..863fd451d --- /dev/null +++ b/Eigen/src/Core/functors/StlFunctors.h @@ -0,0 +1,129 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_STL_FUNCTORS_H +#define EIGEN_STL_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +// default functor traits for STL functors: + +template +struct functor_traits > +{ enum { Cost = NumTraits::MulCost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = NumTraits::MulCost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = NumTraits::AddCost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = functor_traits::Cost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = functor_traits::Cost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1 + functor_traits::Cost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 1 + functor_traits::Cost, PacketAccess = false }; }; + +#ifdef EIGEN_STDEXT_SUPPORT + +template +struct functor_traits > +{ enum { Cost = 0, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = 0, PacketAccess = false }; }; + +template +struct functor_traits > > +{ enum { Cost = 0, PacketAccess = false }; }; + +template +struct functor_traits > > +{ enum { Cost = 0, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = functor_traits::Cost + functor_traits::Cost, PacketAccess = false }; }; + +template +struct functor_traits > +{ enum { Cost = functor_traits::Cost + functor_traits::Cost + functor_traits::Cost, PacketAccess = false }; }; + +#endif // EIGEN_STDEXT_SUPPORT + +// allow to add new functors and specializations of functor_traits from outside Eigen. +// this macro is really needed because functor_traits must be specialized after it is declared but before it is used... +#ifdef EIGEN_FUNCTORS_PLUGIN +#include EIGEN_FUNCTORS_PLUGIN +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_STL_FUNCTORS_H diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h new file mode 100644 index 000000000..a0fcea3f9 --- /dev/null +++ b/Eigen/src/Core/functors/UnaryFunctors.h @@ -0,0 +1,376 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_UNARY_FUNCTORS_H +#define EIGEN_UNARY_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +/** \internal + * \brief Template functor to compute the opposite of a scalar + * + * \sa class CwiseUnaryOp, MatrixBase::operator- + */ +template struct scalar_opposite_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pnegate(a); } +}; +template +struct functor_traits > +{ enum { + Cost = NumTraits::AddCost, + PacketAccess = packet_traits::HasNegate }; +}; + +/** \internal + * \brief Template functor to compute the absolute value of a scalar + * + * \sa class CwiseUnaryOp, Cwise::abs + */ +template struct scalar_abs_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op) + typedef typename NumTraits::Real result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using std::abs; return abs(a); } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pabs(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = NumTraits::AddCost, + PacketAccess = packet_traits::HasAbs + }; +}; + +/** \internal + * \brief Template functor to compute the squared absolute value of a scalar + * + * \sa class CwiseUnaryOp, Cwise::abs2 + */ +template struct scalar_abs2_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op) + typedef typename NumTraits::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const + { return internal::pmul(a,a); } +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasAbs2 }; }; + +/** \internal + * \brief Template functor to compute the conjugate of a complex value + * + * \sa class CwiseUnaryOp, MatrixBase::conjugate() + */ +template struct scalar_conjugate_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op) + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); } + template + EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = NumTraits::IsComplex ? NumTraits::AddCost : 0, + PacketAccess = packet_traits::HasConj + }; +}; + +/** \internal + * \brief Template functor to cast a scalar to another type + * + * \sa class CwiseUnaryOp, MatrixBase::cast() + */ +template +struct scalar_cast_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op) + typedef NewType result_type; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast(a); } +}; +template +struct functor_traits > +{ enum { Cost = is_same::value ? 0 : NumTraits::AddCost, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the real part of a complex + * + * \sa class CwiseUnaryOp, MatrixBase::real() + */ +template +struct scalar_real_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op) + typedef typename NumTraits::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::real(a); } +}; +template +struct functor_traits > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the imaginary part of a complex + * + * \sa class CwiseUnaryOp, MatrixBase::imag() + */ +template +struct scalar_imag_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op) + typedef typename NumTraits::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::imag(a); } +}; +template +struct functor_traits > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the real part of a complex as a reference + * + * \sa class CwiseUnaryOp, MatrixBase::real() + */ +template +struct scalar_real_ref_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op) + typedef typename NumTraits::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::real_ref(*const_cast(&a)); } +}; +template +struct functor_traits > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * \brief Template functor to extract the imaginary part of a complex as a reference + * + * \sa class CwiseUnaryOp, MatrixBase::imag() + */ +template +struct scalar_imag_ref_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op) + typedef typename NumTraits::Real result_type; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::imag_ref(*const_cast(&a)); } +}; +template +struct functor_traits > +{ enum { Cost = 0, PacketAccess = false }; }; + +/** \internal + * + * \brief Template functor to compute the exponential of a scalar + * + * \sa class CwiseUnaryOp, Cwise::exp() + */ +template struct scalar_exp_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::exp; return exp(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::pexp(a); } +}; +template +struct functor_traits > +{ enum { Cost = 5 * NumTraits::MulCost, PacketAccess = packet_traits::HasExp }; }; + +/** \internal + * + * \brief Template functor to compute the logarithm of a scalar + * + * \sa class CwiseUnaryOp, Cwise::log() + */ +template struct scalar_log_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::log; return log(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::plog(a); } +}; +template +struct functor_traits > +{ enum { Cost = 5 * NumTraits::MulCost, PacketAccess = packet_traits::HasLog }; }; + + +/** \internal + * \brief Template functor to compute the square root of a scalar + * \sa class CwiseUnaryOp, Cwise::sqrt() + */ +template struct scalar_sqrt_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sqrt; return sqrt(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); } +}; +template +struct functor_traits > +{ enum { + Cost = 5 * NumTraits::MulCost, + PacketAccess = packet_traits::HasSqrt + }; +}; + +/** \internal + * \brief Template functor to compute the cosine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::cos() + */ +template struct scalar_cos_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::pcos(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = 5 * NumTraits::MulCost, + PacketAccess = packet_traits::HasCos + }; +}; + +/** \internal + * \brief Template functor to compute the sine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::sin() + */ +template struct scalar_sin_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::psin(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = 5 * NumTraits::MulCost, + PacketAccess = packet_traits::HasSin + }; +}; + + +/** \internal + * \brief Template functor to compute the tan of a scalar + * \sa class CwiseUnaryOp, ArrayBase::tan() + */ +template struct scalar_tan_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::ptan(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = 5 * NumTraits::MulCost, + PacketAccess = packet_traits::HasTan + }; +}; + +/** \internal + * \brief Template functor to compute the arc cosine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::acos() + */ +template struct scalar_acos_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::pacos(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = 5 * NumTraits::MulCost, + PacketAccess = packet_traits::HasACos + }; +}; + +/** \internal + * \brief Template functor to compute the arc sine of a scalar + * \sa class CwiseUnaryOp, ArrayBase::asin() + */ +template struct scalar_asin_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::pasin(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = 5 * NumTraits::MulCost, + PacketAccess = packet_traits::HasASin + }; +}; + +/** \internal + * \brief Template functor to compute the inverse of a scalar + * \sa class CwiseUnaryOp, Cwise::inverse() + */ +template +struct scalar_inverse_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; } + template + inline const Packet packetOp(const Packet& a) const + { return internal::pdiv(pset1(Scalar(1)),a); } +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasDiv }; }; + +/** \internal + * \brief Template functor to compute the square of a scalar + * \sa class CwiseUnaryOp, Cwise::square() + */ +template +struct scalar_square_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a; } + template + inline const Packet packetOp(const Packet& a) const + { return internal::pmul(a,a); } +}; +template +struct functor_traits > +{ enum { Cost = NumTraits::MulCost, PacketAccess = packet_traits::HasMul }; }; + +/** \internal + * \brief Template functor to compute the cube of a scalar + * \sa class CwiseUnaryOp, Cwise::cube() + */ +template +struct scalar_cube_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op) + EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a*a; } + template + inline const Packet packetOp(const Packet& a) const + { return internal::pmul(a,pmul(a,a)); } +}; +template +struct functor_traits > +{ enum { Cost = 2*NumTraits::MulCost, PacketAccess = packet_traits::HasMul }; }; + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_FUNCTORS_H From a37bdfc9552374edb022cefa824b8dac56f9b6d6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 6 Nov 2013 11:13:31 +0100 Subject: [PATCH 0175/1103] Fix static/inline order --- Eigen/src/Core/AssignEvaluator.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 8d835b2f6..5f2fb9d38 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -438,7 +438,7 @@ struct copy_using_evaluator_impl struct copy_using_evaluator_impl { - inline static void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -501,7 +501,7 @@ struct copy_using_evaluator_impl struct copy_using_evaluator_impl { - inline static void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -540,7 +540,7 @@ struct copy_using_evaluator_impl struct copy_using_evaluator_impl { - inline static void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -592,7 +592,7 @@ struct copy_using_evaluator_impl struct copy_using_evaluator_impl { - inline static void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; From 8edc9647345445c2b3e8b9ff1c70c625f43755c6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 6 Nov 2013 18:17:59 +0100 Subject: [PATCH 0176/1103] bug #99: refactor assignment and compound assignment mechanism through "assignment functors" and "assignement kernels". The former is very low level and generic. The later abstarct the former for dense expressions. This refactoring permits to get rid of the very ugly SwapWrapper and SelfCwiseBinaryOp classes. In the future, this will also permit to simplify all these evaluation loops and perhaps to reuse them for reduxions. That will also permit to specialize for operations like expr1 += expr2 outside Eigen, and so for any kind of expressions (dense, sparse, tensor, etc.) --- Eigen/Core | 1 + Eigen/src/Core/AssignEvaluator.h | 463 ++++++++++++------- Eigen/src/Core/CoreEvaluators.h | 12 +- Eigen/src/Core/functors/AssignmentFunctors.h | 167 +++++++ test/evaluator_common.h | 0 5 files changed, 478 insertions(+), 165 deletions(-) create mode 100644 Eigen/src/Core/functors/AssignmentFunctors.h create mode 100644 test/evaluator_common.h diff --git a/Eigen/Core b/Eigen/Core index bf2d3a908..722a49030 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -383,6 +383,7 @@ using std::ptrdiff_t; #include "src/Core/ArrayWrapper.h" #ifdef EIGEN_ENABLE_EVALUATORS +#include "src/Core/functors/AssignmentFunctors.h" #include "src/Core/Product.h" #include "src/Core/CoreEvaluators.h" #include "src/Core/AssignEvaluator.h" diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 5f2fb9d38..adea7ef13 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2011 Benoit Jacob -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2013 Gael Guennebaud // Copyright (C) 2011-2012 Jitse Niesen // // This Source Code Form is subject to the terms of the Mozilla @@ -94,7 +94,7 @@ public: enum { Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal)) ? ( - int(MayUnrollCompletely) ? int(CompleteUnrolling) + int(MayUnrollCompletely) ? int(CompleteUnrolling) : int(MayUnrollInner) ? int(InnerUnrolling) : int(NoUnrolling) ) @@ -139,7 +139,7 @@ public: *** Default traversal *** ************************/ -template +template struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling { typedef typename DstEvaluatorType::XprType DstXprType; @@ -150,69 +150,74 @@ struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling }; static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator) + SrcEvaluatorType &srcEvaluator, + const Kernel &kernel + ) { - dstEvaluator.copyCoeffByOuterInner(outer, inner, srcEvaluator); + kernel.assignCoeffByOuterInner(outer, inner, dstEvaluator, srcEvaluator); copy_using_evaluator_DefaultTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); } }; -template -struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling +template +struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&) { } + static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&) { } }; -template +template struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator, + SrcEvaluatorType &srcEvaluator, + const Kernel &kernel, int outer) { - dstEvaluator.copyCoeffByOuterInner(outer, Index, srcEvaluator); + kernel.assignCoeffByOuterInner(outer, Index, dstEvaluator, srcEvaluator); copy_using_evaluator_DefaultTraversal_InnerUnrolling - - ::run(dstEvaluator, srcEvaluator, outer); + + ::run(dstEvaluator, srcEvaluator, kernel, outer); } }; -template -struct copy_using_evaluator_DefaultTraversal_InnerUnrolling +template +struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, int) { } + static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&, int) { } }; /*********************** *** Linear traversal *** ***********************/ -template +template struct copy_using_evaluator_LinearTraversal_CompleteUnrolling { static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator) + SrcEvaluatorType &srcEvaluator, + const Kernel& kernel + ) { - dstEvaluator.copyCoeff(Index, srcEvaluator); + kernel.assignCoeff(Index, dstEvaluator, srcEvaluator); copy_using_evaluator_LinearTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); } }; -template -struct copy_using_evaluator_LinearTraversal_CompleteUnrolling +template +struct copy_using_evaluator_LinearTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&) { } + static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&) { } }; /************************** *** Inner vectorization *** **************************/ -template +template struct copy_using_evaluator_innervec_CompleteUnrolling { typedef typename DstEvaluatorType::XprType DstXprType; @@ -225,63 +230,66 @@ struct copy_using_evaluator_innervec_CompleteUnrolling }; static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator) + SrcEvaluatorType &srcEvaluator, + const Kernel &kernel + ) { - dstEvaluator.template copyPacketByOuterInner(outer, inner, srcEvaluator); + kernel.template assignPacketByOuterInner(outer, inner, dstEvaluator, srcEvaluator); enum { NextIndex = Index + packet_traits::size }; copy_using_evaluator_innervec_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); } }; -template -struct copy_using_evaluator_innervec_CompleteUnrolling +template +struct copy_using_evaluator_innervec_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&) { } + static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&) { } }; -template +template struct copy_using_evaluator_innervec_InnerUnrolling { static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator, + SrcEvaluatorType &srcEvaluator, + const Kernel &kernel, int outer) { - dstEvaluator.template copyPacketByOuterInner(outer, Index, srcEvaluator); + kernel.template assignPacketByOuterInner(outer, Index, dstEvaluator, srcEvaluator); typedef typename DstEvaluatorType::XprType DstXprType; enum { NextIndex = Index + packet_traits::size }; copy_using_evaluator_innervec_InnerUnrolling - - ::run(dstEvaluator, srcEvaluator, outer); + + ::run(dstEvaluator, srcEvaluator, kernel, outer); } }; -template -struct copy_using_evaluator_innervec_InnerUnrolling +template +struct copy_using_evaluator_innervec_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, int) { } + static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel &, int) { } }; /*************************************************************************** * Part 3 : implementation of all cases ***************************************************************************/ -// copy_using_evaluator_impl is based on assign_impl +// dense_assignment_loop is based on assign_impl -template::Traversal, int Unrolling = copy_using_evaluator_traits::Unrolling> -struct copy_using_evaluator_impl; +struct dense_assignment_loop; /************************ *** Default traversal *** ************************/ -template -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static void run(DstXprType& dst, const SrcXprType& src) + static void run(DstXprType& dst, const SrcXprType& src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -292,16 +300,16 @@ struct copy_using_evaluator_impl -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src) + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -310,16 +318,16 @@ struct copy_using_evaluator_impl - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); } }; -template -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { typedef typename DstXprType::Index Index; - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src) + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -330,8 +338,8 @@ struct copy_using_evaluator_impl - ::run(dstEvaluator, srcEvaluator, outer); + + ::run(dstEvaluator, srcEvaluator, kernel, outer); } }; @@ -339,43 +347,50 @@ struct copy_using_evaluator_impl -struct unaligned_copy_using_evaluator_impl +struct unaligned_dense_assignment_loop { // if IsAligned = true, then do nothing - template - static EIGEN_STRONG_INLINE void run(const SrcEvaluatorType&, DstEvaluatorType&, + template + static EIGEN_STRONG_INLINE void run(const SrcEvaluatorType&, DstEvaluatorType&, const Kernel&, typename SrcEvaluatorType::Index, typename SrcEvaluatorType::Index) {} }; template <> -struct unaligned_copy_using_evaluator_impl +struct unaligned_dense_assignment_loop { // MSVC must not inline this functions. If it does, it fails to optimize the // packet access path. + // FIXME check which version exhibits this issue #ifdef _MSC_VER - template + template static EIGEN_DONT_INLINE void run(DstEvaluatorType &dstEvaluator, const SrcEvaluatorType &srcEvaluator, + const Kernel &kernel, typename DstEvaluatorType::Index start, typename DstEvaluatorType::Index end) #else - template + template static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, const SrcEvaluatorType &srcEvaluator, + const Kernel &kernel, typename DstEvaluatorType::Index start, typename DstEvaluatorType::Index end) #endif { for (typename DstEvaluatorType::Index index = start; index < end; ++index) - dstEvaluator.copyCoeff(index, srcEvaluator); + kernel.assignCoeff(index, dstEvaluator, srcEvaluator); } }; -template -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src) + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -395,22 +410,20 @@ struct copy_using_evaluator_impl::run(dstEvaluator, srcEvaluator, 0, alignedStart); + unaligned_dense_assignment_loop::run(dstEvaluator, srcEvaluator, kernel, 0, alignedStart); for(Index index = alignedStart; index < alignedEnd; index += packetSize) - { - dstEvaluator.template copyPacket(index, srcEvaluator); - } + kernel.template assignPacket(index, dstEvaluator, srcEvaluator); - unaligned_copy_using_evaluator_impl<>::run(dstEvaluator, srcEvaluator, alignedEnd, size); + unaligned_dense_assignment_loop<>::run(dstEvaluator, srcEvaluator, kernel, alignedEnd, size); } }; -template -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { typedef typename DstXprType::Index Index; - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src) + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -423,11 +436,11 @@ struct copy_using_evaluator_impl - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); copy_using_evaluator_DefaultTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); } }; @@ -435,10 +448,10 @@ struct copy_using_evaluator_impl -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -451,16 +464,15 @@ struct copy_using_evaluator_impl::size; for(Index outer = 0; outer < outerSize; ++outer) - for(Index inner = 0; inner < innerSize; inner+=packetSize) { - dstEvaluator.template copyPacketByOuterInner(outer, inner, srcEvaluator); - } + for(Index inner = 0; inner < innerSize; inner+=packetSize) + kernel.template assignPacketByOuterInner(outer, inner, dstEvaluator, srcEvaluator); } }; -template -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src) + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -469,16 +481,16 @@ struct copy_using_evaluator_impl - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); } }; -template -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { typedef typename DstXprType::Index Index; - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src) + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -489,8 +501,8 @@ struct copy_using_evaluator_impl - ::run(dstEvaluator, srcEvaluator, outer); + + ::run(dstEvaluator, srcEvaluator, kernel, outer); } }; @@ -498,10 +510,10 @@ struct copy_using_evaluator_impl -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -512,14 +524,14 @@ struct copy_using_evaluator_impl -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src) + static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -528,8 +540,8 @@ struct copy_using_evaluator_impl - ::run(dstEvaluator, srcEvaluator); + + ::run(dstEvaluator, srcEvaluator, kernel); } }; @@ -537,10 +549,10 @@ struct copy_using_evaluator_impl -struct copy_using_evaluator_impl +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -566,19 +578,16 @@ struct copy_using_evaluator_impl(outer, inner, srcEvaluator); - } + for(Index inner = alignedStart; inner(outer, inner, dstEvaluator, srcEvaluator); // do the non-vectorizable part of the assignment - for(Index inner = alignedEnd; inner((alignedStart+alignedStep)%packetSize, innerSize); } @@ -589,10 +598,12 @@ struct copy_using_evaluator_impl -struct copy_using_evaluator_impl +// TODO: this 'AllAtOnceTraversal' should be dropped or caught earlier (Gael) +// Indeed, what to do with the kernel?? +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src) + static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &/*kernel*/) { typedef typename evaluator::type DstEvaluatorType; typedef typename evaluator::type SrcEvaluatorType; @@ -601,23 +612,126 @@ struct copy_using_evaluator_impl +struct generic_dense_assignment_kernel +{ + const Functor &m_functor; + generic_dense_assignment_kernel(const Functor &func) : m_functor(func) {} + + template + void assignCoeff(typename DstEvaluatorType::Index row, typename DstEvaluatorType::Index col, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + m_functor.assignCoeff(dst.coeffRef(row,col), src.coeff(row,col)); + } + + template + void assignCoeff(typename DstEvaluatorType::Index index, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + m_functor.assignCoeff(dst.coeffRef(index), src.coeff(index)); + } + + template + void assignCoeffByOuterInner(typename DstEvaluatorType::Index outer, typename DstEvaluatorType::Index inner, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + typedef typename DstEvaluatorType::Index Index; + Index row = rowIndexByOuterInner(outer, inner); + Index col = colIndexByOuterInner(outer, inner); + assignCoeff(row, col, dst, src); + } + + + template + void assignPacket(typename DstEvaluatorType::Index row, typename DstEvaluatorType::Index col, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + m_functor.assignPacket(&dst.coeffRef(row,col), src.template packet(row,col)); + } + + template + void assignPacket(typename DstEvaluatorType::Index index, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + m_functor.assignPacket(&dst.coeffRef(index), src.template packet(index)); + } + + template + void assignPacketByOuterInner(typename DstEvaluatorType::Index outer, typename DstEvaluatorType::Index inner, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + typedef typename DstEvaluatorType::Index Index; + Index row = rowIndexByOuterInner(outer, inner); + Index col = colIndexByOuterInner(outer, inner); + assignPacket(row, col, dst, src); + } + + template + static Index rowIndexByOuterInner(Index outer, Index inner) + { + typedef typename EvaluatorType::ExpressionTraits Traits; + return int(Traits::RowsAtCompileTime) == 1 ? 0 + : int(Traits::ColsAtCompileTime) == 1 ? inner + : int(Traits::Flags)&RowMajorBit ? outer + : inner; + } + + template + static Index colIndexByOuterInner(Index outer, Index inner) + { + typedef typename EvaluatorType::ExpressionTraits Traits; + return int(Traits::ColsAtCompileTime) == 1 ? 0 + : int(Traits::RowsAtCompileTime) == 1 ? inner + : int(Traits::Flags)&RowMajorBit ? inner + : outer; + } +}; + +template +void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) +{ +#ifdef EIGEN_DEBUG_ASSIGN + internal::copy_using_evaluator_traits::debug(); +#endif + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + typedef generic_dense_assignment_kernel Kernel; + Kernel kernel(func); + + dense_assignment_loop::run(const_cast(dst), src, kernel); +} + +template +void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src) +{ + call_dense_assignment_loop(dst, src, internal::assign_op()); +} + +/*************************************************************************** +* Part 5 : Entry points ***************************************************************************/ // Based on DenseBase::LazyAssign() +// The following functions are just for testing and they are meant to be moved to operator= and the likes. template class StorageBase, typename SrcXprType> EIGEN_STRONG_INLINE const DstXprType& copy_using_evaluator(const NoAlias& dst, const EigenBase& src) { - return noalias_copy_using_evaluator(dst.expression(), src.derived()); + return noalias_copy_using_evaluator(dst.expression(), src.derived(), internal::assign_op()); } template::AssumeAliasing> @@ -641,57 +755,90 @@ struct AddEvalIfAssumingAliasing } }; +template +EIGEN_STRONG_INLINE +const DstXprType& copy_using_evaluator(const EigenBase& dst, const EigenBase& src, const Functor &func) +{ + return noalias_copy_using_evaluator(dst.const_cast_derived(), + AddEvalIfAssumingAliasing::run(src.derived()), + func + ); +} + +// this mimics operator= template EIGEN_STRONG_INLINE const DstXprType& copy_using_evaluator(const EigenBase& dst, const EigenBase& src) { - return noalias_copy_using_evaluator(dst.const_cast_derived(), - AddEvalIfAssumingAliasing::run(src.derived())); + return copy_using_evaluator(dst.const_cast_derived(), src.derived(), internal::assign_op()); } -template +template EIGEN_STRONG_INLINE -const DstXprType& noalias_copy_using_evaluator(const PlainObjectBase& dst, const EigenBase& src) +const DstXprType& noalias_copy_using_evaluator(const PlainObjectBase& dst, const EigenBase& src, const Functor &func) { #ifdef EIGEN_DEBUG_ASSIGN internal::copy_using_evaluator_traits::debug(); #endif #ifdef EIGEN_NO_AUTOMATIC_RESIZING eigen_assert((dst.size()==0 || (IsVectorAtCompileTime ? (dst.size() == src.size()) - : (dst.rows() == src.rows() && dst.cols() == src.cols()))) - && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined"); + : (dst.rows() == src.rows() && dst.cols() == src.cols()))) + && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined"); #else dst.const_cast_derived().resizeLike(src.derived()); #endif - return copy_using_evaluator_without_resizing(dst.const_cast_derived(), src.derived()); + call_dense_assignment_loop(dst.const_cast_derived(), src.derived(), func); + return dst.derived(); } -template +template EIGEN_STRONG_INLINE -const DstXprType& noalias_copy_using_evaluator(const EigenBase& dst, const EigenBase& src) +const DstXprType& noalias_copy_using_evaluator(const EigenBase& dst, const EigenBase& src, const Functor &func) { - return copy_using_evaluator_without_resizing(dst.const_cast_derived(), src.derived()); -} - -template -const DstXprType& copy_using_evaluator_without_resizing(const DstXprType& dst, const SrcXprType& src) -{ -#ifdef EIGEN_DEBUG_ASSIGN - internal::copy_using_evaluator_traits::debug(); -#endif - eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); - copy_using_evaluator_impl::run(const_cast(dst), src); - return dst; + call_dense_assignment_loop(dst.const_cast_derived(), src.derived(), func); + return dst.derived(); } // Based on DenseBase::swap() -// TODO: Chech whether we need to do something special for swapping two -// Arrays or Matrices. +// TODO: Check whether we need to do something special for swapping two +// Arrays or Matrices. (Jitse) +// Overload default assignPacket behavior for swapping them +template +struct swap_kernel : generic_dense_assignment_kernel > +{ + typedef generic_dense_assignment_kernel > Base; + using Base::m_functor; + swap_kernel() : Base(swap_assign_op()) {} + + template + void assignPacket(typename DstEvaluatorType::Index row, typename DstEvaluatorType::Index col, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + m_functor.template swapPacket(&dst.coeffRef(row,col), &const_cast(src).coeffRef(row,col)); + } + + template + void assignPacket(typename DstEvaluatorType::Index index, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + m_functor.template swapPacket(&dst.coeffRef(index), &const_cast(src).coeffRef(index)); + } + + // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael) + template + void assignPacketByOuterInner(typename DstEvaluatorType::Index outer, typename DstEvaluatorType::Index inner, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + { + typedef typename DstEvaluatorType::Index Index; + Index row = Base::template rowIndexByOuterInner(outer, inner); + Index col = Base::template colIndexByOuterInner(outer, inner); + assignPacket(row, col, dst, src); + } +}; + template void swap_using_evaluator(const DstXprType& dst, const SrcXprType& src) { - copy_using_evaluator(SwapWrapper(const_cast(dst)), src); + typedef swap_kernel kernel; + dense_assignment_loop::run(const_cast(dst), src, kernel()); } // Based on MatrixBase::operator+= (in CwiseBinaryOp.h) @@ -699,8 +846,7 @@ template void add_assign_using_evaluator(const MatrixBase& dst, const MatrixBase& src) { typedef typename DstXprType::Scalar Scalar; - SelfCwiseBinaryOp, DstXprType, SrcXprType> tmp(dst.const_cast_derived()); - copy_using_evaluator(tmp, src.derived()); + copy_using_evaluator(dst.derived(), src.derived(), add_assign_op()); } // Based on ArrayBase::operator+= @@ -708,42 +854,37 @@ template void add_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) { typedef typename DstXprType::Scalar Scalar; - SelfCwiseBinaryOp, DstXprType, SrcXprType> tmp(dst.const_cast_derived()); - copy_using_evaluator(tmp, src.derived()); + copy_using_evaluator(dst.derived(), src.derived(), add_assign_op()); } -// TODO: Add add_assign_using_evaluator for EigenBase ? +// TODO: Add add_assign_using_evaluator for EigenBase ? (Jitse) template void subtract_assign_using_evaluator(const MatrixBase& dst, const MatrixBase& src) { typedef typename DstXprType::Scalar Scalar; - SelfCwiseBinaryOp, DstXprType, SrcXprType> tmp(dst.const_cast_derived()); - copy_using_evaluator(tmp, src.derived()); + copy_using_evaluator(dst.derived(), src.derived(), sub_assign_op()); } template void subtract_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) { typedef typename DstXprType::Scalar Scalar; - SelfCwiseBinaryOp, DstXprType, SrcXprType> tmp(dst.const_cast_derived()); - copy_using_evaluator(tmp, src.derived()); + copy_using_evaluator(dst.derived(), src.derived(), sub_assign_op()); } template void multiply_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) { typedef typename DstXprType::Scalar Scalar; - SelfCwiseBinaryOp, DstXprType, SrcXprType> tmp(dst.const_cast_derived()); - copy_using_evaluator(tmp, src.derived()); + copy_using_evaluator(dst.derived(), src.derived(), mul_assign_op()); } template void divide_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) { typedef typename DstXprType::Scalar Scalar; - SelfCwiseBinaryOp, DstXprType, SrcXprType> tmp(dst.const_cast_derived()); - copy_using_evaluator(tmp, src.derived()); + copy_using_evaluator(dst.derived(), src.derived(), div_assign_op()); } diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 3240ec6ed..082c00df4 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -78,6 +78,8 @@ template struct evaluator_impl_base { typedef typename ExpressionType::Index Index; + // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices. + typedef traits ExpressionTraits; template void copyCoeff(Index row, Index col, const OtherEvaluatorType& other) @@ -307,15 +309,17 @@ struct evaluator_impl > evaluator_impl(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()), m_resultImpl(m_result) - { - copy_using_evaluator_without_resizing(m_result, xpr.arg()); + { + // TODO we should simply do m_result(xpr.arg()); + call_dense_assignment_loop(m_result, xpr.arg()); } // This constructor is used when nesting an EvalTo evaluator in another evaluator evaluator_impl(const ArgType& arg) : m_result(arg.rows(), arg.cols()), m_resultImpl(m_result) - { - copy_using_evaluator_without_resizing(m_result, arg); + { + // TODO we should simply do m_result(xpr.arg()); + call_dense_assignment_loop(m_result, arg); } typedef typename PlainObject::Index Index; diff --git a/Eigen/src/Core/functors/AssignmentFunctors.h b/Eigen/src/Core/functors/AssignmentFunctors.h new file mode 100644 index 000000000..ae264aa64 --- /dev/null +++ b/Eigen/src/Core/functors/AssignmentFunctors.h @@ -0,0 +1,167 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2010 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_ASSIGNMENT_FUNCTORS_H +#define EIGEN_ASSIGNMENT_FUNCTORS_H + +namespace Eigen { + +namespace internal { + +/** \internal + * \brief Template functor for scalar/packet assignment + * + */ +template struct assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const { a = b; } + + template + EIGEN_STRONG_INLINE void assignPacket(Scalar* a, const Packet& b) const + { internal::pstoret(a,b); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::ReadCost, + PacketAccess = packet_traits::IsVectorized + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with addition + * + */ +template struct add_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(add_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const { a += b; } + + template + EIGEN_STRONG_INLINE void assignPacket(Scalar* a, const Packet& b) const + { internal::pstoret(a,internal::padd(internal::ploadt(a),b)); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::ReadCost + NumTraits::AddCost, + PacketAccess = packet_traits::HasAdd + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with subtraction + * + */ +template struct sub_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(sub_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const { a -= b; } + + template + EIGEN_STRONG_INLINE void assignPacket(Scalar* a, const Packet& b) const + { internal::pstoret(a,internal::psub(internal::ploadt(a),b)); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::ReadCost + NumTraits::AddCost, + PacketAccess = packet_traits::HasAdd + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with multiplication + * + */ +template struct mul_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(mul_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const { a *= b; } + + template + EIGEN_STRONG_INLINE void assignPacket(Scalar* a, const Packet& b) const + { internal::pstoret(a,internal::pmul(internal::ploadt(a),b)); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::ReadCost + NumTraits::MulCost, + PacketAccess = packet_traits::HasMul + }; +}; + +/** \internal + * \brief Template functor for scalar/packet assignment with diviving + * + */ +template struct div_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(div_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const { a /= b; } + + template + EIGEN_STRONG_INLINE void assignPacket(Scalar* a, const Packet& b) const + { internal::pstoret(a,internal::pdiv(internal::ploadt(a),b)); } +}; +template +struct functor_traits > { + enum { + Cost = NumTraits::ReadCost + NumTraits::MulCost, + PacketAccess = packet_traits::HasMul + }; +}; + + +/** \internal + * \brief Template functor for scalar/packet assignment with swaping + * + * It works as follow. For a non-vectorized evaluation loop, we have: + * for(i) func(A.coeffRef(i), B.coeff(i)); + * where B is a SwapWrapper expression. The trick is to make SwapWrapper::coeff behaves like a non-const coeffRef. + * Actually, SwapWrapper might not even be needed since even if B is a plain expression, since it has to be writable + * B.coeff already returns a const reference to the underlying scalar value. + * + * The case of a vectorized loop is more tricky: + * for(i,j) func.assignPacket(&A.coeffRef(i,j), B.packet(i,j)); + * Here, B must be a SwapWrapper whose packet function actually returns a proxy object holding a Scalar*, + * the actual alignment and Packet type. + * + */ +template struct swap_assign_op { + + EIGEN_EMPTY_STRUCT_CTOR(swap_assign_op) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const + { + using std::swap; + swap(a,const_cast(b)); + } + + template + EIGEN_STRONG_INLINE void swapPacket(Scalar* a, Scalar* b) const + { + Packet tmp = internal::ploadt(b); + internal::pstoret(b, internal::ploadt(a)); + internal::pstoret(a, tmp); + } +}; +template +struct functor_traits > { + enum { + Cost = 3 * NumTraits::ReadCost, + PacketAccess = packet_traits::IsVectorized + }; +}; + +} // namespace internal + +} // namespace Eigen + +#endif // EIGEN_ASSIGNMENT_FUNCTORS_H diff --git a/test/evaluator_common.h b/test/evaluator_common.h new file mode 100644 index 000000000..e69de29bb From 8fe609311d696189d443f42877a09fdf931220b4 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Nov 2013 09:01:26 +0100 Subject: [PATCH 0177/1103] Move internal::swap to numext to fix ambiguous call with std::swap --- Eigen/src/Core/util/Meta.h | 14 +++++++++++--- Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h | 6 +++--- 2 files changed, 14 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h index aea168b46..e4e4d4a87 100644 --- a/Eigen/src/Core/util/Meta.h +++ b/Eigen/src/Core/util/Meta.h @@ -89,9 +89,9 @@ template struct enable_if { typedef T type; }; #if defined(__CUDA_ARCH__) -template EIGEN_DEVICE_FUNC void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; } namespace device { + template struct numeric_limits { EIGEN_DEVICE_FUNC @@ -110,8 +110,6 @@ template<> struct numeric_limits } -#else -template EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); } #endif /** \internal @@ -262,6 +260,16 @@ template struct is_diagonal > } // end namespace internal +namespace numext { + +#if defined(__CUDA_ARCH__) +template EIGEN_DEVICE_FUNC void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; } +#else +template EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); } +#endif + +} // end namespace numext + } // end namespace Eigen #endif // EIGEN_META_H diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index be2f7b452..0c3425069 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -606,12 +606,12 @@ template struct direct_selfadjoint_eigenvalues= roots(1)) - internal::swap(roots(0),roots(1)); + numext::swap(roots(0),roots(1)); if (roots(1) >= roots(2)) { - internal::swap(roots(1),roots(2)); + numext::swap(roots(1),roots(2)); if (roots(0) >= roots(1)) - internal::swap(roots(0),roots(1)); + numext::swap(roots(0),roots(1)); } } From af9851d1d73b3aae5804f62d2d9e690be33bc3a5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Nov 2013 12:03:12 +0100 Subject: [PATCH 0178/1103] bug #99: move the creation of the evaluator to a central place, and make generic_dense_assignment_kernel hold the destination and source evaluators --- Eigen/src/Core/AssignEvaluator.h | 559 +++++++++++++---------------- Eigen/src/Core/ProductEvaluators.h | 2 +- 2 files changed, 254 insertions(+), 307 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index adea7ef13..5b5d29ca9 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -139,9 +139,10 @@ public: *** Default traversal *** ************************/ -template +template struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling { + typedef typename Kernel::DstEvaluatorType DstEvaluatorType; typedef typename DstEvaluatorType::XprType DstXprType; enum { @@ -149,126 +150,101 @@ struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling inner = Index % DstXprType::InnerSizeAtCompileTime }; - static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator, - const Kernel &kernel - ) + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - kernel.assignCoeffByOuterInner(outer, inner, dstEvaluator, srcEvaluator); - copy_using_evaluator_DefaultTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); + kernel.assignCoeffByOuterInner(outer, inner); + copy_using_evaluator_DefaultTraversal_CompleteUnrolling::run(kernel); } }; -template -struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling +template +struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&) { } + static EIGEN_STRONG_INLINE void run(Kernel&) { } }; -template +template struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator, - const Kernel &kernel, - int outer) + static EIGEN_STRONG_INLINE void run(Kernel &kernel, int outer) { - kernel.assignCoeffByOuterInner(outer, Index, dstEvaluator, srcEvaluator); - copy_using_evaluator_DefaultTraversal_InnerUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel, outer); + kernel.assignCoeffByOuterInner(outer, Index); + copy_using_evaluator_DefaultTraversal_InnerUnrolling::run(kernel, outer); } }; -template -struct copy_using_evaluator_DefaultTraversal_InnerUnrolling +template +struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&, int) { } + static EIGEN_STRONG_INLINE void run(Kernel&, int) { } }; /*********************** *** Linear traversal *** ***********************/ -template +template struct copy_using_evaluator_LinearTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator, - const Kernel& kernel - ) + static EIGEN_STRONG_INLINE void run(Kernel& kernel) { - kernel.assignCoeff(Index, dstEvaluator, srcEvaluator); - copy_using_evaluator_LinearTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); + kernel.assignCoeff(Index); + copy_using_evaluator_LinearTraversal_CompleteUnrolling::run(kernel); } }; -template -struct copy_using_evaluator_LinearTraversal_CompleteUnrolling +template +struct copy_using_evaluator_LinearTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&) { } + static EIGEN_STRONG_INLINE void run(Kernel&) { } }; /************************** *** Inner vectorization *** **************************/ -template +template struct copy_using_evaluator_innervec_CompleteUnrolling { + typedef typename Kernel::DstEvaluatorType DstEvaluatorType; typedef typename DstEvaluatorType::XprType DstXprType; - typedef typename SrcEvaluatorType::XprType SrcXprType; enum { outer = Index / DstXprType::InnerSizeAtCompileTime, inner = Index % DstXprType::InnerSizeAtCompileTime, - JointAlignment = copy_using_evaluator_traits::JointAlignment + JointAlignment = Kernel::AssignmentTraits::JointAlignment }; - static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator, - const Kernel &kernel - ) + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - kernel.template assignPacketByOuterInner(outer, inner, dstEvaluator, srcEvaluator); + kernel.template assignPacketByOuterInner(outer, inner); enum { NextIndex = Index + packet_traits::size }; - copy_using_evaluator_innervec_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); + copy_using_evaluator_innervec_CompleteUnrolling::run(kernel); } }; -template -struct copy_using_evaluator_innervec_CompleteUnrolling +template +struct copy_using_evaluator_innervec_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel&) { } + static EIGEN_STRONG_INLINE void run(Kernel&) { } }; -template +template struct copy_using_evaluator_innervec_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - SrcEvaluatorType &srcEvaluator, - const Kernel &kernel, - int outer) + static EIGEN_STRONG_INLINE void run(Kernel &kernel, int outer) { - kernel.template assignPacketByOuterInner(outer, Index, dstEvaluator, srcEvaluator); - typedef typename DstEvaluatorType::XprType DstXprType; + kernel.template assignPacketByOuterInner(outer, Index); + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; enum { NextIndex = Index + packet_traits::size }; - copy_using_evaluator_innervec_InnerUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel, outer); + copy_using_evaluator_innervec_InnerUnrolling::run(kernel, outer); } }; -template -struct copy_using_evaluator_innervec_InnerUnrolling +template +struct copy_using_evaluator_innervec_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(DstEvaluatorType&, SrcEvaluatorType&, const Kernel &, int) { } + static EIGEN_STRONG_INLINE void run(Kernel &, int) { } }; /*************************************************************************** @@ -277,69 +253,51 @@ struct copy_using_evaluator_innervec_InnerUnrolling::Traversal, - int Unrolling = copy_using_evaluator_traits::Unrolling> +template struct dense_assignment_loop; /************************ *** Default traversal *** ************************/ -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - static void run(DstXprType& dst, const SrcXprType& src, const Kernel &kernel) + static void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - typedef typename DstXprType::Index Index; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - for(Index outer = 0; outer < dst.outerSize(); ++outer) { - for(Index inner = 0; inner < dst.innerSize(); ++inner) { - kernel.assignCoeffByOuterInner(outer, inner, dstEvaluator, srcEvaluator); + typedef typename Kernel::Index Index; + + for(Index outer = 0; outer < kernel.outerSize(); ++outer) { + for(Index inner = 0; inner < kernel.innerSize(); ++inner) { + kernel.assignCoeffByOuterInner(outer, inner); } } } }; -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - copy_using_evaluator_DefaultTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + copy_using_evaluator_DefaultTraversal_CompleteUnrolling::run(kernel); } }; -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - typedef typename DstXprType::Index Index; - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + typedef typename Kernel::Index Index; + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - const Index outerSize = dst.outerSize(); + const Index outerSize = kernel.outerSize(); for(Index outer = 0; outer < outerSize; ++outer) - copy_using_evaluator_DefaultTraversal_InnerUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel, outer); + copy_using_evaluator_DefaultTraversal_InnerUnrolling::run(kernel, outer); } }; @@ -355,9 +313,8 @@ template struct unaligned_dense_assignment_loop { // if IsAligned = true, then do nothing - template - static EIGEN_STRONG_INLINE void run(const SrcEvaluatorType&, DstEvaluatorType&, const Kernel&, - typename SrcEvaluatorType::Index, typename SrcEvaluatorType::Index) {} + template + static EIGEN_STRONG_INLINE void run(Kernel&, typename Kernel::Index, typename Kernel::Index) {} }; template <> @@ -367,80 +324,63 @@ struct unaligned_dense_assignment_loop // packet access path. // FIXME check which version exhibits this issue #ifdef _MSC_VER - template - static EIGEN_DONT_INLINE void run(DstEvaluatorType &dstEvaluator, - const SrcEvaluatorType &srcEvaluator, - const Kernel &kernel, - typename DstEvaluatorType::Index start, - typename DstEvaluatorType::Index end) + template + static EIGEN_DONT_INLINE void run(Kernel &kernel, + typename Kernel::Index start, + typename Kernel::Index end) #else - template - static EIGEN_STRONG_INLINE void run(DstEvaluatorType &dstEvaluator, - const SrcEvaluatorType &srcEvaluator, - const Kernel &kernel, - typename DstEvaluatorType::Index start, - typename DstEvaluatorType::Index end) + template + static EIGEN_STRONG_INLINE void run(Kernel &kernel, + typename Kernel::Index start, + typename Kernel::Index end) #endif { - for (typename DstEvaluatorType::Index index = start; index < end; ++index) - kernel.assignCoeff(index, dstEvaluator, srcEvaluator); + for (typename Kernel::Index index = start; index < end; ++index) + kernel.assignCoeff(index); } }; -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - typedef typename DstXprType::Index Index; + typedef typename Kernel::Index Index; - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - const Index size = dst.size(); - typedef packet_traits PacketTraits; + const Index size = kernel.size(); + typedef packet_traits PacketTraits; enum { packetSize = PacketTraits::size, - dstIsAligned = int(copy_using_evaluator_traits::DstIsAligned), + dstIsAligned = int(Kernel::AssignmentTraits::DstIsAligned), dstAlignment = PacketTraits::AlignedOnScalar ? Aligned : dstIsAligned, - srcAlignment = copy_using_evaluator_traits::JointAlignment + srcAlignment = Kernel::AssignmentTraits::JointAlignment }; - const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned(&dstEvaluator.coeffRef(0), size); + const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned(&kernel.dstEvaluator().coeffRef(0), size); const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize; - unaligned_dense_assignment_loop::run(dstEvaluator, srcEvaluator, kernel, 0, alignedStart); + unaligned_dense_assignment_loop::run(kernel, 0, alignedStart); for(Index index = alignedStart; index < alignedEnd; index += packetSize) - kernel.template assignPacket(index, dstEvaluator, srcEvaluator); + kernel.template assignPacket(index); - unaligned_dense_assignment_loop<>::run(dstEvaluator, srcEvaluator, kernel, alignedEnd, size); + unaligned_dense_assignment_loop<>::run(kernel, alignedEnd, size); } }; -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - typedef typename DstXprType::Index Index; - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + typedef typename Kernel::Index Index; + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + enum { size = DstXprType::SizeAtCompileTime, - packetSize = packet_traits::size, + packetSize = packet_traits::size, alignedSize = (size/packetSize)*packetSize }; - copy_using_evaluator_innervec_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); - copy_using_evaluator_DefaultTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); + copy_using_evaluator_innervec_CompleteUnrolling::run(kernel); + copy_using_evaluator_DefaultTraversal_CompleteUnrolling::run(kernel); } }; @@ -448,61 +388,42 @@ struct dense_assignment_loop -struct dense_assignment_loop +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + static inline void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - typedef typename DstXprType::Index Index; + typedef typename Kernel::Index Index; - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - const Index innerSize = dst.innerSize(); - const Index outerSize = dst.outerSize(); - const Index packetSize = packet_traits::size; + const Index innerSize = kernel.innerSize(); + const Index outerSize = kernel.outerSize(); + const Index packetSize = packet_traits::size; for(Index outer = 0; outer < outerSize; ++outer) for(Index inner = 0; inner < innerSize; inner+=packetSize) - kernel.template assignPacketByOuterInner(outer, inner, dstEvaluator, srcEvaluator); + kernel.template assignPacketByOuterInner(outer, inner); } }; -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - copy_using_evaluator_innervec_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + copy_using_evaluator_innervec_CompleteUnrolling::run(kernel); } }; -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - typedef typename DstXprType::Index Index; - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + typedef typename Kernel::Index Index; + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - const Index outerSize = dst.outerSize(); + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + const Index outerSize = kernel.outerSize(); for(Index outer = 0; outer < outerSize; ++outer) - copy_using_evaluator_innervec_InnerUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel, outer); + copy_using_evaluator_innervec_InnerUnrolling::run(kernel, outer); } }; @@ -510,38 +431,25 @@ struct dense_assignment_loop -struct dense_assignment_loop +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + static inline void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - typedef typename DstXprType::Index Index; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - const Index size = dst.size(); + typedef typename Kernel::Index Index; + const Index size = kernel.size(); for(Index i = 0; i < size; ++i) - kernel.assignCoeff(i, dstEvaluator, srcEvaluator); + kernel.assignCoeff(i); } }; -template -struct dense_assignment_loop +template +struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + static EIGEN_STRONG_INLINE void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - copy_using_evaluator_LinearTraversal_CompleteUnrolling - - ::run(dstEvaluator, srcEvaluator, kernel); + typedef typename Kernel::DstEvaluatorType::XprType DstXprType; + copy_using_evaluator_LinearTraversal_CompleteUnrolling::run(kernel); } }; @@ -549,45 +457,39 @@ struct dense_assignment_loop -struct dense_assignment_loop +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &kernel) + static inline void run(Kernel &kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - typedef typename DstXprType::Index Index; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - typedef packet_traits PacketTraits; + typedef typename Kernel::Index Index; + typedef packet_traits PacketTraits; enum { packetSize = PacketTraits::size, alignable = PacketTraits::AlignedOnScalar, - dstAlignment = alignable ? Aligned : int(copy_using_evaluator_traits::DstIsAligned) + dstAlignment = alignable ? Aligned : int(Kernel::AssignmentTraits::DstIsAligned) }; const Index packetAlignedMask = packetSize - 1; - const Index innerSize = dst.innerSize(); - const Index outerSize = dst.outerSize(); - const Index alignedStep = alignable ? (packetSize - dst.outerStride() % packetSize) & packetAlignedMask : 0; - Index alignedStart = ((!alignable) || copy_using_evaluator_traits::DstIsAligned) ? 0 - : internal::first_aligned(&dstEvaluator.coeffRef(0,0), innerSize); + const Index innerSize = kernel.innerSize(); + const Index outerSize = kernel.outerSize(); + const Index alignedStep = alignable ? (packetSize - kernel.outerStride() % packetSize) & packetAlignedMask : 0; + Index alignedStart = ((!alignable) || Kernel::AssignmentTraits::DstIsAligned) ? 0 + : internal::first_aligned(&kernel.dstEvaluator().coeffRef(0,0), innerSize); for(Index outer = 0; outer < outerSize; ++outer) { const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask); // do the non-vectorizable part of the assignment for(Index inner = 0; inner(outer, inner, dstEvaluator, srcEvaluator); + kernel.template assignPacketByOuterInner(outer, inner); // do the non-vectorizable part of the assignment for(Index inner = alignedEnd; inner((alignedStart+alignedStep)%packetSize, innerSize); } @@ -599,21 +501,15 @@ struct dense_assignment_loop -struct dense_assignment_loop +// Indeed, what to do with the kernel's functor?? +template +struct dense_assignment_loop { - static inline void run(DstXprType &dst, const SrcXprType &src, const Kernel &/*kernel*/) + static inline void run(Kernel & kernel) { - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - // Evaluate rhs in temporary to prevent aliasing problems in a = a * a; // TODO: Do not pass the xpr object to evalTo() (Jitse) - srcEvaluator.evalTo(dstEvaluator, dst); + kernel.srcEvaluator().evalTo(kernel.dstEvaluator(), kernel.dstExpression()); } }; @@ -623,94 +519,123 @@ struct dense_assignment_loop -struct generic_dense_assignment_kernel +template +class generic_dense_assignment_kernel { - const Functor &m_functor; - generic_dense_assignment_kernel(const Functor &func) : m_functor(func) {} +protected: + typedef typename DstEvaluatorTypeT::XprType DstXprType; + typedef typename SrcEvaluatorTypeT::XprType SrcXprType; +public: - template - void assignCoeff(typename DstEvaluatorType::Index row, typename DstEvaluatorType::Index col, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + typedef DstEvaluatorTypeT DstEvaluatorType; + typedef SrcEvaluatorTypeT SrcEvaluatorType; + typedef typename DstEvaluatorType::Scalar Scalar; + typedef typename DstEvaluatorType::Index Index; + typedef copy_using_evaluator_traits AssignmentTraits; + + + generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + : m_dst(dst), m_src(src), m_functor(func), m_dstExpr(dstExpr) + {} + + Index size() const { return m_dstExpr.size(); } + Index innerSize() const { return m_dstExpr.innerSize(); } + Index outerSize() const { return m_dstExpr.outerSize(); } + Index outerStride() const { return m_dstExpr.outerStride(); } + + // TODO get rid of this one: + DstXprType& dstExpression() const { return m_dstExpr; } + + DstEvaluatorType& dstEvaluator() { return m_dst; } + const SrcEvaluatorType& srcEvaluator() const { return m_src; } + + void assignCoeff(Index row, Index col) { - m_functor.assignCoeff(dst.coeffRef(row,col), src.coeff(row,col)); + m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col)); } - template - void assignCoeff(typename DstEvaluatorType::Index index, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + void assignCoeff(Index index) { - m_functor.assignCoeff(dst.coeffRef(index), src.coeff(index)); + m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index)); } - template - void assignCoeffByOuterInner(typename DstEvaluatorType::Index outer, typename DstEvaluatorType::Index inner, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + void assignCoeffByOuterInner(Index outer, Index inner) { - typedef typename DstEvaluatorType::Index Index; - Index row = rowIndexByOuterInner(outer, inner); - Index col = colIndexByOuterInner(outer, inner); - assignCoeff(row, col, dst, src); + Index row = rowIndexByOuterInner(outer, inner); + Index col = colIndexByOuterInner(outer, inner); + assignCoeff(row, col); } - template - void assignPacket(typename DstEvaluatorType::Index row, typename DstEvaluatorType::Index col, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + template + void assignPacket(Index row, Index col) { - m_functor.assignPacket(&dst.coeffRef(row,col), src.template packet(row,col)); + m_functor.assignPacket(&m_dst.coeffRef(row,col), m_src.template packet(row,col)); } - template - void assignPacket(typename DstEvaluatorType::Index index, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + template + void assignPacket(Index index) { - m_functor.assignPacket(&dst.coeffRef(index), src.template packet(index)); + m_functor.assignPacket(&m_dst.coeffRef(index), m_src.template packet(index)); } - template - void assignPacketByOuterInner(typename DstEvaluatorType::Index outer, typename DstEvaluatorType::Index inner, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + template + void assignPacketByOuterInner(Index outer, Index inner) { - typedef typename DstEvaluatorType::Index Index; - Index row = rowIndexByOuterInner(outer, inner); - Index col = colIndexByOuterInner(outer, inner); - assignPacket(row, col, dst, src); + Index row = rowIndexByOuterInner(outer, inner); + Index col = colIndexByOuterInner(outer, inner); + assignPacket(row, col); } - template static Index rowIndexByOuterInner(Index outer, Index inner) { - typedef typename EvaluatorType::ExpressionTraits Traits; + typedef typename DstEvaluatorType::ExpressionTraits Traits; return int(Traits::RowsAtCompileTime) == 1 ? 0 : int(Traits::ColsAtCompileTime) == 1 ? inner : int(Traits::Flags)&RowMajorBit ? outer : inner; } - template static Index colIndexByOuterInner(Index outer, Index inner) { - typedef typename EvaluatorType::ExpressionTraits Traits; + typedef typename DstEvaluatorType::ExpressionTraits Traits; return int(Traits::ColsAtCompileTime) == 1 ? 0 : int(Traits::RowsAtCompileTime) == 1 ? inner : int(Traits::Flags)&RowMajorBit ? inner : outer; } + +protected: + DstEvaluatorType& m_dst; + const SrcEvaluatorType& m_src; + const Functor &m_functor; + // TODO find a way to avoid the needs of the original expression + DstXprType& m_dstExpr; }; template void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) { #ifdef EIGEN_DEBUG_ASSIGN + // TODO these traits should be computed from information provided by the evaluators internal::copy_using_evaluator_traits::debug(); #endif eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); - typedef generic_dense_assignment_kernel Kernel; - Kernel kernel(func); + typedef typename evaluator::type DstEvaluatorType; + typedef typename evaluator::type SrcEvaluatorType; + + DstEvaluatorType dstEvaluator(dst); + SrcEvaluatorType srcEvaluator(src); + + typedef generic_dense_assignment_kernel Kernel; + Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived()); - dense_assignment_loop::run(const_cast(dst), src, kernel); + dense_assignment_loop::run(kernel); } template @@ -804,41 +729,63 @@ const DstXprType& noalias_copy_using_evaluator(const EigenBase& dst, // Arrays or Matrices. (Jitse) // Overload default assignPacket behavior for swapping them -template -struct swap_kernel : generic_dense_assignment_kernel > +template +class swap_kernel : public generic_dense_assignment_kernel > { - typedef generic_dense_assignment_kernel > Base; + typedef generic_dense_assignment_kernel > Base; + typedef typename DstEvaluatorTypeT::PacketScalar PacketScalar; + using Base::m_dst; + using Base::m_src; using Base::m_functor; - swap_kernel() : Base(swap_assign_op()) {} - template - void assignPacket(typename DstEvaluatorType::Index row, typename DstEvaluatorType::Index col, DstEvaluatorType &dst, const SrcEvaluatorType &src) const +public: + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + typedef typename Base::DstXprType DstXprType; + + swap_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, DstXprType& dstExpr) + : Base(dst, src, swap_assign_op(), dstExpr) + {} + + template + void assignPacket(Index row, Index col) { - m_functor.template swapPacket(&dst.coeffRef(row,col), &const_cast(src).coeffRef(row,col)); + m_functor.template swapPacket(&m_dst.coeffRef(row,col), &const_cast(m_src).coeffRef(row,col)); } - template - void assignPacket(typename DstEvaluatorType::Index index, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + template + void assignPacket(Index index) { - m_functor.template swapPacket(&dst.coeffRef(index), &const_cast(src).coeffRef(index)); + m_functor.template swapPacket(&m_dst.coeffRef(index), &const_cast(m_src).coeffRef(index)); } // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael) - template - void assignPacketByOuterInner(typename DstEvaluatorType::Index outer, typename DstEvaluatorType::Index inner, DstEvaluatorType &dst, const SrcEvaluatorType &src) const + template + void assignPacketByOuterInner(Index outer, Index inner) { - typedef typename DstEvaluatorType::Index Index; - Index row = Base::template rowIndexByOuterInner(outer, inner); - Index col = Base::template colIndexByOuterInner(outer, inner); - assignPacket(row, col, dst, src); + Index row = Base::rowIndexByOuterInner(outer, inner); + Index col = Base::colIndexByOuterInner(outer, inner); + assignPacket(row, col); } }; template void swap_using_evaluator(const DstXprType& dst, const SrcXprType& src) { - typedef swap_kernel kernel; - dense_assignment_loop::run(const_cast(dst), src, kernel()); + // TODO there is too much redundancy with call_dense_assignment_loop + + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + typedef typename evaluator::type DstEvaluatorType; + typedef typename evaluator::type SrcEvaluatorType; + + DstEvaluatorType dstEvaluator(dst); + SrcEvaluatorType srcEvaluator(src); + + typedef swap_kernel Kernel; + Kernel kernel(dstEvaluator, srcEvaluator, dst.const_cast_derived()); + + dense_assignment_loop::run(kernel); } // Based on MatrixBase::operator+= (in CwiseBinaryOp.h) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 8aed51022..855914f2e 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -95,7 +95,7 @@ struct product_evaluator_dispatcher, GeneralProduct - void evalTo(DstEvaluatorType /* not used */, DstXprType& dst) + void evalTo(DstEvaluatorType /* not used */, DstXprType& dst) const { dst.resize(m_xpr.rows(), m_xpr.cols()); GeneralProduct(m_xpr.lhs(), m_xpr.rhs()).evalTo(dst); From 5887e82729958f2b07ac7564a2929d26f3bf3a83 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Nov 2013 14:02:47 +0100 Subject: [PATCH 0179/1103] Clean evaluator_impl_base. It will probably be removed in the future --- Eigen/src/Core/CoreEvaluators.h | 59 +-------------------------------- 1 file changed, 1 insertion(+), 58 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 082c00df4..3b9dfee7a 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -74,6 +74,7 @@ struct evaluator // ---------- base class for all writable evaluators ---------- +// TODO this class does not seem to be necessary anymore template struct evaluator_impl_base { @@ -81,64 +82,6 @@ struct evaluator_impl_base // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices. typedef traits ExpressionTraits; - template - void copyCoeff(Index row, Index col, const OtherEvaluatorType& other) - { - derived().coeffRef(row, col) = other.coeff(row, col); - } - - template - void copyCoeffByOuterInner(Index outer, Index inner, const OtherEvaluatorType& other) - { - Index row = rowIndexByOuterInner(outer, inner); - Index col = colIndexByOuterInner(outer, inner); - derived().copyCoeff(row, col, other); - } - - template - void copyCoeff(Index index, const OtherEvaluatorType& other) - { - derived().coeffRef(index) = other.coeff(index); - } - - template - void copyPacket(Index row, Index col, const OtherEvaluatorType& other) - { - derived().template writePacket(row, col, - other.template packet(row, col)); - } - - template - void copyPacketByOuterInner(Index outer, Index inner, const OtherEvaluatorType& other) - { - Index row = rowIndexByOuterInner(outer, inner); - Index col = colIndexByOuterInner(outer, inner); - derived().template copyPacket(row, col, other); - } - - template - void copyPacket(Index index, const OtherEvaluatorType& other) - { - derived().template writePacket(index, - other.template packet(index)); - } - - Index rowIndexByOuterInner(Index outer, Index inner) const - { - return int(ExpressionType::RowsAtCompileTime) == 1 ? 0 - : int(ExpressionType::ColsAtCompileTime) == 1 ? inner - : int(ExpressionType::Flags)&RowMajorBit ? outer - : inner; - } - - Index colIndexByOuterInner(Index outer, Index inner) const - { - return int(ExpressionType::ColsAtCompileTime) == 1 ? 0 - : int(ExpressionType::RowsAtCompileTime) == 1 ? inner - : int(ExpressionType::Flags)&RowMajorBit ? inner - : outer; - } - evaluator_impl& derived() { return *static_cast*>(this); From 57327cc2d5aab1cc51dd1308bd0d000f2ccb623c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Nov 2013 14:07:27 +0100 Subject: [PATCH 0180/1103] Drop evaluators for SwapWrapper and SelfCwiseBinaryOp --- Eigen/src/Core/CoreEvaluators.h | 143 +------------------------------- 1 file changed, 3 insertions(+), 140 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 3b9dfee7a..3568cb85f 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -178,9 +178,9 @@ protected: const Scalar *m_data; // We do not need to know the outer stride for vectors - variable_if_dynamic m_outerStride; + variable_if_dynamic m_outerStride; }; template @@ -1114,143 +1114,6 @@ private: EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; } }; - -// ---------- SwapWrapper ---------- - -template -struct evaluator_impl > - : evaluator_impl_base > -{ - typedef SwapWrapper XprType; - - evaluator_impl(const XprType& swapWrapper) - : m_argImpl(swapWrapper.expression()) - { } - - typedef typename XprType::Index Index; - typedef typename XprType::Scalar Scalar; - typedef typename XprType::Packet Packet; - - // This function and the next one are needed by assign to correctly align loads/stores - // TODO make Assign use .data() - Scalar& coeffRef(Index row, Index col) - { - return m_argImpl.coeffRef(row, col); - } - - inline Scalar& coeffRef(Index index) - { - return m_argImpl.coeffRef(index); - } - - template - void copyCoeff(Index row, Index col, const OtherEvaluatorType& other) - { - OtherEvaluatorType& nonconst_other = const_cast(other); - Scalar tmp = m_argImpl.coeff(row, col); - m_argImpl.coeffRef(row, col) = nonconst_other.coeff(row, col); - nonconst_other.coeffRef(row, col) = tmp; - } - - template - void copyCoeff(Index index, const OtherEvaluatorType& other) - { - OtherEvaluatorType& nonconst_other = const_cast(other); - Scalar tmp = m_argImpl.coeff(index); - m_argImpl.coeffRef(index) = nonconst_other.coeff(index); - nonconst_other.coeffRef(index) = tmp; - } - - template - void copyPacket(Index row, Index col, const OtherEvaluatorType& other) - { - OtherEvaluatorType& nonconst_other = const_cast(other); - Packet tmp = m_argImpl.template packet(row, col); - m_argImpl.template writePacket - (row, col, nonconst_other.template packet(row, col)); - nonconst_other.template writePacket(row, col, tmp); - } - - template - void copyPacket(Index index, const OtherEvaluatorType& other) - { - OtherEvaluatorType& nonconst_other = const_cast(other); - Packet tmp = m_argImpl.template packet(index); - m_argImpl.template writePacket - (index, nonconst_other.template packet(index)); - nonconst_other.template writePacket(index, tmp); - } - -protected: - typename evaluator::nestedType m_argImpl; -}; - - -// ---------- SelfCwiseBinaryOp ---------- - -template -struct evaluator_impl > - : evaluator_impl_base > -{ - typedef SelfCwiseBinaryOp XprType; - - evaluator_impl(const XprType& selfCwiseBinaryOp) - : m_argImpl(selfCwiseBinaryOp.expression()), - m_functor(selfCwiseBinaryOp.functor()) - { } - - typedef typename XprType::Index Index; - typedef typename XprType::Scalar Scalar; - typedef typename XprType::Packet Packet; - - // This function and the next one are needed by assign to correctly align loads/stores - // TODO make Assign use .data() - Scalar& coeffRef(Index row, Index col) - { - return m_argImpl.coeffRef(row, col); - } - - inline Scalar& coeffRef(Index index) - { - return m_argImpl.coeffRef(index); - } - - template - void copyCoeff(Index row, Index col, const OtherEvaluatorType& other) - { - Scalar& tmp = m_argImpl.coeffRef(row, col); - tmp = m_functor(tmp, other.coeff(row, col)); - } - - template - void copyCoeff(Index index, const OtherEvaluatorType& other) - { - Scalar& tmp = m_argImpl.coeffRef(index); - tmp = m_functor(tmp, other.coeff(index)); - } - - template - void copyPacket(Index row, Index col, const OtherEvaluatorType& other) - { - const Packet res = m_functor.packetOp(m_argImpl.template packet(row, col), - other.template packet(row, col)); - m_argImpl.template writePacket(row, col, res); - } - - template - void copyPacket(Index index, const OtherEvaluatorType& other) - { - const Packet res = m_functor.packetOp(m_argImpl.template packet(index), - other.template packet(index)); - m_argImpl.template writePacket(index, res); - } - -protected: - typename evaluator::nestedType m_argImpl; - const BinaryOp m_functor; -}; - - } // namespace internal } // end namespace Eigen From 76c230a84d4857722e4a72e9007302a504af0fb7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Nov 2013 16:38:14 +0100 Subject: [PATCH 0181/1103] Add an option to test evaluators globally --- Eigen/Core | 16 +++++------ Eigen/src/Core/Assign.h | 19 ++++++++++-- Eigen/src/Core/GeneralProduct.h | 35 +++++++++++++++++++++++ Eigen/src/Core/MatrixBase.h | 8 ++++++ Eigen/src/Core/Product.h | 2 +- Eigen/src/Core/util/ForwardDeclarations.h | 1 + test/CMakeLists.txt | 7 +++++ test/main.h | 6 ++++ 8 files changed, 83 insertions(+), 11 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 722a49030..468ae0c76 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -311,6 +311,14 @@ using std::ptrdiff_t; #include "src/Core/MatrixBase.h" #include "src/Core/EigenBase.h" +#ifdef EIGEN_ENABLE_EVALUATORS +#include "src/Core/functors/AssignmentFunctors.h" +#include "src/Core/Product.h" +#include "src/Core/CoreEvaluators.h" +#include "src/Core/AssignEvaluator.h" +#include "src/Core/ProductEvaluators.h" +#endif + #ifndef EIGEN_PARSED_BY_DOXYGEN // work around Doxygen bug triggered by Assign.h r814874 // at least confirmed with Doxygen 1.5.5 and 1.5.6 #include "src/Core/Assign.h" @@ -382,14 +390,6 @@ using std::ptrdiff_t; #include "src/Core/Reverse.h" #include "src/Core/ArrayWrapper.h" -#ifdef EIGEN_ENABLE_EVALUATORS -#include "src/Core/functors/AssignmentFunctors.h" -#include "src/Core/Product.h" -#include "src/Core/CoreEvaluators.h" -#include "src/Core/AssignEvaluator.h" -#include "src/Core/ProductEvaluators.h" -#endif - #ifdef EIGEN_USE_BLAS #include "src/Core/products/GeneralMatrixMatrix_MKL.h" #include "src/Core/products/GeneralMatrixVector_MKL.h" diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index 906adcf82..c0b4bbfe0 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -504,12 +504,27 @@ EIGEN_STRONG_INLINE Derived& DenseBase EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Derived,OtherDerived) EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) +#ifdef EIGEN_TEST_EVALUATORS + +#ifdef EIGEN_DEBUG_ASSIGN + internal::copy_using_evaluator_traits::debug(); +#endif + eigen_assert(rows() == other.rows() && cols() == other.cols()); +// internal::copy_using_evaluator_impl::Traversal) +// : int(InvalidTraversal)>::run(derived(),other.derived()); + internal::call_dense_assignment_loop(derived(),other.derived()); + +#else // EIGEN_TEST_EVALUATORS + #ifdef EIGEN_DEBUG_ASSIGN internal::assign_traits::debug(); #endif eigen_assert(rows() == other.rows() && cols() == other.cols()); - internal::assign_impl::Traversal) - : int(InvalidTraversal)>::run(derived(),other.derived()); + internal::assign_impl::Traversal) + : int(InvalidTraversal)>::run(derived(),other.derived()); + +#endif // EIGEN_TEST_EVALUATORS + #ifndef EIGEN_NO_DEBUG checkTransposeAliasing(other.derived()); #endif diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 9d7d18427..e3a165ac6 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -566,6 +566,39 @@ template<> struct gemv_selector * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*() */ #ifndef __CUDACC__ + +#ifdef EIGEN_TEST_EVALUATORS +template +template +inline const Product +MatrixBase::operator*(const MatrixBase &other) const +{ + // A note regarding the function declaration: In MSVC, this function will sometimes + // not be inlined since DenseStorage is an unwindable object for dynamic + // matrices and product types are holding a member to store the result. + // Thus it does not help tagging this function with EIGEN_STRONG_INLINE. + enum { + ProductIsValid = Derived::ColsAtCompileTime==Dynamic + || OtherDerived::RowsAtCompileTime==Dynamic + || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), + AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, + SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) + }; + // note to the lost user: + // * for a dot product use: v1.dot(v2) + // * for a coeff-wise product use: v1.cwiseProduct(v2) + EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), + INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) + EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), + INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) + EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) +#ifdef EIGEN_DEBUG_PRODUCT + internal::product_type::debug(); +#endif + + return Product(derived(), other.derived()); +} +#else template template inline const typename ProductReturnType::Type @@ -595,6 +628,8 @@ MatrixBase::operator*(const MatrixBase &other) const #endif return typename ProductReturnType::Type(derived(), other.derived()); } +#endif + #endif /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation. * diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index e77b49627..61798317e 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -184,9 +184,17 @@ template class MatrixBase operator*(const MatrixBase &other) const { return this->lazyProduct(other); } #else + +#ifdef EIGEN_TEST_EVALUATORS + template + const Product + operator*(const MatrixBase &other) const; +#else template const typename ProductReturnType::Type operator*(const MatrixBase &other) const; +#endif + #endif template diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 3a08c027c..5d3789be7 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -36,7 +36,7 @@ struct traits > // We want A+B*C to be of type Product and not Product // TODO: This flag should eventually go in a separate evaluator traits class enum { - Flags = traits::Type>::Flags & ~EvalBeforeNestingBit + Flags = traits::Type>::Flags & ~(EvalBeforeNestingBit | DirectAccessBit) }; }; } // end namespace internal diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index dd0c18ad3..0a2144c69 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -89,6 +89,7 @@ template class CwiseUnaryView; template class CwiseBinaryOp; template class SelfCwiseBinaryOp; template class ProductBase; +template class Product; template class GeneralProduct; template class CoeffBasedProduct; diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 5b9e92f01..04cb11d60 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -126,6 +126,13 @@ endif(TEST_LIB) set_property(GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT "Official") add_custom_target(BuildOfficial) +option(EIGEN_TEST_EVALUATORS "Enable work in progress evaluators" OFF) + add_definitions("-DEIGEN_TEST_EVALUATORS=1") + add_definitions("-DEIGEN_ENABLE_EVALUATORS=1") +if(EIGEN_TEST_EVALUATORS) + +endif(EIGEN_TEST_EVALUATORS) + ei_add_test(meta) ei_add_test(sizeof) ei_add_test(dynalloc) diff --git a/test/main.h b/test/main.h index c889fa6c5..9dd8bc535 100644 --- a/test/main.h +++ b/test/main.h @@ -271,7 +271,13 @@ inline bool test_isApproxOrLessThan(const long double& a, const long double& b) template inline bool test_isApprox(const Type1& a, const Type2& b) { +#ifdef EIGEN_TEST_EVALUATORS + typename internal::eval::type a_eval(a); + typename internal::eval::type b_eval(b); + return a_eval.isApprox(b_eval, test_precision()); +#else return a.isApprox(b, test_precision()); +#endif } // The idea behind this function is to compare the two scalars a and b where From ae83f5ede96cb1171fb5fd9303777a9035748ca9 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 7 Nov 2013 18:32:24 +0100 Subject: [PATCH 0182/1103] Fixed bug #702 and added unit test. Thanks to Alexander Werner for the report. --- .../Eigen/src/AutoDiff/AutoDiffScalar.h | 8 ++-- unsupported/test/CMakeLists.txt | 1 + unsupported/test/autodiff_scalar.cpp | 44 +++++++++++++++++++ 3 files changed, 48 insertions(+), 5 deletions(-) create mode 100644 unsupported/test/autodiff_scalar.cpp diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h index 8d42e69b9..590797973 100644 --- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h +++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h @@ -599,12 +599,10 @@ atan2(const AutoDiffScalar& a, const AutoDiffScalar& b) PlainADS ret; ret.value() = atan2(a.value(), b.value()); - Scalar tmp2 = a.value() * a.value(); - Scalar tmp3 = b.value() * b.value(); - Scalar tmp4 = tmp3/(tmp2+tmp3); + Scalar squared_hypot = a.value() * a.value() + b.value() * b.value(); - if (tmp4!=0) - ret.derivatives() = (a.derivatives() * b.value() - a.value() * b.derivatives()) * (tmp2+tmp3); + // if (squared_hypot==0) the derivation is undefined and the following results in a NaN: + ret.derivatives() = (a.derivatives() * b.value() - a.value() * b.derivatives()) / squared_hypot; return ret; } diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index a94a3b5e5..38c89a234 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -28,6 +28,7 @@ endif(ADOLC_FOUND) ei_add_test(NonLinearOptimization) ei_add_test(NumericalDiff) +ei_add_test(autodiff_scalar) ei_add_test(autodiff) if (NOT CMAKE_CXX_COMPILER MATCHES "clang\\+\\+$") diff --git a/unsupported/test/autodiff_scalar.cpp b/unsupported/test/autodiff_scalar.cpp new file mode 100644 index 000000000..ba4b5aec4 --- /dev/null +++ b/unsupported/test/autodiff_scalar.cpp @@ -0,0 +1,44 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2013 Christoph Hertzberg +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" +#include + +/* + * In this file scalar derivations are tested for correctness. + * TODO add more tests! + */ + +template void check_atan2() +{ + typedef Matrix Deriv1; + typedef AutoDiffScalar AD; + + AD x(internal::random(-3.0, 3.0), Deriv1::UnitX()); + + using std::exp; + Scalar r = exp(internal::random(-10, 10)); + + AD s = sin(x), c = cos(x); + AD res = atan2(r*s, r*c); + + VERIFY_IS_APPROX(res.value(), x.value()); + VERIFY_IS_APPROX(res.derivatives(), x.derivatives()); +} + + + + +void test_autodiff_scalar() +{ + for(int i = 0; i < g_repeat; i++) { + CALL_SUBTEST_1( check_atan2() ); + CALL_SUBTEST_2( check_atan2() ); + } +} From fe0b44e876b621efe4bccce0a2b6333f3799faed Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Nov 2013 18:48:17 +0100 Subject: [PATCH 0183/1103] Fix stupid mistake in CMakeLists.txt --- test/CMakeLists.txt | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 04cb11d60..3c32f4b21 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -127,10 +127,9 @@ set_property(GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT "Official") add_custom_target(BuildOfficial) option(EIGEN_TEST_EVALUATORS "Enable work in progress evaluators" OFF) +if(EIGEN_TEST_EVALUATORS) add_definitions("-DEIGEN_TEST_EVALUATORS=1") add_definitions("-DEIGEN_ENABLE_EVALUATORS=1") -if(EIGEN_TEST_EVALUATORS) - endif(EIGEN_TEST_EVALUATORS) ei_add_test(meta) From cb8da751a0da5f03eac2bf35e56ab7572c164ac5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 7 Nov 2013 22:44:37 +0100 Subject: [PATCH 0184/1103] fix broken commit --- Eigen/src/Core/Assign.h | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index c0b4bbfe0..cefa6f3cc 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -510,8 +510,6 @@ EIGEN_STRONG_INLINE Derived& DenseBase internal::copy_using_evaluator_traits::debug(); #endif eigen_assert(rows() == other.rows() && cols() == other.cols()); -// internal::copy_using_evaluator_impl::Traversal) -// : int(InvalidTraversal)>::run(derived(),other.derived()); internal::call_dense_assignment_loop(derived(),other.derived()); #else // EIGEN_TEST_EVALUATORS @@ -520,7 +518,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase internal::assign_traits::debug(); #endif eigen_assert(rows() == other.rows() && cols() == other.cols()); - internal::assign_impl::Traversal) + internal::assign_impl::Traversal) : int(InvalidTraversal)>::run(derived(),other.derived()); #endif // EIGEN_TEST_EVALUATORS From b93520b1a52818c4659cafbb4d80aae001ece932 Mon Sep 17 00:00:00 2001 From: Leszek Swirski Date: Fri, 8 Nov 2013 14:07:11 +0000 Subject: [PATCH 0185/1103] Install functor folder with cmake --- Eigen/src/Core/CMakeLists.txt | 1 + Eigen/src/Core/functors/CMakeLists.txt | 6 ++++++ 2 files changed, 7 insertions(+) create mode 100644 Eigen/src/Core/functors/CMakeLists.txt diff --git a/Eigen/src/Core/CMakeLists.txt b/Eigen/src/Core/CMakeLists.txt index 2346fc2bb..38c3afde9 100644 --- a/Eigen/src/Core/CMakeLists.txt +++ b/Eigen/src/Core/CMakeLists.txt @@ -8,3 +8,4 @@ INSTALL(FILES ADD_SUBDIRECTORY(products) ADD_SUBDIRECTORY(util) ADD_SUBDIRECTORY(arch) +ADD_SUBDIRECTORY(functors) diff --git a/Eigen/src/Core/functors/CMakeLists.txt b/Eigen/src/Core/functors/CMakeLists.txt new file mode 100644 index 000000000..f4b99a9c3 --- /dev/null +++ b/Eigen/src/Core/functors/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_Core_Functor_SRCS "*.h") + +INSTALL(FILES + ${Eigen_Core_Functor_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/functors COMPONENT Devel + ) From 5c2d1b4710b5f1ae280bb687b3dd47dfeb62b698 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 10 Nov 2013 15:26:07 +0100 Subject: [PATCH 0186/1103] Add missing nonZeros() overload in Block> --- Eigen/src/SparseCore/SparseBlock.h | 22 ++++++++++++++++------ test/sparse_basic.cpp | 5 ++++- 2 files changed, 20 insertions(+), 7 deletions(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 38c7f1ff6..75b5e37a3 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -52,14 +52,24 @@ public: inline BlockImpl(const XprType& xpr, int i) : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) - {} + {std::cout << __LINE__ << "\n";} inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) - {} + {std::cout << __LINE__ << "\n";} EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } + + Index nonZeros() const + { + Index nnz = 0; + Index end = m_outerStart + m_outerSize.value(); + for(int j=m_outerStart; j inline BlockType& operator=(const SparseMatrixBase& other) @@ -307,13 +317,13 @@ public: m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0), m_blockRows(xpr.rows()), m_blockCols(xpr.cols()) - {} + {std::cout << __LINE__ << "\n";} /** Dynamic-size constructor */ inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) : m_matrix(xpr), m_startRow(startRow), m_startCol(startCol), m_blockRows(blockRows), m_blockCols(blockCols) - {} + {std::cout << __LINE__ << "\n";} inline int rows() const { return m_blockRows.value(); } inline int cols() const { return m_blockCols.value(); } diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index 498ecfe29..d5ff9f80b 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -198,6 +198,8 @@ template void sparse_basic(const SparseMatrixType& re if(j>0) VERIFY(j==numext::real(m3.innerVector(j).lastCoeff())); } + + VERIFY(m3.innerVector(j0).nonZeros() == m3.transpose().innerVector(j0).nonZeros()); //m2.innerVector(j0) = 2*m2.innerVector(j1); //refMat2.col(j0) = 2*refMat2.col(j1); @@ -227,6 +229,8 @@ template void sparse_basic(const SparseMatrixType& re VERIFY_IS_APPROX(m2, refMat2); + VERIFY(m2.innerVectors(j0,n0).nonZeros() == m2.transpose().innerVectors(j0,n0).nonZeros()); + m2.innerVectors(j0,n0) = m2.innerVectors(j0,n0) + m2.innerVectors(j1,n0); if(SparseMatrixType::IsRowMajor) refMat2.middleRows(j0,n0) = (refMat2.middleRows(j0,n0) + refMat2.middleRows(j1,n0)).eval(); @@ -234,7 +238,6 @@ template void sparse_basic(const SparseMatrixType& re refMat2.middleCols(j0,n0) = (refMat2.middleCols(j0,n0) + refMat2.middleCols(j1,n0)).eval(); VERIFY_IS_APPROX(m2, refMat2); - } // test basic computations From 8f2d068e8412a6aa446de974742b633dc3b2af76 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 10 Nov 2013 16:16:50 +0100 Subject: [PATCH 0187/1103] Use the specialization of Block for const matrices too --- Eigen/src/SparseCore/SparseBlock.h | 65 ++++++++++++++++++++++++------ 1 file changed, 52 insertions(+), 13 deletions(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 75b5e37a3..d0436aa7c 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -52,11 +52,11 @@ public: inline BlockImpl(const XprType& xpr, int i) : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) - {std::cout << __LINE__ << "\n";} + {} inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) - {std::cout << __LINE__ << "\n";} + {} EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } @@ -86,11 +86,12 @@ public: * specialization for SparseMatrix ***************************************************************************/ -template -class BlockImpl,BlockRows,BlockCols,true,Sparse> - : public SparseMatrixBase,BlockRows,BlockCols,true> > +namespace internal { + +template +class sparse_matrix_block_impl + : public SparseMatrixBase > { - typedef SparseMatrix<_Scalar, _Options, _Index> SparseMatrixType; typedef typename internal::remove_all::type _MatrixTypeNested; typedef Block BlockType; public: @@ -123,13 +124,13 @@ public: Index m_outer; }; - inline BlockImpl(const SparseMatrixType& xpr, int i) + inline sparse_matrix_block_impl(const SparseMatrixType& xpr, int i) : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) - {std::cout << __LINE__ << "\n";} + {} - inline BlockImpl(const SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) + inline sparse_matrix_block_impl(const SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) - {std::cout << __LINE__ << "\n";} + {} template inline BlockType& operator=(const SparseMatrixBase& other) @@ -237,7 +238,7 @@ public: const Scalar& lastCoeff() const { - EIGEN_STATIC_ASSERT_VECTOR_ONLY(BlockImpl); + EIGEN_STATIC_ASSERT_VECTOR_ONLY(sparse_matrix_block_impl); eigen_assert(nonZeros()>0); if(m_matrix.isCompressed()) return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart+1]-1]; @@ -256,6 +257,44 @@ public: }; +} // namespace internal + +template +class BlockImpl,BlockRows,BlockCols,true,Sparse> + : public internal::sparse_matrix_block_impl,BlockRows,BlockCols> +{ +public: + typedef SparseMatrix<_Scalar, _Options, _Index> SparseMatrixType; + typedef internal::sparse_matrix_block_impl Base; + inline BlockImpl(SparseMatrixType& xpr, int i) + : Base(xpr, i) + {} + + inline BlockImpl(SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) + : Base(xpr, startRow, startCol, blockRows, blockCols) + {} + + using Base::operator=; +}; + +template +class BlockImpl,BlockRows,BlockCols,true,Sparse> + : public internal::sparse_matrix_block_impl,BlockRows,BlockCols> +{ +public: + typedef const SparseMatrix<_Scalar, _Options, _Index> SparseMatrixType; + typedef internal::sparse_matrix_block_impl Base; + inline BlockImpl(SparseMatrixType& xpr, int i) + : Base(xpr, i) + {} + + inline BlockImpl(SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) + : Base(xpr, startRow, startCol, blockRows, blockCols) + {} + + using Base::operator=; +}; + //---------- /** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this @@ -317,13 +356,13 @@ public: m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0), m_blockRows(xpr.rows()), m_blockCols(xpr.cols()) - {std::cout << __LINE__ << "\n";} + {} /** Dynamic-size constructor */ inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) : m_matrix(xpr), m_startRow(startRow), m_startCol(startCol), m_blockRows(blockRows), m_blockCols(blockCols) - {std::cout << __LINE__ << "\n";} + {} inline int rows() const { return m_blockRows.value(); } inline int cols() const { return m_blockCols.value(); } From e59b38abef4dec8e37037284760607d3e108958f Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 13 Nov 2013 16:47:02 +0100 Subject: [PATCH 0188/1103] Implement boolean reductions for zero-sized objects --- Eigen/src/Core/BooleanRedux.h | 8 ++++---- test/zerosized.cpp | 31 ++++++++++++++++++++++++++++--- 2 files changed, 32 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h index 6e37e031a..be9f48a8c 100644 --- a/Eigen/src/Core/BooleanRedux.h +++ b/Eigen/src/Core/BooleanRedux.h @@ -29,9 +29,9 @@ struct all_unroller }; template -struct all_unroller +struct all_unroller { - static inline bool run(const Derived &mat) { return mat.coeff(0, 0); } + static inline bool run(const Derived &/*mat*/) { return true; } }; template @@ -55,9 +55,9 @@ struct any_unroller }; template -struct any_unroller +struct any_unroller { - static inline bool run(const Derived &mat) { return mat.coeff(0, 0); } + static inline bool run(const Derived & /*mat*/) { return false; } }; template diff --git a/test/zerosized.cpp b/test/zerosized.cpp index 6905e584e..da7dd0481 100644 --- a/test/zerosized.cpp +++ b/test/zerosized.cpp @@ -9,12 +9,26 @@ #include "main.h" + +template void zeroReduction(const MatrixType& m) { + // Reductions that must hold for zero sized objects + VERIFY(m.all()); + VERIFY(!m.any()); + VERIFY(m.prod()==1); + VERIFY(m.sum()==0); + VERIFY(m.count()==0); + VERIFY(m.allFinite()); + VERIFY(!m.hasNaN()); +} + + template void zeroSizedMatrix() { MatrixType t1; - if (MatrixType::SizeAtCompileTime == Dynamic) + if (MatrixType::SizeAtCompileTime == Dynamic || MatrixType::SizeAtCompileTime == 0) { + zeroReduction(t1); if (MatrixType::RowsAtCompileTime == Dynamic) VERIFY(t1.rows() == 0); if (MatrixType::ColsAtCompileTime == Dynamic) @@ -22,9 +36,13 @@ template void zeroSizedMatrix() if (MatrixType::RowsAtCompileTime == Dynamic && MatrixType::ColsAtCompileTime == Dynamic) { + MatrixType t2(0, 0); VERIFY(t2.rows() == 0); VERIFY(t2.cols() == 0); + + zeroReduction(t2); + VERIFY(t1==t2); } } } @@ -33,11 +51,15 @@ template void zeroSizedVector() { VectorType t1; - if (VectorType::SizeAtCompileTime == Dynamic) + if (VectorType::SizeAtCompileTime == Dynamic || VectorType::SizeAtCompileTime==0) { + zeroReduction(t1); VERIFY(t1.size() == 0); VectorType t2(DenseIndex(0)); // DenseIndex disambiguates with 0-the-null-pointer (error with gcc 4.4 and MSVC8) VERIFY(t2.size() == 0); + zeroReduction(t2); + + VERIFY(t1==t2); } } @@ -51,9 +73,12 @@ void test_zerosized() zeroSizedMatrix >(); zeroSizedMatrix >(); zeroSizedMatrix >(); - + zeroSizedMatrix >(); + zeroSizedMatrix >(); + zeroSizedVector(); zeroSizedVector(); zeroSizedVector(); zeroSizedVector >(); + zeroSizedVector >(); } From 5e28c41549c5e71011bb9312104d1d30e919e53f Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Thu, 14 Nov 2013 22:27:06 +0100 Subject: [PATCH 0189/1103] C++11: add template metaprogramming helpers Create a new directory CXX11 under unsupported/Eigen that contains code that requires C++11. In that directory, add a few generic templates useful for any module relying on C++11. These templates may be included with #include <[unsupported/]Eigen/CXX11/Core>. At the moment, this will only provide templates in the Eigen::internal namespace. --- unsupported/Eigen/CXX11/Core | 41 ++ .../Eigen/CXX11/src/Core/util/CXX11Meta.h | 504 ++++++++++++++++++ .../CXX11/src/Core/util/CXX11Workarounds.h | 104 ++++ unsupported/test/CMakeLists.txt | 8 + unsupported/test/cxx11_meta.cpp | 343 ++++++++++++ 5 files changed, 1000 insertions(+) create mode 100644 unsupported/Eigen/CXX11/Core create mode 100644 unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h create mode 100644 unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h create mode 100644 unsupported/test/cxx11_meta.cpp diff --git a/unsupported/Eigen/CXX11/Core b/unsupported/Eigen/CXX11/Core new file mode 100644 index 000000000..1b1f1dcc1 --- /dev/null +++ b/unsupported/Eigen/CXX11/Core @@ -0,0 +1,41 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_CORE_MODULE +#define EIGEN_CXX11_CORE_MODULE + +#include + +#include + +/** \defgroup CXX11_Core_Module C++11 Core Module + * + * This module provides common core features for all modules that + * explicitly depend on C++11. Currently, this is only the Tensor + * module. Note that at this stage, you should not need to include + * this module directly. + * + * \code + * #include + * \endcode + */ + +#include + +#include "src/Core/util/CXX11Workarounds.h" +#include "src/Core/util/CXX11Meta.h" + +#include + +#endif // EIGEN_CXX11_CORE_MODULE + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h new file mode 100644 index 000000000..55d3cc37b --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -0,0 +1,504 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11META_H +#define EIGEN_CXX11META_H + +namespace Eigen { + +namespace internal { + +/** \internal + * \file CXX11/Core/util/CXX11Meta.h + * This file contains generic metaprogramming classes which are not specifically related to Eigen. + * This file expands upon Core/util/Meta.h and adds support for C++11 specific features. + */ + +template +struct type_list { constexpr static int count = sizeof...(tt); }; + +template +struct type_list { constexpr static int count = sizeof...(tt) + 1; typedef t first_type; }; + +template +struct numeric_list { constexpr static std::size_t count = sizeof...(nn); }; + +template +struct numeric_list { constexpr static std::size_t count = sizeof...(nn) + 1; constexpr static T first_value = n; }; + +/* numeric list constructors + * + * equivalencies: + * constructor result + * typename gen_numeric_list::type numeric_list + * typename gen_numeric_list_reversed::type numeric_list + * typename gen_numeric_list_swapped_pair::type numeric_list + * typename gen_numeric_list_repeated::type numeric_list + */ + +template struct gen_numeric_list : gen_numeric_list {}; +template struct gen_numeric_list { typedef numeric_list type; }; + +template struct gen_numeric_list_reversed : gen_numeric_list_reversed {}; +template struct gen_numeric_list_reversed { typedef numeric_list type; }; + +template struct gen_numeric_list_swapped_pair : gen_numeric_list_swapped_pair {}; +template struct gen_numeric_list_swapped_pair { typedef numeric_list type; }; + +template struct gen_numeric_list_repeated : gen_numeric_list_repeated {}; +template struct gen_numeric_list_repeated { typedef numeric_list type; }; + +/* list manipulation: concatenate */ + +template struct concat; + +template struct concat, type_list> { typedef type_list type; }; +template struct concat, numeric_list > { typedef numeric_list type; }; + +template struct mconcat; +template struct mconcat { typedef a type; }; +template struct mconcat : concat {}; +template struct mconcat : concat::type> {}; + +/* list manipulation: extract slices */ + +template struct take; +template struct take> : concat, typename take>::type> {}; +template struct take> { typedef type_list<> type; }; +template struct take<0, type_list> { typedef type_list<> type; }; +template<> struct take<0, type_list<>> { typedef type_list<> type; }; + +template struct take> : concat, typename take>::type> {}; +template struct take> { typedef numeric_list type; }; +template struct take<0, numeric_list> { typedef numeric_list type; }; +template struct take<0, numeric_list> { typedef numeric_list type; }; + +template struct h_skip_helper_numeric; +template struct h_skip_helper_numeric : h_skip_helper_numeric {}; +template struct h_skip_helper_numeric { typedef numeric_list type; }; +template struct h_skip_helper_numeric { typedef numeric_list type; }; +template struct h_skip_helper_numeric { typedef numeric_list type; }; + +template struct h_skip_helper_type; +template struct h_skip_helper_type : h_skip_helper_type {}; +template struct h_skip_helper_type<0, t, tt...> { typedef type_list type; }; +template struct h_skip_helper_type { typedef type_list<> type; }; +template<> struct h_skip_helper_type<0> { typedef type_list<> type; }; + +template +struct h_skip { + template + constexpr static inline typename h_skip_helper_numeric::type helper(numeric_list) { return typename h_skip_helper_numeric::type(); } + template + constexpr static inline typename h_skip_helper_type::type helper(type_list) { return typename h_skip_helper_type::type(); } +}; + +template struct skip { typedef decltype(h_skip::helper(a())) type; }; + +template struct slice : take::type> {}; + +/* list manipulation: retrieve single element from list */ + +template struct get; + +template struct get> : get> {}; +template struct get<0, type_list> { typedef a type; }; +template struct get> { static_assert((n - n) < 0, "meta-template get: The element to extract from a list must be smaller than the size of the list."); }; + +template struct get> : get> {}; +template struct get<0, numeric_list> { constexpr static int value = a; }; +template struct get> { static_assert((n - n) < 0, "meta-template get: The element to extract from a list must be smaller than the size of the list."); }; + +/* always get type, regardless of dummy; good for parameter pack expansion */ + +template struct id_numeric { typedef t type; }; +template struct id_type { typedef t type; }; + +/* equality checking, flagged version */ + +template struct is_same_gf : is_same { constexpr static int global_flags = 0; }; + +/* apply_op to list */ + +template< + bool from_left, // false + template class op, + typename additional_param, + typename... values +> +struct h_apply_op_helper { typedef type_list::type...> type; }; +template< + template class op, + typename additional_param, + typename... values +> +struct h_apply_op_helper { typedef type_list::type...> type; }; + +template< + bool from_left, + template class op, + typename additional_param +> +struct h_apply_op +{ + template + constexpr static typename h_apply_op_helper::type helper(type_list) + { return typename h_apply_op_helper::type(); } +}; + +template< + template class op, + typename additional_param, + typename a +> +struct apply_op_from_left { typedef decltype(h_apply_op::helper(a())) type; }; + +template< + template class op, + typename additional_param, + typename a +> +struct apply_op_from_right { typedef decltype(h_apply_op::helper(a())) type; }; + +/* see if an element is in a list */ + +template< + template class test, + typename check_against, + typename h_list, + bool last_check_positive = false +> +struct contained_in_list; + +template< + template class test, + typename check_against, + typename h_list +> +struct contained_in_list +{ + constexpr static bool value = true; +}; + +template< + template class test, + typename check_against, + typename a, + typename... as +> +struct contained_in_list, false> : contained_in_list, test::value> {}; + +template< + template class test, + typename check_against + EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, empty) +> +struct contained_in_list, false> { constexpr static bool value = false; }; + +/* see if an element is in a list and check for global flags */ + +template< + template class test, + typename check_against, + typename h_list, + int default_flags = 0, + bool last_check_positive = false, + int last_check_flags = default_flags +> +struct contained_in_list_gf; + +template< + template class test, + typename check_against, + typename h_list, + int default_flags, + int last_check_flags +> +struct contained_in_list_gf +{ + constexpr static bool value = true; + constexpr static int global_flags = last_check_flags; +}; + +template< + template class test, + typename check_against, + typename a, + typename... as, + int default_flags, + int last_check_flags +> +struct contained_in_list_gf, default_flags, false, last_check_flags> : contained_in_list_gf, default_flags, test::value, test::global_flags> {}; + +template< + template class test, + typename check_against + EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, empty), + int default_flags, + int last_check_flags +> +struct contained_in_list_gf, default_flags, false, last_check_flags> { constexpr static bool value = false; constexpr static int global_flags = default_flags; }; + +/* generic reductions */ + +template< + typename Reducer, + typename... Ts +> struct reduce; + +template< + typename Reducer, + typename A, + typename... Ts +> struct reduce +{ + constexpr static inline A run(A a, Ts...) { return a; } +}; + +template< + typename Reducer, + typename A, + typename B, + typename... Ts +> struct reduce +{ + constexpr static inline auto run(A a, B b, Ts... ts) -> decltype(Reducer::run(a, reduce::run(b, ts...))) { + return Reducer::run(a, reduce::run(b, ts...)); + } +}; + +/* generic binary operations */ + +struct sum_op { template constexpr static inline auto run(A a, B b) -> decltype(a + b) { return a + b; } }; +struct product_op { template constexpr static inline auto run(A a, B b) -> decltype(a * b) { return a * b; } }; + +struct logical_and_op { template constexpr static inline auto run(A a, B b) -> decltype(a && b) { return a && b; } }; +struct logical_or_op { template constexpr static inline auto run(A a, B b) -> decltype(a || b) { return a || b; } }; + +struct equal_op { template constexpr static inline auto run(A a, B b) -> decltype(a == b) { return a == b; } }; +struct not_equal_op { template constexpr static inline auto run(A a, B b) -> decltype(a != b) { return a != b; } }; +struct lesser_op { template constexpr static inline auto run(A a, B b) -> decltype(a < b) { return a < b; } }; +struct lesser_equal_op { template constexpr static inline auto run(A a, B b) -> decltype(a <= b) { return a <= b; } }; +struct greater_op { template constexpr static inline auto run(A a, B b) -> decltype(a < b) { return a < b; } }; +struct greater_equal_op { template constexpr static inline auto run(A a, B b) -> decltype(a >= b) { return a >= b; } }; + +/* generic unary operations */ + +struct not_op { template constexpr static inline auto run(A a) -> decltype(!a) { return !a; } }; +struct negation_op { template constexpr static inline auto run(A a) -> decltype(-a) { return -a; } }; +struct greater_equal_zero_op { template constexpr static inline auto run(A a) -> decltype(a >= 0) { return a >= 0; } }; + + +/* reductions for lists */ + +// using auto -> return value spec makes ICC 13.0 and 13.1 crash here, so we have to hack it +// together in front... (13.0 doesn't work with array_prod/array_reduce/... anyway, but 13.1 +// does... +template +constexpr inline decltype(reduce::run((*((Ts*)0))...)) arg_prod(Ts... ts) +{ + return reduce::run(ts...); +} + +template +constexpr inline decltype(reduce::run((*((Ts*)0))...)) arg_sum(Ts... ts) +{ + return reduce::run(ts...); +} + +/* reverse arrays */ + +template +constexpr inline Array h_array_reverse(Array arr, numeric_list) +{ + return {{std_array_get(arr)...}}; +} + +template +constexpr inline std::array array_reverse(std::array arr) +{ + return h_array_reverse(arr, typename gen_numeric_list::type()); +} + +/* generic array reductions */ + +// can't reuse standard reduce() interface above because Intel's Compiler +// *really* doesn't like it, so we just reimplement the stuff +// (start from N - 1 and work down to 0 because specialization for +// n == N - 1 also doesn't work in Intel's compiler, so it goes into +// an infinite loop) +template +struct h_array_reduce { + constexpr static inline auto run(std::array arr) -> decltype(Reducer::run(h_array_reduce::run(arr), std_array_get(arr))) + { + return Reducer::run(h_array_reduce::run(arr), std_array_get(arr)); + } +}; + +template +struct h_array_reduce +{ + constexpr static inline T run(std::array arr) + { + return std_array_get<0>(arr); + } +}; + +template +constexpr inline auto array_reduce(std::array arr) -> decltype(h_array_reduce::run(arr)) +{ + return h_array_reduce::run(arr); +} + +/* standard array reductions */ + +template +constexpr inline auto array_sum(std::array arr) -> decltype(array_reduce(arr)) +{ + return array_reduce(arr); +} + +template +constexpr inline auto array_prod(std::array arr) -> decltype(array_reduce(arr)) +{ + return array_reduce(arr); +} + +/* zip an array */ + +template +constexpr inline std::array h_array_zip(std::array a, std::array b, numeric_list) +{ + return std::array{{ Op::run(std_array_get(a), std_array_get(b))... }}; +} + +template +constexpr inline std::array array_zip(std::array a, std::array b) +{ + return h_array_zip(a, b, typename gen_numeric_list::type()); +} + +/* zip an array and reduce the result */ + +template +constexpr inline auto h_array_zip_and_reduce(std::array a, std::array b, numeric_list) -> decltype(reduce::type...>::run(Op::run(std_array_get(a), std_array_get(b))...)) +{ + return reduce::type...>::run(Op::run(std_array_get(a), std_array_get(b))...); +} + +template +constexpr inline auto array_zip_and_reduce(std::array a, std::array b) -> decltype(h_array_zip_and_reduce(a, b, typename gen_numeric_list::type())) +{ + return h_array_zip_and_reduce(a, b, typename gen_numeric_list::type()); +} + +/* apply stuff to an array */ + +template +constexpr inline std::array h_array_apply(std::array a, numeric_list) +{ + return std::array{{ Op::run(std_array_get(a))... }}; +} + +template +constexpr inline std::array array_apply(std::array a) +{ + return h_array_apply(a, typename gen_numeric_list::type()); +} + +/* apply stuff to an array and reduce */ + +template +constexpr inline auto h_array_apply_and_reduce(std::array arr, numeric_list) -> decltype(reduce::type...>::run(Op::run(std_array_get(arr))...)) +{ + return reduce::type...>::run(Op::run(std_array_get(arr))...); +} + +template +constexpr inline auto array_apply_and_reduce(std::array a) -> decltype(h_array_apply_and_reduce(a, typename gen_numeric_list::type())) +{ + return h_array_apply_and_reduce(a, typename gen_numeric_list::type()); +} + +/* repeat a value n times (and make an array out of it + * usage: + * std::array = repeat<16>(42); + */ + +template +struct h_repeat +{ + template + constexpr static inline std::array run(t v, numeric_list) + { + return {{ typename id_numeric::type(v)... }}; + } +}; + +template +constexpr std::array repeat(t v) { return h_repeat::run(v, typename gen_numeric_list::type()); } + +/* instantiate a class by a C-style array */ +template +struct h_instantiate_by_c_array; + +template +struct h_instantiate_by_c_array +{ + static InstType run(ArrType* arr, Ps... args) + { + return h_instantiate_by_c_array::run(arr + 1, args..., arr[0]); + } +}; + +template +struct h_instantiate_by_c_array +{ + static InstType run(ArrType* arr, Ps... args) + { + return h_instantiate_by_c_array::run(arr + 1, arr[0], args...); + } +}; + +template +struct h_instantiate_by_c_array +{ + static InstType run(ArrType* arr, Ps... args) + { + (void)arr; + return InstType(args...); + } +}; + +template +struct h_instantiate_by_c_array +{ + static InstType run(ArrType* arr, Ps... args) + { + (void)arr; + return InstType(args...); + } +}; + +template +InstType instantiate_by_c_array(ArrType* arr) +{ + return h_instantiate_by_c_array::run(arr); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_CXX11META_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h new file mode 100644 index 000000000..eec38b343 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -0,0 +1,104 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11WORKAROUNDS_H +#define EIGEN_CXX11WORKAROUNDS_H + +/* COMPATIBILITY CHECKS + * (so users of compilers that are too old get some realistic error messages) + */ +#if defined(__INTEL_COMPILER) && (__INTEL_COMPILER < 1310) +#error Intel Compiler only supports required C++ features since version 13.1. +// note that most stuff in principle works with 13.0 but when combining +// some features, at some point 13.0 will just fail with an internal assertion +#elif defined(__clang__) && (__clang_major__ < 3 || (__clang_major__ == 3 && __clang_minor__ < 1)) +// note that it _should_ work with 3.1 but it was only tested with 3.2 +#error Clang C++ Compiler (clang++) only supports required C++ features since version 3.1. +#elif defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && (__GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 6)) +// G++ < 4.6 by default will continue processing the source files - even if we use #error to make +// it error out. For this reason, we use the pragma to make sure G++ aborts at the first error +// it sees. Unfortunately, that is still not our #error directive, but at least the output is +// short enough the user has a chance to see that the compiler version is not sufficient for +// the funky template mojo we use. +#pragma GCC diagnostic error "-Wfatal-errors" +#error GNU C++ Compiler (g++) only supports required C++ features since version 4.6. +#endif + +/* Check that the compiler at least claims to support C++11. It might not be sufficient + * because the compiler may not implement it correctly, but at least we'll know. + */ +#if __cplusplus <= 199711L +#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) +#pragma GCC diagnostic error "-Wfatal-errors" +#endif +#error This library needs at least a C++11 compliant compiler. If you use g++/clang, please enable the -std=c++11 compiler flag. (-std=c++0x on older versions.) +#endif + +namespace Eigen { + +namespace internal { + +/* std::get is only constexpr in C++14, not yet in C++11 + * - libstdc++ from version 4.7 onwards has it nevertheless, + * so use that + * - libstdc++ older versions: use _M_instance directly + * - libc++ all versions so far: use __elems_ directly + * - all other libs: use std::get to be portable, but + * this may not be constexpr + */ +#if defined(__GLIBCXX__) && __GLIBCXX__ < 20120322 +#define STD_GET_ARR_HACK a._M_instance[I] +#elif defined(_LIBCPP_VERSION) +#define STD_GET_ARR_HACK a.__elems_[I] +#else +#define STD_GET_ARR_HACK std::template get(a) +#endif + +template constexpr inline T& std_array_get(std::array& a) { return (T&) STD_GET_ARR_HACK; } +template constexpr inline T&& std_array_get(std::array&& a) { return (T&&) STD_GET_ARR_HACK; } +template constexpr inline T const& std_array_get(std::array const& a) { return (T const&) STD_GET_ARR_HACK; } + +#undef STD_GET_ARR_HACK + +/* Suppose you have a template of the form + * template struct X; + * And you want to specialize it in such a way: + * template struct X> { ::: }; + * template<> struct X> { ::: }; + * This will work in Intel's compiler 13.0, but only to some extent in g++ 4.6, since + * g++ can only match templates called with parameter packs if the number of template + * arguments is not a fixed size (so inside the first specialization, referencing + * X> will fail in g++). On the other hand, g++ will accept the following: + * template struct X> { ::: }: + * as an additional (!) specialization, which will then only match the empty case. + * But Intel's compiler 13.0 won't accept that, it will only accept the empty syntax, + * so we have to create a workaround for this. + */ +#if defined(__GNUC__) && !defined(__INTEL_COMPILER) +#define EIGEN_TPL_PP_SPEC_HACK_DEF(mt, n) mt... n +#define EIGEN_TPL_PP_SPEC_HACK_DEFC(mt, n) , EIGEN_TPL_PP_SPEC_HACK_DEF(mt, n) +#define EIGEN_TPL_PP_SPEC_HACK_USE(n) n... +#define EIGEN_TPL_PP_SPEC_HACK_USEC(n) , n... +#else +#define EIGEN_TPL_PP_SPEC_HACK_DEF(mt, n) +#define EIGEN_TPL_PP_SPEC_HACK_DEFC(mt, n) +#define EIGEN_TPL_PP_SPEC_HACK_USE(n) +#define EIGEN_TPL_PP_SPEC_HACK_USEC(n) +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_CXX11WORKAROUNDS_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 38c89a234..841cab9d7 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -92,3 +92,11 @@ ei_add_test(gmres) ei_add_test(minres) ei_add_test(levenberg_marquardt) ei_add_test(bdcsvd) + +option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." OFF) +if(EIGEN_TEST_CXX11) + # FIXME: add C++11 compiler switch in some portable way + # (MSVC doesn't need any for example, so this will + # clash there) + ei_add_test(cxx11_meta "-std=c++0x") +endif() diff --git a/unsupported/test/cxx11_meta.cpp b/unsupported/test/cxx11_meta.cpp new file mode 100644 index 000000000..03a9efe54 --- /dev/null +++ b/unsupported/test/cxx11_meta.cpp @@ -0,0 +1,343 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::internal::is_same; +using Eigen::internal::type_list; +using Eigen::internal::numeric_list; +using Eigen::internal::gen_numeric_list; +using Eigen::internal::gen_numeric_list_reversed; +using Eigen::internal::gen_numeric_list_swapped_pair; +using Eigen::internal::gen_numeric_list_repeated; +using Eigen::internal::concat; +using Eigen::internal::mconcat; +using Eigen::internal::take; +using Eigen::internal::skip; +using Eigen::internal::slice; +using Eigen::internal::get; +using Eigen::internal::id_numeric; +using Eigen::internal::id_type; +using Eigen::internal::is_same_gf; +using Eigen::internal::apply_op_from_left; +using Eigen::internal::apply_op_from_right; +using Eigen::internal::contained_in_list; +using Eigen::internal::contained_in_list_gf; +using Eigen::internal::arg_prod; +using Eigen::internal::arg_sum; +using Eigen::internal::sum_op; +using Eigen::internal::product_op; +using Eigen::internal::array_reverse; +using Eigen::internal::array_sum; +using Eigen::internal::array_prod; +using Eigen::internal::array_reduce; +using Eigen::internal::array_zip; +using Eigen::internal::array_zip_and_reduce; +using Eigen::internal::array_apply; +using Eigen::internal::array_apply_and_reduce; +using Eigen::internal::repeat; +using Eigen::internal::instantiate_by_c_array; + +struct dummy_a {}; +struct dummy_b {}; +struct dummy_c {}; +struct dummy_d {}; +struct dummy_e {}; + +// dummy operation for testing apply +template struct dummy_op; +template<> struct dummy_op { typedef dummy_c type; }; +template<> struct dummy_op { typedef dummy_d type; }; +template<> struct dummy_op { typedef dummy_a type; }; +template<> struct dummy_op { typedef dummy_d type; }; +template<> struct dummy_op { typedef dummy_b type; }; +template<> struct dummy_op { typedef dummy_d type; }; +template<> struct dummy_op { typedef dummy_e type; }; +template<> struct dummy_op { typedef dummy_e type; }; +template<> struct dummy_op { typedef dummy_e type; }; + +template struct dummy_test { constexpr static bool value = false; constexpr static int global_flags = 0; }; +template<> struct dummy_test { constexpr static bool value = true; constexpr static int global_flags = 1; }; +template<> struct dummy_test { constexpr static bool value = true; constexpr static int global_flags = 2; }; +template<> struct dummy_test { constexpr static bool value = true; constexpr static int global_flags = 4; }; + +struct times2_op { template static A run(A v) { return v * 2; } }; + +struct dummy_inst +{ + int c; + + dummy_inst() : c(0) {} + explicit dummy_inst(int) : c(1) {} + dummy_inst(int, int) : c(2) {} + dummy_inst(int, int, int) : c(3) {} + dummy_inst(int, int, int, int) : c(4) {} + dummy_inst(int, int, int, int, int) : c(5) {} +}; + +static void test_gen_numeric_list() +{ + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); +} + +static void test_concat() +{ + VERIFY((is_same, type_list<>>::type, type_list>::value)); + VERIFY((is_same, type_list>::type, type_list>::value)); + VERIFY((is_same, type_list>::type, type_list>::value)); + VERIFY((is_same, type_list>::type, type_list>::value)); + VERIFY((is_same, type_list>::type, type_list>::value)); + + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); + + VERIFY((is_same>::type, type_list>::value)); + VERIFY((is_same, type_list>::type, type_list>::value)); + VERIFY((is_same, type_list, type_list>::type, type_list>::value)); + VERIFY((is_same, type_list>::type, type_list>::value)); + VERIFY((is_same, type_list>::type, type_list>::value)); + + VERIFY((is_same>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list, numeric_list>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); + VERIFY((is_same, numeric_list>::type, numeric_list>::value)); +} + +static void test_slice() +{ + typedef type_list tl; + typedef numeric_list il; + + VERIFY((is_same::type, type_list<>>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, type_list<>>::value)); + + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + + VERIFY((is_same::type, typename take<3, tl>::type>::value)); + VERIFY((is_same::type, typename take<3, il>::type>::value)); + VERIFY((is_same::type, type_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); +} + +static void test_get() +{ + typedef type_list tl; + typedef numeric_list il; + + VERIFY((is_same::type, dummy_a>::value)); + VERIFY((is_same::type, dummy_a>::value)); + VERIFY((is_same::type, dummy_b>::value)); + VERIFY((is_same::type, dummy_b>::value)); + VERIFY((is_same::type, dummy_c>::value)); + VERIFY((is_same::type, dummy_c>::value)); + + VERIFY_IS_EQUAL(((int)get<0, il>::value), 4); + VERIFY_IS_EQUAL(((int)get<1, il>::value), 8); + VERIFY_IS_EQUAL(((int)get<2, il>::value), 15); + VERIFY_IS_EQUAL(((int)get<3, il>::value), 16); + VERIFY_IS_EQUAL(((int)get<4, il>::value), 23); + VERIFY_IS_EQUAL(((int)get<5, il>::value), 42); +} + +static void test_id_helper(dummy_a a, dummy_a b, dummy_a c) +{ + (void)a; + (void)b; + (void)c; +} + +template +static void test_id_numeric() +{ + test_id_helper(typename id_numeric::type()...); +} + +template +static void test_id_type() +{ + test_id_helper(typename id_type::type()...); +} + +static void test_id() +{ + // don't call VERIFY here, just assume it works if it compiles + // (otherwise it will complain that it can't find the function) + test_id_numeric<1, 4, 6>(); + test_id_type(); +} + +static void test_is_same_gf() +{ + VERIFY((!is_same_gf::value)); + VERIFY((!!is_same_gf::value)); + VERIFY_IS_EQUAL((!!is_same_gf::global_flags), 0); + VERIFY_IS_EQUAL((!!is_same_gf::global_flags), 0); +} + +static void test_apply_op() +{ + typedef type_list tl; + VERIFY((!!is_same::type, type_list>::value)); + VERIFY((!!is_same::type, type_list>::value)); +} + +static void test_contained_in_list() +{ + typedef type_list tl; + + VERIFY((!!contained_in_list::value)); + VERIFY((!!contained_in_list::value)); + VERIFY((!!contained_in_list::value)); + VERIFY((!contained_in_list::value)); + VERIFY((!contained_in_list::value)); + + VERIFY((!!contained_in_list_gf::value)); + VERIFY((!!contained_in_list_gf::value)); + VERIFY((!!contained_in_list_gf::value)); + VERIFY((!contained_in_list_gf::value)); + VERIFY((!contained_in_list_gf::value)); + + VERIFY_IS_EQUAL(((int)contained_in_list_gf::global_flags), 1); + VERIFY_IS_EQUAL(((int)contained_in_list_gf::global_flags), 2); + VERIFY_IS_EQUAL(((int)contained_in_list_gf::global_flags), 4); + VERIFY_IS_EQUAL(((int)contained_in_list_gf::global_flags), 0); + VERIFY_IS_EQUAL(((int)contained_in_list_gf::global_flags), 0); +} + +static void test_arg_reductions() +{ + VERIFY_IS_EQUAL(arg_sum(1,2,3,4), 10); + VERIFY_IS_EQUAL(arg_prod(1,2,3,4), 24); + VERIFY_IS_APPROX(arg_sum(0.5, 2, 5), 7.5); + VERIFY_IS_APPROX(arg_prod(0.5, 2, 5), 5.0); +} + +static void test_array_reverse_and_reduce() +{ + std::array a{{4, 8, 15, 16, 23, 42}}; + std::array b{{42, 23, 16, 15, 8, 4}}; + + // there is no operator<< for std::array, so VERIFY_IS_EQUAL will + // not compile + VERIFY((array_reverse(a) == b)); + VERIFY((array_reverse(b) == a)); + VERIFY_IS_EQUAL((array_sum(a)), 108); + VERIFY_IS_EQUAL((array_sum(b)), 108); + VERIFY_IS_EQUAL((array_prod(a)), 7418880); + VERIFY_IS_EQUAL((array_prod(b)), 7418880); +} + +static void test_array_zip_and_apply() +{ + std::array a{{4, 8, 15, 16, 23, 42}}; + std::array b{{0, 1, 2, 3, 4, 5}}; + std::array c{{4, 9, 17, 19, 27, 47}}; + std::array d{{0, 8, 30, 48, 92, 210}}; + std::array e{{0, 2, 4, 6, 8, 10}}; + + VERIFY((array_zip(a, b) == c)); + VERIFY((array_zip(a, b) == d)); + VERIFY((array_apply(b) == e)); + VERIFY_IS_EQUAL((array_apply_and_reduce(a)), 216); + VERIFY_IS_EQUAL((array_apply_and_reduce(b)), 30); + VERIFY_IS_EQUAL((array_zip_and_reduce(a, b)), 14755932); + VERIFY_IS_EQUAL((array_zip_and_reduce(a, b)), 388); +} + +static void test_array_misc() +{ + std::array a3{{1, 1, 1}}; + std::array a6{{2, 2, 2, 2, 2, 2}}; + VERIFY((repeat<3, int>(1) == a3)); + VERIFY((repeat<6, int>(2) == a6)); + + int data[5] = { 0, 1, 2, 3, 4 }; + VERIFY_IS_EQUAL((instantiate_by_c_array(data).c), 0); + VERIFY_IS_EQUAL((instantiate_by_c_array(data).c), 1); + VERIFY_IS_EQUAL((instantiate_by_c_array(data).c), 2); + VERIFY_IS_EQUAL((instantiate_by_c_array(data).c), 3); + VERIFY_IS_EQUAL((instantiate_by_c_array(data).c), 4); + VERIFY_IS_EQUAL((instantiate_by_c_array(data).c), 5); +} + +void test_cxx11_meta() +{ + CALL_SUBTEST(test_gen_numeric_list()); + CALL_SUBTEST(test_concat()); + CALL_SUBTEST(test_slice()); + CALL_SUBTEST(test_get()); + CALL_SUBTEST(test_id()); + CALL_SUBTEST(test_is_same_gf()); + CALL_SUBTEST(test_apply_op()); + CALL_SUBTEST(test_contained_in_list()); + CALL_SUBTEST(test_arg_reductions()); + CALL_SUBTEST(test_array_reverse_and_reduce()); + CALL_SUBTEST(test_array_zip_and_apply()); + CALL_SUBTEST(test_array_misc()); +} + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ From f97b3cd0249228820807229a2d529260522ba8c7 Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Thu, 14 Nov 2013 22:52:37 +0100 Subject: [PATCH 0190/1103] CXX11/Tensor: add simple initial tensor implementation This commit adds an initial implementation of a class template Tensor that allows for the storage of objects with more than two indices. Currently, only storing data and setting the object to zero for POD data types are implemented. --- unsupported/Eigen/CXX11/Tensor | 40 +++ unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 314 ++++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorStorage.h | 126 +++++++ unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_simple.cpp | 271 +++++++++++++++ 5 files changed, 752 insertions(+) create mode 100644 unsupported/Eigen/CXX11/Tensor create mode 100644 unsupported/Eigen/CXX11/src/Tensor/Tensor.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h create mode 100644 unsupported/test/cxx11_tensor_simple.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor new file mode 100644 index 000000000..083d8c0a7 --- /dev/null +++ b/unsupported/Eigen/CXX11/Tensor @@ -0,0 +1,40 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_MODULE +#define EIGEN_CXX11_TENSOR_MODULE + +#include + +#include + +/** \defgroup CXX11_Tensor_Module Tensor Module + * + * This module provides a Tensor class for storing arbitrarily indexed + * objects. + * + * \code + * #include + * \endcode + */ + +#include +#include + +#include "src/Tensor/TensorStorage.h" +#include "src/Tensor/Tensor.h" + +#include + +#endif // EIGEN_CXX11_TENSOR_MODULE + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h new file mode 100644 index 000000000..c40905af4 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -0,0 +1,314 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_H +#define EIGEN_CXX11_TENSOR_TENSOR_H + +namespace Eigen { + +/** \class Tensor + * \ingroup CXX11_Tensor_Module + * + * \brief The tensor class. + * + * The %Tensor class is the work-horse for all \em dense tensors within Eigen. + * + * The %Tensor class encompasses only dynamic-size objects so far. + * + * The first two template parameters are required: + * \tparam Scalar_ \anchor tensor_tparam_scalar Numeric type, e.g. float, double, int or std::complex. + * User defined scalar types are supported as well (see \ref user_defined_scalars "here"). + * \tparam NumIndices_ Number of indices (i.e. rank of the tensor) + * + * The remaining template parameters are optional -- in most cases you don't have to worry about them. + * \tparam Options_ \anchor tensor_tparam_options A combination of either \b #RowMajor or \b #ColMajor, and of either + * \b #AutoAlign or \b #DontAlign. + * The former controls \ref TopicStorageOrders "storage order", and defaults to column-major. The latter controls alignment, which is required + * for vectorization. It defaults to aligning tensors. Note that tensors currently do not support any operations that profit from vectorization. + * Support for such operations (i.e. adding two tensors etc.) is planned. + * + * You can access elements of tensors using normal subscripting: + * + * \code + * Eigen::Tensor t(10, 10, 10, 10); + * t(0, 1, 2, 3) = 42.0; + * \endcode + * + * This class can be extended with the help of the plugin mechanism described on the page + * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_TENSOR_PLUGIN. + * + * Some notes: + * + *
+ *
Relation to other parts of Eigen:
+ *
The midterm developement goal for this class is to have a similar hierarchy as Eigen uses for matrices, so that + * taking blocks or using tensors in expressions is easily possible, including an interface with the vector/matrix code + * by providing .asMatrix() and .asVector() (or similar) methods for rank 2 and 1 tensors. However, currently, the %Tensor + * class does not provide any of these features and is only available as a stand-alone class that just allows for + * coefficient access. Also, when fixed-size tensors are implemented, the number of template arguments is likely to + * change dramatically.
+ *
+ * + * \ref TopicStorageOrders + */ +template +class Tensor; + +namespace internal { +template +struct traits> +{ + typedef Scalar_ Scalar; + typedef Dense StorageKind; + typedef DenseIndex Index; + enum { + Options = Options_ + }; +}; + +template +struct tensor_index_linearization_helper +{ + constexpr static inline Index run(std::array const& indices, std::array const& dimensions) + { + return std_array_get(indices) + + std_array_get(dimensions) * + tensor_index_linearization_helper::run(indices, dimensions); + } +}; + +template +struct tensor_index_linearization_helper +{ + constexpr static inline Index run(std::array const& indices, std::array const&) + { + return std_array_get(indices); + } +}; +} // end namespace internal + +template +class Tensor +{ + static_assert(NumIndices_ >= 1, "A tensor must have at least one index."); + + public: + typedef Tensor Self; + typedef typename internal::traits::StorageKind StorageKind; + typedef typename internal::traits::Index Index; + typedef typename internal::traits::Scalar Scalar; + typedef typename internal::packet_traits::type PacketScalar; + typedef typename NumTraits::Real RealScalar; + typedef Self DenseType; + + constexpr static int Options = Options_; + constexpr static std::size_t NumIndices = NumIndices_; + + protected: + TensorStorage m_storage; + + public: + EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return m_storage.dimensions()[n]; } + EIGEN_STRONG_INLINE std::array dimensions() const { return m_storage.dimensions(); } + EIGEN_STRONG_INLINE Index size() const { return internal::array_prod(m_storage.dimensions()); } + EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } + EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } + + // This makes EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + // work, because that uses base().coeffRef() - and we don't yet + // implement a similar class hierarchy + inline Self& base() { return *this; } + inline const Self& base() const { return *this; } + + void setZero() + { + // FIXME: until we have implemented packet access and the + // expression engine w.r.t. nullary ops, use this + // as a kludge. Only works with POD types, but for + // any standard usage, this shouldn't be a problem + memset((void *)data(), 0, size() * sizeof(Scalar)); + } + + inline Self& operator=(Self const& other) + { + m_storage = other.m_storage; + return *this; + } + + template + inline const Scalar& coeff(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const + { + static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + return coeff(std::array{{firstIndex, secondIndex, otherIndices...}}); + } + + inline const Scalar& coeff(const std::array& indices) const + { + eigen_internal_assert(checkIndexRange(indices)); + return m_storage.data()[linearizedIndex(indices)]; + } + + inline const Scalar& coeff(Index index) const + { + eigen_internal_assert(index >= 0 && index < size()); + return m_storage.data()[index]; + } + + template + inline Scalar& coeffRef(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) + { + static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + return coeffRef(std::array{{firstIndex, secondIndex, otherIndices...}}); + } + + inline Scalar& coeffRef(const std::array& indices) + { + eigen_internal_assert(checkIndexRange(indices)); + return m_storage.data()[linearizedIndex(indices)]; + } + + inline Scalar& coeffRef(Index index) + { + eigen_internal_assert(index >= 0 && index < size()); + return m_storage.data()[index]; + } + + template + inline const Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const + { + static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + return this->operator()(std::array{{firstIndex, secondIndex, otherIndices...}}); + } + + inline const Scalar& operator()(const std::array& indices) const + { + eigen_assert(checkIndexRange(indices)); + return coeff(indices); + } + + inline const Scalar& operator()(Index index) const + { + eigen_internal_assert(index >= 0 && index < size()); + return coeff(index); + } + + inline const Scalar& operator[](Index index) const + { + static_assert(NumIndices == 1, "The bracket operator is only for vectors, use the parenthesis operator instead."); + return coeff(index); + } + + template + inline Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) + { + static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + return operator()(std::array{{firstIndex, secondIndex, otherIndices...}}); + } + + inline Scalar& operator()(const std::array& indices) + { + eigen_assert(checkIndexRange(indices)); + return coeffRef(indices); + } + + inline Scalar& operator()(Index index) + { + eigen_assert(index >= 0 && index < size()); + return coeffRef(index); + } + + inline Scalar& operator[](Index index) + { + static_assert(NumIndices == 1, "The bracket operator is only for vectors, use the parenthesis operator instead."); + return coeffRef(index); + } + + inline Tensor() + : m_storage() + { + } + + inline Tensor(const Self& other) + : m_storage(other.m_storage) + { + } + + inline Tensor(Self&& other) + : m_storage(other.m_storage) + { + } + + template + inline Tensor(Index firstDimension, IndexTypes... otherDimensions) + : m_storage() + { + static_assert(sizeof...(otherDimensions) + 1 == NumIndices, "Number of dimensions used to construct a tensor must be equal to the rank of the tensor."); + resize(std::array{{firstDimension, otherDimensions...}}); + } + + inline Tensor(std::array dimensions) + : m_storage(internal::array_prod(dimensions), dimensions) + { + EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + } + + template + void resize(Index firstDimension, IndexTypes... otherDimensions) + { + static_assert(sizeof...(otherDimensions) + 1 == NumIndices, "Number of dimensions used to resize a tensor must be equal to the rank of the tensor."); + resize(std::array{{firstDimension, otherDimensions...}}); + } + + void resize(const std::array& dimensions) + { + std::size_t i; + Index size = Index(1); + for (i = 0; i < NumIndices; i++) { + internal::check_rows_cols_for_overflow::run(size, dimensions[i]); + size *= dimensions[i]; + } + #ifdef EIGEN_INITIALIZE_COEFFS + bool size_changed = size != this->size(); + m_storage.resize(size, dimensions); + if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + #else + m_storage.resize(size, dimensions); + #endif + } + + protected: + bool checkIndexRange(const std::array& indices) const + { + using internal::array_apply_and_reduce; + using internal::array_zip_and_reduce; + using internal::greater_equal_zero_op; + using internal::logical_and_op; + using internal::lesser_op; + + return + // check whether the indices are all >= 0 + array_apply_and_reduce(indices) && + // check whether the indices fit in the dimensions + array_zip_and_reduce(indices, m_storage.dimensions()); + } + + inline Index linearizedIndex(const std::array& indices) const + { + return internal::tensor_index_linearization_helper::run(indices, m_storage.dimensions()); + } +}; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h new file mode 100644 index 000000000..50040147d --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -0,0 +1,126 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSORSTORAGE_H +#define EIGEN_CXX11_TENSOR_TENSORSTORAGE_H + +#ifdef EIGEN_TENSOR_STORAGE_CTOR_PLUGIN + #define EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN EIGEN_TENSOR_STORAGE_CTOR_PLUGIN; +#else + #define EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN +#endif + +namespace Eigen { + +/** \internal + * + * \class TensorStorage + * \ingroup CXX11_Tensor_Module + * + * \brief Stores the data of a tensor + * + * This class stores the data of fixed-size, dynamic-size or mixed tensors + * in a way as compact as possible. + * + * \sa Tensor + */ +template class TensorStorage; + +// pure-dynamic, but without specification of all dimensions explicitly +template +class TensorStorage + : public TensorStorage::type> +{ + typedef TensorStorage::type> Base_; + public: + TensorStorage() = default; + TensorStorage(const TensorStorage&) = default; + TensorStorage(TensorStorage&&) = default; + TensorStorage(internal::constructor_without_unaligned_array_assert) : Base_(internal::constructor_without_unaligned_array_assert()) {} + TensorStorage(DenseIndex size, const std::array& dimensions) : Base_(size, dimensions) {} + TensorStorage& operator=(const TensorStorage&) = default; +}; + +// pure dynamic +template +class TensorStorage::type> +{ + T *m_data; + std::array m_dimensions; + + typedef TensorStorage::type> Self_; + public: + TensorStorage() : m_data(0), m_dimensions(internal::template repeat(0)) {} + TensorStorage(internal::constructor_without_unaligned_array_assert) + : m_data(0), m_dimensions(internal::template repeat(0)) {} + TensorStorage(DenseIndex size, const std::array& dimensions) + : m_data(internal::conditional_aligned_new_auto(size)), m_dimensions(dimensions) + { EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN } + TensorStorage(const Self_& other) + : m_data(internal::conditional_aligned_new_auto(internal::array_prod(other.m_dimensions))) + , m_dimensions(other.m_dimensions) + { + internal::smart_copy(other.m_data, other.m_data+internal::array_prod(other.m_dimensions), m_data); + } + Self_& operator=(const Self_& other) + { + if (this != &other) { + Self_ tmp(other); + this->swap(tmp); + } + return *this; + } + TensorStorage(Self_&& other) + : m_data(std::move(other.m_data)), m_dimensions(std::move(other.m_dimensions)) + { + other.m_data = nullptr; + } + Self_& operator=(Self_&& other) + { + using std::swap; + swap(m_data, other.m_data); + swap(m_dimensions, other.m_dimensions); + return *this; + } + ~TensorStorage() { internal::conditional_aligned_delete_auto(m_data, internal::array_prod(m_dimensions)); } + void swap(Self_& other) + { std::swap(m_data,other.m_data); std::swap(m_dimensions,other.m_dimensions); } + std::array dimensions(void) const {return m_dimensions;} + void conservativeResize(DenseIndex size, const std::array& nbDimensions) + { + m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, internal::array_prod(m_dimensions)); + m_dimensions = nbDimensions; + } + void resize(DenseIndex size, const std::array& nbDimensions) + { + if(size != internal::array_prod(m_dimensions)) + { + internal::conditional_aligned_delete_auto(m_data, internal::array_prod(m_dimensions)); + if (size) + m_data = internal::conditional_aligned_new_auto(size); + else + m_data = 0; + EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN + } + m_dimensions = nbDimensions; + } + const T *data() const { return m_data; } + T *data() { return m_data; } +}; + +// TODO: implement fixed-size stuff + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSORSTORAGE_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 841cab9d7..61e8f56fd 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -99,4 +99,5 @@ if(EIGEN_TEST_CXX11) # (MSVC doesn't need any for example, so this will # clash there) ei_add_test(cxx11_meta "-std=c++0x") + ei_add_test(cxx11_tensor_simple "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_simple.cpp b/unsupported/test/cxx11_tensor_simple.cpp new file mode 100644 index 000000000..6875a4e58 --- /dev/null +++ b/unsupported/test/cxx11_tensor_simple.cpp @@ -0,0 +1,271 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +static void test_1d() +{ + Tensor vec1(6); + Tensor vec2(6); + Tensor vec3; + Tensor vec4; + + vec3.resize(6); + vec4.resize(6); + + vec1(0) = 4; vec2(0) = 0; vec3(0) = 5; + vec1(1) = 8; vec2(1) = 1; vec3(1) = 4; + vec1(2) = 15; vec2(2) = 2; vec3(2) = 3; + vec1(3) = 16; vec2(3) = 3; vec3(3) = 2; + vec1(4) = 23; vec2(4) = 4; vec3(4) = 1; + vec1(5) = 42; vec2(5) = 5; vec3(5) = 0; + vec4.setZero(); + + VERIFY_IS_EQUAL((vec1.size()), 6); + VERIFY_IS_EQUAL((vec1.dimensions()[0]), 6); + + VERIFY_IS_EQUAL((vec1[0]), 4); + VERIFY_IS_EQUAL((vec1[1]), 8); + VERIFY_IS_EQUAL((vec1[2]), 15); + VERIFY_IS_EQUAL((vec1[3]), 16); + VERIFY_IS_EQUAL((vec1[4]), 23); + VERIFY_IS_EQUAL((vec1[5]), 42); + + VERIFY_IS_EQUAL((vec2[0]), 0); + VERIFY_IS_EQUAL((vec2[1]), 1); + VERIFY_IS_EQUAL((vec2[2]), 2); + VERIFY_IS_EQUAL((vec2[3]), 3); + VERIFY_IS_EQUAL((vec2[4]), 4); + VERIFY_IS_EQUAL((vec2[5]), 5); + + VERIFY_IS_EQUAL((vec3[0]), 5); + VERIFY_IS_EQUAL((vec3[1]), 4); + VERIFY_IS_EQUAL((vec3[2]), 3); + VERIFY_IS_EQUAL((vec3[3]), 2); + VERIFY_IS_EQUAL((vec3[4]), 1); + VERIFY_IS_EQUAL((vec3[5]), 0); + + VERIFY_IS_EQUAL((vec4[0]), 0); + VERIFY_IS_EQUAL((vec4[1]), 0); + VERIFY_IS_EQUAL((vec4[2]), 0); + VERIFY_IS_EQUAL((vec4[3]), 0); + VERIFY_IS_EQUAL((vec4[4]), 0); + VERIFY_IS_EQUAL((vec4[5]), 0); + + Tensor vec5(vec1); + + VERIFY_IS_EQUAL((vec5(0)), 4); + VERIFY_IS_EQUAL((vec5(1)), 8); + VERIFY_IS_EQUAL((vec5(2)), 15); + VERIFY_IS_EQUAL((vec5(3)), 16); + VERIFY_IS_EQUAL((vec5(4)), 23); + VERIFY_IS_EQUAL((vec5(5)), 42); + + VERIFY_IS_EQUAL((vec5.data()[0]), 4); + VERIFY_IS_EQUAL((vec5.data()[1]), 8); + VERIFY_IS_EQUAL((vec5.data()[2]), 15); + VERIFY_IS_EQUAL((vec5.data()[3]), 16); + VERIFY_IS_EQUAL((vec5.data()[4]), 23); + VERIFY_IS_EQUAL((vec5.data()[5]), 42); +} + +static void test_2d() +{ + Tensor mat1(2,3); + Tensor mat2(2,3); + + mat1(0,0) = 0; + mat1(0,1) = 1; + mat1(0,2) = 2; + mat1(1,0) = 3; + mat1(1,1) = 4; + mat1(1,2) = 5; + + mat2(0,0) = 0; + mat2(0,1) = 1; + mat2(0,2) = 2; + mat2(1,0) = 3; + mat2(1,1) = 4; + mat2(1,2) = 5; + + VERIFY_IS_EQUAL((mat1.size()), 6); + VERIFY_IS_EQUAL((mat1.dimensions()[0]), 2); + VERIFY_IS_EQUAL((mat1.dimensions()[1]), 3); + + VERIFY_IS_EQUAL((mat2.size()), 6); + VERIFY_IS_EQUAL((mat2.dimensions()[0]), 2); + VERIFY_IS_EQUAL((mat2.dimensions()[1]), 3); + + VERIFY_IS_EQUAL((mat1.data()[0]), 0); + VERIFY_IS_EQUAL((mat1.data()[1]), 3); + VERIFY_IS_EQUAL((mat1.data()[2]), 1); + VERIFY_IS_EQUAL((mat1.data()[3]), 4); + VERIFY_IS_EQUAL((mat1.data()[4]), 2); + VERIFY_IS_EQUAL((mat1.data()[5]), 5); + + VERIFY_IS_EQUAL((mat2.data()[0]), 0); + VERIFY_IS_EQUAL((mat2.data()[1]), 1); + VERIFY_IS_EQUAL((mat2.data()[2]), 2); + VERIFY_IS_EQUAL((mat2.data()[3]), 3); + VERIFY_IS_EQUAL((mat2.data()[4]), 4); + VERIFY_IS_EQUAL((mat2.data()[5]), 5); +} + +static void test_3d() +{ + Tensor epsilon(3,3,3); + epsilon.setZero(); + epsilon(0,1,2) = epsilon(2,0,1) = epsilon(1,2,0) = 1; + epsilon(2,1,0) = epsilon(0,2,1) = epsilon(1,0,2) = -1; + + VERIFY_IS_EQUAL((epsilon.size()), 27); + VERIFY_IS_EQUAL((epsilon.dimensions()[0]), 3); + VERIFY_IS_EQUAL((epsilon.dimensions()[1]), 3); + VERIFY_IS_EQUAL((epsilon.dimensions()[2]), 3); + + VERIFY_IS_EQUAL((epsilon(0,0,0)), 0); + VERIFY_IS_EQUAL((epsilon(0,0,1)), 0); + VERIFY_IS_EQUAL((epsilon(0,0,2)), 0); + VERIFY_IS_EQUAL((epsilon(0,1,0)), 0); + VERIFY_IS_EQUAL((epsilon(0,1,1)), 0); + VERIFY_IS_EQUAL((epsilon(0,2,0)), 0); + VERIFY_IS_EQUAL((epsilon(0,2,2)), 0); + VERIFY_IS_EQUAL((epsilon(1,0,0)), 0); + VERIFY_IS_EQUAL((epsilon(1,0,1)), 0); + VERIFY_IS_EQUAL((epsilon(1,1,0)), 0); + VERIFY_IS_EQUAL((epsilon(1,1,1)), 0); + VERIFY_IS_EQUAL((epsilon(1,1,2)), 0); + VERIFY_IS_EQUAL((epsilon(1,2,1)), 0); + VERIFY_IS_EQUAL((epsilon(1,2,2)), 0); + VERIFY_IS_EQUAL((epsilon(2,0,0)), 0); + VERIFY_IS_EQUAL((epsilon(2,0,2)), 0); + VERIFY_IS_EQUAL((epsilon(2,1,1)), 0); + VERIFY_IS_EQUAL((epsilon(2,1,2)), 0); + VERIFY_IS_EQUAL((epsilon(2,2,0)), 0); + VERIFY_IS_EQUAL((epsilon(2,2,1)), 0); + VERIFY_IS_EQUAL((epsilon(2,2,2)), 0); + + VERIFY_IS_EQUAL((epsilon(0,1,2)), 1); + VERIFY_IS_EQUAL((epsilon(2,0,1)), 1); + VERIFY_IS_EQUAL((epsilon(1,2,0)), 1); + VERIFY_IS_EQUAL((epsilon(2,1,0)), -1); + VERIFY_IS_EQUAL((epsilon(0,2,1)), -1); + VERIFY_IS_EQUAL((epsilon(1,0,2)), -1); + + std::array dims{{2,3,4}}; + Tensor t1(dims); + Tensor t2(dims); + + VERIFY_IS_EQUAL((t1.size()), 24); + VERIFY_IS_EQUAL((t1.dimensions()[0]), 2); + VERIFY_IS_EQUAL((t1.dimensions()[1]), 3); + VERIFY_IS_EQUAL((t1.dimensions()[2]), 4); + + VERIFY_IS_EQUAL((t2.size()), 24); + VERIFY_IS_EQUAL((t2.dimensions()[0]), 2); + VERIFY_IS_EQUAL((t2.dimensions()[1]), 3); + VERIFY_IS_EQUAL((t2.dimensions()[2]), 4); + + for (int i = 0; i < 2; i++) { + for (int j = 0; j < 3; j++) { + for (int k = 0; k < 4; k++) { + t1(i, j, k) = 100 * i + 10 * j + k; + t2(i, j, k) = 100 * i + 10 * j + k; + } + } + } + + VERIFY_IS_EQUAL((t1.data()[0]), 0); + VERIFY_IS_EQUAL((t1.data()[1]), 100); + VERIFY_IS_EQUAL((t1.data()[2]), 10); + VERIFY_IS_EQUAL((t1.data()[3]), 110); + VERIFY_IS_EQUAL((t1.data()[4]), 20); + VERIFY_IS_EQUAL((t1.data()[5]), 120); + VERIFY_IS_EQUAL((t1.data()[6]), 1); + VERIFY_IS_EQUAL((t1.data()[7]), 101); + VERIFY_IS_EQUAL((t1.data()[8]), 11); + VERIFY_IS_EQUAL((t1.data()[9]), 111); + VERIFY_IS_EQUAL((t1.data()[10]), 21); + VERIFY_IS_EQUAL((t1.data()[11]), 121); + VERIFY_IS_EQUAL((t1.data()[12]), 2); + VERIFY_IS_EQUAL((t1.data()[13]), 102); + VERIFY_IS_EQUAL((t1.data()[14]), 12); + VERIFY_IS_EQUAL((t1.data()[15]), 112); + VERIFY_IS_EQUAL((t1.data()[16]), 22); + VERIFY_IS_EQUAL((t1.data()[17]), 122); + VERIFY_IS_EQUAL((t1.data()[18]), 3); + VERIFY_IS_EQUAL((t1.data()[19]), 103); + VERIFY_IS_EQUAL((t1.data()[20]), 13); + VERIFY_IS_EQUAL((t1.data()[21]), 113); + VERIFY_IS_EQUAL((t1.data()[22]), 23); + VERIFY_IS_EQUAL((t1.data()[23]), 123); + + VERIFY_IS_EQUAL((t2.data()[0]), 0); + VERIFY_IS_EQUAL((t2.data()[1]), 1); + VERIFY_IS_EQUAL((t2.data()[2]), 2); + VERIFY_IS_EQUAL((t2.data()[3]), 3); + VERIFY_IS_EQUAL((t2.data()[4]), 10); + VERIFY_IS_EQUAL((t2.data()[5]), 11); + VERIFY_IS_EQUAL((t2.data()[6]), 12); + VERIFY_IS_EQUAL((t2.data()[7]), 13); + VERIFY_IS_EQUAL((t2.data()[8]), 20); + VERIFY_IS_EQUAL((t2.data()[9]), 21); + VERIFY_IS_EQUAL((t2.data()[10]), 22); + VERIFY_IS_EQUAL((t2.data()[11]), 23); + VERIFY_IS_EQUAL((t2.data()[12]), 100); + VERIFY_IS_EQUAL((t2.data()[13]), 101); + VERIFY_IS_EQUAL((t2.data()[14]), 102); + VERIFY_IS_EQUAL((t2.data()[15]), 103); + VERIFY_IS_EQUAL((t2.data()[16]), 110); + VERIFY_IS_EQUAL((t2.data()[17]), 111); + VERIFY_IS_EQUAL((t2.data()[18]), 112); + VERIFY_IS_EQUAL((t2.data()[19]), 113); + VERIFY_IS_EQUAL((t2.data()[20]), 120); + VERIFY_IS_EQUAL((t2.data()[21]), 121); + VERIFY_IS_EQUAL((t2.data()[22]), 122); + VERIFY_IS_EQUAL((t2.data()[23]), 123); +} + +static void test_simple_assign() +{ + Tensor epsilon(3,3,3); + epsilon.setZero(); + epsilon(0,1,2) = epsilon(2,0,1) = epsilon(1,2,0) = 1; + epsilon(2,1,0) = epsilon(0,2,1) = epsilon(1,0,2) = -1; + + Tensor e2(2,3,1); + e2.setZero(); + VERIFY_IS_EQUAL((e2(1,2,0)), 0); + + e2 = epsilon; + VERIFY_IS_EQUAL((e2(1,2,0)), 1); + VERIFY_IS_EQUAL((e2(0,1,2)), 1); + VERIFY_IS_EQUAL((e2(2,0,1)), 1); + VERIFY_IS_EQUAL((e2(2,1,0)), -1); + VERIFY_IS_EQUAL((e2(0,2,1)), -1); + VERIFY_IS_EQUAL((e2(1,0,2)), -1); +} + +void test_cxx11_tensor_simple() +{ + CALL_SUBTEST(test_1d()); + CALL_SUBTEST(test_2d()); + CALL_SUBTEST(test_3d()); + CALL_SUBTEST(test_simple_assign()); +} + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ From 03a956925a969e3759418bd1e0fced82eb5f9d12 Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Thu, 14 Nov 2013 23:35:11 +0100 Subject: [PATCH 0191/1103] CXX11/TensorSymmetry: add symmetry support for Tensor class Add a symCoeff() method to the Tensor class template that allows the user of the class to set multiple elements of a tensor at once if they are connected by a symmetry operation with respect to the tensor's indices (symmetry/antisymmetry/hermiticity/antihermiticity under echange of two indices and combination thereof for different pairs of indices). A compile-time resolution of the required symmetry groups via meta templates is also implemented. For small enough groups this is used to unroll the loop that goes through all the elements of the Tensor that are connected by this group. For larger groups or groups where the symmetries are defined at run time, a standard run-time implementation of the same algorithm is provided. For example, the following code completely initializes all elements of the totally antisymmetric tensor in three dimensions ('epsilon tensor'): SGroup<3, AntiSymmetry<0,1>, AntiSymmetry<1,2>> sym; Eigen::Tensor epsilon(3,3,3); epsilon.setZero(); epsilon.symCoeff(sym, 0, 1, 2) = 1; --- unsupported/Eigen/CXX11/TensorSymmetry | 41 + unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 15 + .../src/TensorSymmetry/DynamicSymmetry.h | 265 ++++++ .../CXX11/src/TensorSymmetry/StaticSymmetry.h | 216 +++++ .../Eigen/CXX11/src/TensorSymmetry/Symmetry.h | 312 +++++++ .../TensorSymmetry/util/TemplateGroupTheory.h | 667 ++++++++++++++ unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_symmetry.cpp | 819 ++++++++++++++++++ 8 files changed, 2336 insertions(+) create mode 100644 unsupported/Eigen/CXX11/TensorSymmetry create mode 100644 unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h create mode 100644 unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h create mode 100644 unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h create mode 100644 unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h create mode 100644 unsupported/test/cxx11_tensor_symmetry.cpp diff --git a/unsupported/Eigen/CXX11/TensorSymmetry b/unsupported/Eigen/CXX11/TensorSymmetry new file mode 100644 index 000000000..5bbab0a80 --- /dev/null +++ b/unsupported/Eigen/CXX11/TensorSymmetry @@ -0,0 +1,41 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSORSYMMETRY_MODULE +#define EIGEN_CXX11_TENSORSYMMETRY_MODULE + +#include + +#include + +/** \defgroup CXX11_TensorSymmetry_Module Tensor Symmetry Module + * + * This module provides a classes that allow for the definition of + * symmetries w.r.t. tensor indices. + * + * Including this module will implicitly include the Tensor module. + * + * \code + * #include + * \endcode + */ + +#include "src/TensorSymmetry/util/TemplateGroupTheory.h" +#include "src/TensorSymmetry/Symmetry.h" +#include "src/TensorSymmetry/StaticSymmetry.h" +#include "src/TensorSymmetry/DynamicSymmetry.h" + +#include + +#endif // EIGEN_CXX11_TENSORSYMMETRY_MODULE + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index c40905af4..ff3d6513e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -92,6 +92,9 @@ struct tensor_index_linearization_helper return std_array_get(indices); } }; + +/* Forward-declaration required for the symmetry support. */ +template class tensor_symmetry_value_setter; } // end namespace internal template @@ -283,6 +286,18 @@ class Tensor #endif } + template + internal::tensor_symmetry_value_setter symCoeff(const Symmetry_& symmetry, Index firstIndex, IndexTypes... otherIndices) + { + return symCoeff(symmetry, std::array{{firstIndex, otherIndices...}}); + } + + template + internal::tensor_symmetry_value_setter symCoeff(const Symmetry_& symmetry, std::array const& indices) + { + return internal::tensor_symmetry_value_setter(*this, symmetry, indices); + } + protected: bool checkIndexRange(const std::array& indices) const { diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h new file mode 100644 index 000000000..87332610d --- /dev/null +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h @@ -0,0 +1,265 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSORSYMMETRY_DYNAMICSYMMETRY_H +#define EIGEN_CXX11_TENSORSYMMETRY_DYNAMICSYMMETRY_H + +namespace Eigen { + +class DynamicSGroup +{ + public: + inline explicit DynamicSGroup(std::size_t numIndices) : m_numIndices(numIndices), m_elements(), m_generators(), m_globalFlags(0) { m_elements.push_back(ge(Generator(0, 0, 0))); } + inline DynamicSGroup(const DynamicSGroup& o) : m_numIndices(o.m_numIndices), m_elements(o.m_elements), m_generators(o.m_generators), m_globalFlags(o.m_globalFlags) { } + inline DynamicSGroup(DynamicSGroup&& o) : m_numIndices(o.m_numIndices), m_elements(), m_generators(o.m_generators), m_globalFlags(o.m_globalFlags) { std::swap(m_elements, o.m_elements); } + inline DynamicSGroup& operator=(const DynamicSGroup& o) { m_numIndices = o.m_numIndices; m_elements = o.m_elements; m_generators = o.m_generators; m_globalFlags = o.m_globalFlags; return *this; } + inline DynamicSGroup& operator=(DynamicSGroup&& o) { m_numIndices = o.m_numIndices; std::swap(m_elements, o.m_elements); m_generators = o.m_generators; m_globalFlags = o.m_globalFlags; return *this; } + + void add(int one, int two, int flags = 0); + + template + inline void add(Gen_) { add(Gen_::One, Gen_::Two, Gen_::Flags); } + inline void addSymmetry(int one, int two) { add(one, two, 0); } + inline void addAntiSymmetry(int one, int two) { add(one, two, NegationFlag); } + inline void addHermiticity(int one, int two) { add(one, two, ConjugationFlag); } + inline void addAntiHermiticity(int one, int two) { add(one, two, NegationFlag | ConjugationFlag); } + + template + inline RV apply(const std::array& idx, RV initial, Args&&... args) const + { + eigen_assert(N == m_numIndices); + for (std::size_t i = 0; i < size(); i++) + initial = Op::run(h_permute(i, idx, typename internal::gen_numeric_list::type()), m_elements[i].flags, initial, std::forward(args)...); + return initial; + } + + template + inline RV apply(const std::vector& idx, RV initial, Args&&... args) const + { + eigen_assert(idx.size() == m_numIndices); + for (std::size_t i = 0; i < size(); i++) + initial = Op::run(h_permute(i, idx), m_elements[i].flags, initial, std::forward(args)...); + return initial; + } + + inline int globalFlags() const { return m_globalFlags; } + inline std::size_t size() const { return m_elements.size(); } + private: + struct GroupElement { + std::vector representation; + int flags; + bool isId() const + { + for (std::size_t i = 0; i < representation.size(); i++) + if (i != (size_t)representation[i]) + return false; + return true; + } + }; + struct Generator { + int one; + int two; + int flags; + constexpr inline Generator(int one_, int two_, int flags_) : one(one_), two(two_), flags(flags_) {} + }; + + std::size_t m_numIndices; + std::vector m_elements; + std::vector m_generators; + int m_globalFlags; + + template + inline std::array h_permute(std::size_t which, const std::array& idx, internal::numeric_list) const + { + return std::array{{ idx[m_elements[which].representation[n]]... }}; + } + + template + inline std::vector h_permute(std::size_t which, std::vector idx) const + { + std::vector result; + result.reserve(idx.size()); + for (auto k : m_elements[which].representation) + result.push_back(idx[k]); + return result; + } + + inline GroupElement ge(Generator const& g) const + { + GroupElement result; + result.representation.reserve(m_numIndices); + result.flags = g.flags; + for (std::size_t k = 0; k < m_numIndices; k++) { + if (k == (std::size_t)g.one) + result.representation.push_back(g.two); + else if (k == (std::size_t)g.two) + result.representation.push_back(g.one); + else + result.representation.push_back(int(k)); + } + return result; + } + + GroupElement mul(GroupElement, GroupElement) const; + inline GroupElement mul(Generator g1, GroupElement g2) const + { + return mul(ge(g1), g2); + } + + inline GroupElement mul(GroupElement g1, Generator g2) const + { + return mul(g1, ge(g2)); + } + + inline GroupElement mul(Generator g1, Generator g2) const + { + return mul(ge(g1), ge(g2)); + } + + inline int findElement(GroupElement e) const + { + for (auto ee : m_elements) { + if (ee.representation == e.representation) + return ee.flags ^ e.flags; + } + return -1; + } + + void updateGlobalFlags(int flagDiffOfSameGenerator); +}; + +// dynamic symmetry group that auto-adds the template parameters in the constructor +template +class DynamicSGroupFromTemplateArgs : public DynamicSGroup +{ + public: + inline DynamicSGroupFromTemplateArgs() : DynamicSGroup(NumIndices) + { + add_all(internal::type_list()); + } + inline DynamicSGroupFromTemplateArgs(DynamicSGroupFromTemplateArgs const& other) : DynamicSGroup(other) { } + inline DynamicSGroupFromTemplateArgs(DynamicSGroupFromTemplateArgs&& other) : DynamicSGroup(other) { } + inline DynamicSGroupFromTemplateArgs& operator=(const DynamicSGroupFromTemplateArgs& o) { DynamicSGroup::operator=(o); return *this; } + inline DynamicSGroupFromTemplateArgs& operator=(DynamicSGroupFromTemplateArgs&& o) { DynamicSGroup::operator=(o); return *this; } + + private: + template + inline void add_all(internal::type_list) + { + add(Gen1()); + add_all(internal::type_list()); + } + + inline void add_all(internal::type_list<>) + { + } +}; + +inline DynamicSGroup::GroupElement DynamicSGroup::mul(GroupElement g1, GroupElement g2) const +{ + eigen_internal_assert(g1.representation.size() == m_numIndices); + eigen_internal_assert(g2.representation.size() == m_numIndices); + + GroupElement result; + result.representation.reserve(m_numIndices); + for (std::size_t i = 0; i < m_numIndices; i++) + result.representation.push_back(g2.representation[g1.representation[i]]); + result.flags = g1.flags ^ g2.flags; + return result; +} + +inline void DynamicSGroup::add(int one, int two, int flags) +{ + eigen_assert(one >= 0 && (std::size_t)one < m_numIndices); + eigen_assert(two >= 0 && (std::size_t)two < m_numIndices); + eigen_assert(one != two); + Generator g{one, two ,flags}; + GroupElement e = ge(g); + + /* special case for first generator */ + if (m_elements.size() == 1) { + while (!e.isId()) { + m_elements.push_back(e); + e = mul(e, g); + } + + if (e.flags > 0) + updateGlobalFlags(e.flags); + + // only add in case we didn't have identity + if (m_elements.size() > 1) + m_generators.push_back(g); + return; + } + + int p = findElement(e); + if (p >= 0) { + updateGlobalFlags(p); + return; + } + + std::size_t coset_order = m_elements.size(); + m_elements.push_back(e); + for (std::size_t i = 1; i < coset_order; i++) + m_elements.push_back(mul(m_elements[i], e)); + m_generators.push_back(g); + + std::size_t coset_rep = coset_order; + do { + for (auto g : m_generators) { + e = mul(m_elements[coset_rep], g); + p = findElement(e); + if (p < 0) { + // element not yet in group + m_elements.push_back(e); + for (std::size_t i = 1; i < coset_order; i++) + m_elements.push_back(mul(m_elements[i], e)); + } else if (p > 0) { + updateGlobalFlags(p); + } + } + coset_rep += coset_order; + } while (coset_rep < m_elements.size()); +} + +inline void DynamicSGroup::updateGlobalFlags(int flagDiffOfSameGenerator) +{ + switch (flagDiffOfSameGenerator) { + case 0: + default: + // nothing happened + break; + case NegationFlag: + // every element is it's own negative => whole tensor is zero + m_globalFlags |= GlobalZeroFlag; + break; + case ConjugationFlag: + // every element is it's own conjugate => whole tensor is real + m_globalFlags |= GlobalRealFlag; + break; + case (NegationFlag | ConjugationFlag): + // every element is it's own negative conjugate => whole tensor is imaginary + m_globalFlags |= GlobalImagFlag; + break; + /* NOTE: + * since GlobalZeroFlag == GlobalRealFlag | GlobalImagFlag, if one generator + * causes the tensor to be real and the next one to be imaginary, this will + * trivially give the correct result + */ + } +} + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSORSYMMETRY_DYNAMICSYMMETRY_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h new file mode 100644 index 000000000..048e753fc --- /dev/null +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h @@ -0,0 +1,216 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSORSYMMETRY_STATICSYMMETRY_H +#define EIGEN_CXX11_TENSORSYMMETRY_STATICSYMMETRY_H + +namespace Eigen { + +namespace internal { + +template struct tensor_static_symgroup_permutate; + +template +struct tensor_static_symgroup_permutate> +{ + constexpr static std::size_t N = sizeof...(nn); + + template + constexpr static inline std::array run(const std::array& indices) + { + return {{indices[nn]...}}; + } +}; + +template +struct tensor_static_symgroup_element +{ + typedef indices_ indices; + constexpr static int flags = flags_; +}; + +template +struct tensor_static_symgroup_element_ctor +{ + typedef tensor_static_symgroup_element< + typename gen_numeric_list_swapped_pair::type, + Gen::Flags + > type; +}; + +template +struct tensor_static_symgroup_identity_ctor +{ + typedef tensor_static_symgroup_element< + typename gen_numeric_list::type, + 0 + > type; +}; + +template +struct tensor_static_symgroup_multiply_helper +{ + template + constexpr static inline numeric_list::value...> helper(numeric_list) { + return numeric_list::value...>(); + } +}; + +template +struct tensor_static_symgroup_multiply +{ + private: + typedef typename A::indices iia; + typedef typename B::indices iib; + constexpr static int ffa = A::flags; + constexpr static int ffb = B::flags; + + public: + static_assert(iia::count == iib::count, "Cannot multiply symmetry elements with different number of indices."); + + typedef tensor_static_symgroup_element< + decltype(tensor_static_symgroup_multiply_helper::helper(iia())), + ffa ^ ffb + > type; +}; + +template +struct tensor_static_symgroup_equality +{ + typedef typename A::indices iia; + typedef typename B::indices iib; + constexpr static int ffa = A::flags; + constexpr static int ffb = B::flags; + static_assert(iia::count == iib::count, "Cannot compare symmetry elements with different number of indices."); + + constexpr static bool value = is_same::value; + + private: + /* this should be zero if they are identical, or else the tensor + * will be forced to be pure real, pure imaginary or even pure zero + */ + constexpr static int flags_cmp_ = ffa ^ ffb; + + /* either they are not equal, then we don't care whether the flags + * match, or they are equal, and then we have to check + */ + constexpr static bool is_zero = value && flags_cmp_ == NegationFlag; + constexpr static bool is_real = value && flags_cmp_ == ConjugationFlag; + constexpr static bool is_imag = value && flags_cmp_ == (NegationFlag | ConjugationFlag); + + public: + constexpr static int global_flags = + (is_real ? GlobalRealFlag : 0) | + (is_imag ? GlobalImagFlag : 0) | + (is_zero ? GlobalZeroFlag : 0); +}; + +template +struct tensor_static_symgroup +{ + typedef StaticSGroup type; + constexpr static std::size_t size = type::static_size; +}; + +template +constexpr static inline std::array tensor_static_symgroup_index_permute(std::array idx, internal::numeric_list) +{ + return {{ idx[ii]... }}; +} + +template +static inline std::vector tensor_static_symgroup_index_permute(std::vector idx, internal::numeric_list) +{ + return {{ idx[ii]... }}; +} + +template struct tensor_static_symgroup_do_apply; + +template +struct tensor_static_symgroup_do_apply> +{ + template + static inline RV run(const std::array& idx, RV initial, Args&&... args) + { + initial = Op::run(tensor_static_symgroup_index_permute(idx, typename first::indices()), first::flags, initial, std::forward(args)...); + return tensor_static_symgroup_do_apply>::template run(idx, initial, args...); + } + + template + static inline RV run(const std::vector& idx, RV initial, Args&&... args) + { + initial = Op::run(tensor_static_symgroup_index_permute(idx, typename first::indices()), first::flags, initial, std::forward(args)...); + return tensor_static_symgroup_do_apply>::template run(idx, initial, args...); + } +}; + +template +struct tensor_static_symgroup_do_apply> +{ + template + static inline RV run(const std::array&, RV initial, Args&&...) + { + // do nothing + return initial; + } + + template + static inline RV run(const std::vector&, RV initial, Args&&...) + { + // do nothing + return initial; + } +}; + +} // end namespace internal + +template +class StaticSGroup +{ + typedef internal::group_theory::enumerate_group_elements< + internal::tensor_static_symgroup_multiply, + internal::tensor_static_symgroup_equality, + typename internal::tensor_static_symgroup_identity_ctor::type, + internal::type_list::type...> + > group_elements; + typedef typename group_elements::type ge; + public: + constexpr inline StaticSGroup() {} + constexpr inline StaticSGroup(const StaticSGroup&) {} + constexpr inline StaticSGroup(StaticSGroup&&) {} + + template + static inline RV apply(const std::array& idx, RV initial, Args&&... args) + { + return internal::tensor_static_symgroup_do_apply::template run(idx, initial, args...); + } + + template + static inline RV apply(const std::vector& idx, RV initial, Args&&... args) + { + eigen_assert(idx.size() == NumIndices); + return internal::tensor_static_symgroup_do_apply::template run(idx, initial, args...); + } + + constexpr static std::size_t static_size = ge::count; + + constexpr static inline std::size_t size() { + return ge::count; + } + constexpr static inline int globalFlags() { return group_elements::global_flags; } +}; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSORSYMMETRY_STATICSYMMETRY_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h new file mode 100644 index 000000000..40ad8d7ff --- /dev/null +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h @@ -0,0 +1,312 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSORSYMMETRY_SYMMETRY_H +#define EIGEN_CXX11_TENSORSYMMETRY_SYMMETRY_H + +namespace Eigen { + +enum { + NegationFlag = 0x01, + ConjugationFlag = 0x02 +}; + +enum { + GlobalRealFlag = 0x01, + GlobalImagFlag = 0x02, + GlobalZeroFlag = 0x03 +}; + +namespace internal { + +template struct tensor_symmetry_pre_analysis; +template struct tensor_static_symgroup; +template struct tensor_static_symgroup_if; +template struct tensor_symmetry_calculate_flags; +template struct tensor_symmetry_assign_value; + +} // end namespace internal + +template +struct Symmetry +{ + static_assert(One_ != Two_, "Symmetries must cover distinct indices."); + constexpr static int One = One_; + constexpr static int Two = Two_; + constexpr static int Flags = 0; +}; + +template +struct AntiSymmetry +{ + static_assert(One_ != Two_, "Symmetries must cover distinct indices."); + constexpr static int One = One_; + constexpr static int Two = Two_; + constexpr static int Flags = NegationFlag; +}; + +template +struct Hermiticity +{ + static_assert(One_ != Two_, "Symmetries must cover distinct indices."); + constexpr static int One = One_; + constexpr static int Two = Two_; + constexpr static int Flags = ConjugationFlag; +}; + +template +struct AntiHermiticity +{ + static_assert(One_ != Two_, "Symmetries must cover distinct indices."); + constexpr static int One = One_; + constexpr static int Two = Two_; + constexpr static int Flags = ConjugationFlag | NegationFlag; +}; + +/** \class DynamicSGroup + * \ingroup TensorSymmetry_Module + * + * \brief Dynamic symmetry group + * + * The %DynamicSGroup class represents a symmetry group that need not be known at + * compile time. It is useful if one wants to support arbitrary run-time defineable + * symmetries for tensors, but it is also instantiated if a symmetry group is defined + * at compile time that would be either too large for the compiler to reasonably + * generate (using templates to calculate this at compile time is very inefficient) + * or that the compiler could generate the group but that it wouldn't make sense to + * unroll the loop for setting coefficients anymore. + */ +class DynamicSGroup; + +/** \internal + * + * \class DynamicSGroupFromTemplateArgs + * \ingroup TensorSymmetry_Module + * + * \brief Dynamic symmetry group, initialized from template arguments + * + * This class is a child class of DynamicSGroup. It uses the template arguments + * specified to initialize itself. + */ +template +class DynamicSGroupFromTemplateArgs; + +/** \class StaticSGroup + * \ingroup TensorSymmetry_Module + * + * \brief Static symmetry group + * + * This class represents a symmetry group that is known and resolved completely + * at compile time. Ideally, no run-time penalty is incurred compared to the + * manual unrolling of the symmetry. + * + * CAUTION: + * + * Do not use this class directly for large symmetry groups. The compiler + * may run into a limit, or segfault or in the very least will take a very, + * very, very long time to compile the code. Use the SGroup class instead + * if you want a static group. That class contains logic that will + * automatically select the DynamicSGroup class instead if the symmetry + * group becomes too large. (In that case, unrolling may not even be + * beneficial.) + */ +template +class StaticSGroup; + +/** \class SGroup + * \ingroup TensorSymmetry_Module + * + * \brief Symmetry group, initialized from template arguments + * + * This class represents a symmetry group whose generators are already + * known at compile time. It may or may not be resolved at compile time, + * depending on the estimated size of the group. + * + * \sa StaticSGroup + * \sa DynamicSGroup + */ +template +class SGroup : public internal::tensor_symmetry_pre_analysis::root_type +{ + public: + typedef typename internal::tensor_symmetry_pre_analysis::root_type Base; + + // make standard constructors + assignment operators public + inline SGroup() : Base() { } + inline SGroup(const SGroup& other) : Base(other) { } + inline SGroup(SGroup&& other) : Base(other) { } + inline SGroup& operator=(const SGroup& other) { Base::operator=(other); return *this; } + inline SGroup& operator=(SGroup&& other) { Base::operator=(other); return *this; } + + // all else is defined in the base class +}; + +namespace internal { + +/** \internal + * + * \class tensor_symmetry_pre_analysis + * \ingroup TensorSymmetry_Module + * + * \brief Pre-select whether to use a static or dynamic symmetry group + * + * When a symmetry group could in principle be determined at compile time, + * this template implements the logic whether to actually do that or whether + * to rather defer that to runtime. + * + * The logic is as follows: + *
+ *
No generators (trivial symmetry):
+ *
Use a trivial static group. Ideally, this has no performance impact + * compared to not using symmetry at all. In practice, this might not + * be the case.
+ *
More than 4 generators:
+ *
Calculate the group at run time, it is likely far too large for the + * compiler to be able to properly generate it in a realistic time.
+ *
Up to and including 4 generators:
+ *
Actually enumerate all group elements, but then check how many there + * are. If there are more than 16, it is unlikely that unrolling the + * loop (as is done in the static compile-time case) is sensible, so + * use a dynamic group instead. If there are at most 16 elements, actually + * use that static group. Note that the largest group with 4 generators + * still compiles with reasonable resources.
+ *
+ * + * Note: Example compile time performance with g++-4.6 on an Intenl Core i5-3470 + * with 16 GiB RAM (all generators non-redundant and the subgroups don't + * factorize): + * + * # Generators -O0 -ggdb -O2 + * ------------------------------------------------------------------- + * 1 0.5 s / 250 MiB 0.45s / 230 MiB + * 2 0.5 s / 260 MiB 0.5 s / 250 MiB + * 3 0.65s / 310 MiB 0.62s / 310 MiB + * 4 2.2 s / 860 MiB 1.7 s / 770 MiB + * 5 130 s / 13000 MiB 120 s / 11000 MiB + * + * It is clear that everything is still very efficient up to 4 generators, then + * the memory and CPU requirements become unreasonable. Thus we only instantiate + * the template group theory logic if the number of generators supplied is 4 or + * lower, otherwise this will be forced to be done during runtime, where the + * algorithm is reasonably fast. + */ +template +struct tensor_symmetry_pre_analysis +{ + typedef StaticSGroup root_type; +}; + +template +struct tensor_symmetry_pre_analysis +{ + constexpr static std::size_t max_static_generators = 4; + constexpr static std::size_t max_static_elements = 16; + typedef tensor_static_symgroup_if<(sizeof...(Gens_) + 1 <= max_static_generators), NumIndices, Gen_, Gens_...> helper; + constexpr static std::size_t possible_size = helper::size; + + typedef typename conditional< + possible_size == 0 || possible_size >= max_static_elements, + DynamicSGroupFromTemplateArgs, + typename helper::type + >::type root_type; +}; + +template +struct tensor_static_symgroup_if +{ + constexpr static std::size_t size = 0; + typedef void type; +}; + +template +struct tensor_static_symgroup_if : tensor_static_symgroup {}; + +template +struct tensor_symmetry_assign_value +{ + typedef typename Tensor_::Index Index; + typedef typename Tensor_::Scalar Scalar; + constexpr static std::size_t NumIndices = Tensor_::NumIndices; + + static inline int run(const std::array& transformed_indices, int transformation_flags, int dummy, Tensor_& tensor, const Scalar& value_) + { + Scalar value(value_); + if (transformation_flags & ConjugationFlag) + value = numext::conj(value); + if (transformation_flags & NegationFlag) + value = -value; + tensor.coeffRef(transformed_indices) = value; + return dummy; + } +}; + +template +struct tensor_symmetry_calculate_flags +{ + typedef typename Tensor_::Index Index; + constexpr static std::size_t NumIndices = Tensor_::NumIndices; + + static inline int run(const std::array& transformed_indices, int transform_flags, int current_flags, const std::array& orig_indices) + { + if (transformed_indices == orig_indices) { + if (transform_flags & (ConjugationFlag | NegationFlag)) + return current_flags | GlobalImagFlag; // anti-hermitian diagonal + else if (transform_flags & ConjugationFlag) + return current_flags | GlobalRealFlag; // hermitian diagonal + else if (transform_flags & NegationFlag) + return current_flags | GlobalZeroFlag; // anti-symmetric diagonal + } + return current_flags; + } +}; + +template +class tensor_symmetry_value_setter +{ + public: + typedef typename Tensor_::Index Index; + typedef typename Tensor_::Scalar Scalar; + constexpr static std::size_t NumIndices = Tensor_::NumIndices; + + inline tensor_symmetry_value_setter(Tensor_& tensor, Symmetry_ const& symmetry, std::array const& indices) + : m_tensor(tensor), m_symmetry(symmetry), m_indices(indices) { } + + inline tensor_symmetry_value_setter& operator=(Scalar const& value) + { + doAssign(value); + return *this; + } + private: + Tensor_& m_tensor; + Symmetry_ m_symmetry; + std::array m_indices; + + inline void doAssign(Scalar const& value) + { + #ifdef EIGEN_TENSOR_SYMMETRY_CHECK_VALUES + int value_flags = m_symmetry.template apply, int>(m_indices, m_symmetry.globalFlags(), m_indices); + if (value_flags & GlobalRealFlag) + eigen_assert(numext::imag(value) == 0); + if (value_flags & GlobalImagFlag) + eigen_assert(numext::real(value) == 0); + #endif + m_symmetry.template apply, int>(m_indices, 0, m_tensor, value); + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSORSYMMETRY_SYMMETRY_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h new file mode 100644 index 000000000..716d7d2e5 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h @@ -0,0 +1,667 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSORSYMMETRY_TEMPLATEGROUPTHEORY_H +#define EIGEN_CXX11_TENSORSYMMETRY_TEMPLATEGROUPTHEORY_H + +namespace Eigen { + +namespace internal { + +namespace group_theory { + +/** \internal + * \file CXX11/Tensor/util/TemplateGroupTheory.h + * This file contains C++ templates that implement group theory algorithms. + * + * The algorithms allow for a compile-time analysis of finite groups. + * + * Currently only Dimino's algorithm is implemented, which returns a list + * of all elements in a group given a set of (possibly redundant) generators. + * (One could also do that with the so-called orbital algorithm, but that + * is much more expensive and usually has no advantages.) + */ + +/********************************************************************** + * "Ok kid, here is where it gets complicated." + * - Amelia Pond in the "Doctor Who" episode + * "The Big Bang" + * + * Dimino's algorithm + * ================== + * + * The following is Dimino's algorithm in sequential form: + * + * Input: identity element, list of generators, equality check, + * multiplication operation + * Output: list of group elements + * + * 1. add identity element + * 2. remove identities from list of generators + * 3. add all powers of first generator that aren't the + * identity element + * 4. go through all remaining generators: + * a. if generator is already in the list of elements + * -> do nothing + * b. otherwise + * i. remember current # of elements + * (i.e. the size of the current subgroup) + * ii. add all current elements (which includes + * the identity) each multiplied from right + * with the current generator to the group + * iii. add all remaining cosets that are generated + * by products of the new generator with itself + * and all other generators seen so far + * + * In functional form, this is implemented as a long set of recursive + * templates that have a complicated relationship. + * + * The main interface for Dimino's algorithm is the template + * enumerate_group_elements. All lists are implemented as variadic + * type_list and numeric_list + * templates. + * + * 'Calling' templates is usually done via typedefs. + * + * This algorithm is an extended version of the basic version. The + * extension consists in the fact that each group element has a set + * of flags associated with it. Multiplication of two group elements + * with each other results in a group element whose flags are the + * XOR of the flags of the previous elements. Each time the algorithm + * notices that a group element it just calculated is already in the + * list of current elements, the flags of both will be compared and + * added to the so-called 'global flags' of the group. + * + * The rationale behind this extension is that this allows not only + * for the description of symmetries between tensor indices, but + * also allows for the description of hermiticity, antisymmetry and + * antihermiticity. Negation and conjugation each are specific bit + * in the flags value and if two different ways to reach a group + * element lead to two different flags, this poses a constraint on + * the allowed values of the resulting tensor. For example, if a + * group element is reach both with and without the conjugation + * flags, it is clear that the resulting tensor has to be real. + * + * Note that this flag mechanism is quite generic and may have other + * uses beyond tensor properties. + * + * IMPORTANT: + * This algorithm assumes the group to be finite. If you try to + * run it with a group that's infinite, the algorithm will only + * terminate once you hit a compiler limit (max template depth). + * Also note that trying to use this implementation to create a + * very large group will probably either make you hit the same + * limit, cause the compiler to segfault or at the very least + * take a *really* long time (hours, days, weeks - sic!) to + * compile. It is not recommended to plug in more than 4 + * generators, unless they are independent of each other. + */ + +/** \internal + * + * \class strip_identities + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Cleanse a list of group elements of the identity element + * + * This template is used to make a first pass through all initial + * generators of Dimino's algorithm and remove the identity + * elements. + * + * \sa enumerate_group_elements + */ +template class Equality, typename id, typename L> struct strip_identities; + +template< + template class Equality, + typename id, + typename t, + typename... ts +> +struct strip_identities> +{ + typedef typename conditional< + Equality::value, + typename strip_identities>::type, + typename concat, typename strip_identities>::type>::type + >::type type; + constexpr static int global_flags = Equality::global_flags | strip_identities>::global_flags; +}; + +template< + template class Equality, + typename id + EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, ts) +> +struct strip_identities> +{ + typedef type_list<> type; + constexpr static int global_flags = 0; +}; + +/** \internal + * + * \class dimino_first_step_elements_helper + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Recursive template that adds powers of the first generator to the list of group elements + * + * This template calls itself recursively to add powers of the first + * generator to the list of group elements. It stops if it reaches + * the identity element again. + * + * \sa enumerate_group_elements, dimino_first_step_elements + */ +template< + template class Multiply, + template class Equality, + typename id, + typename g, + typename current_element, + typename elements, + bool dont_add_current_element // = false +> +struct dimino_first_step_elements_helper : + public dimino_first_step_elements_helper< + Multiply, + Equality, + id, + g, + typename Multiply::type, + typename concat>::type, + Equality::type, id>::value + > {}; + +template< + template class Multiply, + template class Equality, + typename id, + typename g, + typename current_element, + typename elements +> +struct dimino_first_step_elements_helper +{ + typedef elements type; + constexpr static int global_flags = Equality::global_flags; +}; + +/** \internal + * + * \class dimino_first_step_elements + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Add all powers of the first generator to the list of group elements + * + * This template takes the first non-identity generator and generates the initial + * list of elements which consists of all powers of that generator. For a group + * with just one generated, it would be enumerated after this. + * + * \sa enumerate_group_elements + */ +template< + template class Multiply, + template class Equality, + typename id, + typename generators +> +struct dimino_first_step_elements +{ + typedef typename get<0, generators>::type first_generator; + typedef typename skip<1, generators>::type next_generators; + typedef type_list generators_done; + + typedef dimino_first_step_elements_helper< + Multiply, + Equality, + id, + first_generator, + first_generator, + type_list, + false + > helper; + typedef typename helper::type type; + constexpr static int global_flags = helper::global_flags; +}; + +/** \internal + * + * \class dimino_get_coset_elements + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Generate all elements of a specific coset + * + * This template generates all the elements of a specific coset by + * multiplying all elements in the given subgroup with the new + * coset representative. Note that the first element of the + * subgroup is always the identity element, so the first element of + * ther result of this template is going to be the coset + * representative itself. + * + * Note that this template accepts an additional boolean parameter + * that specifies whether to actually generate the coset (true) or + * just return an empty list (false). + * + * \sa enumerate_group_elements, dimino_add_cosets_for_rep + */ +template< + template class Multiply, + typename sub_group_elements, + typename new_coset_rep, + bool generate_coset // = true +> +struct dimino_get_coset_elements +{ + typedef typename apply_op_from_right::type type; +}; + +template< + template class Multiply, + typename sub_group_elements, + typename new_coset_rep +> +struct dimino_get_coset_elements +{ + typedef type_list<> type; +}; + +/** \internal + * + * \class dimino_add_cosets_for_rep + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Recursive template for adding coset spaces + * + * This template multiplies the coset representative with a generator + * from the list of previous generators. If the new element is not in + * the group already, it adds the corresponding coset. Finally it + * proceeds to call itself with the next generator from the list. + * + * \sa enumerate_group_elements, dimino_add_all_coset_spaces + */ +template< + template class Multiply, + template class Equality, + typename id, + typename sub_group_elements, + typename elements, + typename generators, + typename rep_element, + int sub_group_size +> +struct dimino_add_cosets_for_rep; + +template< + template class Multiply, + template class Equality, + typename id, + typename sub_group_elements, + typename elements, + typename g, + typename... gs, + typename rep_element, + int sub_group_size +> +struct dimino_add_cosets_for_rep, rep_element, sub_group_size> +{ + typedef typename Multiply::type new_coset_rep; + typedef contained_in_list_gf _cil; + constexpr static bool add_coset = !_cil::value; + + typedef typename dimino_get_coset_elements< + Multiply, + sub_group_elements, + new_coset_rep, + add_coset + >::type coset_elements; + + typedef dimino_add_cosets_for_rep< + Multiply, + Equality, + id, + sub_group_elements, + typename concat::type, + type_list, + rep_element, + sub_group_size + > _helper; + + typedef typename _helper::type type; + constexpr static int global_flags = _cil::global_flags | _helper::global_flags; + + /* Note that we don't have to update global flags here, since + * we will only add these elements if they are not part of + * the group already. But that only happens if the coset rep + * is not already in the group, so the check for the coset rep + * will catch this. + */ +}; + +template< + template class Multiply, + template class Equality, + typename id, + typename sub_group_elements, + typename elements + EIGEN_TPL_PP_SPEC_HACK_DEFC(typename, empty), + typename rep_element, + int sub_group_size +> +struct dimino_add_cosets_for_rep, rep_element, sub_group_size> +{ + typedef elements type; + constexpr static int global_flags = 0; +}; + +/** \internal + * + * \class dimino_add_all_coset_spaces + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Recursive template for adding all coset spaces for a new generator + * + * This template tries to go through the list of generators (with + * the help of the dimino_add_cosets_for_rep template) as long as + * it still finds elements that are not part of the group and add + * the corresponding cosets. + * + * \sa enumerate_group_elements, dimino_add_cosets_for_rep + */ +template< + template class Multiply, + template class Equality, + typename id, + typename sub_group_elements, + typename elements, + typename generators, + int sub_group_size, + int rep_pos, + bool stop_condition // = false +> +struct dimino_add_all_coset_spaces +{ + typedef typename get::type rep_element; + typedef dimino_add_cosets_for_rep< + Multiply, + Equality, + id, + sub_group_elements, + elements, + generators, + rep_element, + sub_group_elements::count + > _ac4r; + typedef typename _ac4r::type new_elements; + + constexpr static int new_rep_pos = rep_pos + sub_group_elements::count; + constexpr static bool new_stop_condition = new_rep_pos >= new_elements::count; + + typedef dimino_add_all_coset_spaces< + Multiply, + Equality, + id, + sub_group_elements, + new_elements, + generators, + sub_group_size, + new_rep_pos, + new_stop_condition + > _helper; + + typedef typename _helper::type type; + constexpr static int global_flags = _helper::global_flags | _ac4r::global_flags; +}; + +template< + template class Multiply, + template class Equality, + typename id, + typename sub_group_elements, + typename elements, + typename generators, + int sub_group_size, + int rep_pos +> +struct dimino_add_all_coset_spaces +{ + typedef elements type; + constexpr static int global_flags = 0; +}; + +/** \internal + * + * \class dimino_add_generator + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Enlarge the group by adding a new generator. + * + * It accepts a boolean parameter that determines if the generator is redundant, + * i.e. was already seen in the group. In that case, it reduces to a no-op. + * + * \sa enumerate_group_elements, dimino_add_all_coset_spaces + */ +template< + template class Multiply, + template class Equality, + typename id, + typename elements, + typename generators_done, + typename current_generator, + bool redundant // = false +> +struct dimino_add_generator +{ + /* this template is only called if the generator is not redundant + * => all elements of the group multiplied with the new generator + * are going to be new elements of the most trivial coset space + */ + typedef typename apply_op_from_right::type multiplied_elements; + typedef typename concat::type new_elements; + + constexpr static int rep_pos = elements::count; + + typedef dimino_add_all_coset_spaces< + Multiply, + Equality, + id, + elements, // elements of previous subgroup + new_elements, + typename concat>::type, + elements::count, // size of previous subgroup + rep_pos, + false // don't stop (because rep_pos >= new_elements::count is always false at this point) + > _helper; + typedef typename _helper::type type; + constexpr static int global_flags = _helper::global_flags; +}; + +template< + template class Multiply, + template class Equality, + typename id, + typename elements, + typename generators_done, + typename current_generator +> +struct dimino_add_generator +{ + // redundant case + typedef elements type; + constexpr static int global_flags = 0; +}; + +/** \internal + * + * \class dimino_add_remaining_generators + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Recursive template that adds all remaining generators to a group + * + * Loop through the list of generators that remain and successively + * add them to the group. + * + * \sa enumerate_group_elements, dimino_add_generator + */ +template< + template class Multiply, + template class Equality, + typename id, + typename generators_done, + typename remaining_generators, + typename elements +> +struct dimino_add_remaining_generators +{ + typedef typename get<0, remaining_generators>::type first_generator; + typedef typename skip<1, remaining_generators>::type next_generators; + + typedef contained_in_list_gf _cil; + + typedef dimino_add_generator< + Multiply, + Equality, + id, + elements, + generators_done, + first_generator, + _cil::value + > _helper; + + typedef typename _helper::type new_elements; + + typedef dimino_add_remaining_generators< + Multiply, + Equality, + id, + typename concat>::type, + next_generators, + new_elements + > _next_iter; + + typedef typename _next_iter::type type; + constexpr static int global_flags = + _cil::global_flags | + _helper::global_flags | + _next_iter::global_flags; +}; + +template< + template class Multiply, + template class Equality, + typename id, + typename generators_done, + typename elements +> +struct dimino_add_remaining_generators, elements> +{ + typedef elements type; + constexpr static int global_flags = 0; +}; + +/** \internal + * + * \class enumerate_group_elements_noid + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Helper template that implements group element enumeration + * + * This is a helper template that implements the actual enumeration + * of group elements. This has been split so that the list of + * generators can be cleansed of the identity element before + * performing the actual operation. + * + * \sa enumerate_group_elements + */ +template< + template class Multiply, + template class Equality, + typename id, + typename generators, + int initial_global_flags = 0 +> +struct enumerate_group_elements_noid +{ + typedef dimino_first_step_elements first_step; + typedef typename first_step::type first_step_elements; + + typedef dimino_add_remaining_generators< + Multiply, + Equality, + id, + typename first_step::generators_done, + typename first_step::next_generators, // remaining_generators + typename first_step::type // first_step elements + > _helper; + + typedef typename _helper::type type; + constexpr static int global_flags = + initial_global_flags | + first_step::global_flags | + _helper::global_flags; +}; + +// in case when no generators are specified +template< + template class Multiply, + template class Equality, + typename id, + int initial_global_flags +> +struct enumerate_group_elements_noid, initial_global_flags> +{ + typedef type_list type; + constexpr static int global_flags = initial_global_flags; +}; + +/** \internal + * + * \class enumerate_group_elements + * \ingroup CXX11_TensorSymmetry_Module + * + * \brief Enumerate all elements in a finite group + * + * This template enumerates all elements in a finite group. It accepts + * the following template parameters: + * + * \tparam Multiply The multiplication operation that multiplies two group elements + * with each other. + * \tparam Equality The equality check operation that checks if two group elements + * are equal to another. + * \tparam id The identity element + * \tparam _generators A list of (possibly redundant) generators of the group + */ +template< + template class Multiply, + template class Equality, + typename id, + typename _generators +> +struct enumerate_group_elements + : public enumerate_group_elements_noid< + Multiply, + Equality, + id, + typename strip_identities::type, + strip_identities::global_flags + > +{ +}; + +} // end namespace group_theory + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSORSYMMETRY_TEMPLATEGROUPTHEORY_H + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 61e8f56fd..0a6c56c19 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -100,4 +100,5 @@ if(EIGEN_TEST_CXX11) # clash there) ei_add_test(cxx11_meta "-std=c++0x") ei_add_test(cxx11_tensor_simple "-std=c++0x") + ei_add_test(cxx11_tensor_symmetry "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_symmetry.cpp b/unsupported/test/cxx11_tensor_symmetry.cpp new file mode 100644 index 000000000..ebbc8a68d --- /dev/null +++ b/unsupported/test/cxx11_tensor_symmetry.cpp @@ -0,0 +1,819 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2006-2008 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include +#include + +#include +#include + +using Eigen::Tensor; +using Eigen::SGroup; +using Eigen::DynamicSGroup; +using Eigen::StaticSGroup; +using Eigen::Symmetry; +using Eigen::AntiSymmetry; +using Eigen::Hermiticity; +using Eigen::AntiHermiticity; + +using Eigen::NegationFlag; +using Eigen::ConjugationFlag; +using Eigen::GlobalZeroFlag; +using Eigen::GlobalRealFlag; +using Eigen::GlobalImagFlag; + +// helper function to determine if the compiler intantiated a static +// or dynamic symmetry group +template +bool isDynGroup(StaticSGroup const& dummy) +{ + (void)dummy; + return false; +} + +bool isDynGroup(DynamicSGroup const& dummy) +{ + (void)dummy; + return true; +} + +// helper class for checking that the symmetry groups are correct +struct checkIdx { + template + static inline int doCheck_(ArrType e, int flags, int dummy, std::set& found, std::map const& expected) + { + // use decimal representation of value + uint64_t value = e[0]; + for (std::size_t i = 1; i < e.size(); i++) + value = value * 10 + e[i]; + + // we want to make sure that we find each element + auto it = expected.find(value); + VERIFY((it != expected.end())); + VERIFY_IS_EQUAL(it->second, flags); + + // we want to make sure we only have each element once; + // set::insert returns true for the second part of the pair + // if the element was really inserted and not already there + auto p = found.insert(value); + VERIFY((p.second)); + + return dummy; + } + + static inline int run(std::vector e, int flags, int dummy, std::set& found, std::map const& expected) + { + return doCheck_(e, flags, dummy, found, expected); + } + + template + static inline int run(std::array e, int flags, int dummy, std::set& found, std::map const& expected) + { + return doCheck_(e, flags, dummy, found, expected); + } +}; + +static void test_symgroups_static() +{ + std::array identity{{0,1,2,3,4,5,6}}; + + // Simple static symmetry group + StaticSGroup<7, + AntiSymmetry<0,1>, + Hermiticity<0,2> + > group; + + std::set found; + std::map expected; + expected[ 123456] = 0; + expected[1023456] = NegationFlag; + expected[2103456] = ConjugationFlag; + expected[1203456] = ConjugationFlag | NegationFlag; + expected[2013456] = ConjugationFlag | NegationFlag; + expected[ 213456] = ConjugationFlag; + + VERIFY_IS_EQUAL(group.size(), 6u); + VERIFY_IS_EQUAL(group.globalFlags(), GlobalImagFlag); + group.apply(identity, 0, found, expected); + VERIFY_IS_EQUAL(found.size(), 6u); +} + +static void test_symgroups_dynamic() +{ + std::vector identity; + for (int i = 0; i <= 6; i++) + identity.push_back(i); + + // Simple dynamic symmetry group + DynamicSGroup group(7); + group.add(0,1,NegationFlag); + group.add(0,2,ConjugationFlag); + + VERIFY_IS_EQUAL(group.size(), 6u); + VERIFY_IS_EQUAL(group.globalFlags(), GlobalImagFlag); + + std::set found; + std::map expected; + expected[ 123456] = 0; + expected[1023456] = NegationFlag; + expected[2103456] = ConjugationFlag; + expected[1203456] = ConjugationFlag | NegationFlag; + expected[2013456] = ConjugationFlag | NegationFlag; + expected[ 213456] = ConjugationFlag; + + VERIFY_IS_EQUAL(group.size(), 6u); + VERIFY_IS_EQUAL(group.globalFlags(), GlobalImagFlag); + group.apply(identity, 0, found, expected); + VERIFY_IS_EQUAL(found.size(), 6u); +} + +static void test_symgroups_selection() +{ + std::array identity7{{0,1,2,3,4,5,6}}; + std::array identity10{{0,1,2,3,4,5,6,7,8,9}}; + + { + // Do the same test as in test_symgroups_static but + // require selection via SGroup + SGroup<7, + AntiSymmetry<0,1>, + Hermiticity<0,2> + > group; + + std::set found; + std::map expected; + expected[ 123456] = 0; + expected[1023456] = NegationFlag; + expected[2103456] = ConjugationFlag; + expected[1203456] = ConjugationFlag | NegationFlag; + expected[2013456] = ConjugationFlag | NegationFlag; + expected[ 213456] = ConjugationFlag; + + VERIFY(!isDynGroup(group)); + VERIFY_IS_EQUAL(group.size(), 6u); + VERIFY_IS_EQUAL(group.globalFlags(), GlobalImagFlag); + group.apply(identity7, 0, found, expected); + VERIFY_IS_EQUAL(found.size(), 6u); + } + + { + // simple factorizing group: 5 generators, 2^5 = 32 elements + // selection should make this dynamic, although static group + // can still be reasonably generated + SGroup<10, + Symmetry<0,1>, + Symmetry<2,3>, + Symmetry<4,5>, + Symmetry<6,7>, + Symmetry<8,9> + > group; + + std::set found; + std::map expected; + expected[ 123456789] = 0; expected[ 123456798] = 0; expected[ 123457689] = 0; expected[ 123457698] = 0; + expected[ 123546789] = 0; expected[ 123546798] = 0; expected[ 123547689] = 0; expected[ 123547698] = 0; + expected[ 132456789] = 0; expected[ 132456798] = 0; expected[ 132457689] = 0; expected[ 132457698] = 0; + expected[ 132546789] = 0; expected[ 132546798] = 0; expected[ 132547689] = 0; expected[ 132547698] = 0; + expected[1023456789] = 0; expected[1023456798] = 0; expected[1023457689] = 0; expected[1023457698] = 0; + expected[1023546789] = 0; expected[1023546798] = 0; expected[1023547689] = 0; expected[1023547698] = 0; + expected[1032456789] = 0; expected[1032456798] = 0; expected[1032457689] = 0; expected[1032457698] = 0; + expected[1032546789] = 0; expected[1032546798] = 0; expected[1032547689] = 0; expected[1032547698] = 0; + + VERIFY(isDynGroup(group)); + VERIFY_IS_EQUAL(group.size(), 32u); + VERIFY_IS_EQUAL(group.globalFlags(), 0); + group.apply(identity10, 0, found, expected); + VERIFY_IS_EQUAL(found.size(), 32u); + + // no verify that we could also generate a static group + // with these generators + found.clear(); + StaticSGroup<10, + Symmetry<0,1>, + Symmetry<2,3>, + Symmetry<4,5>, + Symmetry<6,7>, + Symmetry<8,9> + > group_static; + VERIFY_IS_EQUAL(group_static.size(), 32u); + VERIFY_IS_EQUAL(group_static.globalFlags(), 0); + group_static.apply(identity10, 0, found, expected); + VERIFY_IS_EQUAL(found.size(), 32u); + } + + { + // try to create a HUGE group + SGroup<7, + Symmetry<0,1>, + Symmetry<1,2>, + Symmetry<2,3>, + Symmetry<3,4>, + Symmetry<4,5>, + Symmetry<5,6> + > group; + + std::set found; + uint64_t pre_expected[5040] = { + 123456, 1023456, 213456, 2013456, 1203456, 2103456, 132456, 1032456, 312456, 3012456, 1302456, 3102456, + 231456, 2031456, 321456, 3021456, 2301456, 3201456, 1230456, 2130456, 1320456, 3120456, 2310456, 3210456, + 124356, 1024356, 214356, 2014356, 1204356, 2104356, 142356, 1042356, 412356, 4012356, 1402356, 4102356, + 241356, 2041356, 421356, 4021356, 2401356, 4201356, 1240356, 2140356, 1420356, 4120356, 2410356, 4210356, + 134256, 1034256, 314256, 3014256, 1304256, 3104256, 143256, 1043256, 413256, 4013256, 1403256, 4103256, + 341256, 3041256, 431256, 4031256, 3401256, 4301256, 1340256, 3140256, 1430256, 4130256, 3410256, 4310256, + 234156, 2034156, 324156, 3024156, 2304156, 3204156, 243156, 2043156, 423156, 4023156, 2403156, 4203156, + 342156, 3042156, 432156, 4032156, 3402156, 4302156, 2340156, 3240156, 2430156, 4230156, 3420156, 4320156, + 1234056, 2134056, 1324056, 3124056, 2314056, 3214056, 1243056, 2143056, 1423056, 4123056, 2413056, 4213056, + 1342056, 3142056, 1432056, 4132056, 3412056, 4312056, 2341056, 3241056, 2431056, 4231056, 3421056, 4321056, + 123546, 1023546, 213546, 2013546, 1203546, 2103546, 132546, 1032546, 312546, 3012546, 1302546, 3102546, + 231546, 2031546, 321546, 3021546, 2301546, 3201546, 1230546, 2130546, 1320546, 3120546, 2310546, 3210546, + 125346, 1025346, 215346, 2015346, 1205346, 2105346, 152346, 1052346, 512346, 5012346, 1502346, 5102346, + 251346, 2051346, 521346, 5021346, 2501346, 5201346, 1250346, 2150346, 1520346, 5120346, 2510346, 5210346, + 135246, 1035246, 315246, 3015246, 1305246, 3105246, 153246, 1053246, 513246, 5013246, 1503246, 5103246, + 351246, 3051246, 531246, 5031246, 3501246, 5301246, 1350246, 3150246, 1530246, 5130246, 3510246, 5310246, + 235146, 2035146, 325146, 3025146, 2305146, 3205146, 253146, 2053146, 523146, 5023146, 2503146, 5203146, + 352146, 3052146, 532146, 5032146, 3502146, 5302146, 2350146, 3250146, 2530146, 5230146, 3520146, 5320146, + 1235046, 2135046, 1325046, 3125046, 2315046, 3215046, 1253046, 2153046, 1523046, 5123046, 2513046, 5213046, + 1352046, 3152046, 1532046, 5132046, 3512046, 5312046, 2351046, 3251046, 2531046, 5231046, 3521046, 5321046, + 124536, 1024536, 214536, 2014536, 1204536, 2104536, 142536, 1042536, 412536, 4012536, 1402536, 4102536, + 241536, 2041536, 421536, 4021536, 2401536, 4201536, 1240536, 2140536, 1420536, 4120536, 2410536, 4210536, + 125436, 1025436, 215436, 2015436, 1205436, 2105436, 152436, 1052436, 512436, 5012436, 1502436, 5102436, + 251436, 2051436, 521436, 5021436, 2501436, 5201436, 1250436, 2150436, 1520436, 5120436, 2510436, 5210436, + 145236, 1045236, 415236, 4015236, 1405236, 4105236, 154236, 1054236, 514236, 5014236, 1504236, 5104236, + 451236, 4051236, 541236, 5041236, 4501236, 5401236, 1450236, 4150236, 1540236, 5140236, 4510236, 5410236, + 245136, 2045136, 425136, 4025136, 2405136, 4205136, 254136, 2054136, 524136, 5024136, 2504136, 5204136, + 452136, 4052136, 542136, 5042136, 4502136, 5402136, 2450136, 4250136, 2540136, 5240136, 4520136, 5420136, + 1245036, 2145036, 1425036, 4125036, 2415036, 4215036, 1254036, 2154036, 1524036, 5124036, 2514036, 5214036, + 1452036, 4152036, 1542036, 5142036, 4512036, 5412036, 2451036, 4251036, 2541036, 5241036, 4521036, 5421036, + 134526, 1034526, 314526, 3014526, 1304526, 3104526, 143526, 1043526, 413526, 4013526, 1403526, 4103526, + 341526, 3041526, 431526, 4031526, 3401526, 4301526, 1340526, 3140526, 1430526, 4130526, 3410526, 4310526, + 135426, 1035426, 315426, 3015426, 1305426, 3105426, 153426, 1053426, 513426, 5013426, 1503426, 5103426, + 351426, 3051426, 531426, 5031426, 3501426, 5301426, 1350426, 3150426, 1530426, 5130426, 3510426, 5310426, + 145326, 1045326, 415326, 4015326, 1405326, 4105326, 154326, 1054326, 514326, 5014326, 1504326, 5104326, + 451326, 4051326, 541326, 5041326, 4501326, 5401326, 1450326, 4150326, 1540326, 5140326, 4510326, 5410326, + 345126, 3045126, 435126, 4035126, 3405126, 4305126, 354126, 3054126, 534126, 5034126, 3504126, 5304126, + 453126, 4053126, 543126, 5043126, 4503126, 5403126, 3450126, 4350126, 3540126, 5340126, 4530126, 5430126, + 1345026, 3145026, 1435026, 4135026, 3415026, 4315026, 1354026, 3154026, 1534026, 5134026, 3514026, 5314026, + 1453026, 4153026, 1543026, 5143026, 4513026, 5413026, 3451026, 4351026, 3541026, 5341026, 4531026, 5431026, + 234516, 2034516, 324516, 3024516, 2304516, 3204516, 243516, 2043516, 423516, 4023516, 2403516, 4203516, + 342516, 3042516, 432516, 4032516, 3402516, 4302516, 2340516, 3240516, 2430516, 4230516, 3420516, 4320516, + 235416, 2035416, 325416, 3025416, 2305416, 3205416, 253416, 2053416, 523416, 5023416, 2503416, 5203416, + 352416, 3052416, 532416, 5032416, 3502416, 5302416, 2350416, 3250416, 2530416, 5230416, 3520416, 5320416, + 245316, 2045316, 425316, 4025316, 2405316, 4205316, 254316, 2054316, 524316, 5024316, 2504316, 5204316, + 452316, 4052316, 542316, 5042316, 4502316, 5402316, 2450316, 4250316, 2540316, 5240316, 4520316, 5420316, + 345216, 3045216, 435216, 4035216, 3405216, 4305216, 354216, 3054216, 534216, 5034216, 3504216, 5304216, + 453216, 4053216, 543216, 5043216, 4503216, 5403216, 3450216, 4350216, 3540216, 5340216, 4530216, 5430216, + 2345016, 3245016, 2435016, 4235016, 3425016, 4325016, 2354016, 3254016, 2534016, 5234016, 3524016, 5324016, + 2453016, 4253016, 2543016, 5243016, 4523016, 5423016, 3452016, 4352016, 3542016, 5342016, 4532016, 5432016, + 1234506, 2134506, 1324506, 3124506, 2314506, 3214506, 1243506, 2143506, 1423506, 4123506, 2413506, 4213506, + 1342506, 3142506, 1432506, 4132506, 3412506, 4312506, 2341506, 3241506, 2431506, 4231506, 3421506, 4321506, + 1235406, 2135406, 1325406, 3125406, 2315406, 3215406, 1253406, 2153406, 1523406, 5123406, 2513406, 5213406, + 1352406, 3152406, 1532406, 5132406, 3512406, 5312406, 2351406, 3251406, 2531406, 5231406, 3521406, 5321406, + 1245306, 2145306, 1425306, 4125306, 2415306, 4215306, 1254306, 2154306, 1524306, 5124306, 2514306, 5214306, + 1452306, 4152306, 1542306, 5142306, 4512306, 5412306, 2451306, 4251306, 2541306, 5241306, 4521306, 5421306, + 1345206, 3145206, 1435206, 4135206, 3415206, 4315206, 1354206, 3154206, 1534206, 5134206, 3514206, 5314206, + 1453206, 4153206, 1543206, 5143206, 4513206, 5413206, 3451206, 4351206, 3541206, 5341206, 4531206, 5431206, + 2345106, 3245106, 2435106, 4235106, 3425106, 4325106, 2354106, 3254106, 2534106, 5234106, 3524106, 5324106, + 2453106, 4253106, 2543106, 5243106, 4523106, 5423106, 3452106, 4352106, 3542106, 5342106, 4532106, 5432106, + 123465, 1023465, 213465, 2013465, 1203465, 2103465, 132465, 1032465, 312465, 3012465, 1302465, 3102465, + 231465, 2031465, 321465, 3021465, 2301465, 3201465, 1230465, 2130465, 1320465, 3120465, 2310465, 3210465, + 124365, 1024365, 214365, 2014365, 1204365, 2104365, 142365, 1042365, 412365, 4012365, 1402365, 4102365, + 241365, 2041365, 421365, 4021365, 2401365, 4201365, 1240365, 2140365, 1420365, 4120365, 2410365, 4210365, + 134265, 1034265, 314265, 3014265, 1304265, 3104265, 143265, 1043265, 413265, 4013265, 1403265, 4103265, + 341265, 3041265, 431265, 4031265, 3401265, 4301265, 1340265, 3140265, 1430265, 4130265, 3410265, 4310265, + 234165, 2034165, 324165, 3024165, 2304165, 3204165, 243165, 2043165, 423165, 4023165, 2403165, 4203165, + 342165, 3042165, 432165, 4032165, 3402165, 4302165, 2340165, 3240165, 2430165, 4230165, 3420165, 4320165, + 1234065, 2134065, 1324065, 3124065, 2314065, 3214065, 1243065, 2143065, 1423065, 4123065, 2413065, 4213065, + 1342065, 3142065, 1432065, 4132065, 3412065, 4312065, 2341065, 3241065, 2431065, 4231065, 3421065, 4321065, + 123645, 1023645, 213645, 2013645, 1203645, 2103645, 132645, 1032645, 312645, 3012645, 1302645, 3102645, + 231645, 2031645, 321645, 3021645, 2301645, 3201645, 1230645, 2130645, 1320645, 3120645, 2310645, 3210645, + 126345, 1026345, 216345, 2016345, 1206345, 2106345, 162345, 1062345, 612345, 6012345, 1602345, 6102345, + 261345, 2061345, 621345, 6021345, 2601345, 6201345, 1260345, 2160345, 1620345, 6120345, 2610345, 6210345, + 136245, 1036245, 316245, 3016245, 1306245, 3106245, 163245, 1063245, 613245, 6013245, 1603245, 6103245, + 361245, 3061245, 631245, 6031245, 3601245, 6301245, 1360245, 3160245, 1630245, 6130245, 3610245, 6310245, + 236145, 2036145, 326145, 3026145, 2306145, 3206145, 263145, 2063145, 623145, 6023145, 2603145, 6203145, + 362145, 3062145, 632145, 6032145, 3602145, 6302145, 2360145, 3260145, 2630145, 6230145, 3620145, 6320145, + 1236045, 2136045, 1326045, 3126045, 2316045, 3216045, 1263045, 2163045, 1623045, 6123045, 2613045, 6213045, + 1362045, 3162045, 1632045, 6132045, 3612045, 6312045, 2361045, 3261045, 2631045, 6231045, 3621045, 6321045, + 124635, 1024635, 214635, 2014635, 1204635, 2104635, 142635, 1042635, 412635, 4012635, 1402635, 4102635, + 241635, 2041635, 421635, 4021635, 2401635, 4201635, 1240635, 2140635, 1420635, 4120635, 2410635, 4210635, + 126435, 1026435, 216435, 2016435, 1206435, 2106435, 162435, 1062435, 612435, 6012435, 1602435, 6102435, + 261435, 2061435, 621435, 6021435, 2601435, 6201435, 1260435, 2160435, 1620435, 6120435, 2610435, 6210435, + 146235, 1046235, 416235, 4016235, 1406235, 4106235, 164235, 1064235, 614235, 6014235, 1604235, 6104235, + 461235, 4061235, 641235, 6041235, 4601235, 6401235, 1460235, 4160235, 1640235, 6140235, 4610235, 6410235, + 246135, 2046135, 426135, 4026135, 2406135, 4206135, 264135, 2064135, 624135, 6024135, 2604135, 6204135, + 462135, 4062135, 642135, 6042135, 4602135, 6402135, 2460135, 4260135, 2640135, 6240135, 4620135, 6420135, + 1246035, 2146035, 1426035, 4126035, 2416035, 4216035, 1264035, 2164035, 1624035, 6124035, 2614035, 6214035, + 1462035, 4162035, 1642035, 6142035, 4612035, 6412035, 2461035, 4261035, 2641035, 6241035, 4621035, 6421035, + 134625, 1034625, 314625, 3014625, 1304625, 3104625, 143625, 1043625, 413625, 4013625, 1403625, 4103625, + 341625, 3041625, 431625, 4031625, 3401625, 4301625, 1340625, 3140625, 1430625, 4130625, 3410625, 4310625, + 136425, 1036425, 316425, 3016425, 1306425, 3106425, 163425, 1063425, 613425, 6013425, 1603425, 6103425, + 361425, 3061425, 631425, 6031425, 3601425, 6301425, 1360425, 3160425, 1630425, 6130425, 3610425, 6310425, + 146325, 1046325, 416325, 4016325, 1406325, 4106325, 164325, 1064325, 614325, 6014325, 1604325, 6104325, + 461325, 4061325, 641325, 6041325, 4601325, 6401325, 1460325, 4160325, 1640325, 6140325, 4610325, 6410325, + 346125, 3046125, 436125, 4036125, 3406125, 4306125, 364125, 3064125, 634125, 6034125, 3604125, 6304125, + 463125, 4063125, 643125, 6043125, 4603125, 6403125, 3460125, 4360125, 3640125, 6340125, 4630125, 6430125, + 1346025, 3146025, 1436025, 4136025, 3416025, 4316025, 1364025, 3164025, 1634025, 6134025, 3614025, 6314025, + 1463025, 4163025, 1643025, 6143025, 4613025, 6413025, 3461025, 4361025, 3641025, 6341025, 4631025, 6431025, + 234615, 2034615, 324615, 3024615, 2304615, 3204615, 243615, 2043615, 423615, 4023615, 2403615, 4203615, + 342615, 3042615, 432615, 4032615, 3402615, 4302615, 2340615, 3240615, 2430615, 4230615, 3420615, 4320615, + 236415, 2036415, 326415, 3026415, 2306415, 3206415, 263415, 2063415, 623415, 6023415, 2603415, 6203415, + 362415, 3062415, 632415, 6032415, 3602415, 6302415, 2360415, 3260415, 2630415, 6230415, 3620415, 6320415, + 246315, 2046315, 426315, 4026315, 2406315, 4206315, 264315, 2064315, 624315, 6024315, 2604315, 6204315, + 462315, 4062315, 642315, 6042315, 4602315, 6402315, 2460315, 4260315, 2640315, 6240315, 4620315, 6420315, + 346215, 3046215, 436215, 4036215, 3406215, 4306215, 364215, 3064215, 634215, 6034215, 3604215, 6304215, + 463215, 4063215, 643215, 6043215, 4603215, 6403215, 3460215, 4360215, 3640215, 6340215, 4630215, 6430215, + 2346015, 3246015, 2436015, 4236015, 3426015, 4326015, 2364015, 3264015, 2634015, 6234015, 3624015, 6324015, + 2463015, 4263015, 2643015, 6243015, 4623015, 6423015, 3462015, 4362015, 3642015, 6342015, 4632015, 6432015, + 1234605, 2134605, 1324605, 3124605, 2314605, 3214605, 1243605, 2143605, 1423605, 4123605, 2413605, 4213605, + 1342605, 3142605, 1432605, 4132605, 3412605, 4312605, 2341605, 3241605, 2431605, 4231605, 3421605, 4321605, + 1236405, 2136405, 1326405, 3126405, 2316405, 3216405, 1263405, 2163405, 1623405, 6123405, 2613405, 6213405, + 1362405, 3162405, 1632405, 6132405, 3612405, 6312405, 2361405, 3261405, 2631405, 6231405, 3621405, 6321405, + 1246305, 2146305, 1426305, 4126305, 2416305, 4216305, 1264305, 2164305, 1624305, 6124305, 2614305, 6214305, + 1462305, 4162305, 1642305, 6142305, 4612305, 6412305, 2461305, 4261305, 2641305, 6241305, 4621305, 6421305, + 1346205, 3146205, 1436205, 4136205, 3416205, 4316205, 1364205, 3164205, 1634205, 6134205, 3614205, 6314205, + 1463205, 4163205, 1643205, 6143205, 4613205, 6413205, 3461205, 4361205, 3641205, 6341205, 4631205, 6431205, + 2346105, 3246105, 2436105, 4236105, 3426105, 4326105, 2364105, 3264105, 2634105, 6234105, 3624105, 6324105, + 2463105, 4263105, 2643105, 6243105, 4623105, 6423105, 3462105, 4362105, 3642105, 6342105, 4632105, 6432105, + 123564, 1023564, 213564, 2013564, 1203564, 2103564, 132564, 1032564, 312564, 3012564, 1302564, 3102564, + 231564, 2031564, 321564, 3021564, 2301564, 3201564, 1230564, 2130564, 1320564, 3120564, 2310564, 3210564, + 125364, 1025364, 215364, 2015364, 1205364, 2105364, 152364, 1052364, 512364, 5012364, 1502364, 5102364, + 251364, 2051364, 521364, 5021364, 2501364, 5201364, 1250364, 2150364, 1520364, 5120364, 2510364, 5210364, + 135264, 1035264, 315264, 3015264, 1305264, 3105264, 153264, 1053264, 513264, 5013264, 1503264, 5103264, + 351264, 3051264, 531264, 5031264, 3501264, 5301264, 1350264, 3150264, 1530264, 5130264, 3510264, 5310264, + 235164, 2035164, 325164, 3025164, 2305164, 3205164, 253164, 2053164, 523164, 5023164, 2503164, 5203164, + 352164, 3052164, 532164, 5032164, 3502164, 5302164, 2350164, 3250164, 2530164, 5230164, 3520164, 5320164, + 1235064, 2135064, 1325064, 3125064, 2315064, 3215064, 1253064, 2153064, 1523064, 5123064, 2513064, 5213064, + 1352064, 3152064, 1532064, 5132064, 3512064, 5312064, 2351064, 3251064, 2531064, 5231064, 3521064, 5321064, + 123654, 1023654, 213654, 2013654, 1203654, 2103654, 132654, 1032654, 312654, 3012654, 1302654, 3102654, + 231654, 2031654, 321654, 3021654, 2301654, 3201654, 1230654, 2130654, 1320654, 3120654, 2310654, 3210654, + 126354, 1026354, 216354, 2016354, 1206354, 2106354, 162354, 1062354, 612354, 6012354, 1602354, 6102354, + 261354, 2061354, 621354, 6021354, 2601354, 6201354, 1260354, 2160354, 1620354, 6120354, 2610354, 6210354, + 136254, 1036254, 316254, 3016254, 1306254, 3106254, 163254, 1063254, 613254, 6013254, 1603254, 6103254, + 361254, 3061254, 631254, 6031254, 3601254, 6301254, 1360254, 3160254, 1630254, 6130254, 3610254, 6310254, + 236154, 2036154, 326154, 3026154, 2306154, 3206154, 263154, 2063154, 623154, 6023154, 2603154, 6203154, + 362154, 3062154, 632154, 6032154, 3602154, 6302154, 2360154, 3260154, 2630154, 6230154, 3620154, 6320154, + 1236054, 2136054, 1326054, 3126054, 2316054, 3216054, 1263054, 2163054, 1623054, 6123054, 2613054, 6213054, + 1362054, 3162054, 1632054, 6132054, 3612054, 6312054, 2361054, 3261054, 2631054, 6231054, 3621054, 6321054, + 125634, 1025634, 215634, 2015634, 1205634, 2105634, 152634, 1052634, 512634, 5012634, 1502634, 5102634, + 251634, 2051634, 521634, 5021634, 2501634, 5201634, 1250634, 2150634, 1520634, 5120634, 2510634, 5210634, + 126534, 1026534, 216534, 2016534, 1206534, 2106534, 162534, 1062534, 612534, 6012534, 1602534, 6102534, + 261534, 2061534, 621534, 6021534, 2601534, 6201534, 1260534, 2160534, 1620534, 6120534, 2610534, 6210534, + 156234, 1056234, 516234, 5016234, 1506234, 5106234, 165234, 1065234, 615234, 6015234, 1605234, 6105234, + 561234, 5061234, 651234, 6051234, 5601234, 6501234, 1560234, 5160234, 1650234, 6150234, 5610234, 6510234, + 256134, 2056134, 526134, 5026134, 2506134, 5206134, 265134, 2065134, 625134, 6025134, 2605134, 6205134, + 562134, 5062134, 652134, 6052134, 5602134, 6502134, 2560134, 5260134, 2650134, 6250134, 5620134, 6520134, + 1256034, 2156034, 1526034, 5126034, 2516034, 5216034, 1265034, 2165034, 1625034, 6125034, 2615034, 6215034, + 1562034, 5162034, 1652034, 6152034, 5612034, 6512034, 2561034, 5261034, 2651034, 6251034, 5621034, 6521034, + 135624, 1035624, 315624, 3015624, 1305624, 3105624, 153624, 1053624, 513624, 5013624, 1503624, 5103624, + 351624, 3051624, 531624, 5031624, 3501624, 5301624, 1350624, 3150624, 1530624, 5130624, 3510624, 5310624, + 136524, 1036524, 316524, 3016524, 1306524, 3106524, 163524, 1063524, 613524, 6013524, 1603524, 6103524, + 361524, 3061524, 631524, 6031524, 3601524, 6301524, 1360524, 3160524, 1630524, 6130524, 3610524, 6310524, + 156324, 1056324, 516324, 5016324, 1506324, 5106324, 165324, 1065324, 615324, 6015324, 1605324, 6105324, + 561324, 5061324, 651324, 6051324, 5601324, 6501324, 1560324, 5160324, 1650324, 6150324, 5610324, 6510324, + 356124, 3056124, 536124, 5036124, 3506124, 5306124, 365124, 3065124, 635124, 6035124, 3605124, 6305124, + 563124, 5063124, 653124, 6053124, 5603124, 6503124, 3560124, 5360124, 3650124, 6350124, 5630124, 6530124, + 1356024, 3156024, 1536024, 5136024, 3516024, 5316024, 1365024, 3165024, 1635024, 6135024, 3615024, 6315024, + 1563024, 5163024, 1653024, 6153024, 5613024, 6513024, 3561024, 5361024, 3651024, 6351024, 5631024, 6531024, + 235614, 2035614, 325614, 3025614, 2305614, 3205614, 253614, 2053614, 523614, 5023614, 2503614, 5203614, + 352614, 3052614, 532614, 5032614, 3502614, 5302614, 2350614, 3250614, 2530614, 5230614, 3520614, 5320614, + 236514, 2036514, 326514, 3026514, 2306514, 3206514, 263514, 2063514, 623514, 6023514, 2603514, 6203514, + 362514, 3062514, 632514, 6032514, 3602514, 6302514, 2360514, 3260514, 2630514, 6230514, 3620514, 6320514, + 256314, 2056314, 526314, 5026314, 2506314, 5206314, 265314, 2065314, 625314, 6025314, 2605314, 6205314, + 562314, 5062314, 652314, 6052314, 5602314, 6502314, 2560314, 5260314, 2650314, 6250314, 5620314, 6520314, + 356214, 3056214, 536214, 5036214, 3506214, 5306214, 365214, 3065214, 635214, 6035214, 3605214, 6305214, + 563214, 5063214, 653214, 6053214, 5603214, 6503214, 3560214, 5360214, 3650214, 6350214, 5630214, 6530214, + 2356014, 3256014, 2536014, 5236014, 3526014, 5326014, 2365014, 3265014, 2635014, 6235014, 3625014, 6325014, + 2563014, 5263014, 2653014, 6253014, 5623014, 6523014, 3562014, 5362014, 3652014, 6352014, 5632014, 6532014, + 1235604, 2135604, 1325604, 3125604, 2315604, 3215604, 1253604, 2153604, 1523604, 5123604, 2513604, 5213604, + 1352604, 3152604, 1532604, 5132604, 3512604, 5312604, 2351604, 3251604, 2531604, 5231604, 3521604, 5321604, + 1236504, 2136504, 1326504, 3126504, 2316504, 3216504, 1263504, 2163504, 1623504, 6123504, 2613504, 6213504, + 1362504, 3162504, 1632504, 6132504, 3612504, 6312504, 2361504, 3261504, 2631504, 6231504, 3621504, 6321504, + 1256304, 2156304, 1526304, 5126304, 2516304, 5216304, 1265304, 2165304, 1625304, 6125304, 2615304, 6215304, + 1562304, 5162304, 1652304, 6152304, 5612304, 6512304, 2561304, 5261304, 2651304, 6251304, 5621304, 6521304, + 1356204, 3156204, 1536204, 5136204, 3516204, 5316204, 1365204, 3165204, 1635204, 6135204, 3615204, 6315204, + 1563204, 5163204, 1653204, 6153204, 5613204, 6513204, 3561204, 5361204, 3651204, 6351204, 5631204, 6531204, + 2356104, 3256104, 2536104, 5236104, 3526104, 5326104, 2365104, 3265104, 2635104, 6235104, 3625104, 6325104, + 2563104, 5263104, 2653104, 6253104, 5623104, 6523104, 3562104, 5362104, 3652104, 6352104, 5632104, 6532104, + 124563, 1024563, 214563, 2014563, 1204563, 2104563, 142563, 1042563, 412563, 4012563, 1402563, 4102563, + 241563, 2041563, 421563, 4021563, 2401563, 4201563, 1240563, 2140563, 1420563, 4120563, 2410563, 4210563, + 125463, 1025463, 215463, 2015463, 1205463, 2105463, 152463, 1052463, 512463, 5012463, 1502463, 5102463, + 251463, 2051463, 521463, 5021463, 2501463, 5201463, 1250463, 2150463, 1520463, 5120463, 2510463, 5210463, + 145263, 1045263, 415263, 4015263, 1405263, 4105263, 154263, 1054263, 514263, 5014263, 1504263, 5104263, + 451263, 4051263, 541263, 5041263, 4501263, 5401263, 1450263, 4150263, 1540263, 5140263, 4510263, 5410263, + 245163, 2045163, 425163, 4025163, 2405163, 4205163, 254163, 2054163, 524163, 5024163, 2504163, 5204163, + 452163, 4052163, 542163, 5042163, 4502163, 5402163, 2450163, 4250163, 2540163, 5240163, 4520163, 5420163, + 1245063, 2145063, 1425063, 4125063, 2415063, 4215063, 1254063, 2154063, 1524063, 5124063, 2514063, 5214063, + 1452063, 4152063, 1542063, 5142063, 4512063, 5412063, 2451063, 4251063, 2541063, 5241063, 4521063, 5421063, + 124653, 1024653, 214653, 2014653, 1204653, 2104653, 142653, 1042653, 412653, 4012653, 1402653, 4102653, + 241653, 2041653, 421653, 4021653, 2401653, 4201653, 1240653, 2140653, 1420653, 4120653, 2410653, 4210653, + 126453, 1026453, 216453, 2016453, 1206453, 2106453, 162453, 1062453, 612453, 6012453, 1602453, 6102453, + 261453, 2061453, 621453, 6021453, 2601453, 6201453, 1260453, 2160453, 1620453, 6120453, 2610453, 6210453, + 146253, 1046253, 416253, 4016253, 1406253, 4106253, 164253, 1064253, 614253, 6014253, 1604253, 6104253, + 461253, 4061253, 641253, 6041253, 4601253, 6401253, 1460253, 4160253, 1640253, 6140253, 4610253, 6410253, + 246153, 2046153, 426153, 4026153, 2406153, 4206153, 264153, 2064153, 624153, 6024153, 2604153, 6204153, + 462153, 4062153, 642153, 6042153, 4602153, 6402153, 2460153, 4260153, 2640153, 6240153, 4620153, 6420153, + 1246053, 2146053, 1426053, 4126053, 2416053, 4216053, 1264053, 2164053, 1624053, 6124053, 2614053, 6214053, + 1462053, 4162053, 1642053, 6142053, 4612053, 6412053, 2461053, 4261053, 2641053, 6241053, 4621053, 6421053, + 125643, 1025643, 215643, 2015643, 1205643, 2105643, 152643, 1052643, 512643, 5012643, 1502643, 5102643, + 251643, 2051643, 521643, 5021643, 2501643, 5201643, 1250643, 2150643, 1520643, 5120643, 2510643, 5210643, + 126543, 1026543, 216543, 2016543, 1206543, 2106543, 162543, 1062543, 612543, 6012543, 1602543, 6102543, + 261543, 2061543, 621543, 6021543, 2601543, 6201543, 1260543, 2160543, 1620543, 6120543, 2610543, 6210543, + 156243, 1056243, 516243, 5016243, 1506243, 5106243, 165243, 1065243, 615243, 6015243, 1605243, 6105243, + 561243, 5061243, 651243, 6051243, 5601243, 6501243, 1560243, 5160243, 1650243, 6150243, 5610243, 6510243, + 256143, 2056143, 526143, 5026143, 2506143, 5206143, 265143, 2065143, 625143, 6025143, 2605143, 6205143, + 562143, 5062143, 652143, 6052143, 5602143, 6502143, 2560143, 5260143, 2650143, 6250143, 5620143, 6520143, + 1256043, 2156043, 1526043, 5126043, 2516043, 5216043, 1265043, 2165043, 1625043, 6125043, 2615043, 6215043, + 1562043, 5162043, 1652043, 6152043, 5612043, 6512043, 2561043, 5261043, 2651043, 6251043, 5621043, 6521043, + 145623, 1045623, 415623, 4015623, 1405623, 4105623, 154623, 1054623, 514623, 5014623, 1504623, 5104623, + 451623, 4051623, 541623, 5041623, 4501623, 5401623, 1450623, 4150623, 1540623, 5140623, 4510623, 5410623, + 146523, 1046523, 416523, 4016523, 1406523, 4106523, 164523, 1064523, 614523, 6014523, 1604523, 6104523, + 461523, 4061523, 641523, 6041523, 4601523, 6401523, 1460523, 4160523, 1640523, 6140523, 4610523, 6410523, + 156423, 1056423, 516423, 5016423, 1506423, 5106423, 165423, 1065423, 615423, 6015423, 1605423, 6105423, + 561423, 5061423, 651423, 6051423, 5601423, 6501423, 1560423, 5160423, 1650423, 6150423, 5610423, 6510423, + 456123, 4056123, 546123, 5046123, 4506123, 5406123, 465123, 4065123, 645123, 6045123, 4605123, 6405123, + 564123, 5064123, 654123, 6054123, 5604123, 6504123, 4560123, 5460123, 4650123, 6450123, 5640123, 6540123, + 1456023, 4156023, 1546023, 5146023, 4516023, 5416023, 1465023, 4165023, 1645023, 6145023, 4615023, 6415023, + 1564023, 5164023, 1654023, 6154023, 5614023, 6514023, 4561023, 5461023, 4651023, 6451023, 5641023, 6541023, + 245613, 2045613, 425613, 4025613, 2405613, 4205613, 254613, 2054613, 524613, 5024613, 2504613, 5204613, + 452613, 4052613, 542613, 5042613, 4502613, 5402613, 2450613, 4250613, 2540613, 5240613, 4520613, 5420613, + 246513, 2046513, 426513, 4026513, 2406513, 4206513, 264513, 2064513, 624513, 6024513, 2604513, 6204513, + 462513, 4062513, 642513, 6042513, 4602513, 6402513, 2460513, 4260513, 2640513, 6240513, 4620513, 6420513, + 256413, 2056413, 526413, 5026413, 2506413, 5206413, 265413, 2065413, 625413, 6025413, 2605413, 6205413, + 562413, 5062413, 652413, 6052413, 5602413, 6502413, 2560413, 5260413, 2650413, 6250413, 5620413, 6520413, + 456213, 4056213, 546213, 5046213, 4506213, 5406213, 465213, 4065213, 645213, 6045213, 4605213, 6405213, + 564213, 5064213, 654213, 6054213, 5604213, 6504213, 4560213, 5460213, 4650213, 6450213, 5640213, 6540213, + 2456013, 4256013, 2546013, 5246013, 4526013, 5426013, 2465013, 4265013, 2645013, 6245013, 4625013, 6425013, + 2564013, 5264013, 2654013, 6254013, 5624013, 6524013, 4562013, 5462013, 4652013, 6452013, 5642013, 6542013, + 1245603, 2145603, 1425603, 4125603, 2415603, 4215603, 1254603, 2154603, 1524603, 5124603, 2514603, 5214603, + 1452603, 4152603, 1542603, 5142603, 4512603, 5412603, 2451603, 4251603, 2541603, 5241603, 4521603, 5421603, + 1246503, 2146503, 1426503, 4126503, 2416503, 4216503, 1264503, 2164503, 1624503, 6124503, 2614503, 6214503, + 1462503, 4162503, 1642503, 6142503, 4612503, 6412503, 2461503, 4261503, 2641503, 6241503, 4621503, 6421503, + 1256403, 2156403, 1526403, 5126403, 2516403, 5216403, 1265403, 2165403, 1625403, 6125403, 2615403, 6215403, + 1562403, 5162403, 1652403, 6152403, 5612403, 6512403, 2561403, 5261403, 2651403, 6251403, 5621403, 6521403, + 1456203, 4156203, 1546203, 5146203, 4516203, 5416203, 1465203, 4165203, 1645203, 6145203, 4615203, 6415203, + 1564203, 5164203, 1654203, 6154203, 5614203, 6514203, 4561203, 5461203, 4651203, 6451203, 5641203, 6541203, + 2456103, 4256103, 2546103, 5246103, 4526103, 5426103, 2465103, 4265103, 2645103, 6245103, 4625103, 6425103, + 2564103, 5264103, 2654103, 6254103, 5624103, 6524103, 4562103, 5462103, 4652103, 6452103, 5642103, 6542103, + 134562, 1034562, 314562, 3014562, 1304562, 3104562, 143562, 1043562, 413562, 4013562, 1403562, 4103562, + 341562, 3041562, 431562, 4031562, 3401562, 4301562, 1340562, 3140562, 1430562, 4130562, 3410562, 4310562, + 135462, 1035462, 315462, 3015462, 1305462, 3105462, 153462, 1053462, 513462, 5013462, 1503462, 5103462, + 351462, 3051462, 531462, 5031462, 3501462, 5301462, 1350462, 3150462, 1530462, 5130462, 3510462, 5310462, + 145362, 1045362, 415362, 4015362, 1405362, 4105362, 154362, 1054362, 514362, 5014362, 1504362, 5104362, + 451362, 4051362, 541362, 5041362, 4501362, 5401362, 1450362, 4150362, 1540362, 5140362, 4510362, 5410362, + 345162, 3045162, 435162, 4035162, 3405162, 4305162, 354162, 3054162, 534162, 5034162, 3504162, 5304162, + 453162, 4053162, 543162, 5043162, 4503162, 5403162, 3450162, 4350162, 3540162, 5340162, 4530162, 5430162, + 1345062, 3145062, 1435062, 4135062, 3415062, 4315062, 1354062, 3154062, 1534062, 5134062, 3514062, 5314062, + 1453062, 4153062, 1543062, 5143062, 4513062, 5413062, 3451062, 4351062, 3541062, 5341062, 4531062, 5431062, + 134652, 1034652, 314652, 3014652, 1304652, 3104652, 143652, 1043652, 413652, 4013652, 1403652, 4103652, + 341652, 3041652, 431652, 4031652, 3401652, 4301652, 1340652, 3140652, 1430652, 4130652, 3410652, 4310652, + 136452, 1036452, 316452, 3016452, 1306452, 3106452, 163452, 1063452, 613452, 6013452, 1603452, 6103452, + 361452, 3061452, 631452, 6031452, 3601452, 6301452, 1360452, 3160452, 1630452, 6130452, 3610452, 6310452, + 146352, 1046352, 416352, 4016352, 1406352, 4106352, 164352, 1064352, 614352, 6014352, 1604352, 6104352, + 461352, 4061352, 641352, 6041352, 4601352, 6401352, 1460352, 4160352, 1640352, 6140352, 4610352, 6410352, + 346152, 3046152, 436152, 4036152, 3406152, 4306152, 364152, 3064152, 634152, 6034152, 3604152, 6304152, + 463152, 4063152, 643152, 6043152, 4603152, 6403152, 3460152, 4360152, 3640152, 6340152, 4630152, 6430152, + 1346052, 3146052, 1436052, 4136052, 3416052, 4316052, 1364052, 3164052, 1634052, 6134052, 3614052, 6314052, + 1463052, 4163052, 1643052, 6143052, 4613052, 6413052, 3461052, 4361052, 3641052, 6341052, 4631052, 6431052, + 135642, 1035642, 315642, 3015642, 1305642, 3105642, 153642, 1053642, 513642, 5013642, 1503642, 5103642, + 351642, 3051642, 531642, 5031642, 3501642, 5301642, 1350642, 3150642, 1530642, 5130642, 3510642, 5310642, + 136542, 1036542, 316542, 3016542, 1306542, 3106542, 163542, 1063542, 613542, 6013542, 1603542, 6103542, + 361542, 3061542, 631542, 6031542, 3601542, 6301542, 1360542, 3160542, 1630542, 6130542, 3610542, 6310542, + 156342, 1056342, 516342, 5016342, 1506342, 5106342, 165342, 1065342, 615342, 6015342, 1605342, 6105342, + 561342, 5061342, 651342, 6051342, 5601342, 6501342, 1560342, 5160342, 1650342, 6150342, 5610342, 6510342, + 356142, 3056142, 536142, 5036142, 3506142, 5306142, 365142, 3065142, 635142, 6035142, 3605142, 6305142, + 563142, 5063142, 653142, 6053142, 5603142, 6503142, 3560142, 5360142, 3650142, 6350142, 5630142, 6530142, + 1356042, 3156042, 1536042, 5136042, 3516042, 5316042, 1365042, 3165042, 1635042, 6135042, 3615042, 6315042, + 1563042, 5163042, 1653042, 6153042, 5613042, 6513042, 3561042, 5361042, 3651042, 6351042, 5631042, 6531042, + 145632, 1045632, 415632, 4015632, 1405632, 4105632, 154632, 1054632, 514632, 5014632, 1504632, 5104632, + 451632, 4051632, 541632, 5041632, 4501632, 5401632, 1450632, 4150632, 1540632, 5140632, 4510632, 5410632, + 146532, 1046532, 416532, 4016532, 1406532, 4106532, 164532, 1064532, 614532, 6014532, 1604532, 6104532, + 461532, 4061532, 641532, 6041532, 4601532, 6401532, 1460532, 4160532, 1640532, 6140532, 4610532, 6410532, + 156432, 1056432, 516432, 5016432, 1506432, 5106432, 165432, 1065432, 615432, 6015432, 1605432, 6105432, + 561432, 5061432, 651432, 6051432, 5601432, 6501432, 1560432, 5160432, 1650432, 6150432, 5610432, 6510432, + 456132, 4056132, 546132, 5046132, 4506132, 5406132, 465132, 4065132, 645132, 6045132, 4605132, 6405132, + 564132, 5064132, 654132, 6054132, 5604132, 6504132, 4560132, 5460132, 4650132, 6450132, 5640132, 6540132, + 1456032, 4156032, 1546032, 5146032, 4516032, 5416032, 1465032, 4165032, 1645032, 6145032, 4615032, 6415032, + 1564032, 5164032, 1654032, 6154032, 5614032, 6514032, 4561032, 5461032, 4651032, 6451032, 5641032, 6541032, + 345612, 3045612, 435612, 4035612, 3405612, 4305612, 354612, 3054612, 534612, 5034612, 3504612, 5304612, + 453612, 4053612, 543612, 5043612, 4503612, 5403612, 3450612, 4350612, 3540612, 5340612, 4530612, 5430612, + 346512, 3046512, 436512, 4036512, 3406512, 4306512, 364512, 3064512, 634512, 6034512, 3604512, 6304512, + 463512, 4063512, 643512, 6043512, 4603512, 6403512, 3460512, 4360512, 3640512, 6340512, 4630512, 6430512, + 356412, 3056412, 536412, 5036412, 3506412, 5306412, 365412, 3065412, 635412, 6035412, 3605412, 6305412, + 563412, 5063412, 653412, 6053412, 5603412, 6503412, 3560412, 5360412, 3650412, 6350412, 5630412, 6530412, + 456312, 4056312, 546312, 5046312, 4506312, 5406312, 465312, 4065312, 645312, 6045312, 4605312, 6405312, + 564312, 5064312, 654312, 6054312, 5604312, 6504312, 4560312, 5460312, 4650312, 6450312, 5640312, 6540312, + 3456012, 4356012, 3546012, 5346012, 4536012, 5436012, 3465012, 4365012, 3645012, 6345012, 4635012, 6435012, + 3564012, 5364012, 3654012, 6354012, 5634012, 6534012, 4563012, 5463012, 4653012, 6453012, 5643012, 6543012, + 1345602, 3145602, 1435602, 4135602, 3415602, 4315602, 1354602, 3154602, 1534602, 5134602, 3514602, 5314602, + 1453602, 4153602, 1543602, 5143602, 4513602, 5413602, 3451602, 4351602, 3541602, 5341602, 4531602, 5431602, + 1346502, 3146502, 1436502, 4136502, 3416502, 4316502, 1364502, 3164502, 1634502, 6134502, 3614502, 6314502, + 1463502, 4163502, 1643502, 6143502, 4613502, 6413502, 3461502, 4361502, 3641502, 6341502, 4631502, 6431502, + 1356402, 3156402, 1536402, 5136402, 3516402, 5316402, 1365402, 3165402, 1635402, 6135402, 3615402, 6315402, + 1563402, 5163402, 1653402, 6153402, 5613402, 6513402, 3561402, 5361402, 3651402, 6351402, 5631402, 6531402, + 1456302, 4156302, 1546302, 5146302, 4516302, 5416302, 1465302, 4165302, 1645302, 6145302, 4615302, 6415302, + 1564302, 5164302, 1654302, 6154302, 5614302, 6514302, 4561302, 5461302, 4651302, 6451302, 5641302, 6541302, + 3456102, 4356102, 3546102, 5346102, 4536102, 5436102, 3465102, 4365102, 3645102, 6345102, 4635102, 6435102, + 3564102, 5364102, 3654102, 6354102, 5634102, 6534102, 4563102, 5463102, 4653102, 6453102, 5643102, 6543102, + 234561, 2034561, 324561, 3024561, 2304561, 3204561, 243561, 2043561, 423561, 4023561, 2403561, 4203561, + 342561, 3042561, 432561, 4032561, 3402561, 4302561, 2340561, 3240561, 2430561, 4230561, 3420561, 4320561, + 235461, 2035461, 325461, 3025461, 2305461, 3205461, 253461, 2053461, 523461, 5023461, 2503461, 5203461, + 352461, 3052461, 532461, 5032461, 3502461, 5302461, 2350461, 3250461, 2530461, 5230461, 3520461, 5320461, + 245361, 2045361, 425361, 4025361, 2405361, 4205361, 254361, 2054361, 524361, 5024361, 2504361, 5204361, + 452361, 4052361, 542361, 5042361, 4502361, 5402361, 2450361, 4250361, 2540361, 5240361, 4520361, 5420361, + 345261, 3045261, 435261, 4035261, 3405261, 4305261, 354261, 3054261, 534261, 5034261, 3504261, 5304261, + 453261, 4053261, 543261, 5043261, 4503261, 5403261, 3450261, 4350261, 3540261, 5340261, 4530261, 5430261, + 2345061, 3245061, 2435061, 4235061, 3425061, 4325061, 2354061, 3254061, 2534061, 5234061, 3524061, 5324061, + 2453061, 4253061, 2543061, 5243061, 4523061, 5423061, 3452061, 4352061, 3542061, 5342061, 4532061, 5432061, + 234651, 2034651, 324651, 3024651, 2304651, 3204651, 243651, 2043651, 423651, 4023651, 2403651, 4203651, + 342651, 3042651, 432651, 4032651, 3402651, 4302651, 2340651, 3240651, 2430651, 4230651, 3420651, 4320651, + 236451, 2036451, 326451, 3026451, 2306451, 3206451, 263451, 2063451, 623451, 6023451, 2603451, 6203451, + 362451, 3062451, 632451, 6032451, 3602451, 6302451, 2360451, 3260451, 2630451, 6230451, 3620451, 6320451, + 246351, 2046351, 426351, 4026351, 2406351, 4206351, 264351, 2064351, 624351, 6024351, 2604351, 6204351, + 462351, 4062351, 642351, 6042351, 4602351, 6402351, 2460351, 4260351, 2640351, 6240351, 4620351, 6420351, + 346251, 3046251, 436251, 4036251, 3406251, 4306251, 364251, 3064251, 634251, 6034251, 3604251, 6304251, + 463251, 4063251, 643251, 6043251, 4603251, 6403251, 3460251, 4360251, 3640251, 6340251, 4630251, 6430251, + 2346051, 3246051, 2436051, 4236051, 3426051, 4326051, 2364051, 3264051, 2634051, 6234051, 3624051, 6324051, + 2463051, 4263051, 2643051, 6243051, 4623051, 6423051, 3462051, 4362051, 3642051, 6342051, 4632051, 6432051, + 235641, 2035641, 325641, 3025641, 2305641, 3205641, 253641, 2053641, 523641, 5023641, 2503641, 5203641, + 352641, 3052641, 532641, 5032641, 3502641, 5302641, 2350641, 3250641, 2530641, 5230641, 3520641, 5320641, + 236541, 2036541, 326541, 3026541, 2306541, 3206541, 263541, 2063541, 623541, 6023541, 2603541, 6203541, + 362541, 3062541, 632541, 6032541, 3602541, 6302541, 2360541, 3260541, 2630541, 6230541, 3620541, 6320541, + 256341, 2056341, 526341, 5026341, 2506341, 5206341, 265341, 2065341, 625341, 6025341, 2605341, 6205341, + 562341, 5062341, 652341, 6052341, 5602341, 6502341, 2560341, 5260341, 2650341, 6250341, 5620341, 6520341, + 356241, 3056241, 536241, 5036241, 3506241, 5306241, 365241, 3065241, 635241, 6035241, 3605241, 6305241, + 563241, 5063241, 653241, 6053241, 5603241, 6503241, 3560241, 5360241, 3650241, 6350241, 5630241, 6530241, + 2356041, 3256041, 2536041, 5236041, 3526041, 5326041, 2365041, 3265041, 2635041, 6235041, 3625041, 6325041, + 2563041, 5263041, 2653041, 6253041, 5623041, 6523041, 3562041, 5362041, 3652041, 6352041, 5632041, 6532041, + 245631, 2045631, 425631, 4025631, 2405631, 4205631, 254631, 2054631, 524631, 5024631, 2504631, 5204631, + 452631, 4052631, 542631, 5042631, 4502631, 5402631, 2450631, 4250631, 2540631, 5240631, 4520631, 5420631, + 246531, 2046531, 426531, 4026531, 2406531, 4206531, 264531, 2064531, 624531, 6024531, 2604531, 6204531, + 462531, 4062531, 642531, 6042531, 4602531, 6402531, 2460531, 4260531, 2640531, 6240531, 4620531, 6420531, + 256431, 2056431, 526431, 5026431, 2506431, 5206431, 265431, 2065431, 625431, 6025431, 2605431, 6205431, + 562431, 5062431, 652431, 6052431, 5602431, 6502431, 2560431, 5260431, 2650431, 6250431, 5620431, 6520431, + 456231, 4056231, 546231, 5046231, 4506231, 5406231, 465231, 4065231, 645231, 6045231, 4605231, 6405231, + 564231, 5064231, 654231, 6054231, 5604231, 6504231, 4560231, 5460231, 4650231, 6450231, 5640231, 6540231, + 2456031, 4256031, 2546031, 5246031, 4526031, 5426031, 2465031, 4265031, 2645031, 6245031, 4625031, 6425031, + 2564031, 5264031, 2654031, 6254031, 5624031, 6524031, 4562031, 5462031, 4652031, 6452031, 5642031, 6542031, + 345621, 3045621, 435621, 4035621, 3405621, 4305621, 354621, 3054621, 534621, 5034621, 3504621, 5304621, + 453621, 4053621, 543621, 5043621, 4503621, 5403621, 3450621, 4350621, 3540621, 5340621, 4530621, 5430621, + 346521, 3046521, 436521, 4036521, 3406521, 4306521, 364521, 3064521, 634521, 6034521, 3604521, 6304521, + 463521, 4063521, 643521, 6043521, 4603521, 6403521, 3460521, 4360521, 3640521, 6340521, 4630521, 6430521, + 356421, 3056421, 536421, 5036421, 3506421, 5306421, 365421, 3065421, 635421, 6035421, 3605421, 6305421, + 563421, 5063421, 653421, 6053421, 5603421, 6503421, 3560421, 5360421, 3650421, 6350421, 5630421, 6530421, + 456321, 4056321, 546321, 5046321, 4506321, 5406321, 465321, 4065321, 645321, 6045321, 4605321, 6405321, + 564321, 5064321, 654321, 6054321, 5604321, 6504321, 4560321, 5460321, 4650321, 6450321, 5640321, 6540321, + 3456021, 4356021, 3546021, 5346021, 4536021, 5436021, 3465021, 4365021, 3645021, 6345021, 4635021, 6435021, + 3564021, 5364021, 3654021, 6354021, 5634021, 6534021, 4563021, 5463021, 4653021, 6453021, 5643021, 6543021, + 2345601, 3245601, 2435601, 4235601, 3425601, 4325601, 2354601, 3254601, 2534601, 5234601, 3524601, 5324601, + 2453601, 4253601, 2543601, 5243601, 4523601, 5423601, 3452601, 4352601, 3542601, 5342601, 4532601, 5432601, + 2346501, 3246501, 2436501, 4236501, 3426501, 4326501, 2364501, 3264501, 2634501, 6234501, 3624501, 6324501, + 2463501, 4263501, 2643501, 6243501, 4623501, 6423501, 3462501, 4362501, 3642501, 6342501, 4632501, 6432501, + 2356401, 3256401, 2536401, 5236401, 3526401, 5326401, 2365401, 3265401, 2635401, 6235401, 3625401, 6325401, + 2563401, 5263401, 2653401, 6253401, 5623401, 6523401, 3562401, 5362401, 3652401, 6352401, 5632401, 6532401, + 2456301, 4256301, 2546301, 5246301, 4526301, 5426301, 2465301, 4265301, 2645301, 6245301, 4625301, 6425301, + 2564301, 5264301, 2654301, 6254301, 5624301, 6524301, 4562301, 5462301, 4652301, 6452301, 5642301, 6542301, + 3456201, 4356201, 3546201, 5346201, 4536201, 5436201, 3465201, 4365201, 3645201, 6345201, 4635201, 6435201, + 3564201, 5364201, 3654201, 6354201, 5634201, 6534201, 4563201, 5463201, 4653201, 6453201, 5643201, 6543201, + 1234560, 2134560, 1324560, 3124560, 2314560, 3214560, 1243560, 2143560, 1423560, 4123560, 2413560, 4213560, + 1342560, 3142560, 1432560, 4132560, 3412560, 4312560, 2341560, 3241560, 2431560, 4231560, 3421560, 4321560, + 1235460, 2135460, 1325460, 3125460, 2315460, 3215460, 1253460, 2153460, 1523460, 5123460, 2513460, 5213460, + 1352460, 3152460, 1532460, 5132460, 3512460, 5312460, 2351460, 3251460, 2531460, 5231460, 3521460, 5321460, + 1245360, 2145360, 1425360, 4125360, 2415360, 4215360, 1254360, 2154360, 1524360, 5124360, 2514360, 5214360, + 1452360, 4152360, 1542360, 5142360, 4512360, 5412360, 2451360, 4251360, 2541360, 5241360, 4521360, 5421360, + 1345260, 3145260, 1435260, 4135260, 3415260, 4315260, 1354260, 3154260, 1534260, 5134260, 3514260, 5314260, + 1453260, 4153260, 1543260, 5143260, 4513260, 5413260, 3451260, 4351260, 3541260, 5341260, 4531260, 5431260, + 2345160, 3245160, 2435160, 4235160, 3425160, 4325160, 2354160, 3254160, 2534160, 5234160, 3524160, 5324160, + 2453160, 4253160, 2543160, 5243160, 4523160, 5423160, 3452160, 4352160, 3542160, 5342160, 4532160, 5432160, + 1234650, 2134650, 1324650, 3124650, 2314650, 3214650, 1243650, 2143650, 1423650, 4123650, 2413650, 4213650, + 1342650, 3142650, 1432650, 4132650, 3412650, 4312650, 2341650, 3241650, 2431650, 4231650, 3421650, 4321650, + 1236450, 2136450, 1326450, 3126450, 2316450, 3216450, 1263450, 2163450, 1623450, 6123450, 2613450, 6213450, + 1362450, 3162450, 1632450, 6132450, 3612450, 6312450, 2361450, 3261450, 2631450, 6231450, 3621450, 6321450, + 1246350, 2146350, 1426350, 4126350, 2416350, 4216350, 1264350, 2164350, 1624350, 6124350, 2614350, 6214350, + 1462350, 4162350, 1642350, 6142350, 4612350, 6412350, 2461350, 4261350, 2641350, 6241350, 4621350, 6421350, + 1346250, 3146250, 1436250, 4136250, 3416250, 4316250, 1364250, 3164250, 1634250, 6134250, 3614250, 6314250, + 1463250, 4163250, 1643250, 6143250, 4613250, 6413250, 3461250, 4361250, 3641250, 6341250, 4631250, 6431250, + 2346150, 3246150, 2436150, 4236150, 3426150, 4326150, 2364150, 3264150, 2634150, 6234150, 3624150, 6324150, + 2463150, 4263150, 2643150, 6243150, 4623150, 6423150, 3462150, 4362150, 3642150, 6342150, 4632150, 6432150, + 1235640, 2135640, 1325640, 3125640, 2315640, 3215640, 1253640, 2153640, 1523640, 5123640, 2513640, 5213640, + 1352640, 3152640, 1532640, 5132640, 3512640, 5312640, 2351640, 3251640, 2531640, 5231640, 3521640, 5321640, + 1236540, 2136540, 1326540, 3126540, 2316540, 3216540, 1263540, 2163540, 1623540, 6123540, 2613540, 6213540, + 1362540, 3162540, 1632540, 6132540, 3612540, 6312540, 2361540, 3261540, 2631540, 6231540, 3621540, 6321540, + 1256340, 2156340, 1526340, 5126340, 2516340, 5216340, 1265340, 2165340, 1625340, 6125340, 2615340, 6215340, + 1562340, 5162340, 1652340, 6152340, 5612340, 6512340, 2561340, 5261340, 2651340, 6251340, 5621340, 6521340, + 1356240, 3156240, 1536240, 5136240, 3516240, 5316240, 1365240, 3165240, 1635240, 6135240, 3615240, 6315240, + 1563240, 5163240, 1653240, 6153240, 5613240, 6513240, 3561240, 5361240, 3651240, 6351240, 5631240, 6531240, + 2356140, 3256140, 2536140, 5236140, 3526140, 5326140, 2365140, 3265140, 2635140, 6235140, 3625140, 6325140, + 2563140, 5263140, 2653140, 6253140, 5623140, 6523140, 3562140, 5362140, 3652140, 6352140, 5632140, 6532140, + 1245630, 2145630, 1425630, 4125630, 2415630, 4215630, 1254630, 2154630, 1524630, 5124630, 2514630, 5214630, + 1452630, 4152630, 1542630, 5142630, 4512630, 5412630, 2451630, 4251630, 2541630, 5241630, 4521630, 5421630, + 1246530, 2146530, 1426530, 4126530, 2416530, 4216530, 1264530, 2164530, 1624530, 6124530, 2614530, 6214530, + 1462530, 4162530, 1642530, 6142530, 4612530, 6412530, 2461530, 4261530, 2641530, 6241530, 4621530, 6421530, + 1256430, 2156430, 1526430, 5126430, 2516430, 5216430, 1265430, 2165430, 1625430, 6125430, 2615430, 6215430, + 1562430, 5162430, 1652430, 6152430, 5612430, 6512430, 2561430, 5261430, 2651430, 6251430, 5621430, 6521430, + 1456230, 4156230, 1546230, 5146230, 4516230, 5416230, 1465230, 4165230, 1645230, 6145230, 4615230, 6415230, + 1564230, 5164230, 1654230, 6154230, 5614230, 6514230, 4561230, 5461230, 4651230, 6451230, 5641230, 6541230, + 2456130, 4256130, 2546130, 5246130, 4526130, 5426130, 2465130, 4265130, 2645130, 6245130, 4625130, 6425130, + 2564130, 5264130, 2654130, 6254130, 5624130, 6524130, 4562130, 5462130, 4652130, 6452130, 5642130, 6542130, + 1345620, 3145620, 1435620, 4135620, 3415620, 4315620, 1354620, 3154620, 1534620, 5134620, 3514620, 5314620, + 1453620, 4153620, 1543620, 5143620, 4513620, 5413620, 3451620, 4351620, 3541620, 5341620, 4531620, 5431620, + 1346520, 3146520, 1436520, 4136520, 3416520, 4316520, 1364520, 3164520, 1634520, 6134520, 3614520, 6314520, + 1463520, 4163520, 1643520, 6143520, 4613520, 6413520, 3461520, 4361520, 3641520, 6341520, 4631520, 6431520, + 1356420, 3156420, 1536420, 5136420, 3516420, 5316420, 1365420, 3165420, 1635420, 6135420, 3615420, 6315420, + 1563420, 5163420, 1653420, 6153420, 5613420, 6513420, 3561420, 5361420, 3651420, 6351420, 5631420, 6531420, + 1456320, 4156320, 1546320, 5146320, 4516320, 5416320, 1465320, 4165320, 1645320, 6145320, 4615320, 6415320, + 1564320, 5164320, 1654320, 6154320, 5614320, 6514320, 4561320, 5461320, 4651320, 6451320, 5641320, 6541320, + 3456120, 4356120, 3546120, 5346120, 4536120, 5436120, 3465120, 4365120, 3645120, 6345120, 4635120, 6435120, + 3564120, 5364120, 3654120, 6354120, 5634120, 6534120, 4563120, 5463120, 4653120, 6453120, 5643120, 6543120, + 2345610, 3245610, 2435610, 4235610, 3425610, 4325610, 2354610, 3254610, 2534610, 5234610, 3524610, 5324610, + 2453610, 4253610, 2543610, 5243610, 4523610, 5423610, 3452610, 4352610, 3542610, 5342610, 4532610, 5432610, + 2346510, 3246510, 2436510, 4236510, 3426510, 4326510, 2364510, 3264510, 2634510, 6234510, 3624510, 6324510, + 2463510, 4263510, 2643510, 6243510, 4623510, 6423510, 3462510, 4362510, 3642510, 6342510, 4632510, 6432510, + 2356410, 3256410, 2536410, 5236410, 3526410, 5326410, 2365410, 3265410, 2635410, 6235410, 3625410, 6325410, + 2563410, 5263410, 2653410, 6253410, 5623410, 6523410, 3562410, 5362410, 3652410, 6352410, 5632410, 6532410, + 2456310, 4256310, 2546310, 5246310, 4526310, 5426310, 2465310, 4265310, 2645310, 6245310, 4625310, 6425310, + 2564310, 5264310, 2654310, 6254310, 5624310, 6524310, 4562310, 5462310, 4652310, 6452310, 5642310, 6542310, + 3456210, 4356210, 3546210, 5346210, 4536210, 5436210, 3465210, 4365210, 3645210, 6345210, 4635210, 6435210, + 3564210, 5364210, 3654210, 6354210, 5634210, 6534210, 4563210, 5463210, 4653210, 6453210, 5643210, 6543210 + }; + std::map expected; + for (std::size_t i = 0; i < 5040; i++) + expected[pre_expected[i]] = 0; // flags are 0, everything is symmetric here + + VERIFY(isDynGroup(group)); + VERIFY_IS_EQUAL(group.size(), 5040u); + VERIFY_IS_EQUAL(group.globalFlags(), 0); + group.apply(identity7, 0, found, expected); + VERIFY_IS_EQUAL(found.size(), 5040u); + } +} + +static void test_tensor_epsilon() +{ + SGroup<3, AntiSymmetry<0,1>, AntiSymmetry<1,2>> sym; + Tensor epsilon(3,3,3); + + epsilon.setZero(); + epsilon.symCoeff(sym, 0, 1, 2) = 1; + + for (int i = 0; i < 3; i++) { + for (int j = 0; j < 3; j++) { + for (int k = 0; k < 3; k++) { + VERIFY_IS_EQUAL((epsilon(i,j,k)), (- (j - i) * (k - j) * (i - k) / 2) ); + } + } + } +} + +static void test_tensor_sym() +{ + SGroup<4, Symmetry<0,1>, Symmetry<2,3>> sym; + Tensor t(10,10,10,10); + + t.setZero(); + + for (int l = 0; l < 10; l++) { + for (int k = l; k < 10; k++) { + for (int j = 0; j < 10; j++) { + for (int i = j; i < 10; i++) { + t.symCoeff(sym, i, j, k, l) = (i + j) * (k + l); + } + } + } + } + + for (int l = 0; l < 10; l++) { + for (int k = 0; k < 10; k++) { + for (int j = 0; j < 10; j++) { + for (int i = 0; i < 10; i++) { + VERIFY_IS_EQUAL((t(i, j, k, l)), ((i + j) * (k + l))); + } + } + } + } + +} + +static void test_tensor_asym() +{ + SGroup<4, AntiSymmetry<0,1>, AntiSymmetry<2,3>> sym; + Tensor t(10,10,10,10); + + t.setZero(); + + for (int l = 0; l < 10; l++) { + for (int k = l + 1; k < 10; k++) { + for (int j = 0; j < 10; j++) { + for (int i = j + 1; i < 10; i++) { + t.symCoeff(sym, i, j, k, l) = ((i * j) + (k * l)); + } + } + } + } + + for (int l = 0; l < 10; l++) { + for (int k = 0; k < 10; k++) { + for (int j = 0; j < 10; j++) { + for (int i = 0; i < 10; i++) { + if (i < j && k < l) + VERIFY_IS_EQUAL((t(i, j, k, l)), (((i * j) + (k * l)))); + else if (i > j && k > l) + VERIFY_IS_EQUAL((t(i, j, k, l)), (((i * j) + (k * l)))); + else if (i < j && k > l) + VERIFY_IS_EQUAL((t(i, j, k, l)), (- ((i * j) + (k * l)))); + else if (i > j && k < l) + VERIFY_IS_EQUAL((t(i, j, k, l)), (- ((i * j) + (k * l)))); + else + VERIFY_IS_EQUAL((t(i, j, k, l)), 0); + } + } + } + } +} + +static void test_tensor_dynsym() +{ + DynamicSGroup sym(4); + sym.addSymmetry(0,1); + sym.addSymmetry(2,3); + Tensor t(10,10,10,10); + + t.setZero(); + + for (int l = 0; l < 10; l++) { + for (int k = l; k < 10; k++) { + for (int j = 0; j < 10; j++) { + for (int i = j; i < 10; i++) { + t.symCoeff(sym, i, j, k, l) = (i + j) * (k + l); + } + } + } + } + + for (int l = 0; l < 10; l++) { + for (int k = 0; k < 10; k++) { + for (int j = 0; j < 10; j++) { + for (int i = 0; i < 10; i++) { + VERIFY_IS_EQUAL((t(i, j, k, l)), ((i + j) * (k + l))); + } + } + } + } +} + +static void test_tensor_randacc() +{ + SGroup<4, Symmetry<0,1>, Symmetry<2,3>> sym; + Tensor t(10,10,10,10); + + t.setZero(); + + // set elements 1 million times, that way we access the + // entire matrix + for (int n = 0; n < 1000000; n++) { + int i = rand() % 10; + int j = rand() % 10; + int k = rand() % 10; + int l = rand() % 10; + // only access those indices in a given order + if (i < j) + std::swap(i, j); + if (k < l) + std::swap(k, l); + t.symCoeff(sym, i, j, k, l) = (i + j) * (k + l); + } + + for (int l = 0; l < 10; l++) { + for (int k = 0; k < 10; k++) { + for (int j = 0; j < 10; j++) { + for (int i = 0; i < 10; i++) { + VERIFY_IS_EQUAL((t(i, j, k, l)), ((i + j) * (k + l))); + } + } + } + } +} + +void test_cxx11_tensor_symmetry() +{ + CALL_SUBTEST(test_symgroups_static()); + CALL_SUBTEST(test_symgroups_dynamic()); + CALL_SUBTEST(test_symgroups_selection()); + CALL_SUBTEST(test_tensor_epsilon()); + CALL_SUBTEST(test_tensor_sym()); + CALL_SUBTEST(test_tensor_asym()); + CALL_SUBTEST(test_tensor_dynsym()); + CALL_SUBTEST(test_tensor_randacc()); +} + +/* + * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; + */ From 6b471f205ef00572abebaf403d55433f2ce40eee Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 15 Nov 2013 10:59:19 +0100 Subject: [PATCH 0192/1103] fix overflow and ambiguity in SparseLU memory allocation --- Eigen/src/SparseLU/SparseLU_Memory.h | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU_Memory.h b/Eigen/src/SparseLU/SparseLU_Memory.h index e1e7de043..d068d62df 100644 --- a/Eigen/src/SparseLU/SparseLU_Memory.h +++ b/Eigen/src/SparseLU/SparseLU_Memory.h @@ -93,7 +93,7 @@ Index SparseLUImpl::expand(VectorType& vec, Index& length, Index { // First time to allocate from LUMemInit() // Let LUMemInit() deals with it. - return 0; + return -1; } if (keep_prev) { @@ -152,10 +152,9 @@ template Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lwork, Index fillratio, Index panel_size, GlobalLU_t& glu) { Index& num_expansions = glu.num_expansions; //No memory expansions so far - num_expansions = 0; - glu.nzumax = glu.nzlumax = (std::min)(fillratio * annz, m*n); // estimated number of nonzeros in U + num_expansions = 0; + glu.nzumax = glu.nzlumax = (std::min)(fillratio * annz / n, m) * n; // estimated number of nonzeros in U glu.nzlmax = (std::max)(Index(4), fillratio) * annz / 4; // estimated nnz in L factor - // Return the estimated size to the user if necessary Index tempSpace; tempSpace = (2*panel_size + 4 + LUNoMarker) * m * sizeof(Index) + (panel_size + 1) * m * sizeof(Scalar); @@ -179,10 +178,10 @@ Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lw // Reserve memory for L/U factors do { - if( (!expand(glu.lusup, glu.nzlumax, 0, 0, num_expansions)) - || (!expand(glu.ucol, glu.nzumax, 0, 0, num_expansions)) - || (!expand (glu.lsub, glu.nzlmax, 0, 0, num_expansions)) - || (!expand (glu.usub, glu.nzumax, 0, 1, num_expansions)) ) + if( (expand(glu.lusup, glu.nzlumax, 0, 0, num_expansions)<0) + || (expand(glu.ucol, glu.nzumax, 0, 0, num_expansions)<0) + || (expand (glu.lsub, glu.nzlmax, 0, 0, num_expansions)<0) + || (expand (glu.usub, glu.nzumax, 0, 1, num_expansions)<0) ) { //Reduce the estimated size and retry glu.nzlumax /= 2; From 46dd1bb1be49cad4e7eecc12813bd0ee5772cefe Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 15 Nov 2013 11:19:19 +0100 Subject: [PATCH 0193/1103] Workaround fixing aliasing issue in x = SparseLU::solve(x) --- Eigen/src/Core/PermutationMatrix.h | 3 ++- Eigen/src/SparseLU/SparseLU.h | 4 +++- 2 files changed, 5 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 4fc5dd318..1297b8413 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -553,7 +553,8 @@ struct permut_matrix_product_retval template inline void evalTo(Dest& dst) const { const Index n = Side==OnTheLeft ? rows() : cols(); - + // FIXME we need an is_same for expression that is not sensitive to constness. For instance + // is_same_xpr, Block >::value should be true. if(is_same::value && extract_data(dst) == extract_data(m_matrix)) { // apply the permutation inplace diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 71fe844ac..7a9aeec2d 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -229,8 +229,10 @@ class SparseLU : public internal::SparseLUImplmatrixL().solveInPlace(X); From f6bac196d55c1658902ec09058023209021cdd1b Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Sat, 16 Nov 2013 00:03:23 +0100 Subject: [PATCH 0194/1103] C++11/Tensor: Fix copyright headers --- unsupported/Eigen/CXX11/Core | 3 +-- unsupported/Eigen/CXX11/Tensor | 3 +-- unsupported/Eigen/CXX11/TensorSymmetry | 3 +-- unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h | 3 +-- unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h | 3 +-- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 3 +-- unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h | 3 +-- unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h | 3 +-- unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h | 3 +-- unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h | 3 +-- .../Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h | 3 +-- unsupported/test/cxx11_meta.cpp | 3 +-- unsupported/test/cxx11_tensor_simple.cpp | 3 +-- unsupported/test/cxx11_tensor_symmetry.cpp | 3 +-- 14 files changed, 14 insertions(+), 28 deletions(-) diff --git a/unsupported/Eigen/CXX11/Core b/unsupported/Eigen/CXX11/Core index 1b1f1dcc1..4dc4ab224 100644 --- a/unsupported/Eigen/CXX11/Core +++ b/unsupported/Eigen/CXX11/Core @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 083d8c0a7..f2c5129b3 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/TensorSymmetry b/unsupported/Eigen/CXX11/TensorSymmetry index 5bbab0a80..027c6087f 100644 --- a/unsupported/Eigen/CXX11/TensorSymmetry +++ b/unsupported/Eigen/CXX11/TensorSymmetry @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h index 55d3cc37b..d6b5d75d9 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index eec38b343..356ae10cf 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index ff3d6513e..c6216e14c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index 50040147d..a34600ee6 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h index 87332610d..b5738b778 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h index 048e753fc..c5a630105 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h index 40ad8d7ff..f0813086a 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h index 716d7d2e5..0fe0b7c46 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/util/TemplateGroupTheory.h @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/test/cxx11_meta.cpp b/unsupported/test/cxx11_meta.cpp index 03a9efe54..a9843e9a9 100644 --- a/unsupported/test/cxx11_meta.cpp +++ b/unsupported/test/cxx11_meta.cpp @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/test/cxx11_tensor_simple.cpp b/unsupported/test/cxx11_tensor_simple.cpp index 6875a4e58..ea512c9cc 100644 --- a/unsupported/test/cxx11_tensor_simple.cpp +++ b/unsupported/test/cxx11_tensor_simple.cpp @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/test/cxx11_tensor_symmetry.cpp b/unsupported/test/cxx11_tensor_symmetry.cpp index ebbc8a68d..e8dfffd92 100644 --- a/unsupported/test/cxx11_tensor_symmetry.cpp +++ b/unsupported/test/cxx11_tensor_symmetry.cpp @@ -1,8 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed From 5d1291a4de78af16c31cf837c0d407007bd0a5a6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 19 Nov 2013 11:51:16 +0100 Subject: [PATCH 0195/1103] Document how to reproduce matlab's rot90 --- doc/AsciiQuickReference.txt | 1 + 1 file changed, 1 insertion(+) diff --git a/doc/AsciiQuickReference.txt b/doc/AsciiQuickReference.txt index 6b0a7cd6a..fc613864f 100644 --- a/doc/AsciiQuickReference.txt +++ b/doc/AsciiQuickReference.txt @@ -77,6 +77,7 @@ R.adjoint() // R' R.transpose() // R.' or conj(R') R.diagonal() // diag(R) x.asDiagonal() // diag(x) +R.transpose().colwise().reverse(); // rot90(R) // All the same as Matlab, but matlab doesn't have *= style operators. // Matrix-vector. Matrix-matrix. Matrix-scalar. From 654eab3bd6dc842c668d8979eb36602cef929165 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 19 Nov 2013 11:53:48 +0100 Subject: [PATCH 0196/1103] Add scaling in JacobiSVD to avoid overflows --- Eigen/src/SVD/JacobiSVD.h | 11 +++++++++-- test/jacobisvd.cpp | 13 +++++++++++-- 2 files changed, 20 insertions(+), 4 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index bf1213656..eee31ca97 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -411,8 +411,8 @@ struct svd_precondition_2x2_block_to_be_real template void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, - JacobiRotation *j_left, - JacobiRotation *j_right) + JacobiRotation *j_left, + JacobiRotation *j_right) { using std::sqrt; Matrix m; @@ -831,6 +831,11 @@ JacobiSVD::compute(const MatrixType& matrix, unsig if(m_computeFullV) m_matrixV.setIdentity(m_cols,m_cols); if(m_computeThinV) m_matrixV.setIdentity(m_cols, m_diagSize); } + + // Scaling factor to reducover/under-flows + RealScalar scale = m_workMatrix.cwiseAbs().maxCoeff(); + if(scale==RealScalar(0)) scale = RealScalar(1); + m_workMatrix /= scale; /*** step 2. The main Jacobi SVD iteration. ***/ @@ -879,6 +884,8 @@ JacobiSVD::compute(const MatrixType& matrix, unsig m_singularValues.coeffRef(i) = a; if(computeU() && (a!=RealScalar(0))) m_matrixU.col(i) *= m_workMatrix.coeff(i,i)/a; } + + m_singularValues *= scale; /*** step 4. Sort singular values in descending order and compute the number of nonzero singular values ***/ diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index a0d8562c1..d441a6eca 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -285,7 +285,7 @@ void jacobisvd_inf_nan() // Regression test for bug 286: JacobiSVD loops indefinitely with some // matrices containing denormal numbers. -void jacobisvd_bug286() +void jacobisvd_underoverflow() { #if defined __INTEL_COMPILER // shut up warning #239: floating point underflow @@ -300,6 +300,15 @@ void jacobisvd_bug286() #endif JacobiSVD svd; svd.compute(M); // just check we don't loop indefinitely + + // Check for overflow: + Matrix3d M3; + M3 << 4.4331978442502944e+307, -5.8585363752028680e+307, 6.4527017443412964e+307, + 3.7841695601406358e+307, 2.4331702789740617e+306, -3.5235707140272905e+307, + -8.7190887618028355e+307, -7.3453213709232193e+307, -2.4367363684472105e+307; + + JacobiSVD svd3; + svd3.compute(M3); // just check we don't loop indefinitely } void jacobisvd_preallocate() @@ -398,5 +407,5 @@ void test_jacobisvd() CALL_SUBTEST_9( jacobisvd_preallocate() ); // Regression check for bug 286 - CALL_SUBTEST_2( jacobisvd_bug286() ); + CALL_SUBTEST_2( jacobisvd_underoverflow() ); } From 28b2abdbea00ab7c37e690679e6e4398ab300978 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 19 Nov 2013 12:53:46 +0100 Subject: [PATCH 0197/1103] Fix FullPivHouseholderQR ctors for non squared fixed size matrix types --- Eigen/src/QR/FullPivHouseholderQR.h | 12 ++++++------ test/qr_fullpivoting.cpp | 4 ++++ 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 13e14db10..e6ab172b3 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -94,10 +94,10 @@ template class FullPivHouseholderQR FullPivHouseholderQR(Index rows, Index cols) : m_qr(rows, cols), m_hCoeffs((std::min)(rows,cols)), - m_rows_transpositions(rows), - m_cols_transpositions(cols), + m_rows_transpositions((std::min)(rows,cols)), + m_cols_transpositions((std::min)(rows,cols)), m_cols_permutation(cols), - m_temp((std::min)(rows,cols)), + m_temp(cols), m_isInitialized(false), m_usePrescribedThreshold(false) {} @@ -116,10 +116,10 @@ template class FullPivHouseholderQR FullPivHouseholderQR(const MatrixType& matrix) : m_qr(matrix.rows(), matrix.cols()), m_hCoeffs((std::min)(matrix.rows(), matrix.cols())), - m_rows_transpositions(matrix.rows()), - m_cols_transpositions(matrix.cols()), + m_rows_transpositions((std::min)(matrix.rows(), matrix.cols())), + m_cols_transpositions((std::min)(matrix.rows(), matrix.cols())), m_cols_permutation(matrix.cols()), - m_temp((std::min)(matrix.rows(), matrix.cols())), + m_temp(matrix.cols()), m_isInitialized(false), m_usePrescribedThreshold(false) { diff --git a/test/qr_fullpivoting.cpp b/test/qr_fullpivoting.cpp index 15d7299d7..511f2473f 100644 --- a/test/qr_fullpivoting.cpp +++ b/test/qr_fullpivoting.cpp @@ -130,4 +130,8 @@ void test_qr_fullpivoting() // Test problem size constructors CALL_SUBTEST_7(FullPivHouseholderQR(10, 20)); + CALL_SUBTEST_7((FullPivHouseholderQR >(10,20))); + CALL_SUBTEST_7((FullPivHouseholderQR >(Matrix::Random()))); + CALL_SUBTEST_7((FullPivHouseholderQR >(20,10))); + CALL_SUBTEST_7((FullPivHouseholderQR >(Matrix::Random()))); } From c550a0e6340e432c87351638561f5b96f32d5d89 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 21 Nov 2013 10:39:47 +0100 Subject: [PATCH 0198/1103] extend Map unit test to check buffers allocated on the stack --- test/mapped_matrix.cpp | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/test/mapped_matrix.cpp b/test/mapped_matrix.cpp index de9dbbde3..c18e687a5 100644 --- a/test/mapped_matrix.cpp +++ b/test/mapped_matrix.cpp @@ -13,6 +13,8 @@ #include "main.h" +#define EIGEN_TESTMAP_MAX_SIZE 256 + template void map_class_vector(const VectorType& m) { typedef typename VectorType::Index Index; @@ -25,15 +27,19 @@ template void map_class_vector(const VectorType& m) Scalar* array2 = internal::aligned_new(size); Scalar* array3 = new Scalar[size+1]; Scalar* array3unaligned = size_t(array3)%16 == 0 ? array3+1 : array3; + Scalar array4[EIGEN_TESTMAP_MAX_SIZE]; Map(array1, size) = VectorType::Random(size); Map(array2, size) = Map(array1, size); Map(array3unaligned, size) = Map(array1, size); + Map(array4, size) = Map(array1, size); VectorType ma1 = Map(array1, size); VectorType ma2 = Map(array2, size); VectorType ma3 = Map(array3unaligned, size); + VectorType ma4 = Map(array4, size); VERIFY_IS_EQUAL(ma1, ma2); VERIFY_IS_EQUAL(ma1, ma3); + VERIFY_IS_EQUAL(ma1, ma4); #ifdef EIGEN_VECTORIZE if(internal::packet_traits::Vectorizable) VERIFY_RAISES_ASSERT((Map(array3unaligned, size))) @@ -120,6 +126,7 @@ void test_mapped_matrix() CALL_SUBTEST_1( map_class_vector(Matrix()) ); CALL_SUBTEST_1( check_const_correctness(Matrix()) ); CALL_SUBTEST_2( map_class_vector(Vector4d()) ); + CALL_SUBTEST_2( map_class_vector(VectorXd(13)) ); CALL_SUBTEST_2( check_const_correctness(Matrix4d()) ); CALL_SUBTEST_3( map_class_vector(RowVector4f()) ); CALL_SUBTEST_4( map_class_vector(VectorXcf(8)) ); From 230f5c3aa91b843cf1864258386e3902ffb6370a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 25 Nov 2013 15:20:31 +0100 Subject: [PATCH 0199/1103] Evaluator: introduce the main Assignment class, add call_assignment to bypass NoAlias and AssumeAliasing, and some bits of cleaning --- Eigen/src/Core/AssignEvaluator.h | 244 +++++++++---------------------- Eigen/src/Core/NoAlias.h | 3 +- Eigen/src/Core/Swap.h | 72 +++++++++ Eigen/src/Core/util/XprHelper.h | 10 ++ test/evaluators.cpp | 76 +++++++++- 5 files changed, 226 insertions(+), 179 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 5b5d29ca9..158dc7842 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -514,7 +514,7 @@ struct dense_assignment_loop }; /*************************************************************************** -* Part 4 : Generic Assignment routine +* Part 4 : Generic dense assignment kernel ***************************************************************************/ // This class generalize the assignment of a coefficient (or packet) from one dense evaluator @@ -617,6 +617,10 @@ protected: DstXprType& m_dstExpr; }; +/*************************************************************************** +* Part 5 : Entry point for dense rectangular assignment +***************************************************************************/ + template void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) { @@ -645,195 +649,81 @@ void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src) } /*************************************************************************** -* Part 5 : Entry points +* Part 6 : Generic assignment ***************************************************************************/ -// Based on DenseBase::LazyAssign() -// The following functions are just for testing and they are meant to be moved to operator= and the likes. -template class StorageBase, typename SrcXprType> -EIGEN_STRONG_INLINE -const DstXprType& copy_using_evaluator(const NoAlias& dst, - const EigenBase& src) +// An evaluator must define its shape. It can be one of the following: +struct DenseShape {}; +struct DiagonalShape {}; +struct BandShape {}; +struct TriangularShape {}; +struct SelfAdjointShape {}; +struct SparseShape {}; + + +// Based on the respective shapes of the destination and source, +// the class AssignmentKind determine the kind of assignment mechanism. +// AssignmentKind must define a Kind typedef. +template struct AssignmentKind; + +// AssignmentKind<.,.>::Kind can be one of the following: + struct Dense2Dense {}; + struct Triangular2Triangular {}; +// struct Diagonal2Diagonal {}; // <=> Dense2Dense + struct Sparse2Dense {}; + struct Sparse2Sparse {}; + +// This is the main assignment class +template< typename DstXprType, typename SrcXprType, typename Functor, + typename Kind = Dense2Dense,//AssignmentKind< evaluator::Shape , evaluator::Shape >::Kind, + typename Scalar = typename DstXprType::Scalar> +struct Assignment; + + +// The only purpose of this call_assignment() function is to deal with noalias() / AssumeAliasing. +// Indeed, I (Gael) think that this concept of AssumeAliasing was a mistake, and it makes thing quite complicated. +// So this intermediate function removes everything related to AssumeAliasing such that Assignment +// does not has to bother about these annoying details. + +template +void call_assignment(Dst& dst, const Src& src, const Func& func) { - return noalias_copy_using_evaluator(dst.expression(), src.derived(), internal::assign_op()); + typedef typename internal::conditional::AssumeAliasing==1, EvalToTemp, Src>::type ActualSrc; + Assignment::run(dst, src, func); } -template::AssumeAliasing> -struct AddEvalIfAssumingAliasing; - -template -struct AddEvalIfAssumingAliasing +// by-pass AssumeAliasing +template class StorageBase, typename Src, typename Func> +void call_assignment(const NoAlias& dst, const Src& src, const Func& func) { - static const XprType& run(const XprType& xpr) - { - return xpr; - } -}; - -template -struct AddEvalIfAssumingAliasing -{ - static const EvalToTemp run(const XprType& xpr) - { - return EvalToTemp(xpr); - } -}; - -template -EIGEN_STRONG_INLINE -const DstXprType& copy_using_evaluator(const EigenBase& dst, const EigenBase& src, const Functor &func) -{ - return noalias_copy_using_evaluator(dst.const_cast_derived(), - AddEvalIfAssumingAliasing::run(src.derived()), - func - ); + Assignment::run(dst.expression(), src, func); } -// this mimics operator= -template -EIGEN_STRONG_INLINE -const DstXprType& copy_using_evaluator(const EigenBase& dst, const EigenBase& src) +// Generic Dense to Dense assignment +template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> +struct Assignment { - return copy_using_evaluator(dst.const_cast_derived(), src.derived(), internal::assign_op()); -} - -template -EIGEN_STRONG_INLINE -const DstXprType& noalias_copy_using_evaluator(const PlainObjectBase& dst, const EigenBase& src, const Functor &func) -{ -#ifdef EIGEN_DEBUG_ASSIGN - internal::copy_using_evaluator_traits::debug(); -#endif -#ifdef EIGEN_NO_AUTOMATIC_RESIZING - eigen_assert((dst.size()==0 || (IsVectorAtCompileTime ? (dst.size() == src.size()) - : (dst.rows() == src.rows() && dst.cols() == src.cols()))) - && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined"); -#else - dst.const_cast_derived().resizeLike(src.derived()); -#endif - call_dense_assignment_loop(dst.const_cast_derived(), src.derived(), func); - return dst.derived(); -} - -template -EIGEN_STRONG_INLINE -const DstXprType& noalias_copy_using_evaluator(const EigenBase& dst, const EigenBase& src, const Functor &func) -{ - call_dense_assignment_loop(dst.const_cast_derived(), src.derived(), func); - return dst.derived(); -} - -// Based on DenseBase::swap() -// TODO: Check whether we need to do something special for swapping two -// Arrays or Matrices. (Jitse) - -// Overload default assignPacket behavior for swapping them -template -class swap_kernel : public generic_dense_assignment_kernel > -{ - typedef generic_dense_assignment_kernel > Base; - typedef typename DstEvaluatorTypeT::PacketScalar PacketScalar; - using Base::m_dst; - using Base::m_src; - using Base::m_functor; - -public: - typedef typename Base::Scalar Scalar; - typedef typename Base::Index Index; - typedef typename Base::DstXprType DstXprType; - - swap_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, DstXprType& dstExpr) - : Base(dst, src, swap_assign_op(), dstExpr) - {} - - template - void assignPacket(Index row, Index col) + static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) { - m_functor.template swapPacket(&m_dst.coeffRef(row,col), &const_cast(m_src).coeffRef(row,col)); - } - - template - void assignPacket(Index index) - { - m_functor.template swapPacket(&m_dst.coeffRef(index), &const_cast(m_src).coeffRef(index)); - } - - // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael) - template - void assignPacketByOuterInner(Index outer, Index inner) - { - Index row = Base::rowIndexByOuterInner(outer, inner); - Index col = Base::colIndexByOuterInner(outer, inner); - assignPacket(row, col); - } -}; - -template -void swap_using_evaluator(const DstXprType& dst, const SrcXprType& src) -{ - // TODO there is too much redundancy with call_dense_assignment_loop - - eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); - - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; + // TODO check whether this is the right place to perform these checks: + enum{ + SameType = internal::is_same::value + }; - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); + EIGEN_STATIC_ASSERT_LVALUE(DstXprType) + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(DstXprType,SrcXprType) + EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - typedef swap_kernel Kernel; - Kernel kernel(dstEvaluator, srcEvaluator, dst.const_cast_derived()); - - dense_assignment_loop::run(kernel); -} - -// Based on MatrixBase::operator+= (in CwiseBinaryOp.h) -template -void add_assign_using_evaluator(const MatrixBase& dst, const MatrixBase& src) -{ - typedef typename DstXprType::Scalar Scalar; - copy_using_evaluator(dst.derived(), src.derived(), add_assign_op()); -} - -// Based on ArrayBase::operator+= -template -void add_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) -{ - typedef typename DstXprType::Scalar Scalar; - copy_using_evaluator(dst.derived(), src.derived(), add_assign_op()); -} - -// TODO: Add add_assign_using_evaluator for EigenBase ? (Jitse) - -template -void subtract_assign_using_evaluator(const MatrixBase& dst, const MatrixBase& src) -{ - typedef typename DstXprType::Scalar Scalar; - copy_using_evaluator(dst.derived(), src.derived(), sub_assign_op()); -} - -template -void subtract_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) -{ - typedef typename DstXprType::Scalar Scalar; - copy_using_evaluator(dst.derived(), src.derived(), sub_assign_op()); -} - -template -void multiply_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) -{ - typedef typename DstXprType::Scalar Scalar; - copy_using_evaluator(dst.derived(), src.derived(), mul_assign_op()); -} - -template -void divide_assign_using_evaluator(const ArrayBase& dst, const ArrayBase& src) -{ - typedef typename DstXprType::Scalar Scalar; - copy_using_evaluator(dst.derived(), src.derived(), div_assign_op()); -} - + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + #ifdef EIGEN_DEBUG_ASSIGN + internal::copy_using_evaluator_traits::debug(); + #endif + + call_dense_assignment_loop(dst, src, func); + } +}; } // namespace internal diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index 0a1c32743..65117a806 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -30,8 +30,9 @@ namespace Eigen { template class StorageBase> class NoAlias { - typedef typename ExpressionType::Scalar Scalar; public: + typedef typename ExpressionType::Scalar Scalar; + NoAlias(ExpressionType& expression) : m_expression(expression) {} /** Behaves like MatrixBase::lazyAssign(other) diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index d602fba65..8fd94b3c7 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -12,6 +12,8 @@ namespace Eigen { +// #ifndef EIGEN_TEST_EVALUATORS + /** \class SwapWrapper * \ingroup Core_Module * @@ -135,6 +137,76 @@ template class SwapWrapper ExpressionType& m_expression; }; +// #endif + +#ifdef EIGEN_TEST_EVALUATORS + +namespace internal { + +// Overload default assignPacket behavior for swapping them +template +class dense_swap_kernel : public generic_dense_assignment_kernel > +{ + typedef generic_dense_assignment_kernel > Base; + typedef typename DstEvaluatorTypeT::PacketScalar PacketScalar; + using Base::m_dst; + using Base::m_src; + using Base::m_functor; + +public: + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + typedef typename Base::DstXprType DstXprType; + + dense_swap_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, DstXprType& dstExpr) + : Base(dst, src, swap_assign_op(), dstExpr) + {} + + template + void assignPacket(Index row, Index col) + { + m_functor.template swapPacket(&m_dst.coeffRef(row,col), &const_cast(m_src).coeffRef(row,col)); + } + + template + void assignPacket(Index index) + { + m_functor.template swapPacket(&m_dst.coeffRef(index), &const_cast(m_src).coeffRef(index)); + } + + // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael) + template + void assignPacketByOuterInner(Index outer, Index inner) + { + Index row = Base::rowIndexByOuterInner(outer, inner); + Index col = Base::colIndexByOuterInner(outer, inner); + assignPacket(row, col); + } +}; + +template +void call_dense_swap_loop(const DstXprType& dst, const SrcXprType& src) +{ + // TODO there is too much redundancy with call_dense_assignment_loop + + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + typedef typename evaluator::type DstEvaluatorType; + typedef typename evaluator::type SrcEvaluatorType; + + DstEvaluatorType dstEvaluator(dst); + SrcEvaluatorType srcEvaluator(src); + + typedef dense_swap_kernel Kernel; + Kernel kernel(dstEvaluator, srcEvaluator, dst.const_cast_derived()); + + dense_assignment_loop::run(kernel); +} + +} // namespace internal + +#endif + } // end namespace Eigen #endif // EIGEN_SWAP_H diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 195d9e2e1..189928c8f 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -293,6 +293,15 @@ struct transfer_constness >::type type; }; +#ifdef EIGEN_TEST_EVALUATORS + +// When using evaluators, we never evaluate when assembling the expression!! +template::type> struct nested +{ + typedef typename ref_selector::type type; +}; + +#else /** \internal Determines how a given expression should be nested into another one. * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be * nested into the bigger product expression. The choice is between nesting the expression b+c as-is, or @@ -339,6 +348,7 @@ template::type> str typename ref_selector::type >::type type; }; +#endif // EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC diff --git a/test/evaluators.cpp b/test/evaluators.cpp index e3922c1be..3fb0d9896 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -1,7 +1,81 @@ + +#ifndef EIGEN_ENABLE_EVALUATORS #define EIGEN_ENABLE_EVALUATORS +#endif + #include "main.h" -using internal::copy_using_evaluator; +namespace Eigen { + + template + EIGEN_STRONG_INLINE + DstXprType& copy_using_evaluator(const EigenBase &dst, const SrcXprType &src) + { + call_assignment(dst.const_cast_derived(), src.derived(), internal::assign_op()); + return dst.const_cast_derived(); + } + + template class StorageBase, typename SrcXprType> + EIGEN_STRONG_INLINE + const DstXprType& copy_using_evaluator(const NoAlias& dst, const SrcXprType &src) + { + call_assignment(dst, src.derived(), internal::assign_op()); + return dst.expression(); + } + + template + EIGEN_STRONG_INLINE + DstXprType& copy_using_evaluator(const PlainObjectBase &dst, const SrcXprType &src) + { + #ifdef EIGEN_NO_AUTOMATIC_RESIZING + eigen_assert((dst.size()==0 || (IsVectorAtCompileTime ? (dst.size() == src.size()) + : (dst.rows() == src.rows() && dst.cols() == src.cols()))) + && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined"); + #else + dst.const_cast_derived().resizeLike(src.derived()); + #endif + + call_assignment(dst.const_cast_derived(), src.derived(), internal::assign_op()); + return dst.const_cast_derived(); + } + + template + void add_assign_using_evaluator(const DstXprType& dst, const SrcXprType& src) + { + typedef typename DstXprType::Scalar Scalar; + call_assignment(dst.const_cast_derived(), src.derived(), internal::add_assign_op()); + } + + template + void subtract_assign_using_evaluator(const DstXprType& dst, const SrcXprType& src) + { + typedef typename DstXprType::Scalar Scalar; + call_assignment(dst.const_cast_derived(), src.derived(), internal::sub_assign_op()); + } + + template + void multiply_assign_using_evaluator(const DstXprType& dst, const SrcXprType& src) + { + typedef typename DstXprType::Scalar Scalar; + call_assignment(dst.const_cast_derived(), src.derived(), internal::mul_assign_op()); + } + + template + void divide_assign_using_evaluator(const DstXprType& dst, const SrcXprType& src) + { + typedef typename DstXprType::Scalar Scalar; + call_assignment(dst.const_cast_derived(), src.derived(), internal::div_assign_op()); + } + + template + void swap_using_evaluator(const DstXprType& dst, const SrcXprType& src) + { + call_dense_swap_loop(dst.const_cast_derived(), src.const_cast_derived()); + } + +} + + using namespace std; #define VERIFY_IS_APPROX_EVALUATOR(DEST,EXPR) VERIFY_IS_APPROX(copy_using_evaluator(DEST,(EXPR)), (EXPR).eval()); From 49034d15709bcad1f5fd3a36a8cd5aa362e7591e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 27 Nov 2013 09:46:59 +0100 Subject: [PATCH 0200/1103] Fix bug #708: add placement new/delete for array --- Eigen/src/Core/util/Memory.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index a292495e5..d177e8b5a 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -664,7 +664,9 @@ template class aligned_stack_memory_handler /* memory allocated we can safely let the default implementation handle */ \ /* this particular case. */ \ static void *operator new(size_t size, void *ptr) { return ::operator new(size,ptr); } \ + static void *operator new[](size_t size, void* ptr) { return ::operator new[](size,ptr); } \ void operator delete(void * memory, void *ptr) throw() { return ::operator delete(memory,ptr); } \ + void operator delete[](void * memory, void *ptr) throw() { return ::operator delete[](memory,ptr); } \ /* nothrow-new (returns zero instead of std::bad_alloc) */ \ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ void operator delete(void *ptr, const std::nothrow_t&) throw() { \ From fc6ecebc69dcd11221233216d70746d495b1f29b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 27 Nov 2013 11:32:07 +0100 Subject: [PATCH 0201/1103] Simplify evaluator of EvalToTemp --- Eigen/src/Core/CoreEvaluators.h | 78 +++++++++------------------------ 1 file changed, 20 insertions(+), 58 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 3568cb85f..57960fcf6 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -105,7 +105,14 @@ struct evaluator_impl > RowsAtCompileTime = PlainObjectType::RowsAtCompileTime, ColsAtCompileTime = PlainObjectType::ColsAtCompileTime }; - + + evaluator_impl() + : m_data(0), + m_outerStride(IsVectorAtCompileTime ? 0 + : int(IsRowMajor) ? ColsAtCompileTime + : RowsAtCompileTime) + {} + evaluator_impl(const PlainObjectType& m) : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) { } @@ -188,6 +195,8 @@ struct evaluator_impl > : evaluator_impl > > { typedef Matrix XprType; + + evaluator_impl() {} evaluator_impl(const XprType& m) : evaluator_impl >(m) @@ -200,6 +209,8 @@ struct evaluator_impl > { typedef Array XprType; + evaluator_impl() {} + evaluator_impl(const XprType& m) : evaluator_impl >(m) { } @@ -246,80 +257,31 @@ class EvalToTemp template struct evaluator_impl > + : public evaluator::type { - typedef EvalToTemp XprType; - typedef typename ArgType::PlainObject PlainObject; + typedef EvalToTemp XprType; + typedef typename ArgType::PlainObject PlainObject; + typedef typename evaluator::type Base; evaluator_impl(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()), m_resultImpl(m_result) + : m_result(xpr.rows(), xpr.cols()) { + ::new (static_cast(this)) Base(m_result); // TODO we should simply do m_result(xpr.arg()); call_dense_assignment_loop(m_result, xpr.arg()); } // This constructor is used when nesting an EvalTo evaluator in another evaluator evaluator_impl(const ArgType& arg) - : m_result(arg.rows(), arg.cols()), m_resultImpl(m_result) + : m_result(arg.rows(), arg.cols()) { + ::new (static_cast(this)) Base(m_result); // TODO we should simply do m_result(xpr.arg()); call_dense_assignment_loop(m_result, arg); } - typedef typename PlainObject::Index Index; - typedef typename PlainObject::Scalar Scalar; - typedef typename PlainObject::CoeffReturnType CoeffReturnType; - typedef typename PlainObject::PacketScalar PacketScalar; - typedef typename PlainObject::PacketReturnType PacketReturnType; - - // All other functions are forwarded to m_resultImpl - - CoeffReturnType coeff(Index row, Index col) const - { - return m_resultImpl.coeff(row, col); - } - - CoeffReturnType coeff(Index index) const - { - return m_resultImpl.coeff(index); - } - - Scalar& coeffRef(Index row, Index col) - { - return m_resultImpl.coeffRef(row, col); - } - - Scalar& coeffRef(Index index) - { - return m_resultImpl.coeffRef(index); - } - - template - PacketReturnType packet(Index row, Index col) const - { - return m_resultImpl.template packet(row, col); - } - - template - PacketReturnType packet(Index index) const - { - return m_resultImpl.packet(index); - } - - template - void writePacket(Index row, Index col, const PacketScalar& x) - { - m_resultImpl.template writePacket(row, col, x); - } - - template - void writePacket(Index index, const PacketScalar& x) - { - m_resultImpl.template writePacket(index, x); - } - protected: PlainObject m_result; - typename evaluator::nestedType m_resultImpl; }; // -------------------- Transpose -------------------- From cc6dd878ee5a80375c587d018c2ae89fc2216dd1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 27 Nov 2013 17:32:57 +0100 Subject: [PATCH 0202/1103] Refactor dense product evaluators --- Eigen/Core | 5 +- Eigen/src/Core/Product.h | 35 +++- Eigen/src/Core/ProductEvaluators.h | 244 ++++++++++++++-------- Eigen/src/Core/util/Constants.h | 2 +- Eigen/src/Core/util/ForwardDeclarations.h | 17 +- 5 files changed, 201 insertions(+), 102 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 468ae0c76..e2c9c69cd 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -316,7 +316,6 @@ using std::ptrdiff_t; #include "src/Core/Product.h" #include "src/Core/CoreEvaluators.h" #include "src/Core/AssignEvaluator.h" -#include "src/Core/ProductEvaluators.h" #endif #ifndef EIGEN_PARSED_BY_DOXYGEN // work around Doxygen bug triggered by Assign.h r814874 @@ -382,6 +381,10 @@ using std::ptrdiff_t; #include "src/Core/BandMatrix.h" #include "src/Core/CoreIterators.h" +#ifdef EIGEN_ENABLE_EVALUATORS +#include "src/Core/ProductEvaluators.h" +#endif + #include "src/Core/BooleanRedux.h" #include "src/Core/Select.h" #include "src/Core/VectorwiseOp.h" diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 5d3789be7..52586e5c0 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -12,8 +12,7 @@ namespace Eigen { -template class Product; -template class ProductImpl; +template class ProductImpl; /** \class Product * \ingroup Core_Module @@ -24,13 +23,17 @@ template class ProductImpl; * \param Rhs the type of the right-hand side expression * * This class represents an expression of the product of two arbitrary matrices. + * + * The other template parameters are: + * \tparam Option can be DefaultProduct or LazyProduct + * \tparam ProductTag can be InnerProduct, OuterProduct, GemvProduct, GemmProduct. It is used to ease expression manipulations. * */ // Use ProductReturnType to get correct traits, in particular vectorization flags namespace internal { -template -struct traits > +template +struct traits > : traits::Type> { // We want A+B*C to be of type Product and not Product @@ -42,14 +45,15 @@ struct traits > } // end namespace internal -template -class Product : public ProductImpl::StorageKind, - typename internal::traits::StorageKind>::ret> +template +class Product : public ProductImpl::StorageKind, + typename internal::traits::StorageKind>::ret> { public: typedef typename ProductImpl< - Lhs, Rhs, + Lhs, Rhs, Option, ProductTag, typename internal::promote_storage_type::ret>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Product) @@ -78,13 +82,13 @@ class Product : public ProductImpl -class ProductImpl : public internal::dense_xpr_base >::type +template +class ProductImpl : public internal::dense_xpr_base >::type { typedef Product Derived; public: - typedef typename internal::dense_xpr_base >::type Base; + typedef typename internal::dense_xpr_base >::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) }; @@ -102,6 +106,15 @@ prod(const Lhs& lhs, const Rhs& rhs) return Product(lhs,rhs); } +/** \internal used to test the evaluator only + */ +template +const Product +lazyprod(const Lhs& lhs, const Rhs& rhs) +{ + return Product(lhs,rhs); +} + } // end namespace Eigen #endif // EIGEN_PRODUCT_H diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 855914f2e..42dd3c7ac 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -17,94 +17,172 @@ namespace Eigen { namespace internal { -// We can evaluate the product either all at once, like GeneralProduct and its evalTo() function, or -// traverse the matrix coefficient by coefficient, like CoeffBasedProduct. Use the existing logic -// in ProductReturnType to decide. + +// Helper class to perform a dense product with the destination at hand. +// Depending on the sizes of the factors, there are different evaluation strategies +// as controlled by internal::product_type. +template::value> +struct dense_product_impl; -template -struct product_evaluator_dispatcher; -template -struct evaluator_impl > - : product_evaluator_dispatcher, typename ProductReturnType::Type> +// The evaluator for default dense products creates a temporary and call dense_product_impl +template +struct evaluator_impl > + : public evaluator::PlainObject>::type { - typedef Product XprType; - typedef product_evaluator_dispatcher::Type> Base; - - evaluator_impl(const XprType& xpr) : Base(xpr) - { } -}; - -template -struct product_evaluator_traits_dispatcher; - -template -struct evaluator_traits > - : product_evaluator_traits_dispatcher, typename ProductReturnType::Type> -{ - static const int AssumeAliasing = 1; -}; - -// Case 1: Evaluate all at once -// -// We can view the GeneralProduct class as a part of the product evaluator. -// Four sub-cases: InnerProduct, OuterProduct, GemmProduct and GemvProduct. -// InnerProduct is special because GeneralProduct does not have an evalTo() method in this case. - -template -struct product_evaluator_traits_dispatcher, GeneralProduct > -{ - static const int HasEvalTo = 0; -}; - -template -struct product_evaluator_dispatcher, GeneralProduct > - : public evaluator::PlainObject>::type -{ - typedef Product XprType; + typedef Product XprType; typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type evaluator_base; + typedef typename evaluator::type Base; - // TODO: Computation is too early (?) - product_evaluator_dispatcher(const XprType& xpr) : evaluator_base(m_result) + evaluator_impl(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) { - m_result.coeffRef(0,0) = (xpr.lhs().transpose().cwiseProduct(xpr.rhs())).sum(); + ::new (static_cast(this)) Base(m_result); + dense_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); } protected: PlainObject m_result; }; -// For the other three subcases, simply call the evalTo() method of GeneralProduct -// TODO: GeneralProduct should take evaluators, not expression objects. -template -struct product_evaluator_traits_dispatcher, GeneralProduct > +template +struct dense_product_impl { - static const int HasEvalTo = 1; -}; - -template -struct product_evaluator_dispatcher, GeneralProduct > -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type evaluator_base; - - product_evaluator_dispatcher(const XprType& xpr) : m_xpr(xpr) - { } - - template - void evalTo(DstEvaluatorType /* not used */, DstXprType& dst) const + template + static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { - dst.resize(m_xpr.rows(), m_xpr.cols()); - GeneralProduct(m_xpr.lhs(), m_xpr.rhs()).evalTo(dst); + dst.coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum(); } -protected: - const XprType& m_xpr; + template + static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + dst.coeffRef(0,0) += (lhs.transpose().cwiseProduct(rhs)).sum(); + } + + template + static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { dst.coeffRef(0,0) -= (lhs.transpose().cwiseProduct(rhs)).sum(); } }; + + +template +struct dense_product_impl +{ + typedef typename Product::Scalar Scalar; + + template + static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + // TODO bypass GeneralProduct class + GeneralProduct(lhs,rhs).evalTo(dst); + } + + template + static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + // TODO bypass GeneralProduct class + GeneralProduct(lhs,rhs).addTo(dst); + } + + template + static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + // TODO bypass GeneralProduct class + GeneralProduct(lhs,rhs).subTo(dst); + } + + template + static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + // TODO bypass GeneralProduct class + GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); + } + +}; + + +// This base class provides default implementations for evalTo, addTo, subTo, in terms of scaleAndAddTo +template +struct dense_product_impl_base +{ + typedef typename Product::Scalar Scalar; + + template + static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { dst.setZero(); scaleAndAddTo(dst, lhs, rhs, Scalar(1)); } + + template + static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { scaleAndAddTo(dst,lhs, rhs, Scalar(1)); } + + template + static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); } + + template + static void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { Derived::scaleAndAddTo(dst,lhs,rhs,alpha); } + +}; + +template +struct dense_product_impl : dense_product_impl_base > +{ + typedef typename Product::Scalar Scalar; + enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight }; + typedef typename internal::conditional::type MatrixType; + + template + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + internal::gemv_selector::HasUsableDirectAccess) + >::run(GeneralProduct(lhs,rhs), dst, alpha); + } +}; + +template +struct dense_product_impl : dense_product_impl_base > +{ + typedef typename Product::Scalar Scalar; + + template + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + // TODO bypass GeneralProduct class + GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); + } +}; + +template +struct dense_product_impl +{ + typedef typename Product::Scalar Scalar; + + template + static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { dst = lazyprod(lhs,rhs); } + + template + static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { dst += lazyprod(lhs,rhs); } + + template + static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { dst -= lazyprod(lhs,rhs); } + + template + static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { dst += alpha * lazyprod(lhs,rhs); } +}; + +template +struct dense_product_impl : dense_product_impl {}; + // Case 2: Evaluate coeff by coeff // // This is mostly taken from CoeffBasedProduct.h @@ -117,20 +195,14 @@ struct etor_product_coeff_impl; template struct etor_product_packet_impl; -template -struct product_evaluator_traits_dispatcher, CoeffBasedProduct > +template +struct evaluator_impl > + : evaluator_impl_base > { - static const int HasEvalTo = 0; -}; + typedef Product XprType; + typedef CoeffBasedProduct CoeffBasedProductType; -template -struct product_evaluator_dispatcher, CoeffBasedProduct > - : evaluator_impl_base > -{ - typedef Product XprType; - typedef CoeffBasedProduct CoeffBasedProductType; - - product_evaluator_dispatcher(const XprType& xpr) + evaluator_impl(const XprType& xpr) : m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()), m_innerDim(xpr.lhs().cols()) @@ -150,11 +222,13 @@ struct product_evaluator_dispatcher, CoeffBasedProduct::InnerSize, CoeffReadCost = traits::CoeffReadCost, Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT, - CanVectorizeInner = traits::CanVectorizeInner + CanVectorizeInner = traits::CanVectorizeInner, + Flags = CoeffBasedProductType::Flags }; typedef typename evaluator::type LhsEtorType; typedef typename evaluator::type RhsEtorType; + typedef etor_product_coeff_impl CoeffImpl; @@ -183,8 +257,8 @@ struct product_evaluator_dispatcher, CoeffBasedProduct PacketImpl; + Unroll ? InnerSize-1 : Dynamic, + LhsEtorType, RhsEtorType, PacketScalar, LoadMode> PacketImpl; PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res); return res; } @@ -197,6 +271,7 @@ protected: Index m_innerDim; }; + /*************************************************************************** * Normal product .coeff() implementation (with meta-unrolling) ***************************************************************************/ @@ -275,7 +350,6 @@ struct etor_product_coeff_impl::run(row, col, lhs, rhs, innerDim, pres); - etor_product_coeff_impl::run(row, col, lhs, rhs, innerDim, res); res = predux(pres); } }; diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 05107fdfe..3178ff06e 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -425,7 +425,7 @@ namespace Architecture /** \internal \ingroup enums * Enum used as template parameter in GeneralProduct. */ -enum { CoeffBasedProductMode, LazyCoeffBasedProductMode, OuterProduct, InnerProduct, GemvProduct, GemmProduct }; +enum { DefaultProduct=0, CoeffBasedProductMode, LazyCoeffBasedProductMode, LazyProduct, OuterProduct, InnerProduct, GemvProduct, GemmProduct }; /** \internal \ingroup enums * Enum used in experimental parallel implementation. */ diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 0a2144c69..459422524 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -87,11 +87,20 @@ template class CwiseNullaryOp; template class CwiseUnaryOp; template class CwiseUnaryView; template class CwiseBinaryOp; -template class SelfCwiseBinaryOp; +template class SelfCwiseBinaryOp; // TODO deprecated template class ProductBase; -template class Product; -template class GeneralProduct; -template class CoeffBasedProduct; + +namespace internal { + template struct product_tag; +} + +template::ret + > class Product; + +template class GeneralProduct; // TODO deprecated +template class CoeffBasedProduct; // TODO deprecated template class DiagonalBase; template class DiagonalWrapper; From 558427532565bf9c81bd4473591cfc6b552deff2 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Nov 2013 13:38:59 +0100 Subject: [PATCH 0203/1103] Remove HasEvalTo and all at once eval mode --- Eigen/src/Core/AssignEvaluator.h | 23 ++--------------------- Eigen/src/Core/CoreEvaluators.h | 22 +--------------------- 2 files changed, 3 insertions(+), 42 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 158dc7842..99ae3f89d 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -32,8 +32,7 @@ public: DstIsAligned = Derived::Flags & AlignedBit, DstHasDirectAccess = Derived::Flags & DirectAccessBit, SrcIsAligned = OtherDerived::Flags & AlignedBit, - JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned, - SrcEvalBeforeAssign = (evaluator_traits::HasEvalTo == 1) + JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned }; private: @@ -68,8 +67,7 @@ private: public: enum { - Traversal = int(SrcEvalBeforeAssign) ? int(AllAtOnceTraversal) - : int(MayInnerVectorize) ? int(InnerVectorizedTraversal) + Traversal = int(MayInnerVectorize) ? int(InnerVectorizedTraversal) : int(MayLinearVectorize) ? int(LinearVectorizedTraversal) : int(MaySliceVectorize) ? int(SliceVectorizedTraversal) : int(MayLinearize) ? int(LinearTraversal) @@ -496,23 +494,6 @@ struct dense_assignment_loop } }; -/**************************** -*** All-at-once traversal *** -****************************/ - -// TODO: this 'AllAtOnceTraversal' should be dropped or caught earlier (Gael) -// Indeed, what to do with the kernel's functor?? -template -struct dense_assignment_loop -{ - static inline void run(Kernel & kernel) - { - // Evaluate rhs in temporary to prevent aliasing problems in a = a * a; - // TODO: Do not pass the xpr object to evalTo() (Jitse) - kernel.srcEvaluator().evalTo(kernel.dstEvaluator(), kernel.dstExpression()); - } -}; - /*************************************************************************** * Part 4 : Generic dense assignment kernel ***************************************************************************/ diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 57960fcf6..c998c37cb 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -22,10 +22,6 @@ namespace internal { template struct evaluator_traits { - // 1 if evaluator_impl::evalTo() exists - // 0 if evaluator_impl allows coefficient-based access - static const int HasEvalTo = 0; - // 1 if assignment A = B assumes aliasing when B is of type T and thus B needs to be evaluated into a // temporary; 0 if not. static const int AssumeAliasing = 0; @@ -37,32 +33,16 @@ template class EvalToTemp; // evaluator::type is type of evaluator for T -// evaluator::nestedType is type of evaluator if T is nested inside another evaluator template struct evaluator_impl { }; - -template::HasEvalTo> -struct evaluator_nested_type; - -template -struct evaluator_nested_type -{ - typedef evaluator_impl type; -}; - -template -struct evaluator_nested_type -{ - typedef evaluator_impl > type; -}; template struct evaluator { typedef evaluator_impl type; - typedef typename evaluator_nested_type::type nestedType; + typedef evaluator_impl nestedType; }; // TODO: Think about const-correctness From d331def6cce11fa4635a0b25cfda294805763955 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Nov 2013 16:18:22 +0100 Subject: [PATCH 0204/1103] add definition of product_tag --- Eigen/src/Core/GeneralProduct.h | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index e3a165ac6..c9ab63782 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -98,6 +98,31 @@ public: #endif }; +template struct product_tag +{ +private: + + typedef typename remove_all::type _Lhs; + typedef typename remove_all::type _Rhs; + enum { + Rows = _Lhs::RowsAtCompileTime, + Cols = _Rhs::ColsAtCompileTime, + Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, _Rhs::RowsAtCompileTime) + }; + + enum { + rows_select = Rows==1 ? int(Rows) : int(Large), + cols_select = Cols==1 ? int(Cols) : int(Large), + depth_select = Depth==1 ? int(Depth) : int(Large) + }; + typedef product_type_selector selector; + +public: + enum { + ret = selector::ret + }; + +}; /* The following allows to select the kind of product at compile time * based on the three dimensions of the product. From fb6e32a62fe288aa7236394fc16eb57f3fc40335 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Nov 2013 16:45:47 +0100 Subject: [PATCH 0205/1103] Get rid of evalautor_impl --- Eigen/src/Core/CoreEvaluators.h | 181 +++++++++++++++-------------- Eigen/src/Core/ProductEvaluators.h | 13 ++- 2 files changed, 102 insertions(+), 92 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index c998c37cb..c63ff8acd 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -17,33 +17,33 @@ namespace Eigen { namespace internal { -// evaluator_traits contains traits for evaluator_impl +template struct evaluator; -template -struct evaluator_traits +// evaluator_traits contains traits for evaluator + + template +struct evaluator_traits_base { + // TODO check whether these two indirections are really needed. + // Basically, if nobody overwrite type and nestedType, then, they can be dropped + typedef evaluator type; + typedef evaluator nestedType; + // 1 if assignment A = B assumes aliasing when B is of type T and thus B needs to be evaluated into a // temporary; 0 if not. static const int AssumeAliasing = 0; }; +template +struct evaluator_traits : public evaluator_traits_base +{ +}; + // expression class for evaluating nested expression to a temporary template class EvalToTemp; -// evaluator::type is type of evaluator for T - -template -struct evaluator_impl -{ }; - -template -struct evaluator -{ - typedef evaluator_impl type; - typedef evaluator_impl nestedType; -}; // TODO: Think about const-correctness @@ -56,26 +56,24 @@ struct evaluator // TODO this class does not seem to be necessary anymore template -struct evaluator_impl_base +struct evaluator_base { + typedef typename evaluator_traits::type type; + typedef typename evaluator_traits::nestedType nestedType; + typedef typename ExpressionType::Index Index; // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices. typedef traits ExpressionTraits; - - evaluator_impl& derived() - { - return *static_cast*>(this); - } }; // -------------------- Matrix and Array -------------------- // -// evaluator_impl is a common base class for the +// evaluator is a common base class for the // Matrix and Array evaluators. template -struct evaluator_impl > - : evaluator_impl_base +struct evaluator > + : evaluator_base { typedef PlainObjectBase PlainObjectType; @@ -86,14 +84,14 @@ struct evaluator_impl > ColsAtCompileTime = PlainObjectType::ColsAtCompileTime }; - evaluator_impl() + evaluator() : m_data(0), m_outerStride(IsVectorAtCompileTime ? 0 : int(IsRowMajor) ? ColsAtCompileTime : RowsAtCompileTime) {} - evaluator_impl(const PlainObjectType& m) + evaluator(const PlainObjectType& m) : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) { } @@ -171,28 +169,28 @@ protected: }; template -struct evaluator_impl > - : evaluator_impl > > +struct evaluator > + : evaluator > > { typedef Matrix XprType; - evaluator_impl() {} + evaluator() {} - evaluator_impl(const XprType& m) - : evaluator_impl >(m) + evaluator(const XprType& m) + : evaluator >(m) { } }; template -struct evaluator_impl > - : evaluator_impl > > +struct evaluator > + : evaluator > > { typedef Array XprType; - evaluator_impl() {} + evaluator() {} - evaluator_impl(const XprType& m) - : evaluator_impl >(m) + evaluator(const XprType& m) + : evaluator >(m) { } }; @@ -236,14 +234,17 @@ class EvalToTemp }; template -struct evaluator_impl > +struct evaluator > : public evaluator::type { typedef EvalToTemp XprType; typedef typename ArgType::PlainObject PlainObject; typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; - evaluator_impl(const XprType& xpr) + evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); @@ -252,7 +253,7 @@ struct evaluator_impl > } // This constructor is used when nesting an EvalTo evaluator in another evaluator - evaluator_impl(const ArgType& arg) + evaluator(const ArgType& arg) : m_result(arg.rows(), arg.cols()) { ::new (static_cast(this)) Base(m_result); @@ -267,12 +268,12 @@ protected: // -------------------- Transpose -------------------- template -struct evaluator_impl > - : evaluator_impl_base > +struct evaluator > + : evaluator_base > { typedef Transpose XprType; - evaluator_impl(const XprType& t) : m_argImpl(t.nestedExpression()) {} + evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -331,11 +332,12 @@ protected: // -------------------- CwiseNullaryOp -------------------- template -struct evaluator_impl > +struct evaluator > + : evaluator_base > { typedef CwiseNullaryOp XprType; - evaluator_impl(const XprType& n) + evaluator(const XprType& n) : m_functor(n.functor()) { } @@ -372,11 +374,12 @@ protected: // -------------------- CwiseUnaryOp -------------------- template -struct evaluator_impl > +struct evaluator > + : evaluator_base > { typedef CwiseUnaryOp XprType; - evaluator_impl(const XprType& op) + evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -415,11 +418,12 @@ protected: // -------------------- CwiseBinaryOp -------------------- template -struct evaluator_impl > +struct evaluator > + : evaluator_base > { typedef CwiseBinaryOp XprType; - evaluator_impl(const XprType& xpr) + evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) @@ -462,12 +466,12 @@ protected: // -------------------- CwiseUnaryView -------------------- template -struct evaluator_impl > - : evaluator_impl_base > +struct evaluator > + : evaluator_base > { typedef CwiseUnaryView XprType; - evaluator_impl(const XprType& op) + evaluator(const XprType& op) : m_unaryOp(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -504,8 +508,8 @@ protected: // -------------------- Map -------------------- template -struct evaluator_impl > - : evaluator_impl_base +struct evaluator > + : evaluator_base { typedef MapBase MapType; typedef Derived XprType; @@ -517,7 +521,7 @@ struct evaluator_impl > typedef typename XprType::PacketScalar PacketScalar; typedef typename XprType::PacketReturnType PacketReturnType; - evaluator_impl(const XprType& map) + evaluator(const XprType& map) : m_data(const_cast(map.data())), m_rowStride(map.rowStride()), m_colStride(map.colStride()) @@ -585,13 +589,13 @@ protected: }; template -struct evaluator_impl > - : public evaluator_impl > > +struct evaluator > + : public evaluator > > { typedef Map XprType; - evaluator_impl(const XprType& map) - : evaluator_impl >(map) + evaluator(const XprType& map) + : evaluator >(map) { } }; @@ -601,17 +605,17 @@ template::ret> struct block_evaluator; template -struct evaluator_impl > +struct evaluator > : block_evaluator { typedef Block XprType; typedef block_evaluator block_evaluator_type; - evaluator_impl(const XprType& block) : block_evaluator_type(block) {} + evaluator(const XprType& block) : block_evaluator_type(block) {} }; template struct block_evaluator - : evaluator_impl_base > + : evaluator_base > { typedef Block XprType; @@ -691,12 +695,12 @@ protected: template struct block_evaluator - : evaluator_impl > > + : evaluator > > { typedef Block XprType; block_evaluator(const XprType& block) - : evaluator_impl >(block) + : evaluator >(block) { } }; @@ -704,11 +708,12 @@ struct block_evaluator -struct evaluator_impl > +struct evaluator > + : evaluator_base > { typedef Select XprType; - evaluator_impl(const XprType& select) + evaluator(const XprType& select) : m_conditionImpl(select.conditionMatrix()), m_thenImpl(select.thenMatrix()), m_elseImpl(select.elseMatrix()) @@ -743,11 +748,12 @@ protected: // -------------------- Replicate -------------------- template -struct evaluator_impl > +struct evaluator > + : evaluator_base > { typedef Replicate XprType; - evaluator_impl(const XprType& replicate) + evaluator(const XprType& replicate) : m_argImpl(replicate.nestedExpression()), m_rows(replicate.nestedExpression().rows()), m_cols(replicate.nestedExpression().cols()) @@ -797,11 +803,12 @@ protected: // the row() and col() member functions. template< typename ArgType, typename MemberOp, int Direction> -struct evaluator_impl > +struct evaluator > + : evaluator_base > { typedef PartialReduxExpr XprType; - evaluator_impl(const XprType expr) + evaluator(const XprType expr) : m_expr(expr) { } @@ -825,16 +832,16 @@ protected: // -------------------- MatrixWrapper and ArrayWrapper -------------------- // -// evaluator_impl_wrapper_base is a common base class for the +// evaluator_wrapper_base is a common base class for the // MatrixWrapper and ArrayWrapper evaluators. template -struct evaluator_impl_wrapper_base - : evaluator_impl_base +struct evaluator_wrapper_base + : evaluator_base { typedef typename remove_all::type ArgType; - evaluator_impl_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} + evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} typedef typename ArgType::Index Index; typedef typename ArgType::Scalar Scalar; @@ -891,24 +898,24 @@ protected: }; template -struct evaluator_impl > - : evaluator_impl_wrapper_base > +struct evaluator > + : evaluator_wrapper_base > { typedef MatrixWrapper XprType; - evaluator_impl(const XprType& wrapper) - : evaluator_impl_wrapper_base >(wrapper.nestedExpression()) + evaluator(const XprType& wrapper) + : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; template -struct evaluator_impl > - : evaluator_impl_wrapper_base > +struct evaluator > + : evaluator_wrapper_base > { typedef ArrayWrapper XprType; - evaluator_impl(const XprType& wrapper) - : evaluator_impl_wrapper_base >(wrapper.nestedExpression()) + evaluator(const XprType& wrapper) + : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; @@ -919,8 +926,8 @@ struct evaluator_impl > template struct reverse_packet_cond; template -struct evaluator_impl > - : evaluator_impl_base > +struct evaluator > + : evaluator_base > { typedef Reverse XprType; typedef typename XprType::Index Index; @@ -943,7 +950,7 @@ struct evaluator_impl > }; typedef internal::reverse_packet_cond reverse_packet; - evaluator_impl(const XprType& reverse) + evaluator(const XprType& reverse) : m_argImpl(reverse.nestedExpression()), m_rows(ReverseRow ? reverse.nestedExpression().rows() : 0), m_cols(ReverseCol ? reverse.nestedExpression().cols() : 0) @@ -1013,12 +1020,12 @@ protected: // -------------------- Diagonal -------------------- template -struct evaluator_impl > - : evaluator_impl_base > +struct evaluator > + : evaluator_base > { typedef Diagonal XprType; - evaluator_impl(const XprType& diagonal) + evaluator(const XprType& diagonal) : m_argImpl(diagonal.nestedExpression()), m_index(diagonal.index()) { } diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 42dd3c7ac..9f5f6eb0c 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -27,14 +27,17 @@ struct dense_product_impl; // The evaluator for default dense products creates a temporary and call dense_product_impl template -struct evaluator_impl > +struct evaluator > : public evaluator::PlainObject>::type { typedef Product XprType; typedef typename XprType::PlainObject PlainObject; typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; - evaluator_impl(const XprType& xpr) + evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); @@ -196,13 +199,13 @@ template -struct evaluator_impl > - : evaluator_impl_base > +struct evaluator > + : evaluator_base > { typedef Product XprType; typedef CoeffBasedProduct CoeffBasedProductType; - evaluator_impl(const XprType& xpr) + evaluator(const XprType& xpr) : m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()), m_innerDim(xpr.lhs().cols()) From c15c65990fc322891b10a6a20db9bc6c127fae26 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Nov 2013 17:50:59 +0100 Subject: [PATCH 0206/1103] First step toward the generalization of evaluators to triangular, sparse and other fancyness. Remove product_tag template parameter to Product. --- Eigen/src/Core/AssignEvaluator.h | 6 ++- Eigen/src/Core/CoreEvaluators.h | 66 ++++++++++++++++++++--- Eigen/src/Core/Product.h | 28 +++++----- Eigen/src/Core/ProductEvaluators.h | 40 ++++++++++---- Eigen/src/Core/util/ForwardDeclarations.h | 5 +- 5 files changed, 109 insertions(+), 36 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 99ae3f89d..2f18cbc95 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -646,7 +646,7 @@ struct SparseShape {}; // Based on the respective shapes of the destination and source, // the class AssignmentKind determine the kind of assignment mechanism. // AssignmentKind must define a Kind typedef. -template struct AssignmentKind; +template struct AssignmentKind; // AssignmentKind<.,.>::Kind can be one of the following: struct Dense2Dense {}; @@ -655,9 +655,11 @@ template struct AssignmentKind; struct Sparse2Dense {}; struct Sparse2Sparse {}; +template<> struct AssignmentKind { typedef Dense2Dense Kind; }; + // This is the main assignment class template< typename DstXprType, typename SrcXprType, typename Functor, - typename Kind = Dense2Dense,//AssignmentKind< evaluator::Shape , evaluator::Shape >::Kind, + typename Kind = typename AssignmentKind< typename evaluator_traits::Shape , typename evaluator_traits::Shape >::Kind, typename Scalar = typename DstXprType::Scalar> struct Assignment; diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index c63ff8acd..9460e675f 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -14,20 +14,70 @@ #define EIGEN_COREEVALUATORS_H namespace Eigen { - + namespace internal { -template struct evaluator; +struct IndexBased {}; +struct IteratorBased {}; + +// This class returns the evaluator kind from the expression storage kind. +// Default assumes index based accessors +template +struct storage_kind_to_evaluator_kind { + typedef IndexBased Kind; +}; + +// TODO to be moved to SparseCore: +/* +template<> +struct storage_kind_to_evaluator_kind { + typedef IteratorBased Kind +}; +*/ + +// This class returns the evaluator shape from the expression storage kind. +// It can be Dense, Sparse, Triangular, Diagonal, SelfAdjoint, Band, etc. +template struct storage_kind_to_shape; + + +template<> +struct storage_kind_to_shape { + typedef Dense Shape; +}; + +// TODO to be moved to SparseCore: +/* +template<> +struct storage_kind_to_shape { + typedef Sparse Shape; +}; +*/ + +template struct evaluator_traits; + +template< typename T, + typename Kind = typename evaluator_traits::Kind, + typename Scalar = typename T::Scalar> struct evaluator; + +template< typename T, + typename LhsKind = typename evaluator_traits::Kind, + typename RhsKind = typename evaluator_traits::Kind, + typename LhsScalar = typename T::Lhs::Scalar, + typename RhsScalar = typename T::Rhs::Scalar> struct binary_evaluator; // evaluator_traits contains traits for evaluator - template +template struct evaluator_traits_base { // TODO check whether these two indirections are really needed. // Basically, if nobody overwrite type and nestedType, then, they can be dropped - typedef evaluator type; - typedef evaluator nestedType; +// typedef evaluator type; +// typedef evaluator nestedType; + + // by default, get evalautor kind and shape from storage + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef typename storage_kind_to_shape::Shape Shape; // 1 if assignment A = B assumes aliasing when B is of type T and thus B needs to be evaluated into a // temporary; 0 if not. @@ -58,8 +108,10 @@ struct evaluator template struct evaluator_base { - typedef typename evaluator_traits::type type; - typedef typename evaluator_traits::nestedType nestedType; +// typedef typename evaluator_traits::type type; +// typedef typename evaluator_traits::nestedType nestedType; + typedef evaluator type; + typedef evaluator nestedType; typedef typename ExpressionType::Index Index; // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices. diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 52586e5c0..970d257a5 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -12,7 +12,7 @@ namespace Eigen { -template class ProductImpl; +template class ProductImpl; /** \class Product * \ingroup Core_Module @@ -26,14 +26,13 @@ template -struct traits > +template +struct traits > : traits::Type> { // We want A+B*C to be of type Product and not Product @@ -45,18 +44,23 @@ struct traits > } // end namespace internal -template -class Product : public ProductImpl::StorageKind, - typename internal::traits::StorageKind>::ret> +template +class Product : public ProductImpl<_Lhs,_Rhs,Option, + typename internal::promote_storage_type::StorageKind, + typename internal::traits<_Rhs>::StorageKind>::ret> { public: + typedef _Lhs Lhs; + typedef _Rhs Rhs; + typedef typename ProductImpl< - Lhs, Rhs, Option, ProductTag, + Lhs, Rhs, Option, typename internal::promote_storage_type::ret>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Product) + + typedef typename Lhs::Nested LhsNested; typedef typename Rhs::Nested RhsNested; @@ -82,13 +86,13 @@ class Product : public ProductImpl -class ProductImpl : public internal::dense_xpr_base >::type +template +class ProductImpl : public internal::dense_xpr_base >::type { typedef Product Derived; public: - typedef typename internal::dense_xpr_base >::type Base; + typedef typename internal::dense_xpr_base >::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) }; diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 9f5f6eb0c..e3a893651 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -16,7 +16,28 @@ namespace Eigen { namespace internal { + +// Like more general binary expressions, products need they own evaluator: +template< typename T, + int ProductTag = internal::product_tag::ret, + typename LhsShape = typename evaluator_traits::Shape, + typename RhsShape = typename evaluator_traits::Shape, + typename LhsScalar = typename T::Lhs::Scalar, + typename RhsScalar = typename T::Rhs::Scalar + > struct product_evaluator; + +template +struct evaluator > + : public product_evaluator > +{ + typedef Product XprType; + typedef product_evaluator Base; + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const XprType& xpr) : Base(xpr) {} +}; // Helper class to perform a dense product with the destination at hand. // Depending on the sizes of the factors, there are different evaluation strategies @@ -27,17 +48,14 @@ struct dense_product_impl; // The evaluator for default dense products creates a temporary and call dense_product_impl template -struct evaluator > - : public evaluator::PlainObject>::type +struct product_evaluator, ProductTag, Dense, Dense, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type { - typedef Product XprType; + typedef Product XprType; typedef typename XprType::PlainObject PlainObject; typedef typename evaluator::type Base; - - typedef evaluator type; - typedef evaluator nestedType; - evaluator(const XprType& xpr) + product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); @@ -199,13 +217,13 @@ template -struct evaluator > - : evaluator_base > +struct product_evaluator, ProductTag, Dense, Dense, typename Lhs::Scalar, typename Rhs::Scalar > + : evaluator_base > { - typedef Product XprType; + typedef Product XprType; typedef CoeffBasedProduct CoeffBasedProductType; - evaluator(const XprType& xpr) + product_evaluator(const XprType& xpr) : m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()), m_innerDim(xpr.lhs().cols()) diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 459422524..776eac587 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -94,10 +94,7 @@ namespace internal { template struct product_tag; } -template::ret - > class Product; +template class Product; template class GeneralProduct; // TODO deprecated template class CoeffBasedProduct; // TODO deprecated From d61345f36671d0cc04378bda0f60ca8378d20b4d Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 29 Nov 2013 19:42:11 +0100 Subject: [PATCH 0207/1103] Fix bug #609: Euler angles are in Range [0:pi]x[-pi:pi]x[-pi:pi]. Now the unit test verifies this (also that it is bijective in this range). --- Eigen/src/Geometry/EulerAngles.h | 2 +- test/geo_eulerangles.cpp | 67 +++++++++++++++++++++----------- 2 files changed, 45 insertions(+), 24 deletions(-) diff --git a/Eigen/src/Geometry/EulerAngles.h b/Eigen/src/Geometry/EulerAngles.h index 97984d590..82802fb43 100644 --- a/Eigen/src/Geometry/EulerAngles.h +++ b/Eigen/src/Geometry/EulerAngles.h @@ -28,7 +28,7 @@ namespace Eigen { * * AngleAxisf(ea[2], Vector3f::UnitZ()); \endcode * This corresponds to the right-multiply conventions (with right hand side frames). * - * The returned angles are in the ranges [0:pi]x[0:pi]x[-pi:pi]. + * The returned angles are in the ranges [0:pi]x[-pi:pi]x[-pi:pi]. * * \sa class AngleAxis */ diff --git a/test/geo_eulerangles.cpp b/test/geo_eulerangles.cpp index 5445cd81a..b4830bd41 100644 --- a/test/geo_eulerangles.cpp +++ b/test/geo_eulerangles.cpp @@ -12,36 +12,48 @@ #include #include -template void check_all_var(const Matrix& ea) + +template +void verify_euler(const Matrix& ea, int i, int j, int k) { typedef Matrix Matrix3; typedef Matrix Vector3; typedef AngleAxis AngleAxisx; using std::abs; + Matrix3 m(AngleAxisx(ea[0], Vector3::Unit(i)) * AngleAxisx(ea[1], Vector3::Unit(j)) * AngleAxisx(ea[2], Vector3::Unit(k))); + Vector3 eabis = m.eulerAngles(i, j, k); + Matrix3 mbis(AngleAxisx(eabis[0], Vector3::Unit(i)) * AngleAxisx(eabis[1], Vector3::Unit(j)) * AngleAxisx(eabis[2], Vector3::Unit(k))); + VERIFY_IS_APPROX(m, mbis); + /* If I==K, and ea[1]==0, then there no unique solution. */ + /* The remark apply in the case where I!=K, and |ea[1]| is close to pi/2. */ + if( (i!=k || ea[1]!=0) && (i==k || !internal::isApprox(abs(ea[1]),Scalar(M_PI/2),test_precision())) ) + VERIFY((ea-eabis).norm() <= test_precision()); - #define VERIFY_EULER(I,J,K, X,Y,Z) { \ - Matrix3 m(AngleAxisx(ea[0], Vector3::Unit##X()) * AngleAxisx(ea[1], Vector3::Unit##Y()) * AngleAxisx(ea[2], Vector3::Unit##Z())); \ - Vector3 eabis = m.eulerAngles(I,J,K); \ - Matrix3 mbis(AngleAxisx(eabis[0], Vector3::Unit##X()) * AngleAxisx(eabis[1], Vector3::Unit##Y()) * AngleAxisx(eabis[2], Vector3::Unit##Z())); \ - VERIFY_IS_APPROX(m, mbis); \ - /* If I==K, and ea[1]==0, then there no unique solution. */ \ - /* The remark apply in the case where I!=K, and |ea[1]| is close to pi/2. */ \ - if( (I!=K || ea[1]!=0) && (I==K || !internal::isApprox(abs(ea[1]),Scalar(M_PI/2),test_precision())) ) VERIFY((ea-eabis).norm() <= test_precision()); \ - } - VERIFY_EULER(0,1,2, X,Y,Z); - VERIFY_EULER(0,1,0, X,Y,X); - VERIFY_EULER(0,2,1, X,Z,Y); - VERIFY_EULER(0,2,0, X,Z,X); + // approx_or_less_than does not work for 0 + VERIFY(0 < eabis[0] || test_isMuchSmallerThan(eabis[0], Scalar(1))); + VERIFY_IS_APPROX_OR_LESS_THAN(eabis[0], Scalar(M_PI)); + VERIFY_IS_APPROX_OR_LESS_THAN(-Scalar(M_PI), eabis[1]); + VERIFY_IS_APPROX_OR_LESS_THAN(eabis[1], Scalar(M_PI)); + VERIFY_IS_APPROX_OR_LESS_THAN(-Scalar(M_PI), eabis[2]); + VERIFY_IS_APPROX_OR_LESS_THAN(eabis[2], Scalar(M_PI)); +} - VERIFY_EULER(1,2,0, Y,Z,X); - VERIFY_EULER(1,2,1, Y,Z,Y); - VERIFY_EULER(1,0,2, Y,X,Z); - VERIFY_EULER(1,0,1, Y,X,Y); +template void check_all_var(const Matrix& ea) +{ + verify_euler(ea, 0,1,2); + verify_euler(ea, 0,1,0); + verify_euler(ea, 0,2,1); + verify_euler(ea, 0,2,0); - VERIFY_EULER(2,0,1, Z,X,Y); - VERIFY_EULER(2,0,2, Z,X,Z); - VERIFY_EULER(2,1,0, Z,Y,X); - VERIFY_EULER(2,1,2, Z,Y,Z); + verify_euler(ea, 1,2,0); + verify_euler(ea, 1,2,1); + verify_euler(ea, 1,0,2); + verify_euler(ea, 1,0,1); + + verify_euler(ea, 2,0,1); + verify_euler(ea, 2,0,2); + verify_euler(ea, 2,1,0); + verify_euler(ea, 2,1,2); } template void eulerangles() @@ -63,7 +75,16 @@ template void eulerangles() ea = m.eulerAngles(0,1,0); check_all_var(ea); - ea = (Array3::Random() + Array3(1,1,0))*Scalar(M_PI)*Array3(0.5,0.5,1); + // Check with purely random Quaternion: + q1.coeffs() = Quaternionx::Coefficients::Random().normalized(); + m = q1; + ea = m.eulerAngles(0,1,2); + check_all_var(ea); + ea = m.eulerAngles(0,1,0); + check_all_var(ea); + + // Check with random angles in range [0:pi]x[-pi:pi]x[-pi:pi]. + ea = (Array3::Random() + Array3(1,0,0))*Scalar(M_PI)*Array3(0.5,1,1); check_all_var(ea); ea[2] = ea[0] = internal::random(0,Scalar(M_PI)); From 276801b25a22ec2fed2bb13d217471e6098aea05 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 29 Nov 2013 19:54:01 +0100 Subject: [PATCH 0208/1103] Fixed and simplified Matlab code and added further block-related examples --- doc/AsciiQuickReference.txt | 40 ++++++++++++++++++++++++------------- 1 file changed, 26 insertions(+), 14 deletions(-) diff --git a/doc/AsciiQuickReference.txt b/doc/AsciiQuickReference.txt index fc613864f..d3bfd439c 100644 --- a/doc/AsciiQuickReference.txt +++ b/doc/AsciiQuickReference.txt @@ -16,8 +16,8 @@ double s; // Basic usage // Eigen // Matlab // comments x.size() // length(x) // vector size -C.rows() // size(C)(1) // number of rows -C.cols() // size(C)(2) // number of columns +C.rows() // size(C,1) // number of rows +C.cols() // size(C,2) // number of columns x(i) // x(i+1) // Matlab is 1-based C(i,j) // C(i+1,j+1) // @@ -51,20 +51,34 @@ v.setLinSpace(size,low,high) // v = linspace(low,high,size)' // Eigen // Matlab x.head(n) // x(1:n) x.head() // x(1:n) -x.tail(n) // N = rows(x); x(N - n: N) -x.tail() // N = rows(x); x(N - n: N) +x.tail(n) // x(end - n + 1: end) +x.tail() // x(end - n + 1: end) x.segment(i, n) // x(i+1 : i+n) x.segment(i) // x(i+1 : i+n) P.block(i, j, rows, cols) // P(i+1 : i+rows, j+1 : j+cols) P.block(i, j) // P(i+1 : i+rows, j+1 : j+cols) +P.row(i) // P(i+1, :) +P.col(j) // P(:, j+1) +P.leftCols() // P(:, 1:cols) +P.leftCols(cols) // P(:, 1:cols) +P.middleCols(j) // P(:, j+1:j+cols) +P.middleCols(j, cols) // P(:, j+1:j+cols) +P.rightCols() // P(:, end-cols+1:end) +P.rightCols(cols) // P(:, end-cols+1:end) +P.topRows() // P(1:rows, :) +P.topRows(rows) // P(1:rows, :) +P.middleRows(i) // P(:, i+1:i+rows) +P.middleRows(i, rows) // P(:, i+1:i+rows) +P.bottomRows() // P(:, end-rows+1:end) +P.bottomRows(rows) // P(:, end-rows+1:end) P.topLeftCorner(rows, cols) // P(1:rows, 1:cols) -P.topRightCorner(rows, cols) // [m n]=size(P); P(1:rows, n-cols+1:n) -P.bottomLeftCorner(rows, cols) // [m n]=size(P); P(m-rows+1:m, 1:cols) -P.bottomRightCorner(rows, cols) // [m n]=size(P); P(m-rows+1:m, n-cols+1:n) +P.topRightCorner(rows, cols) // P(1:rows, end-cols+1:end) +P.bottomLeftCorner(rows, cols) // P(end-rows+1:end, 1:cols) +P.bottomRightCorner(rows, cols) // P(end-rows+1:end, end-cols+1:end) P.topLeftCorner() // P(1:rows, 1:cols) -P.topRightCorner() // [m n]=size(P); P(1:rows, n-cols+1:n) -P.bottomLeftCorner() // [m n]=size(P); P(m-rows+1:m, 1:cols) -P.bottomRightCorner() // [m n]=size(P); P(m-rows+1:m, n-cols+1:n) +P.topRightCorner() // P(1:rows, end-cols+1:end) +P.bottomLeftCorner() // P(end-rows+1:end, 1:cols) +P.bottomRightCorner() // P(end-rows+1:end, end-cols+1:end) // Of particular note is Eigen's swap function which is highly optimized. // Eigen // Matlab @@ -126,10 +140,8 @@ int r, c; // Eigen // Matlab R.minCoeff() // min(R(:)) R.maxCoeff() // max(R(:)) -s = R.minCoeff(&r, &c) // [aa, bb] = min(R); [cc, dd] = min(aa); - // r = bb(dd); c = dd; s = cc -s = R.maxCoeff(&r, &c) // [aa, bb] = max(R); [cc, dd] = max(aa); - // row = bb(dd); col = dd; s = cc +s = R.minCoeff(&r, &c) // [s, i] = min(R(:)); [r, c] = ind2sub(size(R), i); +s = R.maxCoeff(&r, &c) // [s, i] = max(R(:)); [r, c] = ind2sub(size(R), i); R.sum() // sum(R(:)) R.colwise().sum() // sum(R) R.rowwise().sum() // sum(R, 2) or sum(R')' From d0261bd26c3900a5e52da3574fc2aeab3392c30b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 30 Nov 2013 10:42:23 +0100 Subject: [PATCH 0209/1103] Fix swap in DenseBase --- Eigen/src/Core/DenseBase.h | 24 +++++++++++++++++++++++- 1 file changed, 23 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 4794c2f13..50a63c85c 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -387,7 +387,29 @@ template class DenseBase // size types on MSVC. return typename internal::eval::type(derived()); } + +#ifdef EIGEN_TEST_EVALUATORS + /** swaps *this with the expression \a other. + * + */ + template + EIGEN_DEVICE_FUNC + void swap(const DenseBase& other, + int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase) + { + swap_using_evaluator(derived(), other.derived()); + } + /** swaps *this with the matrix or array \a other. + * + */ + template + EIGEN_DEVICE_FUNC + void swap(PlainObjectBase& other) + { + swap_using_evaluator(derived(), other.derived()); + } +#else // EIGEN_TEST_EVALUATORS /** swaps *this with the expression \a other. * */ @@ -408,7 +430,7 @@ template class DenseBase { SwapWrapper(derived()).lazyAssign(other.derived()); } - +#endif // EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC inline const NestByValue nestByValue() const; EIGEN_DEVICE_FUNC inline const ForceAlignedAccess forceAlignedAccess() const; From 27ca9437a1e191f20724492275bf0b1415eee2d6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 14:05:34 +0100 Subject: [PATCH 0210/1103] Fix usage of Dense versus DenseShape --- Eigen/src/Core/AssignEvaluator.h | 31 ++++++++++++++---------------- Eigen/src/Core/CoreEvaluators.h | 4 ++-- Eigen/src/Core/ProductEvaluators.h | 4 ++-- Eigen/src/Core/util/Constants.h | 12 ++++++++++++ 4 files changed, 30 insertions(+), 21 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 2f18cbc95..2c9f2426b 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -526,6 +526,8 @@ public: Index size() const { return m_dstExpr.size(); } Index innerSize() const { return m_dstExpr.innerSize(); } Index outerSize() const { return m_dstExpr.outerSize(); } + Index rows() const { return m_dstExpr.rows(); } + Index cols() const { return m_dstExpr.cols(); } Index outerStride() const { return m_dstExpr.outerStride(); } // TODO get rid of this one: @@ -534,16 +536,25 @@ public: DstEvaluatorType& dstEvaluator() { return m_dst; } const SrcEvaluatorType& srcEvaluator() const { return m_src; } + /// Assign src(row,col) to dst(row,col) through the assignment functor. void assignCoeff(Index row, Index col) { m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col)); } + /// This overload by-passes the source expression, i.e., dst(row,col) ?= value + void assignCoeff(Index row, Index col, const Scalar &value) + { + m_functor.assignCoeff(m_dst.coeffRef(row,col), value); + } + + /// \sa assignCoeff(Index,Index) void assignCoeff(Index index) { m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index)); } + /// \sa assignCoeff(Index,Index) void assignCoeffByOuterInner(Index outer, Index inner) { Index row = rowIndexByOuterInner(outer, inner); @@ -633,29 +644,15 @@ void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src) * Part 6 : Generic assignment ***************************************************************************/ - -// An evaluator must define its shape. It can be one of the following: -struct DenseShape {}; -struct DiagonalShape {}; -struct BandShape {}; -struct TriangularShape {}; -struct SelfAdjointShape {}; -struct SparseShape {}; - - // Based on the respective shapes of the destination and source, // the class AssignmentKind determine the kind of assignment mechanism. // AssignmentKind must define a Kind typedef. template struct AssignmentKind; -// AssignmentKind<.,.>::Kind can be one of the following: - struct Dense2Dense {}; - struct Triangular2Triangular {}; -// struct Diagonal2Diagonal {}; // <=> Dense2Dense - struct Sparse2Dense {}; - struct Sparse2Sparse {}; +// Assignement kind defined in this file: +struct Dense2Dense {}; -template<> struct AssignmentKind { typedef Dense2Dense Kind; }; +template<> struct AssignmentKind { typedef Dense2Dense Kind; }; // This is the main assignment class template< typename DstXprType, typename SrcXprType, typename Functor, diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 9460e675f..961b56e55 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -42,14 +42,14 @@ template struct storage_kind_to_shape; template<> struct storage_kind_to_shape { - typedef Dense Shape; + typedef DenseShape Shape; }; // TODO to be moved to SparseCore: /* template<> struct storage_kind_to_shape { - typedef Sparse Shape; + typedef SparseSpape Shape; }; */ diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index e3a893651..51228be5f 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -48,7 +48,7 @@ struct dense_product_impl; // The evaluator for default dense products creates a temporary and call dense_product_impl template -struct product_evaluator, ProductTag, Dense, Dense, typename Lhs::Scalar, typename Rhs::Scalar> +struct product_evaluator, ProductTag, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> : public evaluator::PlainObject>::type { typedef Product XprType; @@ -217,7 +217,7 @@ template -struct product_evaluator, ProductTag, Dense, Dense, typename Lhs::Scalar, typename Rhs::Scalar > +struct product_evaluator, ProductTag, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar > : evaluator_base > { typedef Product XprType; diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 3178ff06e..14449eb6c 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -440,6 +440,18 @@ struct MatrixXpr {}; /** The type used to identify an array expression */ struct ArrayXpr {}; + +#ifdef EIGEN_ENABLE_EVALUATORS +// An evaluator must define its shape. By default, it can be one of the following: +struct DenseShape { static std::string debugName() { return "DenseShape"; } }; +struct DiagonalShape { static std::string debugName() { return "DiagonalShape"; } }; +struct BandShape { static std::string debugName() { return "BandShape"; } }; +struct TriangularShape { static std::string debugName() { return "TriangularShape"; } }; +struct SelfAdjointShape { static std::string debugName() { return "SelfAdjointShape"; } }; +struct SparseShape { static std::string debugName() { return "SparseShape"; } }; +#endif + + } // end namespace Eigen #endif // EIGEN_CONSTANTS_H From 626821b0e34a624e8fa8980339b771e155722ace Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 14:06:17 +0100 Subject: [PATCH 0211/1103] Add evaluator/assignment to TriangularView expressions --- Eigen/src/Core/TriangularMatrix.h | 317 +++++++++++++++++++++++++++++- test/evaluators.cpp | 35 +++- 2 files changed, 350 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 1d6e34650..96ed9cef9 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -36,7 +36,13 @@ template class TriangularBase : public EigenBase RowsAtCompileTime = internal::traits::RowsAtCompileTime, ColsAtCompileTime = internal::traits::ColsAtCompileTime, MaxRowsAtCompileTime = internal::traits::MaxRowsAtCompileTime, - MaxColsAtCompileTime = internal::traits::MaxColsAtCompileTime + MaxColsAtCompileTime = internal::traits::MaxColsAtCompileTime, + + SizeAtCompileTime = (internal::size_at_compile_time::RowsAtCompileTime, + internal::traits::ColsAtCompileTime>::ret) + /**< This is equal to the number of coefficients, i.e. the number of + * rows times the number of columns, or to \a Dynamic if this is not + * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */ }; typedef typename internal::traits::Scalar Scalar; typedef typename internal::traits::StorageKind StorageKind; @@ -408,15 +414,24 @@ template class TriangularView EIGEN_DEVICE_FUNC void swap(TriangularBase const & other) { +// #ifdef EIGEN_TEST_EVALUATORS +// swap_using_evaluator(this->derived(), other.derived()); +// #else TriangularView,Mode>(const_cast(m_matrix)).lazyAssign(other.derived()); +// #endif } + // TODO: this overload is ambiguous and it should be deprecated (Gael) template EIGEN_DEVICE_FUNC void swap(MatrixBase const & other) { +// #ifdef EIGEN_TEST_EVALUATORS +// swap_using_evaluator(this->derived(), other.derived()); +// #else SwapWrapper swaper(const_cast(m_matrix)); TriangularView,Mode>(swaper).lazyAssign(other.derived()); +// #endif } EIGEN_DEVICE_FUNC @@ -895,6 +910,306 @@ bool MatrixBase::isLowerTriangular(const RealScalar& prec) const return true; } + +#ifdef EIGEN_ENABLE_EVALUATORS + +/*************************************************************************** +**************************************************************************** +* Evaluators and Assignment of triangular expressions +*************************************************************************** +***************************************************************************/ + +namespace internal { + + +// TODO currently a triangular expression has the form TriangularView<.,.> +// in the future triangular-ness should be defined by the expression traits +// such that Transpose > is valid. (currently TriangularBase::transpose() is overloaded to make it work) +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef TriangularShape Shape; + + // 1 if assignment A = B assumes aliasing when B is of type T and thus B needs to be evaluated into a + // temporary; 0 if not. + static const int AssumeAliasing = 0; +}; + +template +struct evaluator, Kind, typename MatrixType::Scalar> + : evaluator +{ + typedef TriangularView XprType; + typedef evaluator Base; + typedef evaluator type; + evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {} +}; + +// Additional assignement kinds: +struct Triangular2Triangular {}; +struct Triangular2Dense {}; +struct Dense2Triangular {}; + + +template struct triangular_assignment_loop; + + +template +void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) +{ +#ifdef EIGEN_DEBUG_ASSIGN + // TODO these traits should be computed from information provided by the evaluators + internal::copy_using_evaluator_traits::debug(); +#endif + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + typedef typename evaluator::type DstEvaluatorType; + typedef typename evaluator::type SrcEvaluatorType; + + DstEvaluatorType dstEvaluator(dst); + SrcEvaluatorType srcEvaluator(src); + + typedef generic_dense_assignment_kernel Kernel; + Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived()); + + enum { + unroll = DstXprType::SizeAtCompileTime != Dynamic + && internal::traits::CoeffReadCost != Dynamic + && DstXprType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT + }; + + triangular_assignment_loop::run(kernel); +} + +template +void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src) +{ + call_triangular_assignment_loop(dst, src, internal::assign_op()); +} + + +template<> struct AssignmentKind { typedef Triangular2Triangular Kind; }; +template<> struct AssignmentKind { typedef Triangular2Dense Kind; }; +template<> struct AssignmentKind { typedef Dense2Triangular Kind; }; + + +template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> +struct Assignment +{ + static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { + eigen_assert(int(DstXprType::Mode) == int(SrcXprType::Mode)); + + call_triangular_assignment_loop(dst, src, func); + } +}; + +template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> +struct Assignment +{ + static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { + call_triangular_assignment_loop(dst, src, func); + } +}; + +template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> +struct Assignment +{ + static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { + call_triangular_assignment_loop(dst, src, func); + } +}; + + +template +struct triangular_assignment_loop +{ + enum { + col = (UnrollCount-1) / Kernel::RowsAtCompileTime, + row = (UnrollCount-1) % Kernel::RowsAtCompileTime + }; + + typedef typename Kernel::Scalar Scalar; + + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + triangular_assignment_loop::run(kernel); + + // TODO should be a static assert + eigen_assert( Mode == Upper || Mode == Lower + || Mode == StrictlyUpper || Mode == StrictlyLower + || Mode == UnitUpper || Mode == UnitLower); + + if((Mode == Upper && row <= col) + || (Mode == Lower && row >= col) + || (Mode == StrictlyUpper && row < col) + || (Mode == StrictlyLower && row > col) + || (Mode == UnitUpper && row < col) + || (Mode == UnitLower && row > col)) + kernel.assignCoeff(row, col); + else if(ClearOpposite) + { + if((Mode&UnitDiag) && row==col) kernel.assignCoeff(row, col, Scalar(1)); + else kernel.assignCoeff(row, col, Scalar(0)); + } + } +}; + +// prevent buggy user code from causing an infinite recursion +template +struct triangular_assignment_loop +{ + EIGEN_DEVICE_FUNC + static inline void run(Kernel &) {} +}; + +// TODO: merge the following 6 variants into a single one (or max two), +// and perhaps write them using sub-expressions + +// TODO: expreiment with a recursive assignment procedure splitting the current +// triangular part into one rectangular and two triangular parts. + +template +struct triangular_assignment_loop +{ + typedef typename Kernel::Index Index; + typedef typename Kernel::Scalar Scalar; + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + for(Index j = 0; j < kernel.cols(); ++j) + { + Index maxi = (std::min)(j, kernel.rows()-1); + for(Index i = 0; i <= maxi; ++i) + kernel.assignCoeff(i, j); + if (ClearOpposite) + for(Index i = maxi+1; i < kernel.rows(); ++i) + kernel.assignCoeff(i, j, Scalar(0)); + } + } +}; + +template +struct triangular_assignment_loop +{ + typedef typename Kernel::Index Index; + typedef typename Kernel::Scalar Scalar; + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + for(Index j = 0; j < kernel.cols(); ++j) + { + for(Index i = j; i < kernel.rows(); ++i) + kernel.assignCoeff(i, j); + Index maxi = (std::min)(j, kernel.rows()); + if (ClearOpposite) + for(Index i = 0; i < maxi; ++i) + kernel.assignCoeff(i, j, Scalar(0)); + } + } +}; + +template +struct triangular_assignment_loop +{ + typedef typename Kernel::Index Index; + typedef typename Kernel::Scalar Scalar; + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + for(Index j = 0; j < kernel.cols(); ++j) + { + Index maxi = (std::min)(j, kernel.rows()); + for(Index i = 0; i < maxi; ++i) + kernel.assignCoeff(i, j); + if (ClearOpposite) + for(Index i = maxi; i < kernel.rows(); ++i) + kernel.assignCoeff(i, j) = Scalar(0); + } + } +}; + +template +struct triangular_assignment_loop +{ + typedef typename Kernel::Index Index; + typedef typename Kernel::Scalar Scalar; + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + for(Index j = 0; j < kernel.cols(); ++j) + { + for(Index i = j+1; i < kernel.rows(); ++i) + kernel.assignCoeff(i, j); + Index maxi = (std::min)(j, kernel.rows()-1); + if (ClearOpposite) + for(Index i = 0; i <= maxi; ++i) + kernel.assignCoeff(i, j, Scalar(0)); + } + } +}; + +template +struct triangular_assignment_loop +{ + typedef typename Kernel::Index Index; + typedef typename Kernel::Scalar Scalar; + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + for(Index j = 0; j < kernel.cols(); ++j) + { + Index maxi = (std::min)(j, kernel.rows()); + Index i = 0; + for(; i < maxi; ++i) + kernel.assignCoeff(i, j); + + if(i +struct triangular_assignment_loop +{ + typedef typename Kernel::Index Index; + typedef typename Kernel::Scalar Scalar; + EIGEN_DEVICE_FUNC + static inline void run(Kernel &kernel) + { + for(Index j = 0; j < kernel.cols(); ++j) + { + Index maxi = (std::min)(j, kernel.rows()); + Index i = 0; + if (ClearOpposite) + { + for(; i < maxi; ++i) + kernel.assignCoeff(i, j, Scalar(0)); + } + + if(i(), MatrixXd(A.triangularView())); + + A.setRandom();B.setRandom(); + VERIFY_IS_APPROX_EVALUATOR2(B, A.triangularView(), MatrixXd(A.triangularView())); + + A.setRandom();B.setRandom(); + VERIFY_IS_APPROX_EVALUATOR2(B, A.triangularView(), MatrixXd(A.triangularView())); + + A.setRandom();B.setRandom(); + C = B; C.triangularView() = A; + copy_using_evaluator(B.triangularView(), A); + VERIFY(B.isApprox(C) && "copy_using_evaluator(B.triangularView(), A)"); + + A.setRandom();B.setRandom(); + C = B; C.triangularView() = A.triangularView(); + copy_using_evaluator(B.triangularView(), A.triangularView()); + VERIFY(B.isApprox(C) && "copy_using_evaluator(B.triangularView(), A.triangularView())"); + + + A.setRandom();B.setRandom(); + C = B; C.triangularView() = A.triangularView().transpose(); + copy_using_evaluator(B.triangularView(), A.triangularView().transpose()); + VERIFY(B.isApprox(C) && "copy_using_evaluator(B.triangularView(), A.triangularView().transpose())"); + } } From c6f73370327e2e91a45ac13e7bbb6c7567179e49 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 14:44:13 +0100 Subject: [PATCH 0212/1103] Get rid of call_dense_swap_loop --- Eigen/src/Core/AssignEvaluator.h | 2 +- Eigen/src/Core/Swap.h | 31 +++++++------------------------ test/evaluators.cpp | 9 +++++---- 3 files changed, 13 insertions(+), 29 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 2c9f2426b..3babc16ba 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -504,7 +504,7 @@ struct dense_assignment_loop // This abstraction level permits to keep the evaluation loops as simple and as generic as possible. // One can customize the assignment using this generic dense_assignment_kernel with different // functors, or by completely overloading it, by-passing a functor. -template +template class generic_dense_assignment_kernel { protected: diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index 8fd94b3c7..117f667f6 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -139,15 +139,16 @@ template class SwapWrapper // #endif -#ifdef EIGEN_TEST_EVALUATORS +#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { // Overload default assignPacket behavior for swapping them template -class dense_swap_kernel : public generic_dense_assignment_kernel > +class generic_dense_assignment_kernel, Specialized> + : public generic_dense_assignment_kernel, BuiltIn> { - typedef generic_dense_assignment_kernel > Base; + typedef generic_dense_assignment_kernel, BuiltIn> Base; typedef typename DstEvaluatorTypeT::PacketScalar PacketScalar; using Base::m_dst; using Base::m_src; @@ -157,9 +158,10 @@ public: typedef typename Base::Scalar Scalar; typedef typename Base::Index Index; typedef typename Base::DstXprType DstXprType; + typedef swap_assign_op Functor; - dense_swap_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, DstXprType& dstExpr) - : Base(dst, src, swap_assign_op(), dstExpr) + generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr) + : Base(dst, src, func, dstExpr) {} template @@ -183,25 +185,6 @@ public: assignPacket(row, col); } }; - -template -void call_dense_swap_loop(const DstXprType& dst, const SrcXprType& src) -{ - // TODO there is too much redundancy with call_dense_assignment_loop - - eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); - - typedef typename evaluator::type DstEvaluatorType; - typedef typename evaluator::type SrcEvaluatorType; - - DstEvaluatorType dstEvaluator(dst); - SrcEvaluatorType srcEvaluator(src); - - typedef dense_swap_kernel Kernel; - Kernel kernel(dstEvaluator, srcEvaluator, dst.const_cast_derived()); - - dense_assignment_loop::run(kernel); -} } // namespace internal diff --git a/test/evaluators.cpp b/test/evaluators.cpp index 29192d7f9..63f940318 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -74,7 +74,8 @@ namespace Eigen { template void swap_using_evaluator(const DstXprType& dst, const SrcXprType& src) { - call_dense_swap_loop(dst.const_cast_derived(), src.const_cast_derived()); + typedef typename DstXprType::Scalar Scalar; + call_assignment(dst.const_cast_derived(), src.const_cast_derived(), internal::swap_assign_op()); } } @@ -193,7 +194,7 @@ void test_evaluators() VERIFY_IS_APPROX_EVALUATOR2(resXX, prod(mX4,m4X), mX4*m4X); VERIFY_IS_APPROX_EVALUATOR2(resXX, prod(mXX,mXX), mXX*mXX); } - +#endif { ArrayXXf a(2,3); ArrayXXf b(3,2); @@ -202,7 +203,7 @@ void test_evaluators() // this does not work because Random is eval-before-nested: // copy_using_evaluator(w, Vector2d::Random().transpose()); - + // test CwiseUnaryOp VERIFY_IS_APPROX_EVALUATOR(v2, 3 * v); VERIFY_IS_APPROX_EVALUATOR(w, (3 * v).transpose()); @@ -405,7 +406,7 @@ void test_evaluators() arr_ref.row(1) /= (arr_ref.row(2) + 1); VERIFY_IS_APPROX(arr, arr_ref); } -#endif + { // test triangular shapes MatrixXd A = MatrixXd::Random(6,6), B(6,6), C(6,6); From 8af1ba534669a223ca69136046690eb6d49ff619 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 15:07:45 +0100 Subject: [PATCH 0213/1103] Make swap unit test work with evaluators --- Eigen/Core | 2 +- Eigen/src/Core/DenseBase.h | 4 ++-- Eigen/src/Core/TriangularMatrix.h | 23 +++++++++++------------ test/evaluators.cpp | 12 +++++++++--- 4 files changed, 23 insertions(+), 18 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index e2c9c69cd..cdc2f2d46 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -305,6 +305,7 @@ using std::ptrdiff_t; #include "src/Core/functors/UnaryFunctors.h" #include "src/Core/functors/NullaryFunctors.h" #include "src/Core/functors/StlFunctors.h" +#include "src/Core/functors/AssignmentFunctors.h" #include "src/Core/DenseCoeffsBase.h" #include "src/Core/DenseBase.h" @@ -312,7 +313,6 @@ using std::ptrdiff_t; #include "src/Core/EigenBase.h" #ifdef EIGEN_ENABLE_EVALUATORS -#include "src/Core/functors/AssignmentFunctors.h" #include "src/Core/Product.h" #include "src/Core/CoreEvaluators.h" #include "src/Core/AssignEvaluator.h" diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 50a63c85c..9bbfacbcf 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -397,7 +397,7 @@ template class DenseBase void swap(const DenseBase& other, int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase) { - swap_using_evaluator(derived(), other.derived()); + call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op()); } /** swaps *this with the matrix or array \a other. @@ -407,7 +407,7 @@ template class DenseBase EIGEN_DEVICE_FUNC void swap(PlainObjectBase& other) { - swap_using_evaluator(derived(), other.derived()); + call_assignment(derived(), other.derived(), internal::swap_assign_op()); } #else // EIGEN_TEST_EVALUATORS /** swaps *this with the expression \a other. diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 96ed9cef9..d1fd21ad7 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -414,11 +414,11 @@ template class TriangularView EIGEN_DEVICE_FUNC void swap(TriangularBase const & other) { -// #ifdef EIGEN_TEST_EVALUATORS -// swap_using_evaluator(this->derived(), other.derived()); -// #else - TriangularView,Mode>(const_cast(m_matrix)).lazyAssign(other.derived()); -// #endif + #ifdef EIGEN_TEST_EVALUATORS + call_assignment(*this, other.const_cast_derived(), internal::swap_assign_op()); + #else + TriangularView,Mode>(const_cast(m_matrix)).lazyAssign(other.const_cast_derived().nestedExpression()); + #endif } // TODO: this overload is ambiguous and it should be deprecated (Gael) @@ -426,12 +426,12 @@ template class TriangularView EIGEN_DEVICE_FUNC void swap(MatrixBase const & other) { -// #ifdef EIGEN_TEST_EVALUATORS -// swap_using_evaluator(this->derived(), other.derived()); -// #else + #ifdef EIGEN_TEST_EVALUATORS + call_assignment(*this, other.const_cast_derived(), internal::swap_assign_op()); + #else SwapWrapper swaper(const_cast(m_matrix)); TriangularView,Mode>(swaper).lazyAssign(other.derived()); -// #endif + #endif } EIGEN_DEVICE_FUNC @@ -988,7 +988,6 @@ void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& sr call_triangular_assignment_loop(dst, src, internal::assign_op()); } - template<> struct AssignmentKind { typedef Triangular2Triangular Kind; }; template<> struct AssignmentKind { typedef Triangular2Dense Kind; }; template<> struct AssignmentKind { typedef Dense2Triangular Kind; }; @@ -1028,8 +1027,8 @@ template(), MatrixXd(A.triangularView())); @@ -434,5 +434,11 @@ void test_evaluators() C = B; C.triangularView() = A.triangularView().transpose(); copy_using_evaluator(B.triangularView(), A.triangularView().transpose()); VERIFY(B.isApprox(C) && "copy_using_evaluator(B.triangularView(), A.triangularView().transpose())"); + + + A.setRandom();B.setRandom(); C = B; D = A; + C.triangularView().swap(D.triangularView()); + swap_using_evaluator(B.triangularView(), A.triangularView()); + VERIFY(B.isApprox(C) && "swap_using_evaluator(B.triangularView(), A.triangularView())"); } } From 7f917807c6a72ae219880fb72e2d644d617427bc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 16:19:14 +0100 Subject: [PATCH 0214/1103] Fix product evaluator when TEST_EVALUATOR in not ON --- Eigen/src/Core/ProductEvaluators.h | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 51228be5f..26145ceff 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -58,7 +58,9 @@ struct product_evaluator, ProductTag, DenseSha product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { + ::new (static_cast(this)) Base(m_result); + dense_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); } @@ -186,7 +188,11 @@ struct dense_product_impl template static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) - { dst = lazyprod(lhs,rhs); } + { + // TODO: use the following instead of calling call_assignment + // dst = lazyprod(lhs,rhs); + call_assignment(dst, lazyprod(lhs,rhs), internal::assign_op()); + } template static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) From 34ca81b1bfe77b11effdd463099b30d12b6466c9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 16:37:58 +0100 Subject: [PATCH 0215/1103] Add direct assignment of products --- Eigen/src/Core/AssignEvaluator.h | 5 ++++ Eigen/src/Core/ProductEvaluators.h | 40 ++++++++++++++++++++++++++++-- test/evaluators.cpp | 15 +++++++++-- 3 files changed, 56 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 3babc16ba..6d3739407 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -679,6 +679,11 @@ void call_assignment(const NoAlias& dst, const Src& src, const { Assignment::run(dst.expression(), src, func); } +template class StorageBase, typename Src, typename Func> +void call_assignment(NoAlias& dst, const Src& src, const Func& func) +{ + Assignment::run(dst.expression(), src, func); +} // Generic Dense to Dense assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 26145ceff..d0dba1644 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -45,6 +45,12 @@ struct evaluator > template::value> struct dense_product_impl; +template +struct evaluator_traits > + : evaluator_traits_base > +{ + enum { AssumeAliasing = 1 }; +}; // The evaluator for default dense products creates a temporary and call dense_product_impl template @@ -58,9 +64,7 @@ struct product_evaluator, ProductTag, DenseSha product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { - ::new (static_cast(this)) Base(m_result); - dense_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); } @@ -68,6 +72,38 @@ protected: PlainObject m_result; }; +// Dense = Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment, internal::assign_op, Dense2Dense, Scalar> +{ + typedef Product SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dense_product_impl::evalTo(dst, src.lhs(), src.rhs()); + } +}; + +// Dense += Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment, internal::add_assign_op, Dense2Dense, Scalar> +{ + typedef Product SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op &) + { + dense_product_impl::addTo(dst, src.lhs(), src.rhs()); + } +}; + +// Dense -= Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment, internal::sub_assign_op, Dense2Dense, Scalar> +{ + typedef Product SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op &) + { + dense_product_impl::subTo(dst, src.lhs(), src.rhs()); + } +}; template struct dense_product_impl diff --git a/test/evaluators.cpp b/test/evaluators.cpp index 7a20014dd..c07146260 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -47,14 +47,14 @@ namespace Eigen { void add_assign_using_evaluator(const DstXprType& dst, const SrcXprType& src) { typedef typename DstXprType::Scalar Scalar; - call_assignment(dst.const_cast_derived(), src.derived(), internal::add_assign_op()); + call_assignment(const_cast(dst), src.derived(), internal::add_assign_op()); } template void subtract_assign_using_evaluator(const DstXprType& dst, const SrcXprType& src) { typedef typename DstXprType::Scalar Scalar; - call_assignment(dst.const_cast_derived(), src.derived(), internal::sub_assign_op()); + call_assignment(const_cast(dst), src.derived(), internal::sub_assign_op()); } template @@ -151,6 +151,17 @@ void test_evaluators() c = a*a; copy_using_evaluator(a, prod(a,a)); VERIFY_IS_APPROX(a,c); + + // check compound assignment of products + d = c; + add_assign_using_evaluator(c.noalias(), prod(a,b)); + d.noalias() += a*b; + VERIFY_IS_APPROX(c, d); + + d = c; + subtract_assign_using_evaluator(c.noalias(), prod(a,b)); + d.noalias() -= a*b; + VERIFY_IS_APPROX(c, d); } { From b5fd774775d31df3af532ab0798d8be3dc8448d9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 17:53:26 +0100 Subject: [PATCH 0216/1103] Fix flags of Product<> --- Eigen/src/Core/Product.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 970d257a5..3b8fd7d9a 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -33,12 +33,12 @@ template class Pro namespace internal { template struct traits > - : traits::Type> + : traits > { // We want A+B*C to be of type Product and not Product // TODO: This flag should eventually go in a separate evaluator traits class enum { - Flags = traits::Type>::Flags & ~(EvalBeforeNestingBit | DirectAccessBit) + Flags = traits >::Flags & ~(EvalBeforeNestingBit | DirectAccessBit) }; }; } // end namespace internal From 6f1a0479b3a8176f4d8db6513cbb39c027c34b7f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 17:54:15 +0100 Subject: [PATCH 0217/1103] fix a typo triangular assignment --- Eigen/src/Core/TriangularMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index d1fd21ad7..602f0c5d1 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -1127,7 +1127,7 @@ struct triangular_assignment_loop kernel.assignCoeff(i, j); if (ClearOpposite) for(Index i = maxi; i < kernel.rows(); ++i) - kernel.assignCoeff(i, j) = Scalar(0); + kernel.assignCoeff(i, j, Scalar(0)); } } }; From f0b82c3ab972a1eafd6aa4f08f4eaffa0d6f1e55 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 2 Dec 2013 17:54:38 +0100 Subject: [PATCH 0218/1103] Make reductions compatible with evaluators --- Eigen/src/Core/Redux.h | 81 ++++++++++++++++++++++++++++++++++++++---- 1 file changed, 74 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index b2c775d90..1b6a4177a 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -174,7 +174,7 @@ struct redux_impl typedef typename Derived::Scalar Scalar; typedef typename Derived::Index Index; EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func) + static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) { eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); Scalar res; @@ -200,10 +200,10 @@ struct redux_impl typedef typename packet_traits::type PacketScalar; typedef typename Derived::Index Index; - static Scalar run(const Derived& mat, const Func& func) + static Scalar run(const Derived &mat, const Func& func) { const Index size = mat.size(); - eigen_assert(size && "you are using an empty matrix"); + const Index packetSize = packet_traits::size; const Index alignedStart = internal::first_aligned(mat); enum { @@ -258,7 +258,7 @@ struct redux_impl typedef typename packet_traits::type PacketScalar; typedef typename Derived::Index Index; - static Scalar run(const Derived& mat, const Func& func) + static Scalar run(const Derived &mat, const Func& func) { eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); const Index innerSize = mat.innerSize(); @@ -300,9 +300,8 @@ struct redux_impl Size = Derived::SizeAtCompileTime, VectorizedSize = (Size / PacketSize) * PacketSize }; - static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func) + static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) { - eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); Scalar res = func.predux(redux_vec_unroller::run(mat,func)); if (VectorizedSize != Size) res = func(res,redux_novec_unroller::run(mat,func)); @@ -310,6 +309,64 @@ struct redux_impl } }; +#ifdef EIGEN_ENABLE_EVALUATORS +// evaluator adaptor +template +class redux_evaluator +{ +public: + redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} + + typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketScalar PacketScalar; + typedef typename XprType::PacketReturnType PacketReturnType; + + enum { + MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = XprType::MaxColsAtCompileTime, + // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator + Flags = XprType::Flags & ~DirectAccessBit, + IsRowMajor = XprType::IsRowMajor, + SizeAtCompileTime = XprType::SizeAtCompileTime, + InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime, + CoeffReadCost = XprType::CoeffReadCost + }; + + Index rows() const { return m_xpr.rows(); } + Index cols() const { return m_xpr.cols(); } + Index size() const { return m_xpr.size(); } + Index innerSize() const { return m_xpr.innerSize(); } + Index outerSize() const { return m_xpr.outerSize(); } + + CoeffReturnType coeff(Index row, Index col) const + { return m_evaluator.coeff(row, col); } + + CoeffReturnType coeff(Index index) const + { return m_evaluator.coeff(index); } + + template + PacketReturnType packet(Index row, Index col) const + { return m_evaluator.template packet(row, col); } + + template + PacketReturnType packet(Index index) const + { return m_evaluator.template packet(index); } + + CoeffReturnType coeffByOuterInner(Index outer, Index inner) const + { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); } + + template + PacketReturnType packetByOuterInner(Index outer, Index inner) const + { return m_evaluator.template packet(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); } + +protected: + typename internal::evaluator::nestedType m_evaluator; + const XprType &m_xpr; +}; +#endif + } // end namespace internal /*************************************************************************** @@ -320,7 +377,7 @@ struct redux_impl /** \returns the result of a full redux operation on the whole matrix or vector using \a func * * The template parameter \a BinaryOp is the type of the functor \a func which must be - * an associative operator. Both current STL and TR1 functor styles are handled. + * an associative operator. Both current C++98 and C++11 functor styles are handled. * * \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise() */ @@ -329,9 +386,19 @@ template EIGEN_STRONG_INLINE typename internal::result_of::Scalar)>::type DenseBase::redux(const Func& func) const { + eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); +#ifdef EIGEN_TEST_EVALUATORS + + typedef typename internal::redux_evaluator ThisEvaluator; + ThisEvaluator thisEval(derived()); + return internal::redux_impl::run(thisEval, func); + +#else typedef typename internal::remove_all::type ThisNested; + return internal::redux_impl ::run(derived(), func); +#endif } /** \returns the minimum of all coefficients of \c *this. From 6c5e915e9a6c79550e7e2db2b53648f163a1411d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 3 Dec 2013 17:17:53 +0100 Subject: [PATCH 0219/1103] Enable use of evaluators for noalias and lazyProduct, add conversion to scalar for inner products --- Eigen/src/Core/CoreEvaluators.h | 4 --- Eigen/src/Core/GeneralProduct.h | 33 +++++++++++++++++++-- Eigen/src/Core/MatrixBase.h | 4 +++ Eigen/src/Core/NoAlias.h | 29 ++++++++++++++++++ Eigen/src/Core/Product.h | 6 ++++ Eigen/src/Core/ProductEvaluators.h | 32 ++++++++++++-------- Eigen/src/Core/products/CoeffBasedProduct.h | 6 +++- Eigen/src/Core/util/ForwardDeclarations.h | 6 ++++ Eigen/src/Core/util/StaticAssert.h | 3 +- test/main.h | 2 ++ 10 files changed, 103 insertions(+), 22 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 961b56e55..faf5aedb5 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -53,11 +53,7 @@ struct storage_kind_to_shape { }; */ -template struct evaluator_traits; -template< typename T, - typename Kind = typename evaluator_traits::Kind, - typename Scalar = typename T::Scalar> struct evaluator; template< typename T, typename LhsKind = typename evaluator_traits::Kind, diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index c9ab63782..675f6ee8d 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -623,7 +623,7 @@ MatrixBase::operator*(const MatrixBase &other) const return Product(derived(), other.derived()); } -#else +#else // EIGEN_TEST_EVALUATORS template template inline const typename ProductReturnType::Type @@ -653,9 +653,10 @@ MatrixBase::operator*(const MatrixBase &other) const #endif return typename ProductReturnType::Type(derived(), other.derived()); } -#endif +#endif // EIGEN_TEST_EVALUATORS + +#endif // __CUDACC__ -#endif /** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation. * * The returned product will behave like any other expressions: the coefficients of the product will be @@ -667,6 +668,31 @@ MatrixBase::operator*(const MatrixBase &other) const * * \sa operator*(const MatrixBase&) */ +#ifdef EIGEN_TEST_EVALUATORS +template +template +const Product +MatrixBase::lazyProduct(const MatrixBase &other) const +{ + enum { + ProductIsValid = Derived::ColsAtCompileTime==Dynamic + || OtherDerived::RowsAtCompileTime==Dynamic + || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), + AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, + SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) + }; + // note to the lost user: + // * for a dot product use: v1.dot(v2) + // * for a coeff-wise product use: v1.cwiseProduct(v2) + EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), + INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) + EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), + INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) + EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) + + return Product(derived(), other.derived()); +} +#else // EIGEN_TEST_EVALUATORS template template const typename LazyProductReturnType::Type @@ -690,6 +716,7 @@ MatrixBase::lazyProduct(const MatrixBase &other) const return typename LazyProductReturnType::Type(derived(), other.derived()); } +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 61798317e..37e7408a4 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -199,7 +199,11 @@ template class MatrixBase template EIGEN_DEVICE_FUNC +#ifdef EIGEN_TEST_EVALUATORS + const Product +#else const typename LazyProductReturnType::Type +#endif lazyProduct(const MatrixBase &other) const; template diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index 65117a806..6ac525336 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -34,6 +34,33 @@ class NoAlias typedef typename ExpressionType::Scalar Scalar; NoAlias(ExpressionType& expression) : m_expression(expression) {} + +#ifdef EIGEN_TEST_EVALUATORS + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase& other) + { + call_assignment(*this, other.derived(), internal::assign_op()); + return m_expression; + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase& other) + { + call_assignment(*this, other.derived(), internal::add_assign_op()); + return m_expression; + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase& other) + { + call_assignment(*this, other.derived(), internal::sub_assign_op()); + return m_expression; + } + +#else /** Behaves like MatrixBase::lazyAssign(other) * \sa MatrixBase::lazyAssign() */ @@ -91,6 +118,8 @@ class NoAlias template ExpressionType& operator=(const ReturnByValue& func) { return m_expression = func; } +#endif + #endif EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 3b8fd7d9a..b37fd2ff7 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -79,6 +79,12 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, const LhsNestedCleaned& lhs() const { return m_lhs; } const RhsNestedCleaned& rhs() const { return m_rhs; } + + /** Convertion to scalar for inner-products */ + operator const Scalar() const { + EIGEN_STATIC_ASSERT(SizeAtCompileTime==1, IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY); + return typename internal::evaluator::type(*this).coeff(0,0); + } protected: diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index d0dba1644..5dd480cad 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -209,12 +209,12 @@ struct dense_product_impl : dense_product_impl_base::Scalar Scalar; - template - static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) - { - // TODO bypass GeneralProduct class - GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); - } +// template +// static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) +// { +// // TODO bypass GeneralProduct class +// GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); +// } }; template @@ -225,22 +225,28 @@ struct dense_product_impl template static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { - // TODO: use the following instead of calling call_assignment + // TODO: use the following instead of calling call_assignment, same for the other methods // dst = lazyprod(lhs,rhs); call_assignment(dst, lazyprod(lhs,rhs), internal::assign_op()); } template static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) - { dst += lazyprod(lhs,rhs); } + { + // dst += lazyprod(lhs,rhs); + call_assignment(dst, lazyprod(lhs,rhs), internal::add_assign_op()); + } template static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) - { dst -= lazyprod(lhs,rhs); } + { + // dst -= lazyprod(lhs,rhs); + call_assignment(dst, lazyprod(lhs,rhs), internal::sub_assign_op()); + } - template - static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) - { dst += alpha * lazyprod(lhs,rhs); } +// template +// static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) +// { dst += alpha * lazyprod(lhs,rhs); } }; template @@ -286,7 +292,7 @@ struct product_evaluator, ProductTag, DenseShape, CoeffReadCost = traits::CoeffReadCost, Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT, CanVectorizeInner = traits::CanVectorizeInner, - Flags = CoeffBasedProductType::Flags + Flags = traits::Flags }; typedef typename evaluator::type LhsEtorType; diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index 637513132..c987a30b0 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -14,7 +14,7 @@ namespace Eigen { namespace internal { - + /********************************************************************************* * Coefficient based product implementation. * It is designed for the following use cases: @@ -110,6 +110,8 @@ struct traits > } // end namespace internal +#ifndef EIGEN_TEST_EVALUATORS + template class CoeffBasedProduct : internal::no_assignment_operator, @@ -447,6 +449,8 @@ struct product_packet_impl } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS + } // end namespace Eigen #endif // EIGEN_COEFFBASED_PRODUCT_H diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 776eac587..3bc151229 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -36,6 +36,12 @@ template struct accessors_level }; }; +template struct evaluator_traits; + +template< typename T, + typename Kind = typename evaluator_traits::Kind, + typename Scalar = typename T::Scalar> struct evaluator; + } // end namespace internal template struct NumTraits; diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h index 8872c5b64..b9b1ee02a 100644 --- a/Eigen/src/Core/util/StaticAssert.h +++ b/Eigen/src/Core/util/StaticAssert.h @@ -90,7 +90,8 @@ YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED, THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE, THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH, - OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG + OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG, + IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY }; }; diff --git a/test/main.h b/test/main.h index 9dd8bc535..dda451a40 100644 --- a/test/main.h +++ b/test/main.h @@ -66,6 +66,8 @@ namespace Eigen static bool g_has_set_repeat, g_has_set_seed; } +#define TRACK std::cerr << __FILE__ << " " << __LINE__ << std::endl + #define EI_PP_MAKE_STRING2(S) #S #define EI_PP_MAKE_STRING(S) EI_PP_MAKE_STRING2(S) From 8d8acc3ab4f0b3f45f2a8bc25c0b7f5f66a22024 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 4 Dec 2013 22:58:19 +0100 Subject: [PATCH 0220/1103] Move inner product special functions to a base class to avoid ambiguous calls --- Eigen/src/Core/Product.h | 50 ++++++++++++++++++++++++++++++++-------- 1 file changed, 41 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index b37fd2ff7..79d09fbb6 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -79,12 +79,6 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, const LhsNestedCleaned& lhs() const { return m_lhs; } const RhsNestedCleaned& rhs() const { return m_rhs; } - - /** Convertion to scalar for inner-products */ - operator const Scalar() const { - EIGEN_STATIC_ASSERT(SizeAtCompileTime==1, IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY); - return typename internal::evaluator::type(*this).coeff(0,0); - } protected: @@ -92,13 +86,51 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, RhsNested m_rhs; }; +namespace internal { + +template::ret> +class dense_product_base + : public internal::dense_xpr_base >::type +{}; + +/** Convertion to scalar for inner-products */ template -class ProductImpl : public internal::dense_xpr_base >::type +class dense_product_base + : public internal::dense_xpr_base >::type +{ + typedef Product ProductXpr; + typedef typename internal::dense_xpr_base::type Base; +public: + using Base::derived; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + + operator const Scalar() const + { + return typename internal::evaluator::type(derived()).coeff(0,0); + } + + Scalar coeff(Index row, Index col) const + { + return typename internal::evaluator::type(derived()).coeff(row,col); + } + + Scalar coeff(Index i) const + { + return typename internal::evaluator::type(derived()).coeff(i); + } +}; + +} // namespace internal + +template +class ProductImpl + : public internal::dense_product_base { typedef Product Derived; public: - - typedef typename internal::dense_xpr_base >::type Base; + + typedef typename internal::dense_product_base Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) }; From 2ca0ccd2f2a88e7bae3c502ec2082178506a3d81 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 7 Dec 2013 17:17:47 +0100 Subject: [PATCH 0221/1103] Add support for triangular products with evaluators --- Eigen/src/Core/NoAlias.h | 2 + Eigen/src/Core/ProductEvaluators.h | 126 ++++++++++++++++---- Eigen/src/Core/TriangularMatrix.h | 24 ++++ Eigen/src/Core/products/CoeffBasedProduct.h | 4 +- test/evaluators.cpp | 5 + 5 files changed, 137 insertions(+), 24 deletions(-) diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index 6ac525336..3c9c951f0 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -40,6 +40,8 @@ class NoAlias EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase& other) { + // TODO either call resize here or call "call_assignment" through m_expression.lazyAssign() ?? + m_expression.resizeLike(other.derived()); call_assignment(*this, other.derived(), internal::assign_op()); return m_expression; } diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 5dd480cad..c3a9f0db4 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -17,7 +17,7 @@ namespace Eigen { namespace internal { -// Like more general binary expressions, products need they own evaluator: +// Like more general binary expressions, products need their own evaluator: template< typename T, int ProductTag = internal::product_tag::ret, typename LhsShape = typename evaluator_traits::Shape, @@ -39,11 +39,14 @@ struct evaluator > evaluator(const XprType& xpr) : Base(xpr) {} }; -// Helper class to perform a dense product with the destination at hand. +// Helper class to perform a matrix product with the destination at hand. // Depending on the sizes of the factors, there are different evaluation strategies // as controlled by internal::product_type. -template::value> -struct dense_product_impl; +template< typename Lhs, typename Rhs, + typename LhsShape = typename evaluator_traits::Shape, + typename RhsShape = typename evaluator_traits::Shape, + int ProductType = internal::product_type::value> +struct generic_product_impl; template struct evaluator_traits > @@ -52,7 +55,7 @@ struct evaluator_traits > enum { AssumeAliasing = 1 }; }; -// The evaluator for default dense products creates a temporary and call dense_product_impl +// The evaluator for default dense products creates a temporary and call generic_product_impl template struct product_evaluator, ProductTag, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> : public evaluator::PlainObject>::type @@ -65,7 +68,7 @@ struct product_evaluator, ProductTag, DenseSha : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); - dense_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); } protected: @@ -79,7 +82,7 @@ struct Assignment, internal::assign_ typedef Product SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { - dense_product_impl::evalTo(dst, src.lhs(), src.rhs()); + generic_product_impl::evalTo(dst, src.lhs(), src.rhs()); } }; @@ -90,7 +93,7 @@ struct Assignment, internal::add_ass typedef Product SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op &) { - dense_product_impl::addTo(dst, src.lhs(), src.rhs()); + generic_product_impl::addTo(dst, src.lhs(), src.rhs()); } }; @@ -101,12 +104,12 @@ struct Assignment, internal::sub_ass typedef Product SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op &) { - dense_product_impl::subTo(dst, src.lhs(), src.rhs()); + generic_product_impl::subTo(dst, src.lhs(), src.rhs()); } }; template -struct dense_product_impl +struct generic_product_impl { template static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) @@ -128,7 +131,7 @@ struct dense_product_impl template -struct dense_product_impl +struct generic_product_impl { typedef typename Product::Scalar Scalar; @@ -165,7 +168,7 @@ struct dense_product_impl // This base class provides default implementations for evalTo, addTo, subTo, in terms of scaleAndAddTo template -struct dense_product_impl_base +struct generic_product_impl_base { typedef typename Product::Scalar Scalar; @@ -188,7 +191,8 @@ struct dense_product_impl_base }; template -struct dense_product_impl : dense_product_impl_base > +struct generic_product_impl + : generic_product_impl_base > { typedef typename Product::Scalar Scalar; enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight }; @@ -205,20 +209,21 @@ struct dense_product_impl : dense_product_impl_base -struct dense_product_impl : dense_product_impl_base > +struct generic_product_impl + : generic_product_impl_base > { typedef typename Product::Scalar Scalar; -// template -// static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) -// { -// // TODO bypass GeneralProduct class -// GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); -// } + template + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + // TODO bypass GeneralProduct class + GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); + } }; template -struct dense_product_impl +struct generic_product_impl { typedef typename Product::Scalar Scalar; @@ -249,8 +254,10 @@ struct dense_product_impl // { dst += alpha * lazyprod(lhs,rhs); } }; +// This specialization enforces the use of a coefficient-based evaluation strategy template -struct dense_product_impl : dense_product_impl {}; +struct generic_product_impl + : generic_product_impl {}; // Case 2: Evaluate coeff by coeff // @@ -547,6 +554,81 @@ struct etor_product_packet_impl } }; + +/*************************************************************************** +* Triangular products +***************************************************************************/ + +template +struct generic_product_impl + : generic_product_impl_base > +{ + typedef typename Product::Scalar Scalar; + + template + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + // TODO bypass TriangularProduct class + TriangularProduct(lhs.nestedExpression(),rhs).scaleAndAddTo(dst, alpha); + } +}; + +template +struct product_evaluator, ProductTag, TriangularShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + + +template +struct generic_product_impl +: generic_product_impl_base > +{ + typedef typename Product::Scalar Scalar; + + template + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + { + // TODO bypass TriangularProduct class + TriangularProduct(lhs,rhs.nestedExpression()).scaleAndAddTo(dst, alpha); + } +}; + +template +struct product_evaluator, ProductTag, DenseShape, TriangularShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + + + + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 602f0c5d1..28191694d 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -49,6 +49,7 @@ template class TriangularBase : public EigenBase typedef typename internal::traits::Index Index; typedef typename internal::traits::DenseMatrixType DenseMatrixType; typedef DenseMatrixType DenseType; + typedef Derived const& Nested; EIGEN_DEVICE_FUNC inline TriangularBase() { eigen_assert(!((Mode&UnitDiag) && (Mode&ZeroDiag))); } @@ -204,6 +205,7 @@ template class TriangularView enum { Mode = _Mode, + Flags = internal::traits::Flags, TransposeMode = (Mode & Upper ? Lower : 0) | (Mode & Lower ? Upper : 0) | (Mode & (UnitDiag)) @@ -326,6 +328,27 @@ template class TriangularView return m_matrix.transpose(); } +#ifdef EIGEN_TEST_EVALUATORS + + /** Efficient triangular matrix times vector/matrix product */ + template + EIGEN_DEVICE_FUNC + const Product + operator*(const MatrixBase& rhs) const + { + return Product(*this, rhs.derived()); + } + + /** Efficient vector/matrix times triangular matrix product */ + template friend + EIGEN_DEVICE_FUNC + const Product + operator*(const MatrixBase& lhs, const TriangularView& rhs) + { + return Product(lhs.derived(),rhs); + } + +#else // EIGEN_TEST_EVALUATORS /** Efficient triangular matrix times vector/matrix product */ template EIGEN_DEVICE_FUNC @@ -347,6 +370,7 @@ template class TriangularView (lhs.derived(),rhs.m_matrix); } +#endif #ifdef EIGEN2_SUPPORT template diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index c987a30b0..5f8182aa9 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -49,8 +49,8 @@ struct traits > enum { LhsCoeffReadCost = _LhsNested::CoeffReadCost, RhsCoeffReadCost = _RhsNested::CoeffReadCost, - LhsFlags = _LhsNested::Flags, - RhsFlags = _RhsNested::Flags, + LhsFlags = traits<_LhsNested>::Flags, + RhsFlags = traits<_RhsNested>::Flags, RowsAtCompileTime = _LhsNested::RowsAtCompileTime, ColsAtCompileTime = _RhsNested::ColsAtCompileTime, diff --git a/test/evaluators.cpp b/test/evaluators.cpp index c07146260..d4b737348 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -451,5 +451,10 @@ void test_evaluators() C.triangularView().swap(D.triangularView()); swap_using_evaluator(B.triangularView(), A.triangularView()); VERIFY(B.isApprox(C) && "swap_using_evaluator(B.triangularView(), A.triangularView())"); + + + VERIFY_IS_APPROX_EVALUATOR2(B, prod(A.triangularView(),A), MatrixXd(A.triangularView()*A)); + + B.col(0).noalias() = prod( (2.1 * A.adjoint()).triangularView() , (A.row(0)).adjoint() ); } } From e94fe4cc3e371f37b39f7b5f824cd3acc74af823 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 13 Dec 2013 18:06:58 +0100 Subject: [PATCH 0222/1103] fix resizing in noalias for blocks, and make -=/+= use evaluators --- Eigen/src/Core/CwiseBinaryOp.h | 8 ++++++++ Eigen/src/Core/NoAlias.h | 2 +- 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index e20daacc8..5624a4718 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -213,8 +213,12 @@ template EIGEN_STRONG_INLINE Derived & MatrixBase::operator-=(const MatrixBase &other) { +#ifdef EIGEN_TEST_EVALUATORS + call_assignment(derived(), other.derived(), internal::sub_assign_op()); +#else SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); tmp = other.derived(); +#endif return derived(); } @@ -227,8 +231,12 @@ template EIGEN_STRONG_INLINE Derived & MatrixBase::operator+=(const MatrixBase& other) { +#ifdef EIGEN_TEST_EVALUATORS + call_assignment(derived(), other.derived(), internal::add_assign_op()); +#else SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); tmp = other.derived(); +#endif return derived(); } diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index 3c9c951f0..412e37258 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -41,7 +41,7 @@ class NoAlias EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase& other) { // TODO either call resize here or call "call_assignment" through m_expression.lazyAssign() ?? - m_expression.resizeLike(other.derived()); + m_expression.resize(other.derived().rows(), other.derived().cols()); call_assignment(*this, other.derived(), internal::assign_op()); return m_expression; } From 27c068e9d6230398b74a1c7b7146d7842c509de7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 13 Dec 2013 18:09:07 +0100 Subject: [PATCH 0223/1103] Make selfqdjoint products use evaluators --- Eigen/src/Core/ProductEvaluators.h | 71 ++++++++++++++++++++++++++++++ Eigen/src/Core/SelfAdjointView.h | 43 ++++++++++++++++-- test/evaluators.cpp | 3 +- 3 files changed, 113 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index c3a9f0db4..f0eb57d67 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -628,6 +628,77 @@ protected: +/*************************************************************************** +* SelfAdjoint products +***************************************************************************/ + +template +struct generic_product_impl + : generic_product_impl_base > +{ + typedef typename Product::Scalar Scalar; + + template + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + {// SelfadjointProductMatrix + // TODO bypass SelfadjointProductMatrix class + SelfadjointProductMatrix(lhs.nestedExpression(),rhs).scaleAndAddTo(dst, alpha); + } +}; + +template +struct product_evaluator, ProductTag, SelfAdjointShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + + +template +struct generic_product_impl +: generic_product_impl_base > +{ + typedef typename Product::Scalar Scalar; + + template + static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) + {//SelfadjointProductMatrix + // TODO bypass SelfadjointProductMatrix class + SelfadjointProductMatrix(lhs,rhs.nestedExpression()).scaleAndAddTo(dst, alpha); + } +}; + +template +struct product_evaluator, ProductTag, DenseShape, SelfAdjointShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + } // end namespace internal diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 8231e3f5c..079b987f8 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -50,11 +50,12 @@ template class SelfAdjointView - : public TriangularBase > +template class SelfAdjointView + : public TriangularBase > { public: + typedef _MatrixType MatrixType; typedef TriangularBase Base; typedef typename internal::traits::MatrixTypeNested MatrixTypeNested; typedef typename internal::traits::MatrixTypeNestedCleaned MatrixTypeNestedCleaned; @@ -65,7 +66,8 @@ template class SelfAdjointView typedef typename MatrixType::Index Index; enum { - Mode = internal::traits::Mode + Mode = internal::traits::Mode, + Flags = internal::traits::Flags }; typedef typename MatrixType::PlainObject PlainObject; @@ -111,6 +113,28 @@ template class SelfAdjointView EIGEN_DEVICE_FUNC MatrixTypeNestedCleaned& nestedExpression() { return *const_cast(&m_matrix); } +#ifdef EIGEN_TEST_EVALUATORS + + /** Efficient triangular matrix times vector/matrix product */ + template + EIGEN_DEVICE_FUNC + const Product + operator*(const MatrixBase& rhs) const + { + return Product(*this, rhs.derived()); + } + + /** Efficient vector/matrix times triangular matrix product */ + template friend + EIGEN_DEVICE_FUNC + const Product + operator*(const MatrixBase& lhs, const SelfAdjointView& rhs) + { + return Product(lhs.derived(),rhs); + } + +#else // EIGEN_TEST_EVALUATORS + /** Efficient self-adjoint matrix times vector/matrix product */ template EIGEN_DEVICE_FUNC @@ -132,6 +156,7 @@ template class SelfAdjointView (lhs.derived(),rhs.m_matrix); } +#endif /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this: * \f$ this = this + \alpha u v^* + conj(\alpha) v u^* \f$ @@ -311,6 +336,18 @@ struct triangular_assignment_selector +// in the future selfadjoint-ness should be defined by the expression traits +// such that Transpose > is valid. (currently TriangularBase::transpose() is overloaded to make it work) +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef SelfAdjointShape Shape; + + static const int AssumeAliasing = 0; +}; + } // end namespace internal /*************************************************************************** diff --git a/test/evaluators.cpp b/test/evaluators.cpp index d4b737348..69a45661f 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -455,6 +455,7 @@ void test_evaluators() VERIFY_IS_APPROX_EVALUATOR2(B, prod(A.triangularView(),A), MatrixXd(A.triangularView()*A)); - B.col(0).noalias() = prod( (2.1 * A.adjoint()).triangularView() , (A.row(0)).adjoint() ); + VERIFY_IS_APPROX_EVALUATOR2(B, prod(A.selfadjointView(),A), MatrixXd(A.selfadjointView()*A)); + } } From d357bbd9c06f4b6088de0a8e47b3e56fdd0b99b3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 14 Dec 2013 22:53:47 +0100 Subject: [PATCH 0224/1103] Fix a few regression regarding temporaries and products --- Eigen/src/Core/GeneralProduct.h | 53 +++++++++++----------- Eigen/src/Core/Product.h | 2 +- Eigen/src/Core/ProductEvaluators.h | 72 ++++++++++++++++++++++++++++-- Eigen/src/Core/SelfAdjointView.h | 2 + 4 files changed, 98 insertions(+), 31 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 675f6ee8d..f823ff251 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -82,7 +82,8 @@ private: public: enum { - value = selector::ret + value = selector::ret, + ret = selector::ret }; #ifdef EIGEN_DEBUG_PRODUCT static void debug() @@ -98,31 +99,31 @@ public: #endif }; -template struct product_tag -{ -private: - - typedef typename remove_all::type _Lhs; - typedef typename remove_all::type _Rhs; - enum { - Rows = _Lhs::RowsAtCompileTime, - Cols = _Rhs::ColsAtCompileTime, - Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, _Rhs::RowsAtCompileTime) - }; - - enum { - rows_select = Rows==1 ? int(Rows) : int(Large), - cols_select = Cols==1 ? int(Cols) : int(Large), - depth_select = Depth==1 ? int(Depth) : int(Large) - }; - typedef product_type_selector selector; - -public: - enum { - ret = selector::ret - }; - -}; +// template struct product_tag +// { +// private: +// +// typedef typename remove_all::type _Lhs; +// typedef typename remove_all::type _Rhs; +// enum { +// Rows = _Lhs::RowsAtCompileTime, +// Cols = _Rhs::ColsAtCompileTime, +// Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, _Rhs::RowsAtCompileTime) +// }; +// +// enum { +// rows_select = Rows==1 ? int(Rows) : int(Large), +// cols_select = Cols==1 ? int(Cols) : int(Large), +// depth_select = Depth==1 ? int(Depth) : int(Large) +// }; +// typedef product_type_selector selector; +// +// public: +// enum { +// ret = selector::ret +// }; +// +// }; /* The following allows to select the kind of product at compile time * based on the three dimensions of the product. diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 79d09fbb6..d64fbae35 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -88,7 +88,7 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, namespace internal { -template::ret> +template::ret> class dense_product_base : public internal::dense_xpr_base >::type {}; diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index f0eb57d67..46048882b 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -19,7 +19,7 @@ namespace internal { // Like more general binary expressions, products need their own evaluator: template< typename T, - int ProductTag = internal::product_tag::ret, + int ProductTag = internal::product_type::ret, typename LhsShape = typename evaluator_traits::Shape, typename RhsShape = typename evaluator_traits::Shape, typename LhsScalar = typename T::Lhs::Scalar, @@ -38,7 +38,43 @@ struct evaluator > evaluator(const XprType& xpr) : Base(xpr) {} }; + +// Catch scalar * ( A * B ) and transform it to (A*scalar) * B +// TODO we should apply that rule if that's really helpful +template +struct evaluator, const Product > > + : public evaluator,const Lhs>, Rhs, DefaultProduct> > +{ + typedef CwiseUnaryOp, const Product > XprType; + typedef evaluator,const Lhs>, Rhs, DefaultProduct> > Base; + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const XprType& xpr) + : Base(xpr.functor().m_other * xpr.nestedExpression().lhs() * xpr.nestedExpression().rhs()) + {} +}; + + +template +struct evaluator, DiagIndex> > + : public evaluator, DiagIndex> > +{ + typedef Diagonal, DiagIndex> XprType; + typedef evaluator, DiagIndex> > Base; + + typedef evaluator type; + typedef evaluator nestedType; +// + evaluator(const XprType& xpr) + : Base(Diagonal, DiagIndex>( + Product(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()), + xpr.index() )) + {} +}; + + // Helper class to perform a matrix product with the destination at hand. // Depending on the sizes of the factors, there are different evaluation strategies // as controlled by internal::product_type. @@ -108,6 +144,23 @@ struct Assignment, internal::sub_ass } }; + +// Dense ?= scalar * Product +// TODO we should apply that rule if that's really helpful +// for instance, this is not good for inner products +template< typename DstXprType, typename Lhs, typename Rhs, typename AssignFunc, typename Scalar, typename ScalarBis> +struct Assignment, + const Product >, AssignFunc, Dense2Dense, Scalar> +{ + typedef CwiseUnaryOp, + const Product > SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const AssignFunc& func) + { + call_assignment(dst.noalias(), (src.functor().m_other * src.nestedExpression().lhs()) * src.nestedExpression().rhs(), func); + } +}; + + template struct generic_product_impl { @@ -255,9 +308,9 @@ struct generic_product_impl }; // This specialization enforces the use of a coefficient-based evaluation strategy -template -struct generic_product_impl - : generic_product_impl {}; +// template +// struct generic_product_impl +// : generic_product_impl {}; // Case 2: Evaluate coeff by coeff // @@ -347,6 +400,17 @@ protected: Index m_innerDim; }; +template +struct product_evaluator, LazyCoeffBasedProductMode, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar > + : product_evaluator, CoeffBasedProductMode, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar > +{ + typedef Product XprType; + typedef Product BaseProduct; + typedef product_evaluator Base; + product_evaluator(const XprType& xpr) + : Base(BaseProduct(xpr.lhs(),xpr.rhs())) + {} +}; /*************************************************************************** * Normal product .coeff() implementation (with meta-unrolling) diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 079b987f8..a5c6d5ceb 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -336,6 +336,7 @@ struct triangular_assignment_selector // in the future selfadjoint-ness should be defined by the expression traits // such that Transpose > is valid. (currently TriangularBase::transpose() is overloaded to make it work) @@ -347,6 +348,7 @@ struct evaluator_traits > static const int AssumeAliasing = 0; }; +#endif // EIGEN_ENABLE_EVALUATORS } // end namespace internal From 6b6071866bf4d7832e623ae194fd752e16832765 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 25 Feb 2014 18:55:16 +0100 Subject: [PATCH 0225/1103] Make pivoting HouseholderQR compatible with custom scalar types --- Eigen/src/QR/ColPivHouseholderQR.h | 2 +- Eigen/src/QR/FullPivHouseholderQR.h | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index 07126a9f4..4824880f5 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -350,7 +350,7 @@ template class ColPivHouseholderQR return m_usePrescribedThreshold ? m_prescribedThreshold // this formula comes from experimenting (see "LU precision tuning" thread on the list) // and turns out to be identical to Higham's formula used already in LDLt. - : NumTraits::epsilon() * m_qr.diagonalSize(); + : NumTraits::epsilon() * RealScalar(m_qr.diagonalSize()); } /** \returns the number of nonzero pivots in the QR decomposition. diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 44eaa1b1a..a7b0fc16f 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -346,7 +346,7 @@ template class FullPivHouseholderQR return m_usePrescribedThreshold ? m_prescribedThreshold // this formula comes from experimenting (see "LU precision tuning" thread on the list) // and turns out to be identical to Higham's formula used already in LDLt. - : NumTraits::epsilon() * m_qr.diagonalSize(); + : NumTraits::epsilon() * RealScalar(m_qr.diagonalSize()); } /** \returns the number of nonzero pivots in the QR decomposition. @@ -417,7 +417,7 @@ FullPivHouseholderQR& FullPivHouseholderQR::compute(cons m_temp.resize(cols); - m_precision = NumTraits::epsilon() * size; + m_precision = NumTraits::epsilon() * RealScalar(size); m_rows_transpositions.resize(size); m_cols_transpositions.resize(size); From ac69d8769f457366581b50ae8fa620634cd9aaa1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 26 Feb 2014 10:12:27 +0100 Subject: [PATCH 0226/1103] Remove early termination in LDLT: the zero on the diagonal of the input matrix does not mean the matrix is not full rank. Typical examples are matrices coming from LS with linear equality constraints. --- Eigen/src/Cholesky/LDLT.h | 14 ++++---------- test/cholesky.cpp | 32 ++++++++++++++++++++++++++++++++ 2 files changed, 36 insertions(+), 10 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index f34d26465..b43e85e7f 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -291,13 +291,6 @@ template<> struct ldlt_inplace cutoff = abs(NumTraits::epsilon() * biggest_in_corner); } - // Finish early if the matrix is not full rank. - if(biggest_in_corner < cutoff) - { - for(Index i = k; i < size; i++) transpositions.coeffRef(i) = i; - break; - } - transpositions.coeffRef(k) = index_of_biggest_in_corner; if(k != index_of_biggest_in_corner) { @@ -333,6 +326,7 @@ template<> struct ldlt_inplace if(rs>0) A21.noalias() -= A20 * temp.head(k); } + if((rs>0) && (abs(mat.coeffRef(k,k)) > cutoff)) A21 /= mat.coeffRef(k,k); @@ -518,12 +512,12 @@ struct solve_retval, Rhs> typedef typename LDLTType::RealScalar RealScalar; const Diagonal vectorD = dec().vectorD(); RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() * NumTraits::epsilon(), - RealScalar(1) / NumTraits::highest()); // motivated by LAPACK's xGELSS + RealScalar(1) / NumTraits::highest()); // motivated by LAPACK's xGELSS for (Index i = 0; i < vectorD.size(); ++i) { if(abs(vectorD(i)) > tolerance) - dst.row(i) /= vectorD(i); + dst.row(i) /= vectorD(i); else - dst.row(i).setZero(); + dst.row(i).setZero(); } // dst = L^-T (D^-1 L^-1 P b) diff --git a/test/cholesky.cpp b/test/cholesky.cpp index b980dc572..d4d90e467 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -179,6 +179,38 @@ template void cholesky(const MatrixType& m) // restore if(sign == -1) symm = -symm; + + // check matrices coming from linear constraints with Lagrange multipliers + if(rows>=3) + { + SquareMatrixType A = symm; + int c = internal::random(0,rows-2); + A.bottomRightCorner(c,c).setZero(); + // Make sure a solution exists: + vecX.setRandom(); + vecB = A * vecX; + vecX.setZero(); + ldltlo.compute(A); + VERIFY_IS_APPROX(A, ldltlo.reconstructedMatrix()); + vecX = ldltlo.solve(vecB); + VERIFY_IS_APPROX(A * vecX, vecB); + } + + // check non-full rank matrices + if(rows>=3) + { + int r = internal::random(1,rows-1); + Matrix a = Matrix::Random(rows,r); + SquareMatrixType A = a * a.adjoint(); + // Make sure a solution exists: + vecX.setRandom(); + vecB = A * vecX; + vecX.setZero(); + ldltlo.compute(A); + VERIFY_IS_APPROX(A, ldltlo.reconstructedMatrix()); + vecX = ldltlo.solve(vecB); + VERIFY_IS_APPROX(A * vecX, vecB); + } } // update/downdate From 41e89c73c755fe70f8babea4a8a3b491057e5c39 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 27 Feb 2014 12:57:24 +0100 Subject: [PATCH 0227/1103] Regression test for bug #752 --- test/block.cpp | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-) diff --git a/test/block.cpp b/test/block.cpp index 189fa5aba..9ed5d7bc5 100644 --- a/test/block.cpp +++ b/test/block.cpp @@ -10,6 +10,26 @@ #define EIGEN_NO_STATIC_ASSERT // otherwise we fail at compile time on unused paths #include "main.h" +template +typename Eigen::internal::enable_if::IsComplex,typename MatrixType::Scalar>::type +block_real_only(const MatrixType &m1, Index r1, Index r2, Index c1, Index c2, const Scalar& s1) { + // check cwise-Functions: + VERIFY_IS_APPROX(m1.row(r1).cwiseMax(s1), m1.cwiseMax(s1).row(r1)); + VERIFY_IS_APPROX(m1.col(c1).cwiseMin(s1), m1.cwiseMin(s1).col(c1)); + + VERIFY_IS_APPROX(m1.block(r1,c1,r2-r1+1,c2-c1+1).cwiseMin(s1), m1.cwiseMin(s1).block(r1,c1,r2-r1+1,c2-c1+1)); + VERIFY_IS_APPROX(m1.block(r1,c1,r2-r1+1,c2-c1+1).cwiseMax(s1), m1.cwiseMax(s1).block(r1,c1,r2-r1+1,c2-c1+1)); + + return Scalar(0); +} + +template +typename Eigen::internal::enable_if::IsComplex,typename MatrixType::Scalar>::type +block_real_only(const MatrixType &, Index, Index, Index, Index, const Scalar&) { + return Scalar(0); +} + + template void block(const MatrixType& m) { typedef typename MatrixType::Index Index; @@ -37,6 +57,8 @@ template void block(const MatrixType& m) Index c1 = internal::random(0,cols-1); Index c2 = internal::random(c1,cols-1); + block_real_only(m1, r1, r2, c1, c1, s1); + //check row() and col() VERIFY_IS_EQUAL(m1.col(c1).transpose(), m1.transpose().row(c1)); //check operator(), both constant and non-constant, on row() and col() @@ -51,7 +73,8 @@ template void block(const MatrixType& m) VERIFY_IS_APPROX(m1.col(c1), m1_copy.col(c1) + s1 * m1_copy.col(c2)); m1.col(c1).col(0) += s1 * m1_copy.col(c2); VERIFY_IS_APPROX(m1.col(c1), m1_copy.col(c1) + Scalar(2) * s1 * m1_copy.col(c2)); - + + //check block() Matrix b1(1,1); b1(0,0) = m1(r1,c1); From 76d2ca27e51179b67a246a51d9f0396991606ac8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 28 Feb 2014 13:11:39 +0100 Subject: [PATCH 0228/1103] Fix PaStiX support for Pastix 5.2 --- Eigen/src/PaStiXSupport/PaStiXSupport.h | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h index a955287d1..8a546dc2f 100644 --- a/Eigen/src/PaStiXSupport/PaStiXSupport.h +++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h @@ -12,6 +12,14 @@ namespace Eigen { +#if defined(DCOMPLEX) + #define PASTIX_COMPLEX COMPLEX + #define PASTIX_DCOMPLEX DCOMPLEX +#else + #define PASTIX_COMPLEX std::complex + #define PASTIX_DCOMPLEX std::complex +#endif + /** \ingroup PaStiXSupport_Module * \brief Interface to the PaStix solver * @@ -74,14 +82,14 @@ namespace internal { if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; } if (nbrhs == 0) {x = NULL; nbrhs=1;} - c_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast(vals), perm, invp, reinterpret_cast(x), nbrhs, iparm, dparm); + c_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast(vals), perm, invp, reinterpret_cast(x), nbrhs, iparm, dparm); } void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex *vals, int *perm, int * invp, std::complex *x, int nbrhs, int *iparm, double *dparm) { if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; } if (nbrhs == 0) {x = NULL; nbrhs=1;} - z_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast(vals), perm, invp, reinterpret_cast(x), nbrhs, iparm, dparm); + z_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast(vals), perm, invp, reinterpret_cast(x), nbrhs, iparm, dparm); } // Convert the matrix to Fortran-style Numbering From 47679c50ae2f51edaa18e4299ae361cfa774be96 Mon Sep 17 00:00:00 2001 From: Olivier Saut Date: Mon, 3 Mar 2014 14:44:39 +0100 Subject: [PATCH 0229/1103] Typo in the example for Eigen::SelfAdjointEigenSolver::eigenvectors, the first eigenvector should be col(0) not col(1) --- doc/snippets/EigenSolver_eigenvectors.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/doc/snippets/EigenSolver_eigenvectors.cpp b/doc/snippets/EigenSolver_eigenvectors.cpp index 0fad4dadb..8355f76c9 100644 --- a/doc/snippets/EigenSolver_eigenvectors.cpp +++ b/doc/snippets/EigenSolver_eigenvectors.cpp @@ -1,4 +1,4 @@ MatrixXd ones = MatrixXd::Ones(3,3); EigenSolver es(ones); -cout << "The first eigenvector of the 3x3 matrix of ones is:" - << endl << es.eigenvectors().col(1) << endl; +cout << "The first eigenvector of the 3x3 matrix of ones is:" + << endl << es.eigenvectors().col(0) << endl; From 04e1e38eedad891862083de85b965b8660dea6f7 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 4 Mar 2014 15:10:29 +0100 Subject: [PATCH 0230/1103] bug #289: Removed useless static keywords --- Eigen/src/Core/util/Memory.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index bc472d720..7bd3056d5 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -464,7 +464,7 @@ template inline void conditional_aligned_delete_auto(T * * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h. */ template -static inline Index first_aligned(const Scalar* array, Index size) +inline Index first_aligned(const Scalar* array, Index size) { enum { PacketSize = packet_traits::size, PacketAlignedMask = PacketSize-1 @@ -492,7 +492,7 @@ static inline Index first_aligned(const Scalar* array, Index size) /** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size */ template -inline static Index first_multiple(Index size, Index base) +inline Index first_multiple(Index size, Index base) { return ((size+base-1)/base)*base; } From 7313f32efa5fe881469e14f81268a98c6d488ee7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 4 Mar 2014 17:24:00 +0100 Subject: [PATCH 0231/1103] Help MSVC to inline some trivial functions --- Eigen/src/Core/ArrayWrapper.h | 2 +- Eigen/src/Core/MatrixBase.h | 8 ++++---- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/ArrayWrapper.h b/Eigen/src/Core/ArrayWrapper.h index 21830745b..4bb648024 100644 --- a/Eigen/src/Core/ArrayWrapper.h +++ b/Eigen/src/Core/ArrayWrapper.h @@ -49,7 +49,7 @@ class ArrayWrapper : public ArrayBase > typedef typename internal::nested::type NestedExpressionType; EIGEN_DEVICE_FUNC - inline ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {} + EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {} EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 49b69f873..172929562 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -356,8 +356,6 @@ template class MatrixBase Scalar trace() const; -/////////// Array module /////////// - template EIGEN_DEVICE_FUNC RealScalar lpNorm() const; EIGEN_DEVICE_FUNC MatrixBase& matrix() { return *this; } @@ -365,8 +363,10 @@ template class MatrixBase /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix * \sa ArrayBase::matrix() */ - EIGEN_DEVICE_FUNC ArrayWrapper array() { return derived(); } - EIGEN_DEVICE_FUNC const ArrayWrapper array() const { return derived(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper array() { return derived(); } + /** \returns a const \link Eigen::ArrayBase Array \endlink expression of this matrix + * \sa ArrayBase::matrix() */ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper array() const { return derived(); } /////////// LU module /////////// From d5cc083782bc59049755696f2d22edfb0172d124 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 5 Mar 2014 14:50:00 +0100 Subject: [PATCH 0232/1103] Fixed bug #754. Only inserted (!defined(_WIN32_WCE)) analog to alloc and free implementation (not tested, but should be correct). --- Eigen/src/Core/util/Memory.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 7bd3056d5..286e963d2 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -274,12 +274,12 @@ inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size) // The defined(_mm_free) is just here to verify that this MSVC version // implements _mm_malloc/_mm_free based on the corresponding _aligned_ // functions. This may not always be the case and we just try to be safe. - #if defined(_MSC_VER) && defined(_mm_free) + #if defined(_MSC_VER) && (!defined(_WIN32_WCE)) && defined(_mm_free) result = _aligned_realloc(ptr,new_size,16); #else result = generic_aligned_realloc(ptr,new_size,old_size); #endif -#elif defined(_MSC_VER) +#elif defined(_MSC_VER) && (!defined(_WIN32_WCE)) result = _aligned_realloc(ptr,new_size,16); #else result = handmade_aligned_realloc(ptr,new_size,old_size); From ce41b72eb8b0657957893275e4936b0961bae80b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 7 Mar 2014 23:09:39 +0100 Subject: [PATCH 0233/1103] Extend sizeof unit test --- test/sizeof.cpp | 2 ++ 1 file changed, 2 insertions(+) diff --git a/test/sizeof.cpp b/test/sizeof.cpp index d9ad35620..7044d2062 100644 --- a/test/sizeof.cpp +++ b/test/sizeof.cpp @@ -21,6 +21,8 @@ template void verifySizeOf(const MatrixType&) void test_sizeof() { CALL_SUBTEST(verifySizeOf(Matrix()) ); + CALL_SUBTEST(verifySizeOf(Vector2d()) ); + CALL_SUBTEST(verifySizeOf(Vector4f()) ); CALL_SUBTEST(verifySizeOf(Matrix4d()) ); CALL_SUBTEST(verifySizeOf(Matrix()) ); CALL_SUBTEST(verifySizeOf(Matrix()) ); From 33ca9b4ee690ed12bd072273ea0d24c5eaae77fd Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 7 Mar 2014 23:11:38 +0100 Subject: [PATCH 0234/1103] Add support for OSX in BTL and fix a few warnings --- bench/btl/CMakeLists.txt | 8 +++- bench/btl/generic_bench/bench.hh | 4 +- bench/btl/generic_bench/btl.hh | 2 +- bench/btl/generic_bench/init/init_matrix.hh | 10 ++-- bench/btl/generic_bench/init/init_vector.hh | 2 +- .../timers/portable_perf_analyzer.hh | 2 +- .../generic_bench/timers/portable_timer.hh | 46 ++++++++++++++++++- bench/btl/libs/eigen3/eigen3_interface.hh | 8 ++-- 8 files changed, 65 insertions(+), 17 deletions(-) diff --git a/bench/btl/CMakeLists.txt b/bench/btl/CMakeLists.txt index 119b470d9..0ac5c15fc 100644 --- a/bench/btl/CMakeLists.txt +++ b/bench/btl/CMakeLists.txt @@ -48,6 +48,12 @@ include_directories( # set(DEFAULT_LIBRARIES ${MKL_LIBRARIES}) # endif (MKL_FOUND) +find_library(EIGEN_BTL_RT_LIBRARY rt) +# if we cannot find it easily, then we don't need it! +if(NOT EIGEN_BTL_RT_LIBRARY) + set(EIGEN_BTL_RT_LIBRARY "") +endif() + MACRO(BTL_ADD_BENCH targetname) foreach(_current_var ${ARGN}) @@ -70,7 +76,7 @@ MACRO(BTL_ADD_BENCH targetname) IF(BUILD_${targetname}) ADD_EXECUTABLE(${targetname} ${_sources}) ADD_TEST(${targetname} "${targetname}") - target_link_libraries(${targetname} ${DEFAULT_LIBRARIES} rt) + target_link_libraries(${targetname} ${DEFAULT_LIBRARIES} ${EIGEN_BTL_RT_LIBRARY}) ENDIF(BUILD_${targetname}) ENDMACRO(BTL_ADD_BENCH) diff --git a/bench/btl/generic_bench/bench.hh b/bench/btl/generic_bench/bench.hh index 005c36395..7b7b951b5 100644 --- a/bench/btl/generic_bench/bench.hh +++ b/bench/btl/generic_bench/bench.hh @@ -102,8 +102,8 @@ BTL_DONT_INLINE void bench( int size_min, int size_max, int nb_point ) // merge the two data std::vector newSizes; std::vector newFlops; - int i=0; - int j=0; + unsigned int i=0; + unsigned int j=0; while (i BTL_DONT_INLINE void init_matrix(Vector & A, int size){ A.resize(size); - for (int row=0; row(A[row],size,row); } } @@ -50,11 +50,11 @@ BTL_DONT_INLINE void init_matrix(Vector & A, int size){ template BTL_DONT_INLINE void init_matrix_symm(Matrix& A, int size){ A.resize(size); - for (int row=0; row @@ -87,6 +87,48 @@ }; // Portable_Timer +#elif defined(__APPLE__) +#include +#include + + +class Portable_Timer +{ + public: + + Portable_Timer() + { + } + + void start() + { + m_start_time = double(mach_absolute_time())*1e-9;; + + } + + void stop() + { + m_stop_time = double(mach_absolute_time())*1e-9;; + + } + + double elapsed() + { + return user_time(); + } + + double user_time() + { + return m_stop_time - m_start_time; + } + + +private: + + double m_stop_time, m_start_time; + +}; // Portable_Timer (Apple) + #else #include @@ -138,7 +180,7 @@ private: int m_clkid; double m_stop_time, m_start_time; -}; // Portable_Timer +}; // Portable_Timer (Linux) #endif diff --git a/bench/btl/libs/eigen3/eigen3_interface.hh b/bench/btl/libs/eigen3/eigen3_interface.hh index 31bcc1f93..0c8aa74da 100644 --- a/bench/btl/libs/eigen3/eigen3_interface.hh +++ b/bench/btl/libs/eigen3/eigen3_interface.hh @@ -52,8 +52,8 @@ public : static BTL_DONT_INLINE void matrix_from_stl(gene_matrix & A, stl_matrix & A_stl){ A.resize(A_stl[0].size(), A_stl.size()); - for (int j=0; j Date: Fri, 7 Mar 2014 23:13:14 +0100 Subject: [PATCH 0235/1103] Fix typo and formating --- Eigen/src/Geometry/Scaling.h | 8 ++++---- doc/TopicMultithreading.dox | 2 +- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Geometry/Scaling.h b/Eigen/src/Geometry/Scaling.h index 1c25f36fe..023fba2ee 100644 --- a/Eigen/src/Geometry/Scaling.h +++ b/Eigen/src/Geometry/Scaling.h @@ -62,10 +62,10 @@ public: template inline Transform operator* (const Transform& t) const { - Transform res = t; - res.prescale(factor()); - return res; -} + Transform res = t; + res.prescale(factor()); + return res; + } /** Concatenates a uniform scaling and a linear transformation matrix */ // TODO returns an expression diff --git a/doc/TopicMultithreading.dox b/doc/TopicMultithreading.dox index fb944af29..ba5e26290 100644 --- a/doc/TopicMultithreading.dox +++ b/doc/TopicMultithreading.dox @@ -39,7 +39,7 @@ int main(int argc, char** argv) } \endcode -\warning note that all functions generating random matrices are \b not re-entrant nor thread-safe. Those include DenseBase::Random(), and DenseBase::setRandom() despite a call to Eigen::initParallel(). This is because these functions are based on std::rand which is not re-entrant. For thread-safe random generator, we recommend the use of boost::random of c++11 random feature. +\warning note that all functions generating random matrices are \b not re-entrant nor thread-safe. Those include DenseBase::Random(), and DenseBase::setRandom() despite a call to Eigen::initParallel(). This is because these functions are based on std::rand which is not re-entrant. For thread-safe random generator, we recommend the use of boost::random or c++11 random feature. In the case your application is parallelized with OpenMP, you might want to disable Eigen's own parallization as detailed in the previous section. From ba2f79e6800ca2762821fe5ab537bc5f10a1268a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 7 Mar 2014 23:18:20 +0100 Subject: [PATCH 0236/1103] Fix selfadjoint_matrix_vector_product for complex with packet size > 2 (e.g., AVX) --- Eigen/src/Core/products/SelfadjointMatrixVector.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index f698f67f9..fdc81205a 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -113,9 +113,9 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product huge speed up) // gcc 4.2 does this optimization automatically. From 033ee7f6d91286920d56de1003f582f4570ff5b3 Mon Sep 17 00:00:00 2001 From: Henry de Valence Date: Sat, 8 Mar 2014 00:44:56 -0500 Subject: [PATCH 0237/1103] Fix typo: 'explicitely' -> 'explicitly' --- doc/TopicLinearAlgebraDecompositions.dox | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/doc/TopicLinearAlgebraDecompositions.dox b/doc/TopicLinearAlgebraDecompositions.dox index 8649cc27b..77f2c92ab 100644 --- a/doc/TopicLinearAlgebraDecompositions.dox +++ b/doc/TopicLinearAlgebraDecompositions.dox @@ -249,7 +249,7 @@ For an introduction on linear solvers and decompositions, check this \link Tutor
Implicit Multi Threading (MT)
Means the algorithm can take advantage of multicore processors via OpenMP. "Implicit" means the algortihm itself is not parallelized, but that it relies on parallelized matrix-matrix product rountines.
Explicit Multi Threading (MT)
-
Means the algorithm is explicitely parallelized to take advantage of multicore processors via OpenMP.
+
Means the algorithm is explicitly parallelized to take advantage of multicore processors via OpenMP.
Meta-unroller
Means the algorithm is automatically and explicitly unrolled for very small fixed size matrices.
From ce99b502ce1d80129a028ddf02fff51f6c51249b Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 17 Dec 2013 10:49:43 -0800 Subject: [PATCH 0238/1103] Use vectorization when packing row-major rhs matrices. (bug #717) --- .../Core/products/GeneralBlockPanelKernel.h | 19 +++++++++++++------ 1 file changed, 13 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 08cc14bd7..686ff84f1 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -1261,6 +1261,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs struct gemm_pack_rhs { + typedef typename packet_traits::type Packet; enum { PacketSize = packet_traits::size }; EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0); }; @@ -1282,12 +1283,18 @@ EIGEN_DONT_INLINE void gemm_pack_rhs(&rhs[k*rhsStride + j2]); + pstoreu(blockB+count, cj.pconj(A)); + count += PacketSize; + } else { + const Scalar* b0 = &rhs[k*rhsStride + j2]; + blockB[count+0] = cj(b0[0]); + blockB[count+1] = cj(b0[1]); + if(nr==4) blockB[count+2] = cj(b0[2]); + if(nr==4) blockB[count+3] = cj(b0[3]); + count += nr; + } } // skip what we have after if(PanelMode) count += nr * (stride-offset-depth); From 8a49dd562624de8ae290b4e38705d5d3540d0f0b Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 19 Dec 2013 11:55:17 +0100 Subject: [PATCH 0239/1103] Fixed typos in comments --- Eigen/src/Geometry/Quaternion.h | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index d036c018a..a6a8df203 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -194,11 +194,11 @@ public: * \brief The quaternion class used to represent 3D orientations and rotations * * \tparam _Scalar the scalar type, i.e., the type of the coefficients - * \tparam _Options controls the memory alignement of the coeffecients. Can be \# AutoAlign or \# DontAlign. Default is AutoAlign. + * \tparam _Options controls the memory alignment of the coefficients. Can be \# AutoAlign or \# DontAlign. Default is AutoAlign. * * This class represents a quaternion \f$ w+xi+yj+zk \f$ that is a convenient representation of * orientations and rotations of objects in three dimensions. Compared to other representations - * like Euler angles or 3x3 matrices, quatertions offer the following advantages: + * like Euler angles or 3x3 matrices, quaternions offer the following advantages: * \li \b compact storage (4 scalars) * \li \b efficient to compose (28 flops), * \li \b stable spherical interpolation @@ -385,7 +385,7 @@ class Map, _Options > /** Constructs a Mapped Quaternion object from the pointer \a coeffs * - * The pointer \a coeffs must reference the four coeffecients of Quaternion in the following order: + * The pointer \a coeffs must reference the four coefficients of Quaternion in the following order: * \code *coeffs == {x, y, z, w} \endcode * * If the template parameter _Options is set to #Aligned, then the pointer coeffs must be aligned. */ @@ -399,16 +399,16 @@ class Map, _Options > }; /** \ingroup Geometry_Module - * Map an unaligned array of single precision scalar as a quaternion */ + * Map an unaligned array of single precision scalars as a quaternion */ typedef Map, 0> QuaternionMapf; /** \ingroup Geometry_Module - * Map an unaligned array of double precision scalar as a quaternion */ + * Map an unaligned array of double precision scalars as a quaternion */ typedef Map, 0> QuaternionMapd; /** \ingroup Geometry_Module - * Map a 16-bits aligned array of double precision scalars as a quaternion */ + * Map a 16-byte aligned array of single precision scalars as a quaternion */ typedef Map, Aligned> QuaternionMapAlignedf; /** \ingroup Geometry_Module - * Map a 16-bits aligned array of double precision scalars as a quaternion */ + * Map a 16-byte aligned array of double precision scalars as a quaternion */ typedef Map, Aligned> QuaternionMapAlignedd; /*************************************************************************** @@ -579,7 +579,7 @@ inline Derived& QuaternionBase::setFromTwoVectors(const MatrixBase accuraletly compute the rotation axis by computing the + // => accurately compute the rotation axis by computing the // intersection of the two planes. This is done by solving: // x^T v0 = 0 // x^T v1 = 0 From 1200bd2ef088da242b8e0d34cb189a36e8754336 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Sat, 21 Dec 2013 21:46:27 +0100 Subject: [PATCH 0240/1103] Grafted from 5725:cdedc9e90d21099e8b3191f95425680ebe710d6f and resolved conflicts --- Eigen/src/plugins/BlockMethods.h | 240 ++++++++++++++++++------------- 1 file changed, 137 insertions(+), 103 deletions(-) diff --git a/Eigen/src/plugins/BlockMethods.h b/Eigen/src/plugins/BlockMethods.h index 3bc345211..9b7fdc4aa 100644 --- a/Eigen/src/plugins/BlockMethods.h +++ b/Eigen/src/plugins/BlockMethods.h @@ -119,13 +119,13 @@ inline const Block topRightCorner() const /** \returns an expression of a top-right corner of *this. * - * \tparam CRows number of rows in corner as specified at compile time - * \tparam CCols number of columns in corner as specified at compile time - * \param cRows number of rows in corner as specified at run time - * \param cCols number of columns in corner as specified at run time + * \tparam CRows number of rows in corner as specified at compile-time + * \tparam CCols number of columns in corner as specified at compile-time + * \param cRows number of rows in corner as specified at run-time + * \param cCols number of columns in corner as specified at run-time * - * This function is mainly useful for corners where the number of rows is specified at compile time - * and the number of columns is specified at run time, or vice versa. The compile-time and run-time + * This function is mainly useful for corners where the number of rows is specified at compile-time + * and the number of columns is specified at run-time, or vice versa. The compile-time and run-time * information should not contradict. In other words, \a cRows should equal \a CRows unless * \a CRows is \a Dynamic, and the same for the number of columns. * @@ -198,13 +198,13 @@ inline const Block topLeftCorner() const /** \returns an expression of a top-left corner of *this. * - * \tparam CRows number of rows in corner as specified at compile time - * \tparam CCols number of columns in corner as specified at compile time - * \param cRows number of rows in corner as specified at run time - * \param cCols number of columns in corner as specified at run time + * \tparam CRows number of rows in corner as specified at compile-time + * \tparam CCols number of columns in corner as specified at compile-time + * \param cRows number of rows in corner as specified at run-time + * \param cCols number of columns in corner as specified at run-time * - * This function is mainly useful for corners where the number of rows is specified at compile time - * and the number of columns is specified at run time, or vice versa. The compile-time and run-time + * This function is mainly useful for corners where the number of rows is specified at compile-time + * and the number of columns is specified at run-time, or vice versa. The compile-time and run-time * information should not contradict. In other words, \a cRows should equal \a CRows unless * \a CRows is \a Dynamic, and the same for the number of columns. * @@ -277,13 +277,13 @@ inline const Block bottomRightCorner() const /** \returns an expression of a bottom-right corner of *this. * - * \tparam CRows number of rows in corner as specified at compile time - * \tparam CCols number of columns in corner as specified at compile time - * \param cRows number of rows in corner as specified at run time - * \param cCols number of columns in corner as specified at run time + * \tparam CRows number of rows in corner as specified at compile-time + * \tparam CCols number of columns in corner as specified at compile-time + * \param cRows number of rows in corner as specified at run-time + * \param cCols number of columns in corner as specified at run-time * - * This function is mainly useful for corners where the number of rows is specified at compile time - * and the number of columns is specified at run time, or vice versa. The compile-time and run-time + * This function is mainly useful for corners where the number of rows is specified at compile-time + * and the number of columns is specified at run-time, or vice versa. The compile-time and run-time * information should not contradict. In other words, \a cRows should equal \a CRows unless * \a CRows is \a Dynamic, and the same for the number of columns. * @@ -356,13 +356,13 @@ inline const Block bottomLeftCorner() const /** \returns an expression of a bottom-left corner of *this. * - * \tparam CRows number of rows in corner as specified at compile time - * \tparam CCols number of columns in corner as specified at compile time - * \param cRows number of rows in corner as specified at run time - * \param cCols number of columns in corner as specified at run time + * \tparam CRows number of rows in corner as specified at compile-time + * \tparam CCols number of columns in corner as specified at compile-time + * \param cRows number of rows in corner as specified at run-time + * \param cCols number of columns in corner as specified at run-time * - * This function is mainly useful for corners where the number of rows is specified at compile time - * and the number of columns is specified at run time, or vice versa. The compile-time and run-time + * This function is mainly useful for corners where the number of rows is specified at compile-time + * and the number of columns is specified at run-time, or vice versa. The compile-time and run-time * information should not contradict. In other words, \a cRows should equal \a CRows unless * \a CRows is \a Dynamic, and the same for the number of columns. * @@ -410,7 +410,11 @@ inline ConstRowsBlockXpr topRows(Index n) const /** \returns a block consisting of the top rows of *this. * - * \tparam N the number of rows in the block + * \tparam N the number of rows in the block as specified at compile-time + * \param n the number of rows in the block as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include MatrixBase_template_int_topRows.cpp * Output: \verbinclude MatrixBase_template_int_topRows.out @@ -419,17 +423,17 @@ inline ConstRowsBlockXpr topRows(Index n) const */ template EIGEN_DEVICE_FUNC -inline typename NRowsBlockXpr::Type topRows() +inline typename NRowsBlockXpr::Type topRows(Index n = N) { - return typename NRowsBlockXpr::Type(derived(), 0, 0, N, cols()); + return typename NRowsBlockXpr::Type(derived(), 0, 0, n, cols()); } /** This is the const version of topRows().*/ template EIGEN_DEVICE_FUNC -inline typename ConstNRowsBlockXpr::Type topRows() const +inline typename ConstNRowsBlockXpr::Type topRows(Index n = N) const { - return typename ConstNRowsBlockXpr::Type(derived(), 0, 0, N, cols()); + return typename ConstNRowsBlockXpr::Type(derived(), 0, 0, n, cols()); } @@ -458,7 +462,11 @@ inline ConstRowsBlockXpr bottomRows(Index n) const /** \returns a block consisting of the bottom rows of *this. * - * \tparam N the number of rows in the block + * \tparam N the number of rows in the block as specified at compile-time + * \param n the number of rows in the block as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include MatrixBase_template_int_bottomRows.cpp * Output: \verbinclude MatrixBase_template_int_bottomRows.out @@ -467,17 +475,17 @@ inline ConstRowsBlockXpr bottomRows(Index n) const */ template EIGEN_DEVICE_FUNC -inline typename NRowsBlockXpr::Type bottomRows() +inline typename NRowsBlockXpr::Type bottomRows(Index n = N) { - return typename NRowsBlockXpr::Type(derived(), rows() - N, 0, N, cols()); + return typename NRowsBlockXpr::Type(derived(), rows() - n, 0, n, cols()); } /** This is the const version of bottomRows().*/ template EIGEN_DEVICE_FUNC -inline typename ConstNRowsBlockXpr::Type bottomRows() const +inline typename ConstNRowsBlockXpr::Type bottomRows(Index n = N) const { - return typename ConstNRowsBlockXpr::Type(derived(), rows() - N, 0, N, cols()); + return typename ConstNRowsBlockXpr::Type(derived(), rows() - n, 0, n, cols()); } @@ -485,7 +493,7 @@ inline typename ConstNRowsBlockXpr::Type bottomRows() const /** \returns a block consisting of a range of rows of *this. * * \param startRow the index of the first row in the block - * \param numRows the number of rows in the block + * \param n the number of rows in the block * * Example: \include DenseBase_middleRows_int.cpp * Output: \verbinclude DenseBase_middleRows_int.out @@ -493,22 +501,26 @@ inline typename ConstNRowsBlockXpr::Type bottomRows() const * \sa class Block, block(Index,Index,Index,Index) */ EIGEN_DEVICE_FUNC -inline RowsBlockXpr middleRows(Index startRow, Index numRows) +inline RowsBlockXpr middleRows(Index startRow, Index n) { - return RowsBlockXpr(derived(), startRow, 0, numRows, cols()); + return RowsBlockXpr(derived(), startRow, 0, n, cols()); } /** This is the const version of middleRows(Index,Index).*/ EIGEN_DEVICE_FUNC -inline ConstRowsBlockXpr middleRows(Index startRow, Index numRows) const +inline ConstRowsBlockXpr middleRows(Index startRow, Index n) const { - return ConstRowsBlockXpr(derived(), startRow, 0, numRows, cols()); + return ConstRowsBlockXpr(derived(), startRow, 0, n, cols()); } /** \returns a block consisting of a range of rows of *this. * - * \tparam N the number of rows in the block + * \tparam N the number of rows in the block as specified at compile-time * \param startRow the index of the first row in the block + * \param n the number of rows in the block as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include DenseBase_template_int_middleRows.cpp * Output: \verbinclude DenseBase_template_int_middleRows.out @@ -517,17 +529,17 @@ inline ConstRowsBlockXpr middleRows(Index startRow, Index numRows) const */ template EIGEN_DEVICE_FUNC -inline typename NRowsBlockXpr::Type middleRows(Index startRow) +inline typename NRowsBlockXpr::Type middleRows(Index startRow, Index n = N) { - return typename NRowsBlockXpr::Type(derived(), startRow, 0, N, cols()); + return typename NRowsBlockXpr::Type(derived(), startRow, 0, n, cols()); } /** This is the const version of middleRows().*/ template EIGEN_DEVICE_FUNC -inline typename ConstNRowsBlockXpr::Type middleRows(Index startRow) const +inline typename ConstNRowsBlockXpr::Type middleRows(Index startRow, Index n = N) const { - return typename ConstNRowsBlockXpr::Type(derived(), startRow, 0, N, cols()); + return typename ConstNRowsBlockXpr::Type(derived(), startRow, 0, n, cols()); } @@ -556,7 +568,11 @@ inline ConstColsBlockXpr leftCols(Index n) const /** \returns a block consisting of the left columns of *this. * - * \tparam N the number of columns in the block + * \tparam N the number of columns in the block as specified at compile-time + * \param n the number of columns in the block as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include MatrixBase_template_int_leftCols.cpp * Output: \verbinclude MatrixBase_template_int_leftCols.out @@ -565,17 +581,17 @@ inline ConstColsBlockXpr leftCols(Index n) const */ template EIGEN_DEVICE_FUNC -inline typename NColsBlockXpr::Type leftCols() +inline typename NColsBlockXpr::Type leftCols(Index n = N) { - return typename NColsBlockXpr::Type(derived(), 0, 0, rows(), N); + return typename NColsBlockXpr::Type(derived(), 0, 0, rows(), n); } /** This is the const version of leftCols().*/ template EIGEN_DEVICE_FUNC -inline typename ConstNColsBlockXpr::Type leftCols() const +inline typename ConstNColsBlockXpr::Type leftCols(Index n = N) const { - return typename ConstNColsBlockXpr::Type(derived(), 0, 0, rows(), N); + return typename ConstNColsBlockXpr::Type(derived(), 0, 0, rows(), n); } @@ -604,7 +620,11 @@ inline ConstColsBlockXpr rightCols(Index n) const /** \returns a block consisting of the right columns of *this. * - * \tparam N the number of columns in the block + * \tparam N the number of columns in the block as specified at compile-time + * \param n the number of columns in the block as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include MatrixBase_template_int_rightCols.cpp * Output: \verbinclude MatrixBase_template_int_rightCols.out @@ -613,17 +633,17 @@ inline ConstColsBlockXpr rightCols(Index n) const */ template EIGEN_DEVICE_FUNC -inline typename NColsBlockXpr::Type rightCols() +inline typename NColsBlockXpr::Type rightCols(Index n = N) { - return typename NColsBlockXpr::Type(derived(), 0, cols() - N, rows(), N); + return typename NColsBlockXpr::Type(derived(), 0, cols() - n, rows(), n); } /** This is the const version of rightCols().*/ template EIGEN_DEVICE_FUNC -inline typename ConstNColsBlockXpr::Type rightCols() const +inline typename ConstNColsBlockXpr::Type rightCols(Index n = N) const { - return typename ConstNColsBlockXpr::Type(derived(), 0, cols() - N, rows(), N); + return typename ConstNColsBlockXpr::Type(derived(), 0, cols() - n, rows(), n); } @@ -653,8 +673,12 @@ inline ConstColsBlockXpr middleCols(Index startCol, Index numCols) const /** \returns a block consisting of a range of columns of *this. * - * \tparam N the number of columns in the block + * \tparam N the number of columns in the block as specified at compile-time * \param startCol the index of the first column in the block + * \param n the number of columns in the block as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include DenseBase_template_int_middleCols.cpp * Output: \verbinclude DenseBase_template_int_middleCols.out @@ -663,17 +687,17 @@ inline ConstColsBlockXpr middleCols(Index startCol, Index numCols) const */ template EIGEN_DEVICE_FUNC -inline typename NColsBlockXpr::Type middleCols(Index startCol) +inline typename NColsBlockXpr::Type middleCols(Index startCol, Index n = N) { - return typename NColsBlockXpr::Type(derived(), 0, startCol, rows(), N); + return typename NColsBlockXpr::Type(derived(), 0, startCol, rows(), n); } /** This is the const version of middleCols().*/ template EIGEN_DEVICE_FUNC -inline typename ConstNColsBlockXpr::Type middleCols(Index startCol) const +inline typename ConstNColsBlockXpr::Type middleCols(Index startCol, Index n = N) const { - return typename ConstNColsBlockXpr::Type(derived(), 0, startCol, rows(), N); + return typename ConstNColsBlockXpr::Type(derived(), 0, startCol, rows(), n); } @@ -711,15 +735,15 @@ inline const Block block(Index startRow, In /** \returns an expression of a block in *this. * - * \tparam BlockRows number of rows in block as specified at compile time - * \tparam BlockCols number of columns in block as specified at compile time + * \tparam BlockRows number of rows in block as specified at compile-time + * \tparam BlockCols number of columns in block as specified at compile-time * \param startRow the first row in the block * \param startCol the first column in the block - * \param blockRows number of rows in block as specified at run time - * \param blockCols number of columns in block as specified at run time + * \param blockRows number of rows in block as specified at run-time + * \param blockCols number of columns in block as specified at run-time * - * This function is mainly useful for blocks where the number of rows is specified at compile time - * and the number of columns is specified at run time, or vice versa. The compile-time and run-time + * This function is mainly useful for blocks where the number of rows is specified at compile-time + * and the number of columns is specified at run-time, or vice versa. The compile-time and run-time * information should not contradict. In other words, \a blockRows should equal \a BlockRows unless * \a BlockRows is \a Dynamic, and the same for the number of columns. * @@ -786,7 +810,7 @@ inline ConstRowXpr row(Index i) const * \only_for_vectors * * \param start the first coefficient in the segment - * \param vecSize the number of coefficients in the segment + * \param n the number of coefficients in the segment * * Example: \include MatrixBase_segment_int_int.cpp * Output: \verbinclude MatrixBase_segment_int_int.out @@ -798,26 +822,26 @@ inline ConstRowXpr row(Index i) const * \sa class Block, segment(Index) */ EIGEN_DEVICE_FUNC -inline SegmentReturnType segment(Index start, Index vecSize) +inline SegmentReturnType segment(Index start, Index n) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return SegmentReturnType(derived(), start, vecSize); + return SegmentReturnType(derived(), start, n); } /** This is the const version of segment(Index,Index).*/ EIGEN_DEVICE_FUNC -inline ConstSegmentReturnType segment(Index start, Index vecSize) const +inline ConstSegmentReturnType segment(Index start, Index n) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return ConstSegmentReturnType(derived(), start, vecSize); + return ConstSegmentReturnType(derived(), start, n); } /** \returns a dynamic-size expression of the first coefficients of *this. * * \only_for_vectors * - * \param vecSize the number of coefficients in the block + * \param n the number of coefficients in the segment * * Example: \include MatrixBase_start_int.cpp * Output: \verbinclude MatrixBase_start_int.out @@ -829,26 +853,25 @@ inline ConstSegmentReturnType segment(Index start, Index vecSize) const * \sa class Block, block(Index,Index) */ EIGEN_DEVICE_FUNC -inline SegmentReturnType head(Index vecSize) +inline SegmentReturnType head(Index n) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return SegmentReturnType(derived(), 0, vecSize); + return SegmentReturnType(derived(), 0, n); } /** This is the const version of head(Index).*/ EIGEN_DEVICE_FUNC -inline ConstSegmentReturnType - head(Index vecSize) const +inline ConstSegmentReturnType head(Index n) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return ConstSegmentReturnType(derived(), 0, vecSize); + return ConstSegmentReturnType(derived(), 0, n); } /** \returns a dynamic-size expression of the last coefficients of *this. * * \only_for_vectors * - * \param vecSize the number of coefficients in the block + * \param n the number of coefficients in the segment * * Example: \include MatrixBase_end_int.cpp * Output: \verbinclude MatrixBase_end_int.out @@ -860,102 +883,113 @@ inline ConstSegmentReturnType * \sa class Block, block(Index,Index) */ EIGEN_DEVICE_FUNC -inline SegmentReturnType tail(Index vecSize) +inline SegmentReturnType tail(Index n) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return SegmentReturnType(derived(), this->size() - vecSize, vecSize); + return SegmentReturnType(derived(), this->size() - n, n); } /** This is the const version of tail(Index).*/ EIGEN_DEVICE_FUNC -inline ConstSegmentReturnType tail(Index vecSize) const +inline ConstSegmentReturnType tail(Index n) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return ConstSegmentReturnType(derived(), this->size() - vecSize, vecSize); + return ConstSegmentReturnType(derived(), this->size() - n, n); } /** \returns a fixed-size expression of a segment (i.e. a vector block) in \c *this * * \only_for_vectors * - * The template parameter \a Size is the number of coefficients in the block + * \tparam N the number of coefficients in the segment as specified at compile-time + * \param start the index of the first element in the segment + * \param n the number of coefficients in the segment as specified at compile-time * - * \param start the index of the first element of the sub-vector + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include MatrixBase_template_int_segment.cpp * Output: \verbinclude MatrixBase_template_int_segment.out * * \sa class Block */ -template +template EIGEN_DEVICE_FUNC -inline typename FixedSegmentReturnType::Type segment(Index start) +inline typename FixedSegmentReturnType::Type segment(Index start, Index n = N) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return typename FixedSegmentReturnType::Type(derived(), start); + return typename FixedSegmentReturnType::Type(derived(), start, n); } /** This is the const version of segment(Index).*/ -template +template EIGEN_DEVICE_FUNC -inline typename ConstFixedSegmentReturnType::Type segment(Index start) const +inline typename ConstFixedSegmentReturnType::Type segment(Index start, Index n = N) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return typename ConstFixedSegmentReturnType::Type(derived(), start); + return typename ConstFixedSegmentReturnType::Type(derived(), start, n); } /** \returns a fixed-size expression of the first coefficients of *this. * * \only_for_vectors * - * The template parameter \a Size is the number of coefficients in the block + * \tparam N the number of coefficients in the segment as specified at compile-time + * \param n the number of coefficients in the segment as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include MatrixBase_template_int_start.cpp * Output: \verbinclude MatrixBase_template_int_start.out * * \sa class Block */ -template +template EIGEN_DEVICE_FUNC -inline typename FixedSegmentReturnType::Type head() +inline typename FixedSegmentReturnType::Type head(Index n = N) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return typename FixedSegmentReturnType::Type(derived(), 0); + return typename FixedSegmentReturnType::Type(derived(), 0, n); } /** This is the const version of head().*/ -template +template EIGEN_DEVICE_FUNC -inline typename ConstFixedSegmentReturnType::Type head() const +inline typename ConstFixedSegmentReturnType::Type head(Index n = N) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return typename ConstFixedSegmentReturnType::Type(derived(), 0); + return typename ConstFixedSegmentReturnType::Type(derived(), 0, n); } /** \returns a fixed-size expression of the last coefficients of *this. * * \only_for_vectors * - * The template parameter \a Size is the number of coefficients in the block + * \tparam N the number of coefficients in the segment as specified at compile-time + * \param n the number of coefficients in the segment as specified at run-time + * + * The compile-time and run-time information should not contradict. In other words, + * \a n should equal \a N unless \a N is \a Dynamic. * * Example: \include MatrixBase_template_int_end.cpp * Output: \verbinclude MatrixBase_template_int_end.out * * \sa class Block */ -template +template EIGEN_DEVICE_FUNC -inline typename FixedSegmentReturnType::Type tail() +inline typename FixedSegmentReturnType::Type tail(Index n = N) { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return typename FixedSegmentReturnType::Type(derived(), size() - Size); + return typename FixedSegmentReturnType::Type(derived(), size() - n); } /** This is the const version of tail.*/ -template +template EIGEN_DEVICE_FUNC -inline typename ConstFixedSegmentReturnType::Type tail() const +inline typename ConstFixedSegmentReturnType::Type tail(Index n = N) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return typename ConstFixedSegmentReturnType::Type(derived(), size() - Size); + return typename ConstFixedSegmentReturnType::Type(derived(), size() - n); } From bbf373bbe9306c701c2488bb6ea05600cf63239e Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Sat, 21 Dec 2013 22:14:03 +0100 Subject: [PATCH 0241/1103] Applied patch from Richard JW Roberts, resolving bug #704 --- Eigen/src/QR/HouseholderQR.h | 94 +++++++++++++++++--------------- Eigen/src/QR/HouseholderQR_MKL.h | 26 +++++---- 2 files changed, 64 insertions(+), 56 deletions(-) diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index ad156396a..352dbf3f0 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -251,56 +251,62 @@ void householder_qr_inplace_unblocked(MatrixQR& mat, HCoeffs& hCoeffs, typename } /** \internal */ -template -void householder_qr_inplace_blocked(MatrixQR& mat, HCoeffs& hCoeffs, - typename MatrixQR::Index maxBlockSize=32, - typename MatrixQR::Scalar* tempData = 0) +template +struct householder_qr_inplace_blocked { - typedef typename MatrixQR::Index Index; - typedef typename MatrixQR::Scalar Scalar; - typedef Block BlockType; - - Index rows = mat.rows(); - Index cols = mat.cols(); - Index size = (std::min)(rows, cols); - - typedef Matrix TempType; - TempType tempVector; - if(tempData==0) + // This is specialized for MKL-supported Scalar types in HouseholderQR_MKL.h + static void run(MatrixQR& mat, HCoeffs& hCoeffs, + typename MatrixQR::Index maxBlockSize=32, + typename MatrixQR::Scalar* tempData = 0) { - tempVector.resize(cols); - tempData = tempVector.data(); - } + typedef typename MatrixQR::Index Index; + typedef typename MatrixQR::Scalar Scalar; + typedef Block BlockType; - Index blockSize = (std::min)(maxBlockSize,size); + Index rows = mat.rows(); + Index cols = mat.cols(); + Index size = (std::min)(rows, cols); - Index k = 0; - for (k = 0; k < size; k += blockSize) - { - Index bs = (std::min)(size-k,blockSize); // actual size of the block - Index tcols = cols - k - bs; // trailing columns - Index brows = rows-k; // rows of the block - - // partition the matrix: - // A00 | A01 | A02 - // mat = A10 | A11 | A12 - // A20 | A21 | A22 - // and performs the qr dec of [A11^T A12^T]^T - // and update [A21^T A22^T]^T using level 3 operations. - // Finally, the algorithm continue on A22 - - BlockType A11_21 = mat.block(k,k,brows,bs); - Block hCoeffsSegment = hCoeffs.segment(k,bs); - - householder_qr_inplace_unblocked(A11_21, hCoeffsSegment, tempData); - - if(tcols) + typedef Matrix TempType; + TempType tempVector; + if(tempData==0) { - BlockType A21_22 = mat.block(k,k+bs,brows,tcols); - apply_block_householder_on_the_left(A21_22,A11_21,hCoeffsSegment.adjoint()); + tempVector.resize(cols); + tempData = tempVector.data(); + } + + Index blockSize = (std::min)(maxBlockSize,size); + + Index k = 0; + for (k = 0; k < size; k += blockSize) + { + Index bs = (std::min)(size-k,blockSize); // actual size of the block + Index tcols = cols - k - bs; // trailing columns + Index brows = rows-k; // rows of the block + + // partition the matrix: + // A00 | A01 | A02 + // mat = A10 | A11 | A12 + // A20 | A21 | A22 + // and performs the qr dec of [A11^T A12^T]^T + // and update [A21^T A22^T]^T using level 3 operations. + // Finally, the algorithm continue on A22 + + BlockType A11_21 = mat.block(k,k,brows,bs); + Block hCoeffsSegment = hCoeffs.segment(k,bs); + + householder_qr_inplace_unblocked(A11_21, hCoeffsSegment, tempData); + + if(tcols) + { + BlockType A21_22 = mat.block(k,k+bs,brows,tcols); + apply_block_householder_on_the_left(A21_22,A11_21,hCoeffsSegment.adjoint()); + } } } -} +}; template struct solve_retval, Rhs> @@ -352,7 +358,7 @@ HouseholderQR& HouseholderQR::compute(const MatrixType& m_temp.resize(cols); - internal::householder_qr_inplace_blocked(m_qr, m_hCoeffs, 48, m_temp.data()); + internal::householder_qr_inplace_blocked::run(m_qr, m_hCoeffs, 48, m_temp.data()); m_isInitialized = true; return *this; diff --git a/Eigen/src/QR/HouseholderQR_MKL.h b/Eigen/src/QR/HouseholderQR_MKL.h index 5313de604..8a3a7e406 100644 --- a/Eigen/src/QR/HouseholderQR_MKL.h +++ b/Eigen/src/QR/HouseholderQR_MKL.h @@ -34,7 +34,7 @@ #ifndef EIGEN_QR_MKL_H #define EIGEN_QR_MKL_H -#include "Eigen/src/Core/util/MKL_support.h" +#include "../Core/util/MKL_support.h" namespace Eigen { @@ -44,18 +44,20 @@ namespace internal { #define EIGEN_MKL_QR_NOPIV(EIGTYPE, MKLTYPE, MKLPREFIX) \ template \ -void householder_qr_inplace_blocked(MatrixQR& mat, HCoeffs& hCoeffs, \ - typename MatrixQR::Index maxBlockSize=32, \ - EIGTYPE* tempData = 0) \ +struct householder_qr_inplace_blocked \ { \ - lapack_int m = mat.rows(); \ - lapack_int n = mat.cols(); \ - lapack_int lda = mat.outerStride(); \ - lapack_int matrix_order = (MatrixQR::IsRowMajor) ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \ - LAPACKE_##MKLPREFIX##geqrf( matrix_order, m, n, (MKLTYPE*)mat.data(), lda, (MKLTYPE*)hCoeffs.data()); \ - hCoeffs.adjointInPlace(); \ -\ -} + static void run(MatrixQR& mat, HCoeffs& hCoeffs, \ + typename MatrixQR::Index = 32, \ + typename MatrixQR::Scalar* = 0) \ + { \ + lapack_int m = (lapack_int) mat.rows(); \ + lapack_int n = (lapack_int) mat.cols(); \ + lapack_int lda = (lapack_int) mat.outerStride(); \ + lapack_int matrix_order = (MatrixQR::IsRowMajor) ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \ + LAPACKE_##MKLPREFIX##geqrf( matrix_order, m, n, (MKLTYPE*)mat.data(), lda, (MKLTYPE*)hCoeffs.data()); \ + hCoeffs.adjointInPlace(); \ + } \ +}; EIGEN_MKL_QR_NOPIV(double, double, d) EIGEN_MKL_QR_NOPIV(float, float, s) From e15cb9f4f83e67096f8f2443afadebba9efab0b4 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 11 Jan 2014 20:04:36 +0100 Subject: [PATCH 0242/1103] Make geo_hyperplane unit test more stable (bug #539) --- test/geo_hyperplane.cpp | 20 +++++++++++++++----- 1 file changed, 15 insertions(+), 5 deletions(-) diff --git a/test/geo_hyperplane.cpp b/test/geo_hyperplane.cpp index f26fc1329..befd7d483 100644 --- a/test/geo_hyperplane.cpp +++ b/test/geo_hyperplane.cpp @@ -18,10 +18,12 @@ template void hyperplane(const HyperplaneType& _plane) /* this test covers the following files: Hyperplane.h */ + using std::abs; typedef typename HyperplaneType::Index Index; const Index dim = _plane.dim(); enum { Options = HyperplaneType::Options }; typedef typename HyperplaneType::Scalar Scalar; + typedef typename HyperplaneType::RealScalar RealScalar; typedef Matrix VectorType; typedef Matrix MatrixType; @@ -42,7 +44,10 @@ template void hyperplane(const HyperplaneType& _plane) VERIFY_IS_APPROX( n1.dot(n1), Scalar(1) ); VERIFY_IS_MUCH_SMALLER_THAN( pl0.absDistance(p0), Scalar(1) ); - VERIFY_IS_APPROX( pl1.signedDistance(p1 + n1 * s0), s0 ); + if(numext::abs2(s0)>RealScalar(1e-6)) + VERIFY_IS_APPROX( pl1.signedDistance(p1 + n1 * s0), s0); + else + VERIFY_IS_MUCH_SMALLER_THAN( abs(pl1.signedDistance(p1 + n1 * s0) - s0), Scalar(1) ); VERIFY_IS_MUCH_SMALLER_THAN( pl1.signedDistance(pl1.projection(p0)), Scalar(1) ); VERIFY_IS_MUCH_SMALLER_THAN( pl1.absDistance(p1 + pl1.normal().unitOrthogonal() * s1), Scalar(1) ); @@ -52,6 +57,8 @@ template void hyperplane(const HyperplaneType& _plane) MatrixType rot = MatrixType::Random(dim,dim).householderQr().householderQ(); DiagonalMatrix scaling(VectorType::Random()); Translation translation(VectorType::Random()); + + while(scaling.diagonal().cwiseAbs().minCoeff() void hyperplane(const HyperplaneType& _plane) VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot*scaling).absDistance((rot*scaling) * p1), Scalar(1) ); pl2 = pl1; VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot*scaling*translation) - .absDistance((rot*scaling*translation) * p1), Scalar(1) ); + .absDistance((rot*scaling*translation) * p1), Scalar(1) ); pl2 = pl1; VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot*translation,Isometry) .absDistance((rot*translation) * p1), Scalar(1) ); @@ -104,12 +111,15 @@ template void lines() Vector result = line_u.intersection(line_v); // the lines should intersect at the point we called "center" - VERIFY_IS_APPROX(result, center); + if(abs(a-1) > 1e-2 && abs(v.normalized().dot(u.normalized()))<0.9) + VERIFY_IS_APPROX(result, center); // check conversions between two types of lines PLine pl(line_u); // gcc 3.3 will commit suicide if we don't name this variable - CoeffsType converted_coeffs = HLine(pl).coeffs(); - converted_coeffs *= (line_u.coeffs()[0])/(converted_coeffs[0]); + HLine line_u2(pl); + CoeffsType converted_coeffs = line_u2.coeffs(); + if(line_u2.normal().dot(line_u.normal())<0.) + converted_coeffs = -line_u2.coeffs(); VERIFY(line_u.coeffs().isApprox(converted_coeffs)); } } From 548216b7ca396dea5ecd55e177e38de0f80e8396 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 12 Jan 2014 11:09:06 +0100 Subject: [PATCH 0243/1103] QuaternionBase::slerp was documented twice and one explanation was ambiguous. --- Eigen/src/Geometry/Quaternion.h | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index a6a8df203..803001272 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -150,10 +150,6 @@ public: /** \returns the conjugated quaternion */ Quaternion conjugate() const; - /** \returns an interpolation for a constant motion between \a other and \c *this - * \a t in [0;1] - * see http://en.wikipedia.org/wiki/Slerp - */ template Quaternion slerp(const Scalar& t, const QuaternionBase& other) const; /** \returns \c true if \c *this is approximately equal to \a other, within the precision @@ -677,8 +673,13 @@ QuaternionBase::angularDistance(const QuaternionBase& oth return static_cast(2 * acos(d)); } + + /** \returns the spherical linear interpolation between the two quaternions - * \c *this and \a other at the parameter \a t + * \c *this and \a other at the parameter \a t in [0;1]. + * + * This represents an interpolation for a constant motion between \c *this and \a other, + * see also http://en.wikipedia.org/wiki/Slerp. */ template template From 4cd4be97a7165e6e45ee60aee23b9342af03c491 Mon Sep 17 00:00:00 2001 From: Anton Gladky Date: Wed, 15 Jan 2014 17:49:52 +0100 Subject: [PATCH 0244/1103] Port unsupported constrained CG to Eigen3 --- .../Eigen/src/IterativeSolvers/ConstrainedConjGrad.h | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h b/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h index 3f18beeeb..dc0093eb9 100644 --- a/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h +++ b/unsupported/Eigen/src/IterativeSolvers/ConstrainedConjGrad.h @@ -58,7 +58,9 @@ void pseudo_inverse(const CMatrix &C, CINVMatrix &CINV) Scalar rho, rho_1, alpha; d.setZero(); - CINV.startFill(); // FIXME estimate the number of non-zeros + typedef Triplet T; + std::vector tripletList; + for (Index i = 0; i < rows; ++i) { d[i] = 1.0; @@ -84,11 +86,12 @@ void pseudo_inverse(const CMatrix &C, CINVMatrix &CINV) // FIXME add a generic "prune/filter" expression for both dense and sparse object to sparse for (Index j=0; j TmpVec; From a58325ac2f7be7326be358ac51c4f0eebcf7fbf9 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Tue, 31 Dec 2013 18:06:28 +0000 Subject: [PATCH 0245/1103] Minor corrections in QR docs. --- Eigen/QR | 4 +++- Eigen/src/QR/FullPivHouseholderQR.h | 8 ++++---- 2 files changed, 7 insertions(+), 5 deletions(-) diff --git a/Eigen/QR b/Eigen/QR index ac5b02693..8c7c6162e 100644 --- a/Eigen/QR +++ b/Eigen/QR @@ -15,7 +15,9 @@ * * This module provides various QR decompositions * This module also provides some MatrixBase methods, including: - * - MatrixBase::qr(), + * - MatrixBase::householderQr() + * - MatrixBase::colPivHouseholderQr() + * - MatrixBase::fullPivHouseholderQr() * * \code * #include diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index e6ab172b3..44eaa1b1a 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -33,13 +33,13 @@ struct traits > * * \param MatrixType the type of the matrix of which we are computing the QR decomposition * - * This class performs a rank-revealing QR decomposition of a matrix \b A into matrices \b P, \b Q and \b R + * This class performs a rank-revealing QR decomposition of a matrix \b A into matrices \b P, \b P', \b Q and \b R * such that * \f[ - * \mathbf{A} \, \mathbf{P} = \mathbf{Q} \, \mathbf{R} + * \mathbf{P} \, \mathbf{A} \, \mathbf{P}' = \mathbf{Q} \, \mathbf{R} * \f] - * by using Householder transformations. Here, \b P is a permutation matrix, \b Q a unitary matrix and \b R an - * upper triangular matrix. + * by using Householder transformations. Here, \b P and \b P' are permutation matrices, \b Q a unitary matrix + * and \b R an upper triangular matrix. * * This decomposition performs a very prudent full pivoting in order to be rank-revealing and achieve optimal * numerical stability. The trade-off is that it is slower than HouseholderQR and ColPivHouseholderQR. From 2a564695f0e9391eb3a0125bd5731c17aabdb680 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Mar 2014 13:28:50 +0100 Subject: [PATCH 0246/1103] Simpler and hopefully more future-proof fix for bug #503 (aligned_allocator with c++11) --- Eigen/src/Core/util/Memory.h | 110 +++++++++-------------------------- 1 file changed, 27 insertions(+), 83 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 286e963d2..0f8ab065a 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -700,98 +700,42 @@ template class aligned_stack_memory_handler * \sa \ref TopicStlContainers. */ template -class aligned_allocator +class aligned_allocator : public std::allocator { public: - typedef size_t size_type; - typedef std::ptrdiff_t difference_type; - typedef T* pointer; - typedef const T* const_pointer; - typedef T& reference; - typedef const T& const_reference; - typedef T value_type; + typedef size_t size_type; + typedef std::ptrdiff_t difference_type; + typedef T* pointer; + typedef const T* const_pointer; + typedef T& reference; + typedef const T& const_reference; + typedef T value_type; - template - struct rebind - { - typedef aligned_allocator other; - }; + template + struct rebind + { + typedef aligned_allocator other; + }; - pointer address( reference value ) const - { - return &value; - } + aligned_allocator() : std::allocator() {} - const_pointer address( const_reference value ) const - { - return &value; - } + aligned_allocator(const aligned_allocator& other) : std::allocator(other) {} - aligned_allocator() - { - } + template + aligned_allocator(const aligned_allocator& other) : std::allocator(other) {} - aligned_allocator( const aligned_allocator& ) - { - } + ~aligned_allocator() {} - template - aligned_allocator( const aligned_allocator& ) - { - } + pointer allocate(size_type num, const void* /*hint*/ = 0) + { + internal::check_size_for_overflow(num); + return static_cast( internal::aligned_malloc(num * sizeof(T)) ); + } - ~aligned_allocator() - { - } - - size_type max_size() const - { - return (std::numeric_limits::max)(); - } - - pointer allocate( size_type num, const void* hint = 0 ) - { - EIGEN_UNUSED_VARIABLE(hint); - internal::check_size_for_overflow(num); - return static_cast( internal::aligned_malloc( num * sizeof(T) ) ); - } - - void construct( pointer p, const T& value ) - { - ::new( p ) T( value ); - } - -#if (__cplusplus >= 201103L) - template - void construct( U* u, Args&&... args) - { - ::new( static_cast(u) ) U( std::forward( args )... ); - } -#endif - - void destroy( pointer p ) - { - p->~T(); - } - -#if (__cplusplus >= 201103L) - template - void destroy( U* u ) - { - u->~U(); - } -#endif - - void deallocate( pointer p, size_type /*num*/ ) - { - internal::aligned_free( p ); - } - - bool operator!=(const aligned_allocator& ) const - { return false; } - - bool operator==(const aligned_allocator& ) const - { return true; } + void deallocate(pointer p, size_type /*num*/) + { + internal::aligned_free(p); + } }; //---------- Cache sizes ---------- From c39a3fa7a1808233ad6556e169e0c08d3bc979e1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 20 Mar 2014 10:14:26 +0100 Subject: [PATCH 0247/1103] Makes gcc to generate a pshufd instruction for pset1 --- Eigen/src/Core/arch/SSE/PacketMath.h | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index f5a3dab52..ea14111e3 100644 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -110,7 +110,20 @@ template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { re template<> EIGEN_STRONG_INLINE Packet2d pset1(const double& from) { return _mm_set_pd(from,from); } template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return _mm_set_epi32(from,from,from,from); } #else -template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { return _mm_set1_ps(from); } + +// GCC generates a shufps instruction for set1_ps instead of the more efficient pshufd instruction. +// However, with AVX, we want it to generate a vbroadcastss. +// Moreover, we cannot use intrinsics here because then gcc generates crappy code in some cases (see bug 203) +#if (defined __GNUC__) && (!defined __INTEL_COMPILER) && (!defined __clang__) && (!defined __AVX__) + template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { + Packet4f res; + asm("pshufd $0, %[a], %[b]" : [b] "=x" (res) : [a] "x" (from)); + return res; + } +#else + template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { return _mm_set_ps1(from); } +#endif + template<> EIGEN_STRONG_INLINE Packet2d pset1(const double& from) { return _mm_set1_pd(from); } template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return _mm_set1_epi32(from); } #endif From cfd3d6ce9c3d1bf5b899b8d15547a81a794a8af3 Mon Sep 17 00:00:00 2001 From: Bo Li Date: Thu, 20 Mar 2014 22:05:40 +0800 Subject: [PATCH 0248/1103] fixed a template type conversion bug in AngleAxis found by Pei Luo --- Eigen/src/Geometry/AngleAxis.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Geometry/AngleAxis.h b/Eigen/src/Geometry/AngleAxis.h index f424e6d7d..b42048c55 100644 --- a/Eigen/src/Geometry/AngleAxis.h +++ b/Eigen/src/Geometry/AngleAxis.h @@ -165,8 +165,8 @@ AngleAxis& AngleAxis::operator=(const QuaternionBase::dummy_precision()*NumTraits::dummy_precision()) { - m_angle = 0; - m_axis << 1, 0, 0; + m_angle = Scalar(0); + m_axis << Scalar(1), Scalar(0), Scalar(0); } else { From 01fd880424f0e937af7841202af67e6e4ee6fc07 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 20 Mar 2014 16:03:46 +0100 Subject: [PATCH 0249/1103] Revert previous change and introduce a new workaround regarding gcc generating a shufps instruction instead of the more efficient pshufd instruction. The trick consists in introducing a new pload1 function to be used in low level product kernels for which bug #203 does not apply. Indeed, it turned out that using inline assembly prevents gcc of doing a good job at instructtion reordering. --- Eigen/src/Core/GenericPacketMath.h | 4 ++++ Eigen/src/Core/arch/SSE/PacketMath.h | 26 ++++++++++++-------------- 2 files changed, 16 insertions(+), 14 deletions(-) mode change 100644 => 100755 Eigen/src/Core/GenericPacketMath.h mode change 100644 => 100755 Eigen/src/Core/arch/SSE/PacketMath.h diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h old mode 100644 new mode 100755 index b0469fa1e..538ab53b2 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -169,6 +169,10 @@ ploaddup(const typename unpacket_traits::type* from) { return *from; } template EIGEN_DEVICE_FUNC inline Packet pset1(const typename unpacket_traits::type& a) { return a; } +/** \internal \returns a packet with constant coefficients \a a[0], e.g.: (a[0],a[0],a[0],a[0]) */ +template EIGEN_DEVICE_FUNC inline Packet +pload1(const typename unpacket_traits::type *a) { return pset1(*a); } + /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */ template inline typename packet_traits::type plset(const Scalar& a) { return a; } diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h old mode 100644 new mode 100755 index ea14111e3..293fb83e4 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -110,24 +110,22 @@ template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { re template<> EIGEN_STRONG_INLINE Packet2d pset1(const double& from) { return _mm_set_pd(from,from); } template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return _mm_set_epi32(from,from,from,from); } #else - -// GCC generates a shufps instruction for set1_ps instead of the more efficient pshufd instruction. -// However, with AVX, we want it to generate a vbroadcastss. -// Moreover, we cannot use intrinsics here because then gcc generates crappy code in some cases (see bug 203) -#if (defined __GNUC__) && (!defined __INTEL_COMPILER) && (!defined __clang__) && (!defined __AVX__) - template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { - Packet4f res; - asm("pshufd $0, %[a], %[b]" : [b] "=x" (res) : [a] "x" (from)); - return res; - } -#else - template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { return _mm_set_ps1(from); } -#endif - +template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { return _mm_set_ps1(from); } template<> EIGEN_STRONG_INLINE Packet2d pset1(const double& from) { return _mm_set1_pd(from); } template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return _mm_set1_epi32(from); } #endif +// GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction. +// However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203) +// Using inline assembly is also not an option because then gcc fails to reorder properly the instructions. +// Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply. +// Also note that with AVX, we want it to generate a vbroadcastss. +#if (defined __GNUC__) && (!defined __INTEL_COMPILER) && (!defined __clang__) && (!defined __AVX__) +template<> EIGEN_STRONG_INLINE Packet4f pload1(const float *from) { + return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0); +} +#endif + template<> EIGEN_STRONG_INLINE Packet4f plset(const float& a) { return _mm_add_ps(pset1(a), _mm_set_ps(3,2,1,0)); } template<> EIGEN_STRONG_INLINE Packet2d plset(const double& a) { return _mm_add_pd(pset1(a),_mm_set_pd(1,0)); } template<> EIGEN_STRONG_INLINE Packet4i plset(const int& a) { return _mm_add_epi32(pset1(a),_mm_set_epi32(3,2,1,0)); } From 60cd361ebea62d973d81436250268e4fd1b86f49 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 26 Mar 2014 17:48:30 +0100 Subject: [PATCH 0250/1103] Fix bug #222. Make temporary matrix column-major independently of EIGEN_DEFAULT_TO_ROW_MAJOR --- Eigen/src/Householder/BlockHouseholder.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Householder/BlockHouseholder.h b/Eigen/src/Householder/BlockHouseholder.h index 1991c6527..60dbea5f5 100644 --- a/Eigen/src/Householder/BlockHouseholder.h +++ b/Eigen/src/Householder/BlockHouseholder.h @@ -48,7 +48,7 @@ void apply_block_householder_on_the_left(MatrixType& mat, const VectorsType& vec typedef typename MatrixType::Index Index; enum { TFactorSize = MatrixType::ColsAtCompileTime }; Index nbVecs = vectors.cols(); - Matrix T(nbVecs,nbVecs); + Matrix T(nbVecs,nbVecs); make_block_householder_triangular_factor(T, vectors, hCoeffs); const TriangularView& V(vectors); From c8c81c1e7454dd824607132c78997adee62101fd Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 2 Jan 2014 16:18:32 -0800 Subject: [PATCH 0251/1103] Improved the efficiency if the block-panel matrix multiplication code: the change reduces the pressure on the L1 cache by removing the calls to gebp_traits::unpackRhs(). Instead the packetization of the rhs blocks is done on the fly in gebp_traits::loadRhs(). This adds numerous calls to pset1 (since we're packetizing on the fly in the inner loop) but this is more than compensated by the fact that we're decreasing the memory transfers by a factor RhsPacketSize. --- .../Core/products/GeneralBlockPanelKernel.h | 76 +++++-------------- Eigen/src/Core/products/GeneralMatrixMatrix.h | 10 +-- .../products/GeneralMatrixMatrixTriangular.h | 19 ++--- .../Core/products/TriangularMatrixMatrix.h | 15 ++-- .../Core/products/TriangularSolverMatrix.h | 13 +--- 5 files changed, 38 insertions(+), 95 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 686ff84f1..ba6fad246 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -10,6 +10,7 @@ #ifndef EIGEN_GENERAL_BLOCK_PANEL_H #define EIGEN_GENERAL_BLOCK_PANEL_H + namespace Eigen { namespace internal { @@ -169,7 +170,7 @@ public: WorkSpaceFactor = nr * RhsPacketSize, LhsProgress = LhsPacketSize, - RhsProgress = RhsPacketSize + RhsProgress = 1 }; typedef typename packet_traits::type _LhsPacket; @@ -187,15 +188,9 @@ public: p = pset1(ResScalar(0)); } - EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const RhsScalar* rhs, RhsScalar* b) - { - for(DenseIndex k=0; k(&b[k*RhsPacketSize], rhs[k]); - } - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const { - dest = pload(b); + dest = pset1(*b); } EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const @@ -240,7 +235,7 @@ public: WorkSpaceFactor = nr*RhsPacketSize, LhsProgress = LhsPacketSize, - RhsProgress = RhsPacketSize + RhsProgress = 1 }; typedef typename packet_traits::type _LhsPacket; @@ -258,15 +253,9 @@ public: p = pset1(ResScalar(0)); } - EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const RhsScalar* rhs, RhsScalar* b) - { - for(DenseIndex k=0; k(&b[k*RhsPacketSize], rhs[k]); - } - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const { - dest = pload(b); + dest = pset1(*b); } EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const @@ -320,7 +309,7 @@ public: WorkSpaceFactor = Vectorizable ? 2*nr*RealPacketSize : nr, LhsProgress = ResPacketSize, - RhsProgress = Vectorizable ? 2*ResPacketSize : 1 + RhsProgress = 1 }; typedef typename packet_traits::type RealPacket; @@ -344,30 +333,15 @@ public: p.second = pset1(RealScalar(0)); } - /* Unpack the rhs coeff such that each complex coefficient is spread into - * two packects containing respectively the real and imaginary coefficient - * duplicated as many time as needed: (x+iy) => [x, ..., x] [y, ..., y] - */ - EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const Scalar* rhs, Scalar* b) + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const { - for(DenseIndex k=0; k((RealScalar*)&b[k*ResPacketSize*2+0], real(rhs[k])); - pstore1((RealScalar*)&b[k*ResPacketSize*2+ResPacketSize], imag(rhs[k])); - } - else - b[k] = rhs[k]; - } + dest = pset1(*b); } - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const { dest = *b; } - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacket& dest) const { - dest.first = pload((const RealScalar*)b); - dest.second = pload((const RealScalar*)(b+ResPacketSize)); + dest.first = pset1(real(*b)); + dest.second = pset1(imag(*b)); } // nothing special here @@ -445,7 +419,7 @@ public: WorkSpaceFactor = nr*RhsPacketSize, LhsProgress = ResPacketSize, - RhsProgress = ResPacketSize + RhsProgress = 1 }; typedef typename packet_traits::type _LhsPacket; @@ -463,15 +437,9 @@ public: p = pset1(ResScalar(0)); } - EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const RhsScalar* rhs, RhsScalar* b) - { - for(DenseIndex k=0; k(&b[k*RhsPacketSize], rhs[k]); - } - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const { - dest = pload(b); + dest = pset1(*b); } EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const @@ -529,14 +497,14 @@ struct gebp_kernel EIGEN_DONT_INLINE void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha, - Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0, RhsScalar* unpackedB=0); + Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0); }; template EIGEN_DONT_INLINE void gebp_kernel ::operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha, - Index strideA, Index strideB, Index offsetA, Index offsetB, RhsScalar* unpackedB) + Index strideA, Index strideB, Index offsetA, Index offsetB) { Traits traits; @@ -550,14 +518,9 @@ void gebp_kernel const Index peeled_mc2 = peeled_mc + (rows-peeled_mc >= LhsProgress ? LhsProgress : 0); const Index peeled_kc = (depth/4)*4; - if(unpackedB==0) - unpackedB = const_cast(blockB - strideB * nr * RhsProgress); - // loops on each micro vertical panel of rhs (depth x nr) for(Index j2=0; j2 we select a mr x nr micro block of res which is entirely // stored into mr/packet_size x nr registers. @@ -590,7 +553,7 @@ void gebp_kernel // performs "inner" product // TODO let's check wether the folowing peeled loop could not be // optimized via optimal prefetching from one loop to the other - const RhsScalar* blB = unpackedB; + const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; for(Index k=0; k do the same but with nr==1 for(Index j2=packet_cols; j20) while(info[j].sync!=k) {} - gebp(res+info[j].rhs_start*resStride, resStride, blockA, blockB+info[j].rhs_start*actual_kc, mc, actual_kc, info[j].rhs_length, alpha, -1,-1,0,0, w); + gebp(res+info[j].rhs_start*resStride, resStride, blockA, blockB+info[j].rhs_start*actual_kc, mc, actual_kc, info[j].rhs_length, alpha, -1,-1,0,0); } // Then keep going as usual with the remaining A' @@ -134,7 +132,7 @@ static void run(Index rows, Index cols, Index depth, pack_lhs(blockA, &lhs(i,k), lhsStride, actual_kc, actual_mc); // C_i += A' * B' - gebp(res+i, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1,-1,0,0, w); + gebp(res+i, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1,-1,0,0); } // Release all the sub blocks B'_j of B' for the current thread, @@ -152,11 +150,9 @@ static void run(Index rows, Index cols, Index depth, // this is the sequential version! std::size_t sizeA = kc*mc; std::size_t sizeB = kc*cols; - std::size_t sizeW = kc*Traits::WorkSpaceFactor; ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA()); ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB()); - ei_declare_aligned_stack_constructed_variable(RhsScalar, blockW, sizeW, blocking.blockW()); // For each horizontal panel of the rhs, and corresponding panel of the lhs... // (==GEMM_VAR1) @@ -182,7 +178,7 @@ static void run(Index rows, Index cols, Index depth, pack_lhs(blockA, &lhs(i2,k2), lhsStride, actual_kc, actual_mc); // Everything is packed, we can now call the block * panel kernel: - gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW); + gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0); } } } diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 5c3763909..ffa871cae 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -73,11 +73,8 @@ struct general_matrix_matrix_triangular_product Traits::nr) mc = (mc/Traits::nr)*Traits::nr; - std::size_t sizeW = kc*Traits::WorkSpaceFactor; - std::size_t sizeB = sizeW + kc*size; ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, kc*mc, 0); - ei_declare_aligned_stack_constructed_variable(RhsScalar, allocatedBlockB, sizeB, 0); - RhsScalar* blockB = allocatedBlockB + sizeW; + ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, kc*size, 0); gemm_pack_lhs pack_lhs; gemm_pack_rhs pack_rhs; @@ -103,15 +100,15 @@ struct general_matrix_matrix_triangular_product processed with gebp or skipped if (UpLo==Lower) gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, (std::min)(size,i2), alpha, - -1, -1, 0, 0, allocatedBlockB); + -1, -1, 0, 0); - sybb(res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha, allocatedBlockB); + sybb(res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha); if (UpLo==Upper) { Index j2 = i2+actual_mc; gebp(res+resStride*j2+i2, resStride, blockA, blockB+actual_kc*j2, actual_mc, actual_kc, (std::max)(Index(0), size-j2), alpha, - -1, -1, 0, 0, allocatedBlockB); + -1, -1, 0, 0); } } } @@ -136,7 +133,7 @@ struct tribb_kernel enum { BlockSize = EIGEN_PLAIN_ENUM_MAX(mr,nr) }; - void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha, RhsScalar* workspace) + void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha) { gebp_kernel gebp_kernel; Matrix buffer; @@ -150,7 +147,7 @@ struct tribb_kernel if(UpLo==Upper) gebp_kernel(res+j*resStride, resStride, blockA, actual_b, j, depth, actualBlockSize, alpha, - -1, -1, 0, 0, workspace); + -1, -1, 0, 0); // selfadjoint micro block { @@ -158,7 +155,7 @@ struct tribb_kernel buffer.setZero(); // 1 - apply the kernel on the temporary buffer gebp_kernel(buffer.data(), BlockSize, blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha, - -1, -1, 0, 0, workspace); + -1, -1, 0, 0); // 2 - triangular accumulation for(Index j1=0; j1 triangularBuffer; triangularBuffer.setZero(); @@ -187,7 +185,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix0) @@ -197,7 +195,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix() (blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc); - gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW); + gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0); } } } @@ -266,11 +264,9 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix triangularBuffer; triangularBuffer.setZero(); @@ -357,14 +353,13 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix conj; gebp_kernel gebp_kernel; @@ -158,7 +156,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix conj; gebp_kernel gebp_kernel; @@ -285,8 +281,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix0) gebp_kernel(_other+i2+startPanel*otherStride, otherStride, blockA, geb, actual_mc, actual_kc, rs, Scalar(-1), - -1, -1, 0, 0, blockW); + -1, -1, 0, 0); } } } From a6a57748ddff8ffc7e00e256385bd83423f68acb Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 5 Jan 2014 14:24:41 +0100 Subject: [PATCH 0252/1103] Fix typo --- CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 4135339b1..a4731c8c1 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -204,7 +204,7 @@ if(NOT MSVC) option(EIGEN_TEST_NEON "Enable/Disable Neon in tests/examples" OFF) if(EIGEN_TEST_NEON) - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon -mcpu=cortex-a"8) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon -mcpu=cortex-a8") message(STATUS "Enabling NEON in tests/examples") endif() From ac409f51f185a4654f9a5c38ab5d1652824836f6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 7 Jan 2014 20:17:59 +0100 Subject: [PATCH 0253/1103] Document the fact that Random and setRandom are not reentrant (so not thread-safe) --- Eigen/src/Core/Random.h | 22 ++++++++++++++++------ doc/Doxyfile.in | 2 ++ doc/TopicMultithreading.dox | 2 ++ 3 files changed, 20 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/Random.h b/Eigen/src/Core/Random.h index a0bd3039e..a4492ddb6 100644 --- a/Eigen/src/Core/Random.h +++ b/Eigen/src/Core/Random.h @@ -34,6 +34,8 @@ struct functor_traits > * The parameters \a rows and \a cols are the number of rows and of columns of * the returned matrix. Must be compatible with this MatrixBase type. * + * \not_reentrant + * * This variant is meant to be used for dynamic-size matrix types. For fixed-size types, * it is redundant to pass \a rows and \a cols as arguments, so Random() should be used * instead. @@ -45,8 +47,8 @@ struct functor_traits > * This expression has the "evaluate before nesting" flag so that it will be evaluated into * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected * behavior with expressions involving random matrices. - * - * \sa MatrixBase::setRandom(), MatrixBase::Random(Index), MatrixBase::Random() + * + * \sa DenseBase::setRandom(), DenseBase::Random(Index), DenseBase::Random() */ template inline const CwiseNullaryOp::Scalar>, Derived> @@ -64,6 +66,7 @@ DenseBase::Random(Index rows, Index cols) * Must be compatible with this MatrixBase type. * * \only_for_vectors + * \not_reentrant * * This variant is meant to be used for dynamic-size vector types. For fixed-size types, * it is redundant to pass \a size as argument, so Random() should be used @@ -76,7 +79,7 @@ DenseBase::Random(Index rows, Index cols) * a temporary vector whenever it is nested in a larger expression. This prevents unexpected * behavior with expressions involving random matrices. * - * \sa MatrixBase::setRandom(), MatrixBase::Random(Index,Index), MatrixBase::Random() + * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random() */ template inline const CwiseNullaryOp::Scalar>, Derived> @@ -99,8 +102,10 @@ DenseBase::Random(Index size) * This expression has the "evaluate before nesting" flag so that it will be evaluated into * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected * behavior with expressions involving random matrices. + * + * \not_reentrant * - * \sa MatrixBase::setRandom(), MatrixBase::Random(Index,Index), MatrixBase::Random(Index) + * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random(Index) */ template inline const CwiseNullaryOp::Scalar>, Derived> @@ -114,6 +119,8 @@ DenseBase::Random() * Numbers are uniformly spread through their whole definition range for integer types, * and in the [-1:1] range for floating point scalar types. * + * \not_reentrant + * * Example: \include MatrixBase_setRandom.cpp * Output: \verbinclude MatrixBase_setRandom.out * @@ -131,11 +138,12 @@ inline Derived& DenseBase::setRandom() * and in the [-1:1] range for floating point scalar types. * * \only_for_vectors + * \not_reentrant * * Example: \include Matrix_setRandom_int.cpp * Output: \verbinclude Matrix_setRandom_int.out * - * \sa MatrixBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, MatrixBase::Random() + * \sa DenseBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, DenseBase::Random() */ template EIGEN_STRONG_INLINE Derived& @@ -150,13 +158,15 @@ PlainObjectBase::setRandom(Index newSize) * Numbers are uniformly spread through their whole definition range for integer types, * and in the [-1:1] range for floating point scalar types. * + * \not_reentrant + * * \param nbRows the new number of rows * \param nbCols the new number of columns * * Example: \include Matrix_setRandom_int_int.cpp * Output: \verbinclude Matrix_setRandom_int_int.out * - * \sa MatrixBase::setRandom(), setRandom(Index), class CwiseNullaryOp, MatrixBase::Random() + * \sa DenseBase::setRandom(), setRandom(Index), class CwiseNullaryOp, DenseBase::Random() */ template EIGEN_STRONG_INLINE Derived& diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in index 1a2603b04..85af9f1d4 100644 --- a/doc/Doxyfile.in +++ b/doc/Doxyfile.in @@ -206,6 +206,7 @@ TAB_SIZE = 8 # You can put \n's in the value part of an alias to insert newlines. ALIASES = "only_for_vectors=This is only for vectors (either row-vectors or column-vectors), i.e. matrices which are known at compile-time to have either one row or one column." \ + "not_reentrant=\warning This function is not re-entrant." \ "array_module=This is defined in the %Array module. \code #include \endcode" \ "cholesky_module=This is defined in the %Cholesky module. \code #include \endcode" \ "eigenvalues_module=This is defined in the %Eigenvalues module. \code #include \endcode" \ @@ -223,6 +224,7 @@ ALIASES = "only_for_vectors=This is only for vectors (either row- "note_about_checking_solutions=This method just tries to find as good a solution as possible. If you want to check whether a solution exists or if it is accurate, just call this function to get a result and then compute the error of this result, or use MatrixBase::isApprox() directly, for instance like this: \code bool a_solution_exists = (A*result).isApprox(b, precision); \endcode This method avoids dividing by zero, so that the non-existence of a solution doesn't by itself mean that you'll get \c inf or \c nan values." \ "note_try_to_help_rvo=This function returns the result by value. In order to make that efficient, it is implemented as just a return statement using a special constructor, hopefully allowing the compiler to perform a RVO (return value optimization)." \ "nonstableyet=\warning This is not considered to be part of the stable public API yet. Changes may happen in future releases. See \ref Experimental \"Experimental parts of Eigen\" + ALIASES += "eigenAutoToc= " ALIASES += "eigenManualPage=\defgroup" diff --git a/doc/TopicMultithreading.dox b/doc/TopicMultithreading.dox index f7d082668..fb944af29 100644 --- a/doc/TopicMultithreading.dox +++ b/doc/TopicMultithreading.dox @@ -39,6 +39,8 @@ int main(int argc, char** argv) } \endcode +\warning note that all functions generating random matrices are \b not re-entrant nor thread-safe. Those include DenseBase::Random(), and DenseBase::setRandom() despite a call to Eigen::initParallel(). This is because these functions are based on std::rand which is not re-entrant. For thread-safe random generator, we recommend the use of boost::random of c++11 random feature. + In the case your application is parallelized with OpenMP, you might want to disable Eigen's own parallization as detailed in the previous section. */ From 92190a1cafc3a8613e1200cd4ed81dbc509a93bd Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 7 Jan 2014 20:23:35 +0100 Subject: [PATCH 0254/1103] Add an example showing how to use C++11 random distributions --- Eigen/src/Core/CwiseNullaryOp.h | 3 +++ Eigen/src/Core/Random.h | 2 ++ doc/special_examples/CMakeLists.txt | 19 +++++++++++++++++++ doc/special_examples/random_cpp11.cpp | 14 ++++++++++++++ 4 files changed, 38 insertions(+) create mode 100644 doc/special_examples/random_cpp11.cpp diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index 1d4ee50a8..124383114 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -138,6 +138,9 @@ DenseBase::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& f * * The template parameter \a CustomNullaryOp is the type of the functor. * + * Here is an example with C++11 random generators: \include random_cpp11.cpp + * Output: \verbinclude random_cpp11.out + * * \sa class CwiseNullaryOp */ template diff --git a/Eigen/src/Core/Random.h b/Eigen/src/Core/Random.h index a4492ddb6..fdd43ed0c 100644 --- a/Eigen/src/Core/Random.h +++ b/Eigen/src/Core/Random.h @@ -48,6 +48,8 @@ struct functor_traits > * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected * behavior with expressions involving random matrices. * + * See DenseBase::NullaryExpr(Index, const CustomNullaryOp&) for an example using C++11 random generators. + * * \sa DenseBase::setRandom(), DenseBase::Random(Index), DenseBase::Random() */ template diff --git a/doc/special_examples/CMakeLists.txt b/doc/special_examples/CMakeLists.txt index 0c9b3c3ba..45e2339e3 100644 --- a/doc/special_examples/CMakeLists.txt +++ b/doc/special_examples/CMakeLists.txt @@ -19,3 +19,22 @@ if(QT4_FOUND) add_dependencies(all_examples Tutorial_sparse_example) endif(QT4_FOUND) + +check_cxx_compiler_flag("-std=c++11" EIGEN_COMPILER_SUPPORT_CPP11) +if(EIGEN_COMPILER_SUPPORT_CPP11) + add_executable(random_cpp11 random_cpp11.cpp) + target_link_libraries(random_cpp11 ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO}) + add_dependencies(all_examples random_cpp11) + ei_add_target_property(random_cpp11 COMPILE_FLAGS "-std=c++11") + + get_target_property(random_cpp11_exec + random_cpp11 LOCATION) + add_custom_command( + TARGET random_cpp11 + POST_BUILD + COMMAND ${random_cpp11_exec} + ARGS >${CMAKE_CURRENT_BINARY_DIR}/random_cpp11.out + ) + + +endif() \ No newline at end of file diff --git a/doc/special_examples/random_cpp11.cpp b/doc/special_examples/random_cpp11.cpp new file mode 100644 index 000000000..ccd7c77d0 --- /dev/null +++ b/doc/special_examples/random_cpp11.cpp @@ -0,0 +1,14 @@ +#include +#include +#include + +using namespace Eigen; + +int main() { + std::default_random_engine generator; + std::poisson_distribution distribution(4.1); + auto poisson = [&] (int) {return distribution(generator);}; + + RowVectorXi v = RowVectorXi::NullaryExpr(10, poisson ); + std::cout << v << "\n"; +} From 3a4616d6e3d7d60af5ba569b0d36e954d45a6fb9 Mon Sep 17 00:00:00 2001 From: Martinho Fernandes Date: Fri, 10 Jan 2014 11:02:11 +0100 Subject: [PATCH 0255/1103] Add C++11 allocator overloads to avoid implicit conversions. --- Eigen/src/Core/util/Memory.h | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index d177e8b5a..c1050e1ea 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -761,11 +761,27 @@ public: ::new( p ) T( value ); } +#if (__cplusplus >= 201103L) + template + void construct( U* u, Args&&... args) + { + ::new( u ) U( std::forward( args )... ); + } +#endif + void destroy( pointer p ) { p->~T(); } +#if (__cplusplus >= 201103L) + template + void destroy( U* u ) + { + u->~U(); + } +#endif + void deallocate( pointer p, size_type /*num*/ ) { internal::aligned_free( p ); From 4ccff2d028e9e39c66c4f003c7c077a0805d3023 Mon Sep 17 00:00:00 2001 From: Martinho Fernandes Date: Fri, 10 Jan 2014 11:20:40 +0100 Subject: [PATCH 0256/1103] Placement new must use void* to avoid user-specific overloads. --- Eigen/src/Core/util/Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index c1050e1ea..d1b08ed4c 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -765,7 +765,7 @@ public: template void construct( U* u, Args&&... args) { - ::new( u ) U( std::forward( args )... ); + ::new( static_cast(u) ) U( std::forward( args )... ); } #endif From aa0db35185f7eda94eb103b6bb92630c432512e5 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sat, 18 Jan 2014 01:16:17 +0000 Subject: [PATCH 0257/1103] Add doc page on computing Least Squares. --- doc/LeastSquares.dox | 70 ++++++++++++++++++++ doc/Manual.dox | 2 + doc/TutorialLinearAlgebra.dox | 11 +-- doc/snippets/LeastSquaresNormalEquations.cpp | 4 ++ doc/snippets/LeastSquaresQR.cpp | 4 ++ 5 files changed, 86 insertions(+), 5 deletions(-) create mode 100644 doc/LeastSquares.dox create mode 100644 doc/snippets/LeastSquaresNormalEquations.cpp create mode 100644 doc/snippets/LeastSquaresQR.cpp diff --git a/doc/LeastSquares.dox b/doc/LeastSquares.dox new file mode 100644 index 000000000..e2191a22f --- /dev/null +++ b/doc/LeastSquares.dox @@ -0,0 +1,70 @@ +namespace Eigen { + +/** \eigenManualPage LeastSquares Solving linear least squares systems + +This page describes how to solve linear least squares systems using %Eigen. An overdetermined system +of equations, say \a Ax = \a b, has no solutions. In this case, it makes sense to search for the +vector \a x which is closest to being a solution, in the sense that the difference \a Ax - \a b is +as small as possible. This \a x is called the least square solution (if the Euclidean norm is used). + +The three methods discussed on this page are the SVD decomposition, the QR decomposition and normal +equations. Of these, the SVD decomposition is generally the most accurate but the slowest, normal +equations is the fastest but least accurate, and the QR decomposition is in between. + +\eigenAutoToc + + +\section LeastSquaresSVD Using the SVD decomposition + +The \link JacobiSVD::solve() solve() \endlink method in the JacobiSVD class can be directly used to +solve linear squares systems. It is not enough to compute only the singular values (the default for +this class); you also need the singular vectors but the thin SVD decomposition suffices for +computing least squares solutions: + +
Apply a \em general transformation \n to a \b normal \b vector -(explanations)\code +(explanations)\code VectorNf n1, n2; MatrixNf normalMatrix = t.linear().inverse().transpose(); n2 = (normalMatrix * n1).normalized();\endcode
+ + + + + +
Example:Output:
\include TutorialLinAlgSVDSolve.cpp \verbinclude TutorialLinAlgSVDSolve.out
+ +This is example from the page \link TutorialLinearAlgebra Linear algebra and decompositions \endlink. + + +\section LeastSquaresQR Using the QR decomposition + +The solve() method in QR decomposition classes also computes the least squares solution. There are +three QR decomposition classes: HouseholderQR (no pivoting, so fast but unstable), +ColPivHouseholderQR (column pivoting, thus a bit slower but more accurate) and FullPivHouseholderQR +(full pivoting, so slowest and most stable). Here is an example with column pivoting: + + + + + + + +
Example:Output:
\include LeastSquaresQR.cpp \verbinclude LeastSquaresQR.out
+ + +\section LeastSquaresNormalEquations Using normal equations + +Finding the least squares solution of \a Ax = \a b is equivalent to solving the normal equation +ATAx = ATb. This leads to the following code + + + + + + + +
Example:Output:
\include LeastSquaresNormalEquations.cpp \verbinclude LeastSquaresNormalEquations.out
+ +If the matrix \a A is ill-conditioned, then this is not a good method, because the condition number +of ATA is the square of the condition number of \a A. This means that you +lose twice as many digits using normal equation than if you use the other methods. + +*/ + +} \ No newline at end of file diff --git a/doc/Manual.dox b/doc/Manual.dox index 3367982ca..0d3787675 100644 --- a/doc/Manual.dox +++ b/doc/Manual.dox @@ -96,6 +96,8 @@ namespace Eigen { \ingroup DenseLinearSolvers_chapter */ /** \addtogroup TopicLinearAlgebraDecompositions \ingroup DenseLinearSolvers_chapter */ +/** \addtogroup LeastSquares + \ingroup DenseLinearSolvers_chapter */ /** \addtogroup DenseLinearSolvers_Reference \ingroup DenseLinearSolvers_chapter */ diff --git a/doc/TutorialLinearAlgebra.dox b/doc/TutorialLinearAlgebra.dox index b09f3543e..e6c41fd70 100644 --- a/doc/TutorialLinearAlgebra.dox +++ b/doc/TutorialLinearAlgebra.dox @@ -167,8 +167,8 @@ Here is an example: \section TutorialLinAlgLeastsquares Least squares solving -The best way to do least squares solving is with a SVD decomposition. Eigen provides one as the JacobiSVD class, and its solve() -is doing least-squares solving. +The most accurate method to do least squares solving is with a SVD decomposition. Eigen provides one +as the JacobiSVD class, and its solve() is doing least-squares solving. Here is an example: @@ -179,9 +179,10 @@ Here is an example:
-Another way, potentially faster but less reliable, is to use a LDLT decomposition -of the normal matrix. In any case, just read any reference text on least squares, and it will be very easy for you -to implement any linear least squares computation on top of Eigen. +Another methods, potentially faster but less reliable, are to use a Cholesky decomposition of the +normal matrix or a QR decomposition. Our page on \link LeastSquares least squares solving \endlink +has more details. + \section TutorialLinAlgSeparateComputation Separating the computation from the construction diff --git a/doc/snippets/LeastSquaresNormalEquations.cpp b/doc/snippets/LeastSquaresNormalEquations.cpp new file mode 100644 index 000000000..997cf1715 --- /dev/null +++ b/doc/snippets/LeastSquaresNormalEquations.cpp @@ -0,0 +1,4 @@ +MatrixXf A = MatrixXf::Random(3, 2); +VectorXf b = VectorXf::Random(3); +cout << "The solution using normal equations is:\n" + << (A.transpose() * A).ldlt().solve(A.transpose() * b) << endl; diff --git a/doc/snippets/LeastSquaresQR.cpp b/doc/snippets/LeastSquaresQR.cpp new file mode 100644 index 000000000..6c9704547 --- /dev/null +++ b/doc/snippets/LeastSquaresQR.cpp @@ -0,0 +1,4 @@ +MatrixXf A = MatrixXf::Random(3, 2); +VectorXf b = VectorXf::Random(3); +cout << "The solution using the QR decomposition is:\n" + << A.colPivHouseholderQr().solve(b) << endl; From 66f1c56aabc4ec4789405d11f544806312d49cd6 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Sun, 19 Jan 2014 03:04:51 +0100 Subject: [PATCH 0258/1103] sparse_solve_retval_base::defaultEvalTo created extremely oversized temporary matrices in some cases --- Eigen/src/misc/SparseSolve.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/Eigen/src/misc/SparseSolve.h b/Eigen/src/misc/SparseSolve.h index 244bb8ec7..05caa9266 100644 --- a/Eigen/src/misc/SparseSolve.h +++ b/Eigen/src/misc/SparseSolve.h @@ -54,8 +54,10 @@ template struct sparse_solve_retval_b static const int NbColsAtOnce = 4; int rhsCols = m_rhs.cols(); int size = m_rhs.rows(); - Eigen::Matrix tmp(size,rhsCols); - Eigen::Matrix tmpX(size,rhsCols); + // the temporary matrices do not need more columns than NbColsAtOnce: + int tmpCols = (std::min)(rhsCols, NbColsAtOnce); + Eigen::Matrix tmp(size,tmpCols); + Eigen::Matrix tmpX(size,tmpCols); for(int k=0; k(rhsCols-k, NbColsAtOnce); From deab937d4526c4542d13aa2654e8d898aab9a28b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 24 Jan 2014 12:50:29 +0100 Subject: [PATCH 0259/1103] NVCC: fix closed-form eigenvalue decomposition, workaround gcc4.7/nvcc5.5 issue --- Eigen/src/Core/Diagonal.h | 7 ++++- .../src/Eigenvalues/SelfAdjointEigenSolver.h | 6 ++-- test/cuda_basic.cu | 31 ++++++++++++++++--- 3 files changed, 36 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index 4436c6a69..b160479ab 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -77,7 +77,12 @@ template class Diagonal EIGEN_DEVICE_FUNC inline Index rows() const - { return m_index.value()<0 ? (std::min)(m_matrix.cols(),m_matrix.rows()+m_index.value()) : (std::min)(m_matrix.rows(),m_matrix.cols()-m_index.value()); } + { + EIGEN_USING_STD_MATH(min); + return m_index.value()<0 ? (min)(Index(m_matrix.cols()),Index(m_matrix.rows()+m_index.value())) + : (min)(Index(m_matrix.rows()),Index(m_matrix.cols()-m_index.value())); + + } EIGEN_DEVICE_FUNC inline Index cols() const { return 1; } diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index 0c3425069..b8146d04d 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -756,7 +756,9 @@ struct direct_selfadjoint_eigenvalues EIGEN_DEVICE_FUNC static inline void run(SolverType& solver, const MatrixType& mat, int options) { - using std::sqrt; + EIGEN_USING_STD_MATH(max) + EIGEN_USING_STD_MATH(sqrt); + eigen_assert(mat.cols() == 2 && mat.cols() == mat.rows()); eigen_assert((options&~(EigVecMask|GenEigMask))==0 && (options&EigVecMask)!=EigVecMask @@ -768,7 +770,7 @@ struct direct_selfadjoint_eigenvalues // map the matrix coefficients to [-1:1] to avoid over- and underflow. Scalar scale = mat.cwiseAbs().maxCoeff(); - scale = (std::max)(scale,Scalar(1)); + scale = (max)(scale,Scalar(1)); MatrixType scaledMat = mat / scale; // Compute the eigenvalues diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu index a062947a8..aa7f7a599 100644 --- a/test/cuda_basic.cu +++ b/test/cuda_basic.cu @@ -1,8 +1,16 @@ +// workaround issue between gcc >= 4.7 and cuda 5.5 +#if (defined __GNUC__) && (__GNUC__>4 || __GNUC_MINOR__>=7) + #undef _GLIBCXX_ATOMIC_BUILTINS + #undef _GLIBCXX_USE_INT128 +#endif + #define EIGEN_TEST_NO_LONGDOUBLE #define EIGEN_TEST_NO_COMPLEX #define EIGEN_TEST_FUNC cuda_basic +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int + #include "main.h" #include "cuda_common.h" @@ -70,6 +78,17 @@ struct prod { } }; +template +struct diagonal { + EIGEN_DEVICE_FUNC + void operator()(int i, const typename T1::Scalar* in, typename T1::Scalar* out) const + { + using namespace Eigen; + T1 x1(in+i); + Map res(out+i*T2::MaxSizeAtCompileTime); + res += x1.diagonal(); + } +}; template struct eigenvalues { @@ -82,12 +101,11 @@ struct eigenvalues { Map res(out+i*Vec::MaxSizeAtCompileTime); T A = M*M.adjoint(); SelfAdjointEigenSolver eig; - eig.computeDirect(A); - res = A.eigenvalues(); + eig.computeDirect(M); + res = eig.eigenvalues(); } }; - void test_cuda_basic() { ei_test_init_cuda(); @@ -110,7 +128,10 @@ void test_cuda_basic() CALL_SUBTEST( run_and_compare_to_cuda(prod(), nthreads, in, out) ); CALL_SUBTEST( run_and_compare_to_cuda(prod(), nthreads, in, out) ); -// CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); -// CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(diagonal(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_c(), nthreads, in, out) ); + + CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); } From 40c42d9788a66ac355c4206612e386229f7fc581 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 24 Jan 2014 12:51:05 +0100 Subject: [PATCH 0260/1103] NVCC: no need to enforce host compiler --- test/CMakeLists.txt | 1 - 1 file changed, 1 deletion(-) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 3c32f4b21..e1bff179d 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -302,7 +302,6 @@ find_package(CUDA) if(CUDA_FOUND) set(CUDA_PROPAGATE_HOST_FLAGS OFF) - set(CUDA_HOST_COMPILER ${CMAKE_CXX_COMPILER}) cuda_include_directories(${CMAKE_CURRENT_BINARY_DIR}) set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu") From afcfb560a2b92ad8b73f03fff80e691050fd5adc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 24 Jan 2014 12:51:33 +0100 Subject: [PATCH 0261/1103] NVCC: add more debug info --- test/cuda_common.h | 3 +++ 1 file changed, 3 insertions(+) diff --git a/test/cuda_common.h b/test/cuda_common.h index 0dd870b9c..9737693ac 100644 --- a/test/cuda_common.h +++ b/test/cuda_common.h @@ -86,12 +86,15 @@ void ei_test_init_cuda() cudaDeviceProp deviceProp; cudaGetDeviceProperties(&deviceProp, device); std::cout << "CUDA device info:\n"; + std::cout << " name: " << deviceProp.name << "\n"; std::cout << " capability: " << deviceProp.major << "." << deviceProp.minor << "\n"; std::cout << " multiProcessorCount: " << deviceProp.multiProcessorCount << "\n"; std::cout << " maxThreadsPerMultiProcessor: " << deviceProp.maxThreadsPerMultiProcessor << "\n"; + std::cout << " warpSize: " << deviceProp.warpSize << "\n"; std::cout << " regsPerBlock: " << deviceProp.regsPerBlock << "\n"; std::cout << " concurrentKernels: " << deviceProp.concurrentKernels << "\n"; std::cout << " clockRate: " << deviceProp.clockRate << "\n"; + std::cout << " canMapHostMemory: " << deviceProp.canMapHostMemory << "\n"; std::cout << " computeMode: " << deviceProp.computeMode << "\n"; } From 6cf938df53b5f26690a0e8034e10d11640d3a5c2 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 24 Jan 2014 13:24:30 +0100 Subject: [PATCH 0262/1103] Add a minimalistic page on CUDA with Eigen. --- doc/Manual.dox | 1 + doc/UsingNVCC.dox | 32 ++++++++++++++++++++++++++++++++ 2 files changed, 33 insertions(+) create mode 100644 doc/UsingNVCC.dox diff --git a/doc/Manual.dox b/doc/Manual.dox index 0d3787675..0cbe792d5 100644 --- a/doc/Manual.dox +++ b/doc/Manual.dox @@ -11,6 +11,7 @@ namespace Eigen { - \subpage TopicCustomizingEigen - \subpage TopicMultiThreading - \subpage TopicUsingIntelMKL + - \subpage TopicCUDA - \subpage TopicTemplateKeyword - \subpage UserManual_UnderstandingEigen */ diff --git a/doc/UsingNVCC.dox b/doc/UsingNVCC.dox new file mode 100644 index 000000000..e9df5de04 --- /dev/null +++ b/doc/UsingNVCC.dox @@ -0,0 +1,32 @@ + +namespace Eigen { + +/** \page TopicCUDA Using Eigen in CUDA kernels + +\b Disclaimer: this page is about an \b experimental feature in %Eigen. + +Staring from CUDA 5.0, the CUDA compiler, \c nvcc, is able to properly parse %Eigen's code (almost). +A few adaptations of the %Eigen's code already allows to use some parts of %Eigen in your own CUDA kernels. +To this end you need the devel branch of %Eigen, CUDA 5.0 or greater with GCC. + +Known issues: + + - \c nvcc with MS Visual Studio does not work (patch welcome) + + - \c nvcc with \c clang does not work (patch welcome) + + - \c nvcc 5.5 with gcc-4.7 (or greater) has issues with the standard \c header file. To workaround this, you can add the following before including any other files: + \code + // workaround issue between gcc >= 4.7 and cuda 5.5 + #if (defined __GNUC__) && (__GNUC__>4 || __GNUC_MINOR__>=7) + #undef _GLIBCXX_ATOMIC_BUILTINS + #undef _GLIBCXX_USE_INT128 + #endif + \endcode + + - On 64bits system Eigen uses \c long \c int as the default type for indexes and sizes. On CUDA device, it would make sense to default to 32 bits \c int. + However, to keep host and CUDA code compatible, this cannot be done automatically by %Eigen, and the user is thus required to define \c EIGEN_DEFAULT_DENSE_INDEX_TYPE to \c int throughout his code (or only for CUDA code if there is no interaction between host and CUDA code through %Eigen's object). + +*/ + +} From a7621809fe0fc7c65446ab9a83739ebe313004d4 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 25 Jan 2014 16:54:13 +0100 Subject: [PATCH 0263/1103] Remove useless register keyword, and optimize predux_min/max for SSE4 --- Eigen/src/Core/arch/SSE/PacketMath.h | 18 ++++++++++++++---- .../Core/products/SelfadjointMatrixVector.h | 6 +++--- 2 files changed, 17 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index f85d2e06e..f5a3dab52 100644 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -504,13 +504,18 @@ template<> EIGEN_STRONG_INLINE double predux_min(const Packet2d& a) } template<> EIGEN_STRONG_INLINE int predux_min(const Packet4i& a) { +#ifdef EIGEN_VECTORIZE_SSE4_1 + Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); + return pfirst(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); +#else // after some experiments, it is seems this is the fastest way to implement it // for GCC (eg., it does not like using std::min after the pstore !!) EIGEN_ALIGN16 int aux[4]; pstore(aux, a); - register int aux0 = aux[0] EIGEN_STRONG_INLINE double predux_max(const Packet2d& a) } template<> EIGEN_STRONG_INLINE int predux_max(const Packet4i& a) { +#ifdef EIGEN_VECTORIZE_SSE4_1 + Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); + return pfirst(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); +#else // after some experiments, it is seems this is the fastest way to implement it // for GCC (eg., it does not like using std::min after the pstore !!) EIGEN_ALIGN16 int aux[4]; pstore(aux, a); - register int aux0 = aux[0]>aux[1] ? aux[0] : aux[1]; - register int aux2 = aux[2]>aux[3] ? aux[2] : aux[3]; + int aux0 = aux[0]>aux[1] ? aux[0] : aux[1]; + int aux2 = aux[2]>aux[3] ? aux[2] : aux[3]; return aux0>aux2 ? aux0 : aux2; +#endif // EIGEN_VECTORIZE_SSE4_1 } #if (defined __GNUC__) diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index c40e80f53..f698f67f9 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -79,8 +79,8 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product(t0); @@ -147,7 +147,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product Date: Sat, 25 Jan 2014 19:06:07 +0100 Subject: [PATCH 0264/1103] fix scalar * prod in evaluators unit test --- Eigen/src/Core/ProductEvaluators.h | 3 ++- Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h | 2 +- 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 46048882b..2279ec33b 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -156,7 +156,8 @@ struct Assignment > SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const AssignFunc& func) { - call_assignment(dst.noalias(), (src.functor().m_other * src.nestedExpression().lhs()) * src.nestedExpression().rhs(), func); + // TODO use operator* instead of prod() once we have made enough progress + call_assignment(dst.noalias(), prod(src.functor().m_other * src.nestedExpression().lhs(), src.nestedExpression().rhs()), func); } }; diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 5c3763909..87bd36bc3 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -20,7 +20,7 @@ namespace internal { /********************************************************************** * This file implements a general A * B product while * evaluating only one triangular part of the product. -* This is more general version of self adjoint product (C += A A^T) +* This is a more general version of self adjoint product (C += A A^T) * as the level 3 SYRK Blas routine. **********************************************************************/ From 5fa7262e4c9fdbac2e5ddc4e2bb2241e04a10bef Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 25 Jan 2014 23:02:14 +0100 Subject: [PATCH 0265/1103] Refactor triangular assignment --- Eigen/src/Core/AssignEvaluator.h | 13 +- Eigen/src/Core/SelfAdjointView.h | 5 + Eigen/src/Core/Swap.h | 1 + Eigen/src/Core/TriangularMatrix.h | 301 ++++++++++++++++-------------- 4 files changed, 167 insertions(+), 153 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 6d3739407..86cd1cd68 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -140,6 +140,7 @@ public: template struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling { + // FIXME: this is not very clean, perhaps this information should be provided by the kernel? typedef typename Kernel::DstEvaluatorType DstEvaluatorType; typedef typename DstEvaluatorType::XprType DstXprType; @@ -204,9 +205,10 @@ struct copy_using_evaluator_LinearTraversal_CompleteUnrolling struct copy_using_evaluator_innervec_CompleteUnrolling { + // FIXME: this is not very clean, perhaps this information should be provided by the kernel? typedef typename Kernel::DstEvaluatorType DstEvaluatorType; typedef typename DstEvaluatorType::XprType DstXprType; - + enum { outer = Index / DstXprType::InnerSizeAtCompileTime, inner = Index % DstXprType::InnerSizeAtCompileTime, @@ -233,8 +235,7 @@ struct copy_using_evaluator_innervec_InnerUnrolling static EIGEN_STRONG_INLINE void run(Kernel &kernel, int outer) { kernel.template assignPacketByOuterInner(outer, Index); - typedef typename Kernel::DstEvaluatorType::XprType DstXprType; - enum { NextIndex = Index + packet_traits::size }; + enum { NextIndex = Index + packet_traits::size }; copy_using_evaluator_innervec_InnerUnrolling::run(kernel, outer); } }; @@ -542,12 +543,6 @@ public: m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col)); } - /// This overload by-passes the source expression, i.e., dst(row,col) ?= value - void assignCoeff(Index row, Index col, const Scalar &value) - { - m_functor.assignCoeff(m_dst.coeffRef(row,col), value); - } - /// \sa assignCoeff(Index,Index) void assignCoeff(Index index) { diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index a5c6d5ceb..f34de8b39 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -244,6 +244,8 @@ template class SelfAdjointView namespace internal { +#ifndef EIGEN_TEST_EVALUATORS + template struct triangular_assignment_selector { @@ -336,6 +338,8 @@ struct triangular_assignment_selector // in the future selfadjoint-ness should be defined by the expression traits @@ -348,6 +352,7 @@ struct evaluator_traits > static const int AssumeAliasing = 0; }; + #endif // EIGEN_ENABLE_EVALUATORS } // end namespace internal diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index 117f667f6..e7d525572 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -148,6 +148,7 @@ template class generic_dense_assignment_kernel, Specialized> : public generic_dense_assignment_kernel, BuiltIn> { +protected: typedef generic_dense_assignment_kernel, BuiltIn> Base; typedef typename DstEvaluatorTypeT::PacketScalar PacketScalar; using Base::m_dst; diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 28191694d..c658e2290 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -298,8 +298,8 @@ template class TriangularView template EIGEN_DEVICE_FUNC - void lazyAssign(const MatrixBase& other); - + void lazyAssign(const MatrixBase& other); + /** \sa MatrixBase::conjugate() */ EIGEN_DEVICE_FUNC inline TriangularView conjugate() @@ -469,6 +469,8 @@ template class TriangularView return m_matrix.diagonal().prod(); } +#ifndef EIGEN_TEST_EVALUATORS + // TODO simplify the following: template EIGEN_DEVICE_FUNC @@ -514,7 +516,9 @@ template class TriangularView { return assignProduct(other,-other.alpha()); } - + +#endif // EIGEN_TEST_EVALUATORS + protected: template @@ -530,6 +534,8 @@ template class TriangularView namespace internal { +#ifndef EIGEN_TEST_EVALUATORS + template struct triangular_assignment_selector { @@ -694,8 +700,52 @@ struct triangular_assignment_selector +template +inline TriangularView& +TriangularView::operator=(const MatrixBase& other) +{ + internal::call_assignment(*this, other.template triangularView(), internal::assign_op()); + return *this; +} + +// FIXME should we keep that possibility +template +template +void TriangularView::lazyAssign(const MatrixBase& other) +{ + internal::call_assignment(this->noalias(), other.template triangularView()); +} + + + +template +template +inline TriangularView& +TriangularView::operator=(const TriangularBase& other) +{ + eigen_assert(Mode == int(OtherDerived::Mode)); + internal::call_assignment(*this, other.derived()); + return *this; +} + +template +template +void TriangularView::lazyAssign(const TriangularBase& other) +{ + eigen_assert(Mode == int(OtherDerived::Mode)); + internal::call_assignment(this->noalias(), other.derived()); +} + +#else + // FIXME should we keep that possibility template template @@ -770,6 +820,8 @@ void TriangularView::lazyAssign(const TriangularBase::run(m_matrix.const_cast_derived(), other.derived().nestedExpression()); } +#endif // EIGEN_TEST_EVALUATORS + /*************************************************************************** * Implementation of TriangularBase methods ***************************************************************************/ @@ -790,6 +842,8 @@ void TriangularBase::evalTo(MatrixBase &other) const evalToLazy(other.derived()); } +#ifndef EIGEN_TEST_EVALUATORS + /** Assigns a triangular or selfadjoint matrix to a dense matrix. * If the matrix is triangular, the opposite part is set to zero. */ template @@ -811,6 +865,8 @@ void TriangularBase::evalToLazy(MatrixBase &other) const >::run(other.derived(), derived().nestedExpression()); } +#endif // EIGEN_TEST_EVALUATORS + /*************************************************************************** * Implementation of TriangularView methods ***************************************************************************/ @@ -962,15 +1018,15 @@ struct evaluator_traits > template struct evaluator, Kind, typename MatrixType::Scalar> - : evaluator + : evaluator::type> { typedef TriangularView XprType; - typedef evaluator Base; + typedef evaluator::type> Base; typedef evaluator type; evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {} }; -// Additional assignement kinds: +// Additional assignment kinds: struct Triangular2Triangular {}; struct Triangular2Dense {}; struct Dense2Triangular {}; @@ -978,6 +1034,59 @@ struct Dense2Triangular {}; template struct triangular_assignment_loop; + +// Specialization of the dense assignment kernel for triangular matrices. +// The main additions are: +// - An overload of assignCoeff(i, j, val) that by-passes the source expression. Needed to set the opposite triangular part to 0. +// - When calling assignCoeff(i,j...), we guarantee that i!=j +// - New assignDiagonalCoeff(i), and assignDiagonalCoeff(i, val) for assigning coefficients on the diagonal. +template +class triangular_dense_assignment_kernel : public generic_dense_assignment_kernel +{ +protected: + typedef generic_dense_assignment_kernel Base; + typedef typename Base::DstXprType DstXprType; + typedef typename Base::SrcXprType SrcXprType; + using Base::m_dst; + using Base::m_src; + using Base::m_functor; +public: + + typedef typename Base::DstEvaluatorType DstEvaluatorType; + typedef typename Base::SrcEvaluatorType SrcEvaluatorType; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + typedef typename Base::AssignmentTraits AssignmentTraits; + + + triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + : Base(dst, src, func, dstExpr) + {} + +#ifdef EIGEN_INTERNAL_DEBUGGING + void assignCoeff(Index row, Index col) + { + eigen_internal_assert(row!=col); + Base::assignCoeff(row,col); + } +#else + using Base::assignCoeff; +#endif + + void assignDiagonalCoeff(Index id) + { + if((Mode&UnitDiag) && ClearOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(1)); + else if((Mode&ZeroDiag) && ClearOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(0)); + else if((Mode&(UnitDiag|ZeroDiag))==0) Base::assignCoeff(id,id); + } + + void assignOppositeCoeff(Index row, Index col) + { + eigen_internal_assert(row!=col); + if(ClearOpposite) + m_functor.assignCoeff(m_dst.coeffRef(row,col), Scalar(0)); + } +}; template void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) @@ -994,13 +1103,13 @@ void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& sr DstEvaluatorType dstEvaluator(dst); SrcEvaluatorType srcEvaluator(src); - typedef generic_dense_assignment_kernel Kernel; + typedef triangular_dense_assignment_kernel Kernel; Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived()); enum { unroll = DstXprType::SizeAtCompileTime != Dynamic - && internal::traits::CoeffReadCost != Dynamic - && DstXprType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT +// && internal::traits::CoeffReadCost != Dynamic +// && DstXprType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT }; triangular_assignment_loop::run(kernel); @@ -1050,9 +1159,13 @@ struct Assignment template struct triangular_assignment_loop { + // FIXME: this is not very clean, perhaps this information should be provided by the kernel? + typedef typename Kernel::DstEvaluatorType DstEvaluatorType; + typedef typename DstEvaluatorType::XprType DstXprType; + enum { - col = (UnrollCount-1) / Kernel::DstEvaluatorType::RowsAtCompileTime, - row = (UnrollCount-1) % Kernel::DstEvaluatorType::RowsAtCompileTime + col = (UnrollCount-1) / DstXprType::RowsAtCompileTime, + row = (UnrollCount-1) % DstXprType::RowsAtCompileTime }; typedef typename Kernel::Scalar Scalar; @@ -1061,24 +1174,13 @@ struct triangular_assignment_loop static inline void run(Kernel &kernel) { triangular_assignment_loop::run(kernel); - - // TODO should be a static assert - eigen_assert( Mode == Upper || Mode == Lower - || Mode == StrictlyUpper || Mode == StrictlyLower - || Mode == UnitUpper || Mode == UnitLower); - if((Mode == Upper && row <= col) - || (Mode == Lower && row >= col) - || (Mode == StrictlyUpper && row < col) - || (Mode == StrictlyLower && row > col) - || (Mode == UnitUpper && row < col) - || (Mode == UnitLower && row > col)) - kernel.assignCoeff(row, col); - else if(ClearOpposite) - { - if((Mode&UnitDiag) && row==col) kernel.assignCoeff(row, col, Scalar(1)); - else kernel.assignCoeff(row, col, Scalar(0)); - } + if(row==col) + kernel.assignDiagonalCoeff(row); + else if( ((Mode&Lower) && row>col) || ((Mode&Upper) && row static inline void run(Kernel &) {} }; -// TODO: merge the following 6 variants into a single one (or max two), -// and perhaps write them using sub-expressions -// TODO: expreiment with a recursive assignment procedure splitting the current + +// TODO: experiment with a recursive assignment procedure splitting the current // triangular part into one rectangular and two triangular parts. -template -struct triangular_assignment_loop -{ - typedef typename Kernel::Index Index; - typedef typename Kernel::Scalar Scalar; - EIGEN_DEVICE_FUNC - static inline void run(Kernel &kernel) - { - for(Index j = 0; j < kernel.cols(); ++j) - { - Index maxi = (std::min)(j, kernel.rows()-1); - for(Index i = 0; i <= maxi; ++i) - kernel.assignCoeff(i, j); - if (ClearOpposite) - for(Index i = maxi+1; i < kernel.rows(); ++i) - kernel.assignCoeff(i, j, Scalar(0)); - } - } -}; -template -struct triangular_assignment_loop -{ - typedef typename Kernel::Index Index; - typedef typename Kernel::Scalar Scalar; - EIGEN_DEVICE_FUNC - static inline void run(Kernel &kernel) - { - for(Index j = 0; j < kernel.cols(); ++j) - { - for(Index i = j; i < kernel.rows(); ++i) - kernel.assignCoeff(i, j); - Index maxi = (std::min)(j, kernel.rows()); - if (ClearOpposite) - for(Index i = 0; i < maxi; ++i) - kernel.assignCoeff(i, j, Scalar(0)); - } - } -}; - -template -struct triangular_assignment_loop -{ - typedef typename Kernel::Index Index; - typedef typename Kernel::Scalar Scalar; - EIGEN_DEVICE_FUNC - static inline void run(Kernel &kernel) - { - for(Index j = 0; j < kernel.cols(); ++j) - { - Index maxi = (std::min)(j, kernel.rows()); - for(Index i = 0; i < maxi; ++i) - kernel.assignCoeff(i, j); - if (ClearOpposite) - for(Index i = maxi; i < kernel.rows(); ++i) - kernel.assignCoeff(i, j, Scalar(0)); - } - } -}; - -template -struct triangular_assignment_loop -{ - typedef typename Kernel::Index Index; - typedef typename Kernel::Scalar Scalar; - EIGEN_DEVICE_FUNC - static inline void run(Kernel &kernel) - { - for(Index j = 0; j < kernel.cols(); ++j) - { - for(Index i = j+1; i < kernel.rows(); ++i) - kernel.assignCoeff(i, j); - Index maxi = (std::min)(j, kernel.rows()-1); - if (ClearOpposite) - for(Index i = 0; i <= maxi; ++i) - kernel.assignCoeff(i, j, Scalar(0)); - } - } -}; - -template -struct triangular_assignment_loop +template +struct triangular_assignment_loop { typedef typename Kernel::Index Index; typedef typename Kernel::Scalar Scalar; @@ -1188,50 +1210,41 @@ struct triangular_assignment_loop { Index maxi = (std::min)(j, kernel.rows()); Index i = 0; - for(; i < maxi; ++i) - kernel.assignCoeff(i, j); - - if(i -struct triangular_assignment_loop -{ - typedef typename Kernel::Index Index; - typedef typename Kernel::Scalar Scalar; - EIGEN_DEVICE_FUNC - static inline void run(Kernel &kernel) - { - for(Index j = 0; j < kernel.cols(); ++j) - { - Index maxi = (std::min)(j, kernel.rows()); - Index i = 0; - if (ClearOpposite) + if (((Mode&Lower) && ClearOpposite) || (Mode&Upper)) { for(; i < maxi; ++i) - kernel.assignCoeff(i, j, Scalar(0)); + if(Mode&Upper) kernel.assignCoeff(i, j); + else kernel.assignOppositeCoeff(i, j); } + else + i = maxi; if(i +template +void TriangularBase::evalToLazy(MatrixBase &other) const +{ + other.derived().resize(this->rows(), this->cols()); + internal::call_triangular_assignment_loop(other.derived(), derived().nestedExpression()); +} + +#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen From f54e62e4a941ad5d1a5daaa0cc1cd6835cb7dd72 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 26 Jan 2014 12:18:36 +0100 Subject: [PATCH 0266/1103] Port evaluation from selfadjoint to full to evaluators --- Eigen/src/Core/SelfAdjointView.h | 41 +++++++++++++++++++++ Eigen/src/Core/TriangularMatrix.h | 61 ++++++++++++++++--------------- 2 files changed, 72 insertions(+), 30 deletions(-) diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index f34de8b39..dd8eb25f8 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -353,6 +353,47 @@ struct evaluator_traits > static const int AssumeAliasing = 0; }; +template +class triangular_dense_assignment_kernel + : public generic_dense_assignment_kernel +{ +protected: + typedef generic_dense_assignment_kernel Base; + typedef typename Base::DstXprType DstXprType; + typedef typename Base::SrcXprType SrcXprType; + using Base::m_dst; + using Base::m_src; + using Base::m_functor; +public: + + typedef typename Base::DstEvaluatorType DstEvaluatorType; + typedef typename Base::SrcEvaluatorType SrcEvaluatorType; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + typedef typename Base::AssignmentTraits AssignmentTraits; + + + triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + : Base(dst, src, func, dstExpr) + {} + + void assignCoeff(Index row, Index col) + { + eigen_internal_assert(row!=col); + Scalar tmp = m_src.coeff(row,col); + m_functor.assignCoeff(m_dst.coeffRef(row,col), tmp); + m_functor.assignCoeff(m_dst.coeffRef(col,row), numext::conj(tmp)); + } + + void assignDiagonalCoeff(Index id) + { + Base::assignCoeff(id,id); + } + + void assignOppositeCoeff(Index, Index) + { eigen_internal_assert(false && "should never be called"); } +}; + #endif // EIGEN_ENABLE_EVALUATORS } // end namespace internal diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index c658e2290..d413ec2e9 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -1035,12 +1035,12 @@ struct Dense2Triangular {}; template struct triangular_assignment_loop; -// Specialization of the dense assignment kernel for triangular matrices. -// The main additions are: -// - An overload of assignCoeff(i, j, val) that by-passes the source expression. Needed to set the opposite triangular part to 0. -// - When calling assignCoeff(i,j...), we guarantee that i!=j -// - New assignDiagonalCoeff(i), and assignDiagonalCoeff(i, val) for assigning coefficients on the diagonal. -template +/** \internal Specialization of the dense assignment kernel for triangular matrices. + * The main difference is that the triangular, diagonal, and opposite parts are processed through three different functions. + * \tparam UpLo must be either Lower or Upper + * \tparam Mode must be either 0, UnitDiag, ZeroDiag, or SelfAdjoint + */ +template class triangular_dense_assignment_kernel : public generic_dense_assignment_kernel { protected: @@ -1075,20 +1075,20 @@ public: void assignDiagonalCoeff(Index id) { - if((Mode&UnitDiag) && ClearOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(1)); - else if((Mode&ZeroDiag) && ClearOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(0)); - else if((Mode&(UnitDiag|ZeroDiag))==0) Base::assignCoeff(id,id); + if(Mode==UnitDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(1)); + else if(Mode==ZeroDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(0)); + else if(Mode==0) Base::assignCoeff(id,id); } void assignOppositeCoeff(Index row, Index col) { eigen_internal_assert(row!=col); - if(ClearOpposite) + if(SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(row,col), Scalar(0)); } }; -template +template void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) { #ifdef EIGEN_DEBUG_ASSIGN @@ -1103,22 +1103,23 @@ void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& sr DstEvaluatorType dstEvaluator(dst); SrcEvaluatorType srcEvaluator(src); - typedef triangular_dense_assignment_kernel Kernel; + typedef triangular_dense_assignment_kernel< Mode&(Lower|Upper),Mode&(UnitDiag|ZeroDiag|SelfAdjoint),SetOpposite, + DstEvaluatorType,SrcEvaluatorType,Functor> Kernel; Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived()); enum { unroll = DstXprType::SizeAtCompileTime != Dynamic -// && internal::traits::CoeffReadCost != Dynamic -// && DstXprType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT + && internal::traits::CoeffReadCost != Dynamic + && DstXprType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT }; - triangular_assignment_loop::run(kernel); + triangular_assignment_loop::run(kernel); } -template +template void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src) { - call_triangular_assignment_loop(dst, src, internal::assign_op()); + call_triangular_assignment_loop(dst, src, internal::assign_op()); } template<> struct AssignmentKind { typedef Triangular2Triangular Kind; }; @@ -1141,8 +1142,8 @@ template< typename DstXprType, typename SrcXprType, typename Functor, typename S struct Assignment { static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) - { - call_triangular_assignment_loop(dst, src, func); + { + call_triangular_assignment_loop(dst, src, func); } }; @@ -1150,13 +1151,13 @@ template< typename DstXprType, typename SrcXprType, typename Functor, typename S struct Assignment { static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) - { + { call_triangular_assignment_loop(dst, src, func); } }; -template +template struct triangular_assignment_loop { // FIXME: this is not very clean, perhaps this information should be provided by the kernel? @@ -1173,20 +1174,20 @@ struct triangular_assignment_loop EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel) { - triangular_assignment_loop::run(kernel); + triangular_assignment_loop::run(kernel); if(row==col) kernel.assignDiagonalCoeff(row); else if( ((Mode&Lower) && row>col) || ((Mode&Upper) && row -struct triangular_assignment_loop +template +struct triangular_assignment_loop { EIGEN_DEVICE_FUNC static inline void run(Kernel &) {} @@ -1198,8 +1199,8 @@ struct triangular_assignment_loop // triangular part into one rectangular and two triangular parts. -template -struct triangular_assignment_loop +template +struct triangular_assignment_loop { typedef typename Kernel::Index Index; typedef typename Kernel::Scalar Scalar; @@ -1210,7 +1211,7 @@ struct triangular_assignment_loop { Index maxi = (std::min)(j, kernel.rows()); Index i = 0; - if (((Mode&Lower) && ClearOpposite) || (Mode&Upper)) + if (((Mode&Lower) && SetOpposite) || (Mode&Upper)) { for(; i < maxi; ++i) if(Mode&Upper) kernel.assignCoeff(i, j); @@ -1222,7 +1223,7 @@ struct triangular_assignment_loop if(i void TriangularBase::evalToLazy(MatrixBase &other) const { other.derived().resize(this->rows(), this->cols()); - internal::call_triangular_assignment_loop(other.derived(), derived().nestedExpression()); + internal::call_triangular_assignment_loop(other.derived(), derived().nestedExpression()); } #endif // EIGEN_ENABLE_EVALUATORS From ee1c55f9234a5782846035028ddfc22b92ee3732 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 26 Jan 2014 14:55:25 +0100 Subject: [PATCH 0267/1103] Add missing template keyword --- Eigen/src/Core/AssignEvaluator.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 86cd1cd68..fe6fbdb9a 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -561,13 +561,13 @@ public: template void assignPacket(Index row, Index col) { - m_functor.assignPacket(&m_dst.coeffRef(row,col), m_src.template packet(row,col)); + m_functor.template assignPacket(&m_dst.coeffRef(row,col), m_src.template packet(row,col)); } template void assignPacket(Index index) { - m_functor.assignPacket(&m_dst.coeffRef(index), m_src.template packet(index)); + m_functor.template assignPacket(&m_dst.coeffRef(index), m_src.template packet(index)); } template From 34694d88285f69131eb96e9f496f35da5165c36d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 26 Jan 2014 15:34:26 +0100 Subject: [PATCH 0268/1103] Fix evaluator for fixed size objects --- Eigen/src/Core/CoreEvaluators.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index faf5aedb5..ef900e236 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -839,8 +839,8 @@ struct evaluator > protected: typename evaluator::nestedType m_argImpl; - const variable_if_dynamic m_rows; - const variable_if_dynamic m_cols; + const variable_if_dynamic m_rows; + const variable_if_dynamic m_cols; }; From 94acccc126d430bf34587527d84ff9b389219c2f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 26 Jan 2014 15:35:44 +0100 Subject: [PATCH 0269/1103] Fix Random().normalized() by introducing a nested_eval helper (recall that the old nested<> class is deprecated) --- Eigen/src/Core/Dot.h | 7 ++++++- Eigen/src/Core/util/XprHelper.h | 29 +++++++++++++++++++++++++++++ 2 files changed, 35 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h index 718de5d1a..38f6fbf44 100644 --- a/Eigen/src/Core/Dot.h +++ b/Eigen/src/Core/Dot.h @@ -141,8 +141,13 @@ template inline const typename MatrixBase::PlainObject MatrixBase::normalized() const { - typedef typename internal::nested::type Nested; +#ifndef EIGEN_TEST_EVALUATORS + typedef typename internal::nested::type Nested; typedef typename internal::remove_reference::type _Nested; +#else + typedef typename internal::nested_eval::type _Nested; +// typedef typename internal::remove_reference::type _Nested; +#endif // EIGEN_TEST_EVALUATORS _Nested n(derived()); return n / n.norm(); } diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 189928c8f..f210344d3 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -296,11 +296,40 @@ struct transfer_constness #ifdef EIGEN_TEST_EVALUATORS // When using evaluators, we never evaluate when assembling the expression!! +// TODO: get rid of this nested class since it's just an alias for ref_selector. template::type> struct nested { typedef typename ref_selector::type type; }; +// However, we still need a mechanism to detect whether an expression which is evaluated multiple time +// has to be evaluated into a temporary. +// That's the purpose of this new nested_eval helper: +template::type> struct nested_eval +{ + enum { + // For the purpose of this test, to keep it reasonably simple, we arbitrarily choose a value of Dynamic values. + // the choice of 10000 makes it larger than any practical fixed value and even most dynamic values. + // in extreme cases where these assumptions would be wrong, we would still at worst suffer performance issues + // (poor choice of temporaries). + // It's important that this value can still be squared without integer overflowing. + DynamicAsInteger = 10000, + ScalarReadCost = NumTraits::Scalar>::ReadCost, + ScalarReadCostAsInteger = ScalarReadCost == Dynamic ? int(DynamicAsInteger) : int(ScalarReadCost), + CoeffReadCost = traits::CoeffReadCost, + CoeffReadCostAsInteger = CoeffReadCost == Dynamic ? int(DynamicAsInteger) : int(CoeffReadCost), + NAsInteger = n == Dynamic ? int(DynamicAsInteger) : n, + CostEvalAsInteger = (NAsInteger+1) * ScalarReadCostAsInteger + CoeffReadCostAsInteger, + CostNoEvalAsInteger = NAsInteger * CoeffReadCostAsInteger + }; + + typedef typename conditional< + int(CostEvalAsInteger) < int(CostNoEvalAsInteger), + PlainObject, + typename ref_selector::type + >::type type; +}; + #else /** \internal Determines how a given expression should be nested into another one. * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be From 64a85800bd3573a7da7a396fde9707dce87a58d9 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 29 Jan 2014 11:43:05 -0800 Subject: [PATCH 0270/1103] Added support for AVX to Eigen. --- CMakeLists.txt | 6 + Eigen/Core | 22 +- Eigen/Geometry | 4 +- Eigen/LU | 4 +- Eigen/src/Core/DenseStorage.h | 4 +- Eigen/src/Core/Redux.h | 13 +- Eigen/src/Core/arch/AVX/CMakeLists.txt | 6 + Eigen/src/Core/arch/AVX/Complex.h | 443 ++++++++++++++++++ Eigen/src/Core/arch/AVX/PacketMath.h | 423 +++++++++++++++++ Eigen/src/Core/arch/SSE/PacketMath.h | 8 + Eigen/src/Core/products/GeneralMatrixMatrix.h | 6 +- Eigen/src/Core/products/GeneralMatrixVector.h | 11 +- Eigen/src/Core/util/Macros.h | 9 +- Eigen/src/Core/util/Memory.h | 38 +- cmake/EigenTesting.cmake | 6 + test/packetmath.cpp | 35 +- test/unalignedcount.cpp | 9 +- 17 files changed, 993 insertions(+), 54 deletions(-) create mode 100644 Eigen/src/Core/arch/AVX/CMakeLists.txt create mode 100644 Eigen/src/Core/arch/AVX/Complex.h create mode 100644 Eigen/src/Core/arch/AVX/PacketMath.h diff --git a/CMakeLists.txt b/CMakeLists.txt index a4731c8c1..16a9b5bcb 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -196,6 +196,12 @@ if(NOT MSVC) message(STATUS "Enabling SSE4.2 in tests/examples") endif() + option(EIGEN_TEST_AVX "Enable/Disable AVX in tests/examples" ON) + if(EIGEN_TEST_AVX) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx") + message(STATUS "Enabling AVX in tests/examples") + endif() + option(EIGEN_TEST_ALTIVEC "Enable/Disable AltiVec in tests/examples" OFF) if(EIGEN_TEST_ALTIVEC) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -maltivec -mabi=altivec") diff --git a/Eigen/Core b/Eigen/Core index 468ae0c76..bd20d5ac5 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -110,7 +110,13 @@ #ifdef __SSE4_2__ #define EIGEN_VECTORIZE_SSE4_2 #endif - + #ifdef __AVX__ + #define EIGEN_VECTORIZE_AVX + #define EIGEN_VECTORIZE_SSE3 + #define EIGEN_VECTORIZE_SSSE3 + #define EIGEN_VECTORIZE_SSE4_1 + #define EIGEN_VECTORIZE_SSE4_2 + #endif // include files // This extern "C" works around a MINGW-w64 compilation issue @@ -140,6 +146,9 @@ #ifdef EIGEN_VECTORIZE_SSE4_2 #include #endif + #ifdef EIGEN_VECTORIZE_AVX + #include + #endif #endif } // end extern "C" #elif defined __ALTIVEC__ @@ -209,7 +218,9 @@ namespace Eigen { inline static const char *SimdInstructionSetsInUse(void) { -#if defined(EIGEN_VECTORIZE_SSE4_2) +#if defined(EIGEN_VECTORIZE_AVX) + return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2"; +#elif defined(EIGEN_VECTORIZE_SSE4_2) return "SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2"; #elif defined(EIGEN_VECTORIZE_SSE4_1) return "SSE, SSE2, SSE3, SSSE3, SSE4.1"; @@ -287,7 +298,12 @@ using std::ptrdiff_t; #include "src/Core/MathFunctions.h" #include "src/Core/GenericPacketMath.h" -#if defined EIGEN_VECTORIZE_SSE +#if defined EIGEN_VECTORIZE_AVX + // Use AVX for floats and doubles, SSE for integers + #include "src/Core/arch/AVX/PacketMath.h" + #include "src/Core/arch/AVX/Complex.h" + #include "src/Core/arch/SSE/PacketMath.h" // For integers +#elif defined EIGEN_VECTORIZE_SSE #include "src/Core/arch/SSE/PacketMath.h" #include "src/Core/arch/SSE/MathFunctions.h" #include "src/Core/arch/SSE/Complex.h" diff --git a/Eigen/Geometry b/Eigen/Geometry index efd9d4504..f9bc6fc57 100644 --- a/Eigen/Geometry +++ b/Eigen/Geometry @@ -47,7 +47,9 @@ #include "src/Geometry/AlignedBox.h" #include "src/Geometry/Umeyama.h" - #if defined EIGEN_VECTORIZE_SSE + // Use the SSE optimized version whenever possible. At the moment the + // SSE version doesn't compile when AVX is enabled + #if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX #include "src/Geometry/arch/Geometry_SSE.h" #endif #endif diff --git a/Eigen/LU b/Eigen/LU index db5795504..e5c3f32f7 100644 --- a/Eigen/LU +++ b/Eigen/LU @@ -27,7 +27,9 @@ #include "src/LU/Determinant.h" #include "src/LU/Inverse.h" -#if defined EIGEN_VECTORIZE_SSE +// Use the SSE optimized version whenever possible. At the moment the +// SSE version doesn't compile when AVX is enabled +#if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX #include "src/LU/arch/Inverse_SSE.h" #endif diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index b9dd75ade..2342b08a1 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -83,7 +83,7 @@ struct plain_array template struct plain_array { - EIGEN_USER_ALIGN16 T array[Size]; + EIGEN_USER_ALIGN32 T array[Size]; EIGEN_DEVICE_FUNC plain_array() @@ -102,7 +102,7 @@ struct plain_array template struct plain_array { - EIGEN_USER_ALIGN16 T array[1]; + EIGEN_USER_ALIGN32 T array[1]; EIGEN_DEVICE_FUNC plain_array() {} EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {} }; diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index b2c775d90..5b82c9a65 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -303,10 +303,15 @@ struct redux_impl static EIGEN_STRONG_INLINE Scalar run(const Derived& mat, const Func& func) { eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); - Scalar res = func.predux(redux_vec_unroller::run(mat,func)); - if (VectorizedSize != Size) - res = func(res,redux_novec_unroller::run(mat,func)); - return res; + if (VectorizedSize > 0) { + Scalar res = func.predux(redux_vec_unroller::run(mat,func)); + if (VectorizedSize != Size) + res = func(res,redux_novec_unroller::run(mat,func)); + return res; + } + else { + return redux_novec_unroller::run(mat,func); + } } }; diff --git a/Eigen/src/Core/arch/AVX/CMakeLists.txt b/Eigen/src/Core/arch/AVX/CMakeLists.txt new file mode 100644 index 000000000..bdb71ab99 --- /dev/null +++ b/Eigen/src/Core/arch/AVX/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_Core_arch_AVX_SRCS "*.h") + +INSTALL(FILES + ${Eigen_Core_arch_AVX_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Core/arch/AVX COMPONENT Devel +) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h new file mode 100644 index 000000000..9fb44ecab --- /dev/null +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -0,0 +1,443 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com) +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_COMPLEX_AVX_H +#define EIGEN_COMPLEX_AVX_H + +namespace Eigen { + +namespace internal { + +//---------- float ---------- +struct Packet4cf +{ + EIGEN_STRONG_INLINE Packet4cf() {} + EIGEN_STRONG_INLINE explicit Packet4cf(const __m256& a) : v(a) {} + __m256 v; +}; + +template<> struct packet_traits > : default_packet_traits +{ + typedef Packet4cf type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 4, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits { typedef std::complex type; enum {size=4}; }; + +template<> EIGEN_STRONG_INLINE Packet4cf padd(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_add_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cf psub(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_sub_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cf pnegate(const Packet4cf& a) +{ + return Packet4cf(pnegate(a.v)); +} +template<> EIGEN_STRONG_INLINE Packet4cf pconj(const Packet4cf& a) +{ + const __m256 mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000)); + return Packet4cf(_mm256_xor_ps(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) +{ + __m256 tmp1 = _mm256_mul_ps(_mm256_moveldup_ps(a.v), b.v); + __m256 tmp2 = _mm256_mul_ps(_mm256_movehdup_ps(a.v), _mm256_permute_ps(b.v, _MM_SHUFFLE(2,3,0,1))); + __m256 result = _mm256_addsub_ps(tmp1, tmp2); + return Packet4cf(result); +} + +template<> EIGEN_STRONG_INLINE Packet4cf pand (const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_and_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cf por (const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_or_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cf pxor (const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_xor_ps(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet4cf pandnot(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_andnot_ps(a.v,b.v)); } + +template<> EIGEN_STRONG_INLINE Packet4cf pload (const std::complex* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet4cf(pload(&numext::real_ref(*from))); } +template<> EIGEN_STRONG_INLINE Packet4cf ploadu(const std::complex* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet4cf(ploadu(&numext::real_ref(*from))); } + + +template<> EIGEN_STRONG_INLINE Packet4cf pset1(const std::complex& from) +{ + __m256 result; + for (int i = 0; i < 8; i+=2) { + result[i] = std::real(from); + result[i+1] = std::imag(from); + } + return Packet4cf(result); +} + +template<> EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex* from) +{ + __m256 result; + for (int i = 0; i < 2; ++i) { + result[4*i] = std::real(from[i]); + result[4*i+1] = std::imag(from[i]); + result[4*i+2] = std::real(from[i]); + result[4*i+3] = std::imag(from[i]); + } + return Packet4cf(result); +} + +template<> EIGEN_STRONG_INLINE void pstore >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + + +template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet4cf& a) +{ + return std::complex(a.v[0], a.v[1]); +} + +template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { + __m256 result; + result[0] = a.v[6]; + result[1] = a.v[7]; + result[2] = a.v[4]; + result[3] = a.v[5]; + result[4] = a.v[2]; + result[5] = a.v[3]; + result[6] = a.v[0]; + result[7] = a.v[1]; + return Packet4cf(result); +} + +template<> EIGEN_STRONG_INLINE std::complex predux(const Packet4cf& a) +{ + return std::complex(a.v[0]+a.v[2]+a.v[4]+a.v[6], a.v[1]+a.v[3]+a.v[5]+a.v[7]); +} + +template<> EIGEN_STRONG_INLINE Packet4cf preduxp(const Packet4cf* vecs) +{ + __m256 result = _mm256_setzero_ps(); + for (int i = 0; i < 4; ++i) { + for (int j = 0; j < 8; j+=2) { + result[2*i] += vecs[i].v[j]; + result[2*i+1] += vecs[i].v[j+1]; + } + } + return Packet4cf(result); +} + +template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet4cf& a) +{ + std::complex result(a.v[0], a.v[1]); + for (int i = 2; i < 8; i+=2) { + result *= std::complex(a.v[i], a.v[i+1]); + } + return result; +} + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet4cf& first, const Packet4cf& second) + { + if (Offset==0) return; + for (int i = 0; i < 4-Offset; ++i) + { + first.v[2*i] = first.v[2*(i+Offset)]; + first.v[2*i+1] = first.v[2*(i+Offset)+1]; + } + for (int i = 4-Offset; i < 4; ++i) + { + first.v[2*i] = second.v[2*(i-4+Offset)]; + first.v[2*i+1] = second.v[2*(i-4+Offset)+1]; + } + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) const + { return Packet4cf(Eigen::internal::pmul(x, y.v)); } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet8f& y, const Packet4cf& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& x, const Packet8f& y) const + { return Packet4cf(Eigen::internal::pmul(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet4cf pdiv(const Packet4cf& a, const Packet4cf& b) +{ + Packet4cf res; + for (int i = 0; i < 8; i+=2) { + std::complex result = std::complex(a.v[i], a.v[i+1]) / std::complex(b.v[i], b.v[i+1]); + res.v[i] = std::real(result); + res.v[i+1] = std::imag(result); + } + return res; +} + +template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip(const Packet4cf& x) +{ + Packet4cf res; + for (int i = 0; i < 8; i+=2) { + res.v[i] = x.v[i+1]; + res.v[i+1] = x.v[i]; + } + return res; +} + +//---------- double ---------- +struct Packet2cd +{ + EIGEN_STRONG_INLINE Packet2cd() {} + EIGEN_STRONG_INLINE explicit Packet2cd(const __m256d& a) : v(a) {} + __m256d v; +}; + +template<> struct packet_traits > : default_packet_traits +{ + typedef Packet2cd type; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 2, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; }; + +template<> EIGEN_STRONG_INLINE Packet2cd padd(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_add_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd psub(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_sub_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd pnegate(const Packet2cd& a) { return Packet2cd(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd pconj(const Packet2cd& a) +{ + const __m256d mask = _mm256_castsi256_pd(_mm256_set_epi32(0x80000000,0x0,0x0,0x0,0x80000000,0x0,0x0,0x0)); + return Packet2cd(_mm256_xor_pd(a.v,mask)); +} + +template<> EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) +{ + __m256d tmp1 = _mm256_mul_pd(_mm256_permute_pd(a.v, 0), b.v); + // FIXME: _mm256_permute_pd(b.v, _MM_SHUFFLE2(1,0) won't work as expected, figure out an alternative. + __m256d op = {b.v[1], b.v[0], b.v[3], b.v[2]}; + __m256d tmp2 = _mm256_mul_pd(_mm256_permute_pd(a.v, 15), op); + __m256d result = _mm256_addsub_pd(tmp1, tmp2); + + return Packet2cd(result); +} + +template<> EIGEN_STRONG_INLINE Packet2cd pand (const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd por (const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_or_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd pxor (const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_xor_pd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet2cd pandnot(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_andnot_pd(a.v,b.v)); } + +template<> EIGEN_STRONG_INLINE Packet2cd pload (const std::complex* from) +{ EIGEN_DEBUG_ALIGNED_LOAD return Packet2cd(pload((const double*)from)); } +template<> EIGEN_STRONG_INLINE Packet2cd ploadu(const std::complex* from) +{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cd(ploadu((const double*)from)); } + +template<> EIGEN_STRONG_INLINE Packet2cd pset1(const std::complex& from) +{ + __m256d result; + for (int i = 0; i < 4; i+=2) { + result[i] = std::real(from); + result[i+1] = std::imag(from); + } + return Packet2cd(result); +} + +template<> EIGEN_STRONG_INLINE Packet2cd ploaddup(const std::complex* from) { return pset1(*from); } + +template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cd& a) +{ + return std::complex(a.v[0],a.v[1]); +} + +template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) { + __m256d result; + result[0] = a.v[2]; + result[1] = a.v[3]; + result[2] = a.v[0]; + result[3] = a.v[1]; + return Packet2cd(result); +} + +template<> EIGEN_STRONG_INLINE std::complex predux(const Packet2cd& a) +{ + return std::complex(a.v[0]+a.v[2], a.v[1]+a.v[3]); +} + +template<> EIGEN_STRONG_INLINE Packet2cd preduxp(const Packet2cd* vecs) +{ + __m256d result = _mm256_setzero_pd(); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 4; j+=2) { + result[2*i] += vecs[i].v[j]; + result[2*i+1] += vecs[i].v[j+1]; + } + } + return Packet2cd(result); +} + +template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet2cd& a) +{ + return std::complex(a.v[0], a.v[1]) * std::complex(a.v[2], a.v[3]); +} + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet2cd& first, const Packet2cd& second) + { + if (Offset==0) return; + first.v[0] = first.v[2]; + first.v[1] = first.v[3]; + first.v[2] = second.v[0]; + first.v[3] = second.v[1]; + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet4d& x, const Packet2cd& y, const Packet2cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet4d& x, const Packet2cd& y) const + { return Packet2cd(Eigen::internal::pmul(x, y.v)); } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet4d& y, const Packet2cd& c) const + { return padd(c, pmul(x,y)); } + + EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& x, const Packet4d& y) const + { return Packet2cd(Eigen::internal::pmul(x.v, y)); } +}; + +template<> EIGEN_STRONG_INLINE Packet2cd pdiv(const Packet2cd& a, const Packet2cd& b) +{ + Packet2cd res; + for (int i = 0; i < 4; i+=2) { + std::complex result = std::complex(a.v[i], a.v[i+1]) / std::complex(b.v[i], b.v[i+1]); + res.v[i] = std::real(result); + res.v[i+1] = std::imag(result); + } + return res; +} + +template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip(const Packet2cd& x) +{ + Packet2cd res; + for (int i = 0; i < 4; i+=2) { + res.v[i] = x.v[i+1]; + res.v[i+1] = x.v[i]; + } + return res; +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_COMPLEX_AVX_H diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h new file mode 100644 index 000000000..244e63e74 --- /dev/null +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -0,0 +1,423 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com) +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_PACKET_MATH_AVX_H +#define EIGEN_PACKET_MATH_AVX_H + +namespace Eigen { + +namespace internal { + +#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD +#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 +#endif + +#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) +#endif + +typedef __m256 Packet8f; +typedef __m256i Packet8i; +typedef __m256d Packet4d; + +template<> struct is_arithmetic<__m256> { enum { value = true }; }; +template<> struct is_arithmetic<__m256i> { enum { value = true }; }; +template<> struct is_arithmetic<__m256d> { enum { value = true }; }; + +#define _EIGEN_DECLARE_CONST_Packet8f(NAME,X) \ + const Packet8f p8f_##NAME = pset1(X) + +#define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \ + const Packet4d p4d_##NAME = pset1(X) + + +template<> struct packet_traits : default_packet_traits +{ + typedef Packet8f type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=8, + + HasDiv = 1, + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 0, + HasSqrt = 0 + }; + }; +template<> struct packet_traits : default_packet_traits +{ + typedef Packet4d type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + + HasDiv = 1, + HasExp = 0 + }; +}; + +/* Proper support for integers is only provided by AVX2. In the meantime, we'll + use SSE instructions and packets to deal with integers. +template<> struct packet_traits : default_packet_traits +{ + typedef Packet8i type; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=8 + }; +}; +*/ + +template<> struct unpacket_traits { typedef float type; enum {size=8}; }; +template<> struct unpacket_traits { typedef double type; enum {size=4}; }; +template<> struct unpacket_traits { typedef int type; enum {size=8}; }; + +template<> EIGEN_STRONG_INLINE Packet8f pset1(const float& from) { return _mm256_set1_ps(from); } +template<> EIGEN_STRONG_INLINE Packet4d pset1(const double& from) { return _mm256_set1_pd(from); } +template<> EIGEN_STRONG_INLINE Packet8i pset1(const int& from) { return _mm256_set1_epi32(from); } + +template<> EIGEN_STRONG_INLINE Packet8f plset(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); } +template<> EIGEN_STRONG_INLINE Packet4d plset(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); } + +template<> EIGEN_STRONG_INLINE Packet8f padd(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d padd(const Packet4d& a, const Packet4d& b) { return _mm256_add_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f psub(const Packet8f& a, const Packet8f& b) { return _mm256_sub_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d psub(const Packet4d& a, const Packet4d& b) { return _mm256_sub_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f pnegate(const Packet8f& a) +{ + return _mm256_sub_ps(_mm256_set1_ps(0.0),a); +} +template<> EIGEN_STRONG_INLINE Packet4d pnegate(const Packet4d& a) +{ + return _mm256_sub_pd(_mm256_set1_pd(0.0),a); +} + +template<> EIGEN_STRONG_INLINE Packet8f pconj(const Packet8f& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet4d pconj(const Packet4d& a) { return a; } +template<> EIGEN_STRONG_INLINE Packet8i pconj(const Packet8i& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet8f pmul(const Packet8f& a, const Packet8f& b) { return _mm256_mul_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pmul(const Packet4d& a, const Packet4d& b) { return _mm256_mul_pd(a,b); } + + +template<> EIGEN_STRONG_INLINE Packet8f pdiv(const Packet8f& a, const Packet8f& b) { return _mm256_div_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pdiv(const Packet4d& a, const Packet4d& b) { return _mm256_div_pd(a,b); } +template<> EIGEN_STRONG_INLINE Packet8i pdiv(const Packet8i& /*a*/, const Packet8i& /*b*/) +{ eigen_assert(false && "packet integer division are not supported by AVX"); + return pset1(0); +} + +template<> EIGEN_STRONG_INLINE Packet8f pmin(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pmin(const Packet4d& a, const Packet4d& b) { return _mm256_min_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f pmax(const Packet8f& a, const Packet8f& b) { return _mm256_max_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pmax(const Packet4d& a, const Packet4d& b) { return _mm256_max_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f pand(const Packet8f& a, const Packet8f& b) { return _mm256_and_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pand(const Packet4d& a, const Packet4d& b) { return _mm256_and_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f por(const Packet8f& a, const Packet8f& b) { return _mm256_or_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d por(const Packet4d& a, const Packet4d& b) { return _mm256_or_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f pxor(const Packet8f& a, const Packet8f& b) { return _mm256_xor_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pxor(const Packet4d& a, const Packet4d& b) { return _mm256_xor_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f pandnot(const Packet8f& a, const Packet8f& b) { return _mm256_andnot_ps(a,b); } +template<> EIGEN_STRONG_INLINE Packet4d pandnot(const Packet4d& a, const Packet4d& b) { return _mm256_andnot_pd(a,b); } + +template<> EIGEN_STRONG_INLINE Packet8f pload(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_ps(from); } +template<> EIGEN_STRONG_INLINE Packet4d pload(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_pd(from); } +template<> EIGEN_STRONG_INLINE Packet8i pload(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(reinterpret_cast(from)); } + +template<> EIGEN_STRONG_INLINE Packet8f ploadu(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_ps(from); } +template<> EIGEN_STRONG_INLINE Packet4d ploadu(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_pd(from); } +template<> EIGEN_STRONG_INLINE Packet8i ploadu(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(reinterpret_cast(from)); } + +// Loads 4 floats from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, a3} +template<> EIGEN_STRONG_INLINE Packet8f ploaddup(const float* from) +{ + Packet8f tmp = ploadu(from); + Packet8f tmp1 = _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2)); + Packet8f tmp2 = _mm256_permute_ps(tmp, _MM_SHUFFLE(1,1,0,0)); + return _mm256_blend_ps(_mm256_permute2f128_ps(tmp1,tmp1,1),tmp2,15); +} +// Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1} +template<> EIGEN_STRONG_INLINE Packet4d ploaddup(const double* from) +{ + Packet4d tmp = ploadu(from); + Packet4d tmp1 = _mm256_permute_pd(tmp,0); + Packet4d tmp2 = _mm256_permute_pd(tmp,3); + return _mm256_blend_pd(tmp1,_mm256_permute2f128_pd(tmp2,tmp2,1),12); +} + +template<> EIGEN_STRONG_INLINE void pstore(float* to, const Packet8f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstore(double* to, const Packet4d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet8i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } + +template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet8f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_ps(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); } +template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } + +template<> EIGEN_STRONG_INLINE void pstore1(float* to, const float& a) +{ + Packet8f pa = pset1(a); + pstore(to, pa); +} +template<> EIGEN_STRONG_INLINE void pstore1(double* to, const double& a) +{ + Packet4d pa = pset1(a); + pstore(to, pa); +} +template<> EIGEN_STRONG_INLINE void pstore1(int* to, const int& a) +{ + Packet8i pa = pset1(a); + pstore(to, pa); +} + +template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } + +template<> EIGEN_STRONG_INLINE float pfirst(const Packet8f& a) { + return _mm_cvtss_f32(_mm256_castps256_ps128(a)); +} +template<> EIGEN_STRONG_INLINE double pfirst(const Packet4d& a) { + return _mm_cvtsd_f64(_mm256_castpd256_pd128(a)); +} +template<> EIGEN_STRONG_INLINE int pfirst(const Packet8i& a) { + return _mm_cvtsi128_si32(_mm256_castsi256_si128(a)); +} + + +template<> EIGEN_STRONG_INLINE Packet8f preverse(const Packet8f& a) +{ + __m256 tmp = _mm256_shuffle_ps(a,a,0x1b); + return _mm256_permute2f128_ps(tmp, tmp, 1); +} +template<> EIGEN_STRONG_INLINE Packet4d preverse(const Packet4d& a) +{ + __m256d tmp = _mm256_shuffle_pd(a,a,5); + return _mm256_permute2f128_pd(tmp, tmp, 1); + + __m256d swap_halves = _mm256_permute2f128_pd(a,a,1); + return _mm256_permute_pd(swap_halves,5); +} + +// pabs should be ok +template<> EIGEN_STRONG_INLINE Packet8f pabs(const Packet8f& a) +{ + const Packet8f mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); + return _mm256_and_ps(a,mask); +} +template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a) +{ + const Packet4d mask = _mm256_castsi256_pd(_mm256_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); + return _mm256_and_pd(a,mask); +} + +// preduxp should be ok +// FIXME: why is this ok? why isn't the simply implementation working as expected? +template<> EIGEN_STRONG_INLINE Packet8f preduxp(const Packet8f* vecs) +{ + __m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]); + __m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]); + __m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]); + __m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]); + + __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1); + __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2); + __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3); + __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4); + + __m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); + __m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); + __m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); + __m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); + + __m256 sum1 = _mm256_add_ps(perm1, hsum5); + __m256 sum2 = _mm256_add_ps(perm2, hsum6); + __m256 sum3 = _mm256_add_ps(perm3, hsum7); + __m256 sum4 = _mm256_add_ps(perm4, hsum8); + + __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); + __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); + + __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0); + return final; +} +template<> EIGEN_STRONG_INLINE Packet4d preduxp(const Packet4d* vecs) +{ + Packet4d tmp0, tmp1; + + tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]); + tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); + + tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]); + tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); + + return _mm256_blend_pd(tmp0, tmp1, 0xC); +} + +template<> EIGEN_STRONG_INLINE float predux(const Packet8f& a) +{ + Packet8f tmp0 = _mm256_hadd_ps(a,_mm256_permute2f128_ps(a,a,1)); + tmp0 = _mm256_hadd_ps(tmp0,tmp0); + return pfirst(_mm256_hadd_ps(tmp0, tmp0)); +} +template<> EIGEN_STRONG_INLINE double predux(const Packet4d& a) +{ + Packet4d tmp0 = _mm256_hadd_pd(a,_mm256_permute2f128_pd(a,a,1)); + return pfirst(_mm256_hadd_pd(tmp0,tmp0)); +} + +template<> EIGEN_STRONG_INLINE float predux_mul(const Packet8f& a) +{ + Packet8f tmp; + tmp = _mm256_mul_ps(a, _mm256_permute2f128_ps(a,a,1)); + tmp = _mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); + return pfirst(_mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); +} +template<> EIGEN_STRONG_INLINE double predux_mul(const Packet4d& a) +{ + Packet4d tmp; + tmp = _mm256_mul_pd(a, _mm256_permute2f128_pd(a,a,1)); + return pfirst(_mm256_mul_pd(tmp, _mm256_shuffle_pd(tmp,tmp,1))); +} + +template<> EIGEN_STRONG_INLINE float predux_min(const Packet8f& a) +{ + float result = a[0]; + for (int i = 1; i < 8; ++i) { + if (a[i] < result) result = a[i]; + } + return result; +} +template<> EIGEN_STRONG_INLINE double predux_min(const Packet4d& a) +{ + double result = a[0]; + for (int i = 1; i < 4; ++i) { + if (a[i] < result) result = a[i]; + } + return result; +} + +template<> EIGEN_STRONG_INLINE float predux_max(const Packet8f& a) +{ + float result = a[0]; + for (int i = 1; i < 8; ++i) { + if (a[i] > result) result = a[i]; + } + return result; +} + +template<> EIGEN_STRONG_INLINE double predux_max(const Packet4d& a) +{ + double result = a[0]; + for (int i = 1; i < 4; ++i) { + if (a[i] > result) result = a[i]; + } + return result; +} + + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet8f& first, const Packet8f& second) + { + if (Offset==1) + { + first = _mm256_blend_ps(first, second, 1); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); + first = _mm256_blend_ps(tmp, _mm256_permute2f128_ps (tmp, tmp, 1), 0x88); + } + else if (Offset==2) + { + first = _mm256_blend_ps(first, second, 3); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); + first = _mm256_blend_ps(tmp, _mm256_permute2f128_ps (tmp, tmp, 1), 0xcc); + } + else if (Offset==3) + { + first = _mm256_blend_ps(first, second, 7); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); + first = _mm256_blend_ps(tmp, _mm256_permute2f128_ps (tmp, tmp, 1), 0xee); + } + else if (Offset==4) + { + first = _mm256_blend_ps(first, second, 15); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0)); + first = _mm256_permute_ps(_mm256_permute2f128_ps (tmp, tmp, 1), _MM_SHUFFLE(3,2,1,0)); + } + else if (Offset==5) + { + first = _mm256_blend_ps(first, second, 31); + first = _mm256_permute2f128_ps(first, first, 1); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); + first = _mm256_permute2f128_ps(tmp, tmp, 1); + first = _mm256_blend_ps(tmp, first, 0x88); + } + else if (Offset==6) + { + first = _mm256_blend_ps(first, second, 63); + first = _mm256_permute2f128_ps(first, first, 1); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); + first = _mm256_permute2f128_ps(tmp, tmp, 1); + first = _mm256_blend_ps(tmp, first, 0xcc); + } + else if (Offset==7) + { + first = _mm256_blend_ps(first, second, 127); + first = _mm256_permute2f128_ps(first, first, 1); + Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); + first = _mm256_permute2f128_ps(tmp, tmp, 1); + first = _mm256_blend_ps(tmp, first, 0xee); + } + } +}; + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet4d& first, const Packet4d& second) + { + if (Offset==1) + { + first = _mm256_blend_pd(first, second, 1); + __m256d tmp = _mm256_permute_pd(first, 5); + first = _mm256_permute2f128_pd(tmp, tmp, 1); + first = _mm256_blend_pd(tmp, first, 0xA); + } + else if (Offset==2) + { + first = _mm256_blend_pd(first, second, 3); + first = _mm256_permute2f128_pd(first, first, 1); + } + else if (Offset==3) + { + first = _mm256_blend_pd(first, second, 7); + __m256d tmp = _mm256_permute_pd(first, 5); + first = _mm256_permute2f128_pd(tmp, tmp, 1); + first = _mm256_blend_pd(tmp, first, 5); + } + } +}; + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_PACKET_MATH_AVX_H diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index f5a3dab52..e913af650 100644 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -58,6 +58,9 @@ template<> struct is_arithmetic<__m128d> { enum { value = true }; }; const Packet4i p4i_##NAME = pset1(X) +// Use the packet_traits defined in AVX/PacketMath.h instead if we're going +// to leverage AVX instructions. +#ifndef EIGEN_VECTORIZE_AVX template<> struct packet_traits : default_packet_traits { typedef Packet4f type; @@ -87,6 +90,7 @@ template<> struct packet_traits : default_packet_traits HasSqrt = 1 }; }; +#endif template<> struct packet_traits : default_packet_traits { typedef Packet4i type; @@ -115,8 +119,10 @@ template<> EIGEN_STRONG_INLINE Packet2d pset1(const double& from) { re template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return _mm_set1_epi32(from); } #endif +#ifndef EIGEN_VECTORIZE_AVX template<> EIGEN_STRONG_INLINE Packet4f plset(const float& a) { return _mm_add_ps(pset1(a), _mm_set_ps(3,2,1,0)); } template<> EIGEN_STRONG_INLINE Packet2d plset(const double& a) { return _mm_add_pd(pset1(a),_mm_set_pd(1,0)); } +#endif template<> EIGEN_STRONG_INLINE Packet4i plset(const int& a) { return _mm_add_epi32(pset1(a),_mm_set_epi32(3,2,1,0)); } template<> EIGEN_STRONG_INLINE Packet4f padd(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); } @@ -331,9 +337,11 @@ template<> EIGEN_STRONG_INLINE void pstore1(double* to, const double& pstore(to, vec2d_swizzle1(pa,0,0)); } +#ifndef EIGEN_VECTORIZE_AVX template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } +#endif #if defined(_MSC_VER) && defined(_WIN64) && !defined(__INTEL_COMPILER) // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 3f5ffcf51..eb399a824 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -286,9 +286,9 @@ class gemm_blocking_space 4. + if( (size_t(lhs)%sizeof(LhsScalar)) || (size_t(res)%sizeof(ResScalar)) || LhsPacketSize > 4) { alignedSize = 0; alignedStart = 0; @@ -404,7 +405,9 @@ EIGEN_DONT_INLINE void general_matrix_vector_product 4. + if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (size_t(lhs)%sizeof(LhsScalar)) || (size_t(rhs)%sizeof(RhsScalar)) || + (LhsPacketSize > 4)) { alignedSize = 0; alignedStart = 0; @@ -446,7 +449,7 @@ EIGEN_DONT_INLINE void general_matrix_vector_product(tmp0); const LhsScalar* lhs0 = lhs + i*lhsStride; // process first unaligned result's coeffs diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index debc04f3f..787f800b8 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -72,7 +72,11 @@ #endif #define EIGEN_ALIGN 0 #else - #define EIGEN_ALIGN 1 + #if !defined(EIGEN_DONT_VECTORIZE) && defined(__AVX__) + #define EIGEN_ALIGN 32 + #else + #define EIGEN_ALIGN 16 + #endif #endif // EIGEN_ALIGN_STATICALLY is the true test whether we want to align arrays on the stack or not. It takes into account both the user choice to explicitly disable @@ -281,13 +285,16 @@ #endif #define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16) +#define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32) #if EIGEN_ALIGN_STATICALLY #define EIGEN_USER_ALIGN_TO_BOUNDARY(n) EIGEN_ALIGN_TO_BOUNDARY(n) #define EIGEN_USER_ALIGN16 EIGEN_ALIGN16 +#define EIGEN_USER_ALIGN32 EIGEN_ALIGN32 #else #define EIGEN_USER_ALIGN_TO_BOUNDARY(n) #define EIGEN_USER_ALIGN16 +#define EIGEN_USER_ALIGN32 #endif #ifdef EIGEN_DONT_USE_RESTRICT_KEYWORD diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index d177e8b5a..76bdb6cfc 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -32,7 +32,7 @@ // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed // quite safe, at least within the context of glibc, to equate 64-bit with LP64. #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \ - && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) + && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_ALIGN == 16) #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1 #else #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0 @@ -42,14 +42,14 @@ // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup -#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__) +#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__) && (EIGEN_ALIGN == 16) #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1 #else #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0 #endif -#if defined(__APPLE__) \ - || defined(_WIN64) \ +#if (defined(__APPLE__) && (EIGEN_ALIGN == 16)) \ + || (defined(_WIN64) && (EIGEN_ALIGN == 16)) \ || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \ || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED #define EIGEN_MALLOC_ALREADY_ALIGNED 1 @@ -73,7 +73,7 @@ #define EIGEN_HAS_POSIX_MEMALIGN 0 #endif -#ifdef EIGEN_VECTORIZE_SSE +#if defined EIGEN_VECTORIZE_SSE || defined EIGEN_VECTORIZE_AVX #define EIGEN_HAS_MM_MALLOC 1 #else #define EIGEN_HAS_MM_MALLOC 0 @@ -105,9 +105,9 @@ inline void throw_std_bad_alloc() */ inline void* handmade_aligned_malloc(std::size_t size) { - void *original = std::malloc(size+16); + void *original = std::malloc(size+EIGEN_ALIGN); if (original == 0) return 0; - void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(15))) + 16); + void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(EIGEN_ALIGN-1))) + EIGEN_ALIGN); *(reinterpret_cast(aligned) - 1) = original; return aligned; } @@ -128,9 +128,9 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = if (ptr == 0) return handmade_aligned_malloc(size); void *original = *(reinterpret_cast(ptr) - 1); std::ptrdiff_t previous_offset = static_cast(ptr)-static_cast(original); - original = std::realloc(original,size+16); + original = std::realloc(original,size+EIGEN_ALIGN); if (original == 0) return 0; - void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(15))) + 16); + void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(EIGEN_ALIGN-1))) + EIGEN_ALIGN); void *previous_aligned = static_cast(original)+previous_offset; if(aligned!=previous_aligned) std::memmove(aligned, previous_aligned, size); @@ -208,7 +208,7 @@ inline void check_that_malloc_is_allowed() {} #endif -/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 bytes alignment. +/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements. * On allocation error, the returned pointer is null, and std::bad_alloc is thrown. */ inline void* aligned_malloc(size_t size) @@ -221,11 +221,11 @@ inline void* aligned_malloc(size_t size) #elif EIGEN_MALLOC_ALREADY_ALIGNED result = std::malloc(size); #elif EIGEN_HAS_POSIX_MEMALIGN - if(posix_memalign(&result, 16, size)) result = 0; + if(posix_memalign(&result, EIGEN_ALIGN, size)) result = 0; #elif EIGEN_HAS_MM_MALLOC - result = _mm_malloc(size, 16); + result = _mm_malloc(size, EIGEN_ALIGN); #elif defined(_MSC_VER) && (!defined(_WIN32_WCE)) - result = _aligned_malloc(size, 16); + result = _aligned_malloc(size, EIGEN_ALIGN); #else result = handmade_aligned_malloc(size); #endif @@ -275,12 +275,12 @@ inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size) // implements _mm_malloc/_mm_free based on the corresponding _aligned_ // functions. This may not always be the case and we just try to be safe. #if defined(_MSC_VER) && defined(_mm_free) - result = _aligned_realloc(ptr,new_size,16); + result = _aligned_realloc(ptr,new_size,EIGEN_ALIGN); #else result = generic_aligned_realloc(ptr,new_size,old_size); #endif #elif defined(_MSC_VER) - result = _aligned_realloc(ptr,new_size,16); + result = _aligned_realloc(ptr,new_size,EIGEN_ALIGN); #else result = handmade_aligned_realloc(ptr,new_size,old_size); #endif @@ -607,9 +607,9 @@ template class aligned_stack_memory_handler * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token. */ #ifdef EIGEN_ALLOCA - - #ifdef __arm__ - #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast((reinterpret_cast(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16) + // The native alloca() that comes with llvm aligns buffer on 16 bytes even when AVX is enabled. + #if defined(__arm__) || EIGEN_ALIGN > 16 + #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast((reinterpret_cast(EIGEN_ALLOCA(SIZE+EIGEN_ALIGN)) & ~(size_t(EIGEN_ALIGN-1))) + EIGEN_ALIGN) #else #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA #endif @@ -679,7 +679,7 @@ template class aligned_stack_memory_handler #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true) #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \ - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%16==0))) + EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_ALIGN==0))) /****************************************************************************/ diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index fdc166bf8..d1ea4b9d2 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -264,6 +264,12 @@ macro(ei_testing_print_summary) message(STATUS "SSE4.2: Using architecture defaults") endif() + if(EIGEN_TEST_AVX) + message(STATUS "AVX: ON") + else() + message(STATUS "AVX: Using architecture defaults") + endif() + if(EIGEN_TEST_ALTIVEC) message(STATUS "Altivec: ON") else() diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 2c0519c41..d7c336c22 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -104,11 +104,12 @@ template void packetmath() const int PacketSize = internal::packet_traits::size; typedef typename NumTraits::Real RealScalar; - const int size = PacketSize*4; - EIGEN_ALIGN16 Scalar data1[internal::packet_traits::size*4]; - EIGEN_ALIGN16 Scalar data2[internal::packet_traits::size*4]; - EIGEN_ALIGN16 Packet packets[PacketSize*2]; - EIGEN_ALIGN16 Scalar ref[internal::packet_traits::size*4]; + const int max_size = PacketSize > 4 ? PacketSize : 4; + const int size = PacketSize*max_size; + EIGEN_ALIGN32 Scalar data1[size]; + EIGEN_ALIGN32 Scalar data2[size]; + EIGEN_ALIGN32 Packet packets[PacketSize*2]; + EIGEN_ALIGN32 Scalar ref[size]; RealScalar refvalue = 0; for (int i=0; i void packetmath() else if (offset==1) internal::palign<1>(packets[0], packets[1]); else if (offset==2) internal::palign<2>(packets[0], packets[1]); else if (offset==3) internal::palign<3>(packets[0], packets[1]); + else if (offset==4) internal::palign<4>(packets[0], packets[1]); + else if (offset==5) internal::palign<5>(packets[0], packets[1]); + else if (offset==6) internal::palign<6>(packets[0], packets[1]); + else if (offset==7) internal::palign<7>(packets[0], packets[1]); internal::pstore(data2, packets[0]); for (int i=0; i void packetmath_real() const int PacketSize = internal::packet_traits::size; const int size = PacketSize*4; - EIGEN_ALIGN16 Scalar data1[internal::packet_traits::size*4]; - EIGEN_ALIGN16 Scalar data2[internal::packet_traits::size*4]; - EIGEN_ALIGN16 Scalar ref[internal::packet_traits::size*4]; + EIGEN_ALIGN32 Scalar data1[internal::packet_traits::size*4]; + EIGEN_ALIGN32 Scalar data2[internal::packet_traits::size*4]; + EIGEN_ALIGN32 Scalar ref[internal::packet_traits::size*4]; for (int i=0; i void packetmath_notcomplex() typedef typename internal::packet_traits::type Packet; const int PacketSize = internal::packet_traits::size; - EIGEN_ALIGN16 Scalar data1[internal::packet_traits::size*4]; - EIGEN_ALIGN16 Scalar data2[internal::packet_traits::size*4]; - EIGEN_ALIGN16 Scalar ref[internal::packet_traits::size*4]; + EIGEN_ALIGN32 Scalar data1[internal::packet_traits::size*4]; + EIGEN_ALIGN32 Scalar data2[internal::packet_traits::size*4]; + EIGEN_ALIGN32 Scalar ref[internal::packet_traits::size*4]; Array::Map(data1, internal::packet_traits::size*4).setRandom(); @@ -317,10 +322,10 @@ template void packetmath_complex() const int PacketSize = internal::packet_traits::size; const int size = PacketSize*4; - EIGEN_ALIGN16 Scalar data1[PacketSize*4]; - EIGEN_ALIGN16 Scalar data2[PacketSize*4]; - EIGEN_ALIGN16 Scalar ref[PacketSize*4]; - EIGEN_ALIGN16 Scalar pval[PacketSize*4]; + EIGEN_ALIGN32 Scalar data1[PacketSize*4]; + EIGEN_ALIGN32 Scalar data2[PacketSize*4]; + EIGEN_ALIGN32 Scalar ref[PacketSize*4]; + EIGEN_ALIGN32 Scalar pval[PacketSize*4]; for (int i=0; i Date: Wed, 29 Jan 2014 22:05:39 +0100 Subject: [PATCH 0271/1103] Fix bug #730: Path of OpenGL headers is different on MacOS --- unsupported/Eigen/OpenGLSupport | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/OpenGLSupport b/unsupported/Eigen/OpenGLSupport index 4454bf820..c4090ab11 100644 --- a/unsupported/Eigen/OpenGLSupport +++ b/unsupported/Eigen/OpenGLSupport @@ -11,7 +11,12 @@ #define EIGEN_OPENGL_MODULE #include -#include + +#if defined(__APPLE_CC__) + #include +#else + #include +#endif namespace Eigen { From e722f36ffa80f131e19428bab275dbc2df522093 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Sat, 1 Feb 2014 20:49:48 +0100 Subject: [PATCH 0272/1103] =?UTF-8?q?Fixed=20issue=20#734=20(thanks=20to?= =?UTF-8?q?=20Philipp=20B=C3=BCttgenbach=20for=20reporting=20the=20issue?= =?UTF-8?q?=20and=20proposing=20a=20fix).=20Kept=20ColMajor=20layout=20if?= =?UTF-8?q?=20possible=20in=20order=20to=20keep=20derivatives=20of=20the?= =?UTF-8?q?=20same=20order=20adjacent=20in=20memory.?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- unsupported/Eigen/src/Splines/SplineFwd.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/src/Splines/SplineFwd.h b/unsupported/Eigen/src/Splines/SplineFwd.h index 49db8d35d..5d9bfacf8 100644 --- a/unsupported/Eigen/src/Splines/SplineFwd.h +++ b/unsupported/Eigen/src/Splines/SplineFwd.h @@ -31,6 +31,8 @@ namespace Eigen enum { OrderAtCompileTime = _Degree==Dynamic ? Dynamic : _Degree+1 /*!< The spline curve's order at compile-time. */ }; enum { NumOfDerivativesAtCompileTime = OrderAtCompileTime /*!< The number of derivatives defined for the current spline. */ }; + + enum { DerivativeMemoryLayout = Dimension==1 ? RowMajor : ColMajor /*!< The derivative type's memory layout. */ }; /** \brief The data type used to store non-zero basis functions. */ typedef Array BasisVectorType; @@ -39,7 +41,7 @@ namespace Eigen typedef Array BasisDerivativeType; /** \brief The data type used to store the spline's derivative values. */ - typedef Array DerivativeType; + typedef Array DerivativeType; /** \brief The point type the spline is representing. */ typedef Array PointType; @@ -67,7 +69,7 @@ namespace Eigen typedef Array<_Scalar,Dynamic,Dynamic,RowMajor,NumOfDerivativesAtCompileTime,OrderAtCompileTime> BasisDerivativeType; /** \brief The data type used to store the spline's derivative values. */ - typedef Array<_Scalar,_Dim,Dynamic,ColMajor,_Dim,NumOfDerivativesAtCompileTime> DerivativeType; + typedef Array<_Scalar,_Dim,Dynamic,DerivativeMemoryLayout,_Dim,NumOfDerivativesAtCompileTime> DerivativeType; }; /** \brief 2D float B-spline with dynamic degree. */ From 6c527bd811a4bbb122093fcc9ba9b7bd86963855 Mon Sep 17 00:00:00 2001 From: Hauke Heibel Date: Tue, 4 Feb 2014 07:02:34 +0100 Subject: [PATCH 0273/1103] Fixed assignment from QMatrix to Transform for compact storage. --- Eigen/src/Geometry/Transform.h | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 419f90148..020868103 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -699,9 +699,13 @@ template Transform& Transform::operator=(const QMatrix& other) { EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE) - m_matrix << other.m11(), other.m21(), other.dx(), - other.m12(), other.m22(), other.dy(), - 0, 0, 1; + if (Mode == int(AffineCompact)) + m_matrix << other.m11(), other.m21(), other.dx(), + other.m12(), other.m22(), other.dy(); + else + m_matrix << other.m11(), other.m21(), other.dx(), + other.m12(), other.m22(), other.dy(), + 0, 0, 1; return *this; } From ff8d81762d6612b812db6385b1096c6a8b1006cc Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Thu, 6 Feb 2014 11:06:06 +0000 Subject: [PATCH 0274/1103] Fix bug #736: LDLT isPositive returns false for a positive semidefinite matrix Add unit test covering this case. --- Eigen/src/Cholesky/LDLT.h | 52 +++++++++++++++++++++++---------------- test/cholesky.cpp | 24 +++++++++++++++--- 2 files changed, 52 insertions(+), 24 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 4faedd257..f34d26465 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -16,7 +16,10 @@ namespace Eigen { namespace internal { -template struct LDLT_Traits; + template struct LDLT_Traits; + + // PositiveSemiDef means positive semi-definite and non-zero; same for NegativeSemiDef + enum SignMatrix { PositiveSemiDef, NegativeSemiDef, ZeroSign, Indefinite }; } /** \ingroup Cholesky_Module @@ -69,7 +72,12 @@ template class LDLT * The default constructor is useful in cases in which the user intends to * perform decompositions via LDLT::compute(const MatrixType&). */ - LDLT() : m_matrix(), m_transpositions(), m_isInitialized(false) {} + LDLT() + : m_matrix(), + m_transpositions(), + m_sign(internal::ZeroSign), + m_isInitialized(false) + {} /** \brief Default Constructor with memory preallocation * @@ -81,6 +89,7 @@ template class LDLT : m_matrix(size, size), m_transpositions(size), m_temporary(size), + m_sign(internal::ZeroSign), m_isInitialized(false) {} @@ -93,6 +102,7 @@ template class LDLT : m_matrix(matrix.rows(), matrix.cols()), m_transpositions(matrix.rows()), m_temporary(matrix.rows()), + m_sign(internal::ZeroSign), m_isInitialized(false) { compute(matrix); @@ -139,7 +149,7 @@ template class LDLT inline bool isPositive() const { eigen_assert(m_isInitialized && "LDLT is not initialized."); - return m_sign == 1; + return m_sign == internal::PositiveSemiDef || m_sign == internal::ZeroSign; } #ifdef EIGEN2_SUPPORT @@ -153,7 +163,7 @@ template class LDLT inline bool isNegative(void) const { eigen_assert(m_isInitialized && "LDLT is not initialized."); - return m_sign == -1; + return m_sign == internal::NegativeSemiDef || m_sign == internal::ZeroSign; } /** \returns a solution x of \f$ A x = b \f$ using the current decomposition of A. @@ -235,7 +245,7 @@ template class LDLT MatrixType m_matrix; TranspositionType m_transpositions; TmpMatrixType m_temporary; - int m_sign; + internal::SignMatrix m_sign; bool m_isInitialized; }; @@ -246,7 +256,7 @@ template struct ldlt_inplace; template<> struct ldlt_inplace { template - static bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, int* sign=0) + static bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, SignMatrix& sign) { using std::abs; typedef typename MatrixType::Scalar Scalar; @@ -258,8 +268,9 @@ template<> struct ldlt_inplace if (size <= 1) { transpositions.setIdentity(); - if(sign) - *sign = numext::real(mat.coeff(0,0))>0 ? 1:-1; + if (numext::real(mat.coeff(0,0)) > 0) sign = PositiveSemiDef; + else if (numext::real(mat.coeff(0,0)) < 0) sign = NegativeSemiDef; + else sign = ZeroSign; return true; } @@ -284,7 +295,6 @@ template<> struct ldlt_inplace if(biggest_in_corner < cutoff) { for(Index i = k; i < size; i++) transpositions.coeffRef(i) = i; - if(sign) *sign = 0; break; } @@ -325,15 +335,15 @@ template<> struct ldlt_inplace } if((rs>0) && (abs(mat.coeffRef(k,k)) > cutoff)) A21 /= mat.coeffRef(k,k); - - if(sign) - { - // LDLT is not guaranteed to work for indefinite matrices, but let's try to get the sign right - int newSign = numext::real(mat.diagonal().coeff(index_of_biggest_in_corner)) > 0; - if(k == 0) - *sign = newSign; - else if(*sign != newSign) - *sign = 0; + + RealScalar realAkk = numext::real(mat.coeffRef(k,k)); + if (sign == PositiveSemiDef) { + if (realAkk < 0) sign = Indefinite; + } else if (sign == NegativeSemiDef) { + if (realAkk > 0) sign = Indefinite; + } else if (sign == ZeroSign) { + if (realAkk > 0) sign = PositiveSemiDef; + else if (realAkk < 0) sign = NegativeSemiDef; } } @@ -399,7 +409,7 @@ template<> struct ldlt_inplace template<> struct ldlt_inplace { template - static EIGEN_STRONG_INLINE bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, int* sign=0) + static EIGEN_STRONG_INLINE bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, SignMatrix& sign) { Transpose matt(mat); return ldlt_inplace::unblocked(matt, transpositions, temp, sign); @@ -445,7 +455,7 @@ LDLT& LDLT::compute(const MatrixType& a) m_isInitialized = false; m_temporary.resize(size); - internal::ldlt_inplace::unblocked(m_matrix, m_transpositions, m_temporary, &m_sign); + internal::ldlt_inplace::unblocked(m_matrix, m_transpositions, m_temporary, m_sign); m_isInitialized = true; return *this; @@ -473,7 +483,7 @@ LDLT& LDLT::rankUpdate(const MatrixBase=0 ? 1 : -1; + m_sign = sigma>=0 ? internal::PositiveSemiDef : internal::NegativeSemiDef; m_isInitialized = true; } diff --git a/test/cholesky.cpp b/test/cholesky.cpp index 378525a83..b980dc572 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -263,8 +263,8 @@ template void cholesky_bug241(const MatrixType& m) // LDLT is not guaranteed to work for indefinite matrices, but happens to work fine if matrix is diagonal. // This test checks that LDLT reports correctly that matrix is indefinite. -// See http://forum.kde.org/viewtopic.php?f=74&t=106942 -template void cholesky_indefinite(const MatrixType& m) +// See http://forum.kde.org/viewtopic.php?f=74&t=106942 and bug 736 +template void cholesky_definiteness(const MatrixType& m) { eigen_assert(m.rows() == 2 && m.cols() == 2); MatrixType mat; @@ -280,6 +280,24 @@ template void cholesky_indefinite(const MatrixType& m) VERIFY(!ldlt.isNegative()); VERIFY(!ldlt.isPositive()); } + { + mat << 0, 0, 0, 0; + LDLT ldlt(mat); + VERIFY(ldlt.isNegative()); + VERIFY(ldlt.isPositive()); + } + { + mat << 0, 0, 0, 1; + LDLT ldlt(mat); + VERIFY(!ldlt.isNegative()); + VERIFY(ldlt.isPositive()); + } + { + mat << -1, 0, 0, 0; + LDLT ldlt(mat); + VERIFY(ldlt.isNegative()); + VERIFY(!ldlt.isPositive()); + } } template void cholesky_verify_assert() @@ -309,7 +327,7 @@ void test_cholesky() CALL_SUBTEST_1( cholesky(Matrix()) ); CALL_SUBTEST_3( cholesky(Matrix2d()) ); CALL_SUBTEST_3( cholesky_bug241(Matrix2d()) ); - CALL_SUBTEST_3( cholesky_indefinite(Matrix2d()) ); + CALL_SUBTEST_3( cholesky_definiteness(Matrix2d()) ); CALL_SUBTEST_4( cholesky(Matrix3f()) ); CALL_SUBTEST_5( cholesky(Matrix4d()) ); s = internal::random(1,EIGEN_TEST_MAX_SIZE); From 9e71ecbeec4df51b4f25985a6fabe5a06604a0d0 Mon Sep 17 00:00:00 2001 From: Naumov Maks Date: Sat, 8 Feb 2014 10:40:51 +0000 Subject: [PATCH 0275/1103] fix typo in evalSolverSugarFunction() --- unsupported/test/polynomialsolver.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/test/polynomialsolver.cpp b/unsupported/test/polynomialsolver.cpp index 13f92169e..3c9265d98 100644 --- a/unsupported/test/polynomialsolver.cpp +++ b/unsupported/test/polynomialsolver.cpp @@ -120,7 +120,7 @@ void evalSolverSugarFunction( const POLYNOMIAL& pols, const ROOTS& roots, const bool found = false; for( size_t j=0; j Date: Sat, 8 Feb 2014 20:31:35 +0000 Subject: [PATCH 0276/1103] Remove unused typedef in polynomialsolver test. --- unsupported/test/polynomialsolver.cpp | 1 - 1 file changed, 1 deletion(-) diff --git a/unsupported/test/polynomialsolver.cpp b/unsupported/test/polynomialsolver.cpp index 3c9265d98..680ff6d31 100644 --- a/unsupported/test/polynomialsolver.cpp +++ b/unsupported/test/polynomialsolver.cpp @@ -106,7 +106,6 @@ void evalSolverSugarFunction( const POLYNOMIAL& pols, const ROOTS& roots, const typedef typename POLYNOMIAL::Scalar Scalar; typedef typename REAL_ROOTS::Scalar Real; - typedef PolynomialSolver PolynomialSolverType; //Test realRoots std::vector< Real > calc_realRoots; From e170e7070babbdb164cbda9289419e264d3c2645 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 11 Feb 2014 17:27:14 +0100 Subject: [PATCH 0277/1103] Added examples for casting, made better examples for Maps --- doc/AsciiQuickReference.txt | 22 +++++++++++++++++----- 1 file changed, 17 insertions(+), 5 deletions(-) diff --git a/doc/AsciiQuickReference.txt b/doc/AsciiQuickReference.txt index d3bfd439c..4c8fe2f47 100644 --- a/doc/AsciiQuickReference.txt +++ b/doc/AsciiQuickReference.txt @@ -163,13 +163,25 @@ x.squaredNorm() // dot(x, x) Note the equivalence is not true for co x.dot(y) // dot(x, y) x.cross(y) // cross(x, y) Requires #include +//// Type conversion +// Eigen // Matlab +A.cast(); // double(A) +A.cast(); // single(A) +A.cast(); // int32(A) +// if the original type equals destination type, no work is done + +// Note that for most operations Eigen requires all operands to have the same type: +MatrixXf F = MatrixXf::Zero(3,3); +A += F; // illegal in Eigen. In Matlab A = A+F is allowed +A += F.cast(); // F converted to double and then added (generally, conversion happens on-the-fly) + // Eigen can map existing memory into Eigen matrices. float array[3]; -Map(array, 3).fill(10); -int data[4] = 1, 2, 3, 4; -Matrix2i mat2x2(data); -MatrixXi mat2x2 = Map(data); -MatrixXi mat2x2 = Map(data, 2, 2); +Vector3f::Map(array).fill(10); // create a temporary Map over array and sets entries to 10 +int data[4] = {1, 2, 3, 4}; +Matrix2i mat2x2(data); // copies data into mat2x2 +Matrix2i::Map(data) = 2*mat2x2; // overwrite elements of data with 2*mat2x2 +MatrixXi::Map(data, 2, 2) += mat2x2; // adds mat2x2 to elements of data (alternative syntax if size is not know at compile time) // Solve Ax = b. Result stored in x. Matlab: x = A \ b. x = A.ldlt().solve(b)); // A sym. p.s.d. #include From 31c63ef0b42ae9d6249281c70a0268ddb4eba10d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Feb 2014 13:37:23 +0100 Subject: [PATCH 0278/1103] fix compilation of Transform * UniformScaling --- Eigen/src/Geometry/Transform.h | 4 ++-- test/geo_transformations.cpp | 7 +++++++ 2 files changed, 9 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 020868103..0a5012cfe 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -530,9 +530,9 @@ public: inline Transform& operator=(const UniformScaling& t); inline Transform& operator*=(const UniformScaling& s) { return scale(s.factor()); } - inline Transform operator*(const UniformScaling& s) const + inline Transform operator*(const UniformScaling& s) const { - Transform res = *this; + Transform res = *this; res.scale(s.factor()); return res; } diff --git a/test/geo_transformations.cpp b/test/geo_transformations.cpp index 35ae67ebe..ee3030b5d 100644 --- a/test/geo_transformations.cpp +++ b/test/geo_transformations.cpp @@ -279,6 +279,13 @@ template void transformations() t1 = Eigen::Scaling(s0,s0,s0) * t1; VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); + t0 = t3; + t0.scale(s0); + t1 = t3 * Eigen::Scaling(s0); + VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); + t0.prescale(s0); + t1 = Eigen::Scaling(s0) * t1; + VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); t0.setIdentity(); t0.prerotate(q1).prescale(v0).pretranslate(v0); From 7ea6ef89693b0d70db29ccad1a3cc7771195e318 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Wed, 12 Feb 2014 14:03:39 +0000 Subject: [PATCH 0279/1103] Fix documentation of MatrixBase::applyOnTheLeft (bug #739) Add examples; move methods from EigenBase.h to MatrixBase.h --- Eigen/src/Core/EigenBase.h | 30 -------------- Eigen/src/Core/MatrixBase.h | 45 +++++++++++++++++++++ doc/snippets/MatrixBase_applyOnTheLeft.cpp | 7 ++++ doc/snippets/MatrixBase_applyOnTheRight.cpp | 9 +++++ 4 files changed, 61 insertions(+), 30 deletions(-) create mode 100644 doc/snippets/MatrixBase_applyOnTheLeft.cpp create mode 100644 doc/snippets/MatrixBase_applyOnTheRight.cpp diff --git a/Eigen/src/Core/EigenBase.h b/Eigen/src/Core/EigenBase.h index a25e823ab..1a577c2dc 100644 --- a/Eigen/src/Core/EigenBase.h +++ b/Eigen/src/Core/EigenBase.h @@ -141,36 +141,6 @@ Derived& DenseBase::operator-=(const EigenBase &other) return derived(); } -/** replaces \c *this by \c *this * \a other. - * - * \returns a reference to \c *this - */ -template -template -inline Derived& -MatrixBase::operator*=(const EigenBase &other) -{ - other.derived().applyThisOnTheRight(derived()); - return derived(); -} - -/** replaces \c *this by \c *this * \a other. It is equivalent to MatrixBase::operator*=(). - */ -template -template -inline void MatrixBase::applyOnTheRight(const EigenBase &other) -{ - other.derived().applyThisOnTheRight(derived()); -} - -/** replaces \c *this by \c *this * \a other. */ -template -template -inline void MatrixBase::applyOnTheLeft(const EigenBase &other) -{ - other.derived().applyThisOnTheLeft(derived()); -} - } // end namespace Eigen #endif // EIGEN_EIGENBASE_H diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 61798317e..49b69f873 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -564,6 +564,51 @@ template class MatrixBase {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;} }; + +/*************************************************************************** +* Implementation of matrix base methods +***************************************************************************/ + +/** replaces \c *this by \c *this * \a other. + * + * \returns a reference to \c *this + * + * Example: \include MatrixBase_applyOnTheRight.cpp + * Output: \verbinclude MatrixBase_applyOnTheRight.out + */ +template +template +inline Derived& +MatrixBase::operator*=(const EigenBase &other) +{ + other.derived().applyThisOnTheRight(derived()); + return derived(); +} + +/** replaces \c *this by \c *this * \a other. It is equivalent to MatrixBase::operator*=(). + * + * Example: \include MatrixBase_applyOnTheRight.cpp + * Output: \verbinclude MatrixBase_applyOnTheRight.out + */ +template +template +inline void MatrixBase::applyOnTheRight(const EigenBase &other) +{ + other.derived().applyThisOnTheRight(derived()); +} + +/** replaces \c *this by \a other * \c *this. + * + * Example: \include MatrixBase_applyOnTheLeft.cpp + * Output: \verbinclude MatrixBase_applyOnTheLeft.out + */ +template +template +inline void MatrixBase::applyOnTheLeft(const EigenBase &other) +{ + other.derived().applyThisOnTheLeft(derived()); +} + } // end namespace Eigen #endif // EIGEN_MATRIXBASE_H diff --git a/doc/snippets/MatrixBase_applyOnTheLeft.cpp b/doc/snippets/MatrixBase_applyOnTheLeft.cpp new file mode 100644 index 000000000..6398c873a --- /dev/null +++ b/doc/snippets/MatrixBase_applyOnTheLeft.cpp @@ -0,0 +1,7 @@ +Matrix3f A = Matrix3f::Random(3,3), B; +B << 0,1,0, + 0,0,1, + 1,0,0; +cout << "At start, A = " << endl << A << endl; +A.applyOnTheLeft(B); +cout << "After applyOnTheLeft, A = " << endl << A << endl; diff --git a/doc/snippets/MatrixBase_applyOnTheRight.cpp b/doc/snippets/MatrixBase_applyOnTheRight.cpp new file mode 100644 index 000000000..e4b71b2d8 --- /dev/null +++ b/doc/snippets/MatrixBase_applyOnTheRight.cpp @@ -0,0 +1,9 @@ +Matrix3f A = Matrix3f::Random(3,3), B; +B << 0,1,0, + 0,0,1, + 1,0,0; +cout << "At start, A = " << endl << A << endl; +A *= B; +cout << "After A *= B, A = " << endl << A << endl; +A.applyOnTheRight(B); // equivalent to A *= B +cout << "After applyOnTheRight, A = " << endl << A << endl; From 14422decc2583e556352408983f4f8fda3054c70 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 13 Feb 2014 09:21:13 +0100 Subject: [PATCH 0280/1103] Fix Fortran compiler detection --- blas/CMakeLists.txt | 3 +++ cmake/language_support.cmake | 2 +- lapack/CMakeLists.txt | 3 +++ 3 files changed, 7 insertions(+), 1 deletion(-) diff --git a/blas/CMakeLists.txt b/blas/CMakeLists.txt index c35a2fdbe..a9bc05137 100644 --- a/blas/CMakeLists.txt +++ b/blas/CMakeLists.txt @@ -7,6 +7,9 @@ workaround_9220(Fortran EIGEN_Fortran_COMPILER_WORKS) if(EIGEN_Fortran_COMPILER_WORKS) enable_language(Fortran OPTIONAL) + if(NOT CMAKE_Fortran_COMPILER) + set(EIGEN_Fortran_COMPILER_WORKS OFF) + endif() endif() add_custom_target(blas) diff --git a/cmake/language_support.cmake b/cmake/language_support.cmake index 2ca303c92..d687b71f6 100644 --- a/cmake/language_support.cmake +++ b/cmake/language_support.cmake @@ -23,7 +23,7 @@ function(workaround_9220 language language_works) #message("DEBUG: language = ${language}") set(text "project(test NONE) - cmake_minimum_required(VERSION 2.6.0) + cmake_minimum_required(VERSION 2.8.0) set (CMAKE_Fortran_FLAGS \"${CMAKE_Fortran_FLAGS}\") set (CMAKE_EXE_LINKER_FLAGS \"${CMAKE_EXE_LINKER_FLAGS}\") enable_language(${language} OPTIONAL) diff --git a/lapack/CMakeLists.txt b/lapack/CMakeLists.txt index 7e7444326..9883d4c72 100644 --- a/lapack/CMakeLists.txt +++ b/lapack/CMakeLists.txt @@ -7,6 +7,9 @@ workaround_9220(Fortran EIGEN_Fortran_COMPILER_WORKS) if(EIGEN_Fortran_COMPILER_WORKS) enable_language(Fortran OPTIONAL) + if(NOT CMAKE_Fortran_COMPILER) + set(EIGEN_Fortran_COMPILER_WORKS OFF) + endif() endif() add_custom_target(lapack) From 32915806305081d837711305bcf57508714d0068 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 13 Feb 2014 15:44:01 +0100 Subject: [PATCH 0281/1103] Fix bug #740: overflow issue in stableNorm --- Eigen/src/Core/StableNorm.h | 27 ++++++++++++++++++++------- test/stable_norm.cpp | 14 +++++++++++--- 2 files changed, 31 insertions(+), 10 deletions(-) diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h index c219e2f53..525620bad 100644 --- a/Eigen/src/Core/StableNorm.h +++ b/Eigen/src/Core/StableNorm.h @@ -17,16 +17,29 @@ namespace internal { template inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale) { - Scalar max = bl.cwiseAbs().maxCoeff(); - if (max>scale) + using std::max; + Scalar maxCoeff = bl.cwiseAbs().maxCoeff(); + + if (maxCoeff>scale) { - ssq = ssq * numext::abs2(scale/max); - scale = max; - invScale = Scalar(1)/scale; + ssq = ssq * numext::abs2(scale/maxCoeff); + Scalar tmp = Scalar(1)/maxCoeff; + if(tmp > NumTraits::highest()) + { + invScale = NumTraits::highest(); + scale = Scalar(1)/invScale; + } + else + { + scale = maxCoeff; + invScale = tmp; + } } - // TODO if the max is much much smaller than the current scale, + + // TODO if the maxCoeff is much much smaller than the current scale, // then we can neglect this sub vector - ssq += (bl*invScale).squaredNorm(); + if(scale>Scalar(0)) // if scale==0, then bl is 0 + ssq += (bl*invScale).squaredNorm(); } template diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp index c09fc17b7..364170acd 100644 --- a/test/stable_norm.cpp +++ b/test/stable_norm.cpp @@ -55,8 +55,16 @@ template void stable_norm(const MatrixType& m) Index rows = m.rows(); Index cols = m.cols(); - Scalar big = internal::random() * ((std::numeric_limits::max)() * RealScalar(1e-4)); - Scalar small = internal::random() * ((std::numeric_limits::min)() * RealScalar(1e4)); + // get a non-zero random factor + Scalar factor = internal::random(); + while(factor(); + Scalar big = factor * ((std::numeric_limits::max)() * RealScalar(1e-4)); + + factor = internal::random(); + while(factor(); + Scalar small = factor * ((std::numeric_limits::min)() * RealScalar(1e4)); MatrixType vzero = MatrixType::Zero(rows, cols), vrand = MatrixType::Random(rows, cols), @@ -91,7 +99,7 @@ template void stable_norm(const MatrixType& m) VERIFY_IS_APPROX(vsmall.blueNorm(), sqrt(size)*abs(small)); VERIFY_IS_APPROX(vsmall.hypotNorm(), sqrt(size)*abs(small)); -// Test compilation of cwise() version + // Test compilation of cwise() version VERIFY_IS_APPROX(vrand.colwise().stableNorm(), vrand.colwise().norm()); VERIFY_IS_APPROX(vrand.colwise().blueNorm(), vrand.colwise().norm()); VERIFY_IS_APPROX(vrand.colwise().hypotNorm(), vrand.colwise().norm()); From 0715d49908e59487e040d49e03319b0a9093b969 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 13 Feb 2014 15:49:54 +0100 Subject: [PATCH 0282/1103] Fix stable_norm unit test for complexes --- test/stable_norm.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp index 364170acd..549f91fbf 100644 --- a/test/stable_norm.cpp +++ b/test/stable_norm.cpp @@ -57,12 +57,12 @@ template void stable_norm(const MatrixType& m) // get a non-zero random factor Scalar factor = internal::random(); - while(factor(); Scalar big = factor * ((std::numeric_limits::max)() * RealScalar(1e-4)); factor = internal::random(); - while(factor(); Scalar small = factor * ((std::numeric_limits::min)() * RealScalar(1e4)); From c0e08e9e4b9307553ce1d5b321c6a1b2590ec56e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 13 Feb 2014 15:53:51 +0100 Subject: [PATCH 0283/1103] fix stable norm benchmark --- bench/bench_norm.cpp | 84 +++++++++++++++++++++++--------------------- 1 file changed, 44 insertions(+), 40 deletions(-) diff --git a/bench/bench_norm.cpp b/bench/bench_norm.cpp index 806db292c..398fef835 100644 --- a/bench/bench_norm.cpp +++ b/bench/bench_norm.cpp @@ -32,25 +32,25 @@ EIGEN_DONT_INLINE typename T::Scalar lapackNorm(T& v) Scalar ssq = 1; for (int i=0;i= ax) { - ssq += internal::abs2(ax/scale); + ssq += numext::abs2(ax/scale); } else { - ssq = Scalar(1) + ssq * internal::abs2(scale/ax); + ssq = Scalar(1) + ssq * numext::abs2(scale/ax); scale = ax; } } - return scale * internal::sqrt(ssq); + return scale * std::sqrt(ssq); } template EIGEN_DONT_INLINE typename T::Scalar twopassNorm(T& v) { typedef typename T::Scalar Scalar; - Scalar s = v.cwise().abs().maxCoeff(); + Scalar s = v.array().abs().maxCoeff(); return s*(v/s).norm(); } @@ -73,16 +73,20 @@ EIGEN_DONT_INLINE typename T::Scalar divacNorm(T& v) v(i) = v(2*i) + v(2*i+1); n = n/2; } - return internal::sqrt(v(0)); + return std::sqrt(v(0)); } +namespace Eigen { +namespace internal { #ifdef EIGEN_VECTORIZE -Packet4f internal::plt(const Packet4f& a, Packet4f& b) { return _mm_cmplt_ps(a,b); } -Packet2d internal::plt(const Packet2d& a, Packet2d& b) { return _mm_cmplt_pd(a,b); } +Packet4f plt(const Packet4f& a, Packet4f& b) { return _mm_cmplt_ps(a,b); } +Packet2d plt(const Packet2d& a, Packet2d& b) { return _mm_cmplt_pd(a,b); } -Packet4f internal::pandnot(const Packet4f& a, Packet4f& b) { return _mm_andnot_ps(a,b); } -Packet2d internal::pandnot(const Packet2d& a, Packet2d& b) { return _mm_andnot_pd(a,b); } +Packet4f pandnot(const Packet4f& a, Packet4f& b) { return _mm_andnot_ps(a,b); } +Packet2d pandnot(const Packet2d& a, Packet2d& b) { return _mm_andnot_pd(a,b); } #endif +} +} template EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) @@ -126,7 +130,7 @@ EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) overfl = rbig*s2m; // overfow boundary for abig eps = std::pow(ibeta, 1-it); - relerr = internal::sqrt(eps); // tolerance for neglecting asml + relerr = std::sqrt(eps); // tolerance for neglecting asml abig = 1.0/eps - 1.0; if (Scalar(nbig)>abig) nmax = abig; // largest safe n else nmax = nbig; @@ -134,13 +138,13 @@ EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) typedef typename internal::packet_traits::type Packet; const int ps = internal::packet_traits::size; - Packet pasml = internal::pset1(Scalar(0)); - Packet pamed = internal::pset1(Scalar(0)); - Packet pabig = internal::pset1(Scalar(0)); - Packet ps2m = internal::pset1(s2m); - Packet ps1m = internal::pset1(s1m); - Packet pb2 = internal::pset1(b2); - Packet pb1 = internal::pset1(b1); + Packet pasml = internal::pset1(Scalar(0)); + Packet pamed = internal::pset1(Scalar(0)); + Packet pabig = internal::pset1(Scalar(0)); + Packet ps2m = internal::pset1(s2m); + Packet ps1m = internal::pset1(s1m); + Packet pb2 = internal::pset1(b2); + Packet pb1 = internal::pset1(b1); for(int j=0; j(j)); @@ -170,7 +174,7 @@ EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) Scalar amed = internal::predux(pamed); if(abig > Scalar(0)) { - abig = internal::sqrt(abig); + abig = std::sqrt(abig); if(abig > overfl) { eigen_assert(false && "overflow"); @@ -179,7 +183,7 @@ EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) if(amed > Scalar(0)) { abig = abig/s2m; - amed = internal::sqrt(amed); + amed = std::sqrt(amed); } else { @@ -191,24 +195,24 @@ EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) { if (amed > Scalar(0)) { - abig = internal::sqrt(amed); - amed = internal::sqrt(asml) / s1m; + abig = std::sqrt(amed); + amed = std::sqrt(asml) / s1m; } else { - return internal::sqrt(asml)/s1m; + return std::sqrt(asml)/s1m; } } else { - return internal::sqrt(amed); + return std::sqrt(amed); } asml = std::min(abig, amed); abig = std::max(abig, amed); if(asml <= abig*relerr) return abig; else - return abig * internal::sqrt(Scalar(1) + internal::abs2(asml/abig)); + return abig * std::sqrt(Scalar(1) + numext::abs2(asml/abig)); #endif } @@ -224,22 +228,22 @@ EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) for (int i=0; i()); - double yd = based * internal::abs(internal::random()); + double yf = basef * std::abs(internal::random()); + double yd = based * std::abs(internal::random()); VectorXf vf = VectorXf::Ones(s) * yf; VectorXd vd = VectorXd::Ones(s) * yd; - std::cout << "reference\t" << internal::sqrt(double(s))*yf << "\t" << internal::sqrt(double(s))*yd << "\n"; + std::cout << "reference\t" << std::sqrt(double(s))*yf << "\t" << std::sqrt(double(s))*yd << "\n"; std::cout << "sqsumNorm\t" << sqsumNorm(vf) << "\t" << sqsumNorm(vd) << "\n"; std::cout << "hypotNorm\t" << hypotNorm(vf) << "\t" << hypotNorm(vd) << "\n"; std::cout << "blueNorm\t" << blueNorm(vf) << "\t" << blueNorm(vd) << "\n"; @@ -255,8 +259,8 @@ void check_accuracy_var(int ef0, int ef1, int ed0, int ed1, int s) VectorXd vd(s); for (int i=0; i()) * std::pow(double(10), internal::random(ef0,ef1)); - vd[i] = internal::abs(internal::random()) * std::pow(double(10), internal::random(ed0,ed1)); + vf[i] = std::abs(internal::random()) * std::pow(double(10), internal::random(ef0,ef1)); + vd[i] = std::abs(internal::random()) * std::pow(double(10), internal::random(ed0,ed1)); } //std::cout << "reference\t" << internal::sqrt(double(s))*yf << "\t" << internal::sqrt(double(s))*yd << "\n"; @@ -321,10 +325,10 @@ int main(int argc, char** argv) VectorXcf vcf = VectorXcf::Random(1024*1024*32) * y; BENCH_PERF(sqsumNorm); BENCH_PERF(blueNorm); -// BENCH_PERF(pblueNorm); -// BENCH_PERF(lapackNorm); -// BENCH_PERF(hypotNorm); -// BENCH_PERF(twopassNorm); + BENCH_PERF(pblueNorm); + BENCH_PERF(lapackNorm); + BENCH_PERF(hypotNorm); + BENCH_PERF(twopassNorm); BENCH_PERF(bl2passNorm); } @@ -336,10 +340,10 @@ int main(int argc, char** argv) VectorXcf vcf = VectorXcf::Random(512) * y; BENCH_PERF(sqsumNorm); BENCH_PERF(blueNorm); -// BENCH_PERF(pblueNorm); -// BENCH_PERF(lapackNorm); -// BENCH_PERF(hypotNorm); -// BENCH_PERF(twopassNorm); + BENCH_PERF(pblueNorm); + BENCH_PERF(lapackNorm); + BENCH_PERF(hypotNorm); + BENCH_PERF(twopassNorm); BENCH_PERF(bl2passNorm); } } From 0b1430ae10a5d6e865424d9fb9e0182bf60b3df8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 13 Feb 2014 23:58:28 +0100 Subject: [PATCH 0284/1103] Fix propagation of index type --- .../ConservativeSparseSparseProduct.h | 20 +++++++++---------- Eigen/src/SparseCore/SparseMatrixBase.h | 4 ++-- Eigen/src/SparseCore/SparseProduct.h | 5 +++-- .../SparseSparseProductWithPruning.h | 11 +++++----- 4 files changed, 21 insertions(+), 19 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 4b13f08d4..cd996d5ba 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -134,8 +134,8 @@ struct conservative_sparse_sparse_product_selector RowMajorMatrix; - typedef SparseMatrix ColMajorMatrix; + typedef SparseMatrix RowMajorMatrix; + typedef SparseMatrix ColMajorMatrix; ColMajorMatrix resCol(lhs.rows(),rhs.cols()); internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol); // sort the non zeros: @@ -149,7 +149,7 @@ struct conservative_sparse_sparse_product_selector RowMajorMatrix; + typedef SparseMatrix RowMajorMatrix; RowMajorMatrix rhsRow = rhs; RowMajorMatrix resRow(lhs.rows(), rhs.cols()); internal::conservative_sparse_sparse_product_impl(rhsRow, lhs, resRow); @@ -162,7 +162,7 @@ struct conservative_sparse_sparse_product_selector RowMajorMatrix; + typedef SparseMatrix RowMajorMatrix; RowMajorMatrix lhsRow = lhs; RowMajorMatrix resRow(lhs.rows(), rhs.cols()); internal::conservative_sparse_sparse_product_impl(rhs, lhsRow, resRow); @@ -175,7 +175,7 @@ struct conservative_sparse_sparse_product_selector RowMajorMatrix; + typedef SparseMatrix RowMajorMatrix; RowMajorMatrix resRow(lhs.rows(), rhs.cols()); internal::conservative_sparse_sparse_product_impl(rhs, lhs, resRow); res = resRow; @@ -190,7 +190,7 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrix; + typedef SparseMatrix ColMajorMatrix; ColMajorMatrix resCol(lhs.rows(), rhs.cols()); internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol); res = resCol; @@ -202,7 +202,7 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrix; + typedef SparseMatrix ColMajorMatrix; ColMajorMatrix lhsCol = lhs; ColMajorMatrix resCol(lhs.rows(), rhs.cols()); internal::conservative_sparse_sparse_product_impl(lhsCol, rhs, resCol); @@ -215,7 +215,7 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrix; + typedef SparseMatrix ColMajorMatrix; ColMajorMatrix rhsCol = rhs; ColMajorMatrix resCol(lhs.rows(), rhs.cols()); internal::conservative_sparse_sparse_product_impl(lhs, rhsCol, resCol); @@ -228,8 +228,8 @@ struct conservative_sparse_sparse_product_selector RowMajorMatrix; - typedef SparseMatrix ColMajorMatrix; + typedef SparseMatrix RowMajorMatrix; + typedef SparseMatrix ColMajorMatrix; RowMajorMatrix resRow(lhs.rows(),rhs.cols()); internal::conservative_sparse_sparse_product_impl(rhs, lhs, resRow); // sort the non zeros: diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index 706f699b8..bbcf7fb1c 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -302,8 +302,8 @@ template class SparseMatrixBase : public EigenBase } else { - SparseMatrix trans = m; - s << static_cast >&>(trans); + SparseMatrix trans = m; + s << static_cast >&>(trans); } } return s; diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index 70b6480ef..cf7663070 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -16,6 +16,7 @@ template struct SparseSparseProductReturnType { typedef typename internal::traits::Scalar Scalar; + typedef typename internal::traits::Index Index; enum { LhsRowMajor = internal::traits::Flags & RowMajorBit, RhsRowMajor = internal::traits::Flags & RowMajorBit, @@ -24,11 +25,11 @@ struct SparseSparseProductReturnType }; typedef typename internal::conditional, + SparseMatrix, typename internal::nested::type>::type LhsNested; typedef typename internal::conditional, + SparseMatrix, typename internal::nested::type>::type RhsNested; typedef SparseSparseProduct Type; diff --git a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h index 70857c7b6..1dccdfe1b 100644 --- a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h +++ b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h @@ -100,7 +100,7 @@ struct sparse_sparse_product_with_pruning_selector SparseTemporaryType; + typedef SparseMatrix SparseTemporaryType; SparseTemporaryType _res(res.rows(), res.cols()); internal::sparse_sparse_product_with_pruning_impl(lhs, rhs, _res, tolerance); res = _res; @@ -126,10 +126,11 @@ struct sparse_sparse_product_with_pruning_selector ColMajorMatrix; - ColMajorMatrix colLhs(lhs); - ColMajorMatrix colRhs(rhs); - internal::sparse_sparse_product_with_pruning_impl(colLhs, colRhs, res, tolerance); + typedef SparseMatrix ColMajorMatrixLhs; + typedef SparseMatrix ColMajorMatrixRhs; + ColMajorMatrixLhs colLhs(lhs); + ColMajorMatrixRhs colRhs(rhs); + internal::sparse_sparse_product_with_pruning_impl(colLhs, colRhs, res, tolerance); // let's transpose the product to get a column x column product // typedef SparseMatrix SparseTemporaryType; From 97965dde9bc079d042d3c4037e1927bee0df443f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 14 Feb 2014 00:04:38 +0100 Subject: [PATCH 0285/1103] alloca is not necessarily alligned on windows --- Eigen/src/Core/util/Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index d177e8b5a..1f30fba1c 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -608,7 +608,7 @@ template class aligned_stack_memory_handler */ #ifdef EIGEN_ALLOCA - #ifdef __arm__ + #if defined(__arm__) || defined(_WIN32) #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast((reinterpret_cast(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16) #else #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA From 6df3bee68751bd0b25c19cf4aae48c41259a4066 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 14 Feb 2014 09:58:30 +0100 Subject: [PATCH 0286/1103] reduce false negative in the qr unit test --- test/qr.cpp | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/test/qr.cpp b/test/qr.cpp index a79e0dd34..98738777f 100644 --- a/test/qr.cpp +++ b/test/qr.cpp @@ -54,6 +54,8 @@ template void qr_invertible() { using std::log; using std::abs; + using std::pow; + using std::max; typedef typename NumTraits::Real RealScalar; typedef typename MatrixType::Scalar Scalar; @@ -65,7 +67,7 @@ template void qr_invertible() if (internal::is_same::value) { // let's build a matrix more stable to inverse - MatrixType a = MatrixType::Random(size,size*2); + MatrixType a = MatrixType::Random(size,size*4); m1 += a * a.adjoint(); } @@ -81,8 +83,11 @@ template void qr_invertible() m3 = qr.householderQ(); // get a unitary m1 = m3 * m1 * m3; qr.compute(m1); - VERIFY_IS_APPROX(absdet, qr.absDeterminant()); VERIFY_IS_APPROX(log(absdet), qr.logAbsDeterminant()); + // This test is tricky if the determinant becomes too small. + // Since we generate random numbers with magnitude rrange [0,1], the average determinant is 0.5^size + VERIFY_IS_MUCH_SMALLER_THAN( abs(absdet-qr.absDeterminant()), (max)(RealScalar(pow(0.5,size)),(max)(abs(absdet),abs(qr.absDeterminant()))) ); + } template void qr_verify_assert() From 0d3f496233ceb0e96da0a39e360e5bdd5c89c0e3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 14 Feb 2014 10:18:14 +0100 Subject: [PATCH 0287/1103] Upload the 3.2 testing result to its own CDash project --- CTestConfig.cmake | 8 ++------ 1 file changed, 2 insertions(+), 6 deletions(-) diff --git a/CTestConfig.cmake b/CTestConfig.cmake index 4c0027824..0557c491a 100644 --- a/CTestConfig.cmake +++ b/CTestConfig.cmake @@ -4,14 +4,10 @@ ## # The following are required to uses Dart and the Cdash dashboard ## ENABLE_TESTING() ## INCLUDE(CTest) -set(CTEST_PROJECT_NAME "Eigen") +set(CTEST_PROJECT_NAME "Eigen3.2") set(CTEST_NIGHTLY_START_TIME "00:00:00 UTC") set(CTEST_DROP_METHOD "http") set(CTEST_DROP_SITE "manao.inria.fr") -set(CTEST_DROP_LOCATION "/CDash/submit.php?project=Eigen") +set(CTEST_DROP_LOCATION "/CDash/submit.php?project=Eigen3.2") set(CTEST_DROP_SITE_CDASH TRUE) -set(CTEST_PROJECT_SUBPROJECTS -Official -Unsupported -) From 3283d98d1378ccd3a8c89eec1d88108ddd517d95 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 14 Feb 2014 14:46:01 +0100 Subject: [PATCH 0288/1103] optimize sparse-sparse Kronecker product --- .../KroneckerProduct/KroneckerTensorProduct.h | 17 +++++++++- unsupported/test/kronecker_product.cpp | 34 +++++++++++++++++++ 2 files changed, 50 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index a4516056d..b8f2cba17 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -153,7 +153,22 @@ void KroneckerProductSparse::evalTo(Dest& dst) const Bc = m_B.cols(); dst.resize(this->rows(), this->cols()); dst.resizeNonZeros(0); - dst.reserve(m_A.nonZeros() * m_B.nonZeros()); + + // compute number of non-zeros per innervectors of dst + { + VectorXi nnzA = VectorXi::Zero(Dest::IsRowMajor ? m_A.rows() : m_A.cols()); + for (Index kA=0; kA < m_A.outerSize(); ++kA) + for (typename Lhs::InnerIterator itA(m_A,kA); itA; ++itA) + nnzA(Dest::IsRowMajor ? itA.row() : itA.col())++; + + VectorXi nnzB = VectorXi::Zero(Dest::IsRowMajor ? m_B.rows() : m_B.cols()); + for (Index kB=0; kB < m_B.outerSize(); ++kB) + for (typename Rhs::InnerIterator itB(m_B,kB); itB; ++itB) + nnzB(Dest::IsRowMajor ? itB.row() : itB.col())++; + + Matrix nnzAB = nnzB * nnzA.transpose(); + dst.reserve(VectorXi::Map(nnzAB.data(), nnzAB.size())); + } for (Index kA=0; kA < m_A.outerSize(); ++kA) { diff --git a/unsupported/test/kronecker_product.cpp b/unsupported/test/kronecker_product.cpp index c68a07de8..753a2d417 100644 --- a/unsupported/test/kronecker_product.cpp +++ b/unsupported/test/kronecker_product.cpp @@ -183,4 +183,38 @@ void test_kronecker_product() DM_b2.resize(4,8); DM_ab2 = kroneckerProduct(DM_a2,DM_b2); CALL_SUBTEST(check_dimension(DM_ab2,10*4,9*8)); + + for(int i = 0; i < g_repeat; i++) + { + double density = Eigen::internal::random(0.01,0.5); + int ra = Eigen::internal::random(1,50); + int ca = Eigen::internal::random(1,50); + int rb = Eigen::internal::random(1,50); + int cb = Eigen::internal::random(1,50); + SparseMatrix sA(ra,ca), sB(rb,cb), sC; + SparseMatrix sC2; + MatrixXf dA(ra,ca), dB(rb,cb), dC; + initSparse(density, dA, sA); + initSparse(density, dB, sB); + + sC = kroneckerProduct(sA,sB); + dC = kroneckerProduct(dA,dB); + VERIFY_IS_APPROX(MatrixXf(sC),dC); + + sC = kroneckerProduct(sA.transpose(),sB); + dC = kroneckerProduct(dA.transpose(),dB); + VERIFY_IS_APPROX(MatrixXf(sC),dC); + + sC = kroneckerProduct(sA.transpose(),sB.transpose()); + dC = kroneckerProduct(dA.transpose(),dB.transpose()); + VERIFY_IS_APPROX(MatrixXf(sC),dC); + + sC = kroneckerProduct(sA,sB.transpose()); + dC = kroneckerProduct(dA,dB.transpose()); + VERIFY_IS_APPROX(MatrixXf(sC),dC); + + sC2 = kroneckerProduct(sA,sB); + dC = kroneckerProduct(dA,dB); + VERIFY_IS_APPROX(MatrixXf(sC2),dC); + } } From cd606bbc943c4862cf0cf596d3ef97aae32c2887 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 14 Feb 2014 23:10:16 +0100 Subject: [PATCH 0289/1103] Fix infinite loop in sparselu --- Eigen/src/SparseLU/SparseLU_Memory.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU_Memory.h b/Eigen/src/SparseLU/SparseLU_Memory.h index d068d62df..1ffa7d54e 100644 --- a/Eigen/src/SparseLU/SparseLU_Memory.h +++ b/Eigen/src/SparseLU/SparseLU_Memory.h @@ -70,7 +70,7 @@ Index SparseLUImpl::expand(VectorType& vec, Index& length, Index if(num_expansions == 0 || keep_prev) new_len = length ; // First time allocate requested else - new_len = Index(alpha * length); + new_len = (std::max)(length+1,Index(alpha * length)); VectorType old_vec; // Temporary vector to hold the previous values if (nbElts > 0 ) @@ -107,7 +107,7 @@ Index SparseLUImpl::expand(VectorType& vec, Index& length, Index do { alpha = (alpha + 1)/2; - new_len = Index(alpha * length); + new_len = (std::max)(length+1,Index(alpha * length)); #ifdef EIGEN_EXCEPTIONS try #endif From 4b6b3f310f5b651d82a65eb546e61d70c09407c5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 15 Feb 2014 09:35:23 +0100 Subject: [PATCH 0290/1103] Fix a few Index to int buggy conversions --- .../SparseCore/ConservativeSparseSparseProduct.h | 14 +++++++------- Eigen/src/SparseCore/SparseDenseProduct.h | 4 ++-- Eigen/src/SparseCore/SparseMatrix.h | 10 +++++----- .../SparseCore/SparseSparseProductWithPruning.h | 2 +- test/sparse.h | 6 +++--- test/sparse_product.cpp | 1 + test/sparse_vector.cpp | 13 +++++++------ 7 files changed, 26 insertions(+), 24 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index cd996d5ba..5c320e2d2 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -66,9 +66,9 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r } // unordered insertion - for(int k=0; k1) std::sort(indices.data(),indices.data()+nnz); - for(int k=0; k::InnerIterator : public _LhsNes inline Scalar value() const { return Base::value() * m_factor; } protected: - int m_outer; + Index m_outer; Scalar m_factor; }; @@ -156,7 +156,7 @@ struct sparse_time_dense_product_impl(std::malloc(m_outerSize * sizeof(Index))); - for (int i = 0; i < m_outerSize; i++) + for (Index i = 0; i < m_outerSize; i++) { m_innerNonZeros[i] = m_outerIndex[i+1] - m_outerIndex[i]; } @@ -752,8 +752,8 @@ class SparseMatrix else for (Index i=0; i::sumupDuplicates() wi.fill(-1); Index count = 0; // for each inner-vector, wi[inner_index] will hold the position of first element into the index/value buffers - for(int j=0; j void +template void initSparse(double density, Matrix& refVec, - SparseVector& sparseVec, + SparseVector& sparseVec, std::vector* zeroCoords = 0, std::vector* nonzeroCoords = 0) { sparseVec.reserve(int(refVec.size()*density)); sparseVec.setZero(); - for(int i=0; i(0,1) < density) ? internal::random() : Scalar(0); if (v!=Scalar(0)) diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index 664e33887..948b0026b 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -244,6 +244,7 @@ void test_sparse_product() CALL_SUBTEST_1( (sparse_product >()) ); CALL_SUBTEST_2( (sparse_product, ColMajor > >()) ); CALL_SUBTEST_2( (sparse_product, RowMajor > >()) ); + CALL_SUBTEST_3( (sparse_product >()) ); CALL_SUBTEST_4( (sparse_product_regression_test, Matrix >()) ); } } diff --git a/test/sparse_vector.cpp b/test/sparse_vector.cpp index ec5877b6a..0c9476803 100644 --- a/test/sparse_vector.cpp +++ b/test/sparse_vector.cpp @@ -9,14 +9,14 @@ #include "sparse.h" -template void sparse_vector(int rows, int cols) +template void sparse_vector(int rows, int cols) { double densityMat = (std::max)(8./(rows*cols), 0.01); double densityVec = (std::max)(8./float(rows), 0.1); typedef Matrix DenseMatrix; typedef Matrix DenseVector; - typedef SparseVector SparseVectorType; - typedef SparseMatrix SparseMatrixType; + typedef SparseVector SparseVectorType; + typedef SparseMatrix SparseMatrixType; Scalar eps = 1e-6; SparseMatrixType m1(rows,rows); @@ -101,9 +101,10 @@ template void sparse_vector(int rows, int cols) void test_sparse_vector() { for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( sparse_vector(8, 8) ); - CALL_SUBTEST_2( sparse_vector >(16, 16) ); - CALL_SUBTEST_1( sparse_vector(299, 535) ); + CALL_SUBTEST_1(( sparse_vector(8, 8) )); + CALL_SUBTEST_2(( sparse_vector, int>(16, 16) )); + CALL_SUBTEST_1(( sparse_vector(299, 535) )); + CALL_SUBTEST_1(( sparse_vector(299, 535) )); } } From 3bb57e21a8e56917e490a514ab7ebfd3cd3811ab Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 17 Feb 2014 09:56:46 +0100 Subject: [PATCH 0291/1103] Fix FFTW unit test with clang --- unsupported/test/FFTW.cpp | 13 +++++-------- 1 file changed, 5 insertions(+), 8 deletions(-) diff --git a/unsupported/test/FFTW.cpp b/unsupported/test/FFTW.cpp index a07bf274b..d3718e2d2 100644 --- a/unsupported/test/FFTW.cpp +++ b/unsupported/test/FFTW.cpp @@ -16,9 +16,6 @@ std::complex RandomCpx() { return std::complex( (T)(rand()/(T)RAND_MAX - . using namespace std; using namespace Eigen; -float norm(float x) {return x*x;} -double norm(double x) {return x*x;} -long double norm(long double x) {return x*x;} template < typename T> complex promote(complex x) { return complex(x.real(),x.imag()); } @@ -40,11 +37,11 @@ complex promote(long double x) { return complex( x); for (size_t k1=0;k1<(size_t)timebuf.size();++k1) { acc += promote( timebuf[k1] ) * exp( complex(0,k1*phinc) ); } - totalpower += norm(acc); + totalpower += numext::abs2(acc); complex x = promote(fftbuf[k0]); complex dif = acc - x; - difpower += norm(dif); - //cerr << k0 << "\t" << acc << "\t" << x << "\t" << sqrt(norm(dif)) << endl; + difpower += numext::abs2(dif); + //cerr << k0 << "\t" << acc << "\t" << x << "\t" << sqrt(numext::abs2(dif)) << endl; } cerr << "rmse:" << sqrt(difpower/totalpower) << endl; return sqrt(difpower/totalpower); @@ -57,8 +54,8 @@ complex promote(long double x) { return complex( x); long double difpower=0; size_t n = (min)( buf1.size(),buf2.size() ); for (size_t k=0;k Date: Mon, 17 Feb 2014 09:57:47 +0100 Subject: [PATCH 0292/1103] Fix sparse_product/sparse_extra unit tests --- test/sparse.h | 4 ++-- test/sparse_product.cpp | 16 +++++++++------- 2 files changed, 11 insertions(+), 9 deletions(-) diff --git a/test/sparse.h b/test/sparse.h index 433d48ec4..e19a76316 100644 --- a/test/sparse.h +++ b/test/sparse.h @@ -178,10 +178,10 @@ initSparse(double density, } } -template void +template void initSparse(double density, Matrix& refVec, - SparseVector& sparseVec, + SparseVector& sparseVec, std::vector* zeroCoords = 0, std::vector* nonzeroCoords = 0) { diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index 948b0026b..a2ea9d5b7 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -13,8 +13,9 @@ template struct test_outer { static void run(SparseMatrixType& m2, SparseMatrixType& m4, DenseMatrix& refMat2, DenseMatrix& refMat4) { - int c = internal::random(0,m2.cols()-1); - int c1 = internal::random(0,m2.cols()-1); + typedef typename SparseMatrixType::Index Index; + Index c = internal::random(0,m2.cols()-1); + Index c1 = internal::random(0,m2.cols()-1); VERIFY_IS_APPROX(m4=m2.col(c)*refMat2.col(c1).transpose(), refMat4=refMat2.col(c)*refMat2.col(c1).transpose()); VERIFY_IS_APPROX(m4=refMat2.col(c1)*m2.col(c).transpose(), refMat4=refMat2.col(c1)*refMat2.col(c).transpose()); } @@ -22,8 +23,9 @@ template struct test_outer struct test_outer { static void run(SparseMatrixType& m2, SparseMatrixType& m4, DenseMatrix& refMat2, DenseMatrix& refMat4) { - int r = internal::random(0,m2.rows()-1); - int c1 = internal::random(0,m2.cols()-1); + typedef typename SparseMatrixType::Index Index; + Index r = internal::random(0,m2.rows()-1); + Index c1 = internal::random(0,m2.cols()-1); VERIFY_IS_APPROX(m4=m2.row(r).transpose()*refMat2.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat2.col(c1).transpose()); VERIFY_IS_APPROX(m4=refMat2.col(c1)*m2.row(r), refMat4=refMat2.col(c1)*refMat2.row(r)); } @@ -37,9 +39,9 @@ template void sparse_product() { typedef typename SparseMatrixType::Index Index; Index n = 100; - const Index rows = internal::random(1,n); - const Index cols = internal::random(1,n); - const Index depth = internal::random(1,n); + const Index rows = internal::random(1,n); + const Index cols = internal::random(1,n); + const Index depth = internal::random(1,n); typedef typename SparseMatrixType::Scalar Scalar; enum { Flags = SparseMatrixType::Flags }; From bd6eca059d1026743188adecebd33ece12a23184 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 17 Feb 2014 10:00:38 +0100 Subject: [PATCH 0293/1103] Fix compilation of SPlines module --- unsupported/Eigen/src/Splines/SplineFwd.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/unsupported/Eigen/src/Splines/SplineFwd.h b/unsupported/Eigen/src/Splines/SplineFwd.h index 5d9bfacf8..9ea23a9a1 100644 --- a/unsupported/Eigen/src/Splines/SplineFwd.h +++ b/unsupported/Eigen/src/Splines/SplineFwd.h @@ -64,6 +64,8 @@ namespace Eigen { enum { OrderAtCompileTime = _Degree==Dynamic ? Dynamic : _Degree+1 /*!< The spline curve's order at compile-time. */ }; enum { NumOfDerivativesAtCompileTime = _DerivativeOrder==Dynamic ? Dynamic : _DerivativeOrder+1 /*!< The number of derivatives defined for the current spline. */ }; + + enum { DerivativeMemoryLayout = _Dim==1 ? RowMajor : ColMajor /*!< The derivative type's memory layout. */ }; /** \brief The data type used to store the values of the basis function derivatives. */ typedef Array<_Scalar,Dynamic,Dynamic,RowMajor,NumOfDerivativesAtCompileTime,OrderAtCompileTime> BasisDerivativeType; From 3573a107120c3762c1f9b6b17f62d817ee1a9ec0 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 17 Feb 2014 13:46:17 +0100 Subject: [PATCH 0294/1103] Fix support for row (resp. column) of a column-major (resp. row-major) sparse matrix --- Eigen/src/SparseCore/SparseBlock.h | 133 +++++++++++++++++++++++------ test/sparse_basic.cpp | 8 ++ 2 files changed, 115 insertions(+), 26 deletions(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index d0436aa7c..4fa98c01c 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -335,6 +335,14 @@ const Block SparseMatrixBase::inner } +namespace internal { + +template< typename XprType, int BlockRows, int BlockCols, bool InnerPanel, + bool OuterVector = (BlockCols==1 && XprType::IsRowMajor) || (BlockRows==1 && !XprType::IsRowMajor)> +class GenericSparseBlockInnerIteratorImpl; + +} + /** Generic implementation of sparse Block expression. * Real-only. */ @@ -354,8 +362,8 @@ public: : m_matrix(xpr), m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0), m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0), - m_blockRows(xpr.rows()), - m_blockCols(xpr.cols()) + m_blockRows(BlockRows==1 ? 1 : xpr.rows()), + m_blockCols(BlockCols==1 ? 1 : xpr.cols()) {} /** Dynamic-size constructor @@ -394,29 +402,8 @@ public: inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } - class InnerIterator : public _MatrixTypeNested::InnerIterator - { - typedef typename _MatrixTypeNested::InnerIterator Base; - const BlockType& m_block; - Index m_end; - public: - - EIGEN_STRONG_INLINE InnerIterator(const BlockType& block, Index outer) - : Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())), - m_block(block), - m_end(IsRowMajor ? block.m_startCol.value()+block.m_blockCols.value() : block.m_startRow.value()+block.m_blockRows.value()) - { - while( (Base::operator bool()) && (Base::index() < (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value())) ) - Base::operator++(); - } - - inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } - inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } - inline Index row() const { return Base::row() - m_block.m_startRow.value(); } - inline Index col() const { return Base::col() - m_block.m_startCol.value(); } - - inline operator bool() const { return Base::operator bool() && Base::index() < m_end; } - }; + typedef internal::GenericSparseBlockInnerIteratorImpl InnerIterator; + class ReverseInnerIterator : public _MatrixTypeNested::ReverseInnerIterator { typedef typename _MatrixTypeNested::ReverseInnerIterator Base; @@ -441,7 +428,7 @@ public: inline operator bool() const { return Base::operator bool() && Base::index() >= m_begin; } }; protected: - friend class InnerIterator; + friend class internal::GenericSparseBlockInnerIteratorImpl; friend class ReverseInnerIterator; EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) @@ -454,6 +441,100 @@ public: }; +namespace internal { + template + class GenericSparseBlockInnerIteratorImpl : public Block::_MatrixTypeNested::InnerIterator + { + typedef Block BlockType; + enum { + IsRowMajor = BlockType::IsRowMajor + }; + typedef typename BlockType::_MatrixTypeNested _MatrixTypeNested; + typedef typename BlockType::Index Index; + typedef typename _MatrixTypeNested::InnerIterator Base; + const BlockType& m_block; + Index m_end; + public: + + EIGEN_STRONG_INLINE GenericSparseBlockInnerIteratorImpl(const BlockType& block, Index outer) + : Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())), + m_block(block), + m_end(IsRowMajor ? block.m_startCol.value()+block.m_blockCols.value() : block.m_startRow.value()+block.m_blockRows.value()) + { + while( (Base::operator bool()) && (Base::index() < (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value())) ) + Base::operator++(); + } + + inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } + inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } + inline Index row() const { return Base::row() - m_block.m_startRow.value(); } + inline Index col() const { return Base::col() - m_block.m_startCol.value(); } + + inline operator bool() const { return Base::operator bool() && Base::index() < m_end; } + }; + + // Row vector of a column-major sparse matrix or column of a row-major one. + template + class GenericSparseBlockInnerIteratorImpl + { + typedef Block BlockType; + enum { + IsRowMajor = BlockType::IsRowMajor + }; + typedef typename BlockType::_MatrixTypeNested _MatrixTypeNested; + typedef typename BlockType::Index Index; + typedef typename BlockType::Scalar Scalar; + const BlockType& m_block; + Index m_outerPos; + Index m_innerIndex; + Scalar m_value; + Index m_end; + public: + + EIGEN_STRONG_INLINE GenericSparseBlockInnerIteratorImpl(const BlockType& block, Index outer = 0) + : + m_block(block), + m_outerPos( (IsRowMajor ? block.m_startCol.value() : block.m_startRow.value()) - 1), // -1 so that operator++ finds the first non-zero entry + m_innerIndex(IsRowMajor ? block.m_startRow.value() : block.m_startCol.value()), + m_end(IsRowMajor ? block.m_startCol.value()+block.m_blockCols.value() : block.m_startRow.value()+block.m_blockRows.value()) + { + EIGEN_UNUSED_VARIABLE(outer); + eigen_assert(outer==0); + + ++(*this); + } + + inline Index index() const { return m_outerPos - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } + inline Index outer() const { return 0; } + inline Index row() const { return IsRowMajor ? 0 : index(); } + inline Index col() const { return IsRowMajor ? index() : 0; } + + inline Scalar value() const { return m_value; } + + inline GenericSparseBlockInnerIteratorImpl& operator++() + { + // search next non-zero entry + while(m_outerPos void sparse_basic(const SparseMatrixType& re VERIFY_IS_APPROX(m1.innerVector(0).dot(refM2.row(0)), refM1.row(0).dot(refM2.row(0))); else VERIFY_IS_APPROX(m1.innerVector(0).dot(refM2.row(0)), refM1.col(0).dot(refM2.row(0))); + + DenseVector rv = DenseVector::Random(m1.cols()); + DenseVector cv = DenseVector::Random(m1.rows()); + Index r = internal::random(0,m1.rows()-2); + Index c = internal::random(0,m1.cols()-1); + VERIFY_IS_APPROX(( m1.template block<1,Dynamic>(r,0,1,m1.cols()).dot(rv)) , refM1.row(r).dot(rv)); + VERIFY_IS_APPROX(m1.row(r).dot(rv), refM1.row(r).dot(rv)); + VERIFY_IS_APPROX(m1.col(c).dot(cv), refM1.col(c).dot(cv)); VERIFY_IS_APPROX(m1.conjugate(), refM1.conjugate()); VERIFY_IS_APPROX(m1.real(), refM1.real()); From b14a4628aff4b92761229c862a331342a2bb518b Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 17 Feb 2014 13:48:00 +0100 Subject: [PATCH 0295/1103] Relaxed umeyama test. Problem was ill-posed if linear part was scaled with very small number. This should fix bug #744. --- test/umeyama.cpp | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/test/umeyama.cpp b/test/umeyama.cpp index 738d0af70..2e8092434 100644 --- a/test/umeyama.cpp +++ b/test/umeyama.cpp @@ -130,10 +130,11 @@ void run_fixed_size_test(int num_elements) // MUST be positive because in any other case det(cR_t) may become negative for // odd dimensions! - const Scalar c = abs(internal::random()); + // Also if c is to small compared to t.norm(), problem is ill-posed (cf. Bug 744) + const Scalar c = internal::random(0.5, 2.0); FixedMatrix R = randMatrixSpecialUnitary(dim); - FixedVector t = Scalar(50)*FixedVector::Random(dim,1); + FixedVector t = Scalar(32)*FixedVector::Random(dim,1); HomMatrix cR_t = HomMatrix::Identity(dim+1,dim+1); cR_t.block(0,0,dim,dim) = c*R; @@ -149,9 +150,9 @@ void run_fixed_size_test(int num_elements) HomMatrix cR_t_umeyama = umeyama(src_block, dst_block); - const Scalar error = ( cR_t_umeyama*src - dst ).array().square().sum(); + const Scalar error = ( cR_t_umeyama*src - dst ).squaredNorm(); - VERIFY(error < Scalar(10)*std::numeric_limits::epsilon()); + VERIFY(error < Scalar(16)*std::numeric_limits::epsilon()); } void test_umeyama() From bffa15142c4271313a70801e6bb7d01365a00bc9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 17 Feb 2014 16:10:55 +0100 Subject: [PATCH 0296/1103] Add evaluator support for diagonal products --- Eigen/src/Core/DiagonalMatrix.h | 36 ++++++- Eigen/src/Core/ProductEvaluators.h | 146 ++++++++++++++++++++++++++++- test/evaluators.cpp | 17 +++- 3 files changed, 192 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index f7ac22f8b..8df636928 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -66,6 +66,7 @@ class DiagonalBase : public EigenBase EIGEN_DEVICE_FUNC inline Index cols() const { return diagonal().size(); } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the diagonal matrix product of \c *this by the matrix \a matrix. */ template @@ -75,6 +76,15 @@ class DiagonalBase : public EigenBase { return DiagonalProduct(matrix.derived(), derived()); } +#else + template + EIGEN_DEVICE_FUNC + const Product + operator*(const MatrixBase &matrix) const + { + return Product(derived(),matrix.derived()); + } +#endif // EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC inline const DiagonalWrapper, const DiagonalVectorType> > @@ -270,7 +280,8 @@ struct traits > ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, MaxRowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, MaxColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, - Flags = traits::Flags & LvalueBit + Flags = traits::Flags & LvalueBit, + CoeffReadCost = traits<_DiagonalVectorType>::CoeffReadCost }; }; } @@ -341,6 +352,29 @@ bool MatrixBase::isDiagonal(const RealScalar& prec) const return true; } +#ifdef EIGEN_ENABLE_EVALUATORS +namespace internal { + +// TODO currently a diagonal expression has the form DiagonalMatrix<> or DiagonalWrapper +// in the future diagonal-ness should be defined by the expression traits +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef DiagonalShape Shape; + static const int AssumeAliasing = 0; +}; +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef DiagonalShape Shape; + static const int AssumeAliasing = 0; +}; + +} // namespace internal +#endif // EIGEN_ENABLE_EVALUATORS + } // end namespace Eigen #endif // EIGEN_DIAGONALMATRIX_H diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 2279ec33b..d991ff8b5 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -309,9 +309,9 @@ struct generic_product_impl }; // This specialization enforces the use of a coefficient-based evaluation strategy -// template -// struct generic_product_impl -// : generic_product_impl {}; +template +struct generic_product_impl + : generic_product_impl {}; // Case 2: Evaluate coeff by coeff // @@ -764,6 +764,146 @@ protected: PlainObject m_result; }; +/*************************************************************************** +* Diagonal products +***************************************************************************/ + +template +struct diagonal_product_evaluator_base + : evaluator_base +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::PacketScalar PacketScalar; +public: + diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag) + : m_diagImpl(diag), m_matImpl(mat) + { + } + + EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const + { + return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx); + } + +protected: + template + EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::true_type) const + { + return internal::pmul(m_matImpl.template packet(row, col), + internal::pset1(m_diagImpl.coeff(id))); + } + + template + EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::false_type) const + { + enum { + InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime, + DiagonalPacketLoadMode = (LoadMode == Aligned && (((InnerSize%16) == 0) || (int(DiagonalType::Flags)&AlignedBit)==AlignedBit) ? Aligned : Unaligned) + }; + return internal::pmul(m_matImpl.template packet(row, col), + m_diagImpl.template packet(id)); + } + + typename evaluator::nestedType m_diagImpl; + typename evaluator::nestedType m_matImpl; +}; + +// diagonal * dense +template +struct product_evaluator, ProductTag, DiagonalShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : diagonal_product_evaluator_base > +{ + typedef diagonal_product_evaluator_base > Base; + using Base::m_diagImpl; + using Base::m_matImpl; + using Base::coeff; + using Base::packet_impl; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + typedef typename Base::PacketScalar PacketScalar; + + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + + product_evaluator(const XprType& xpr) + : Base(xpr.rhs(), xpr.lhs().diagonal()) + { + } + + EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const + { + return m_diagImpl.coeff(row) * m_matImpl.coeff(row, col); + } + + template + EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const + { + enum { + StorageOrder = Rhs::Flags & RowMajorBit ? RowMajor : ColMajor + }; + return this->template packet_impl(row,col, row, + typename internal::conditional::type()); + } + + template + EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const + { + enum { + StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor + }; + return packet(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); + } + +}; + +// dense * diagonal +template +struct product_evaluator, ProductTag, DenseShape, DiagonalShape, typename Lhs::Scalar, typename Rhs::Scalar> + : diagonal_product_evaluator_base > +{ + typedef diagonal_product_evaluator_base > Base; + using Base::m_diagImpl; + using Base::m_matImpl; + using Base::coeff; + using Base::packet_impl; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + typedef typename Base::PacketScalar PacketScalar; + + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + + product_evaluator(const XprType& xpr) + : Base(xpr.lhs(), xpr.rhs().diagonal()) + { + } + + EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const + { + return m_matImpl.coeff(row, col) * m_diagImpl.coeff(col); + } + + template + EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const + { + enum { + StorageOrder = Rhs::Flags & RowMajorBit ? RowMajor : ColMajor + }; + return this->template packet_impl(row,col, col, + typename internal::conditional::type()); + } + + template + EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const + { + enum { + StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor + }; + return packet(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); + } + +}; } // end namespace internal diff --git a/test/evaluators.cpp b/test/evaluators.cpp index 69a45661f..305047a6a 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -151,19 +151,19 @@ void test_evaluators() c = a*a; copy_using_evaluator(a, prod(a,a)); VERIFY_IS_APPROX(a,c); - + // check compound assignment of products d = c; add_assign_using_evaluator(c.noalias(), prod(a,b)); d.noalias() += a*b; VERIFY_IS_APPROX(c, d); - + d = c; subtract_assign_using_evaluator(c.noalias(), prod(a,b)); d.noalias() -= a*b; VERIFY_IS_APPROX(c, d); } - + { // test product with all possible sizes int s = internal::random(1,100); @@ -458,4 +458,15 @@ void test_evaluators() VERIFY_IS_APPROX_EVALUATOR2(B, prod(A.selfadjointView(),A), MatrixXd(A.selfadjointView()*A)); } + + { + // test diagonal shapes + VectorXd d = VectorXd::Random(6); + MatrixXd A = MatrixXd::Random(6,6), B(6,6); + A.setRandom();B.setRandom(); + + VERIFY_IS_APPROX_EVALUATOR2(B, lazyprod(d.asDiagonal(),A), MatrixXd(d.asDiagonal()*A)); + VERIFY_IS_APPROX_EVALUATOR2(B, lazyprod(A,d.asDiagonal()), MatrixXd(A*d.asDiagonal())); + + } } From d595fd31f544009e62a55d6ffc26e0b62f3147d5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 17 Feb 2014 16:11:55 +0100 Subject: [PATCH 0297/1103] Deal with automatic transposition in call_assignment, fix a few shortcomings --- Eigen/src/Core/Assign.h | 27 ++++++++- Eigen/src/Core/AssignEvaluator.h | 61 ++++++++++++++++++--- Eigen/src/Core/PlainObjectBase.h | 19 ++++++- Eigen/src/Core/TriangularMatrix.h | 2 + Eigen/src/Core/products/CoeffBasedProduct.h | 4 +- 5 files changed, 102 insertions(+), 11 deletions(-) diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index cefa6f3cc..6080a83f6 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -531,6 +531,23 @@ EIGEN_STRONG_INLINE Derived& DenseBase namespace internal { +#ifdef EIGEN_TEST_EVALUATORS + +// TODO remove this class which is now useless + +template +struct assign_selector { + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { + call_assignment(dst, other.derived(), internal::assign_op()); + return dst; + } + template + EIGEN_DEVICE_FUNC + static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; } +}; + +#else // EIGEN_TEST_EVALUATORS template::Flags) & EvalBeforeAssigningBit) != 0, bool NeedToTranspose = ((int(Derived::RowsAtCompileTime) == 1 && int(OtherDerived::ColsAtCompileTime) == 1) @@ -566,7 +583,7 @@ struct assign_selector { EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); } }; - +#endif // EIGEN_TEST_EVALUATORS } // end namespace internal template @@ -604,7 +621,11 @@ template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const EigenBase& other) { +#ifdef EIGEN_TEST_EVALUATORS + return internal::assign_selector::evalTo(derived(), other.derived()); +#else return internal::assign_selector::evalTo(derived(), other.derived()); +#endif } template @@ -612,7 +633,11 @@ template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const ReturnByValue& other) { +#ifdef EIGEN_TEST_EVALUATORS + return internal::assign_selector::evalTo(derived(), other.derived()); +#else return internal::assign_selector::evalTo(derived(), other.derived()); +#endif } } // end namespace Eigen diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index fe6fbdb9a..0a3475d83 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -656,28 +656,75 @@ template< typename DstXprType, typename SrcXprType, typename Functor, struct Assignment; -// The only purpose of this call_assignment() function is to deal with noalias() / AssumeAliasing. +// The only purpose of this call_assignment() function is to deal with noalias() / AssumeAliasing and automatic transposition. // Indeed, I (Gael) think that this concept of AssumeAliasing was a mistake, and it makes thing quite complicated. // So this intermediate function removes everything related to AssumeAliasing such that Assignment // does not has to bother about these annoying details. -template -void call_assignment(Dst& dst, const Src& src, const Func& func) +template struct transpose_to_match { - typedef typename internal::conditional::AssumeAliasing==1, EvalToTemp, Src>::type ActualSrc; - Assignment::run(dst, src, func); + enum { + NeedToTranspose = ( (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1) + | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". + // revert to || as soon as not needed anymore. + (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1)) + && int(Dst::SizeAtCompileTime) != 1 + }; + + typedef typename internal::conditional, Dst>::type type; +}; + +template +void call_assignment(Dst& dst, const Src& src) +{ + call_assignment(dst, src, internal::assign_op()); +} +template +void call_assignment(const Dst& dst, const Src& src) +{ + call_assignment(dst, src, internal::assign_op()); +} + +// Deal with AssumeAliasing +template +void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) +{ + // The following initial implementation through an EvalToTemp object does not permit to + // perform deferred resizing as in 'A = A * B' when the size of 'A' as to be changed + // typedef typename internal::conditional::AssumeAliasing==1, EvalToTemp, Src>::type ActualSrc; + // Assignment::run(dst, src, func); + + // TODO we should simply do tmp(src); +#ifdef EIGEN_TEST_EVALUATORS + typename Src::PlainObject tmp(src); +#else + typename Src::PlainObject tmp(src.rows(), src.cols()); + call_assignment(tmp.noalias(), src); +#endif + + // resizing + dst.resize(tmp.rows(), tmp.cols()); + call_assignment(dst.noalias(), tmp, func); + TRACK; +} + +template +void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==0, void*>::type = 0) +{ + Assignment::type,Src,Func>::run(dst, src, func); } // by-pass AssumeAliasing +// FIXME the const version should probably not be needed template class StorageBase, typename Src, typename Func> void call_assignment(const NoAlias& dst, const Src& src, const Func& func) { - Assignment::run(dst.expression(), src, func); + Assignment::type,Src,Func>::run(dst.expression(), src, func); } template class StorageBase, typename Src, typename Func> void call_assignment(NoAlias& dst, const Src& src, const Func& func) { - Assignment::run(dst.expression(), src, func); + Assignment::type,Src,Func>::run(dst.expression(), src, func); } // Generic Dense to Dense assignment diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 0305066ba..8eccbfbd0 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -639,6 +639,18 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * \internal */ +#ifdef EIGEN_TEST_EVALUATORS + // aliasing is dealt once in internall::call_assignment + // so at this stage we have to assume aliasing... and resising has to be done later. + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& _set(const DenseBase& other) + { + internal::call_assignment(this->derived(), other.derived()); + return this->derived(); + return this->derived(); + } +#else template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& _set(const DenseBase& other) @@ -654,7 +666,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::false_type&) { _set_noalias(other); } - +#endif /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which * is the case when creating a new matrix) so one can enforce lazy evaluation. * @@ -669,7 +681,12 @@ class PlainObjectBase : public internal::dense_xpr_base::type //_resize_to_match(other); // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because // it wouldn't allow to copy a row-vector into a column-vector. +#ifdef EIGEN_TEST_EVALUATORS + internal::call_assignment(this->noalias(), other.derived()); + return this->derived(); +#else return internal::assign_selector::run(this->derived(), other.derived()); +#endif } template diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index d413ec2e9..2b1c236b9 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -1235,6 +1235,7 @@ struct triangular_assignment_loop } // end namespace internal +#ifdef EIGEN_TEST_EVALUATORS /** Assigns a triangular or selfadjoint matrix to a dense matrix. * If the matrix is triangular, the opposite part is set to zero. */ template @@ -1244,6 +1245,7 @@ void TriangularBase::evalToLazy(MatrixBase &other) const other.derived().resize(this->rows(), this->cols()); internal::call_triangular_assignment_loop(other.derived(), derived().nestedExpression()); } +#endif #endif // EIGEN_ENABLE_EVALUATORS diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index 5f8182aa9..de06108da 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -47,8 +47,8 @@ struct traits > typename traits<_RhsNested>::Index>::type Index; enum { - LhsCoeffReadCost = _LhsNested::CoeffReadCost, - RhsCoeffReadCost = _RhsNested::CoeffReadCost, + LhsCoeffReadCost = traits<_LhsNested>::CoeffReadCost, + RhsCoeffReadCost = traits<_RhsNested>::CoeffReadCost, LhsFlags = traits<_LhsNested>::Flags, RhsFlags = traits<_RhsNested>::Flags, From 873401032bf91e092413cdd7b56848f7b485490a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 17 Feb 2014 19:00:45 +0100 Subject: [PATCH 0298/1103] Fix scalar * product optimization when 'product' includes a selfadjoint matrix --- Eigen/src/Core/SelfAdjointView.h | 7 +++++++ Eigen/src/Core/TriangularMatrix.h | 8 ++++++++ 2 files changed, 15 insertions(+) diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index dd8eb25f8..2cc1815fd 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -132,6 +132,13 @@ template class SelfAdjointView { return Product(lhs.derived(),rhs); } + + friend EIGEN_DEVICE_FUNC + const SelfAdjointView,MatrixType>,UpLo> + operator*(const Scalar& s, const SelfAdjointView& mat) + { + return (s*mat.nestedExpression()).template selfadjointView(); + } #else // EIGEN_TEST_EVALUATORS diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 2b1c236b9..c5f137ab8 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -62,6 +62,14 @@ template class TriangularBase : public EigenBase inline Index outerStride() const { return derived().outerStride(); } EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().innerStride(); } + + // dummy resize function + void resize(Index nbRows, Index nbCols) + { + EIGEN_UNUSED_VARIABLE(nbRows); + EIGEN_UNUSED_VARIABLE(nbCols); + eigen_assert(nbRows==rows() && nbCols==nbCols); + } EIGEN_DEVICE_FUNC inline Scalar coeff(Index row, Index col) const { return derived().coeff(row,col); } From 2d136d3d7f0623189e42be44d45e1353d1cde93e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 10:52:00 +0100 Subject: [PATCH 0299/1103] Get rid of SeflCwiseBinaryOp --- Eigen/src/Core/ArrayBase.h | 54 ++++++++++++++++++++++++++++++ Eigen/src/Core/SelfCwiseBinaryOp.h | 48 ++++++++++++++++++++++++++ 2 files changed, 102 insertions(+) diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index 2c9ace4a7..c528de733 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -177,6 +177,59 @@ template class ArrayBase {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;} }; +#ifdef EIGEN_TEST_EVALUATORS +/** replaces \c *this by \c *this - \a other. + * + * \returns a reference to \c *this + */ +template +template +EIGEN_STRONG_INLINE Derived & +ArrayBase::operator-=(const ArrayBase &other) +{ + call_assignment(derived(), other.derived(), internal::sub_assign_op()); + return derived(); +} + +/** replaces \c *this by \c *this + \a other. + * + * \returns a reference to \c *this + */ +template +template +EIGEN_STRONG_INLINE Derived & +ArrayBase::operator+=(const ArrayBase& other) +{ + call_assignment(derived(), other.derived(), internal::add_assign_op()); + return derived(); +} + +/** replaces \c *this by \c *this * \a other coefficient wise. + * + * \returns a reference to \c *this + */ +template +template +EIGEN_STRONG_INLINE Derived & +ArrayBase::operator*=(const ArrayBase& other) +{ + call_assignment(derived(), other.derived(), internal::mul_assign_op()); + return derived(); +} + +/** replaces \c *this by \c *this / \a other coefficient wise. + * + * \returns a reference to \c *this + */ +template +template +EIGEN_STRONG_INLINE Derived & +ArrayBase::operator/=(const ArrayBase& other) +{ + call_assignment(derived(), other.derived(), internal::div_assign_op()); + return derived(); +} +#else // EIGEN_TEST_EVALUATORS /** replaces \c *this by \c *this - \a other. * * \returns a reference to \c *this @@ -232,6 +285,7 @@ ArrayBase::operator/=(const ArrayBase& other) tmp = other.derived(); return derived(); } +#endif } // end namespace Eigen diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h index 65864adf8..ae7f9b887 100644 --- a/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -12,6 +12,8 @@ namespace Eigen { +#ifndef EIGEN_TEST_EVALUATORS + /** \class SelfCwiseBinaryOp * \ingroup Core_Module * @@ -179,6 +181,51 @@ template class SelfCwiseBinaryOp SelfCwiseBinaryOp& operator=(const SelfCwiseBinaryOp&); }; +#endif // EIGEN_TEST_EVALUATORS + +#ifdef EIGEN_TEST_EVALUATORS +template +inline Derived& DenseBase::operator*=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op()); + return derived(); +} + +template +inline Derived& ArrayBase::operator+=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op()); + return derived(); +} + +template +inline Derived& ArrayBase::operator-=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op()); + return derived(); +} + +template +inline Derived& DenseBase::operator/=(const Scalar& other) +{ + typedef typename Derived::PlainObject PlainObject; + + typedef typename internal::conditional::IsInteger, + internal::div_assign_op, + internal::mul_assign_op >::type AssignOp; + + Scalar actual_other; + if(NumTraits::IsInteger) actual_other = other; + else actual_other = Scalar(1)/other; + + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),actual_other), AssignOp()); + + return derived(); +} +#else template inline Derived& DenseBase::operator*=(const Scalar& other) { @@ -220,6 +267,7 @@ inline Derived& DenseBase::operator/=(const Scalar& other) tmp = PlainObject::Constant(rows(),cols(), actual_other); return derived(); } +#endif } // end namespace Eigen From 551bf5c66a3f29c8ac1f24f8439a6b85d0da79ac Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 10:52:26 +0100 Subject: [PATCH 0300/1103] Get rid of DiagonalProduct --- Eigen/src/Core/DiagonalProduct.h | 12 ++++++++++++ Eigen/src/Core/MatrixBase.h | 7 +++++++ 2 files changed, 19 insertions(+) diff --git a/Eigen/src/Core/DiagonalProduct.h b/Eigen/src/Core/DiagonalProduct.h index c03a0c2e1..f5539df13 100644 --- a/Eigen/src/Core/DiagonalProduct.h +++ b/Eigen/src/Core/DiagonalProduct.h @@ -13,6 +13,7 @@ namespace Eigen { +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template struct traits > @@ -124,6 +125,17 @@ MatrixBase::operator*(const DiagonalBase &a_diagonal) { return DiagonalProduct(derived(), a_diagonal.derived()); } +#else // EIGEN_TEST_EVALUATORS +/** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal. + */ +template +template +inline const Product +MatrixBase::operator*(const DiagonalBase &a_diagonal) const +{ + return Product(derived(),a_diagonal.derived()); +} +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 37e7408a4..6453145d0 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -215,10 +215,17 @@ template class MatrixBase template void applyOnTheRight(const EigenBase& other); +#ifndef EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC const DiagonalProduct operator*(const DiagonalBase &diagonal) const; +#else // EIGEN_TEST_EVALUATORS + template + EIGEN_DEVICE_FUNC + const Product + operator*(const DiagonalBase &diagonal) const; +#endif // EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC From 573c587e3d66c8151f6523433d853664e1fa7667 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 10:53:14 +0100 Subject: [PATCH 0301/1103] New design for handling automatic transposition --- Eigen/src/Core/AssignEvaluator.h | 61 ++++++++++++++++++-------------- Eigen/src/Core/PlainObjectBase.h | 2 +- 2 files changed, 36 insertions(+), 27 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 0a3475d83..d5745f20c 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -661,34 +661,21 @@ struct Assignment; // So this intermediate function removes everything related to AssumeAliasing such that Assignment // does not has to bother about these annoying details. -template struct transpose_to_match -{ - enum { - NeedToTranspose = ( (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1) - | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". - // revert to || as soon as not needed anymore. - (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1)) - && int(Dst::SizeAtCompileTime) != 1 - }; - - typedef typename internal::conditional, Dst>::type type; -}; - template void call_assignment(Dst& dst, const Src& src) -{ +{TRACK; call_assignment(dst, src, internal::assign_op()); } template void call_assignment(const Dst& dst, const Src& src) -{ +{TRACK; call_assignment(dst, src, internal::assign_op()); } // Deal with AssumeAliasing template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) -{ +{TRACK; // The following initial implementation through an EvalToTemp object does not permit to // perform deferred resizing as in 'A = A * B' when the size of 'A' as to be changed // typedef typename internal::conditional::AssumeAliasing==1, EvalToTemp, Src>::type ActualSrc; @@ -699,40 +686,62 @@ void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable typename Src::PlainObject tmp(src); #else typename Src::PlainObject tmp(src.rows(), src.cols()); - call_assignment(tmp.noalias(), src); + call_assignment_no_alias(tmp, src, internal::assign_op()); #endif // resizing dst.resize(tmp.rows(), tmp.cols()); - call_assignment(dst.noalias(), tmp, func); - TRACK; + call_assignment_no_alias(dst, tmp, func); } template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==0, void*>::type = 0) -{ - Assignment::type,Src,Func>::run(dst, src, func); +{TRACK; + // There is no explicit no-aliasing, so we must resize here: + dst.resize(src.rows(), src.cols()); + call_assignment_no_alias(dst, src, func); } // by-pass AssumeAliasing // FIXME the const version should probably not be needed +// When there is no aliasing, we require that 'dst' has been properly resized template class StorageBase, typename Src, typename Func> void call_assignment(const NoAlias& dst, const Src& src, const Func& func) -{ - Assignment::type,Src,Func>::run(dst.expression(), src, func); +{TRACK; + call_assignment_no_alias(dst.expression(), src, func); } template class StorageBase, typename Src, typename Func> void call_assignment(NoAlias& dst, const Src& src, const Func& func) -{ - Assignment::type,Src,Func>::run(dst.expression(), src, func); +{TRACK; + call_assignment_no_alias(dst.expression(), src, func); } + +template +void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) +{TRACK; + enum { + NeedToTranspose = ( (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1) + | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". + // revert to || as soon as not needed anymore. + (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1)) + && int(Dst::SizeAtCompileTime) != 1 + }; + + typedef typename internal::conditional, Dst>::type ActualDstTypeCleaned; + typedef typename internal::conditional, Dst&>::type ActualDstType; + ActualDstType actualDst(dst); + + Assignment::run(actualDst, src, func); +} + + // Generic Dense to Dense assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> struct Assignment { static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) - { + {TRACK; // TODO check whether this is the right place to perform these checks: enum{ SameType = internal::is_same::value diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 8eccbfbd0..4f456804c 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -682,7 +682,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because // it wouldn't allow to copy a row-vector into a column-vector. #ifdef EIGEN_TEST_EVALUATORS - internal::call_assignment(this->noalias(), other.derived()); + internal::call_assignment_no_alias(this->derived(), other.derived(), internal::assign_op()); return this->derived(); #else return internal::assign_selector::run(this->derived(), other.derived()); From a08cba6b5f51db2338da58d54cda7e04cc24e372 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 11:03:59 +0100 Subject: [PATCH 0302/1103] Move is_diagonal to XprHelper, forward declare Ref --- Eigen/Core | 6 +++--- Eigen/src/Core/Ref.h | 4 ---- Eigen/src/Core/Stride.h | 4 ++-- Eigen/src/Core/util/ForwardDeclarations.h | 5 +++++ Eigen/src/Core/util/Meta.h | 12 ------------ Eigen/src/Core/util/XprHelper.h | 12 ++++++++++++ 6 files changed, 22 insertions(+), 21 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index cdc2f2d46..c1d168ec4 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -277,8 +277,8 @@ using std::ptrdiff_t; */ #include "src/Core/util/Constants.h" -#include "src/Core/util/ForwardDeclarations.h" #include "src/Core/util/Meta.h" +#include "src/Core/util/ForwardDeclarations.h" #include "src/Core/util/StaticAssert.h" #include "src/Core/util/XprHelper.h" #include "src/Core/util/Memory.h" @@ -340,12 +340,12 @@ using std::ptrdiff_t; #include "src/Core/SelfCwiseBinaryOp.h" #include "src/Core/Dot.h" #include "src/Core/StableNorm.h" -#include "src/Core/MapBase.h" #include "src/Core/Stride.h" +#include "src/Core/MapBase.h" #include "src/Core/Map.h" +#include "src/Core/Ref.h" #include "src/Core/Block.h" #include "src/Core/VectorBlock.h" -#include "src/Core/Ref.h" #include "src/Core/Transpose.h" #include "src/Core/DiagonalMatrix.h" #include "src/Core/Diagonal.h" diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index 00d9e6d2b..ce9e82913 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -12,10 +12,6 @@ namespace Eigen { -template class RefBase; -template,OuterStride<> >::type > class Ref; - /** \class Ref * \ingroup Core_Module * diff --git a/Eigen/src/Core/Stride.h b/Eigen/src/Core/Stride.h index d3d454e4e..187774978 100644 --- a/Eigen/src/Core/Stride.h +++ b/Eigen/src/Core/Stride.h @@ -86,7 +86,7 @@ class Stride /** \brief Convenience specialization of Stride to specify only an inner stride * See class Map for some examples */ -template +template class InnerStride : public Stride<0, Value> { typedef Stride<0, Value> Base; @@ -98,7 +98,7 @@ class InnerStride : public Stride<0, Value> /** \brief Convenience specialization of Stride to specify only an outer stride * See class Map for some examples */ -template +template class OuterStride : public Stride { typedef Stride Base; diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 3bc151229..db0e2b033 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -121,7 +121,12 @@ template::has_write_access ? WriteAccessors : ReadOnlyAccessors > class MapBase; template class Stride; +template class InnerStride; +template class OuterStride; template > class Map; +template class RefBase; +template,OuterStride<> >::type > class Ref; template class TriangularBase; template class TriangularView; diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h index e4e4d4a87..559928a92 100644 --- a/Eigen/src/Core/util/Meta.h +++ b/Eigen/src/Core/util/Meta.h @@ -246,18 +246,6 @@ template struct scalar_product_traits, T> // typedef typename scalar_product_traits::type, typename remove_all::type>::ReturnType type; // }; -template struct is_diagonal -{ enum { ret = false }; }; - -template struct is_diagonal > -{ enum { ret = true }; }; - -template struct is_diagonal > -{ enum { ret = true }; }; - -template struct is_diagonal > -{ enum { ret = true }; }; - } // end namespace internal namespace numext { diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index f210344d3..7ea70450f 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -502,6 +502,18 @@ struct is_lvalue bool(traits::Flags & LvalueBit) }; }; +template struct is_diagonal +{ enum { ret = false }; }; + +template struct is_diagonal > +{ enum { ret = true }; }; + +template struct is_diagonal > +{ enum { ret = true }; }; + +template struct is_diagonal > +{ enum { ret = true }; }; + } // end namespace internal } // end namespace Eigen From 1b5de5a37b47ef2738e5bdf4135777f00ac5967d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 13:30:16 +0100 Subject: [PATCH 0303/1103] Add evaluator for Ref --- Eigen/src/Core/CoreEvaluators.h | 23 ++++++++++++++++++----- 1 file changed, 18 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index ef900e236..d02bac0a8 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -587,7 +587,7 @@ struct evaluator > CoeffReturnType coeff(Index index) const { return coeff(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + RowsAtCompileTime == 1 ? index : 0); } Scalar& coeffRef(Index row, Index col) @@ -598,7 +598,7 @@ struct evaluator > Scalar& coeffRef(Index index) { return coeffRef(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + RowsAtCompileTime == 1 ? index : 0); } template @@ -612,7 +612,7 @@ struct evaluator > PacketReturnType packet(Index index) const { return packet(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + RowsAtCompileTime == 1 ? index : 0); } template @@ -626,8 +626,8 @@ struct evaluator > void writePacket(Index index, const PacketScalar& x) { return writePacket(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0, - x); + RowsAtCompileTime == 1 ? index : 0, + x); } protected: @@ -647,6 +647,19 @@ struct evaluator > { } }; +// -------------------- Ref -------------------- + +template +struct evaluator > + : public evaluator > > +{ + typedef Ref XprType; + + evaluator(const XprType& map) + : evaluator >(map) + { } +}; + // -------------------- Block -------------------- template Date: Tue, 18 Feb 2014 13:31:44 +0100 Subject: [PATCH 0304/1103] Finally, the simplest remains to deffer resizing at the latest --- Eigen/src/Core/AssignEvaluator.h | 10 +++------- Eigen/src/Core/NoAlias.h | 8 +++----- 2 files changed, 6 insertions(+), 12 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index d5745f20c..f6e1abccd 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -676,12 +676,7 @@ void call_assignment(const Dst& dst, const Src& src) template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) {TRACK; - // The following initial implementation through an EvalToTemp object does not permit to - // perform deferred resizing as in 'A = A * B' when the size of 'A' as to be changed - // typedef typename internal::conditional::AssumeAliasing==1, EvalToTemp, Src>::type ActualSrc; - // Assignment::run(dst, src, func); - // TODO we should simply do tmp(src); #ifdef EIGEN_TEST_EVALUATORS typename Src::PlainObject tmp(src); #else @@ -697,8 +692,6 @@ void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==0, void*>::type = 0) {TRACK; - // There is no explicit no-aliasing, so we must resize here: - dst.resize(src.rows(), src.cols()); call_assignment_no_alias(dst, src, func); } @@ -728,6 +721,9 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) && int(Dst::SizeAtCompileTime) != 1 }; + dst.resize(NeedToTranspose ? src.cols() : src.rows(), + NeedToTranspose ? src.rows() : src.cols()); + typedef typename internal::conditional, Dst>::type ActualDstTypeCleaned; typedef typename internal::conditional, Dst&>::type ActualDstType; ActualDstType actualDst(dst); diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index 412e37258..fe6dded60 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -40,9 +40,7 @@ class NoAlias EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase& other) { - // TODO either call resize here or call "call_assignment" through m_expression.lazyAssign() ?? - m_expression.resize(other.derived().rows(), other.derived().cols()); - call_assignment(*this, other.derived(), internal::assign_op()); + call_assignment_no_alias(m_expression, other.derived(), internal::assign_op()); return m_expression; } @@ -50,7 +48,7 @@ class NoAlias EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase& other) { - call_assignment(*this, other.derived(), internal::add_assign_op()); + call_assignment_no_alias(m_expression, other.derived(), internal::add_assign_op()); return m_expression; } @@ -58,7 +56,7 @@ class NoAlias EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase& other) { - call_assignment(*this, other.derived(), internal::sub_assign_op()); + call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op()); return m_expression; } From 7002aa858f3bd42e98ae1c2317aefe39338fab3e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 13:32:30 +0100 Subject: [PATCH 0305/1103] Support Product::coeff(0,0) even for dynamic matrices --- Eigen/src/Core/Product.h | 26 ++++++++++++++++---------- Eigen/src/Core/ProductEvaluators.h | 22 +++++++--------------- 2 files changed, 23 insertions(+), 25 deletions(-) diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index d64fbae35..12726ed12 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -109,16 +109,6 @@ public: { return typename internal::evaluator::type(derived()).coeff(0,0); } - - Scalar coeff(Index row, Index col) const - { - return typename internal::evaluator::type(derived()).coeff(row,col); - } - - Scalar coeff(Index i) const - { - return typename internal::evaluator::type(derived()).coeff(i); - } }; } // namespace internal @@ -132,6 +122,22 @@ class ProductImpl typedef typename internal::dense_product_base Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + + Scalar coeff(Index row, Index col) const + { + EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) + eigen_assert(this->rows() == 1 && this->cols() == 1); + + return typename internal::evaluator::type(derived()).coeff(row,col); + } + + Scalar coeff(Index i) const + { + EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) + eigen_assert(this->rows() == 1 && this->cols() == 1); + + return typename internal::evaluator::type(derived()).coeff(i); + } }; /*************************************************************************** diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index d991ff8b5..e0cfb56bb 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -270,7 +270,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) - { + {TRACK; // TODO bypass GeneralProduct class GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); } @@ -705,7 +705,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) - {// SelfadjointProductMatrix + { // TODO bypass SelfadjointProductMatrix class SelfadjointProductMatrix(lhs.nestedExpression(),rhs).scaleAndAddTo(dst, alpha); } @@ -739,7 +739,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) - {//SelfadjointProductMatrix + { // TODO bypass SelfadjointProductMatrix class SelfadjointProductMatrix(lhs,rhs.nestedExpression()).scaleAndAddTo(dst, alpha); } @@ -825,6 +825,8 @@ struct product_evaluator, ProductTag, DiagonalSha typedef Product XprType; typedef typename XprType::PlainObject PlainObject; + + enum { StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) @@ -839,9 +841,6 @@ struct product_evaluator, ProductTag, DiagonalSha template EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const { - enum { - StorageOrder = Rhs::Flags & RowMajorBit ? RowMajor : ColMajor - }; return this->template packet_impl(row,col, row, typename internal::conditional::type()); } @@ -849,9 +848,6 @@ struct product_evaluator, ProductTag, DiagonalSha template EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const { - enum { - StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor - }; return packet(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); } @@ -873,6 +869,8 @@ struct product_evaluator, ProductTag, DenseShape, typedef Product XprType; typedef typename XprType::PlainObject PlainObject; + + enum { StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal()) @@ -887,9 +885,6 @@ struct product_evaluator, ProductTag, DenseShape, template EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const { - enum { - StorageOrder = Rhs::Flags & RowMajorBit ? RowMajor : ColMajor - }; return this->template packet_impl(row,col, col, typename internal::conditional::type()); } @@ -897,9 +892,6 @@ struct product_evaluator, ProductTag, DenseShape, template EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const { - enum { - StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor - }; return packet(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); } From 06545058bbdeec4713b9b31c5018335660796076 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 13:33:04 +0100 Subject: [PATCH 0306/1103] Temporary workaround for permutations --- Eigen/src/Core/PermutationMatrix.h | 1 + 1 file changed, 1 insertion(+) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 1297b8413..6a26af3a5 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -532,6 +532,7 @@ namespace internal { template struct traits > + : traits { typedef typename MatrixType::PlainObject ReturnType; }; From 37a1d736bfd6fdaa0e1a93aee585d1861ab4004a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 13:35:24 +0100 Subject: [PATCH 0307/1103] _MatrixTypeNested must be public in sparse Block --- Eigen/src/SparseCore/SparseBlock.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 4fa98c01c..6f615e250 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -350,11 +350,12 @@ template class BlockImpl : public SparseMatrixBase >, internal::no_assignment_operator { - typedef typename internal::remove_all::type _MatrixTypeNested; typedef Block BlockType; public: enum { IsRowMajor = internal::traits::IsRowMajor }; EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) + + typedef typename internal::remove_all::type _MatrixTypeNested; /** Column or Row constructor */ From 99e27916cfd1381aab50850611905335c288ee40 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 14:26:25 +0100 Subject: [PATCH 0308/1103] Fix all()/any() for evaluators --- Eigen/src/Core/BooleanRedux.h | 46 ++++++++++++++++++++++++++++++++++- 1 file changed, 45 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h index be9f48a8c..f0b68f470 100644 --- a/Eigen/src/Core/BooleanRedux.h +++ b/Eigen/src/Core/BooleanRedux.h @@ -17,10 +17,18 @@ namespace internal { template struct all_unroller { +#ifdef EIGEN_TEST_EVALUATORS + typedef typename Derived::ExpressionTraits Traits; + enum { + col = (UnrollCount-1) / Traits::RowsAtCompileTime, + row = (UnrollCount-1) % Traits::RowsAtCompileTime + }; +#else enum { col = (UnrollCount-1) / Derived::RowsAtCompileTime, row = (UnrollCount-1) % Derived::RowsAtCompileTime }; +#endif static inline bool run(const Derived &mat) { @@ -43,11 +51,19 @@ struct all_unroller template struct any_unroller { +#ifdef EIGEN_TEST_EVALUATORS + typedef typename Derived::ExpressionTraits Traits; + enum { + col = (UnrollCount-1) / Traits::RowsAtCompileTime, + row = (UnrollCount-1) % Traits::RowsAtCompileTime + }; +#else enum { col = (UnrollCount-1) / Derived::RowsAtCompileTime, row = (UnrollCount-1) % Derived::RowsAtCompileTime }; - +#endif + static inline bool run(const Derived &mat) { return any_unroller::run(mat) || mat.coeff(row, col); @@ -84,6 +100,19 @@ inline bool DenseBase::all() const && NumTraits::AddCost != Dynamic && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT }; +#ifdef EIGEN_TEST_EVALUATORS + typedef typename internal::evaluator::type Evaluator; + Evaluator evaluator(derived()); + if(unroll) + return internal::all_unroller::run(evaluator); + else + { + for(Index j = 0; j < cols(); ++j) + for(Index i = 0; i < rows(); ++i) + if (!evaluator.coeff(i, j)) return false; + return true; + } +#else if(unroll) return internal::all_unroller::run(derived()); else @@ -93,6 +122,7 @@ inline bool DenseBase::all() const if (!coeff(i, j)) return false; return true; } +#endif } /** \returns true if at least one coefficient is true @@ -108,6 +138,19 @@ inline bool DenseBase::any() const && NumTraits::AddCost != Dynamic && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT }; +#ifdef EIGEN_TEST_EVALUATORS + typedef typename internal::evaluator::type Evaluator; + Evaluator evaluator(derived()); + if(unroll) + return internal::any_unroller::run(evaluator); + else + { + for(Index j = 0; j < cols(); ++j) + for(Index i = 0; i < rows(); ++i) + if (evaluator.coeff(i, j)) return true; + return false; + } +#else if(unroll) return internal::any_unroller::run(derived()); else @@ -117,6 +160,7 @@ inline bool DenseBase::any() const if (coeff(i, j)) return true; return false; } +#endif } /** \returns the number of coefficients which evaluate to true From 0543cb51b58bfc5130bcb0b83969f2854f30d58e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 14:51:41 +0100 Subject: [PATCH 0309/1103] Product::coeff method are also OK for lazy products (including diagonal products) --- Eigen/src/Core/Product.h | 21 ++++++++++++++++----- Eigen/src/Core/util/StaticAssert.h | 1 + 2 files changed, 17 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 12726ed12..31730983f 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -117,27 +117,38 @@ template class ProductImpl : public internal::dense_product_base { - typedef Product Derived; + typedef Product Derived; + public: typedef typename internal::dense_product_base Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + protected: + enum { + IsOneByOne = (RowsAtCompileTime == 1 || RowsAtCompileTime == Dynamic) && + (ColsAtCompileTime == 1 || ColsAtCompileTime == Dynamic), + EnableCoeff = IsOneByOne || Option==LazyProduct + }; + public: + Scalar coeff(Index row, Index col) const { - EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) - eigen_assert(this->rows() == 1 && this->cols() == 1); + EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); + eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); return typename internal::evaluator::type(derived()).coeff(row,col); } Scalar coeff(Index i) const { - EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) - eigen_assert(this->rows() == 1 && this->cols() == 1); + EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); + eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); return typename internal::evaluator::type(derived()).coeff(i); } + + }; /*************************************************************************** diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h index b9b1ee02a..f5691826e 100644 --- a/Eigen/src/Core/util/StaticAssert.h +++ b/Eigen/src/Core/util/StaticAssert.h @@ -84,6 +84,7 @@ THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY, YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT, THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS, + THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS, THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL, THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES, YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED, From 82c066b3c4a0d64962ad0bf268bf7a4a28b2704b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 15:44:32 +0100 Subject: [PATCH 0310/1103] Cleaning --- Eigen/src/Core/AssignEvaluator.h | 16 ++++++++-------- Eigen/src/Core/ProductEvaluators.h | 2 +- 2 files changed, 9 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index f6e1abccd..cf4db8de5 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -663,19 +663,19 @@ struct Assignment; template void call_assignment(Dst& dst, const Src& src) -{TRACK; +{ call_assignment(dst, src, internal::assign_op()); } template void call_assignment(const Dst& dst, const Src& src) -{TRACK; +{ call_assignment(dst, src, internal::assign_op()); } // Deal with AssumeAliasing template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) -{TRACK; +{ #ifdef EIGEN_TEST_EVALUATORS typename Src::PlainObject tmp(src); @@ -691,7 +691,7 @@ void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==0, void*>::type = 0) -{TRACK; +{ call_assignment_no_alias(dst, src, func); } @@ -700,19 +700,19 @@ void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable // When there is no aliasing, we require that 'dst' has been properly resized template class StorageBase, typename Src, typename Func> void call_assignment(const NoAlias& dst, const Src& src, const Func& func) -{TRACK; +{ call_assignment_no_alias(dst.expression(), src, func); } template class StorageBase, typename Src, typename Func> void call_assignment(NoAlias& dst, const Src& src, const Func& func) -{TRACK; +{ call_assignment_no_alias(dst.expression(), src, func); } template void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) -{TRACK; +{ enum { NeedToTranspose = ( (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1) | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". @@ -737,7 +737,7 @@ template< typename DstXprType, typename SrcXprType, typename Functor, typename S struct Assignment { static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) - {TRACK; + { // TODO check whether this is the right place to perform these checks: enum{ SameType = internal::is_same::value diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index e0cfb56bb..113bcd618 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -270,7 +270,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) - {TRACK; + { // TODO bypass GeneralProduct class GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); } From 8169c6ac59ce78a7340ebae40f8eabe9df1c8f62 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 16:57:25 +0100 Subject: [PATCH 0311/1103] Simplify implementation of coeff-based products to fully exploit our reduxion mechanisms. If this results in performance regressions, then we should optimize reduxion rather than somehow duplicate the code. --- Eigen/src/Core/ProductEvaluators.h | 180 +++++------------------------ 1 file changed, 27 insertions(+), 153 deletions(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 113bcd618..67d0dc80c 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -333,8 +333,11 @@ struct product_evaluator, ProductTag, DenseShape, typedef CoeffBasedProduct CoeffBasedProductType; product_evaluator(const XprType& xpr) - : m_lhsImpl(xpr.lhs()), - m_rhsImpl(xpr.rhs()), + : m_lhs(xpr.lhs()), + m_rhs(xpr.rhs()), + m_lhsImpl(m_lhs), // FIXME the creation of the evaluator objects should result in a no-op, but check that! + m_rhsImpl(m_rhs), // Moreover, they are only useful for the packet path, so we could completely disable them when not needed, + // or perhaps declare them on the fly on the packet method... We have experiment to check what's best. m_innerDim(xpr.lhs().cols()) { } @@ -355,31 +358,32 @@ struct product_evaluator, ProductTag, DenseShape, CanVectorizeInner = traits::CanVectorizeInner, Flags = traits::Flags }; - - typedef typename evaluator::type LhsEtorType; - typedef typename evaluator::type RhsEtorType; - typedef etor_product_coeff_impl CoeffImpl; + typedef typename internal::nested_eval::type LhsNested; + typedef typename internal::nested_eval::type RhsNested; + + typedef typename internal::remove_all::type LhsNestedCleaned; + typedef typename internal::remove_all::type RhsNestedCleaned; + typedef typename evaluator::type LhsEtorType; + typedef typename evaluator::type RhsEtorType; + const CoeffReturnType coeff(Index row, Index col) const { - Scalar res; - CoeffImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res); - return res; + // TODO check performance regression wrt to Eigen 3.2 which has special handling of this function + return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum(); } /* Allow index-based non-packet access. It is impossible though to allow index-based packed access, * which is why we don't set the LinearAccessBit. + * TODO: this seems possible when the result is a vector */ const CoeffReturnType coeff(Index index) const { - Scalar res; const Index row = RowsAtCompileTime == 1 ? 0 : index; const Index col = RowsAtCompileTime == 1 ? index : 0; - CoeffImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res); - return res; + // TODO check performance regression wrt to Eigen 3.2 which has special handling of this function + return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum(); } template @@ -389,13 +393,17 @@ struct product_evaluator, ProductTag, DenseShape, typedef etor_product_packet_impl PacketImpl; + PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res); return res; } protected: - typename evaluator::type m_lhsImpl; - typename evaluator::type m_rhsImpl; + const LhsNested m_lhs; + const RhsNested m_rhs; + + LhsEtorType m_lhsImpl; + RhsEtorType m_rhsImpl; // TODO: Get rid of m_innerDim if known at compile time Index m_innerDim; @@ -413,143 +421,9 @@ struct product_evaluator, LazyCoeffBasedProduc {} }; -/*************************************************************************** -* Normal product .coeff() implementation (with meta-unrolling) -***************************************************************************/ - -/************************************** -*** Scalar path - no vectorization *** -**************************************/ - -template -struct etor_product_coeff_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, RetScalar &res) - { - etor_product_coeff_impl::run(row, col, lhs, rhs, innerDim, res); - res += lhs.coeff(row, UnrollingIndex) * rhs.coeff(UnrollingIndex, col); - } -}; - -template -struct etor_product_coeff_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, RetScalar &res) - { - res = lhs.coeff(row, 0) * rhs.coeff(0, col); - } -}; - -template -struct etor_product_coeff_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, RetScalar& res) - { - eigen_assert(innerDim>0 && "you are using a non initialized matrix"); - res = lhs.coeff(row, 0) * rhs.coeff(0, col); - for(Index i = 1; i < innerDim; ++i) - res += lhs.coeff(row, i) * rhs.coeff(i, col); - } -}; - -/******************************************* -*** Scalar path with inner vectorization *** -*******************************************/ - -template -struct etor_product_coeff_vectorized_unroller -{ - typedef typename Lhs::Index Index; - enum { PacketSize = packet_traits::size }; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, typename Lhs::PacketScalar &pres) - { - etor_product_coeff_vectorized_unroller::run(row, col, lhs, rhs, innerDim, pres); - pres = padd(pres, pmul( lhs.template packet(row, UnrollingIndex) , rhs.template packet(UnrollingIndex, col) )); - } -}; - -template -struct etor_product_coeff_vectorized_unroller<0, Lhs, Rhs, Packet> -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, typename Lhs::PacketScalar &pres) - { - pres = pmul(lhs.template packet(row, 0) , rhs.template packet(0, col)); - } -}; - -template -struct etor_product_coeff_impl -{ - typedef typename Lhs::PacketScalar Packet; - typedef typename Lhs::Index Index; - enum { PacketSize = packet_traits::size }; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, RetScalar &res) - { - Packet pres; - etor_product_coeff_vectorized_unroller::run(row, col, lhs, rhs, innerDim, pres); - res = predux(pres); - } -}; - -template -struct etor_product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, typename Lhs::Scalar &res) - { - res = lhs.row(row).transpose().cwiseProduct(rhs.col(col)).sum(); - } -}; - -// NOTE the 3 following specializations are because taking .col(0) on a vector is a bit slower -// NOTE maybe they are now useless since we have a specialization for Block -template -struct etor_product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index /*row*/, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, typename Lhs::Scalar &res) - { - res = lhs.transpose().cwiseProduct(rhs.col(col)).sum(); - } -}; - -template -struct etor_product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index /*col*/, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, typename Lhs::Scalar &res) - { - res = lhs.row(row).transpose().cwiseProduct(rhs).sum(); - } -}; - -template -struct etor_product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, typename Lhs::Scalar &res) - { - res = lhs.transpose().cwiseProduct(rhs).sum(); - } -}; - -template -struct etor_product_coeff_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, typename Lhs::Scalar &res) - { - etor_product_coeff_vectorized_dyn_selector::run(row, col, lhs, rhs, innerDim, res); - } -}; - -/******************* -*** Packet path *** -*******************/ +/**************************************** +*** Coeff based product, Packet path *** +****************************************/ template struct etor_product_packet_impl From 5b78780deffd4a5728cfc6ae5371f56e038e2ac0 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 18 Feb 2014 17:43:16 +0100 Subject: [PATCH 0312/1103] Add evaluator shortcut for triangular ?= product --- Eigen/src/Core/TriangularMatrix.h | 70 ++++++++++++++++++- .../products/GeneralMatrixMatrixTriangular.h | 12 +++- test/product_notemporary.cpp | 3 +- 3 files changed, 79 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index c5f137ab8..ffa2faeaa 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -233,6 +233,24 @@ template class TriangularView EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.innerStride(); } +#ifdef EIGEN_TEST_EVALUATORS + + /** \sa MatrixBase::operator+=() */ + template + EIGEN_DEVICE_FUNC + TriangularView& operator+=(const DenseBase& other) { + internal::call_assignment_no_alias(*this, other.derived(), internal::add_assign_op()); + return *this; + } + /** \sa MatrixBase::operator-=() */ + template + EIGEN_DEVICE_FUNC + TriangularView& operator-=(const DenseBase& other) { + internal::call_assignment_no_alias(*this, other.derived(), internal::sub_assign_op()); + return *this; + } + +#else /** \sa MatrixBase::operator+=() */ template EIGEN_DEVICE_FUNC @@ -241,6 +259,7 @@ template class TriangularView template EIGEN_DEVICE_FUNC TriangularView& operator-=(const DenseBase& other) { return *this = m_matrix - other.derived(); } +#endif /** \sa MatrixBase::operator*=() */ EIGEN_DEVICE_FUNC TriangularView& operator*=(const typename internal::traits::Scalar& other) { return *this = m_matrix * other; } @@ -527,12 +546,18 @@ template class TriangularView #endif // EIGEN_TEST_EVALUATORS - protected: +#ifdef EIGEN_TEST_EVALUATORS + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TriangularView& _assignProduct(const ProductType& prod, const Scalar& alpha); + protected: +#else + protected: template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& assignProduct(const ProductBase& prod, const Scalar& alpha); - +#endif MatrixTypeNested m_matrix; }; @@ -720,7 +745,7 @@ template inline TriangularView& TriangularView::operator=(const MatrixBase& other) { - internal::call_assignment(*this, other.template triangularView(), internal::assign_op()); + internal::call_assignment_no_alias(*this, other.derived(), internal::assign_op()); return *this; } @@ -1253,6 +1278,45 @@ void TriangularBase::evalToLazy(MatrixBase &other) const other.derived().resize(this->rows(), this->cols()); internal::call_triangular_assignment_loop(other.derived(), derived().nestedExpression()); } + +namespace internal { + +// Triangular = Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment, internal::assign_op, Dense2Triangular, Scalar> +{ + typedef Product SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dst.setZero(); + dst._assignProduct(src, 1); + } +}; + +// Triangular += Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment, internal::add_assign_op, Dense2Triangular, Scalar> +{ + typedef Product SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op &) + { + dst._assignProduct(src, 1); + } +}; + +// Triangular -= Product +template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar> +struct Assignment, internal::sub_assign_op, Dense2Triangular, Scalar> +{ + typedef Product SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op &) + { + dst._assignProduct(src, -1); + } +}; + + +} // end namespace internal #endif #endif // EIGEN_ENABLE_EVALUATORS diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 87bd36bc3..46be46f85 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -264,6 +264,16 @@ struct general_product_to_triangular_selector } }; +#ifdef EIGEN_TEST_EVALUATORS +template +template +TriangularView& TriangularView::_assignProduct(const ProductType& prod, const Scalar& alpha) +{ + general_product_to_triangular_selector::InnerSize==1>::run(m_matrix.const_cast_derived(), prod, alpha); + + return *this; +} +#else template template TriangularView& TriangularView::assignProduct(const ProductBase& prod, const Scalar& alpha) @@ -272,7 +282,7 @@ TriangularView& TriangularView::assignProduct( return *this; } - +#endif } // end namespace Eigen #endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp index 258d238e2..deb494c96 100644 --- a/test/product_notemporary.cpp +++ b/test/product_notemporary.cpp @@ -114,8 +114,7 @@ template void product_notemporary(const MatrixType& m) VERIFY_EVALUATION_COUNT( Scalar tmp = 0; tmp += Scalar(RealScalar(1)) / (m3.transpose() * m3).diagonal().array().abs().sum(), 0 ); // Zero temporaries for ... CoeffBasedProductMode - // - does not work with GCC because of the <..>, we'ld need variadic macros ... - //VERIFY_EVALUATION_COUNT( m3.col(0).head<5>() * m3.col(0).transpose() + m3.col(0).head<5>() * m3.col(0).transpose(), 0 ); + VERIFY_EVALUATION_COUNT( m3.col(0).template head<5>() * m3.col(0).transpose() + m3.col(0).template head<5>() * m3.col(0).transpose(), 0 ); // Check matrix * vectors VERIFY_EVALUATION_COUNT( cvres.noalias() = m1 * cv1, 0 ); From 7ed9441ea472031bb9357b5bec80151cae7ed2cb Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 18 Feb 2014 18:06:44 -0800 Subject: [PATCH 0313/1103] Reverted the definition of the EIGEN_ALIGN to its former meaning (i.e. a boolean) Created a new EIGEN_ALIGN_BYTES define to encode how the data should be aligned Fixed a few remaining alignment issues exposed when the Eigen code is compiled with avx enabled. Created a new EIGEN_ALIGN_DEFAULT define, which is set to the minimum alignment value required for the chosen instruction set. Use this value instead of EIGEN_ALIGN32 to preserve the existing alignment on SSE/Altivec/Neon. --- Eigen/src/Core/Block.h | 2 +- Eigen/src/Core/DenseStorage.h | 4 +-- Eigen/src/Core/GeneralProduct.h | 2 +- Eigen/src/Core/Map.h | 2 +- Eigen/src/Core/MapBase.h | 2 +- Eigen/src/Core/products/GeneralMatrixMatrix.h | 6 ++-- Eigen/src/Core/products/GeneralMatrixVector.h | 4 +-- Eigen/src/Core/util/Macros.h | 18 ++++++++--- Eigen/src/Core/util/Memory.h | 32 +++++++++---------- Eigen/src/Core/util/XprHelper.h | 2 +- test/geo_parametrizedline.cpp | 6 ++-- test/mapped_matrix.cpp | 6 ++-- test/packetmath.cpp | 28 ++++++++-------- 13 files changed, 61 insertions(+), 53 deletions(-) diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 31cd5c72c..e948e14aa 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -83,7 +83,7 @@ struct traits > : traits::size) == 0) && (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0, - MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % 16) == 0)) ? AlignedBit : 0, + MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) ? AlignedBit : 0, FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1) ? LinearAccessBit : 0, FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 2342b08a1..7264b44c7 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -40,7 +40,7 @@ void check_static_allocation_size() */ template struct plain_array { @@ -81,7 +81,7 @@ struct plain_array #endif template -struct plain_array +struct plain_array { EIGEN_USER_ALIGN32 T array[Size]; diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index e3a165ac6..adda6f784 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -397,7 +397,7 @@ struct gemv_static_vector_if internal::plain_array m_data; EIGEN_STRONG_INLINE Scalar* data() { return ForceAlignment - ? reinterpret_cast((reinterpret_cast(m_data.array) & ~(size_t(15))) + 16) + ? reinterpret_cast((reinterpret_cast(m_data.array) & ~(size_t(EIGEN_ALIGN_BYTES-1))) + EIGEN_ALIGN_BYTES) : m_data.array; } #endif diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index 8ea13cfb7..c75a5e95f 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -88,7 +88,7 @@ struct traits > && ( bool(IsDynamicSize) || HasNoOuterStride || ( OuterStrideAtCompileTime!=Dynamic - && ((static_cast(sizeof(Scalar))*OuterStrideAtCompileTime)%16)==0 ) ), + && ((static_cast(sizeof(Scalar))*OuterStrideAtCompileTime)%EIGEN_ALIGN_BYTES)==0 ) ), Flags0 = TraitsBase::Flags & (~NestByRefBit), Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit), Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime)) diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index ffa1371c2..a45a0b374 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -164,7 +164,7 @@ template class MapBase EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(internal::traits::Flags&PacketAccessBit, internal::inner_stride_at_compile_time::ret==1), PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1); - eigen_assert(EIGEN_IMPLIES(internal::traits::Flags&AlignedBit, (size_t(m_data) % 16) == 0) + eigen_assert(EIGEN_IMPLIES(internal::traits::Flags&AlignedBit, (size_t(m_data) % EIGEN_ALIGN_BYTES) == 0) && "data is not aligned"); } diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index eb399a824..3dfd239c1 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -286,9 +286,9 @@ class gemm_blocking_space(tmp0); const LhsScalar* lhs0 = lhs + i*lhsStride; // process first unaligned result's coeffs diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 787f800b8..d6d5bfa23 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -66,17 +66,22 @@ #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0 #endif +// Defined the boundary (in bytes) on which the data needs to be aligned. Note +// that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be +// aligned at all regardless of the value of this #define. +#define EIGEN_ALIGN_BYTES 16 + #ifdef EIGEN_DONT_ALIGN #ifndef EIGEN_DONT_ALIGN_STATICALLY #define EIGEN_DONT_ALIGN_STATICALLY #endif #define EIGEN_ALIGN 0 -#else - #if !defined(EIGEN_DONT_VECTORIZE) && defined(__AVX__) - #define EIGEN_ALIGN 32 - #else - #define EIGEN_ALIGN 16 +#elif !defined(EIGEN_DONT_VECTORIZE) + #if defined(__AVX__) + #undef EIGEN_ALIGN_BYTES + #define EIGEN_ALIGN_BYTES 32 #endif + #define EIGEN_ALIGN 1 #endif // EIGEN_ALIGN_STATICALLY is the true test whether we want to align arrays on the stack or not. It takes into account both the user choice to explicitly disable @@ -286,15 +291,18 @@ #define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16) #define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32) +#define EIGEN_ALIGN_DEFAULT EIGEN_ALIGN_TO_BOUNDARY(EIGEN_ALIGN_BYTES) #if EIGEN_ALIGN_STATICALLY #define EIGEN_USER_ALIGN_TO_BOUNDARY(n) EIGEN_ALIGN_TO_BOUNDARY(n) #define EIGEN_USER_ALIGN16 EIGEN_ALIGN16 #define EIGEN_USER_ALIGN32 EIGEN_ALIGN32 +#define EIGEN_USER_ALIGN_DEFAULT EIGEN_ALIGN_DEFAULT #else #define EIGEN_USER_ALIGN_TO_BOUNDARY(n) #define EIGEN_USER_ALIGN16 #define EIGEN_USER_ALIGN32 +#define EIGEN_USER_ALIGN_DEFAULT #endif #ifdef EIGEN_DONT_USE_RESTRICT_KEYWORD diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 76bdb6cfc..2f2398bbf 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -32,7 +32,7 @@ // page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed // quite safe, at least within the context of glibc, to equate 64-bit with LP64. #if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \ - && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_ALIGN == 16) + && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_ALIGN_BYTES == 16) #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1 #else #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0 @@ -42,14 +42,14 @@ // See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup // FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures // See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup -#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__) && (EIGEN_ALIGN == 16) +#if defined(__FreeBSD__) && !defined(__arm__) && !defined(__mips__) && (EIGEN_ALIGN_BYTES == 16) #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1 #else #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0 #endif -#if (defined(__APPLE__) && (EIGEN_ALIGN == 16)) \ - || (defined(_WIN64) && (EIGEN_ALIGN == 16)) \ +#if (defined(__APPLE__) && (EIGEN_ALIGN_BYTES == 16)) \ + || (defined(_WIN64) && (EIGEN_ALIGN_BYTES == 16)) \ || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED \ || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED #define EIGEN_MALLOC_ALREADY_ALIGNED 1 @@ -105,9 +105,9 @@ inline void throw_std_bad_alloc() */ inline void* handmade_aligned_malloc(std::size_t size) { - void *original = std::malloc(size+EIGEN_ALIGN); + void *original = std::malloc(size+EIGEN_ALIGN_BYTES); if (original == 0) return 0; - void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(EIGEN_ALIGN-1))) + EIGEN_ALIGN); + void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(EIGEN_ALIGN_BYTES-1))) + EIGEN_ALIGN_BYTES); *(reinterpret_cast(aligned) - 1) = original; return aligned; } @@ -128,9 +128,9 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = if (ptr == 0) return handmade_aligned_malloc(size); void *original = *(reinterpret_cast(ptr) - 1); std::ptrdiff_t previous_offset = static_cast(ptr)-static_cast(original); - original = std::realloc(original,size+EIGEN_ALIGN); + original = std::realloc(original,size+EIGEN_ALIGN_BYTES); if (original == 0) return 0; - void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(EIGEN_ALIGN-1))) + EIGEN_ALIGN); + void *aligned = reinterpret_cast((reinterpret_cast(original) & ~(std::size_t(EIGEN_ALIGN_BYTES-1))) + EIGEN_ALIGN_BYTES); void *previous_aligned = static_cast(original)+previous_offset; if(aligned!=previous_aligned) std::memmove(aligned, previous_aligned, size); @@ -221,11 +221,11 @@ inline void* aligned_malloc(size_t size) #elif EIGEN_MALLOC_ALREADY_ALIGNED result = std::malloc(size); #elif EIGEN_HAS_POSIX_MEMALIGN - if(posix_memalign(&result, EIGEN_ALIGN, size)) result = 0; + if(posix_memalign(&result, EIGEN_ALIGN_BYTES, size)) result = 0; #elif EIGEN_HAS_MM_MALLOC - result = _mm_malloc(size, EIGEN_ALIGN); + result = _mm_malloc(size, EIGEN_ALIGN_BYTES); #elif defined(_MSC_VER) && (!defined(_WIN32_WCE)) - result = _aligned_malloc(size, EIGEN_ALIGN); + result = _aligned_malloc(size, EIGEN_ALIGN_BYTES); #else result = handmade_aligned_malloc(size); #endif @@ -275,12 +275,12 @@ inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size) // implements _mm_malloc/_mm_free based on the corresponding _aligned_ // functions. This may not always be the case and we just try to be safe. #if defined(_MSC_VER) && defined(_mm_free) - result = _aligned_realloc(ptr,new_size,EIGEN_ALIGN); + result = _aligned_realloc(ptr,new_size,EIGEN_ALIGN_BYTES); #else result = generic_aligned_realloc(ptr,new_size,old_size); #endif #elif defined(_MSC_VER) - result = _aligned_realloc(ptr,new_size,EIGEN_ALIGN); + result = _aligned_realloc(ptr,new_size,EIGEN_ALIGN_BYTES); #else result = handmade_aligned_realloc(ptr,new_size,old_size); #endif @@ -608,8 +608,8 @@ template class aligned_stack_memory_handler */ #ifdef EIGEN_ALLOCA // The native alloca() that comes with llvm aligns buffer on 16 bytes even when AVX is enabled. - #if defined(__arm__) || EIGEN_ALIGN > 16 - #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast((reinterpret_cast(EIGEN_ALLOCA(SIZE+EIGEN_ALIGN)) & ~(size_t(EIGEN_ALIGN-1))) + EIGEN_ALIGN) + #if defined(__arm__) || EIGEN_ALIGN_BYTES > 16 + #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast((reinterpret_cast(EIGEN_ALLOCA(SIZE+EIGEN_ALIGN_BYTES)) & ~(size_t(EIGEN_ALIGN_BYTES-1))) + EIGEN_ALIGN_BYTES) #else #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA #endif @@ -679,7 +679,7 @@ template class aligned_stack_memory_handler #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true) #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \ - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_ALIGN==0))) + EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_ALIGN_BYTES==0))) /****************************************************************************/ diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 195d9e2e1..a08538aff 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -136,7 +136,7 @@ class compute_matrix_flags ((Options&DontAlign)==0) && ( #if EIGEN_ALIGN_STATICALLY - ((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % 16) == 0)) + ((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) #else 0 #endif diff --git a/test/geo_parametrizedline.cpp b/test/geo_parametrizedline.cpp index f0462d40a..5a72b3575 100644 --- a/test/geo_parametrizedline.cpp +++ b/test/geo_parametrizedline.cpp @@ -66,9 +66,9 @@ template void parametrizedline_alignment() typedef ParametrizedLine Line4a; typedef ParametrizedLine Line4u; - EIGEN_ALIGN16 Scalar array1[8]; - EIGEN_ALIGN16 Scalar array2[8]; - EIGEN_ALIGN16 Scalar array3[8+1]; + EIGEN_ALIGN_DEFAULT Scalar array1[8]; + EIGEN_ALIGN_DEFAULT Scalar array2[8]; + EIGEN_ALIGN_DEFAULT Scalar array3[8+1]; Scalar* array3u = array3+1; Line4a *p1 = ::new(reinterpret_cast(array1)) Line4a; diff --git a/test/mapped_matrix.cpp b/test/mapped_matrix.cpp index c18e687a5..5eba3ecb3 100644 --- a/test/mapped_matrix.cpp +++ b/test/mapped_matrix.cpp @@ -26,7 +26,7 @@ template void map_class_vector(const VectorType& m) Scalar* array1 = internal::aligned_new(size); Scalar* array2 = internal::aligned_new(size); Scalar* array3 = new Scalar[size+1]; - Scalar* array3unaligned = size_t(array3)%16 == 0 ? array3+1 : array3; + Scalar* array3unaligned = size_t(array3)%EIGEN_ALIGN_BYTES == 0 ? array3+1 : array3; Scalar array4[EIGEN_TESTMAP_MAX_SIZE]; Map(array1, size) = VectorType::Random(size); @@ -64,7 +64,7 @@ template void map_class_matrix(const MatrixType& m) for(int i = 0; i < size; i++) array2[i] = Scalar(1); Scalar* array3 = new Scalar[size+1]; for(int i = 0; i < size+1; i++) array3[i] = Scalar(1); - Scalar* array3unaligned = size_t(array3)%16 == 0 ? array3+1 : array3; + Scalar* array3unaligned = size_t(array3)%EIGEN_ALIGN_BYTES == 0 ? array3+1 : array3; Map(array1, rows, cols) = MatrixType::Ones(rows,cols); Map(array2, rows, cols) = Map(array1, rows, cols); Map(array3unaligned, rows, cols) = Map(array1, rows, cols); @@ -90,7 +90,7 @@ template void map_static_methods(const VectorType& m) Scalar* array1 = internal::aligned_new(size); Scalar* array2 = internal::aligned_new(size); Scalar* array3 = new Scalar[size+1]; - Scalar* array3unaligned = size_t(array3)%16 == 0 ? array3+1 : array3; + Scalar* array3unaligned = size_t(array3)%EIGEN_ALIGN_BYTES == 0 ? array3+1 : array3; VectorType::MapAligned(array1, size) = VectorType::Random(size); VectorType::Map(array2, size) = VectorType::Map(array1, size); diff --git a/test/packetmath.cpp b/test/packetmath.cpp index d7c336c22..5a680d1ee 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -106,10 +106,10 @@ template void packetmath() const int max_size = PacketSize > 4 ? PacketSize : 4; const int size = PacketSize*max_size; - EIGEN_ALIGN32 Scalar data1[size]; - EIGEN_ALIGN32 Scalar data2[size]; - EIGEN_ALIGN32 Packet packets[PacketSize*2]; - EIGEN_ALIGN32 Scalar ref[size]; + EIGEN_ALIGN_DEFAULT Scalar data1[size]; + EIGEN_ALIGN_DEFAULT Scalar data2[size]; + EIGEN_ALIGN_DEFAULT Packet packets[PacketSize*2]; + EIGEN_ALIGN_DEFAULT Scalar ref[size]; RealScalar refvalue = 0; for (int i=0; i void packetmath_real() const int PacketSize = internal::packet_traits::size; const int size = PacketSize*4; - EIGEN_ALIGN32 Scalar data1[internal::packet_traits::size*4]; - EIGEN_ALIGN32 Scalar data2[internal::packet_traits::size*4]; - EIGEN_ALIGN32 Scalar ref[internal::packet_traits::size*4]; + EIGEN_ALIGN_DEFAULT Scalar data1[internal::packet_traits::size*4]; + EIGEN_ALIGN_DEFAULT Scalar data2[internal::packet_traits::size*4]; + EIGEN_ALIGN_DEFAULT Scalar ref[internal::packet_traits::size*4]; for (int i=0; i void packetmath_notcomplex() typedef typename internal::packet_traits::type Packet; const int PacketSize = internal::packet_traits::size; - EIGEN_ALIGN32 Scalar data1[internal::packet_traits::size*4]; - EIGEN_ALIGN32 Scalar data2[internal::packet_traits::size*4]; - EIGEN_ALIGN32 Scalar ref[internal::packet_traits::size*4]; + EIGEN_ALIGN_DEFAULT Scalar data1[internal::packet_traits::size*4]; + EIGEN_ALIGN_DEFAULT Scalar data2[internal::packet_traits::size*4]; + EIGEN_ALIGN_DEFAULT Scalar ref[internal::packet_traits::size*4]; Array::Map(data1, internal::packet_traits::size*4).setRandom(); @@ -322,10 +322,10 @@ template void packetmath_complex() const int PacketSize = internal::packet_traits::size; const int size = PacketSize*4; - EIGEN_ALIGN32 Scalar data1[PacketSize*4]; - EIGEN_ALIGN32 Scalar data2[PacketSize*4]; - EIGEN_ALIGN32 Scalar ref[PacketSize*4]; - EIGEN_ALIGN32 Scalar pval[PacketSize*4]; + EIGEN_ALIGN_DEFAULT Scalar data1[PacketSize*4]; + EIGEN_ALIGN_DEFAULT Scalar data2[PacketSize*4]; + EIGEN_ALIGN_DEFAULT Scalar ref[PacketSize*4]; + EIGEN_ALIGN_DEFAULT Scalar pval[PacketSize*4]; for (int i=0; i Date: Wed, 19 Feb 2014 11:30:58 +0100 Subject: [PATCH 0314/1103] isApprox must honors nested_eval --- Eigen/src/Core/Fuzzy.h | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/Eigen/src/Core/Fuzzy.h b/Eigen/src/Core/Fuzzy.h index f9a88dd3c..9c8d10683 100644 --- a/Eigen/src/Core/Fuzzy.h +++ b/Eigen/src/Core/Fuzzy.h @@ -23,8 +23,13 @@ struct isApprox_selector static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) { EIGEN_USING_STD_MATH(min); +#ifdef EIGEN_TEST_EVALUATORS + typename internal::nested_eval::type nested(x); + typename internal::nested_eval::type otherNested(y); +#else typename internal::nested::type nested(x); typename internal::nested::type otherNested(y); +#endif return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum()); } }; From b3a07eecc5ad19f418f15d70d1517f8b7e07dc1b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 11:32:04 +0100 Subject: [PATCH 0315/1103] Fix CoeffReadCost of products to handle Dynamic costs --- Eigen/src/Core/products/CoeffBasedProduct.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index de06108da..77f291ab9 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -90,7 +90,7 @@ struct traits > // TODO enable vectorization for mixed types | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0), - CoeffReadCost = InnerSize == Dynamic ? Dynamic + CoeffReadCost = (InnerSize == Dynamic || LhsCoeffReadCost==Dynamic || RhsCoeffReadCost==Dynamic || NumTraits::AddCost==Dynamic || NumTraits::MulCost==Dynamic) ? Dynamic : InnerSize * (NumTraits::MulCost + LhsCoeffReadCost + RhsCoeffReadCost) + (InnerSize - 1) * NumTraits::AddCost, From ccc41128fb936563651a3c5b25bfc80f303710bf Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 11:33:29 +0100 Subject: [PATCH 0316/1103] Add a Solve expression for uniform treatment of solve() methods. --- Eigen/Core | 3 + Eigen/src/Core/CoreEvaluators.h | 4 +- Eigen/src/Core/Solve.h | 145 ++++++++++++++++++++++ Eigen/src/Core/TriangularMatrix.h | 17 +++ Eigen/src/Core/util/ForwardDeclarations.h | 3 +- 5 files changed, 169 insertions(+), 3 deletions(-) create mode 100644 Eigen/src/Core/Solve.h diff --git a/Eigen/Core b/Eigen/Core index c1d168ec4..e00cd7a16 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -361,6 +361,9 @@ using std::ptrdiff_t; #include "src/Core/Flagged.h" #include "src/Core/ProductBase.h" #include "src/Core/GeneralProduct.h" +#ifdef EIGEN_ENABLE_EVALUATORS +#include "src/Core/Solve.h" +#endif #include "src/Core/TriangularMatrix.h" #include "src/Core/SelfAdjointView.h" #include "src/Core/products/GeneralBlockPanelKernel.h" diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index d02bac0a8..726a6854a 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -495,14 +495,14 @@ struct evaluator > PacketScalar packet(Index row, Index col) const { return m_functor.packetOp(m_lhsImpl.template packet(row, col), - m_rhsImpl.template packet(row, col)); + m_rhsImpl.template packet(row, col)); } template PacketScalar packet(Index index) const { return m_functor.packetOp(m_lhsImpl.template packet(index), - m_rhsImpl.template packet(index)); + m_rhsImpl.template packet(index)); } protected: diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h new file mode 100644 index 000000000..2da2aff56 --- /dev/null +++ b/Eigen/src/Core/Solve.h @@ -0,0 +1,145 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_INVERSE_H +#define EIGEN_INVERSE_H + +namespace Eigen { + +template class SolveImpl; + +/** \class Solve + * \ingroup Core_Module + * + * \brief Pseudo expression representing a solving operation + * + * \tparam Decomposition the type of the matrix or decomposion object + * \tparam Rhstype the type of the right-hand side + * + * This class represents an expression of A.solve(B) + * and most of the time this is the only way it is used. + * + */ +namespace internal { + +// this solve_traits class permits to determine the evaluation type with respect to storage kind (Dense vs Sparse) +template struct solve_traits; + +template +struct solve_traits +{ + typedef typename Decomposition::MatrixType MatrixType; + typedef Matrix PlainObject; +}; + +template +struct traits > + : traits::StorageKind>::PlainObject> +{ + typedef typename solve_traits::StorageKind>::PlainObject PlainObject; + typedef traits BaseTraits; + enum { + Flags = BaseTraits::Flags & RowMajorBit, + CoeffReadCost = Dynamic + }; +}; + +} + + +template +class Solve : public SolveImpl::StorageKind> +{ +public: + typedef typename RhsType::Index Index; + typedef typename internal::traits::PlainObject PlainObject; + + Solve(const Decomposition &dec, const RhsType &rhs) + : m_dec(dec), m_rhs(rhs) + {} + + EIGEN_DEVICE_FUNC Index rows() const { return m_dec.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); } + + EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; } + EIGEN_DEVICE_FUNC const RhsType& rhs() const { return m_rhs; } + +protected: + const Decomposition &m_dec; + const RhsType &m_rhs; +}; + + +// Specilaization of the Solve expression for dense results +template +class SolveImpl + : public MatrixBase > +{ + typedef Solve Derived; + +public: + + typedef MatrixBase > Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + +private: + + Scalar coeff(Index row, Index col) const; + Scalar coeff(Index i) const; +}; + + +namespace internal { + +// Evaluator of Solve -> eval into a temporary +template +struct evaluator > + : public evaluator::PlainObject>::type +{ + typedef Solve SolveType; + typedef typename SolveType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const SolveType& solve) + : m_result(solve.rows(), solve.cols()) + { + ::new (static_cast(this)) Base(m_result); + solve.dec()._solve_impl(solve.rhs(), m_result); + } + +protected: + PlainObject m_result; +}; + +// Specialization for "dst = dec.solve(rhs)" +// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere +template +struct Assignment, internal::assign_op, Dense2Dense, Scalar> +{ + typedef Solve SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + // FIXME shall we resize dst here? + src.dec()._solve_impl(src.rhs(), dst); + } +}; + +} // end namepsace internal + +} // end namespace Eigen + +#endif // EIGEN_SOLVE_H diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index ffa2faeaa..76ba76d5f 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -437,11 +437,19 @@ template class TriangularView EIGEN_DEVICE_FUNC void solveInPlace(const MatrixBase& other) const; +#ifdef EIGEN_TEST_EVALUATORS + template + EIGEN_DEVICE_FUNC + inline const Solve + solve(const MatrixBase& other) const + { return Solve(*this, other.derived()); } +#else // EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC inline const internal::triangular_solve_retval solve(const MatrixBase& other) const { return solve(other); } +#endif // EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC @@ -547,6 +555,15 @@ template class TriangularView #endif // EIGEN_TEST_EVALUATORS #ifdef EIGEN_TEST_EVALUATORS + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const { + if(!(internal::is_same::value && internal::extract_data(dst) == internal::extract_data(rhs))) + dst = rhs; + this->template solveInPlace(dst); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularView& _assignProduct(const ProductType& prod, const Scalar& alpha); diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index db0e2b033..8cf13abd9 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -94,7 +94,8 @@ template class CwiseUnaryOp; template class CwiseUnaryView; template class CwiseBinaryOp; template class SelfCwiseBinaryOp; // TODO deprecated -template class ProductBase; +template class ProductBase; // TODO deprecated +template class Solve; namespace internal { template struct product_tag; From 3a735a6cf17db3f7aab0eee34412d3b711e64044 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 13:17:41 +0100 Subject: [PATCH 0317/1103] Fix lazy evaluation in Ref --- Eigen/src/Core/Ref.h | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index ce9e82913..665b7497a 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -239,7 +239,11 @@ template class Ref< template void construct(const Expression& expr, internal::false_type) { +#ifdef EIGEN_TEST_EVALUATORS + internal::call_assignment_no_alias(m_object,expr,internal::assign_op()); +#else m_object.lazyAssign(expr); +#endif Base::construct(m_object); } From 68e8ddaf94035827f9865ed0e402682f0292d53f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 13:26:07 +0100 Subject: [PATCH 0318/1103] Fix vectorization logic wrt assignment functors --- Eigen/src/Core/AssignEvaluator.h | 11 ++++++----- Eigen/src/Core/functors/AssignmentFunctors.h | 8 ++++---- 2 files changed, 10 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index cf4db8de5..8c6e31500 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -24,7 +24,7 @@ namespace internal { // copy_using_evaluator_traits is based on assign_traits -template +template struct copy_using_evaluator_traits { public: @@ -50,7 +50,8 @@ private: enum { StorageOrdersAgree = (int(Derived::IsRowMajor) == int(OtherDerived::IsRowMajor)), MightVectorize = StorageOrdersAgree - && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit), + && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit) + && (functor_traits::PacketAccess), MayInnerVectorize = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0 && int(DstIsAligned) && int(SrcIsAligned), MayLinearize = StorageOrdersAgree && (int(Derived::Flags) & int(OtherDerived::Flags) & LinearAccessBit), @@ -517,7 +518,7 @@ public: typedef SrcEvaluatorTypeT SrcEvaluatorType; typedef typename DstEvaluatorType::Scalar Scalar; typedef typename DstEvaluatorType::Index Index; - typedef copy_using_evaluator_traits AssignmentTraits; + typedef copy_using_evaluator_traits AssignmentTraits; generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) @@ -613,7 +614,7 @@ void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, co { #ifdef EIGEN_DEBUG_ASSIGN // TODO these traits should be computed from information provided by the evaluators - internal::copy_using_evaluator_traits::debug(); + internal::copy_using_evaluator_traits::debug(); #endif eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); @@ -750,7 +751,7 @@ struct Assignment eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); #ifdef EIGEN_DEBUG_ASSIGN - internal::copy_using_evaluator_traits::debug(); + internal::copy_using_evaluator_traits::debug(); #endif call_dense_assignment_loop(dst, src, func); diff --git a/Eigen/src/Core/functors/AssignmentFunctors.h b/Eigen/src/Core/functors/AssignmentFunctors.h index ae264aa64..1ff0eac29 100644 --- a/Eigen/src/Core/functors/AssignmentFunctors.h +++ b/Eigen/src/Core/functors/AssignmentFunctors.h @@ -31,7 +31,7 @@ template struct functor_traits > { enum { Cost = NumTraits::ReadCost, - PacketAccess = packet_traits::IsVectorized + PacketAccess = packet_traits::Vectorizable }; }; @@ -73,7 +73,7 @@ template struct functor_traits > { enum { Cost = NumTraits::ReadCost + NumTraits::AddCost, - PacketAccess = packet_traits::HasAdd + PacketAccess = packet_traits::HasSub }; }; @@ -115,7 +115,7 @@ template struct functor_traits > { enum { Cost = NumTraits::ReadCost + NumTraits::MulCost, - PacketAccess = packet_traits::HasMul + PacketAccess = packet_traits::HasDiv }; }; @@ -156,7 +156,7 @@ template struct functor_traits > { enum { Cost = 3 * NumTraits::ReadCost, - PacketAccess = packet_traits::IsVectorized + PacketAccess = packet_traits::Vectorizable }; }; From 61cff286189144e4dc84631c04db2a3b8fd43fad Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 14:05:56 +0100 Subject: [PATCH 0319/1103] Disable Flagged and ForceAlignedAccess --- Eigen/Core | 4 +++- Eigen/src/Core/DenseBase.h | 7 +++++++ Eigen/src/Core/MatrixBase.h | 9 +++++++++ 3 files changed, 19 insertions(+), 1 deletion(-) diff --git a/Eigen/Core b/Eigen/Core index e00cd7a16..bcf354a48 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -327,7 +327,10 @@ using std::ptrdiff_t; #include "src/Core/util/BlasUtil.h" #include "src/Core/DenseStorage.h" #include "src/Core/NestByValue.h" +#ifndef EIGEN_ENABLE_EVALUATORS #include "src/Core/ForceAlignedAccess.h" +#include "src/Core/Flagged.h" +#endif #include "src/Core/ReturnByValue.h" #include "src/Core/NoAlias.h" #include "src/Core/PlainObjectBase.h" @@ -358,7 +361,6 @@ using std::ptrdiff_t; #include "src/Core/IO.h" #include "src/Core/Swap.h" #include "src/Core/CommaInitializer.h" -#include "src/Core/Flagged.h" #include "src/Core/ProductBase.h" #include "src/Core/GeneralProduct.h" #ifdef EIGEN_ENABLE_EVALUATORS diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 9bbfacbcf..bd7c6d6a0 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -287,8 +287,15 @@ template class DenseBase EIGEN_DEVICE_FUNC CommaInitializer operator<< (const Scalar& s); +#ifndef EIGEN_TEST_EVALUATORS template const Flagged flagged() const; +#else + // TODO flagged is temporarly disabled. It seems useless now + template + const Derived& flagged() const + { return derived(); } +#endif template EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 851fa28d5..5a3675a20 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -360,10 +360,19 @@ template class MatrixBase NoAlias noalias(); +#ifndef EIGEN_TEST_EVALUATORS inline const ForceAlignedAccess forceAlignedAccess() const; inline ForceAlignedAccess forceAlignedAccess(); template inline typename internal::add_const_on_value_type,Derived&>::type>::type forceAlignedAccessIf() const; template inline typename internal::conditional,Derived&>::type forceAlignedAccessIf(); +#else + // TODO forceAlignedAccess is temporarly disabled + // Need to find a nicer workaround. + inline const Derived& forceAlignedAccess() const { return derived(); } + inline Derived& forceAlignedAccess() { return derived(); } + template inline const Derived& forceAlignedAccessIf() const { return derived(); } + template inline Derived& forceAlignedAccessIf() { return derived(); } +#endif Scalar trace() const; From b1ab6a8e0b1f1ecaffb6257fe41aecf46709348c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 14:06:35 +0100 Subject: [PATCH 0320/1103] Add missing assertion in swap() --- Eigen/src/Core/DenseBase.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index bd7c6d6a0..2fea574dd 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -404,6 +404,7 @@ template class DenseBase void swap(const DenseBase& other, int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase) { + eigen_assert(rows()==other.rows() && cols()==other.cols()); call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op()); } @@ -414,6 +415,7 @@ template class DenseBase EIGEN_DEVICE_FUNC void swap(PlainObjectBase& other) { + eigen_assert(rows()==other.rows() && cols()==other.cols()); call_assignment(derived(), other.derived(), internal::swap_assign_op()); } #else // EIGEN_TEST_EVALUATORS From 95b0a6707b45b78bad11fa7a090cd6439072c554 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 14:51:46 +0100 Subject: [PATCH 0321/1103] evaluator must evaluate its argument to avoid redundant evaluations --- Eigen/src/Core/CoreEvaluators.h | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 726a6854a..92b50d190 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -815,7 +815,8 @@ struct evaluator > typedef Replicate XprType; evaluator(const XprType& replicate) - : m_argImpl(replicate.nestedExpression()), + : m_arg(replicate.nestedExpression()), + m_argImpl(m_arg), m_rows(replicate.nestedExpression().rows()), m_cols(replicate.nestedExpression().cols()) { } @@ -851,7 +852,14 @@ struct evaluator > } protected: - typename evaluator::nestedType m_argImpl; + enum { + Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor + }; + typedef typename internal::nested_eval::type ArgTypeNested; + typedef typename internal::remove_all::type ArgTypeNestedCleaned; + + const ArgTypeNested m_arg; // FIXME is it OK to store both the argument and its evaluator?? (we have the same situation in evalautor_product) + typename evaluator::nestedType m_argImpl; const variable_if_dynamic m_rows; const variable_if_dynamic m_cols; }; From 8af02d19b2441024823e261dc71125b9d42fe909 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 15:16:11 +0100 Subject: [PATCH 0322/1103] ExprType::Nested has a new meaning now... --- test/nesting_ops.cpp | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/test/nesting_ops.cpp b/test/nesting_ops.cpp index 1e8523283..114dd5e41 100644 --- a/test/nesting_ops.cpp +++ b/test/nesting_ops.cpp @@ -11,7 +11,12 @@ template void run_nesting_ops(const MatrixType& _m) { +#ifndef EIGEN_TEST_EVALUATORS + // TODO, with evaluator, the following is not correct anymore: typename MatrixType::Nested m(_m); +#else + typename internal::nested_eval::type m(_m); +#endif // Make really sure that we are in debug mode! VERIFY_RAISES_ASSERT(eigen_assert(false)); From 2eee6eaf3c073fabb214e4e524a58148f4013c2c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 16:30:17 +0100 Subject: [PATCH 0323/1103] Fix mixing scalar types with evaluators --- Eigen/src/Core/ArrayBase.h | 2 +- Eigen/src/Core/AssignEvaluator.h | 14 +++++--------- Eigen/src/Core/CwiseBinaryOp.h | 13 ------------- Eigen/src/Core/ProductEvaluators.h | 4 ++-- Eigen/src/Core/functors/AssignmentFunctors.h | 18 ++++++++++-------- Eigen/src/Core/util/XprHelper.h | 13 +++++++++++++ test/mixingtypes.cpp | 5 ++++- 7 files changed, 35 insertions(+), 34 deletions(-) diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index c528de733..e67ca34cd 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -213,7 +213,7 @@ template EIGEN_STRONG_INLINE Derived & ArrayBase::operator*=(const ArrayBase& other) { - call_assignment(derived(), other.derived(), internal::mul_assign_op()); + call_assignment(derived(), other.derived(), internal::mul_assign_op()); return derived(); } diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 8c6e31500..084d650d8 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -729,6 +729,11 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) typedef typename internal::conditional, Dst&>::type ActualDstType; ActualDstType actualDst(dst); + // TODO check whether this is the right place to perform these checks: + EIGEN_STATIC_ASSERT_LVALUE(Dst) + EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(ActualDstTypeCleaned,Src) + EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar); + Assignment::run(actualDst, src, func); } @@ -739,15 +744,6 @@ struct Assignment { static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) { - // TODO check whether this is the right place to perform these checks: - enum{ - SameType = internal::is_same::value - }; - - EIGEN_STATIC_ASSERT_LVALUE(DstXprType) - EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(DstXprType,SrcXprType) - EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); #ifdef EIGEN_DEBUG_ASSIGN diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index 5624a4718..c78067a88 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -86,19 +86,6 @@ struct traits > }; } // end namespace internal -// we require Lhs and Rhs to have the same scalar type. Currently there is no example of a binary functor -// that would take two operands of different types. If there were such an example, then this check should be -// moved to the BinaryOp functors, on a per-case basis. This would however require a change in the BinaryOp functors, as -// currently they take only one typename Scalar template parameter. -// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths. -// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to -// add together a float matrix and a double matrix. -#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \ - EIGEN_STATIC_ASSERT((internal::functor_is_product_like::ret \ - ? int(internal::scalar_product_traits::Defined) \ - : int(internal::is_same::value)), \ - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - template class CwiseBinaryOpImpl; diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 67d0dc80c..cf612d58a 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -647,8 +647,8 @@ struct diagonal_product_evaluator_base : evaluator_base { typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::PacketScalar PacketScalar; + typedef typename scalar_product_traits::ReturnType Scalar; + typedef typename internal::packet_traits::type PacketScalar; public: diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag) : m_diagImpl(diag), m_matImpl(mat) diff --git a/Eigen/src/Core/functors/AssignmentFunctors.h b/Eigen/src/Core/functors/AssignmentFunctors.h index 1ff0eac29..d4d85a1ca 100644 --- a/Eigen/src/Core/functors/AssignmentFunctors.h +++ b/Eigen/src/Core/functors/AssignmentFunctors.h @@ -81,22 +81,24 @@ struct functor_traits > { * \brief Template functor for scalar/packet assignment with multiplication * */ -template struct mul_assign_op { +template +struct mul_assign_op { EIGEN_EMPTY_STRUCT_CTOR(mul_assign_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const { a *= b; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a *= b; } template - EIGEN_STRONG_INLINE void assignPacket(Scalar* a, const Packet& b) const - { internal::pstoret(a,internal::pmul(internal::ploadt(a),b)); } + EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const + { internal::pstoret(a,internal::pmul(internal::ploadt(a),b)); } }; -template -struct functor_traits > { +template +struct functor_traits > { enum { - Cost = NumTraits::ReadCost + NumTraits::MulCost, - PacketAccess = packet_traits::HasMul + Cost = NumTraits::ReadCost + NumTraits::MulCost, + PacketAccess = is_same::value && packet_traits::HasMul }; }; +template struct functor_is_product_like > { enum { ret = 1 }; }; /** \internal * \brief Template functor for scalar/packet assignment with diviving diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 7ea70450f..76e979ba0 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -516,6 +516,19 @@ template struct is_diagonal > } // end namespace internal +// we require Lhs and Rhs to have the same scalar type. Currently there is no example of a binary functor +// that would take two operands of different types. If there were such an example, then this check should be +// moved to the BinaryOp functors, on a per-case basis. This would however require a change in the BinaryOp functors, as +// currently they take only one typename Scalar template parameter. +// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths. +// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to +// add together a float matrix and a double matrix. +#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \ + EIGEN_STATIC_ASSERT((internal::functor_is_product_like::ret \ + ? int(internal::scalar_product_traits::Defined) \ + : int(internal::is_same::value)), \ + YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + } // end namespace Eigen #endif // EIGEN_XPRHELPER_H diff --git a/test/mixingtypes.cpp b/test/mixingtypes.cpp index 6c2f74875..ffedfb1f4 100644 --- a/test/mixingtypes.cpp +++ b/test/mixingtypes.cpp @@ -53,10 +53,13 @@ template void mixingtypes(int size = SizeAtCompileType) mf+mf; VERIFY_RAISES_ASSERT(mf+md); VERIFY_RAISES_ASSERT(mf+mcf); +#ifndef EIGEN_TEST_EVALUATORS + // they do not even compile when using evaluators VERIFY_RAISES_ASSERT(vf=vd); VERIFY_RAISES_ASSERT(vf+=vd); VERIFY_RAISES_ASSERT(mcd=md); - +#endif + // check scalar products VERIFY_IS_APPROX(vcf * sf , vcf * complex(sf)); VERIFY_IS_APPROX(sd * vcd, complex(sd) * vcd); From 5960befc206ac7405841a9da2436f377e4df694f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 19 Feb 2014 21:42:29 +0100 Subject: [PATCH 0324/1103] More int versus Index fixes --- Eigen/src/SparseCore/SparseMatrix.h | 9 +++++---- Eigen/src/SparseCore/SparsePermutation.h | 4 ++-- 2 files changed, 7 insertions(+), 6 deletions(-) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 0e7e760fb..9ac18bcf6 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -223,7 +223,7 @@ class SparseMatrix if(isCompressed()) { - reserve(VectorXi::Constant(outerSize(), 2)); + reserve(Matrix::Constant(outerSize(), 2)); } return insertUncompressed(row,col); } @@ -939,12 +939,13 @@ void set_from_triplets(const InputIterator& begin, const InputIterator& end, Spa EIGEN_UNUSED_VARIABLE(Options); enum { IsRowMajor = SparseMatrixType::IsRowMajor }; typedef typename SparseMatrixType::Scalar Scalar; + typedef typename SparseMatrixType::Index Index; SparseMatrix trMat(mat.rows(),mat.cols()); if(begin!=end) { // pass 1: count the nnz per inner-vector - VectorXi wi(trMat.outerSize()); + Matrix wi(trMat.outerSize()); wi.setZero(); for(InputIterator it(begin); it!=end; ++it) { @@ -1018,7 +1019,7 @@ void SparseMatrix::sumupDuplicates() { eigen_assert(!isCompressed()); // TODO, in practice we should be able to use m_innerNonZeros for that task - VectorXi wi(innerSize()); + Matrix wi(innerSize()); wi.fill(-1); Index count = 0; // for each inner-vector, wi[inner_index] will hold the position of first element into the index/value buffers @@ -1081,7 +1082,7 @@ EIGEN_DONT_INLINE SparseMatrix& SparseMatrix positions(dest.outerSize()); for (Index j=0; j tmp(m_matrix.rows(), m_matrix.cols()); - VectorXi sizes(m_matrix.outerSize()); + Matrix sizes(m_matrix.outerSize()); for(Index j=0; j tmp(m_matrix.rows(), m_matrix.cols()); - VectorXi sizes(tmp.outerSize()); + Matrix sizes(tmp.outerSize()); sizes.setZero(); PermutationMatrix perm; if((Side==OnTheLeft) ^ Transposed) From ecd2c8f37b8023b56d00cec4aebec7d2f3157e3f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 20 Feb 2014 14:18:24 +0100 Subject: [PATCH 0325/1103] Add general Inverse<> expression with evaluator --- Eigen/src/Core/MatrixBase.h | 5 + Eigen/src/Core/Solve.h | 6 +- Eigen/src/Core/util/ForwardDeclarations.h | 1 + Eigen/src/LU/Determinant.h | 4 + Eigen/src/LU/Inverse.h | 152 +++++++++++++++++++++- 5 files changed, 163 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 5a3675a20..72aa75da8 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -412,8 +412,13 @@ template class MatrixBase } #endif + #ifdef EIGEN_TEST_EVALUATORS + EIGEN_DEVICE_FUNC + const Inverse inverse() const; + #else EIGEN_DEVICE_FUNC const internal::inverse_impl inverse() const; + #endif template void computeInverseAndDetWithCheck( ResultType& inverse, diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index 2da2aff56..cab0e916e 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -7,8 +7,8 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. -#ifndef EIGEN_INVERSE_H -#define EIGEN_INVERSE_H +#ifndef EIGEN_SOLVE_H +#define EIGEN_SOLVE_H namespace Eigen { @@ -81,7 +81,7 @@ protected: }; -// Specilaization of the Solve expression for dense results +// Specialization of the Solve expression for dense results template class SolveImpl : public MatrixBase > diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 8cf13abd9..ecb811910 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -96,6 +96,7 @@ template class CwiseBinaryOp; template class SelfCwiseBinaryOp; // TODO deprecated template class ProductBase; // TODO deprecated template class Solve; +template class Inverse; namespace internal { template struct product_tag; diff --git a/Eigen/src/LU/Determinant.h b/Eigen/src/LU/Determinant.h index bb8e78a8a..9726bd96a 100644 --- a/Eigen/src/LU/Determinant.h +++ b/Eigen/src/LU/Determinant.h @@ -92,7 +92,11 @@ template inline typename internal::traits::Scalar MatrixBase::determinant() const { eigen_assert(rows() == cols()); +#ifdef EIGEN_TEST_EVALUATORS + typedef typename internal::nested_eval::type Nested; +#else typedef typename internal::nested::type Nested; +#endif return internal::determinant_impl::type>::run(derived()); } diff --git a/Eigen/src/LU/Inverse.h b/Eigen/src/LU/Inverse.h index 8d1364e0a..593ce6f8f 100644 --- a/Eigen/src/LU/Inverse.h +++ b/Eigen/src/LU/Inverse.h @@ -42,7 +42,12 @@ struct compute_inverse static inline void run(const MatrixType& matrix, ResultType& result) { typedef typename MatrixType::Scalar Scalar; +#ifdef EIGEN_TEST_EVALUATORS + typename internal::evaluator::type matrixEval(matrix); + result.coeffRef(0,0) = Scalar(1) / matrixEval.coeff(0,0); +#else result.coeffRef(0,0) = Scalar(1) / matrix.coeff(0,0); +#endif } }; @@ -75,10 +80,10 @@ inline void compute_inverse_size2_helper( const MatrixType& matrix, const typename ResultType::Scalar& invdet, ResultType& result) { - result.coeffRef(0,0) = matrix.coeff(1,1) * invdet; + result.coeffRef(0,0) = matrix.coeff(1,1) * invdet; result.coeffRef(1,0) = -matrix.coeff(1,0) * invdet; result.coeffRef(0,1) = -matrix.coeff(0,1) * invdet; - result.coeffRef(1,1) = matrix.coeff(0,0) * invdet; + result.coeffRef(1,1) = matrix.coeff(0,0) * invdet; } template @@ -279,6 +284,7 @@ struct compute_inverse_and_det_with_check *** MatrixBase methods *** *************************/ +#ifndef EIGEN_TEST_EVALUATORS template struct traits > { @@ -313,9 +319,141 @@ struct inverse_impl : public ReturnByValue > compute_inverse::run(m_matrix, dst); } }; +#endif +} // end namespace internal + +#ifdef EIGEN_TEST_EVALUATORS + +// TODO move the general declaration in Core, and rename this file DenseInverseImpl.h, or something like this... + +template class InverseImpl; + +namespace internal { + +template +struct traits > + : traits +{ + typedef typename XprType::PlainObject PlainObject; + typedef traits BaseTraits; + enum { + Flags = BaseTraits::Flags & RowMajorBit, + CoeffReadCost = Dynamic + }; +}; } // end namespace internal +/** \class Inverse + * \ingroup LU_Module + * + * \brief Expression of the inverse of another expression + * + * \tparam XprType the type of the expression we are taking the inverse + * + * This class represents an expression of A.inverse() + * and most of the time this is the only way it is used. + * + */ +template +class Inverse : public InverseImpl::StorageKind> +{ +public: + typedef typename XprType::Index Index; + typedef typename XprType::PlainObject PlainObject; + typedef typename internal::nested::type XprTypeNested; + typedef typename internal::remove_all::type XprTypeNestedCleaned; + + Inverse(const XprType &xpr) + : m_xpr(xpr) + {} + + EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); } + + EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; } + +protected: + XprTypeNested &m_xpr; +}; + +// Specialization of the Inverse expression for dense expressions +template +class InverseImpl + : public MatrixBase > +{ + typedef Inverse Derived; + +public: + + typedef MatrixBase Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + +private: + + Scalar coeff(Index row, Index col) const; + Scalar coeff(Index i) const; +}; + +namespace internal { + +// Evaluator of Inverse -> eval into a temporary +template +struct evaluator > + : public evaluator::PlainObject>::type +{ + typedef Inverse InverseType; + typedef typename InverseType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const InverseType& inv_xpr) + : m_result(inv_xpr.rows(), inv_xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + + typedef typename internal::nested_eval::type ActualXprType; + typedef typename internal::remove_all::type ActualXprTypeCleanded; + + ActualXprType actual_xpr(inv_xpr.nestedExpression()); + + compute_inverse::run(actual_xpr, m_result); + } + +protected: + PlainObject m_result; +}; + +// Specialization for "dst = xpr.inverse()" +// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere +template +struct Assignment, internal::assign_op, Dense2Dense, Scalar> +{ + typedef Inverse SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + // FIXME shall we resize dst here? + const int Size = EIGEN_PLAIN_ENUM_MIN(XprType::ColsAtCompileTime,DstXprType::ColsAtCompileTime); + EIGEN_ONLY_USED_FOR_DEBUG(Size); + eigen_assert(( (Size<=1) || (Size>4) || (extract_data(src.nestedExpression())!=extract_data(dst))) + && "Aliasing problem detected in inverse(), you need to do inverse().eval() here."); + + typedef typename internal::nested_eval::type ActualXprType; + typedef typename internal::remove_all::type ActualXprTypeCleanded; + + ActualXprType actual_xpr(src.nestedExpression()); + + compute_inverse::run(actual_xpr, dst); + } +}; + + +} // end namespace internal + +#endif + /** \lu_module * * \returns the matrix inverse of this matrix. @@ -333,6 +471,15 @@ struct inverse_impl : public ReturnByValue > * * \sa computeInverseAndDetWithCheck() */ +#ifdef EIGEN_TEST_EVALUATORS +template +inline const Inverse MatrixBase::inverse() const +{ + EIGEN_STATIC_ASSERT(!NumTraits::IsInteger,THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES) + eigen_assert(rows() == cols()); + return Inverse(derived()); +} +#else template inline const internal::inverse_impl MatrixBase::inverse() const { @@ -340,6 +487,7 @@ inline const internal::inverse_impl MatrixBase::inverse() cons eigen_assert(rows() == cols()); return internal::inverse_impl(derived()); } +#endif /** \lu_module * From 5f6ec95291a7adfe990854469d7697cc65642c8c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 20 Feb 2014 15:24:00 +0100 Subject: [PATCH 0326/1103] Propagate LvalueBit flag to TriangularView --- Eigen/src/Core/TriangularMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 76ba76d5f..96c81f933 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -174,7 +174,7 @@ struct traits > : traits typedef typename MatrixType::PlainObject DenseMatrixType; enum { Mode = _Mode, - Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode, + Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode, CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost }; }; From 96213335451c821f1b969ffbce5b6e73e5e55ca3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 20 Feb 2014 15:24:21 +0100 Subject: [PATCH 0327/1103] Fix dimension of Solve expression --- Eigen/src/Core/Solve.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index cab0e916e..c28789968 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -69,7 +69,7 @@ public: : m_dec(dec), m_rhs(rhs) {} - EIGEN_DEVICE_FUNC Index rows() const { return m_dec.rows(); } + EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); } EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); } EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; } From b2e1453e1e81132381bfb5cd46ec5a66ec55b3c6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 20 Feb 2014 15:25:06 +0100 Subject: [PATCH 0328/1103] Some bit flags and internal structures are deprecated --- Eigen/src/Core/util/Constants.h | 8 ++++---- Eigen/src/Core/util/ForwardDeclarations.h | 2 ++ 2 files changed, 6 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 14449eb6c..d9e51ffea 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -53,14 +53,14 @@ const int Infinity = -1; const unsigned int RowMajorBit = 0x1; /** \ingroup flags - * + * \deprecated * means the expression should be evaluated by the calling expression */ -const unsigned int EvalBeforeNestingBit = 0x2; +const unsigned int EvalBeforeNestingBit = 0x2; // FIXME deprecated /** \ingroup flags - * + * \deprecated * means the expression should be evaluated before any assignment */ -const unsigned int EvalBeforeAssigningBit = 0x4; +const unsigned int EvalBeforeAssigningBit = 0x4; // FIXME deprecated /** \ingroup flags * diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index ecb811910..975fdbf2a 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -141,8 +141,10 @@ template class ArrayWrapper; template class MatrixWrapper; namespace internal { +#ifndef EIGEN_TEST_EVALUATROS template struct solve_retval_base; template struct solve_retval; +#endif template struct kernel_retval_base; template struct kernel_retval; template struct image_retval_base; From 93125e372df80f07bb7d74abdebb592425ddba7b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 20 Feb 2014 15:26:15 +0100 Subject: [PATCH 0329/1103] Port LU module to evaluators (except image() and kernel()) --- Eigen/src/Core/AssignEvaluator.h | 2 +- Eigen/src/LU/FullPivLU.h | 152 ++++++++++++++++++++++--------- Eigen/src/LU/PartialPivLU.h | 85 ++++++++++++++--- 3 files changed, 178 insertions(+), 61 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 084d650d8..73f20c1a1 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -721,7 +721,7 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1)) && int(Dst::SizeAtCompileTime) != 1 }; - + dst.resize(NeedToTranspose ? src.cols() : src.rows(), NeedToTranspose ? src.rows() : src.cols()); diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index 44699b763..86de91ccb 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -12,6 +12,19 @@ namespace Eigen { +namespace internal { +template struct traits > + : traits<_MatrixType> +{ + typedef traits<_MatrixType> BaseTraits; + enum { + Flags = BaseTraits::Flags & RowMajorBit, + CoeffReadCost = Dynamic + }; +}; + +} // end namespace internal + /** \ingroup LU_Module * * \class FullPivLU @@ -61,6 +74,7 @@ template class FullPivLU typedef typename internal::plain_col_type::type IntColVectorType; typedef PermutationMatrix PermutationQType; typedef PermutationMatrix PermutationPType; + typedef typename MatrixType::PlainObject PlainObject; /** * \brief Default Constructor. @@ -209,6 +223,15 @@ template class FullPivLU * * \sa TriangularView::solve(), kernel(), inverse() */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "LU is not initialized."); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -216,6 +239,7 @@ template class FullPivLU eigen_assert(m_isInitialized && "LU is not initialized."); return internal::solve_retval(*this, b.derived()); } +#endif /** \returns the determinant of the matrix of which * *this is the LU decomposition. It has only linear complexity @@ -359,6 +383,14 @@ template class FullPivLU * * \sa MatrixBase::inverse() */ +#ifdef EIGEN_TEST_EVALUATORS + inline const Inverse inverse() const + { + eigen_assert(m_isInitialized && "LU is not initialized."); + eigen_assert(m_lu.rows() == m_lu.cols() && "You can't take the inverse of a non-square matrix!"); + return Inverse(*this); + } +#else inline const internal::solve_retval inverse() const { eigen_assert(m_isInitialized && "LU is not initialized."); @@ -366,11 +398,18 @@ template class FullPivLU return internal::solve_retval (*this, MatrixType::Identity(m_lu.rows(), m_lu.cols())); } +#endif MatrixType reconstructedMatrix() const; inline Index rows() const { return m_lu.rows(); } inline Index cols() const { return m_lu.cols(); } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif protected: MatrixType m_lu; @@ -662,6 +701,61 @@ struct image_retval > /***** Implementation of solve() *****************************************************/ +} // end namespace internal + +#ifndef EIGEN_PARSED_BY_DOXYGEN +template +template +void FullPivLU<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const { + + /* The decomposition PAQ = LU can be rewritten as A = P^{-1} L U Q^{-1}. + * So we proceed as follows: + * Step 1: compute c = P * rhs. + * Step 2: replace c by the solution x to Lx = c. Exists because L is invertible. + * Step 3: replace c by the solution x to Ux = c. May or may not exist. + * Step 4: result = Q * c; + */ + + const Index rows = this->rows(), + cols = this->cols(), + nonzero_pivots = this->nonzeroPivots(); + eigen_assert(rhs.rows() == rows); + const Index smalldim = (std::min)(rows, cols); + + if(nonzero_pivots == 0) + { + dst.setZero(); + return; + } + + typename RhsType::PlainObject c(rhs.rows(), rhs.cols()); + + // Step 1 + c = permutationP() * rhs; + + // Step 2 + m_lu.topLeftCorner(smalldim,smalldim) + .template triangularView() + .solveInPlace(c.topRows(smalldim)); + if(rows>cols) + c.bottomRows(rows-cols) -= m_lu.bottomRows(rows-cols) * c.topRows(cols); + + // Step 3 + m_lu.topLeftCorner(nonzero_pivots, nonzero_pivots) + .template triangularView() + .solveInPlace(c.topRows(nonzero_pivots)); + + // Step 4 + for(Index i = 0; i < nonzero_pivots; ++i) + dst.row(permutationQ().indices().coeff(i)) = c.row(i); + for(Index i = nonzero_pivots; i < m_lu.cols(); ++i) + dst.row(permutationQ().indices().coeff(i)).setZero(); +} +#endif + +namespace internal { + +#ifdef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -670,53 +764,21 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - /* The decomposition PAQ = LU can be rewritten as A = P^{-1} L U Q^{-1}. - * So we proceed as follows: - * Step 1: compute c = P * rhs. - * Step 2: replace c by the solution x to Lx = c. Exists because L is invertible. - * Step 3: replace c by the solution x to Ux = c. May or may not exist. - * Step 4: result = Q * c; - */ + dec()._solve_impl(rhs(), dst); + } +}; +#endif - const Index rows = dec().rows(), cols = dec().cols(), - nonzero_pivots = dec().nonzeroPivots(); - eigen_assert(rhs().rows() == rows); - const Index smalldim = (std::min)(rows, cols); +/***** Implementation of inverse() *****************************************************/ - if(nonzero_pivots == 0) - { - dst.setZero(); - return; - } - - typename Rhs::PlainObject c(rhs().rows(), rhs().cols()); - - // Step 1 - c = dec().permutationP() * rhs(); - - // Step 2 - dec().matrixLU() - .topLeftCorner(smalldim,smalldim) - .template triangularView() - .solveInPlace(c.topRows(smalldim)); - if(rows>cols) - { - c.bottomRows(rows-cols) - -= dec().matrixLU().bottomRows(rows-cols) - * c.topRows(cols); - } - - // Step 3 - dec().matrixLU() - .topLeftCorner(nonzero_pivots, nonzero_pivots) - .template triangularView() - .solveInPlace(c.topRows(nonzero_pivots)); - - // Step 4 - for(Index i = 0; i < nonzero_pivots; ++i) - dst.row(dec().permutationQ().indices().coeff(i)) = c.row(i); - for(Index i = nonzero_pivots; i < dec().matrixLU().cols(); ++i) - dst.row(dec().permutationQ().indices().coeff(i)).setZero(); +template +struct Assignment >, internal::assign_op, Dense2Dense, Scalar> +{ + typedef FullPivLU LuType; + typedef Inverse SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index 1d389ecac..ac53f7ab5 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -13,6 +13,19 @@ namespace Eigen { +namespace internal { +template struct traits > + : traits<_MatrixType> +{ + typedef traits<_MatrixType> BaseTraits; + enum { + Flags = BaseTraits::Flags & RowMajorBit, + CoeffReadCost = Dynamic + }; +}; + +} // end namespace internal + /** \ingroup LU_Module * * \class PartialPivLU @@ -62,6 +75,7 @@ template class PartialPivLU typedef typename MatrixType::Index Index; typedef PermutationMatrix PermutationType; typedef Transpositions TranspositionType; + typedef typename MatrixType::PlainObject PlainObject; /** @@ -128,6 +142,15 @@ template class PartialPivLU * * \sa TriangularView::solve(), inverse(), computeInverse() */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -135,6 +158,7 @@ template class PartialPivLU eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); return internal::solve_retval(*this, b.derived()); } +#endif /** \returns the inverse of the matrix of which *this is the LU decomposition. * @@ -143,12 +167,20 @@ template class PartialPivLU * * \sa MatrixBase::inverse(), LU::inverse() */ +#ifdef EIGEN_TEST_EVALUATORS + inline const Inverse inverse() const + { + eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); + return Inverse(*this); + } +#else inline const internal::solve_retval inverse() const { eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); return internal::solve_retval (*this, MatrixType::Identity(m_lu.rows(), m_lu.cols())); } +#endif /** \returns the determinant of the matrix of which * *this is the LU decomposition. It has only linear complexity @@ -169,6 +201,30 @@ template class PartialPivLU inline Index rows() const { return m_lu.rows(); } inline Index cols() const { return m_lu.cols(); } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const { + /* The decomposition PA = LU can be rewritten as A = P^{-1} L U. + * So we proceed as follows: + * Step 1: compute c = Pb. + * Step 2: replace c by the solution x to Lx = c. + * Step 3: replace c by the solution x to Ux = c. + */ + + eigen_assert(rhs.rows() == m_lu.rows()); + + // Step 1 + dst = permutationP() * rhs; + + // Step 2 + m_lu.template triangularView().solveInPlace(dst); + + // Step 3 + m_lu.template triangularView().solveInPlace(dst); + } + #endif protected: MatrixType m_lu; @@ -434,6 +490,7 @@ MatrixType PartialPivLU::reconstructedMatrix() const namespace internal { +#ifndef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -442,23 +499,21 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - /* The decomposition PA = LU can be rewritten as A = P^{-1} L U. - * So we proceed as follows: - * Step 1: compute c = Pb. - * Step 2: replace c by the solution x to Lx = c. - * Step 3: replace c by the solution x to Ux = c. - */ + dec()._solve_impl(rhs(), dst); + } +}; +#endif - eigen_assert(rhs().rows() == dec().matrixLU().rows()); +/***** Implementation of inverse() *****************************************************/ - // Step 1 - dst = dec().permutationP() * rhs(); - - // Step 2 - dec().matrixLU().template triangularView().solveInPlace(dst); - - // Step 3 - dec().matrixLU().template triangularView().solveInPlace(dst); +template +struct Assignment >, internal::assign_op, Dense2Dense, Scalar> +{ + typedef PartialPivLU LuType; + typedef Inverse SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; From af31b6c37a3b4b32c8075d94b39a78108f12fd31 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 21 Feb 2014 15:22:08 +0100 Subject: [PATCH 0330/1103] Generalize evaluator> such that there is no need to specialize it --- Eigen/src/Core/AssignEvaluator.h | 6 +++++- Eigen/src/LU/Inverse.h | 27 +++++++++++++++++---------- 2 files changed, 22 insertions(+), 11 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 73f20c1a1..14effd1f2 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -736,7 +736,11 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) Assignment::run(actualDst, src, func); } - +template +void call_assignment_no_alias(Dst& dst, const Src& src) +{ + call_assignment_no_alias(dst, src, internal::assign_op()); +} // Generic Dense to Dense assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> diff --git a/Eigen/src/LU/Inverse.h b/Eigen/src/LU/Inverse.h index 593ce6f8f..db053406e 100644 --- a/Eigen/src/LU/Inverse.h +++ b/Eigen/src/LU/Inverse.h @@ -351,7 +351,7 @@ struct traits > * * \tparam XprType the type of the expression we are taking the inverse * - * This class represents an expression of A.inverse() + * This class represents an abstract expression of A.inverse() * and most of the time this is the only way it is used. * */ @@ -377,7 +377,11 @@ protected: XprTypeNested &m_xpr; }; -// Specialization of the Inverse expression for dense expressions +/** \internal + * Specialization of the Inverse expression for dense expressions. + * Direct access to the coefficients are discared. + * FIXME this intermediate class is probably not needed anymore. + */ template class InverseImpl : public MatrixBase > @@ -397,7 +401,16 @@ private: namespace internal { -// Evaluator of Inverse -> eval into a temporary +/** \internal + * \brief Default evaluator for Inverse expression. + * + * This default evaluator for Inverse expression simply evaluate the inverse into a temporary + * by a call to internal::call_assignment_no_alias. + * Therefore, inverse implementers only have to specialize Assignment, ...> for + * there own nested expression. + * + * \sa class Inverse + */ template struct evaluator > : public evaluator::PlainObject>::type @@ -413,13 +426,7 @@ struct evaluator > : m_result(inv_xpr.rows(), inv_xpr.cols()) { ::new (static_cast(this)) Base(m_result); - - typedef typename internal::nested_eval::type ActualXprType; - typedef typename internal::remove_all::type ActualXprTypeCleanded; - - ActualXprType actual_xpr(inv_xpr.nestedExpression()); - - compute_inverse::run(actual_xpr, m_result); + internal::call_assignment_no_alias(m_result, inv_xpr); } protected: From 728c3d2cb955a255cae5515197ae65dc83209509 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 21 Feb 2014 16:27:24 +0100 Subject: [PATCH 0331/1103] Get rid of GeneralProduct for outer-products, and get rid of ScaledProduct --- Eigen/src/Core/GeneralProduct.h | 3 ++ Eigen/src/Core/ProductBase.h | 3 ++ Eigen/src/Core/ProductEvaluators.h | 71 +++++++++++++++++++++++++----- 3 files changed, 66 insertions(+), 11 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index f823ff251..4c0fc7f63 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -247,6 +247,7 @@ class GeneralProduct * Implementation of Outer Vector Vector Product ***********************************************************************/ +#ifndef EIGEN_TEST_EVALUATORS namespace internal { // Column major @@ -326,6 +327,8 @@ class GeneralProduct } }; +#endif // EIGEN_TEST_EVALUATORS + /*********************************************************************** * Implementation of General Matrix Vector Product ***********************************************************************/ diff --git a/Eigen/src/Core/ProductBase.h b/Eigen/src/Core/ProductBase.h index a494b5f87..f6b719d19 100644 --- a/Eigen/src/Core/ProductBase.h +++ b/Eigen/src/Core/ProductBase.h @@ -174,6 +174,7 @@ class ProductBase : public MatrixBase mutable PlainObject m_result; }; +#ifndef EIGEN_TEST_EVALUATORS // here we need to overload the nested rule for products // such that the nested type is a const reference to a plain matrix namespace internal { @@ -263,6 +264,8 @@ class ScaledProduct Scalar m_alpha; }; +#endif // EIGEN_TEST_EVALUATORS + /** \internal * Overloaded to perform an efficient C = (A*B).lazy() */ template diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index cf612d58a..93ae5f5f5 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -17,7 +17,11 @@ namespace Eigen { namespace internal { -// Like more general binary expressions, products need their own evaluator: +/** \internal + * \class product_evaluator + * Products need their own evaluator with more template arguments allowing for + * easier partial template specializations. + */ template< typename T, int ProductTag = internal::product_type::ret, typename LhsShape = typename evaluator_traits::Shape, @@ -26,6 +30,14 @@ template< typename T, typename RhsScalar = typename T::Rhs::Scalar > struct product_evaluator; +/** \internal + * Evaluator of a product expression. + * Since products require special treatments to handle all possible cases, + * we simply deffer the evaluation logic to a product_evaluator class + * which offers more partial specialization possibilities. + * + * \sa class product_evaluator + */ template struct evaluator > : public product_evaluator > @@ -40,7 +52,7 @@ struct evaluator > }; // Catch scalar * ( A * B ) and transform it to (A*scalar) * B -// TODO we should apply that rule if that's really helpful +// TODO we should apply that rule only if that's really helpful template struct evaluator, const Product > > : public evaluator,const Lhs>, Rhs, DefaultProduct> > @@ -66,7 +78,7 @@ struct evaluator, DiagIndex> > typedef evaluator type; typedef evaluator nestedType; -// + evaluator(const XprType& xpr) : Base(Diagonal, DiagIndex>( Product(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()), @@ -183,38 +195,75 @@ struct generic_product_impl }; +/*********************************************************************** +* Implementation of outer dense * dense vector product +***********************************************************************/ + +// Column major result +template +EIGEN_DONT_INLINE void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&) +{ + typedef typename Dst::Index Index; + // FIXME make sure lhs is sequentially stored + // FIXME not very good if rhs is real and lhs complex while alpha is real too + // FIXME we should probably build an evaluator for dst and rhs + const Index cols = dst.cols(); + for (Index j=0; j +EIGEN_DONT_INLINE void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&) { + typedef typename Dst::Index Index; + // FIXME make sure rhs is sequentially stored + // FIXME not very good if lhs is real and rhs complex while alpha is real too + // FIXME we should probably build an evaluator for dst and lhs + const Index rows = dst.rows(); + for (Index i=0; i struct generic_product_impl { + template struct IsRowMajor : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {}; typedef typename Product::Scalar Scalar; + // TODO it would be nice to be able to exploit our *_assign_op functors for that purpose + struct set { template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() = src; } }; + struct add { template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } }; + struct sub { template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } }; + struct adds { + Scalar m_scale; + adds(const Scalar& s) : m_scale(s) {} + template void operator()(const Dst& dst, const Src& src) const { + dst.const_cast_derived() += m_scale * src; + } + }; + template static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { - // TODO bypass GeneralProduct class - GeneralProduct(lhs,rhs).evalTo(dst); + internal::outer_product_selector_run(dst, lhs, rhs, set(), IsRowMajor()); } template static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { - // TODO bypass GeneralProduct class - GeneralProduct(lhs,rhs).addTo(dst); + internal::outer_product_selector_run(dst, lhs, rhs, add(), IsRowMajor()); } template static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) { - // TODO bypass GeneralProduct class - GeneralProduct(lhs,rhs).subTo(dst); + internal::outer_product_selector_run(dst, lhs, rhs, sub(), IsRowMajor()); } template static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - // TODO bypass GeneralProduct class - GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); + internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), IsRowMajor()); } }; From 6c7ab508117d84671054808c921980a4908efb20 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 21 Feb 2014 16:43:03 +0100 Subject: [PATCH 0332/1103] Get rid of GeneralProduct<> for GemmProduct --- Eigen/Core | 7 +-- Eigen/src/Core/ProductEvaluators.h | 14 ----- Eigen/src/Core/products/GeneralMatrixMatrix.h | 57 +++++++++++++++++++ 3 files changed, 60 insertions(+), 18 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index bcf354a48..245604465 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -371,6 +371,9 @@ using std::ptrdiff_t; #include "src/Core/products/GeneralBlockPanelKernel.h" #include "src/Core/products/Parallelizer.h" #include "src/Core/products/CoeffBasedProduct.h" +#ifdef EIGEN_ENABLE_EVALUATORS +#include "src/Core/ProductEvaluators.h" +#endif #include "src/Core/products/GeneralMatrixVector.h" #include "src/Core/products/GeneralMatrixMatrix.h" #include "src/Core/SolveTriangular.h" @@ -386,10 +389,6 @@ using std::ptrdiff_t; #include "src/Core/BandMatrix.h" #include "src/Core/CoreIterators.h" -#ifdef EIGEN_ENABLE_EVALUATORS -#include "src/Core/ProductEvaluators.h" -#endif - #include "src/Core/BooleanRedux.h" #include "src/Core/Select.h" #include "src/Core/VectorwiseOp.h" diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 93ae5f5f5..99c48ae52 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -311,20 +311,6 @@ struct generic_product_impl } }; -template -struct generic_product_impl - : generic_product_impl_base > -{ - typedef typename Product::Scalar Scalar; - - template - static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) - { - // TODO bypass GeneralProduct class - GeneralProduct(lhs,rhs).scaleAndAddTo(dst, alpha); - } -}; - template struct generic_product_impl { diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 3f5ffcf51..1c8940e1c 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -374,6 +374,7 @@ class gemm_blocking_space class GeneralProduct : public ProductBase, Lhs, Rhs> @@ -421,6 +422,62 @@ class GeneralProduct internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit); } }; +#else // EIGEN_TEST_EVALUATORS +namespace internal { + +template +struct generic_product_impl + : generic_product_impl_base > +{ + typedef typename Product::Scalar Scalar; + typedef typename Product::Index Index; + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef typename internal::remove_all::type ActualLhsTypeCleaned; + + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all::type ActualRhsTypeCleaned; + + enum { + MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime) + }; + + template + static void scaleAndAddTo(Dest& dst, const Lhs& a_lhs, const Rhs& a_rhs, const Scalar& alpha) + { + eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols()); + + typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar, + Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType; + + typedef internal::gemm_functor< + Scalar, Index, + internal::general_matrix_matrix_product< + Index, + LhsScalar, (ActualLhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate), + RhsScalar, (ActualRhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate), + (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>, + ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor; + + BlockingType blocking(dst.rows(), dst.cols(), lhs.cols()); + + internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)> + (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), Dest::Flags&RowMajorBit); + } +}; + +} // end namespace internal +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen From d67548f345d01c69d9dbba5869d8cc0159e96464 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 21 Feb 2014 17:13:28 +0100 Subject: [PATCH 0333/1103] Get rid of GeneralProduct<> for GemvProduct --- Eigen/src/Core/GeneralProduct.h | 208 ++++++++++++++++++++++++++--- Eigen/src/Core/ProductBase.h | 3 +- Eigen/src/Core/ProductEvaluators.h | 2 +- 3 files changed, 196 insertions(+), 17 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 4c0fc7f63..57d5d3c38 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -342,16 +342,19 @@ class GeneralProduct */ namespace internal { +#ifndef EIGEN_TEST_EVALUATORS template struct traits > : traits, Lhs, Rhs> > {}; +#endif template struct gemv_selector; } // end namespace internal +#ifndef EIGEN_TEST_EVALUATORS template class GeneralProduct : public ProductBase, Lhs, Rhs> @@ -378,24 +381,10 @@ class GeneralProduct bool(internal::blas_traits::HasUsableDirectAccess)>::run(*this, dst, alpha); } }; +#endif namespace internal { -// The vector is on the left => transposition -template -struct gemv_selector -{ - template - static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha) - { - Transpose destT(dest); - enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; - gemv_selector - ::run(GeneralProduct,Transpose, GemvProduct> - (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha); - } -}; - template struct gemv_static_vector_if; template @@ -432,6 +421,23 @@ struct gemv_static_vector_if #endif }; +#ifndef EIGEN_TEST_EVALUATORS + +// The vector is on the left => transposition +template +struct gemv_selector +{ + template + static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha) + { + Transpose destT(dest); + enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; + gemv_selector + ::run(GeneralProduct,Transpose, GemvProduct> + (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha); + } +}; + template<> struct gemv_selector { template @@ -582,6 +588,178 @@ template<> struct gemv_selector } }; +#else // EIGEN_TEST_EVALUATORS + +// The vector is on the left => transposition +template +struct gemv_selector +{ + template + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + Transpose destT(dest); + enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; + gemv_selector + ::run(rhs.transpose(), lhs.transpose(), destT, alpha); + } +}; + +template<> struct gemv_selector +{ + template + static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Dest::Index Index; + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + typedef typename Dest::RealScalar RealScalar; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + + typedef Map, Aligned> MappedDest; + + ActualLhsType actualLhs = LhsBlasTraits::extract(lhs); + ActualRhsType actualRhs = RhsBlasTraits::extract(rhs); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); + + enum { + // FIXME find a way to allow an inner stride on the result if packet_traits::size==1 + // on, the other hand it is good for the cache to pack the vector anyways... + EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1, + ComplexByReal = (NumTraits::IsComplex) && (!NumTraits::IsComplex), + MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal + }; + + gemv_static_vector_if static_dest; + + bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0)); + bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible; + + RhsScalar compatibleAlpha = get_factor::run(actualAlpha); + + ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), + evalToDest ? dest.data() : static_dest.data()); + + if(!evalToDest) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + int size = dest.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + if(!alphaIsCompatible) + { + MappedDest(actualDestPtr, dest.size()).setZero(); + compatibleAlpha = RhsScalar(1); + } + else + MappedDest(actualDestPtr, dest.size()) = dest; + } + + general_matrix_vector_product + ::run( + actualLhs.rows(), actualLhs.cols(), + actualLhs.data(), actualLhs.outerStride(), + actualRhs.data(), actualRhs.innerStride(), + actualDestPtr, 1, + compatibleAlpha); + + if (!evalToDest) + { + if(!alphaIsCompatible) + dest += actualAlpha * MappedDest(actualDestPtr, dest.size()); + else + dest = MappedDest(actualDestPtr, dest.size()); + } + } +}; + +template<> struct gemv_selector +{ + template + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Dest::Index Index; + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + typedef typename Dest::RealScalar RealScalar; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all::type ActualRhsTypeCleaned; + + typename add_const::type actualLhs = LhsBlasTraits::extract(lhs); + typename add_const::type actualRhs = RhsBlasTraits::extract(rhs); + + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); + + enum { + // FIXME find a way to allow an inner stride on the result if packet_traits::size==1 + // on, the other hand it is good for the cache to pack the vector anyways... + DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 + }; + + gemv_static_vector_if static_rhs; + + ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), + DirectlyUseRhs ? const_cast(actualRhs.data()) : static_rhs.data()); + + if(!DirectlyUseRhs) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + int size = actualRhs.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + Map(actualRhsPtr, actualRhs.size()) = actualRhs; + } + + general_matrix_vector_product + ::run( + actualLhs.rows(), actualLhs.cols(), + actualLhs.data(), actualLhs.outerStride(), + actualRhsPtr, 1, + dest.data(), dest.innerStride(), + actualAlpha); + } +}; + +template<> struct gemv_selector +{ + template + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Dest::Index Index; + // TODO makes sure dest is sequentially stored in memory, otherwise use a temp + const Index size = rhs.rows(); + for(Index k=0; k struct gemv_selector +{ + template + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) + { + typedef typename Dest::Index Index; + // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp + const Index rows = dest.rows(); + for(Index i=0; i }; #ifndef EIGEN_TEST_EVALUATORS + // here we need to overload the nested rule for products // such that the nested type is a const reference to a plain matrix namespace internal { diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 99c48ae52..5f9cf69a2 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -307,7 +307,7 @@ struct generic_product_impl internal::gemv_selector::HasUsableDirectAccess) - >::run(GeneralProduct(lhs,rhs), dst, alpha); + >::run(lhs, rhs, dst, alpha); } }; From c98881e1306c94c53ad3de77f9e9036d98dbcf2a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 23 Feb 2014 22:51:13 +0100 Subject: [PATCH 0334/1103] By-pass ProductBase for triangular and selfadjoint products and get rid of ProductBase --- Eigen/src/Core/GeneralProduct.h | 1 - Eigen/src/Core/ProductBase.h | 4 +- Eigen/src/Core/ProductEvaluators.h | 20 +-- Eigen/src/Core/SelfAdjointView.h | 2 + .../Core/products/SelfadjointMatrixMatrix.h | 52 ++++++++ .../Core/products/SelfadjointMatrixVector.h | 104 ++++++++++++++++ .../Core/products/TriangularMatrixMatrix.h | 53 +++++++- .../Core/products/TriangularMatrixVector.h | 117 +++++++++++------- 8 files changed, 299 insertions(+), 54 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 57d5d3c38..06aa05ee6 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -688,7 +688,6 @@ template<> struct gemv_selector typedef typename Lhs::Scalar LhsScalar; typedef typename Rhs::Scalar RhsScalar; typedef typename Dest::Scalar ResScalar; - typedef typename Dest::RealScalar RealScalar; typedef internal::blas_traits LhsBlasTraits; typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; diff --git a/Eigen/src/Core/ProductBase.h b/Eigen/src/Core/ProductBase.h index f7cef9a9e..3b2246fd8 100644 --- a/Eigen/src/Core/ProductBase.h +++ b/Eigen/src/Core/ProductBase.h @@ -11,6 +11,8 @@ #define EIGEN_PRODUCTBASE_H namespace Eigen { + +#ifndef EIGEN_TEST_EVALUATORS /** \class ProductBase * \ingroup Core_Module @@ -174,8 +176,6 @@ class ProductBase : public MatrixBase mutable PlainObject m_result; }; -#ifndef EIGEN_TEST_EVALUATORS - // here we need to overload the nested rule for products // such that the nested type is a const reference to a plain matrix namespace internal { diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 5f9cf69a2..186ae4a34 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -532,6 +532,10 @@ struct etor_product_packet_impl /*************************************************************************** * Triangular products ***************************************************************************/ +template +struct triangular_product_impl; template struct generic_product_impl @@ -542,8 +546,8 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - // TODO bypass TriangularProduct class - TriangularProduct(lhs.nestedExpression(),rhs).scaleAndAddTo(dst, alpha); + triangular_product_impl + ::run(dst, lhs.nestedExpression(), rhs, alpha); } }; @@ -576,8 +580,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - // TODO bypass TriangularProduct class - TriangularProduct(lhs,rhs.nestedExpression()).scaleAndAddTo(dst, alpha); + triangular_product_impl::run(dst, lhs, rhs.nestedExpression(), alpha); } }; @@ -605,6 +608,9 @@ protected: /*************************************************************************** * SelfAdjoint products ***************************************************************************/ +template +struct selfadjoint_product_impl; template struct generic_product_impl @@ -615,8 +621,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - // TODO bypass SelfadjointProductMatrix class - SelfadjointProductMatrix(lhs.nestedExpression(),rhs).scaleAndAddTo(dst, alpha); + selfadjoint_product_impl::run(dst, lhs.nestedExpression(), rhs, alpha); } }; @@ -649,8 +654,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - // TODO bypass SelfadjointProductMatrix class - SelfadjointProductMatrix(lhs,rhs.nestedExpression()).scaleAndAddTo(dst, alpha); + selfadjoint_product_impl::run(dst, lhs, rhs.nestedExpression(), alpha); } }; diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 2cc1815fd..f7f512cf4 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -45,9 +45,11 @@ struct traits > : traits }; } +#ifndef EIGEN_TEST_EVALUATORS template struct SelfadjointProductMatrix; +#endif // FIXME could also be called SelfAdjointWrapper to be consistent with DiagonalWrapper ?? template class SelfAdjointView diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index 99cf9e0ae..f252aef85 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -381,6 +381,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix struct traits > @@ -430,6 +431,57 @@ struct SelfadjointProductMatrix ); } }; +#else // EIGEN_TEST_EVALUATORS +namespace internal { + +template +struct selfadjoint_product_impl +{ + typedef typename Product::Scalar Scalar; + typedef typename Product::Index Index; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + + enum { + LhsIsUpper = (LhsMode&(Upper|Lower))==Upper, + LhsIsSelfAdjoint = (LhsMode&SelfAdjoint)==SelfAdjoint, + RhsIsUpper = (RhsMode&(Upper|Lower))==Upper, + RhsIsSelfAdjoint = (RhsMode&SelfAdjoint)==SelfAdjoint + }; + + template + static void run(Dest &dst, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) + { + eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols()); + + typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + internal::product_selfadjoint_matrix::Flags &RowMajorBit) ? RowMajor : ColMajor, LhsIsSelfAdjoint, + NumTraits::IsComplex && EIGEN_LOGICAL_XOR(LhsIsUpper,bool(LhsBlasTraits::NeedToConjugate)), + EIGEN_LOGICAL_XOR(RhsIsUpper,internal::traits::Flags &RowMajorBit) ? RowMajor : ColMajor, RhsIsSelfAdjoint, + NumTraits::IsComplex && EIGEN_LOGICAL_XOR(RhsIsUpper,bool(RhsBlasTraits::NeedToConjugate)), + internal::traits::Flags&RowMajorBit ? RowMajor : ColMajor> + ::run( + lhs.rows(), rhs.cols(), // sizes + &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info + &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info + &dst.coeffRef(0,0), dst.outerStride(), // result info + actualAlpha // alpha + ); + } +}; + +} // end namespace internal + +#endif } // end namespace Eigen diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index f698f67f9..ddc07d535 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -168,6 +168,7 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product struct traits > @@ -276,6 +277,109 @@ struct SelfadjointProductMatrix } }; +#else // EIGEN_TEST_EVALUATORS + +namespace internal { + +template +struct selfadjoint_product_impl +{ + typedef typename Product::Scalar Scalar; + typedef typename Product::Index Index; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef typename internal::remove_all::type ActualLhsTypeCleaned; + + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all::type ActualRhsTypeCleaned; + + enum { LhsUpLo = LhsMode&(Upper|Lower) }; + + template + static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) + { + typedef typename Dest::Scalar ResScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef Map, Aligned> MappedDest; + + eigen_assert(dest.rows()==a_lhs.rows() && dest.cols()==a_rhs.cols()); + + typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + enum { + EvalToDest = (Dest::InnerStrideAtCompileTime==1), + UseRhs = (ActualRhsTypeCleaned::InnerStrideAtCompileTime==1) + }; + + internal::gemv_static_vector_if static_dest; + internal::gemv_static_vector_if static_rhs; + + ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), + EvalToDest ? dest.data() : static_dest.data()); + + ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,rhs.size(), + UseRhs ? const_cast(rhs.data()) : static_rhs.data()); + + if(!EvalToDest) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + int size = dest.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + MappedDest(actualDestPtr, dest.size()) = dest; + } + + if(!UseRhs) + { + #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN + int size = rhs.size(); + EIGEN_DENSE_STORAGE_CTOR_PLUGIN + #endif + Map(actualRhsPtr, rhs.size()) = rhs; + } + + + internal::selfadjoint_matrix_vector_product::Flags&RowMajorBit) ? RowMajor : ColMajor, + int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run + ( + lhs.rows(), // size + &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info + actualRhsPtr, 1, // rhs info + actualDestPtr, // result info + actualAlpha // scale factor + ); + + if(!EvalToDest) + dest = MappedDest(actualDestPtr, dest.size()); + } +}; + +template +struct selfadjoint_product_impl +{ + typedef typename Product::Scalar Scalar; + enum { RhsUpLo = RhsMode&(Upper|Lower) }; + + template + static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha) + { + // let's simply transpose the product + Transpose destT(dest); + selfadjoint_product_impl, int(RhsUpLo)==Upper ? Lower : Upper, false, + Transpose, 0, true>::run(destT, a_rhs.transpose(), a_lhs.transpose(), alpha); + } +}; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS + } // end namespace Eigen #endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index 8110507b5..e654b45b1 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -372,7 +372,6 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix struct traits > : traits, Lhs, Rhs> > @@ -380,6 +379,7 @@ struct traits > } // end namespace internal +#ifndef EIGEN_TEST_EVALUATORS template struct TriangularProduct : public ProductBase, Lhs, Rhs > @@ -421,6 +421,57 @@ struct TriangularProduct ); } }; +#else // EIGEN_TEST_EVALUATORS +namespace internal { +template +struct triangular_product_impl +{ + template static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha) + { + typedef typename Dest::Index Index; + typedef typename Dest::Scalar Scalar; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef typename internal::remove_all::type ActualLhsTypeCleaned; + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all::type ActualRhsTypeCleaned; + + typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(a_lhs); + typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(a_rhs); + + Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs) + * RhsBlasTraits::extractScalarFactor(a_rhs); + + typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, + Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType; + + enum { IsLower = (Mode&Lower) == Lower }; + Index stripedRows = ((!LhsIsTriangular) || (IsLower)) ? lhs.rows() : (std::min)(lhs.rows(),lhs.cols()); + Index stripedCols = ((LhsIsTriangular) || (!IsLower)) ? rhs.cols() : (std::min)(rhs.cols(),rhs.rows()); + Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows())) + : ((IsLower) ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols())); + + BlockingType blocking(stripedRows, stripedCols, stripedDepth); + + internal::product_triangular_matrix_matrix::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate, + (internal::traits::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate, + (internal::traits::Flags&RowMajorBit) ? RowMajor : ColMajor> + ::run( + stripedRows, stripedCols, stripedDepth, // sizes + &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info + &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info + &dst.coeffRef(0,0), dst.outerStride(), // result info + actualAlpha, blocking + ); + } +}; + +} // end namespace internal +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/products/TriangularMatrixVector.h b/Eigen/src/Core/products/TriangularMatrixVector.h index 6117d5a82..eed7f4258 100644 --- a/Eigen/src/Core/products/TriangularMatrixVector.h +++ b/Eigen/src/Core/products/TriangularMatrixVector.h @@ -168,11 +168,12 @@ struct traits > {}; -template +template struct trmv_selector; } // end namespace internal +#ifndef EIGEN_TEST_EVALUATORS template struct TriangularProduct : public ProductBase, Lhs, Rhs > @@ -185,7 +186,7 @@ struct TriangularProduct { eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols()); - internal::trmv_selector<(int(internal::traits::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(*this, dst, alpha); + internal::trmv_selector::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(m_lhs, m_rhs, dst, alpha); } }; @@ -201,39 +202,71 @@ struct TriangularProduct { eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols()); - typedef TriangularProduct<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower),true,Transpose,false,Transpose,true> TriangularProductTranspose; Transpose dstT(dst); - internal::trmv_selector<(int(internal::traits::Flags)&RowMajorBit) ? ColMajor : RowMajor>::run( - TriangularProductTranspose(m_rhs.transpose(),m_lhs.transpose()), dstT, alpha); + internal::trmv_selector<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower), + (int(internal::traits::Flags)&RowMajorBit) ? ColMajor : RowMajor> + ::run(m_rhs.transpose(),m_lhs.transpose(), dstT, alpha); } }; +#else // EIGEN_TEST_EVALUATORS +namespace internal { + +template +struct triangular_product_impl +{ + template static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha) + { + eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols()); + + internal::trmv_selector::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(lhs, rhs, dst, alpha); + } +}; + +template +struct triangular_product_impl +{ + template static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha) + { + eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols()); + + Transpose dstT(dst); + internal::trmv_selector<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower), + (int(internal::traits::Flags)&RowMajorBit) ? ColMajor : RowMajor> + ::run(rhs.transpose(),lhs.transpose(), dstT, alpha); + } +}; + +} // end namespace internal +#endif // EIGEN_TEST_EVALUATORS + namespace internal { // TODO: find a way to factorize this piece of code with gemv_selector since the logic is exactly the same. -template<> struct trmv_selector +template struct trmv_selector { - template - static void run(const TriangularProduct& prod, Dest& dest, const typename TriangularProduct::Scalar& alpha) + template + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef TriangularProduct ProductType; - typedef typename ProductType::Index Index; - typedef typename ProductType::LhsScalar LhsScalar; - typedef typename ProductType::RhsScalar RhsScalar; - typedef typename ProductType::Scalar ResScalar; - typedef typename ProductType::RealScalar RealScalar; - typedef typename ProductType::ActualLhsType ActualLhsType; - typedef typename ProductType::ActualRhsType ActualRhsType; - typedef typename ProductType::LhsBlasTraits LhsBlasTraits; - typedef typename ProductType::RhsBlasTraits RhsBlasTraits; + typedef typename Dest::Index Index; + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + typedef typename Dest::RealScalar RealScalar; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef Map, Aligned> MappedDest; - typename internal::add_const_on_value_type::type actualLhs = LhsBlasTraits::extract(prod.lhs()); - typename internal::add_const_on_value_type::type actualRhs = RhsBlasTraits::extract(prod.rhs()); + typename internal::add_const_on_value_type::type actualLhs = LhsBlasTraits::extract(lhs); + typename internal::add_const_on_value_type::type actualRhs = RhsBlasTraits::extract(rhs); - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) - * RhsBlasTraits::extractScalarFactor(prod.rhs()); + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); enum { // FIXME find a way to allow an inner stride on the result if packet_traits::size==1 @@ -288,33 +321,33 @@ template<> struct trmv_selector } }; -template<> struct trmv_selector +template struct trmv_selector { - template - static void run(const TriangularProduct& prod, Dest& dest, const typename TriangularProduct::Scalar& alpha) + template + static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef TriangularProduct ProductType; - typedef typename ProductType::LhsScalar LhsScalar; - typedef typename ProductType::RhsScalar RhsScalar; - typedef typename ProductType::Scalar ResScalar; - typedef typename ProductType::Index Index; - typedef typename ProductType::ActualLhsType ActualLhsType; - typedef typename ProductType::ActualRhsType ActualRhsType; - typedef typename ProductType::_ActualRhsType _ActualRhsType; - typedef typename ProductType::LhsBlasTraits LhsBlasTraits; - typedef typename ProductType::RhsBlasTraits RhsBlasTraits; + typedef typename Dest::Index Index; + typedef typename Lhs::Scalar LhsScalar; + typedef typename Rhs::Scalar RhsScalar; + typedef typename Dest::Scalar ResScalar; + + typedef internal::blas_traits LhsBlasTraits; + typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; + typedef internal::blas_traits RhsBlasTraits; + typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; + typedef typename internal::remove_all::type ActualRhsTypeCleaned; - typename add_const::type actualLhs = LhsBlasTraits::extract(prod.lhs()); - typename add_const::type actualRhs = RhsBlasTraits::extract(prod.rhs()); + typename add_const::type actualLhs = LhsBlasTraits::extract(lhs); + typename add_const::type actualRhs = RhsBlasTraits::extract(rhs); - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) - * RhsBlasTraits::extractScalarFactor(prod.rhs()); + ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs) + * RhsBlasTraits::extractScalarFactor(rhs); enum { - DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1 + DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1 }; - gemv_static_vector_if static_rhs; + gemv_static_vector_if static_rhs; ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), DirectlyUseRhs ? const_cast(actualRhs.data()) : static_rhs.data()); @@ -325,7 +358,7 @@ template<> struct trmv_selector int size = actualRhs.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif - Map(actualRhsPtr, actualRhs.size()) = actualRhs; + Map(actualRhsPtr, actualRhs.size()) = actualRhs; } internal::triangular_matrix_vector_product From 1e0c2f6ddb60c85440fceb3503769bbee80452c2 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 24 Feb 2014 11:41:19 +0100 Subject: [PATCH 0335/1103] Hide some deprecated classes. --- Eigen/src/Core/GeneralProduct.h | 8 +++++++- Eigen/src/Core/Swap.h | 4 ++-- Eigen/src/LU/FullPivLU.h | 4 ++-- Eigen/src/LU/PartialPivLU.h | 4 ++-- 4 files changed, 13 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 06aa05ee6..c37526b72 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -13,6 +13,7 @@ namespace Eigen { +#ifndef EIGEN_TEST_EVALUATORS /** \class GeneralProduct * \ingroup Core_Module * @@ -34,6 +35,8 @@ namespace Eigen { */ template::value> class GeneralProduct; +#endif // EIGEN_TEST_EVALUATORS + enum { Large = 2, @@ -154,6 +157,7 @@ template<> struct product_type_selector { enum } // end namespace internal +#ifndef EIGEN_TEST_EVALUATORS /** \class ProductReturnType * \ingroup Core_Module * @@ -201,6 +205,7 @@ struct ProductReturnType template struct LazyProductReturnType : public ProductReturnType {}; +#endif /*********************************************************************** * Implementation of Inner Vector Vector Product @@ -212,6 +217,7 @@ struct LazyProductReturnType : public ProductReturnType with: operator=(Scalar x); +#ifndef EIGEN_TEST_EVALUATORS namespace internal { @@ -242,7 +248,7 @@ class GeneralProduct return Base::coeff(0,0); } }; - +#endif // EIGEN_TEST_EVALUATORS /*********************************************************************** * Implementation of Outer Vector Vector Product ***********************************************************************/ diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index e7d525572..9a1c5f4f8 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -12,7 +12,7 @@ namespace Eigen { -// #ifndef EIGEN_TEST_EVALUATORS +#ifndef EIGEN_TEST_EVALUATORS /** \class SwapWrapper * \ingroup Core_Module @@ -137,7 +137,7 @@ template class SwapWrapper ExpressionType& m_expression; }; -// #endif +#endif #ifdef EIGEN_ENABLE_EVALUATORS diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index 86de91ccb..e0236fd9e 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -770,7 +770,7 @@ struct solve_retval, Rhs> #endif /***** Implementation of inverse() *****************************************************/ - +#ifdef EIGEN_TEST_EVALUATORS template struct Assignment >, internal::assign_op, Dense2Dense, Scalar> { @@ -781,7 +781,7 @@ struct Assignment >, internal::assign_ dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; - +#endif } // end namespace internal /******* MatrixBase methods *****************************************************************/ diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index ac53f7ab5..65312f7d1 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -505,7 +505,7 @@ struct solve_retval, Rhs> #endif /***** Implementation of inverse() *****************************************************/ - +#ifdef EIGEN_TEST_EVALUATORS template struct Assignment >, internal::assign_op, Dense2Dense, Scalar> { @@ -516,7 +516,7 @@ struct Assignment >, internal::assi dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; - +#endif } // end namespace internal /******** MatrixBase methods *******/ From cbc572caf7c01854076d6ff84e5ae864b490458a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 24 Feb 2014 11:49:30 +0100 Subject: [PATCH 0336/1103] Split LU/Inverse.h to Core/Inverse.h for the generic Inverse expression, and LU/InverseImpl.h for the dense implementation of dense.inverse() --- Eigen/Core | 1 + Eigen/LU | 2 +- Eigen/src/Core/Inverse.h | 131 ++++++++++++++++++++++ Eigen/src/LU/{Inverse.h => InverseImpl.h} | 117 +------------------ 4 files changed, 138 insertions(+), 113 deletions(-) create mode 100644 Eigen/src/Core/Inverse.h rename Eigen/src/LU/{Inverse.h => InverseImpl.h} (83%) diff --git a/Eigen/Core b/Eigen/Core index 245604465..474ef5f8f 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -365,6 +365,7 @@ using std::ptrdiff_t; #include "src/Core/GeneralProduct.h" #ifdef EIGEN_ENABLE_EVALUATORS #include "src/Core/Solve.h" +#include "src/Core/Inverse.h" #endif #include "src/Core/TriangularMatrix.h" #include "src/Core/SelfAdjointView.h" diff --git a/Eigen/LU b/Eigen/LU index db5795504..fac80024f 100644 --- a/Eigen/LU +++ b/Eigen/LU @@ -25,7 +25,7 @@ #include "src/LU/PartialPivLU_MKL.h" #endif #include "src/LU/Determinant.h" -#include "src/LU/Inverse.h" +#include "src/LU/InverseImpl.h" #if defined EIGEN_VECTORIZE_SSE #include "src/LU/arch/Inverse_SSE.h" diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h new file mode 100644 index 000000000..0cfc71d34 --- /dev/null +++ b/Eigen/src/Core/Inverse.h @@ -0,0 +1,131 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_INVERSE_H +#define EIGEN_INVERSE_H + +namespace Eigen { + +#ifdef EIGEN_TEST_EVALUATORS + +// TODO move the general declaration in Core, and rename this file DenseInverseImpl.h, or something like this... + +template class InverseImpl; + +namespace internal { + +template +struct traits > + : traits +{ + typedef typename XprType::PlainObject PlainObject; + typedef traits BaseTraits; + enum { + Flags = BaseTraits::Flags & RowMajorBit, + CoeffReadCost = Dynamic + }; +}; + +} // end namespace internal + +/** \class Inverse + * + * \brief Expression of the inverse of another expression + * + * \tparam XprType the type of the expression we are taking the inverse + * + * This class represents an abstract expression of A.inverse() + * and most of the time this is the only way it is used. + * + */ +template +class Inverse : public InverseImpl::StorageKind> +{ +public: + typedef typename XprType::Index Index; + typedef typename XprType::PlainObject PlainObject; + typedef typename internal::nested::type XprTypeNested; + typedef typename internal::remove_all::type XprTypeNestedCleaned; + + Inverse(const XprType &xpr) + : m_xpr(xpr) + {} + + EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); } + + EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; } + +protected: + XprTypeNested &m_xpr; +}; + +/** \internal + * Specialization of the Inverse expression for dense expressions. + * Direct access to the coefficients are discared. + * FIXME this intermediate class is probably not needed anymore. + */ +template +class InverseImpl + : public MatrixBase > +{ + typedef Inverse Derived; + +public: + + typedef MatrixBase Base; + EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + +private: + + Scalar coeff(Index row, Index col) const; + Scalar coeff(Index i) const; +}; + +namespace internal { + +/** \internal + * \brief Default evaluator for Inverse expression. + * + * This default evaluator for Inverse expression simply evaluate the inverse into a temporary + * by a call to internal::call_assignment_no_alias. + * Therefore, inverse implementers only have to specialize Assignment, ...> for + * there own nested expression. + * + * \sa class Inverse + */ +template +struct evaluator > + : public evaluator::PlainObject>::type +{ + typedef Inverse InverseType; + typedef typename InverseType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const InverseType& inv_xpr) + : m_result(inv_xpr.rows(), inv_xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + internal::call_assignment_no_alias(m_result, inv_xpr); + } + +protected: + PlainObject m_result; +}; + +} // end namespace internal + +#endif + +} // end namespace Eigen + +#endif // EIGEN_INVERSE_H diff --git a/Eigen/src/LU/Inverse.h b/Eigen/src/LU/InverseImpl.h similarity index 83% rename from Eigen/src/LU/Inverse.h rename to Eigen/src/LU/InverseImpl.h index db053406e..174dfbac5 100644 --- a/Eigen/src/LU/Inverse.h +++ b/Eigen/src/LU/InverseImpl.h @@ -2,13 +2,14 @@ // for linear algebra. // // Copyright (C) 2008-2010 Benoit Jacob +// Copyright (C) 2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. -#ifndef EIGEN_INVERSE_H -#define EIGEN_INVERSE_H +#ifndef EIGEN_INVERSE_IMPL_H +#define EIGEN_INVERSE_IMPL_H namespace Eigen { @@ -324,117 +325,9 @@ struct inverse_impl : public ReturnByValue > #ifdef EIGEN_TEST_EVALUATORS -// TODO move the general declaration in Core, and rename this file DenseInverseImpl.h, or something like this... - -template class InverseImpl; - namespace internal { -template -struct traits > - : traits -{ - typedef typename XprType::PlainObject PlainObject; - typedef traits BaseTraits; - enum { - Flags = BaseTraits::Flags & RowMajorBit, - CoeffReadCost = Dynamic - }; -}; - -} // end namespace internal - -/** \class Inverse - * \ingroup LU_Module - * - * \brief Expression of the inverse of another expression - * - * \tparam XprType the type of the expression we are taking the inverse - * - * This class represents an abstract expression of A.inverse() - * and most of the time this is the only way it is used. - * - */ -template -class Inverse : public InverseImpl::StorageKind> -{ -public: - typedef typename XprType::Index Index; - typedef typename XprType::PlainObject PlainObject; - typedef typename internal::nested::type XprTypeNested; - typedef typename internal::remove_all::type XprTypeNestedCleaned; - - Inverse(const XprType &xpr) - : m_xpr(xpr) - {} - - EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); } - EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); } - - EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; } - -protected: - XprTypeNested &m_xpr; -}; - -/** \internal - * Specialization of the Inverse expression for dense expressions. - * Direct access to the coefficients are discared. - * FIXME this intermediate class is probably not needed anymore. - */ -template -class InverseImpl - : public MatrixBase > -{ - typedef Inverse Derived; - -public: - - typedef MatrixBase Base; - EIGEN_DENSE_PUBLIC_INTERFACE(Derived) - -private: - - Scalar coeff(Index row, Index col) const; - Scalar coeff(Index i) const; -}; - -namespace internal { - -/** \internal - * \brief Default evaluator for Inverse expression. - * - * This default evaluator for Inverse expression simply evaluate the inverse into a temporary - * by a call to internal::call_assignment_no_alias. - * Therefore, inverse implementers only have to specialize Assignment, ...> for - * there own nested expression. - * - * \sa class Inverse - */ -template -struct evaluator > - : public evaluator::PlainObject>::type -{ - typedef Inverse InverseType; - typedef typename InverseType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - typedef evaluator type; - typedef evaluator nestedType; - - evaluator(const InverseType& inv_xpr) - : m_result(inv_xpr.rows(), inv_xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - internal::call_assignment_no_alias(m_result, inv_xpr); - } - -protected: - PlainObject m_result; -}; - -// Specialization for "dst = xpr.inverse()" -// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere +// Specialization for "dense = dense_xpr.inverse()" template struct Assignment, internal::assign_op, Dense2Dense, Scalar> { @@ -569,4 +462,4 @@ inline void MatrixBase::computeInverseWithCheck( } // end namespace Eigen -#endif // EIGEN_INVERSE_H +#endif // EIGEN_INVERSE_IMPL_H From 3e439889e0dd10f67d328f3c688178d1e6c091d8 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 24 Feb 2014 13:12:10 +0100 Subject: [PATCH 0337/1103] Specify what non-resizeable objects are in transposeInPlace and adjointInPlace (cf bug #749) --- Eigen/src/Core/Transpose.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 976708a0f..aba3f6670 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -293,7 +293,8 @@ struct inplace_transpose_selector { // non square matrix * Notice however that this method is only useful if you want to replace a matrix by its own transpose. * If you just need the transpose of a matrix, use transpose(). * - * \note if the matrix is not square, then \c *this must be a resizable matrix. + * \note if the matrix is not square, then \c *this must be a resizable matrix. + * This excludes (non-square) fixed-size matrices, block-expressions and maps. * * \sa transpose(), adjoint(), adjointInPlace() */ template @@ -324,6 +325,7 @@ inline void DenseBase::transposeInPlace() * If you just need the adjoint of a matrix, use adjoint(). * * \note if the matrix is not square, then \c *this must be a resizable matrix. + * This excludes (non-square) fixed-size matrices, block-expressions and maps. * * \sa transpose(), adjoint(), transposeInPlace() */ template From 6fecb6f1b6d17883525f3f34dd5c1860faeecbd8 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Mon, 24 Feb 2014 14:10:17 +0000 Subject: [PATCH 0338/1103] Fix bug #748 - array_5 test fails for seed 1392781168. --- test/array.cpp | 13 +++---------- 1 file changed, 3 insertions(+), 10 deletions(-) diff --git a/test/array.cpp b/test/array.cpp index f1deda7e3..5f49fc1ea 100644 --- a/test/array.cpp +++ b/test/array.cpp @@ -173,21 +173,14 @@ template void array_real(const ArrayType& m) Scalar s1 = internal::random(); // these tests are mostly to check possible compilation issues. -// VERIFY_IS_APPROX(m1.sin(), std::sin(m1)); VERIFY_IS_APPROX(m1.sin(), sin(m1)); -// VERIFY_IS_APPROX(m1.cos(), std::cos(m1)); VERIFY_IS_APPROX(m1.cos(), cos(m1)); -// VERIFY_IS_APPROX(m1.asin(), std::asin(m1)); VERIFY_IS_APPROX(m1.asin(), asin(m1)); -// VERIFY_IS_APPROX(m1.acos(), std::acos(m1)); VERIFY_IS_APPROX(m1.acos(), acos(m1)); -// VERIFY_IS_APPROX(m1.tan(), std::tan(m1)); VERIFY_IS_APPROX(m1.tan(), tan(m1)); VERIFY_IS_APPROX(cos(m1+RealScalar(3)*m2), cos((m1+RealScalar(3)*m2).eval())); -// VERIFY_IS_APPROX(std::cos(m1+RealScalar(3)*m2), std::cos((m1+RealScalar(3)*m2).eval())); -// VERIFY_IS_APPROX(m1.abs().sqrt(), std::sqrt(std::abs(m1))); VERIFY_IS_APPROX(m1.abs().sqrt(), sqrt(abs(m1))); VERIFY_IS_APPROX(m1.abs(), sqrt(numext::abs2(m1))); @@ -196,9 +189,10 @@ template void array_real(const ArrayType& m) if(!NumTraits::IsComplex) VERIFY_IS_APPROX(numext::real(m1), m1); - VERIFY_IS_APPROX(m1.abs().log() , log(abs(m1))); + // shift argument of logarithm so that it is not zero + Scalar smallNumber = NumTraits::dummy_precision(); + VERIFY_IS_APPROX((m1.abs() + smallNumber).log() , log(abs(m1) + smallNumber)); -// VERIFY_IS_APPROX(m1.exp(), std::exp(m1)); VERIFY_IS_APPROX(m1.exp() * m2.exp(), exp(m1+m2)); VERIFY_IS_APPROX(m1.exp(), exp(m1)); VERIFY_IS_APPROX(m1.exp() / m2.exp(),(m1-m2).exp()); @@ -242,7 +236,6 @@ template void array_complex(const ArrayType& m) m2(i,j) = sqrt(m1(i,j)); VERIFY_IS_APPROX(m1.sqrt(), m2); -// VERIFY_IS_APPROX(m1.sqrt(), std::sqrt(m1)); VERIFY_IS_APPROX(m1.sqrt(), Eigen::sqrt(m1)); } From 21fecd5252061c05cfb9f91e4e9becf25089949d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 24 Feb 2014 17:12:17 +0100 Subject: [PATCH 0339/1103] Workaround clang ABI change with unsed arguments (ugly fix) --- .../Core/products/GeneralBlockPanelKernel.h | 12 +++++++ Eigen/src/Core/products/GeneralMatrixVector.h | 31 ++++++++++++------- 2 files changed, 31 insertions(+), 12 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 780fa74d3..10cd34e9e 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -1130,6 +1130,10 @@ EIGEN_DONT_INLINE void gemm_pack_lhs=depth && offset<=stride)); eigen_assert( (StorageOrder==RowMajor) || ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) ); +#ifdef __clang__ + // Workaround clang ABI change with unsed arguments + if(!PanelMode) depth += stride + offset; +#endif conj_if::IsComplex && Conjugate> cj; const_blas_data_mapper lhs(_lhs,lhsStride); Index count = 0; @@ -1216,6 +1220,10 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=depth && offset<=stride)); +#ifdef __clang__ + // Workaround clang ABI change with unsed arguments + if(!PanelMode) depth += stride + offset; +#endif conj_if::IsComplex && Conjugate> cj; Index packet_cols = (cols/nr) * nr; Index count = 0; @@ -1267,6 +1275,10 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=depth && offset<=stride)); +#ifdef __clang__ + // Workaround clang ABI change with unsed arguments + if(!PanelMode) depth += stride + offset; +#endif conj_if::IsComplex && Conjugate> cj; Index packet_cols = (cols/nr) * nr; Index count = 0; diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h index c8697c913..c3e1e7142 100644 --- a/Eigen/src/Core/products/GeneralMatrixVector.h +++ b/Eigen/src/Core/products/GeneralMatrixVector.h @@ -80,11 +80,8 @@ EIGEN_DONT_INLINE static void run( Index rows, Index cols, const LhsScalar* lhs, Index lhsStride, const RhsScalar* rhs, Index rhsIncr, - ResScalar* res, Index - #ifdef EIGEN_INTERNAL_DEBUGGING - resIncr - #endif - , RhsScalar alpha); + ResScalar* res, Index resIncr, + RhsScalar alpha); }; template @@ -92,13 +89,17 @@ EIGEN_DONT_INLINE void general_matrix_vector_product Date: Mon, 24 Feb 2014 18:13:49 +0100 Subject: [PATCH 0340/1103] Implement bug #317: use a template function call to suppress unused variable warnings. This also fix the issue of the previous changeset in a much nicer way. --- .../Core/products/GeneralBlockPanelKernel.h | 18 ++++++------------ Eigen/src/Core/products/GeneralMatrixVector.h | 15 ++------------- Eigen/src/Core/util/Macros.h | 7 ++++++- 3 files changed, 14 insertions(+), 26 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 10cd34e9e..08cc14bd7 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -1128,12 +1128,10 @@ EIGEN_DONT_INLINE void gemm_pack_lhs::size }; EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS"); + EIGEN_UNUSED_VARIABLE(stride); + EIGEN_UNUSED_VARIABLE(offset); eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride)); eigen_assert( (StorageOrder==RowMajor) || ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) ); -#ifdef __clang__ - // Workaround clang ABI change with unsed arguments - if(!PanelMode) depth += stride + offset; -#endif conj_if::IsComplex && Conjugate> cj; const_blas_data_mapper lhs(_lhs,lhsStride); Index count = 0; @@ -1219,11 +1217,9 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=depth && offset<=stride)); -#ifdef __clang__ - // Workaround clang ABI change with unsed arguments - if(!PanelMode) depth += stride + offset; -#endif conj_if::IsComplex && Conjugate> cj; Index packet_cols = (cols/nr) * nr; Index count = 0; @@ -1274,11 +1270,9 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=depth && offset<=stride)); -#ifdef __clang__ - // Workaround clang ABI change with unsed arguments - if(!PanelMode) depth += stride + offset; -#endif conj_if::IsComplex && Conjugate> cj; Index packet_cols = (cols/nr) * nr; Index count = 0; diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h index c3e1e7142..a73ce5ff0 100644 --- a/Eigen/src/Core/products/GeneralMatrixVector.h +++ b/Eigen/src/Core/products/GeneralMatrixVector.h @@ -92,14 +92,8 @@ EIGEN_DONT_INLINE void general_matrix_vector_product void ignore_unused_variable(const T&) {} + } +} +#define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var); #if !defined(EIGEN_ASM_COMMENT) #if (defined __GNUC__) && ( defined(__i386__) || defined(__x86_64__) ) From db7d49efbb60b30f06f5fa97c767589f01e3df88 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 24 Feb 2014 13:45:32 -0800 Subject: [PATCH 0341/1103] Added support for FMA instructions --- Eigen/Core | 3 +++ Eigen/src/Core/arch/AVX/PacketMath.h | 5 +++++ Eigen/src/Core/arch/SSE/PacketMath.h | 4 ++++ .../src/Core/products/GeneralBlockPanelKernel.h | 16 ++++++++++++++++ 4 files changed, 28 insertions(+) diff --git a/Eigen/Core b/Eigen/Core index bd20d5ac5..dbe68586d 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -117,6 +117,9 @@ #define EIGEN_VECTORIZE_SSE4_1 #define EIGEN_VECTORIZE_SSE4_2 #endif + #ifdef __FMA__ + #define EIGEN_VECTORIZE_FMA + #endif // include files // This extern "C" works around a MINGW-w64 compilation issue diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 244e63e74..d1a134087 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -120,6 +120,11 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv(const Packet8i& /*a*/, co return pset1(0); } +#ifdef EIGEN_VECTORIZE_FMA +template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { return _mm256_fmadd_ps(a,b,c); } +template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { return _mm256_fmadd_pd(a,b,c); } +#endif + template<> EIGEN_STRONG_INLINE Packet8f pmin(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet4d pmin(const Packet4d& a, const Packet4d& b) { return _mm256_min_pd(a,b); } diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index e913af650..a35f9ce98 100644 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -179,6 +179,10 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv(const Packet4i& /*a*/, co // for some weird raisons, it has to be overloaded for packet of integers template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); } +#ifdef EIGEN_VECTORIZE_FMA +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); } +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); } +#endif template<> EIGEN_STRONG_INLINE Packet4f pmin(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); } template<> EIGEN_STRONG_INLINE Packet2d pmin(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); } diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 780fa74d3..9caa15081 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -205,7 +205,15 @@ public: EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, AccPacket& tmp) const { + // It would be a lot cleaner to call pmadd all the time. Unfortunately if we + // let gcc allocate the register in which to store the result of the pmul + // (in the case where there is no FMA) gcc fails to figure out how to avoid + // spilling register. +#ifdef EIGEN_VECTORIZE_FMA + c = pmadd(a,b,c); +#else tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp); +#endif } EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const @@ -281,7 +289,11 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { +#ifdef EIGEN_VECTORIZE_FMA + c.v = pmadd(a.v,b,c.v); +#else tmp = b; tmp = pmul(a.v,tmp); c.v = padd(c.v,tmp); +#endif } EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const @@ -486,7 +498,11 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { +#ifdef EIGEN_VECTORIZE_FMA + c = pmadd(a,b,c); +#else tmp = b; tmp.v = pmul(a,tmp.v); c = padd(c,tmp); +#endif } EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const From a6b130c63c4e1cb6be6bf71414f80feaea922b23 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 5 Mar 2014 10:07:44 +0100 Subject: [PATCH 0342/1103] swap 3.2 <-> default CTestConfig.cmake file --- CTestConfig.cmake | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/CTestConfig.cmake b/CTestConfig.cmake index 0557c491a..4c0027824 100644 --- a/CTestConfig.cmake +++ b/CTestConfig.cmake @@ -4,10 +4,14 @@ ## # The following are required to uses Dart and the Cdash dashboard ## ENABLE_TESTING() ## INCLUDE(CTest) -set(CTEST_PROJECT_NAME "Eigen3.2") +set(CTEST_PROJECT_NAME "Eigen") set(CTEST_NIGHTLY_START_TIME "00:00:00 UTC") set(CTEST_DROP_METHOD "http") set(CTEST_DROP_SITE "manao.inria.fr") -set(CTEST_DROP_LOCATION "/CDash/submit.php?project=Eigen3.2") +set(CTEST_DROP_LOCATION "/CDash/submit.php?project=Eigen") set(CTEST_DROP_SITE_CDASH TRUE) +set(CTEST_PROJECT_SUBPROJECTS +Official +Unsupported +) From 804ef2350d51f6fe4d22f18036dfc7e10a4bb711 Mon Sep 17 00:00:00 2001 From: Abraham Bachrach Date: Sun, 9 Mar 2014 16:56:44 -0700 Subject: [PATCH 0343/1103] Move the Base typedef's from private to public scope Move the Quaternion::Base typedef from private to public scope so that one may create child classes of Quaternion. NOTE: This matches the semantics of MatrixBase. --- Eigen/src/Geometry/Quaternion.h | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index 803001272..6ee0e4f75 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -34,8 +34,9 @@ struct quaternionbase_assign_impl; template class QuaternionBase : public RotationBase { + public: typedef RotationBase Base; -public: + using Base::operator*; using Base::derived; @@ -223,10 +224,10 @@ struct traits > template class Quaternion : public QuaternionBase > { +public: typedef QuaternionBase > Base; enum { IsAligned = internal::traits::IsAligned }; -public: typedef _Scalar Scalar; EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Quaternion) @@ -334,9 +335,9 @@ template class Map, _Options > : public QuaternionBase, _Options> > { + public: typedef QuaternionBase, _Options> > Base; - public: typedef _Scalar Scalar; typedef typename internal::traits::Coefficients Coefficients; EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Map) @@ -371,9 +372,9 @@ template class Map, _Options > : public QuaternionBase, _Options> > { + public: typedef QuaternionBase, _Options> > Base; - public: typedef _Scalar Scalar; typedef typename internal::traits::Coefficients Coefficients; EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Map) From 5c0f294098f628439cf9341edb286d92f7fa138b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 10 Mar 2014 09:28:00 +0100 Subject: [PATCH 0344/1103] Fix evaluators unit test (i.e., when only EIGEN_ENABLE_EVALUATORS is defined --- Eigen/src/Core/GeneralProduct.h | 25 +++--- Eigen/src/Core/ProductEvaluators.h | 2 +- Eigen/src/Core/products/GeneralMatrixMatrix.h | 6 +- .../Core/products/SelfadjointMatrixMatrix.h | 5 +- .../Core/products/TriangularMatrixMatrix.h | 5 +- Eigen/src/Core/util/XprHelper.h | 76 ++++++++++--------- 6 files changed, 67 insertions(+), 52 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index c37526b72..f28d2e46c 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -353,10 +353,13 @@ template struct traits > : traits, Lhs, Rhs> > {}; -#endif - template struct gemv_selector; +#endif +#ifdef EIGEN_ENABLE_EVALUATORS +template +struct gemv_dense_sense_selector; +#endif } // end namespace internal @@ -594,23 +597,25 @@ template<> struct gemv_selector } }; -#else // EIGEN_TEST_EVALUATORS +#endif // EIGEN_TEST_EVALUATORS + +#ifdef EIGEN_ENABLE_EVALUATORS // The vector is on the left => transposition template -struct gemv_selector +struct gemv_dense_sense_selector { template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { Transpose destT(dest); enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; - gemv_selector + gemv_dense_sense_selector ::run(rhs.transpose(), lhs.transpose(), destT, alpha); } }; -template<> struct gemv_selector +template<> struct gemv_dense_sense_selector { template static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) @@ -685,7 +690,7 @@ template<> struct gemv_selector } }; -template<> struct gemv_selector +template<> struct gemv_dense_sense_selector { template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) @@ -737,7 +742,7 @@ template<> struct gemv_selector } }; -template<> struct gemv_selector +template<> struct gemv_dense_sense_selector { template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) @@ -750,7 +755,7 @@ template<> struct gemv_selector } }; -template<> struct gemv_selector +template<> struct gemv_dense_sense_selector { template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) @@ -763,7 +768,7 @@ template<> struct gemv_selector } }; -#endif // EIGEN_TEST_EVALUATORS +#endif // EIGEN_ENABLE_EVALUATORS } // end namespace internal diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 186ae4a34..b16b86d8f 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -304,7 +304,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - internal::gemv_selector::HasUsableDirectAccess) >::run(lhs, rhs, dst, alpha); diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 1c8940e1c..1726f98ed 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -422,7 +422,9 @@ class GeneralProduct internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit); } }; -#else // EIGEN_TEST_EVALUATORS +#endif // EIGEN_TEST_EVALUATORS + +#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template @@ -477,7 +479,7 @@ struct generic_product_impl }; } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS +#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index f252aef85..afa8af43c 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -431,7 +431,8 @@ struct SelfadjointProductMatrix ); } }; -#else // EIGEN_TEST_EVALUATORS +#endif // EIGEN_TEST_EVALUATORS +#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template @@ -481,7 +482,7 @@ struct selfadjoint_product_impl } // end namespace internal -#endif +#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index e654b45b1..3f0618410 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -421,7 +421,8 @@ struct TriangularProduct ); } }; -#else // EIGEN_TEST_EVALUATORS +#endif // EIGEN_TEST_EVALUATORS +#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template struct triangular_product_impl @@ -471,7 +472,7 @@ struct triangular_product_impl }; } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS +#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 76e979ba0..a4baa921c 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -293,44 +293,10 @@ struct transfer_constness >::type type; }; -#ifdef EIGEN_TEST_EVALUATORS -// When using evaluators, we never evaluate when assembling the expression!! -// TODO: get rid of this nested class since it's just an alias for ref_selector. -template::type> struct nested -{ - typedef typename ref_selector::type type; -}; -// However, we still need a mechanism to detect whether an expression which is evaluated multiple time -// has to be evaluated into a temporary. -// That's the purpose of this new nested_eval helper: -template::type> struct nested_eval -{ - enum { - // For the purpose of this test, to keep it reasonably simple, we arbitrarily choose a value of Dynamic values. - // the choice of 10000 makes it larger than any practical fixed value and even most dynamic values. - // in extreme cases where these assumptions would be wrong, we would still at worst suffer performance issues - // (poor choice of temporaries). - // It's important that this value can still be squared without integer overflowing. - DynamicAsInteger = 10000, - ScalarReadCost = NumTraits::Scalar>::ReadCost, - ScalarReadCostAsInteger = ScalarReadCost == Dynamic ? int(DynamicAsInteger) : int(ScalarReadCost), - CoeffReadCost = traits::CoeffReadCost, - CoeffReadCostAsInteger = CoeffReadCost == Dynamic ? int(DynamicAsInteger) : int(CoeffReadCost), - NAsInteger = n == Dynamic ? int(DynamicAsInteger) : n, - CostEvalAsInteger = (NAsInteger+1) * ScalarReadCostAsInteger + CoeffReadCostAsInteger, - CostNoEvalAsInteger = NAsInteger * CoeffReadCostAsInteger - }; +#ifndef EIGEN_TEST_EVALUATORS - typedef typename conditional< - int(CostEvalAsInteger) < int(CostNoEvalAsInteger), - PlainObject, - typename ref_selector::type - >::type type; -}; - -#else /** \internal Determines how a given expression should be nested into another one. * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be * nested into the bigger product expression. The choice is between nesting the expression b+c as-is, or @@ -377,8 +343,48 @@ template::type> str typename ref_selector::type >::type type; }; + +#else + +// When using evaluators, we never evaluate when assembling the expression!! +// TODO: get rid of this nested class since it's just an alias for ref_selector. +template::type> struct nested +{ + typedef typename ref_selector::type type; +}; + #endif // EIGEN_TEST_EVALUATORS +#ifdef EIGEN_ENABLE_EVALUATORS +// However, we still need a mechanism to detect whether an expression which is evaluated multiple time +// has to be evaluated into a temporary. +// That's the purpose of this new nested_eval helper: +template::type> struct nested_eval +{ + enum { + // For the purpose of this test, to keep it reasonably simple, we arbitrarily choose a value of Dynamic values. + // the choice of 10000 makes it larger than any practical fixed value and even most dynamic values. + // in extreme cases where these assumptions would be wrong, we would still at worst suffer performance issues + // (poor choice of temporaries). + // It's important that this value can still be squared without integer overflowing. + DynamicAsInteger = 10000, + ScalarReadCost = NumTraits::Scalar>::ReadCost, + ScalarReadCostAsInteger = ScalarReadCost == Dynamic ? int(DynamicAsInteger) : int(ScalarReadCost), + CoeffReadCost = traits::CoeffReadCost, + CoeffReadCostAsInteger = CoeffReadCost == Dynamic ? int(DynamicAsInteger) : int(CoeffReadCost), + NAsInteger = n == Dynamic ? int(DynamicAsInteger) : n, + CostEvalAsInteger = (NAsInteger+1) * ScalarReadCostAsInteger + CoeffReadCostAsInteger, + CostNoEvalAsInteger = NAsInteger * CoeffReadCostAsInteger + }; + + typedef typename conditional< + int(CostEvalAsInteger) < int(CostNoEvalAsInteger), + PlainObject, + typename ref_selector::type + >::type type; +}; +#endif + template EIGEN_DEVICE_FUNC T* const_cast_ptr(const T* ptr) From 354bd8a42837ac405a142281fca2c4ac59ff701f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 10 Mar 2014 09:30:58 +0100 Subject: [PATCH 0345/1103] Hide legacy dense assignment routines with EIGEN_TEST_EVALUATORS --- Eigen/src/Core/Assign.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index 6080a83f6..95306a9a7 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -13,7 +13,7 @@ #define EIGEN_ASSIGN_H namespace Eigen { - +#ifndef EIGEN_TEST_EVALUATORS namespace internal { /*************************************************************************** @@ -487,6 +487,8 @@ struct assign_impl Date: Mon, 10 Mar 2014 23:24:40 +0100 Subject: [PATCH 0346/1103] Move CoeffReadCost mechanism to evaluators --- Eigen/src/Core/ArrayBase.h | 4 +- Eigen/src/Core/Assign.h | 3 - Eigen/src/Core/AssignEvaluator.h | 64 +++++------ Eigen/src/Core/BooleanRedux.h | 37 ++++--- Eigen/src/Core/CoreEvaluators.h | 111 ++++++++++++++------ Eigen/src/Core/CwiseBinaryOp.h | 9 +- Eigen/src/Core/CwiseNullaryOp.h | 5 +- Eigen/src/Core/CwiseUnaryOp.h | 5 +- Eigen/src/Core/CwiseUnaryView.h | 2 + Eigen/src/Core/DenseBase.h | 2 + Eigen/src/Core/Diagonal.h | 2 + Eigen/src/Core/DiagonalMatrix.h | 5 +- Eigen/src/Core/DiagonalProduct.h | 5 +- Eigen/src/Core/Matrix.h | 2 + Eigen/src/Core/MatrixBase.h | 4 +- Eigen/src/Core/ProductEvaluators.h | 48 ++++++--- Eigen/src/Core/Redux.h | 2 +- Eigen/src/Core/Replicate.h | 5 +- Eigen/src/Core/Reverse.h | 6 +- Eigen/src/Core/Select.h | 5 +- Eigen/src/Core/Transpose.h | 2 + Eigen/src/Core/TriangularMatrix.h | 15 +-- Eigen/src/Core/VectorwiseOp.h | 2 + Eigen/src/Core/Visitor.h | 43 ++++++++ Eigen/src/Core/products/CoeffBasedProduct.h | 8 +- Eigen/src/Core/util/Macros.h | 41 ++++++++ Eigen/src/Core/util/XprHelper.h | 4 +- 27 files changed, 321 insertions(+), 120 deletions(-) diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index e67ca34cd..d1c422836 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -64,8 +64,10 @@ template class ArrayBase using Base::MaxSizeAtCompileTime; using Base::IsVectorAtCompileTime; using Base::Flags; +#ifndef EIGEN_TEST_EVALUATORS using Base::CoeffReadCost; - +#endif + using Base::derived; using Base::const_cast_derived; using Base::rows; diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index 95306a9a7..95eb37dd5 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -508,9 +508,6 @@ EIGEN_STRONG_INLINE Derived& DenseBase #ifdef EIGEN_TEST_EVALUATORS -#ifdef EIGEN_DEBUG_ASSIGN - internal::copy_using_evaluator_traits::debug(); -#endif eigen_assert(rows() == other.rows() && cols() == other.cols()); internal::call_dense_assignment_loop(derived(),other.derived()); diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 14effd1f2..2ea1cc126 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -24,37 +24,45 @@ namespace internal { // copy_using_evaluator_traits is based on assign_traits -template +template struct copy_using_evaluator_traits { + typedef typename DstEvaluator::XprType Dst; + typedef typename SrcEvaluator::XprType Src; + // TODO, we should get these flags from the evaluators + enum { + DstFlags = Dst::Flags, + SrcFlags = Src::Flags + }; + public: enum { - DstIsAligned = Derived::Flags & AlignedBit, - DstHasDirectAccess = Derived::Flags & DirectAccessBit, - SrcIsAligned = OtherDerived::Flags & AlignedBit, + DstIsAligned = DstFlags & AlignedBit, + DstHasDirectAccess = DstFlags & DirectAccessBit, + SrcIsAligned = SrcFlags & AlignedBit, JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned }; private: enum { - InnerSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::SizeAtCompileTime) - : int(Derived::Flags)&RowMajorBit ? int(Derived::ColsAtCompileTime) - : int(Derived::RowsAtCompileTime), - InnerMaxSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::MaxSizeAtCompileTime) - : int(Derived::Flags)&RowMajorBit ? int(Derived::MaxColsAtCompileTime) - : int(Derived::MaxRowsAtCompileTime), - MaxSizeAtCompileTime = Derived::SizeAtCompileTime, - PacketSize = packet_traits::size + InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime) + : int(DstFlags)&RowMajorBit ? int(Dst::ColsAtCompileTime) + : int(Dst::RowsAtCompileTime), + InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime) + : int(DstFlags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime) + : int(Dst::MaxRowsAtCompileTime), + MaxSizeAtCompileTime = Dst::SizeAtCompileTime, + PacketSize = packet_traits::size }; enum { - StorageOrdersAgree = (int(Derived::IsRowMajor) == int(OtherDerived::IsRowMajor)), + StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)), MightVectorize = StorageOrdersAgree - && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit) + && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit) && (functor_traits::PacketAccess), MayInnerVectorize = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0 && int(DstIsAligned) && int(SrcIsAligned), - MayLinearize = StorageOrdersAgree && (int(Derived::Flags) & int(OtherDerived::Flags) & LinearAccessBit), + MayLinearize = StorageOrdersAgree && (int(DstFlags) & int(SrcFlags) & LinearAccessBit), MayLinearVectorize = MightVectorize && MayLinearize && DstHasDirectAccess && (DstIsAligned || MaxSizeAtCompileTime == Dynamic), /* If the destination isn't aligned, we have to do runtime checks and we don't unroll, @@ -81,12 +89,12 @@ public: private: enum { UnrollingLimit = EIGEN_UNROLLING_LIMIT * (Vectorized ? int(PacketSize) : 1), - MayUnrollCompletely = int(Derived::SizeAtCompileTime) != Dynamic - && int(OtherDerived::CoeffReadCost) != Dynamic - && int(Derived::SizeAtCompileTime) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit), + MayUnrollCompletely = int(Dst::SizeAtCompileTime) != Dynamic + && int(SrcEvaluator::CoeffReadCost) != Dynamic + && int(Dst::SizeAtCompileTime) * int(SrcEvaluator::CoeffReadCost) <= int(UnrollingLimit), MayUnrollInner = int(InnerSize) != Dynamic - && int(OtherDerived::CoeffReadCost) != Dynamic - && int(InnerSize) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit) + && int(SrcEvaluator::CoeffReadCost) != Dynamic + && int(InnerSize) * int(SrcEvaluator::CoeffReadCost) <= int(UnrollingLimit) }; public: @@ -518,12 +526,16 @@ public: typedef SrcEvaluatorTypeT SrcEvaluatorType; typedef typename DstEvaluatorType::Scalar Scalar; typedef typename DstEvaluatorType::Index Index; - typedef copy_using_evaluator_traits AssignmentTraits; + typedef copy_using_evaluator_traits AssignmentTraits; generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) : m_dst(dst), m_src(src), m_functor(func), m_dstExpr(dstExpr) - {} + { + #ifdef EIGEN_DEBUG_ASSIGN + AssignmentTraits::debug(); + #endif + } Index size() const { return m_dstExpr.size(); } Index innerSize() const { return m_dstExpr.innerSize(); } @@ -612,10 +624,6 @@ protected: template void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) { -#ifdef EIGEN_DEBUG_ASSIGN - // TODO these traits should be computed from information provided by the evaluators - internal::copy_using_evaluator_traits::debug(); -#endif eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); typedef typename evaluator::type DstEvaluatorType; @@ -750,10 +758,6 @@ struct Assignment { eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); - #ifdef EIGEN_DEBUG_ASSIGN - internal::copy_using_evaluator_traits::debug(); - #endif - call_dense_assignment_loop(dst, src, func); } }; diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h index f0b68f470..192e1db53 100644 --- a/Eigen/src/Core/BooleanRedux.h +++ b/Eigen/src/Core/BooleanRedux.h @@ -94,14 +94,14 @@ struct any_unroller template inline bool DenseBase::all() const { - enum { - unroll = SizeAtCompileTime != Dynamic - && CoeffReadCost != Dynamic - && NumTraits::AddCost != Dynamic - && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT - }; #ifdef EIGEN_TEST_EVALUATORS typedef typename internal::evaluator::type Evaluator; + enum { + unroll = SizeAtCompileTime != Dynamic + && Evaluator::CoeffReadCost != Dynamic + && NumTraits::AddCost != Dynamic + && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT + }; Evaluator evaluator(derived()); if(unroll) return internal::all_unroller::run(evaluator); @@ -113,6 +113,13 @@ inline bool DenseBase::all() const return true; } #else + enum { + unroll = SizeAtCompileTime != Dynamic + && CoeffReadCost != Dynamic + && NumTraits::AddCost != Dynamic + && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT + }; + if(unroll) return internal::all_unroller::run(derived()); else @@ -132,14 +139,14 @@ inline bool DenseBase::all() const template inline bool DenseBase::any() const { - enum { - unroll = SizeAtCompileTime != Dynamic - && CoeffReadCost != Dynamic - && NumTraits::AddCost != Dynamic - && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT - }; #ifdef EIGEN_TEST_EVALUATORS typedef typename internal::evaluator::type Evaluator; + enum { + unroll = SizeAtCompileTime != Dynamic + && Evaluator::CoeffReadCost != Dynamic + && NumTraits::AddCost != Dynamic + && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT + }; Evaluator evaluator(derived()); if(unroll) return internal::any_unroller::run(evaluator); @@ -151,6 +158,12 @@ inline bool DenseBase::any() const return false; } #else + enum { + unroll = SizeAtCompileTime != Dynamic + && CoeffReadCost != Dynamic + && NumTraits::AddCost != Dynamic + && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT + }; if(unroll) return internal::any_unroller::run(derived()); else diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 92b50d190..33c89c2d4 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -71,7 +71,7 @@ struct evaluator_traits_base // typedef evaluator type; // typedef evaluator nestedType; - // by default, get evalautor kind and shape from storage + // by default, get evaluator kind and shape from storage typedef typename storage_kind_to_evaluator_kind::Kind Kind; typedef typename storage_kind_to_shape::Shape Shape; @@ -124,12 +124,19 @@ struct evaluator > : evaluator_base { typedef PlainObjectBase PlainObjectType; + typedef typename PlainObjectType::Index Index; + typedef typename PlainObjectType::Scalar Scalar; + typedef typename PlainObjectType::CoeffReturnType CoeffReturnType; + typedef typename PlainObjectType::PacketScalar PacketScalar; + typedef typename PlainObjectType::PacketReturnType PacketReturnType; enum { IsRowMajor = PlainObjectType::IsRowMajor, IsVectorAtCompileTime = PlainObjectType::IsVectorAtCompileTime, RowsAtCompileTime = PlainObjectType::RowsAtCompileTime, - ColsAtCompileTime = PlainObjectType::ColsAtCompileTime + ColsAtCompileTime = PlainObjectType::ColsAtCompileTime, + + CoeffReadCost = NumTraits::ReadCost }; evaluator() @@ -143,12 +150,6 @@ struct evaluator > : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) { } - typedef typename PlainObjectType::Index Index; - typedef typename PlainObjectType::Scalar Scalar; - typedef typename PlainObjectType::CoeffReturnType CoeffReturnType; - typedef typename PlainObjectType::PacketScalar PacketScalar; - typedef typename PlainObjectType::PacketReturnType PacketReturnType; - CoeffReturnType coeff(Index row, Index col) const { if (IsRowMajor) @@ -320,6 +321,10 @@ struct evaluator > : evaluator_base > { typedef Transpose XprType; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + }; evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} @@ -384,6 +389,10 @@ struct evaluator > : evaluator_base > { typedef CwiseNullaryOp XprType; + + enum { + CoeffReadCost = internal::functor_traits::Cost + }; evaluator(const XprType& n) : m_functor(n.functor()) @@ -426,6 +435,10 @@ struct evaluator > : evaluator_base > { typedef CwiseUnaryOp XprType; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost + }; evaluator(const XprType& op) : m_functor(op.functor()), @@ -470,6 +483,10 @@ struct evaluator > : evaluator_base > { typedef CwiseBinaryOp XprType; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + evaluator::CoeffReadCost + functor_traits::Cost + }; evaluator(const XprType& xpr) : m_functor(xpr.functor()), @@ -518,6 +535,10 @@ struct evaluator > : evaluator_base > { typedef CwiseUnaryView XprType; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost + }; evaluator(const XprType& op) : m_unaryOp(op.functor()), @@ -561,7 +582,6 @@ struct evaluator > { typedef MapBase MapType; typedef Derived XprType; - typedef typename XprType::PointerType PointerType; typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -569,16 +589,17 @@ struct evaluator > typedef typename XprType::PacketScalar PacketScalar; typedef typename XprType::PacketReturnType PacketReturnType; + enum { + RowsAtCompileTime = XprType::RowsAtCompileTime, + CoeffReadCost = NumTraits::ReadCost + }; + evaluator(const XprType& map) : m_data(const_cast(map.data())), m_rowStride(map.rowStride()), m_colStride(map.colStride()) { } - enum { - RowsAtCompileTime = XprType::RowsAtCompileTime - }; - CoeffReturnType coeff(Index row, Index col) const { return m_data[col * m_colStride + row * m_rowStride]; @@ -670,6 +691,9 @@ struct evaluator > : block_evaluator { typedef Block XprType; + enum { + CoeffReadCost = evaluator::CoeffReadCost + }; typedef block_evaluator block_evaluator_type; evaluator(const XprType& block) : block_evaluator_type(block) {} }; @@ -703,8 +727,7 @@ struct block_evaluator @@ -728,7 +750,7 @@ struct block_evaluator(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + RowsAtCompileTime == 1 ? index : 0); } template @@ -741,8 +763,8 @@ struct block_evaluator(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0, - x); + RowsAtCompileTime == 1 ? index : 0, + x); } protected: @@ -773,6 +795,11 @@ struct evaluator > : evaluator_base > { typedef Select XprType; + enum { + CoeffReadCost = evaluator::CoeffReadCost + + EIGEN_SIZE_MAX(evaluator::CoeffReadCost, + evaluator::CoeffReadCost) + }; evaluator(const XprType& select) : m_conditionImpl(select.conditionMatrix()), @@ -813,6 +840,18 @@ struct evaluator > : evaluator_base > { typedef Replicate XprType; + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + enum { + Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor + }; + typedef typename internal::nested_eval::type ArgTypeNested; + typedef typename internal::remove_all::type ArgTypeNestedCleaned; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + }; evaluator(const XprType& replicate) : m_arg(replicate.nestedExpression()), @@ -821,10 +860,6 @@ struct evaluator > m_cols(replicate.nestedExpression().cols()) { } - typedef typename XprType::Index Index; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; - CoeffReturnType coeff(Index row, Index col) const { // try to avoid using modulo; this is a pure optimization strategy @@ -852,13 +887,7 @@ struct evaluator > } protected: - enum { - Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor - }; - typedef typename internal::nested_eval::type ArgTypeNested; - typedef typename internal::remove_all::type ArgTypeNestedCleaned; - - const ArgTypeNested m_arg; // FIXME is it OK to store both the argument and its evaluator?? (we have the same situation in evalautor_product) + const ArgTypeNested m_arg; // FIXME is it OK to store both the argument and its evaluator?? (we have the same situation in evaluator_product) typename evaluator::nestedType m_argImpl; const variable_if_dynamic m_rows; const variable_if_dynamic m_cols; @@ -876,6 +905,15 @@ struct evaluator > : evaluator_base > { typedef PartialReduxExpr XprType; + typedef typename XprType::Scalar InputScalar; + enum { + TraversalSize = Direction==Vertical ? XprType::RowsAtCompileTime : XprType::ColsAtCompileTime + }; + typedef typename MemberOp::template Cost CostOpType; + enum { + CoeffReadCost = TraversalSize==Dynamic ? Dynamic + : TraversalSize * evaluator::CoeffReadCost + int(CostOpType::value) + }; evaluator(const XprType expr) : m_expr(expr) @@ -909,6 +947,9 @@ struct evaluator_wrapper_base : evaluator_base { typedef typename remove_all::type ArgType; + enum { + CoeffReadCost = evaluator::CoeffReadCost + }; evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} @@ -1015,7 +1056,9 @@ struct evaluator > OffsetCol = ReverseCol && IsRowMajor ? PacketSize : 1, ReversePacket = (Direction == BothDirections) || ((Direction == Vertical) && IsColMajor) - || ((Direction == Horizontal) && IsRowMajor) + || ((Direction == Horizontal) && IsRowMajor), + + CoeffReadCost = evaluator::CoeffReadCost }; typedef internal::reverse_packet_cond reverse_packet; @@ -1093,6 +1136,10 @@ struct evaluator > : evaluator_base > { typedef Diagonal XprType; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + }; evaluator(const XprType& diagonal) : m_argImpl(diagonal.nestedExpression()), diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index c78067a88..105e7fb11 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -65,8 +65,6 @@ struct traits > typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; enum { - LhsCoeffReadCost = _LhsNested::CoeffReadCost, - RhsCoeffReadCost = _RhsNested::CoeffReadCost, LhsFlags = _LhsNested::Flags, RhsFlags = _RhsNested::Flags, SameType = is_same::value, @@ -80,8 +78,13 @@ struct traits > ) ) ), - Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit), + Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit) +#ifndef EIGEN_TEST_EVALUATORS + , + LhsCoeffReadCost = _LhsNested::CoeffReadCost, + RhsCoeffReadCost = _RhsNested::CoeffReadCost, CoeffReadCost = LhsCoeffReadCost + RhsCoeffReadCost + functor_traits::Cost +#endif }; }; } // end namespace internal diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index 124383114..560e03f12 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -39,8 +39,11 @@ struct traits > : traits::ret ? LinearAccessBit : 0) | (functor_traits::PacketAccess ? PacketAccessBit : 0))) - | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit), + | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit) +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = functor_traits::Cost +#endif }; }; } diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index aa7df197f..25da52ab7 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -46,8 +46,11 @@ struct traits > enum { Flags = _XprTypeNested::Flags & ( HereditaryBits | LinearAccessBit | AlignedBit - | (functor_traits::PacketAccess ? PacketAccessBit : 0)), + | (functor_traits::PacketAccess ? PacketAccessBit : 0)) +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = _XprTypeNested::CoeffReadCost + functor_traits::Cost +#endif }; }; } diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index b2638d326..a0bd80fb9 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -38,7 +38,9 @@ struct traits > typedef typename remove_all::type _MatrixTypeNested; enum { Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)), +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = traits<_MatrixTypeNested>::CoeffReadCost + functor_traits::Cost, +#endif MatrixTypeInnerStride = inner_stride_at_compile_time::ret, // need to cast the sizeof's from size_t to int explicitly, otherwise: // "error: no integral type can represent all of the enumerator values diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 2fea574dd..f3c79aa31 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -169,10 +169,12 @@ template class DenseBase InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime) : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime), +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = internal::traits::CoeffReadCost, /**< This is a rough measure of how expensive it is to read one coefficient from * this expression. */ +#endif InnerStrideAtCompileTime = internal::inner_stride_at_compile_time::ret, OuterStrideAtCompileTime = internal::outer_stride_at_compile_time::ret diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index b160479ab..02ab04980 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -53,7 +53,9 @@ struct traits > MaxColsAtCompileTime = 1, MaskLvalueBit = is_lvalue::value ? LvalueBit : 0, Flags = (unsigned int)_MatrixTypeNested::Flags & (HereditaryBits | LinearAccessBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = _MatrixTypeNested::CoeffReadCost, +#endif MatrixTypeOuterStride = outer_stride_at_compile_time::ret, InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1, OuterStrideAtCompileTime = 0 diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index 8df636928..784e4b1ce 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -280,8 +280,11 @@ struct traits > ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, MaxRowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, MaxColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, - Flags = traits::Flags & LvalueBit, + Flags = traits::Flags & LvalueBit +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = traits<_DiagonalVectorType>::CoeffReadCost +#endif }; }; } diff --git a/Eigen/src/Core/DiagonalProduct.h b/Eigen/src/Core/DiagonalProduct.h index f5539df13..840b70dbb 100644 --- a/Eigen/src/Core/DiagonalProduct.h +++ b/Eigen/src/Core/DiagonalProduct.h @@ -35,8 +35,11 @@ struct traits > _Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))), _LinearAccessMask = (RowsAtCompileTime==1 || ColsAtCompileTime==1) ? LinearAccessBit : 0, - Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit,//(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit), + Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit //(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit), +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = NumTraits::MulCost + MatrixType::CoeffReadCost + DiagonalType::DiagonalVectorType::CoeffReadCost +#endif }; }; } diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index c2cedbf6a..f84db0dc2 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -115,7 +115,9 @@ struct traits > MaxRowsAtCompileTime = _MaxRows, MaxColsAtCompileTime = _MaxCols, Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret, +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = NumTraits::ReadCost, +#endif Options = _Options, InnerStrideAtCompileTime = 1, OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 72aa75da8..1f9f295c9 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -66,8 +66,10 @@ template class MatrixBase using Base::MaxSizeAtCompileTime; using Base::IsVectorAtCompileTime; using Base::Flags; +#ifndef EIGEN_TEST_EVALUATORS using Base::CoeffReadCost; - +#endif + using Base::derived; using Base::const_cast_derived; using Base::rows; diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index b16b86d8f..7ebf31696 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -109,6 +109,9 @@ struct product_evaluator, ProductTag, DenseSha : public evaluator::PlainObject>::type { typedef Product XprType; +// enum { +// CoeffReadCost = 0 // FIXME why is it needed? (this was already the case before the evaluators, see traits) +// }; typedef typename XprType::PlainObject PlainObject; typedef typename evaluator::type Base; @@ -366,6 +369,11 @@ struct product_evaluator, ProductTag, DenseShape, { typedef Product XprType; typedef CoeffBasedProduct CoeffBasedProductType; + typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketScalar PacketScalar; + typedef typename XprType::PacketReturnType PacketReturnType; product_evaluator(const XprType& xpr) : m_lhs(xpr.lhs()), @@ -376,24 +384,8 @@ struct product_evaluator, ProductTag, DenseShape, m_innerDim(xpr.lhs().cols()) { } - typedef typename XprType::Index Index; - typedef typename XprType::Scalar Scalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketScalar PacketScalar; - typedef typename XprType::PacketReturnType PacketReturnType; - // Everything below here is taken from CoeffBasedProduct.h - enum { - RowsAtCompileTime = traits::RowsAtCompileTime, - PacketSize = packet_traits::size, - InnerSize = traits::InnerSize, - CoeffReadCost = traits::CoeffReadCost, - Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT, - CanVectorizeInner = traits::CanVectorizeInner, - Flags = traits::Flags - }; - typedef typename internal::nested_eval::type LhsNested; typedef typename internal::nested_eval::type RhsNested; @@ -403,6 +395,22 @@ struct product_evaluator, ProductTag, DenseShape, typedef typename evaluator::type LhsEtorType; typedef typename evaluator::type RhsEtorType; + enum { + RowsAtCompileTime = traits::RowsAtCompileTime, + PacketSize = packet_traits::size, + InnerSize = traits::InnerSize, + + LhsCoeffReadCost = LhsEtorType::CoeffReadCost, + RhsCoeffReadCost = RhsEtorType::CoeffReadCost, + CoeffReadCost = (InnerSize == Dynamic || LhsCoeffReadCost==Dynamic || RhsCoeffReadCost==Dynamic || NumTraits::AddCost==Dynamic || NumTraits::MulCost==Dynamic) ? Dynamic + : InnerSize * (NumTraits::MulCost + LhsCoeffReadCost + RhsCoeffReadCost) + + (InnerSize - 1) * NumTraits::AddCost, + + Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT, + CanVectorizeInner = traits::CanVectorizeInner, + Flags = traits::Flags + }; + const CoeffReturnType coeff(Index row, Index col) const { // TODO check performance regression wrt to Eigen 3.2 which has special handling of this function @@ -689,6 +697,10 @@ struct diagonal_product_evaluator_base typedef typename scalar_product_traits::ReturnType Scalar; typedef typename internal::packet_traits::type PacketScalar; public: + enum { + CoeffReadCost = NumTraits::MulCost + evaluator::CoeffReadCost + evaluator::CoeffReadCost + }; + diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag) : m_diagImpl(diag), m_matImpl(mat) { @@ -739,7 +751,9 @@ struct product_evaluator, ProductTag, DiagonalSha typedef Product XprType; typedef typename XprType::PlainObject PlainObject; - enum { StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; + enum { + StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor + }; product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 1b6a4177a..41290323f 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -331,7 +331,7 @@ public: IsRowMajor = XprType::IsRowMajor, SizeAtCompileTime = XprType::SizeAtCompileTime, InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime, - CoeffReadCost = XprType::CoeffReadCost + CoeffReadCost = evaluator::CoeffReadCost }; Index rows() const { return m_xpr.rows(); } diff --git a/Eigen/src/Core/Replicate.h b/Eigen/src/Core/Replicate.h index dde86a834..1e640d8aa 100644 --- a/Eigen/src/Core/Replicate.h +++ b/Eigen/src/Core/Replicate.h @@ -53,8 +53,11 @@ struct traits > IsRowMajor = MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1 ? 1 : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0 : (MatrixType::Flags & RowMajorBit) ? 1 : 0, - Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0), + Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0) +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = _MatrixTypeNested::CoeffReadCost +#endif }; }; } diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h index e30ae3d28..495b44cc4 100644 --- a/Eigen/src/Core/Reverse.h +++ b/Eigen/src/Core/Reverse.h @@ -49,9 +49,11 @@ struct traits > LinearAccess = ( (Direction==BothDirections) && (int(_MatrixTypeNested::Flags)&PacketAccessBit) ) ? LinearAccessBit : 0, - Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess), - + Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess) +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = _MatrixTypeNested::CoeffReadCost +#endif }; }; diff --git a/Eigen/src/Core/Select.h b/Eigen/src/Core/Select.h index 87993bbb5..abcba2d15 100644 --- a/Eigen/src/Core/Select.h +++ b/Eigen/src/Core/Select.h @@ -43,10 +43,13 @@ struct traits > ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime, MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime, - Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits, + Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = traits::type>::CoeffReadCost + EIGEN_SIZE_MAX(traits::type>::CoeffReadCost, traits::type>::CoeffReadCost) +#endif }; }; } diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 976708a0f..98f9e888f 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -45,7 +45,9 @@ struct traits > : traits Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit), Flags1 = Flags0 | FlagsLvalueBit, Flags = Flags1 ^ RowMajorBit, +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = MatrixTypeNestedPlain::CoeffReadCost, +#endif InnerStrideAtCompileTime = inner_stride_at_compile_time::ret, OuterStrideAtCompileTime = outer_stride_at_compile_time::ret }; diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 96c81f933..064e0e10d 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -32,7 +32,9 @@ template class TriangularBase : public EigenBase enum { Mode = internal::traits::Mode, +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = internal::traits::CoeffReadCost, +#endif RowsAtCompileTime = internal::traits::RowsAtCompileTime, ColsAtCompileTime = internal::traits::ColsAtCompileTime, MaxRowsAtCompileTime = internal::traits::MaxRowsAtCompileTime, @@ -174,8 +176,11 @@ struct traits > : traits typedef typename MatrixType::PlainObject DenseMatrixType; enum { Mode = _Mode, - Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode, + Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost +#endif }; }; } @@ -1141,10 +1146,6 @@ public: template void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) { -#ifdef EIGEN_DEBUG_ASSIGN - // TODO these traits should be computed from information provided by the evaluators - internal::copy_using_evaluator_traits::debug(); -#endif eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); typedef typename evaluator::type DstEvaluatorType; @@ -1159,8 +1160,8 @@ void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& sr enum { unroll = DstXprType::SizeAtCompileTime != Dynamic - && internal::traits::CoeffReadCost != Dynamic - && DstXprType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT + && SrcEvaluatorType::CoeffReadCost != Dynamic + && DstXprType::SizeAtCompileTime * SrcEvaluatorType::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT }; triangular_assignment_loop::run(kernel); diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index f25ddca17..702d0006d 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -52,6 +52,7 @@ struct traits > Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0), TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime : MatrixType::ColsAtCompileTime }; +#ifndef EIGEN_TEST_EVALUATORS #if EIGEN_GNUC_AT_LEAST(3,4) typedef typename MemberOp::template Cost CostOpType; #else @@ -61,6 +62,7 @@ struct traits > CoeffReadCost = TraversalSize==Dynamic ? Dynamic : TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value) }; +#endif }; } diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h index 64867b7a2..fba0e2a76 100644 --- a/Eigen/src/Core/Visitor.h +++ b/Eigen/src/Core/Visitor.h @@ -53,6 +53,35 @@ struct visitor_impl } }; +#ifdef EIGEN_ENABLE_EVALUATORS +// evaluator adaptor +template +class visitor_evaluator +{ +public: + visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} + + typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + enum { + RowsAtCompileTime = XprType::RowsAtCompileTime, + CoeffReadCost = internal::evaluator::CoeffReadCost + }; + + Index rows() const { return m_xpr.rows(); } + Index cols() const { return m_xpr.cols(); } + Index size() const { return m_xpr.size(); } + + CoeffReturnType coeff(Index row, Index col) const + { return m_evaluator.coeff(row, col); } + +protected: + typename internal::evaluator::nestedType m_evaluator; + const XprType &m_xpr; +}; +#endif } // end namespace internal /** Applies the visitor \a visitor to the whole coefficients of the matrix or vector. @@ -76,6 +105,19 @@ template template void DenseBase::visit(Visitor& visitor) const { +#ifdef EIGEN_TEST_EVALUATORS + typedef typename internal::visitor_evaluator ThisEvaluator; + ThisEvaluator thisEval(derived()); + + enum { unroll = SizeAtCompileTime != Dynamic + && ThisEvaluator::CoeffReadCost != Dynamic + && (SizeAtCompileTime == 1 || internal::functor_traits::Cost != Dynamic) + && SizeAtCompileTime * ThisEvaluator::CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits::Cost + <= EIGEN_UNROLLING_LIMIT }; + return internal::visitor_impl::run(thisEval, visitor); +#else enum { unroll = SizeAtCompileTime != Dynamic && CoeffReadCost != Dynamic && (SizeAtCompileTime == 1 || internal::functor_traits::Cost != Dynamic) @@ -84,6 +126,7 @@ void DenseBase::visit(Visitor& visitor) const return internal::visitor_impl::run(derived(), visitor); +#endif } namespace internal { diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index 77f291ab9..94099b7d3 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -47,8 +47,6 @@ struct traits > typename traits<_RhsNested>::Index>::type Index; enum { - LhsCoeffReadCost = traits<_LhsNested>::CoeffReadCost, - RhsCoeffReadCost = traits<_RhsNested>::CoeffReadCost, LhsFlags = traits<_LhsNested>::Flags, RhsFlags = traits<_RhsNested>::Flags, @@ -89,11 +87,13 @@ struct traits > | (CanVectorizeRhs ? (RhsFlags & AlignedBit) : 0) // TODO enable vectorization for mixed types | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0), - +#ifndef EIGEN_TEST_EVALUATORS + LhsCoeffReadCost = traits<_LhsNested>::CoeffReadCost, + RhsCoeffReadCost = traits<_RhsNested>::CoeffReadCost, CoeffReadCost = (InnerSize == Dynamic || LhsCoeffReadCost==Dynamic || RhsCoeffReadCost==Dynamic || NumTraits::AddCost==Dynamic || NumTraits::MulCost==Dynamic) ? Dynamic : InnerSize * (NumTraits::MulCost + LhsCoeffReadCost + RhsCoeffReadCost) + (InnerSize - 1) * NumTraits::AddCost, - +#endif /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index debc04f3f..3a1dec129 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -345,6 +345,8 @@ * documentation in a single line. **/ +#ifndef EIGEN_TEST_EVALUATORS + #define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \ typedef typename Eigen::internal::traits::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex. */ \ typedef typename Eigen::NumTraits::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex, T were corresponding to RealScalar. */ \ @@ -381,7 +383,46 @@ using Base::derived; \ using Base::const_cast_derived; +#else +// TODO The EIGEN_DENSE_PUBLIC_INTERFACE should not exists anymore + +#define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \ + typedef typename Eigen::internal::traits::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex. */ \ + typedef typename Eigen::NumTraits::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex, T were corresponding to RealScalar. */ \ + typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \ + typedef typename Eigen::internal::nested::type Nested; \ + typedef typename Eigen::internal::traits::StorageKind StorageKind; \ + typedef typename Eigen::internal::traits::Index Index; \ + enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ + ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ + Flags = Eigen::internal::traits::Flags, \ + SizeAtCompileTime = Base::SizeAtCompileTime, \ + MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \ + IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; + + +#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \ + typedef typename Eigen::internal::traits::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex. */ \ + typedef typename Eigen::NumTraits::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex, T were corresponding to RealScalar. */ \ + typedef typename Base::PacketScalar PacketScalar; \ + typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \ + typedef typename Eigen::internal::nested::type Nested; \ + typedef typename Eigen::internal::traits::StorageKind StorageKind; \ + typedef typename Eigen::internal::traits::Index Index; \ + enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ + ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ + MaxRowsAtCompileTime = Eigen::internal::traits::MaxRowsAtCompileTime, \ + MaxColsAtCompileTime = Eigen::internal::traits::MaxColsAtCompileTime, \ + Flags = Eigen::internal::traits::Flags, \ + SizeAtCompileTime = Base::SizeAtCompileTime, \ + MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \ + IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \ + using Base::derived; \ + using Base::const_cast_derived; + +#endif // EIGEN_TEST_EVALUATORS + #define EIGEN_PLAIN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b) #define EIGEN_PLAIN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b) diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index a4baa921c..8931c5a2d 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -370,7 +370,9 @@ template::type> struc DynamicAsInteger = 10000, ScalarReadCost = NumTraits::Scalar>::ReadCost, ScalarReadCostAsInteger = ScalarReadCost == Dynamic ? int(DynamicAsInteger) : int(ScalarReadCost), - CoeffReadCost = traits::CoeffReadCost, + CoeffReadCost = evaluator::CoeffReadCost, // TODO What if an evaluator evaluate itself into a tempory? + // Then CoeffReadCost will be small but we still have to evaluate if n>1... + // The solution might be to ask the evaluator if it creates a temp. Perhaps we could even ask the number of temps? CoeffReadCostAsInteger = CoeffReadCost == Dynamic ? int(DynamicAsInteger) : int(CoeffReadCost), NAsInteger = n == Dynamic ? int(DynamicAsInteger) : n, CostEvalAsInteger = (NAsInteger+1) * ScalarReadCostAsInteger + CoeffReadCostAsInteger, From 2bf63c6b4a6fb9f69df184dc69f3427ffd81c4a3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 11 Mar 2014 11:42:07 +0100 Subject: [PATCH 0347/1103] Even ReturnByValue should not evaluate when assembling the expression --- Eigen/src/Core/ReturnByValue.h | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index 7834f6cbc..d63b96138 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -33,6 +33,7 @@ struct traits > }; }; +#ifndef EIGEN_TEST_EVALUATORS /* The ReturnByValue object doesn't even have a coeff() method. * So the only way that nesting it in an expression can work, is by evaluating it into a plain matrix. * So internal::nested always gives the plain return matrix type. @@ -44,6 +45,13 @@ struct nested, n, PlainObject> { typedef typename traits::ReturnType type; }; +#else +template +struct nested_eval, n, PlainObject> +{ + typedef typename traits::ReturnType type; +}; +#endif } // end namespace internal From 5806e7380026ab4ebd8f1aa4792b83cfc60f58a9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 11 Mar 2014 11:44:11 +0100 Subject: [PATCH 0348/1103] It is not clear what XprType::Nested should be, so let's use nested::type as much as possible --- Eigen/src/Core/Product.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 31730983f..cac90bc1f 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -62,8 +62,8 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, - typedef typename Lhs::Nested LhsNested; - typedef typename Rhs::Nested RhsNested; + typedef typename internal::nested::type LhsNested; + typedef typename internal::nested::type RhsNested; typedef typename internal::remove_all::type LhsNestedCleaned; typedef typename internal::remove_all::type RhsNestedCleaned; From ae405839652dc72935821bdaed163e7be04b3082 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 11 Mar 2014 11:47:14 +0100 Subject: [PATCH 0349/1103] Fix CoeffReadCost issues --- Eigen/src/Core/PermutationMatrix.h | 9 ++++++++- Eigen/src/LU/FullPivLU.h | 8 ++------ 2 files changed, 10 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 6a26af3a5..9add80c54 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -60,7 +60,9 @@ class PermutationBase : public EigenBase typedef typename Traits::IndicesType IndicesType; enum { Flags = Traits::Flags, +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = Traits::CoeffReadCost, +#endif RowsAtCompileTime = Traits::RowsAtCompileTime, ColsAtCompileTime = Traits::ColsAtCompileTime, MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime, @@ -473,8 +475,11 @@ struct traits > ColsAtCompileTime = _IndicesType::SizeAtCompileTime, MaxRowsAtCompileTime = IndicesType::MaxRowsAtCompileTime, MaxColsAtCompileTime = IndicesType::MaxColsAtCompileTime, - Flags = 0, + Flags = 0 +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = _IndicesType::CoeffReadCost +#endif }; }; } @@ -626,7 +631,9 @@ class Transpose > typedef typename Derived::DenseMatrixType DenseMatrixType; enum { Flags = Traits::Flags, +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = Traits::CoeffReadCost, +#endif RowsAtCompileTime = Traits::RowsAtCompileTime, ColsAtCompileTime = Traits::ColsAtCompileTime, MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime, diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index e0236fd9e..a06f5702c 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -16,11 +16,7 @@ namespace internal { template struct traits > : traits<_MatrixType> { - typedef traits<_MatrixType> BaseTraits; - enum { - Flags = BaseTraits::Flags & RowMajorBit, - CoeffReadCost = Dynamic - }; + enum { Flags = 0 }; }; } // end namespace internal @@ -755,7 +751,7 @@ void FullPivLU<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const namespace internal { -#ifdef EIGEN_TEST_EVALUATORS +#ifndef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> From 9be72cda2ab25650ce97eacd6dc2e994c988741b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 11 Mar 2014 11:47:32 +0100 Subject: [PATCH 0350/1103] Port QR module to Solve/Inverse --- Eigen/src/QR/ColPivHouseholderQR.h | 112 +++++++++++++++++------ Eigen/src/QR/FullPivHouseholderQR.h | 135 ++++++++++++++++++++-------- Eigen/src/QR/HouseholderQR.h | 65 ++++++++++---- 3 files changed, 229 insertions(+), 83 deletions(-) diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index 07126a9f4..1bf19d19e 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -13,6 +13,15 @@ namespace Eigen { +namespace internal { +template struct traits > + : traits<_MatrixType> +{ + enum { Flags = 0 }; +}; + +} // end namespace internal + /** \ingroup QR_Module * * \class ColPivHouseholderQR @@ -56,6 +65,7 @@ template class ColPivHouseholderQR typedef typename internal::plain_row_type::type RowVectorType; typedef typename internal::plain_row_type::type RealRowVectorType; typedef HouseholderSequence::type> HouseholderSequenceType; + typedef typename MatrixType::PlainObject PlainObject; private: @@ -137,6 +147,15 @@ template class ColPivHouseholderQR * Example: \include ColPivHouseholderQR_solve.cpp * Output: \verbinclude ColPivHouseholderQR_solve.out */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -144,9 +163,10 @@ template class ColPivHouseholderQR eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); return internal::solve_retval(*this, b.derived()); } +#endif - HouseholderSequenceType householderQ(void) const; - HouseholderSequenceType matrixQ(void) const + HouseholderSequenceType householderQ() const; + HouseholderSequenceType matrixQ() const { return householderQ(); } @@ -284,6 +304,13 @@ template class ColPivHouseholderQR * \note If this matrix is not invertible, the returned matrix has undefined coefficients. * Use isInvertible() to first determine whether this matrix is invertible. */ +#ifdef EIGEN_TEST_EVALUATORS + inline const Inverse inverse() const + { + eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); + return Inverse(*this); + } +#else inline const internal::solve_retval inverse() const @@ -292,6 +319,7 @@ template class ColPivHouseholderQR return internal::solve_retval (*this, MatrixType::Identity(m_qr.rows(), m_qr.cols())); } +#endif inline Index rows() const { return m_qr.rows(); } inline Index cols() const { return m_qr.cols(); } @@ -382,6 +410,12 @@ template class ColPivHouseholderQR eigen_assert(m_isInitialized && "Decomposition is not initialized."); return Success; } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif protected: MatrixType m_qr; @@ -514,8 +548,41 @@ ColPivHouseholderQR& ColPivHouseholderQR::compute(const return *this; } +#ifndef EIGEN_PARSED_BY_DOXYGEN +template +template +void ColPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const +{ + eigen_assert(rhs.rows() == rows()); + + const Index nonzero_pivots = nonzeroPivots(); + + if(nonzero_pivots == 0) + { + dst.setZero(); + return; + } + + typename RhsType::PlainObject c(rhs); + + // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T + c.applyOnTheLeft(householderSequence(m_qr, m_hCoeffs) + .setLength(nonzero_pivots) + .transpose() + ); + + m_qr.topLeftCorner(nonzero_pivots, nonzero_pivots) + .template triangularView() + .solveInPlace(c.topRows(nonzero_pivots)); + + for(Index i = 0; i < nonzero_pivots; ++i) dst.row(m_colsPermutation.indices().coeff(i)) = c.row(i); + for(Index i = nonzero_pivots; i < cols(); ++i) dst.row(m_colsPermutation.indices().coeff(i)).setZero(); +} +#endif + namespace internal { +#ifndef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -524,34 +591,23 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - eigen_assert(rhs().rows() == dec().rows()); - - const Index cols = dec().cols(), - nonzero_pivots = dec().nonzeroPivots(); - - if(nonzero_pivots == 0) - { - dst.setZero(); - return; - } - - typename Rhs::PlainObject c(rhs()); - - // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T - c.applyOnTheLeft(householderSequence(dec().matrixQR(), dec().hCoeffs()) - .setLength(dec().nonzeroPivots()) - .transpose() - ); - - dec().matrixR() - .topLeftCorner(nonzero_pivots, nonzero_pivots) - .template triangularView() - .solveInPlace(c.topRows(nonzero_pivots)); - - for(Index i = 0; i < nonzero_pivots; ++i) dst.row(dec().colsPermutation().indices().coeff(i)) = c.row(i); - for(Index i = nonzero_pivots; i < cols; ++i) dst.row(dec().colsPermutation().indices().coeff(i)).setZero(); + dec()._solve_impl(rhs(), dst); } }; +#endif + +#ifdef EIGEN_TEST_EVALUATORS +template +struct Assignment >, internal::assign_op, Dense2Dense, Scalar> +{ + typedef ColPivHouseholderQR QrType; + typedef Inverse SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); + } +}; +#endif } // end namespace internal diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 44eaa1b1a..8cdf14e4b 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -15,6 +15,12 @@ namespace Eigen { namespace internal { +template struct traits > + : traits<_MatrixType> +{ + enum { Flags = 0 }; +}; + template struct FullPivHouseholderQRMatrixQReturnType; template @@ -23,7 +29,7 @@ struct traits > typedef typename MatrixType::PlainObject ReturnType; }; -} +} // end namespace internal /** \ingroup QR_Module * @@ -69,6 +75,7 @@ template class FullPivHouseholderQR typedef PermutationMatrix PermutationType; typedef typename internal::plain_row_type::type RowVectorType; typedef typename internal::plain_col_type::type ColVectorType; + typedef typename MatrixType::PlainObject PlainObject; /** \brief Default Constructor. * @@ -144,6 +151,15 @@ template class FullPivHouseholderQR * Example: \include FullPivHouseholderQR_solve.cpp * Output: \verbinclude FullPivHouseholderQR_solve.out */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -151,6 +167,7 @@ template class FullPivHouseholderQR eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); return internal::solve_retval(*this, b.derived()); } +#endif /** \returns Expression object representing the matrix Q */ @@ -280,7 +297,15 @@ template class FullPivHouseholderQR * * \note If this matrix is not invertible, the returned matrix has undefined coefficients. * Use isInvertible() to first determine whether this matrix is invertible. - */ inline const + */ +#ifdef EIGEN_TEST_EVALUATORS + inline const Inverse inverse() const + { + eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); + return Inverse(*this); + } +#else + inline const internal::solve_retval inverse() const { @@ -288,6 +313,7 @@ template class FullPivHouseholderQR return internal::solve_retval (*this, MatrixType::Identity(m_qr.rows(), m_qr.cols())); } +#endif inline Index rows() const { return m_qr.rows(); } inline Index cols() const { return m_qr.cols(); } @@ -366,6 +392,12 @@ template class FullPivHouseholderQR * diagonal coefficient of U. */ RealScalar maxPivot() const { return m_maxpivot; } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif protected: MatrixType m_qr; @@ -485,8 +517,46 @@ FullPivHouseholderQR& FullPivHouseholderQR::compute(cons return *this; } -namespace internal { +#ifndef EIGEN_PARSED_BY_DOXYGEN +template +template +void FullPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const +{ + eigen_assert(rhs.rows() == rows()); + const Index l_rank = rank(); + // FIXME introduce nonzeroPivots() and use it here. and more generally, + // make the same improvements in this dec as in FullPivLU. + if(l_rank==0) + { + dst.setZero(); + return; + } + + typename RhsType::PlainObject c(rhs); + + Matrix temp(rhs.cols()); + for (Index k = 0; k < l_rank; ++k) + { + Index remainingSize = rows()-k; + c.row(k).swap(c.row(m_rows_transpositions.coeff(k))); + c.bottomRightCorner(remainingSize, rhs.cols()) + .applyHouseholderOnTheLeft(m_qr.col(k).tail(remainingSize-1), + m_hCoeffs.coeff(k), &temp.coeffRef(0)); + } + + m_qr.topLeftCorner(l_rank, l_rank) + .template triangularView() + .solveInPlace(c.topRows(l_rank)); + + for(Index i = 0; i < l_rank; ++i) dst.row(m_cols_permutation.indices().coeff(i)) = c.row(i); + for(Index i = l_rank; i < cols(); ++i) dst.row(m_cols_permutation.indices().coeff(i)).setZero(); +} +#endif + +namespace internal { + +#ifndef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -495,38 +565,23 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - const Index rows = dec().rows(), cols = dec().cols(); - eigen_assert(rhs().rows() == rows); - - // FIXME introduce nonzeroPivots() and use it here. and more generally, - // make the same improvements in this dec as in FullPivLU. - if(dec().rank()==0) - { - dst.setZero(); - return; - } - - typename Rhs::PlainObject c(rhs()); - - Matrix temp(rhs().cols()); - for (Index k = 0; k < dec().rank(); ++k) - { - Index remainingSize = rows-k; - c.row(k).swap(c.row(dec().rowsTranspositions().coeff(k))); - c.bottomRightCorner(remainingSize, rhs().cols()) - .applyHouseholderOnTheLeft(dec().matrixQR().col(k).tail(remainingSize-1), - dec().hCoeffs().coeff(k), &temp.coeffRef(0)); - } - - dec().matrixQR() - .topLeftCorner(dec().rank(), dec().rank()) - .template triangularView() - .solveInPlace(c.topRows(dec().rank())); - - for(Index i = 0; i < dec().rank(); ++i) dst.row(dec().colsPermutation().indices().coeff(i)) = c.row(i); - for(Index i = dec().rank(); i < cols; ++i) dst.row(dec().colsPermutation().indices().coeff(i)).setZero(); + dec()._solve_impl(rhs(), dst); } }; +#endif // EIGEN_TEST_EVALUATORS + +#ifdef EIGEN_TEST_EVALUATORS +template +struct Assignment >, internal::assign_op, Dense2Dense, Scalar> +{ + typedef FullPivHouseholderQR QrType; + typedef Inverse SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); + } +}; +#endif /** \ingroup QR_Module * @@ -534,6 +589,7 @@ struct solve_retval, Rhs> * * \tparam MatrixType type of underlying dense matrix */ +// #ifndef EIGEN_TEST_EVALUATORS template struct FullPivHouseholderQRMatrixQReturnType : public ReturnByValue > { @@ -550,7 +606,7 @@ public: : m_qr(qr), m_hCoeffs(hCoeffs), m_rowsTranspositions(rowsTranspositions) - {} + {} template void evalTo(ResultType& result) const @@ -580,8 +636,8 @@ public: } } - Index rows() const { return m_qr.rows(); } - Index cols() const { return m_qr.rows(); } + Index rows() const { return m_qr.rows(); } + Index cols() const { return m_qr.rows(); } protected: typename MatrixType::Nested m_qr; @@ -589,6 +645,13 @@ protected: typename IntDiagSizeVectorType::Nested m_rowsTranspositions; }; +// #ifdef EIGEN_TEST_EVALUATORS +// template +// struct evaluator > +// : public evaluator > > +// {}; +// #endif + } // end namespace internal template diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index 352dbf3f0..8808e6c0d 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -117,6 +117,15 @@ template class HouseholderQR * Example: \include HouseholderQR_solve.cpp * Output: \verbinclude HouseholderQR_solve.out */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "HouseholderQR is not initialized."); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -124,6 +133,7 @@ template class HouseholderQR eigen_assert(m_isInitialized && "HouseholderQR is not initialized."); return internal::solve_retval(*this, b.derived()); } +#endif /** This method returns an expression of the unitary matrix Q as a sequence of Householder transformations. * @@ -187,6 +197,12 @@ template class HouseholderQR * For advanced uses only. */ const HCoeffsType& hCoeffs() const { return m_hCoeffs; } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif protected: MatrixType m_qr; @@ -308,6 +324,35 @@ struct householder_qr_inplace_blocked } }; +} // end namespace internal + +#ifndef EIGEN_PARSED_BY_DOXYGEN +template +template +void HouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const +{ + const Index rank = (std::min)(rows(), cols()); + eigen_assert(rhs.rows() == rows()); + + typename RhsType::PlainObject c(rhs); + + // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T + c.applyOnTheLeft(householderSequence( + m_qr.leftCols(rank), + m_hCoeffs.head(rank)).transpose() + ); + + m_qr.topLeftCorner(rank, rank) + .template triangularView() + .solveInPlace(c.topRows(rank)); + + dst.topRows(rank) = c.topRows(rank); + dst.bottomRows(cols()-rank).setZero(); +} +#endif + +namespace internal { + template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -316,25 +361,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - const Index rows = dec().rows(), cols = dec().cols(); - const Index rank = (std::min)(rows, cols); - eigen_assert(rhs().rows() == rows); - - typename Rhs::PlainObject c(rhs()); - - // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T - c.applyOnTheLeft(householderSequence( - dec().matrixQR().leftCols(rank), - dec().hCoeffs().head(rank)).transpose() - ); - - dec().matrixQR() - .topLeftCorner(rank, rank) - .template triangularView() - .solveInPlace(c.topRows(rank)); - - dst.topRows(rank) = c.topRows(rank); - dst.bottomRows(cols-rank).setZero(); + dec()._solve_impl(rhs(), dst); } }; From bbc0ada12a3ae1531860205b7f1e21f991bb6273 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 11 Mar 2014 12:18:32 +0100 Subject: [PATCH 0351/1103] Avoid stupid "enumeral mismatch in conditional expression" warnings in GCC --- Eigen/src/Geometry/Transform.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 0a5012cfe..f40644011 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -530,9 +530,9 @@ public: inline Transform& operator=(const UniformScaling& t); inline Transform& operator*=(const UniformScaling& s) { return scale(s.factor()); } - inline Transform operator*(const UniformScaling& s) const + inline TransformTimeDiagonalReturnType operator*(const UniformScaling& s) const { - Transform res = *this; + TransformTimeDiagonalReturnType res = *this; res.scale(s.factor()); return res; } From 082f7ddc3745160c57d8a5a185a2a22e4d781b5f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 11 Mar 2014 13:33:44 +0100 Subject: [PATCH 0352/1103] Port Cholesky module to evaluators --- Eigen/src/Cholesky/LDLT.h | 86 +++++++++++++++++++++----------- Eigen/src/Cholesky/LLT.h | 32 +++++++++++- Eigen/src/Core/SelfAdjointView.h | 5 +- Eigen/src/Core/util/XprHelper.h | 2 +- 4 files changed, 93 insertions(+), 32 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index f34d26465..c5ae2c87e 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -181,6 +181,17 @@ template class LDLT * * \sa MatrixBase::ldlt() */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "LDLT is not initialized."); + eigen_assert(m_matrix.rows()==b.rows() + && "LDLT::solve(): invalid number of rows of the right hand side matrix b"); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -190,6 +201,7 @@ template class LDLT && "LDLT::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } +#endif #ifdef EIGEN2_SUPPORT template @@ -233,6 +245,12 @@ template class LDLT eigen_assert(m_isInitialized && "LDLT is not initialized."); return Success; } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif protected: @@ -492,7 +510,44 @@ LDLT& LDLT::rankUpdate(const MatrixBase +template +void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const +{ + eigen_assert(rhs.rows() == rows()); + // dst = P b + dst = m_transpositions * rhs; + + // dst = L^-1 (P b) + matrixL().solveInPlace(dst); + + // dst = D^-1 (L^-1 P b) + // more precisely, use pseudo-inverse of D (see bug 241) + using std::abs; + EIGEN_USING_STD_MATH(max); + const Diagonal vecD = vectorD(); + RealScalar tolerance = (max)( vecD.array().abs().maxCoeff() * NumTraits::epsilon(), + RealScalar(1) / NumTraits::highest()); // motivated by LAPACK's xGELSS + + for (Index i = 0; i < vecD.size(); ++i) + { + if(abs(vecD(i)) > tolerance) + dst.row(i) /= vecD(i); + else + dst.row(i).setZero(); + } + + // dst = L^-T (D^-1 L^-1 P b) + matrixU().solveInPlace(dst); + + // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b + dst = m_transpositions.transpose() * dst; +} +#endif + namespace internal { +#ifndef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -502,37 +557,10 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - eigen_assert(rhs().rows() == dec().matrixLDLT().rows()); - // dst = P b - dst = dec().transpositionsP() * rhs(); - - // dst = L^-1 (P b) - dec().matrixL().solveInPlace(dst); - - // dst = D^-1 (L^-1 P b) - // more precisely, use pseudo-inverse of D (see bug 241) - using std::abs; - EIGEN_USING_STD_MATH(max); - typedef typename LDLTType::MatrixType MatrixType; - typedef typename LDLTType::Scalar Scalar; - typedef typename LDLTType::RealScalar RealScalar; - const Diagonal vectorD = dec().vectorD(); - RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() * NumTraits::epsilon(), - RealScalar(1) / NumTraits::highest()); // motivated by LAPACK's xGELSS - for (Index i = 0; i < vectorD.size(); ++i) { - if(abs(vectorD(i)) > tolerance) - dst.row(i) /= vectorD(i); - else - dst.row(i).setZero(); - } - - // dst = L^-T (D^-1 L^-1 P b) - dec().matrixU().solveInPlace(dst); - - // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b - dst = dec().transpositionsP().transpose() * dst; + dec()._solve_impl(rhs(),dst); } }; +#endif } /** \internal use x = ldlt_object.solve(x); diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index 2201c641e..d9a8ef1fb 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -117,6 +117,17 @@ template class LLT * * \sa solveInPlace(), MatrixBase::llt() */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "LLT is not initialized."); + eigen_assert(m_matrix.rows()==b.rows() + && "LLT::solve(): invalid number of rows of the right hand side matrix b"); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -126,6 +137,7 @@ template class LLT && "LLT::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } +#endif #ifdef EIGEN2_SUPPORT template @@ -172,6 +184,12 @@ template class LLT template LLT rankUpdate(const VectorType& vec, const RealScalar& sigma = 1); + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif protected: /** \internal @@ -415,8 +433,19 @@ LLT<_MatrixType,_UpLo> LLT<_MatrixType,_UpLo>::rankUpdate(const VectorType& v, c return *this; } - + +#ifndef EIGEN_PARSED_BY_DOXYGEN +template +template +void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const +{ + dst = rhs; + solveInPlace(dst); +} +#endif + namespace internal { +#ifndef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -430,6 +459,7 @@ struct solve_retval, Rhs> dec().solveInPlace(dst); } }; +#endif } /** \internal use x = llt_object.solve(x); diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index f7f512cf4..b300c6a48 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -39,8 +39,11 @@ struct traits > : traits enum { Mode = UpLo | SelfAdjoint, Flags = MatrixTypeNestedCleaned::Flags & (HereditaryBits) - & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)), // FIXME these flags should be preserved + & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost +#endif }; }; } diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 8931c5a2d..bcd6183e2 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -348,7 +348,7 @@ template::type> str // When using evaluators, we never evaluate when assembling the expression!! // TODO: get rid of this nested class since it's just an alias for ref_selector. -template::type> struct nested +template struct nested { typedef typename ref_selector::type type; }; From 7eefdb948c1ff372f85991ff3f9d998e66a554d9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 11 Mar 2014 13:43:46 +0100 Subject: [PATCH 0353/1103] Migrate JacobiSVD to Solver --- Eigen/src/LU/FullPivLU.h | 4 ++-- Eigen/src/SVD/JacobiSVD.h | 50 ++++++++++++++++++++++++++++++--------- 2 files changed, 41 insertions(+), 13 deletions(-) diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index a06f5702c..78cd9b90f 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -702,8 +702,8 @@ struct image_retval > #ifndef EIGEN_PARSED_BY_DOXYGEN template template -void FullPivLU<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const { - +void FullPivLU<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const +{ /* The decomposition PAQ = LU can be rewritten as A = P^{-1} L U Q^{-1}. * So we proceed as follows: * Step 1: compute c = P * rhs. diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index eee31ca97..d78583ecc 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -653,6 +653,16 @@ template class JacobiSVD * \note SVD solving is implicitly least-squares. Thus, this method serves both purposes of exact solving and least-squares solving. * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$. */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + eigen_assert(computeU() && computeV() && "JacobiSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); + return Solve(*this, b.derived()); + } +#else template inline const internal::solve_retval solve(const MatrixBase& b) const @@ -661,6 +671,7 @@ template class JacobiSVD eigen_assert(computeU() && computeV() && "JacobiSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); return internal::solve_retval(*this, b.derived()); } +#endif /** \returns the number of singular values that are not exactly 0 */ Index nonzeroSingularValues() const @@ -734,6 +745,12 @@ template class JacobiSVD inline Index rows() const { return m_rows; } inline Index cols() const { return m_cols; } + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif private: void allocate(Index rows, Index cols, unsigned int computationOptions); @@ -912,7 +929,27 @@ JacobiSVD::compute(const MatrixType& matrix, unsig return *this; } +#ifndef EIGEN_PARSED_BY_DOXYGEN +template +template +void JacobiSVD<_MatrixType,QRPreconditioner>::_solve_impl(const RhsType &rhs, DstType &dst) const +{ + eigen_assert(rhs.rows() == rows()); + + // A = U S V^* + // So A^{-1} = V S^{-1} U^* + + Matrix tmp; + Index l_rank = rank(); + + tmp.noalias() = m_matrixU.leftCols(l_rank).adjoint() * rhs; + tmp = m_singularValues.head(l_rank).asDiagonal().inverse() * tmp; + dst = m_matrixV.leftCols(l_rank) * tmp; +} +#endif + namespace internal { +#ifndef EIGEN_TEST_EVALUATORS template struct solve_retval, Rhs> : solve_retval_base, Rhs> @@ -922,19 +959,10 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - eigen_assert(rhs().rows() == dec().rows()); - - // A = U S V^* - // So A^{-1} = V S^{-1} U^* - - Matrix tmp; - Index rank = dec().rank(); - - tmp.noalias() = dec().matrixU().leftCols(rank).adjoint() * rhs(); - tmp = dec().singularValues().head(rank).asDiagonal().inverse() * tmp; - dst = dec().matrixV().leftCols(rank) * tmp; + dec()._solve_impl(rhs(), dst); } }; +#endif } // end namespace internal #ifndef __CUDACC__ From 8dd3b716e39d4b4b472b948de1af20838bf17493 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 13:34:11 +0100 Subject: [PATCH 0354/1103] Move evaluation related flags from traits to evaluator and fix evaluators of MapBase and Replicate --- Eigen/src/Core/AssignEvaluator.h | 15 +- Eigen/src/Core/Block.h | 11 + Eigen/src/Core/CoreEvaluators.h | 231 +++++++++++++----- Eigen/src/Core/CwiseBinaryOp.h | 8 +- Eigen/src/Core/CwiseNullaryOp.h | 7 +- Eigen/src/Core/CwiseUnaryOp.h | 7 +- Eigen/src/Core/CwiseUnaryView.h | 4 +- Eigen/src/Core/Diagonal.h | 5 +- Eigen/src/Core/DiagonalMatrix.h | 1 + Eigen/src/Core/DiagonalProduct.h | 8 +- Eigen/src/Core/Map.h | 7 +- Eigen/src/Core/MapBase.h | 9 +- Eigen/src/Core/Product.h | 27 +- Eigen/src/Core/ProductEvaluators.h | 111 +++++++-- Eigen/src/Core/Redux.h | 11 + Eigen/src/Core/Replicate.h | 7 +- Eigen/src/Core/Reverse.h | 7 +- Eigen/src/Core/Select.h | 5 +- Eigen/src/Core/Transpose.h | 7 +- Eigen/src/Core/VectorwiseOp.h | 4 + .../Core/products/TriangularMatrixVector.h | 2 +- Eigen/src/Core/util/ForwardDeclarations.h | 12 + Eigen/src/Core/util/XprHelper.h | 52 ++++ 23 files changed, 433 insertions(+), 125 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 2ea1cc126..05816094c 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -28,11 +28,10 @@ template struct copy_using_evaluator_traits { typedef typename DstEvaluator::XprType Dst; - typedef typename SrcEvaluator::XprType Src; - // TODO, we should get these flags from the evaluators + enum { - DstFlags = Dst::Flags, - SrcFlags = Src::Flags + DstFlags = DstEvaluator::Flags, + SrcFlags = SrcEvaluator::Flags }; public: @@ -56,7 +55,9 @@ private: }; enum { - StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)), + DstIsRowMajor = DstEvaluator::Flags&RowMajorBit, + SrcIsRowMajor = SrcEvaluator::Flags&RowMajorBit, + StorageOrdersAgree = (int(DstIsRowMajor) == int(SrcIsRowMajor)), MightVectorize = StorageOrdersAgree && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit) && (functor_traits::PacketAccess), @@ -596,7 +597,7 @@ public: typedef typename DstEvaluatorType::ExpressionTraits Traits; return int(Traits::RowsAtCompileTime) == 1 ? 0 : int(Traits::ColsAtCompileTime) == 1 ? inner - : int(Traits::Flags)&RowMajorBit ? outer + : int(DstEvaluatorType::Flags)&RowMajorBit ? outer : inner; } @@ -605,7 +606,7 @@ public: typedef typename DstEvaluatorType::ExpressionTraits Traits; return int(Traits::ColsAtCompileTime) == 1 ? 0 : int(Traits::RowsAtCompileTime) == 1 ? inner - : int(Traits::Flags)&RowMajorBit ? inner + : int(DstEvaluatorType::Flags)&RowMajorBit ? inner : outer; } diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 31cd5c72c..d92797a98 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -68,6 +68,7 @@ struct traits > : traits::MaxColsAtCompileTime), + XprTypeIsRowMajor = (int(traits::Flags)&RowMajorBit) != 0, IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1 : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0 @@ -80,6 +81,10 @@ struct traits > : traits::ret) : int(inner_stride_at_compile_time::ret), + // IsAligned is needed by MapBase's assertions + // We can sefely set it to false here. Internal alignment errors will be detected by an eigen_internal_assert in the respective evaluator + IsAligned = 0, +#ifndef EIGEN_TEST_EVALUATORS MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits::size) == 0) && (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0, @@ -92,6 +97,12 @@ struct traits > : traits::value ? LvalueBit : 0, + FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, + Flags = (traits::Flags & DirectAccessBit) | FlagsLvalueBit | FlagsRowMajorBit // FIXME DirectAccessBit should not be handled by expressions +#endif }; }; diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 33c89c2d4..a5de3593c 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -136,7 +136,9 @@ struct evaluator > RowsAtCompileTime = PlainObjectType::RowsAtCompileTime, ColsAtCompileTime = PlainObjectType::ColsAtCompileTime, - CoeffReadCost = NumTraits::ReadCost + CoeffReadCost = NumTraits::ReadCost, + Flags = compute_matrix_evaluator_flags< Scalar,Derived::RowsAtCompileTime,Derived::ColsAtCompileTime, + Derived::Options,Derived::MaxRowsAtCompileTime,Derived::MaxColsAtCompileTime>::ret }; evaluator() @@ -323,7 +325,8 @@ struct evaluator > typedef Transpose XprType; enum { - CoeffReadCost = evaluator::CoeffReadCost + CoeffReadCost = evaluator::CoeffReadCost, + Flags = evaluator::Flags ^ RowMajorBit }; evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} @@ -389,9 +392,16 @@ struct evaluator > : evaluator_base > { typedef CwiseNullaryOp XprType; + typedef typename internal::remove_all::type PlainObjectTypeCleaned; enum { - CoeffReadCost = internal::functor_traits::Cost + CoeffReadCost = internal::functor_traits::Cost, + + Flags = (evaluator::Flags + & ( HereditaryBits + | (functor_has_linear_access::ret ? LinearAccessBit : 0) + | (functor_traits::PacketAccess ? PacketAccessBit : 0))) + | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit) // FIXME EvalBeforeNestingBit should be needed anymore }; evaluator(const XprType& n) @@ -437,7 +447,11 @@ struct evaluator > typedef CwiseUnaryOp XprType; enum { - CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost + CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost, + + Flags = evaluator::Flags & ( + HereditaryBits | LinearAccessBit | AlignedBit + | (functor_traits::PacketAccess ? PacketAccessBit : 0)) }; evaluator(const XprType& op) @@ -485,7 +499,22 @@ struct evaluator > typedef CwiseBinaryOp XprType; enum { - CoeffReadCost = evaluator::CoeffReadCost + evaluator::CoeffReadCost + functor_traits::Cost + CoeffReadCost = evaluator::CoeffReadCost + evaluator::CoeffReadCost + functor_traits::Cost, + + LhsFlags = evaluator::Flags, + RhsFlags = evaluator::Flags, + SameType = is_same::value, + StorageOrdersAgree = (int(LhsFlags)&RowMajorBit)==(int(RhsFlags)&RowMajorBit), + Flags0 = (int(LhsFlags) | int(RhsFlags)) & ( + HereditaryBits + | (int(LhsFlags) & int(RhsFlags) & + ( AlignedBit + | (StorageOrdersAgree ? LinearAccessBit : 0) + | (functor_traits::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0) + ) + ) + ), + Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit) }; evaluator(const XprType& xpr) @@ -537,7 +566,9 @@ struct evaluator > typedef CwiseUnaryView XprType; enum { - CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost + CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost, + + Flags = (evaluator::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit)) }; evaluator(const XprType& op) @@ -576,12 +607,15 @@ protected: // -------------------- Map -------------------- -template -struct evaluator > - : evaluator_base +// FIXME perhaps the PlainObjectType could be provided by Derived::PlainObject ? +// but that might complicate template specialization +template +struct mapbase_evaluator; + +template +struct mapbase_evaluator : evaluator_base { - typedef MapBase MapType; - typedef Derived XprType; + typedef Derived XprType; typedef typename XprType::PointerType PointerType; typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -590,81 +624,103 @@ struct evaluator > typedef typename XprType::PacketReturnType PacketReturnType; enum { - RowsAtCompileTime = XprType::RowsAtCompileTime, + IsRowMajor = XprType::RowsAtCompileTime, + ColsAtCompileTime = XprType::ColsAtCompileTime, CoeffReadCost = NumTraits::ReadCost }; - evaluator(const XprType& map) + mapbase_evaluator(const XprType& map) : m_data(const_cast(map.data())), - m_rowStride(map.rowStride()), - m_colStride(map.colStride()) - { } + m_xpr(map) + { + EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(evaluator::Flags&PacketAccessBit, internal::inner_stride_at_compile_time::ret==1), + PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1); + } CoeffReturnType coeff(Index row, Index col) const - { - return m_data[col * m_colStride + row * m_rowStride]; + { + return m_data[col * m_xpr.colStride() + row * m_xpr.rowStride()]; } CoeffReturnType coeff(Index index) const - { - return coeff(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + { + return m_data[index * m_xpr.innerStride()]; } Scalar& coeffRef(Index row, Index col) - { - return m_data[col * m_colStride + row * m_rowStride]; + { + return m_data[col * m_xpr.colStride() + row * m_xpr.rowStride()]; } Scalar& coeffRef(Index index) - { - return coeffRef(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + { + return m_data[index * m_xpr.innerStride()]; } template PacketReturnType packet(Index row, Index col) const - { - PointerType ptr = m_data + row * m_rowStride + col * m_colStride; + { + PointerType ptr = m_data + row * m_xpr.rowStride() + col * m_xpr.colStride(); return internal::ploadt(ptr); } template PacketReturnType packet(Index index) const - { - return packet(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0); + { + return internal::ploadt(m_data + index * m_xpr.innerStride()); } template void writePacket(Index row, Index col, const PacketScalar& x) - { - PointerType ptr = m_data + row * m_rowStride + col * m_colStride; + { + PointerType ptr = m_data + row * m_xpr.rowStride() + col * m_xpr.colStride(); return internal::pstoret(ptr, x); } template void writePacket(Index index, const PacketScalar& x) - { - return writePacket(RowsAtCompileTime == 1 ? 0 : index, - RowsAtCompileTime == 1 ? index : 0, - x); + { + internal::pstoret(m_data + index * m_xpr.innerStride(), x); } protected: PointerType m_data; - int m_rowStride; - int m_colStride; + const XprType& m_xpr; }; template struct evaluator > - : public evaluator > > + : public mapbase_evaluator, PlainObjectType> { typedef Map XprType; + typedef typename XprType::Scalar Scalar; + + enum { + InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0 + ? int(PlainObjectType::InnerStrideAtCompileTime) + : int(StrideType::InnerStrideAtCompileTime), + OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0 + ? int(PlainObjectType::OuterStrideAtCompileTime) + : int(StrideType::OuterStrideAtCompileTime), + HasNoInnerStride = InnerStrideAtCompileTime == 1, + HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0, + HasNoStride = HasNoInnerStride && HasNoOuterStride, + IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned), + IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic, + KeepsPacketAccess = bool(HasNoInnerStride) + && ( bool(IsDynamicSize) + || HasNoOuterStride + || ( OuterStrideAtCompileTime!=Dynamic + && ((static_cast(sizeof(Scalar))*OuterStrideAtCompileTime)%16)==0 ) ), + Flags0 = evaluator::Flags, + Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit), + Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime)) + ? int(Flags1) : int(Flags1 & ~LinearAccessBit), + Flags = KeepsPacketAccess ? int(Flags2) : (int(Flags2) & ~PacketAccessBit) + }; evaluator(const XprType& map) - : evaluator >(map) + : mapbase_evaluator(map) { } }; @@ -672,12 +728,16 @@ struct evaluator > template struct evaluator > - : public evaluator > > + : public mapbase_evaluator, PlainObjectType> { typedef Ref XprType; + + enum { + Flags = evaluator >::Flags + }; - evaluator(const XprType& map) - : evaluator >(map) + evaluator(const XprType& ref) + : mapbase_evaluator(ref) { } }; @@ -691,8 +751,39 @@ struct evaluator > : block_evaluator { typedef Block XprType; + typedef typename XprType::Scalar Scalar; + enum { - CoeffReadCost = evaluator::CoeffReadCost + CoeffReadCost = evaluator::CoeffReadCost, + + RowsAtCompileTime = traits::RowsAtCompileTime, + ColsAtCompileTime = traits::ColsAtCompileTime, + MaxRowsAtCompileTime = traits::MaxRowsAtCompileTime, + MaxColsAtCompileTime = traits::MaxColsAtCompileTime, + + XprTypeIsRowMajor = (int(traits::Flags)&RowMajorBit) != 0, + IsRowMajor = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? 1 + : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0 + : XprTypeIsRowMajor, + HasSameStorageOrderAsXprType = (IsRowMajor == XprTypeIsRowMajor), + InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime), + InnerStrideAtCompileTime = HasSameStorageOrderAsXprType + ? int(inner_stride_at_compile_time::ret) + : int(outer_stride_at_compile_time::ret), + OuterStrideAtCompileTime = HasSameStorageOrderAsXprType + ? int(outer_stride_at_compile_time::ret) + : int(inner_stride_at_compile_time::ret), + MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits::size) == 0) + && (InnerStrideAtCompileTime == 1) + ? PacketAccessBit : 0, + MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % 16) == 0)) ? AlignedBit : 0, + FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1) ? LinearAccessBit : 0, + FlagsRowMajorBit = XprType::Flags&RowMajorBit, + Flags0 = traits::Flags & ( (HereditaryBits & ~RowMajorBit) | + DirectAccessBit | + MaskPacketAccessBit | + MaskAlignedBit), + Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit }; typedef block_evaluator block_evaluator_type; evaluator(const XprType& block) : block_evaluator_type(block) {} @@ -778,18 +869,23 @@ protected: template struct block_evaluator - : evaluator > > + : mapbase_evaluator, + typename Block::PlainObject> { typedef Block XprType; block_evaluator(const XprType& block) - : evaluator >(block) - { } + : mapbase_evaluator(block) + { + // FIXME this should be an internal assertion + eigen_assert(EIGEN_IMPLIES(evaluator::Flags&AlignedBit, (size_t(block.data()) % 16) == 0) && "data is not aligned"); + } }; // -------------------- Select -------------------- +// TODO enable vectorization for Select template struct evaluator > : evaluator_base > @@ -798,7 +894,9 @@ struct evaluator > enum { CoeffReadCost = evaluator::CoeffReadCost + EIGEN_SIZE_MAX(evaluator::CoeffReadCost, - evaluator::CoeffReadCost) + evaluator::CoeffReadCost), + + Flags = (unsigned int)evaluator::Flags & evaluator::Flags & HereditaryBits }; evaluator(const XprType& select) @@ -850,7 +948,9 @@ struct evaluator > typedef typename internal::remove_all::type ArgTypeNestedCleaned; enum { - CoeffReadCost = evaluator::CoeffReadCost + CoeffReadCost = evaluator::CoeffReadCost, + + Flags = (evaluator::Flags & HereditaryBits & ~RowMajorBit) | (traits::Flags & RowMajorBit) }; evaluator(const XprType& replicate) @@ -858,7 +958,7 @@ struct evaluator > m_argImpl(m_arg), m_rows(replicate.nestedExpression().rows()), m_cols(replicate.nestedExpression().cols()) - { } + {} CoeffReturnType coeff(Index row, Index col) const { @@ -907,17 +1007,19 @@ struct evaluator > typedef PartialReduxExpr XprType; typedef typename XprType::Scalar InputScalar; enum { - TraversalSize = Direction==Vertical ? XprType::RowsAtCompileTime : XprType::ColsAtCompileTime + TraversalSize = Direction==Vertical ? ArgType::RowsAtCompileTime : XprType::ColsAtCompileTime }; typedef typename MemberOp::template Cost CostOpType; enum { CoeffReadCost = TraversalSize==Dynamic ? Dynamic - : TraversalSize * evaluator::CoeffReadCost + int(CostOpType::value) + : TraversalSize * evaluator::CoeffReadCost + int(CostOpType::value), + + Flags = (traits::Flags&RowMajorBit) | (evaluator::Flags&HereditaryBits) }; evaluator(const XprType expr) : m_expr(expr) - { } + {} typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; @@ -948,7 +1050,8 @@ struct evaluator_wrapper_base { typedef typename remove_all::type ArgType; enum { - CoeffReadCost = evaluator::CoeffReadCost + CoeffReadCost = evaluator::CoeffReadCost, + Flags = evaluator::Flags }; evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} @@ -1058,7 +1161,15 @@ struct evaluator > || ((Direction == Vertical) && IsColMajor) || ((Direction == Horizontal) && IsRowMajor), - CoeffReadCost = evaluator::CoeffReadCost + CoeffReadCost = evaluator::CoeffReadCost, + + // let's enable LinearAccess only with vectorization because of the product overhead + // FIXME enable DirectAccess with negative strides? + Flags0 = evaluator::Flags, + LinearAccess = ( (Direction==BothDirections) && (int(Flags0)&PacketAccessBit) ) + ? LinearAccessBit : 0, + + Flags = int(Flags0) & (HereditaryBits | PacketAccessBit | LinearAccess) }; typedef internal::reverse_packet_cond reverse_packet; @@ -1071,7 +1182,7 @@ struct evaluator > CoeffReturnType coeff(Index row, Index col) const { return m_argImpl.coeff(ReverseRow ? m_rows.value() - row - 1 : row, - ReverseCol ? m_cols.value() - col - 1 : col); + ReverseCol ? m_cols.value() - col - 1 : col); } CoeffReturnType coeff(Index index) const @@ -1082,7 +1193,7 @@ struct evaluator > Scalar& coeffRef(Index row, Index col) { return m_argImpl.coeffRef(ReverseRow ? m_rows.value() - row - 1 : row, - ReverseCol ? m_cols.value() - col - 1 : col); + ReverseCol ? m_cols.value() - col - 1 : col); } Scalar& coeffRef(Index index) @@ -1138,7 +1249,9 @@ struct evaluator > typedef Diagonal XprType; enum { - CoeffReadCost = evaluator::CoeffReadCost + CoeffReadCost = evaluator::CoeffReadCost, + + Flags = (unsigned int)evaluator::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit) & ~RowMajorBit }; evaluator(const XprType& diagonal) diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index 105e7fb11..07861dbc9 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -65,6 +65,7 @@ struct traits > typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; enum { +#ifndef EIGEN_TEST_EVALUATORS LhsFlags = _LhsNested::Flags, RhsFlags = _RhsNested::Flags, SameType = is_same::value, @@ -78,12 +79,13 @@ struct traits > ) ) ), - Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit) -#ifndef EIGEN_TEST_EVALUATORS - , + Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit), + LhsCoeffReadCost = _LhsNested::CoeffReadCost, RhsCoeffReadCost = _RhsNested::CoeffReadCost, CoeffReadCost = LhsCoeffReadCost + RhsCoeffReadCost + functor_traits::Cost +#else + Flags = _LhsNested::Flags & RowMajorBit #endif }; }; diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index 560e03f12..f9f127cc2 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -35,14 +35,15 @@ template struct traits > : traits { enum { +#ifndef EIGEN_TEST_EVALUATORS Flags = (traits::Flags & ( HereditaryBits | (functor_has_linear_access::ret ? LinearAccessBit : 0) | (functor_traits::PacketAccess ? PacketAccessBit : 0))) - | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit) -#ifndef EIGEN_TEST_EVALUATORS - , + | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit), CoeffReadCost = functor_traits::Cost +#else + Flags = traits::Flags & RowMajorBit #endif }; }; diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index 25da52ab7..af05a9108 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -44,12 +44,13 @@ struct traits > typedef typename XprType::Nested XprTypeNested; typedef typename remove_reference::type _XprTypeNested; enum { +#ifndef EIGEN_TEST_EVALUATORS Flags = _XprTypeNested::Flags & ( HereditaryBits | LinearAccessBit | AlignedBit - | (functor_traits::PacketAccess ? PacketAccessBit : 0)) -#ifndef EIGEN_TEST_EVALUATORS - , + | (functor_traits::PacketAccess ? PacketAccessBit : 0)), CoeffReadCost = _XprTypeNested::CoeffReadCost + functor_traits::Cost +#else + Flags = _XprTypeNested::Flags & RowMajorBit #endif }; }; diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index a0bd80fb9..9cdebb8e7 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -37,9 +37,11 @@ struct traits > typedef typename MatrixType::Nested MatrixTypeNested; typedef typename remove_all::type _MatrixTypeNested; enum { - Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)), #ifndef EIGEN_TEST_EVALUATORS + Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)), CoeffReadCost = traits<_MatrixTypeNested>::CoeffReadCost + functor_traits::Cost, +#else + Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | LvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions #endif MatrixTypeInnerStride = inner_stride_at_compile_time::ret, // need to cast the sizeof's from size_t to int explicitly, otherwise: diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index 02ab04980..3ff6a3e66 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -51,10 +51,13 @@ struct traits > : (EIGEN_PLAIN_ENUM_MIN(MatrixType::MaxRowsAtCompileTime - EIGEN_PLAIN_ENUM_MAX(-DiagIndex, 0), MatrixType::MaxColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))), MaxColsAtCompileTime = 1, +#ifndef EIGEN_TEST_EVALUATORS MaskLvalueBit = is_lvalue::value ? LvalueBit : 0, Flags = (unsigned int)_MatrixTypeNested::Flags & (HereditaryBits | LinearAccessBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, -#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = _MatrixTypeNested::CoeffReadCost, +#else + MaskLvalueBit = is_lvalue::value ? LvalueBit : 0, + Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions #endif MatrixTypeOuterStride = outer_stride_at_compile_time::ret, InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1, diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index 784e4b1ce..ba0042ba4 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -275,6 +275,7 @@ struct traits > typedef typename DiagonalVectorType::Scalar Scalar; typedef typename DiagonalVectorType::Index Index; typedef typename DiagonalVectorType::StorageKind StorageKind; + typedef typename traits::XprKind XprKind; enum { RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, diff --git a/Eigen/src/Core/DiagonalProduct.h b/Eigen/src/Core/DiagonalProduct.h index 840b70dbb..c6dafdddc 100644 --- a/Eigen/src/Core/DiagonalProduct.h +++ b/Eigen/src/Core/DiagonalProduct.h @@ -26,6 +26,7 @@ struct traits > MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, +#ifndef EIGEN_TEST_EVALUATORS _StorageOrder = MatrixType::Flags & RowMajorBit ? RowMajor : ColMajor, _ScalarAccessOnDiag = !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft) ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)), @@ -34,11 +35,10 @@ struct traits > //_Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))), _Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))), _LinearAccessMask = (RowsAtCompileTime==1 || ColsAtCompileTime==1) ? LinearAccessBit : 0, - - Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit //(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit), -#ifndef EIGEN_TEST_EVALUATORS - , + Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit, //(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit), CoeffReadCost = NumTraits::MulCost + MatrixType::CoeffReadCost + DiagonalType::DiagonalVectorType::CoeffReadCost +#else + Flags = RowMajorBit & (unsigned int)(MatrixType::Flags) #endif }; }; diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index 8ea13cfb7..23bbb46bf 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -79,10 +79,11 @@ struct traits > OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0 ? int(PlainObjectType::OuterStrideAtCompileTime) : int(StrideType::OuterStrideAtCompileTime), + IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned), +#ifndef EIGEN_TEST_EVALUATORS HasNoInnerStride = InnerStrideAtCompileTime == 1, HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0, HasNoStride = HasNoInnerStride && HasNoOuterStride, - IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned), IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic, KeepsPacketAccess = bool(HasNoInnerStride) && ( bool(IsDynamicSize) @@ -95,6 +96,10 @@ struct traits > ? int(Flags1) : int(Flags1 & ~LinearAccessBit), Flags3 = is_lvalue::value ? int(Flags2) : (int(Flags2) & ~LvalueBit), Flags = KeepsPacketAccess ? int(Flags3) : (int(Flags3) & ~PacketAccessBit) +#else + Flags0 = TraitsBase::Flags & (~NestByRefBit), + Flags = is_lvalue::value ? int(Flags0) : (int(Flags0) & ~LvalueBit) +#endif }; private: enum { Options }; // Expressions don't have Options diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index ffa1371c2..de1424b09 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -161,11 +161,16 @@ template class MapBase EIGEN_DEVICE_FUNC void checkSanity() const { +#ifndef EIGEN_TEST_EVALUATORS + // moved to evaluator EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(internal::traits::Flags&PacketAccessBit, internal::inner_stride_at_compile_time::ret==1), PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1); - eigen_assert(EIGEN_IMPLIES(internal::traits::Flags&AlignedBit, (size_t(m_data) % 16) == 0) - && "data is not aligned"); + eigen_assert(EIGEN_IMPLIES(internal::traits::Flags&AlignedBit, (size_t(m_data) % 16) == 0) && "data is not aligned"); +#else + eigen_assert(EIGEN_IMPLIES(internal::traits::IsAligned, (size_t(m_data) % 16) == 0) && "data is not aligned"); +#endif + } PointerType m_data; diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index cac90bc1f..453180049 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -33,14 +33,29 @@ template class Pro namespace internal { template struct traits > - : traits > -{ - // We want A+B*C to be of type Product and not Product - // TODO: This flag should eventually go in a separate evaluator traits class +{ + typedef typename remove_all::type LhsCleaned; + typedef typename remove_all::type RhsCleaned; + + typedef MatrixXpr XprKind; + + typedef typename scalar_product_traits::ReturnType Scalar; + typedef typename promote_storage_type::StorageKind, + typename traits::StorageKind>::ret StorageKind; + typedef typename promote_index_type::Index, + typename traits::Index>::type Index; + enum { - Flags = traits >::Flags & ~(EvalBeforeNestingBit | DirectAccessBit) + RowsAtCompileTime = LhsCleaned::RowsAtCompileTime, + ColsAtCompileTime = RhsCleaned::ColsAtCompileTime, + MaxRowsAtCompileTime = LhsCleaned::MaxRowsAtCompileTime, + MaxColsAtCompileTime = RhsCleaned::MaxColsAtCompileTime, + + // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator. + Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0) }; }; + } // end namespace internal @@ -59,8 +74,6 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, typename internal::promote_storage_type::ret>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Product) - - typedef typename internal::nested::type LhsNested; typedef typename internal::nested::type RhsNested; diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 7ebf31696..1159c2f44 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -17,19 +17,6 @@ namespace Eigen { namespace internal { -/** \internal - * \class product_evaluator - * Products need their own evaluator with more template arguments allowing for - * easier partial template specializations. - */ -template< typename T, - int ProductTag = internal::product_type::ret, - typename LhsShape = typename evaluator_traits::Shape, - typename RhsShape = typename evaluator_traits::Shape, - typename LhsScalar = typename T::Lhs::Scalar, - typename RhsScalar = typename T::Rhs::Scalar - > struct product_evaluator; - /** \internal * Evaluator of a product expression. * Since products require special treatments to handle all possible cases, @@ -119,6 +106,18 @@ struct product_evaluator, ProductTag, DenseSha : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); + +// FIXME shall we handle nested_eval here? +// typedef typename internal::nested_eval::type LhsNested; +// typedef typename internal::nested_eval::type RhsNested; +// typedef typename internal::remove_all::type LhsNestedCleaned; +// typedef typename internal::remove_all::type RhsNestedCleaned; +// +// const LhsNested lhs(xpr.lhs()); +// const RhsNested rhs(xpr.rhs()); +// +// generic_product_impl::evalTo(m_result, lhs, rhs); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); } @@ -133,6 +132,7 @@ struct Assignment, internal::assign_ typedef Product SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { + // FIXME shall we handle nested_eval here? generic_product_impl::evalTo(dst, src.lhs(), src.rhs()); } }; @@ -144,6 +144,7 @@ struct Assignment, internal::add_ass typedef Product SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op &) { + // FIXME shall we handle nested_eval here? generic_product_impl::addTo(dst, src.lhs(), src.rhs()); } }; @@ -155,6 +156,7 @@ struct Assignment, internal::sub_ass typedef Product SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op &) { + // FIXME shall we handle nested_eval here? generic_product_impl::subTo(dst, src.lhs(), src.rhs()); } }; @@ -368,7 +370,6 @@ struct product_evaluator, ProductTag, DenseShape, : evaluator_base > { typedef Product XprType; - typedef CoeffBasedProduct CoeffBasedProductType; typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; @@ -396,9 +397,13 @@ struct product_evaluator, ProductTag, DenseShape, typedef typename evaluator::type RhsEtorType; enum { - RowsAtCompileTime = traits::RowsAtCompileTime, + RowsAtCompileTime = LhsNestedCleaned::RowsAtCompileTime, + ColsAtCompileTime = RhsNestedCleaned::ColsAtCompileTime, + InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsNestedCleaned::ColsAtCompileTime, RhsNestedCleaned::RowsAtCompileTime), + MaxRowsAtCompileTime = LhsNestedCleaned::MaxRowsAtCompileTime, + MaxColsAtCompileTime = RhsNestedCleaned::MaxColsAtCompileTime, + PacketSize = packet_traits::size, - InnerSize = traits::InnerSize, LhsCoeffReadCost = LhsEtorType::CoeffReadCost, RhsCoeffReadCost = RhsEtorType::CoeffReadCost, @@ -407,8 +412,51 @@ struct product_evaluator, ProductTag, DenseShape, + (InnerSize - 1) * NumTraits::AddCost, Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT, - CanVectorizeInner = traits::CanVectorizeInner, - Flags = traits::Flags + + LhsFlags = LhsEtorType::Flags, + RhsFlags = RhsEtorType::Flags, + + LhsRowMajor = LhsFlags & RowMajorBit, + RhsRowMajor = RhsFlags & RowMajorBit, + + SameType = is_same::value, + + CanVectorizeRhs = RhsRowMajor && (RhsFlags & PacketAccessBit) + && (ColsAtCompileTime == Dynamic + || ( (ColsAtCompileTime % packet_traits::size) == 0 + && (RhsFlags&AlignedBit) + ) + ), + + CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit) + && (RowsAtCompileTime == Dynamic + || ( (RowsAtCompileTime % packet_traits::size) == 0 + && (LhsFlags&AlignedBit) + ) + ), + + EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1 + : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0 + : (RhsRowMajor && !CanVectorizeLhs), + + Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit) + | (EvalToRowMajor ? RowMajorBit : 0) + | (CanVectorizeLhs ? (LhsFlags & AlignedBit) : 0) + | (CanVectorizeRhs ? (RhsFlags & AlignedBit) : 0) + // TODO enable vectorization for mixed types + | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0), + + /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside + * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner + * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect + * the Flags, it is safe to make this value depend on ActualPacketAccessBit, that doesn't affect the ABI. + */ + CanVectorizeInner = SameType + && LhsRowMajor + && (!RhsRowMajor) + && (LhsFlags & RhsFlags & ActualPacketAccessBit) + && (LhsFlags & RhsFlags & AlignedBit) + && (InnerSize % packet_traits::size == 0) }; const CoeffReturnType coeff(Index row, Index col) const @@ -689,7 +737,7 @@ protected: * Diagonal products ***************************************************************************/ -template +template struct diagonal_product_evaluator_base : evaluator_base { @@ -698,7 +746,20 @@ struct diagonal_product_evaluator_base typedef typename internal::packet_traits::type PacketScalar; public: enum { - CoeffReadCost = NumTraits::MulCost + evaluator::CoeffReadCost + evaluator::CoeffReadCost + CoeffReadCost = NumTraits::MulCost + evaluator::CoeffReadCost + evaluator::CoeffReadCost, + + MatrixFlags = evaluator::Flags, + DiagFlags = evaluator::Flags, + _StorageOrder = MatrixFlags & RowMajorBit ? RowMajor : ColMajor, + _ScalarAccessOnDiag = !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft) + ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)), + _SameTypes = is_same::value, + // FIXME currently we need same types, but in the future the next rule should be the one + //_Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagFlags)&PacketAccessBit))), + _Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))), + _LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0, + Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit + //(int(MatrixFlags)&int(DiagFlags)&AlignedBit), }; diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag) @@ -724,7 +785,7 @@ protected: { enum { InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime, - DiagonalPacketLoadMode = (LoadMode == Aligned && (((InnerSize%16) == 0) || (int(DiagonalType::Flags)&AlignedBit)==AlignedBit) ? Aligned : Unaligned) + DiagonalPacketLoadMode = (LoadMode == Aligned && (((InnerSize%16) == 0) || (int(DiagFlags)&AlignedBit)==AlignedBit) ? Aligned : Unaligned) }; return internal::pmul(m_matImpl.template packet(row, col), m_diagImpl.template packet(id)); @@ -737,9 +798,9 @@ protected: // diagonal * dense template struct product_evaluator, ProductTag, DiagonalShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> - : diagonal_product_evaluator_base > + : diagonal_product_evaluator_base, OnTheLeft> { - typedef diagonal_product_evaluator_base > Base; + typedef diagonal_product_evaluator_base, OnTheLeft> Base; using Base::m_diagImpl; using Base::m_matImpl; using Base::coeff; @@ -783,9 +844,9 @@ struct product_evaluator, ProductTag, DiagonalSha // dense * diagonal template struct product_evaluator, ProductTag, DenseShape, DiagonalShape, typename Lhs::Scalar, typename Rhs::Scalar> - : diagonal_product_evaluator_base > + : diagonal_product_evaluator_base, OnTheRight> { - typedef diagonal_product_evaluator_base > Base; + typedef diagonal_product_evaluator_base, OnTheRight> Base; using Base::m_diagImpl; using Base::m_matImpl; using Base::coeff; diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 41290323f..6c8c58e95 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -389,8 +389,19 @@ DenseBase::redux(const Func& func) const eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); #ifdef EIGEN_TEST_EVALUATORS + // FIXME, eval_nest should be handled by redux_evaluator, however: + // - it is currently difficult to provide the right Flags since they are still handled by the expressions + // - handling it here might reduce the number of template instantiations +// typedef typename internal::nested_eval::type ThisNested; +// typedef typename internal::remove_all::type ThisNestedCleaned; +// typedef typename internal::redux_evaluator ThisEvaluator; +// +// ThisNested thisNested(derived()); +// ThisEvaluator thisEval(thisNested); + typedef typename internal::redux_evaluator ThisEvaluator; ThisEvaluator thisEval(derived()); + return internal::redux_impl::run(thisEval, func); #else diff --git a/Eigen/src/Core/Replicate.h b/Eigen/src/Core/Replicate.h index 1e640d8aa..2dff03ea3 100644 --- a/Eigen/src/Core/Replicate.h +++ b/Eigen/src/Core/Replicate.h @@ -53,10 +53,13 @@ struct traits > IsRowMajor = MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1 ? 1 : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0 : (MatrixType::Flags & RowMajorBit) ? 1 : 0, - Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0) + #ifndef EIGEN_TEST_EVALUATORS - , + Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0), CoeffReadCost = _MatrixTypeNested::CoeffReadCost +#else + // FIXME enable DirectAccess with negative strides? + Flags = IsRowMajor ? RowMajorBit : 0 #endif }; }; diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h index 495b44cc4..4969bb4fc 100644 --- a/Eigen/src/Core/Reverse.h +++ b/Eigen/src/Core/Reverse.h @@ -45,14 +45,15 @@ struct traits > MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, +#ifndef EIGEN_TEST_EVALUATORS // let's enable LinearAccess only with vectorization because of the product overhead LinearAccess = ( (Direction==BothDirections) && (int(_MatrixTypeNested::Flags)&PacketAccessBit) ) ? LinearAccessBit : 0, - Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess) -#ifndef EIGEN_TEST_EVALUATORS - , + Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess), CoeffReadCost = _MatrixTypeNested::CoeffReadCost +#else + Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit) #endif }; }; diff --git a/Eigen/src/Core/Select.h b/Eigen/src/Core/Select.h index abcba2d15..d4fd88e62 100644 --- a/Eigen/src/Core/Select.h +++ b/Eigen/src/Core/Select.h @@ -43,12 +43,13 @@ struct traits > ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime, MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime, - Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits #ifndef EIGEN_TEST_EVALUATORS - , + Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits, CoeffReadCost = traits::type>::CoeffReadCost + EIGEN_SIZE_MAX(traits::type>::CoeffReadCost, traits::type>::CoeffReadCost) +#else + Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & RowMajorBit #endif }; }; diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 98f9e888f..11b0e45a8 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -41,12 +41,17 @@ struct traits > : traits ColsAtCompileTime = MatrixType::RowsAtCompileTime, MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime, MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime, +#ifndef EIGEN_TEST_EVALUATORS FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit), Flags1 = Flags0 | FlagsLvalueBit, Flags = Flags1 ^ RowMajorBit, -#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = MatrixTypeNestedPlain::CoeffReadCost, +#else + FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, + Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit), + Flags1 = Flags0 | FlagsLvalueBit, + Flags = Flags1 ^ RowMajorBit, #endif InnerStrideAtCompileTime = inner_stride_at_compile_time::ret, OuterStrideAtCompileTime = outer_stride_at_compile_time::ret diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index 702d0006d..672b9662f 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -48,8 +48,12 @@ struct traits > ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime, MaxRowsAtCompileTime = Direction==Vertical ? 1 : MatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime, +#ifndef EIGEN_TEST_EVALUATORS Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits, Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0), +#else + Flags = RowsAtCompileTime == 1 ? RowMajorBit : 0, +#endif TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime : MatrixType::ColsAtCompileTime }; #ifndef EIGEN_TEST_EVALUATORS diff --git a/Eigen/src/Core/products/TriangularMatrixVector.h b/Eigen/src/Core/products/TriangularMatrixVector.h index eed7f4258..771613b11 100644 --- a/Eigen/src/Core/products/TriangularMatrixVector.h +++ b/Eigen/src/Core/products/TriangularMatrixVector.h @@ -259,7 +259,7 @@ template struct trmv_selector typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; typedef internal::blas_traits RhsBlasTraits; typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; - + typedef Map, Aligned> MappedDest; typename internal::add_const_on_value_type::type actualLhs = LhsBlasTraits::extract(lhs); diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 975fdbf2a..092ba758e 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -157,6 +157,18 @@ template struct product_type; +/** \internal + * \class product_evaluator + * Products need their own evaluator with more template arguments allowing for + * easier partial template specializations. + */ +template< typename T, + int ProductTag = internal::product_type::ret, + typename LhsShape = typename evaluator_traits::Shape, + typename RhsShape = typename evaluator_traits::Shape, + typename LhsScalar = typename T::Lhs::Scalar, + typename RhsScalar = typename T::Rhs::Scalar + > struct product_evaluator; } template type; }; +#ifndef EIGEN_TEST_EVALUATORS template class compute_matrix_flags { @@ -158,6 +159,57 @@ class compute_matrix_flags enum { ret = LinearAccessBit | LvalueBit | DirectAccessBit | NestByRefBit | packet_access_bit | row_major_bit | aligned_bit }; }; +#else // EIGEN_TEST_EVALUATORS + +template +class compute_matrix_flags +{ + enum { row_major_bit = Options&RowMajor ? RowMajorBit : 0 }; + public: + // FIXME currently we still have to handle DirectAccessBit at the expression level to handle DenseCoeffsBase<> + // and then propagate this information to the evaluator's flags. + // However, I (Gael) think that DirectAccessBit should only matter at the evaluation stage. + enum { ret = DirectAccessBit | LvalueBit | NestByRefBit | row_major_bit }; +}; +#endif + +#ifdef EIGEN_ENABLE_EVALUATORS +template +class compute_matrix_evaluator_flags +{ + enum { + row_major_bit = Options&RowMajor ? RowMajorBit : 0, + is_dynamic_size_storage = MaxRows==Dynamic || MaxCols==Dynamic, + + aligned_bit = + ( + ((Options&DontAlign)==0) + && ( +#if EIGEN_ALIGN_STATICALLY + ((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % 16) == 0)) +#else + 0 +#endif + + || + +#if EIGEN_ALIGN + is_dynamic_size_storage +#else + 0 +#endif + + ) + ) ? AlignedBit : 0, + packet_access_bit = packet_traits::Vectorizable && aligned_bit ? PacketAccessBit : 0 + }; + + public: + enum { ret = LinearAccessBit | DirectAccessBit | packet_access_bit | row_major_bit | aligned_bit }; +}; + +#endif // EIGEN_ENABLE_EVALUATORS + template struct size_at_compile_time { enum { ret = (_Rows==Dynamic || _Cols==Dynamic) ? Dynamic : _Rows * _Cols }; From 0bd5671b9e0fe85a7de50d316688b83927900bca Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 13:35:44 +0100 Subject: [PATCH 0355/1103] Fix Eigenvalues module --- Eigen/src/Eigenvalues/Tridiagonalization.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Eigenvalues/Tridiagonalization.h b/Eigen/src/Eigenvalues/Tridiagonalization.h index 192278d68..e3a27f275 100644 --- a/Eigen/src/Eigenvalues/Tridiagonalization.h +++ b/Eigen/src/Eigenvalues/Tridiagonalization.h @@ -18,8 +18,10 @@ namespace internal { template struct TridiagonalizationMatrixTReturnType; template struct traits > + : public traits { - typedef typename MatrixType::PlainObject ReturnType; + typedef typename MatrixType::PlainObject ReturnType; // FIXME shall it be a BandMatrix? + enum { Flags = 0 }; }; template From 88aa18df641f8235c88661e41ea128f33d88a8f6 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 12 Mar 2014 13:43:19 +0100 Subject: [PATCH 0356/1103] bug #759: Removed hard-coded double-math from Quaternion::angularDistance. Some documentation improvements --- Eigen/src/Geometry/Quaternion.h | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index 6ee0e4f75..a712692e5 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -204,6 +204,8 @@ class QuaternionBase : public RotationBase * \li \c Quaternionf for \c float * \li \c Quaterniond for \c double * + * \warning Operations interpreting the quaternion as rotation have undefined behavior if the quaternion is not normalized. + * * \sa class AngleAxis, class Transform */ @@ -345,7 +347,7 @@ class Map, _Options > /** Constructs a Mapped Quaternion object from the pointer \a coeffs * - * The pointer \a coeffs must reference the four coeffecients of Quaternion in the following order: + * The pointer \a coeffs must reference the four coefficients of Quaternion in the following order: * \code *coeffs == {x, y, z, w} \endcode * * If the template parameter _Options is set to #Aligned, then the pointer coeffs must be aligned. */ @@ -465,7 +467,7 @@ QuaternionBase::_transformVector(Vector3 v) const // Note that this algorithm comes from the optimization by hand // of the conversion to a Matrix followed by a Matrix/Vector product. // It appears to be much faster than the common algorithm found - // in the litterature (30 versus 39 flops). It also requires two + // in the literature (30 versus 39 flops). It also requires two // Vector3 as temporaries. Vector3 uv = this->vec().cross(v); uv += uv; @@ -668,10 +670,10 @@ QuaternionBase::angularDistance(const QuaternionBase& oth { using std::acos; using std::abs; - double d = abs(this->dot(other)); - if (d>=1.0) + Scalar d = abs(this->dot(other)); + if (d>=Scalar(1)) return Scalar(0); - return static_cast(2 * acos(d)); + return Scalar(2) * acos(d); } From 2379ccffcb8f1af10f8ed91353d5cb8b3a9a7847 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 12 Mar 2014 13:48:09 +0100 Subject: [PATCH 0357/1103] bug #755: CommaInitializer produced wrong assertions in absence of ReturnValueOptimization. --- Eigen/src/Core/CommaInitializer.h | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/CommaInitializer.h b/Eigen/src/Core/CommaInitializer.h index 2bbf74b05..70cbfeff5 100644 --- a/Eigen/src/Core/CommaInitializer.h +++ b/Eigen/src/Core/CommaInitializer.h @@ -45,6 +45,18 @@ struct CommaInitializer m_xpr.block(0, 0, other.rows(), other.cols()) = other; } + /* Copy/Move constructor which transfers ownership. This is crucial in + * absence of return value optimization to avoid assertions during destruction. */ + // FIXME in C++11 mode this could be replaced by a proper RValue constructor + EIGEN_DEVICE_FUNC + inline CommaInitializer(const CommaInitializer& o) + : m_xpr(o.m_xpr), m_row(o.m_row), m_col(o.m_col), m_currentBlockRows(o.m_currentBlockRows) { + // Mark original object as finished. In absence of R-value references we need to const_cast: + const_cast(o).m_row = m_xpr.rows(); + const_cast(o).m_col = m_xpr.cols(); + const_cast(o).m_currentBlockRows = 0; + } + /* inserts a scalar value in the target matrix */ EIGEN_DEVICE_FUNC CommaInitializer& operator,(const Scalar& s) @@ -110,7 +122,7 @@ struct CommaInitializer EIGEN_DEVICE_FUNC inline XprType& finished() { return m_xpr; } - XprType& m_xpr; // target expression + XprType& m_xpr; // target expression Index m_row; // current row id Index m_col; // current col id Index m_currentBlockRows; // current block height From a6be1952f4d345b10011c605a833e609bcc695e7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 16:18:34 +0100 Subject: [PATCH 0358/1103] Fix a few regression when moving the flags --- Eigen/Core | 7 +++++++ Eigen/src/Core/AssignEvaluator.h | 4 ++++ Eigen/src/Core/CoreEvaluators.h | 8 ++++---- Eigen/src/Core/Matrix.h | 3 +++ Eigen/src/Core/PlainObjectBase.h | 4 ++++ Eigen/src/Core/Product.h | 8 ++++++++ 6 files changed, 30 insertions(+), 4 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 474ef5f8f..6d5571875 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -11,6 +11,13 @@ #ifndef EIGEN_CORE_H #define EIGEN_CORE_H +// EIGEN_TEST_EVALUATORS => EIGEN_ENABLE_EVALUATORS +#ifdef EIGEN_TEST_EVALUATORS +#ifndef EIGEN_ENABLE_EVALUATORS +#define EIGEN_ENABLE_EVALUATORS +#endif +#endif + // first thing Eigen does: stop the compiler from committing suicide #include "src/Core/util/DisableStupidWarnings.h" diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 05816094c..62ab531f0 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -126,6 +126,10 @@ public: EIGEN_DEBUG_VAR(PacketSize) EIGEN_DEBUG_VAR(StorageOrdersAgree) EIGEN_DEBUG_VAR(MightVectorize) + std::cerr.setf(std::ios::hex, std::ios::basefield); + EIGEN_DEBUG_VAR(DstFlags) + EIGEN_DEBUG_VAR(SrcFlags) + std::cerr.unsetf(std::ios::hex); EIGEN_DEBUG_VAR(MayLinearize) EIGEN_DEBUG_VAR(MayInnerVectorize) EIGEN_DEBUG_VAR(MayLinearVectorize) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index a5de3593c..5f2d742f8 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -779,10 +779,10 @@ struct evaluator > MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % 16) == 0)) ? AlignedBit : 0, FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1) ? LinearAccessBit : 0, FlagsRowMajorBit = XprType::Flags&RowMajorBit, - Flags0 = traits::Flags & ( (HereditaryBits & ~RowMajorBit) | - DirectAccessBit | - MaskPacketAccessBit | - MaskAlignedBit), + Flags0 = evaluator::Flags & ( (HereditaryBits & ~RowMajorBit) | + DirectAccessBit | + MaskPacketAccessBit | + MaskAlignedBit), Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit }; typedef block_evaluator block_evaluator_type; diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index f84db0dc2..cd4913525 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -117,6 +117,9 @@ struct traits > Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret, #ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = NumTraits::ReadCost, +#else + // FIXME, the following flag in only used to define NeedsToAlign in PlainObjectBase + EvaluatorFlags = compute_matrix_evaluator_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret, #endif Options = _Options, InnerStrideAtCompileTime = 1, diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 4f456804c..c31bfbc8c 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -128,7 +128,11 @@ class PlainObjectBase : public internal::dense_xpr_base::type DenseStorage m_storage; public: +#ifndef EIGEN_TEST_EVALUATORS enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits::Flags & AlignedBit) != 0 }; +#else + enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits::EvaluatorFlags & AlignedBit) != 0 }; +#endif EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 453180049..2785847a6 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -51,6 +51,14 @@ struct traits > MaxRowsAtCompileTime = LhsCleaned::MaxRowsAtCompileTime, MaxColsAtCompileTime = RhsCleaned::MaxColsAtCompileTime, + // FIXME: only needed by GeneralMatrixMatrixTriangular + InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsCleaned::ColsAtCompileTime, RhsCleaned::RowsAtCompileTime), + +#ifndef EIGEN_TEST_EVALUATORS + // dummy, for evaluators unit test only + CoeffReadCost = Dynamic, +#endif + // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator. Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0) }; From 0b362e0c9ab1670e73fc8ae951d765f686342882 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 16:18:54 +0100 Subject: [PATCH 0359/1103] This file is not needed anymore --- Eigen/src/Core/products/CoeffBasedProduct.h | 3 +++ 1 file changed, 3 insertions(+) diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index 94099b7d3..85e8ac2c7 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -13,6 +13,7 @@ namespace Eigen { +#ifndef EIGEN_TEST_EVALUATORS namespace internal { /********************************************************************************* @@ -451,6 +452,8 @@ struct product_packet_impl #endif // EIGEN_TEST_EVALUATORS +#endif // Date: Wed, 12 Mar 2014 18:13:18 +0100 Subject: [PATCH 0360/1103] Extend evaluation traits debuging info --- Eigen/src/Core/AssignEvaluator.h | 11 +++++++---- Eigen/src/Core/Redux.h | 26 +++++++++++++++++++++++++- 2 files changed, 32 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 62ab531f0..a083e340e 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -118,6 +118,12 @@ public: #ifdef EIGEN_DEBUG_ASSIGN static void debug() { + std::cerr << "DstXpr: " << typeid(typename DstEvaluator::XprType).name() << std::endl; + std::cerr << "SrcXpr: " << typeid(typename SrcEvaluator::XprType).name() << std::endl; + std::cerr.setf(std::ios::hex, std::ios::basefield); + EIGEN_DEBUG_VAR(DstFlags) + EIGEN_DEBUG_VAR(SrcFlags) + std::cerr.unsetf(std::ios::hex); EIGEN_DEBUG_VAR(DstIsAligned) EIGEN_DEBUG_VAR(SrcIsAligned) EIGEN_DEBUG_VAR(JointAlignment) @@ -126,10 +132,6 @@ public: EIGEN_DEBUG_VAR(PacketSize) EIGEN_DEBUG_VAR(StorageOrdersAgree) EIGEN_DEBUG_VAR(MightVectorize) - std::cerr.setf(std::ios::hex, std::ios::basefield); - EIGEN_DEBUG_VAR(DstFlags) - EIGEN_DEBUG_VAR(SrcFlags) - std::cerr.unsetf(std::ios::hex); EIGEN_DEBUG_VAR(MayLinearize) EIGEN_DEBUG_VAR(MayInnerVectorize) EIGEN_DEBUG_VAR(MayLinearVectorize) @@ -139,6 +141,7 @@ public: EIGEN_DEBUG_VAR(MayUnrollCompletely) EIGEN_DEBUG_VAR(MayUnrollInner) EIGEN_DEBUG_VAR(Unrolling) + std::cerr << std::endl; } #endif }; diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 6c8c58e95..b9d59af47 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -65,6 +65,29 @@ public: ? CompleteUnrolling : NoUnrolling }; + +#ifdef EIGEN_DEBUG_ASSIGN + static void debug() + { +#ifdef EIGEN_TEST_EVALUATORS + std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl; +#else + std::cerr << "Xpr: " << typeid(Derived).name() << std::endl; +#endif + std::cerr.setf(std::ios::hex, std::ios::basefield); + EIGEN_DEBUG_VAR(Derived::Flags) + std::cerr.unsetf(std::ios::hex); + EIGEN_DEBUG_VAR(InnerMaxSize) + EIGEN_DEBUG_VAR(PacketSize) + EIGEN_DEBUG_VAR(MightVectorize) + EIGEN_DEBUG_VAR(MayLinearVectorize) + EIGEN_DEBUG_VAR(MaySliceVectorize) + EIGEN_DEBUG_VAR(Traversal) + EIGEN_DEBUG_VAR(UnrollingLimit) + EIGEN_DEBUG_VAR(Unrolling) + std::cerr << std::endl; + } +#endif }; /*************************************************************************** @@ -311,10 +334,11 @@ struct redux_impl #ifdef EIGEN_ENABLE_EVALUATORS // evaluator adaptor -template +template class redux_evaluator { public: + typedef _XprType XprType; redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} typedef typename XprType::Index Index; From f74ed345395b57f299d597c835177222120d9992 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 18:14:08 +0100 Subject: [PATCH 0361/1103] Fix regressions in redux_evaluator flags and evaluator flags --- Eigen/src/Core/Block.h | 2 +- Eigen/src/Core/CoreEvaluators.h | 26 +++++++++++++------------- Eigen/src/Core/Redux.h | 2 +- 3 files changed, 15 insertions(+), 15 deletions(-) diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index d92797a98..e0b24e199 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -98,7 +98,7 @@ struct traits > : traits::value ? LvalueBit : 0, FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, Flags = (traits::Flags & DirectAccessBit) | FlagsLvalueBit | FlagsRowMajorBit // FIXME DirectAccessBit should not be handled by expressions diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 5f2d742f8..47f50d548 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -756,23 +756,23 @@ struct evaluator > enum { CoeffReadCost = evaluator::CoeffReadCost, - RowsAtCompileTime = traits::RowsAtCompileTime, - ColsAtCompileTime = traits::ColsAtCompileTime, - MaxRowsAtCompileTime = traits::MaxRowsAtCompileTime, - MaxColsAtCompileTime = traits::MaxColsAtCompileTime, + RowsAtCompileTime = traits::RowsAtCompileTime, + ColsAtCompileTime = traits::ColsAtCompileTime, + MaxRowsAtCompileTime = traits::MaxRowsAtCompileTime, + MaxColsAtCompileTime = traits::MaxColsAtCompileTime, - XprTypeIsRowMajor = (int(traits::Flags)&RowMajorBit) != 0, + ArgTypeIsRowMajor = (int(evaluator::Flags)&RowMajorBit) != 0, IsRowMajor = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? 1 : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0 - : XprTypeIsRowMajor, - HasSameStorageOrderAsXprType = (IsRowMajor == XprTypeIsRowMajor), + : ArgTypeIsRowMajor, + HasSameStorageOrderAsArgType = (IsRowMajor == ArgTypeIsRowMajor), InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime), - InnerStrideAtCompileTime = HasSameStorageOrderAsXprType - ? int(inner_stride_at_compile_time::ret) - : int(outer_stride_at_compile_time::ret), - OuterStrideAtCompileTime = HasSameStorageOrderAsXprType - ? int(outer_stride_at_compile_time::ret) - : int(inner_stride_at_compile_time::ret), + InnerStrideAtCompileTime = HasSameStorageOrderAsArgType + ? int(inner_stride_at_compile_time::ret) + : int(outer_stride_at_compile_time::ret), + OuterStrideAtCompileTime = HasSameStorageOrderAsArgType + ? int(outer_stride_at_compile_time::ret) + : int(inner_stride_at_compile_time::ret), MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits::size) == 0) && (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0, diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index b9d59af47..1e2270d34 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -351,7 +351,7 @@ public: MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime, MaxColsAtCompileTime = XprType::MaxColsAtCompileTime, // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator - Flags = XprType::Flags & ~DirectAccessBit, + Flags = evaluator::Flags & ~DirectAccessBit, IsRowMajor = XprType::IsRowMajor, SizeAtCompileTime = XprType::SizeAtCompileTime, InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime, From a395024d4496d6ee5f6876c41a7988e340b392e1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 18:14:58 +0100 Subject: [PATCH 0362/1103] More debug info and use lazyProd instead of operator* to query the right flags --- test/vectorization_logic.cpp | 69 ++++++++++++++++++++++++++++++------ 1 file changed, 58 insertions(+), 11 deletions(-) diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp index aee68a87f..23be0ffba 100644 --- a/test/vectorization_logic.cpp +++ b/test/vectorization_logic.cpp @@ -27,19 +27,43 @@ std::string demangle_unrolling(int t) if(t==CompleteUnrolling) return "CompleteUnrolling"; return "?"; } +std::string demangle_flags(int f) +{ + std::string res; + if(f&RowMajorBit) res += " | RowMajor"; + if(f&PacketAccessBit) res += " | Packet"; + if(f&LinearAccessBit) res += " | Linear"; + if(f&LvalueBit) res += " | Lvalue"; + if(f&DirectAccessBit) res += " | Direct"; + if(f&AlignedBit) res += " | Aligned"; + if(f&NestByRefBit) res += " | NestByRef"; + return res; +} template bool test_assign(const Dst&, const Src&, int traversal, int unrolling) { - internal::assign_traits::debug(); - bool res = internal::assign_traits::Traversal==traversal - && internal::assign_traits::Unrolling==unrolling; +#ifdef EIGEN_TEST_EVALUATORS + typedef internal::copy_using_evaluator_traits,internal::evaluator, internal::assign_op > traits; +#else + typedef internal::assign_traits traits; +#endif + bool res = traits::Traversal==traversal && traits::Unrolling==unrolling; if(!res) { + std::cerr << "Src: " << demangle_flags(Src::Flags) << std::endl; +#ifdef EIGEN_TEST_EVALUATORS + std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; +#endif + std::cerr << "Dst: " << demangle_flags(Dst::Flags) << std::endl; +#ifdef EIGEN_TEST_EVALUATORS + std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; +#endif + traits::debug(); std::cerr << " Expected Traversal == " << demangle_traversal(traversal) - << " got " << demangle_traversal(internal::assign_traits::Traversal) << "\n"; + << " got " << demangle_traversal(traits::Traversal) << "\n"; std::cerr << " Expected Unrolling == " << demangle_unrolling(unrolling) - << " got " << demangle_unrolling(internal::assign_traits::Unrolling) << "\n"; + << " got " << demangle_unrolling(traits::Unrolling) << "\n"; } return res; } @@ -47,15 +71,27 @@ bool test_assign(const Dst&, const Src&, int traversal, int unrolling) template bool test_assign(int traversal, int unrolling) { - internal::assign_traits::debug(); - bool res = internal::assign_traits::Traversal==traversal - && internal::assign_traits::Unrolling==unrolling; +#ifdef EIGEN_TEST_EVALUATORS + typedef internal::copy_using_evaluator_traits,internal::evaluator, internal::assign_op > traits; +#else + typedef internal::assign_traits traits; +#endif + bool res = traits::Traversal==traversal && traits::Unrolling==unrolling; if(!res) { + std::cerr << "Src: " << demangle_flags(Src::Flags) << std::endl; +#ifdef EIGEN_TEST_EVALUATORS + std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; +#endif + std::cerr << "Dst: " << demangle_flags(Dst::Flags) << std::endl; +#ifdef EIGEN_TEST_EVALUATORS + std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; +#endif + traits::debug(); std::cerr << " Expected Traversal == " << demangle_traversal(traversal) - << " got " << demangle_traversal(internal::assign_traits::Traversal) << "\n"; + << " got " << demangle_traversal(traits::Traversal) << "\n"; std::cerr << " Expected Unrolling == " << demangle_unrolling(unrolling) - << " got " << demangle_unrolling(internal::assign_traits::Unrolling) << "\n"; + << " got " << demangle_unrolling(traits::Unrolling) << "\n"; } return res; } @@ -63,10 +99,21 @@ bool test_assign(int traversal, int unrolling) template bool test_redux(const Xpr&, int traversal, int unrolling) { +#ifdef EIGEN_TEST_EVALUATORS + typedef internal::redux_traits,internal::redux_evaluator > traits; +#else typedef internal::redux_traits,Xpr> traits; +#endif + bool res = traits::Traversal==traversal && traits::Unrolling==unrolling; if(!res) { + std::cerr << demangle_flags(Xpr::Flags) << std::endl; +#ifdef EIGEN_TEST_EVALUATORS + std::cerr << demangle_flags(internal::evaluator::Flags) << std::endl; +#endif + traits::debug(); + std::cerr << " Expected Traversal == " << demangle_traversal(traversal) << " got " << demangle_traversal(traits::Traversal) << "\n"; std::cerr << " Expected Unrolling == " << demangle_unrolling(unrolling) @@ -185,7 +232,7 @@ template::Vectori Matrix22 >(DefaultTraversal,CompleteUnrolling))); - VERIFY((test_assign(Matrix11(), Matrix11()*Matrix11(), InnerVectorizedTraversal, CompleteUnrolling))); + VERIFY((test_assign(Matrix11(), Matrix11().lazyProduct(Matrix11()), InnerVectorizedTraversal, CompleteUnrolling))); #endif VERIFY(test_assign(MatrixXX(10,10),MatrixXX(20,20).block(10,10,2,3), From 16d4c7a5e80cc1bcda3eb5d3eb9c5e890d47b2b5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 20:23:44 +0100 Subject: [PATCH 0363/1103] Conditionally disable unit tests that are not supported by evaluators yet --- test/CMakeLists.txt | 66 ++++++++++++++++++++++++--------------------- 1 file changed, 35 insertions(+), 31 deletions(-) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index e1bff179d..83cdb40b6 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -129,7 +129,6 @@ add_custom_target(BuildOfficial) option(EIGEN_TEST_EVALUATORS "Enable work in progress evaluators" OFF) if(EIGEN_TEST_EVALUATORS) add_definitions("-DEIGEN_TEST_EVALUATORS=1") - add_definitions("-DEIGEN_ENABLE_EVALUATORS=1") endif(EIGEN_TEST_EVALUATORS) ei_add_test(meta) @@ -137,6 +136,7 @@ ei_add_test(sizeof) ei_add_test(dynalloc) ei_add_test(nomalloc) ei_add_test(first_aligned) +ei_add_test(nullary) ei_add_test(mixingtypes) ei_add_test(packetmath) ei_add_test(unalignedassert) @@ -144,13 +144,15 @@ ei_add_test(vectorization_logic) ei_add_test(basicstuff) ei_add_test(linearstructure) ei_add_test(integer_types) -ei_add_test(cwiseop) ei_add_test(unalignedcount) ei_add_test(exceptions) ei_add_test(redux) ei_add_test(visitor) ei_add_test(block) ei_add_test(corners) +ei_add_test(swap) +ei_add_test(resize) +ei_add_test(conservative_resize) ei_add_test(product_small) ei_add_test(product_large) ei_add_test(product_extra) @@ -179,6 +181,7 @@ ei_add_test(product_trsolve) ei_add_test(product_mmtr) ei_add_test(product_notemporary) ei_add_test(stable_norm) +ei_add_test(permutationmatrices) ei_add_test(bandmatrix) ei_add_test(cholesky) ei_add_test(lu) @@ -198,30 +201,33 @@ ei_add_test(real_qz) ei_add_test(eigensolver_generalized_real) ei_add_test(jacobi) ei_add_test(jacobisvd) -ei_add_test(geo_orthomethods) -ei_add_test(geo_homogeneous) -ei_add_test(geo_quaternion) -ei_add_test(geo_transformations) -ei_add_test(geo_eulerangles) -ei_add_test(geo_hyperplane) -ei_add_test(geo_parametrizedline) -ei_add_test(geo_alignedbox) -ei_add_test(stdvector) -ei_add_test(stdvector_overload) -ei_add_test(stdlist) -ei_add_test(stddeque) -ei_add_test(resize) -ei_add_test(sparse_vector) -ei_add_test(sparse_basic) -ei_add_test(sparse_product) -ei_add_test(sparse_solvers) -ei_add_test(umeyama) ei_add_test(householder) -ei_add_test(swap) -ei_add_test(conservative_resize) -ei_add_test(permutationmatrices) -ei_add_test(sparse_permutations) -ei_add_test(nullary) +if(NOT EIGEN_TEST_EVALUATORS) + ei_add_test(cwiseop) # Eigen2 related + ei_add_test(geo_orthomethods) + ei_add_test(geo_homogeneous) + ei_add_test(geo_quaternion) + ei_add_test(geo_transformations) + ei_add_test(geo_eulerangles) + ei_add_test(geo_hyperplane) + ei_add_test(geo_parametrizedline) + ei_add_test(geo_alignedbox) + ei_add_test(stdvector) + ei_add_test(stdvector_overload) + ei_add_test(stdlist) + ei_add_test(stddeque) + ei_add_test(sparse_vector) + ei_add_test(sparse_basic) + ei_add_test(sparse_product) + ei_add_test(sparse_solvers) + ei_add_test(sparse_permutations) + ei_add_test(simplicial_cholesky) + ei_add_test(conjugate_gradient) + ei_add_test(bicgstab) + ei_add_test(sparselu) + ei_add_test(sparseqr) +endif(NOT EIGEN_TEST_EVALUATORS) +ei_add_test(umeyama) ei_add_test(nesting_ops "${CMAKE_CXX_FLAGS_DEBUG}") ei_add_test(zerosized) ei_add_test(dontalign) @@ -233,13 +239,9 @@ ei_add_test(special_numbers) ei_add_test(rvalue_types) ei_add_test(dense_storage) -ei_add_test(simplicial_cholesky) -ei_add_test(conjugate_gradient) -ei_add_test(bicgstab) -ei_add_test(sparselu) -ei_add_test(sparseqr) +# # ei_add_test(denseLM) -# ei_add_test(denseLM) +if(NOT EIGEN_TEST_EVALUATORS) if(QT4_FOUND) ei_add_test(qtvector "" "${QT_QTCORE_LIBRARY}") @@ -275,6 +277,8 @@ if(METIS_FOUND) ei_add_test(metis_support "" "${METIS_LIBRARIES}") endif() +endif(NOT EIGEN_TEST_EVALUATORS) + string(TOLOWER "${CMAKE_CXX_COMPILER}" cmake_cxx_compiler_tolower) if(cmake_cxx_compiler_tolower MATCHES "qcc") set(CXX_IS_QCC "ON") From aceae8314b80fbf96e8dc1b0d45c1e99951e770c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 20:25:36 +0100 Subject: [PATCH 0364/1103] Resurect EvalBeforeNestingBit to control nested_eval --- Eigen/src/Core/Inverse.h | 2 ++ Eigen/src/Core/util/Constants.h | 3 +-- Eigen/src/Core/util/XprHelper.h | 3 ++- 3 files changed, 5 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 0cfc71d34..57eaf99f1 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -110,6 +110,8 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; + + enum { Flags = Base::Flags | EvalBeforeNestingBit }; evaluator(const InverseType& inv_xpr) : m_result(inv_xpr.rows(), inv_xpr.cols()) diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index d9e51ffea..64efac8e9 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -53,9 +53,8 @@ const int Infinity = -1; const unsigned int RowMajorBit = 0x1; /** \ingroup flags - * \deprecated * means the expression should be evaluated by the calling expression */ -const unsigned int EvalBeforeNestingBit = 0x2; // FIXME deprecated +const unsigned int EvalBeforeNestingBit = 0x2; /** \ingroup flags * \deprecated diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 016b37f71..0be5029c9 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -432,7 +432,8 @@ template::type> struc }; typedef typename conditional< - int(CostEvalAsInteger) < int(CostNoEvalAsInteger), + ( (int(evaluator::Flags) & EvalBeforeNestingBit) || + (int(CostEvalAsInteger) < int(CostNoEvalAsInteger)) ), PlainObject, typename ref_selector::type >::type type; From 847d801a4c91d3770fa06eb56772b64ed576ce1c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 12 Mar 2014 21:33:45 +0100 Subject: [PATCH 0365/1103] Fix bug #760: complete Eigen's lapack interface with default Lapack for SPQR if there is no fortran compiler. --- test/CMakeLists.txt | 65 +++++++++++++++++++++++++++------------------ 1 file changed, 39 insertions(+), 26 deletions(-) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index e1bff179d..62cbedae7 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -13,11 +13,26 @@ if(NOT EXISTS ${CMAKE_CURRENT_BINARY_DIR}/split_test_helper.h) endforeach() endif() +# check if we have a Fortran compiler +include("../cmake/language_support.cmake") + +workaround_9220(Fortran EIGEN_Fortran_COMPILER_WORKS) + +if(EIGEN_Fortran_COMPILER_WORKS) + enable_language(Fortran OPTIONAL) + if(NOT CMAKE_Fortran_COMPILER) + set(EIGEN_Fortran_COMPILER_WORKS OFF) + endif() +endif() + +if(NOT EIGEN_Fortran_COMPILER_WORKS) + # search for a default Lapack library to complete Eigen's one + find_package(LAPACK) +endif() + # configure blas/lapack (use Eigen's ones) -set(BLAS_FOUND TRUE) -set(LAPACK_FOUND TRUE) -set(BLAS_LIBRARIES eigen_blas) -set(LAPACK_LIBRARIES eigen_lapack) +set(EIGEN_BLAS_LIBRARIES eigen_blas) +set(EIGEN_LAPACK_LIBRARIES eigen_lapack) set(EIGEN_TEST_MATRIX_DIR "" CACHE STRING "Enable testing of realword sparse matrices contained in the specified path") if(EIGEN_TEST_MATRIX_DIR) @@ -32,33 +47,33 @@ endif(EIGEN_TEST_MATRIX_DIR) set(SPARSE_LIBS " ") find_package(Cholmod) -if(CHOLMOD_FOUND AND BLAS_FOUND AND LAPACK_FOUND) +if(CHOLMOD_FOUND) add_definitions("-DEIGEN_CHOLMOD_SUPPORT") include_directories(${CHOLMOD_INCLUDES}) - set(SPARSE_LIBS ${SPARSE_LIBS} ${CHOLMOD_LIBRARIES} ${BLAS_LIBRARIES} ${LAPACK_LIBRARIES}) - set(CHOLMOD_ALL_LIBS ${CHOLMOD_LIBRARIES} ${BLAS_LIBRARIES} ${LAPACK_LIBRARIES}) + set(SPARSE_LIBS ${SPARSE_LIBS} ${CHOLMOD_LIBRARIES} ${EIGEN_BLAS_LIBRARIES} ${EIGEN_LAPACK_LIBRARIES}) + set(CHOLMOD_ALL_LIBS ${CHOLMOD_LIBRARIES} ${EIGEN_BLAS_LIBRARIES} ${EIGEN_LAPACK_LIBRARIES}) ei_add_property(EIGEN_TESTED_BACKENDS "Cholmod, ") else() ei_add_property(EIGEN_MISSING_BACKENDS "Cholmod, ") endif() find_package(Umfpack) -if(UMFPACK_FOUND AND BLAS_FOUND) +if(UMFPACK_FOUND) add_definitions("-DEIGEN_UMFPACK_SUPPORT") include_directories(${UMFPACK_INCLUDES}) - set(SPARSE_LIBS ${SPARSE_LIBS} ${UMFPACK_LIBRARIES} ${BLAS_LIBRARIES}) - set(UMFPACK_ALL_LIBS ${UMFPACK_LIBRARIES} ${BLAS_LIBRARIES}) + set(SPARSE_LIBS ${SPARSE_LIBS} ${UMFPACK_LIBRARIES} ${EIGEN_BLAS_LIBRARIES}) + set(UMFPACK_ALL_LIBS ${UMFPACK_LIBRARIES} ${EIGEN_BLAS_LIBRARIES}) ei_add_property(EIGEN_TESTED_BACKENDS "UmfPack, ") else() ei_add_property(EIGEN_MISSING_BACKENDS "UmfPack, ") endif() find_package(SuperLU) -if(SUPERLU_FOUND AND BLAS_FOUND) +if(SUPERLU_FOUND) add_definitions("-DEIGEN_SUPERLU_SUPPORT") include_directories(${SUPERLU_INCLUDES}) - set(SPARSE_LIBS ${SPARSE_LIBS} ${SUPERLU_LIBRARIES} ${BLAS_LIBRARIES}) - set(SUPERLU_ALL_LIBS ${SUPERLU_LIBRARIES} ${BLAS_LIBRARIES}) + set(SPARSE_LIBS ${SPARSE_LIBS} ${SUPERLU_LIBRARIES} ${EIGEN_BLAS_LIBRARIES}) + set(SUPERLU_ALL_LIBS ${SUPERLU_LIBRARIES} ${EIGEN_BLAS_LIBRARIES}) ei_add_property(EIGEN_TESTED_BACKENDS "SuperLU, ") else() ei_add_property(EIGEN_MISSING_BACKENDS "SuperLU, ") @@ -68,7 +83,7 @@ endif() find_package(Pastix) find_package(Scotch) find_package(Metis) -if(PASTIX_FOUND AND BLAS_FOUND) +if(PASTIX_FOUND) add_definitions("-DEIGEN_PASTIX_SUPPORT") include_directories(${PASTIX_INCLUDES}) if(SCOTCH_FOUND) @@ -80,8 +95,8 @@ if(PASTIX_FOUND AND BLAS_FOUND) else(SCOTCH_FOUND) ei_add_property(EIGEN_MISSING_BACKENDS "PaStiX, ") endif(SCOTCH_FOUND) - set(SPARSE_LIBS ${SPARSE_LIBS} ${PASTIX_LIBRARIES} ${ORDERING_LIBRARIES} ${BLAS_LIBRARIES}) - set(PASTIX_ALL_LIBS ${PASTIX_LIBRARIES} ${BLAS_LIBRARIES}) + set(SPARSE_LIBS ${SPARSE_LIBS} ${PASTIX_LIBRARIES} ${ORDERING_LIBRARIES} ${EIGEN_BLAS_LIBRARIES}) + set(PASTIX_ALL_LIBS ${PASTIX_LIBRARIES} ${EIGEN_BLAS_LIBRARIES}) ei_add_property(EIGEN_TESTED_BACKENDS "PaStiX, ") else() ei_add_property(EIGEN_MISSING_BACKENDS "PaStiX, ") @@ -96,16 +111,14 @@ else() endif() find_package(SPQR) -if(SPQR_FOUND AND BLAS_FOUND AND LAPACK_FOUND) - if(CHOLMOD_FOUND) - add_definitions("-DEIGEN_SPQR_SUPPORT") - include_directories(${SPQR_INCLUDES}) - set(SPQR_ALL_LIBS ${SPQR_LIBRARIES} ${CHOLMOD_LIBRARIES} ${LAPACK_LIBRARIES} ${BLAS_LIBRARIES}) - set(SPARSE_LIBS ${SPARSE_LIBS} ${SPQR_ALL_LIBS}) - ei_add_property(EIGEN_TESTED_BACKENDS "SPQR, ") - else(CHOLMOD_FOUND) - ei_add_property(EIGEN_MISSING_BACKENDS "SPQR, ") - endif(CHOLMOD_FOUND) +if(SPQR_FOUND AND CHOLMOD_FOUND AND (EIGEN_Fortran_COMPILER_WORKS OR LAPACK_FOUND) ) + add_definitions("-DEIGEN_SPQR_SUPPORT") + include_directories(${SPQR_INCLUDES}) + set(SPQR_ALL_LIBS ${SPQR_LIBRARIES} ${CHOLMOD_LIBRARIES} ${EIGEN_LAPACK_LIBRARIES} ${EIGEN_BLAS_LIBRARIES} ${LAPACK_LIBRARIES}) + set(SPARSE_LIBS ${SPARSE_LIBS} ${SPQR_ALL_LIBS}) + ei_add_property(EIGEN_TESTED_BACKENDS "SPQR, ") +else() + ei_add_property(EIGEN_MISSING_BACKENDS "SPQR, ") endif() option(EIGEN_TEST_NOQT "Disable Qt support in unit tests" OFF) From 2db792852ffe8d03b23375da75330c90f490e035 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 13 Mar 2014 12:58:57 +0100 Subject: [PATCH 0366/1103] Silence stupid parenthesis warnings for old GCC versions (<= 4.6.x) --- Eigen/src/SparseCore/SparseBlock.h | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 6f615e250..5b95cc33f 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -338,7 +338,10 @@ const Block SparseMatrixBase::inner namespace internal { template< typename XprType, int BlockRows, int BlockCols, bool InnerPanel, - bool OuterVector = (BlockCols==1 && XprType::IsRowMajor) || (BlockRows==1 && !XprType::IsRowMajor)> + bool OuterVector = (BlockCols==1 && XprType::IsRowMajor) + | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". + // revert to || as soon as not needed anymore. + (BlockRows==1 && !XprType::IsRowMajor)> class GenericSparseBlockInnerIteratorImpl; } From 0a6c472335b593a227c3adbcf1d770187449a30f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 13 Mar 2014 15:44:20 +0100 Subject: [PATCH 0367/1103] A bit of cleaning --- Eigen/src/Core/Assign.h | 82 ++++++++++++++++++++++++++++------------- 1 file changed, 56 insertions(+), 26 deletions(-) diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index 95eb37dd5..0e10f125f 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -509,7 +509,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase #ifdef EIGEN_TEST_EVALUATORS eigen_assert(rows() == other.rows() && cols() == other.cols()); - internal::call_dense_assignment_loop(derived(),other.derived()); + internal::call_assignment_no_alias(derived(),other.derived()); #else // EIGEN_TEST_EVALUATORS @@ -530,23 +530,7 @@ EIGEN_STRONG_INLINE Derived& DenseBase namespace internal { -#ifdef EIGEN_TEST_EVALUATORS - -// TODO remove this class which is now useless - -template -struct assign_selector { - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { - call_assignment(dst, other.derived(), internal::assign_op()); - return dst; - } - template - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; } -}; - -#else // EIGEN_TEST_EVALUATORS +#ifndef EIGEN_TEST_EVALUATORS template::Flags) & EvalBeforeAssigningBit) != 0, bool NeedToTranspose = ((int(Derived::RowsAtCompileTime) == 1 && int(OtherDerived::ColsAtCompileTime) == 1) @@ -585,6 +569,59 @@ struct assign_selector { #endif // EIGEN_TEST_EVALUATORS } // end namespace internal +#ifdef EIGEN_TEST_EVALUATORS +template +template +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& DenseBase::operator=(const DenseBase& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& DenseBase::operator=(const DenseBase& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const MatrixBase& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template +template +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const DenseBase& other) +{ + internal::call_assignment(derived(), other.derived()); + return derived(); +} + +template +template +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const EigenBase& other) +{ + other.derived().evalTo(derived()); + return derived(); +} + +template +template +EIGEN_DEVICE_FUNC +EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const ReturnByValue& other) +{ + other.derived().evalTo(derived()); + return derived(); +} +#else // EIGEN_TEST_EVALUATORS template template EIGEN_DEVICE_FUNC @@ -620,11 +657,7 @@ template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const EigenBase& other) { -#ifdef EIGEN_TEST_EVALUATORS - return internal::assign_selector::evalTo(derived(), other.derived()); -#else return internal::assign_selector::evalTo(derived(), other.derived()); -#endif } template @@ -632,12 +665,9 @@ template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const ReturnByValue& other) { -#ifdef EIGEN_TEST_EVALUATORS - return internal::assign_selector::evalTo(derived(), other.derived()); -#else return internal::assign_selector::evalTo(derived(), other.derived()); -#endif } +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen From bb4b67cf39b2a0aae2c912e1fbad50c1cf3b1ab6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 13 Mar 2014 18:04:19 +0100 Subject: [PATCH 0368/1103] Relax Ref such that Ref accepts a RowVectorXf which can be seen as a degenerate MatrixXf(1,N) --- Eigen/src/Core/Ref.h | 10 +++-- test/ref.cpp | 104 ++++++++++++++++++++++++++----------------- 2 files changed, 69 insertions(+), 45 deletions(-) diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index 00d9e6d2b..cd6d949c4 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -101,7 +101,7 @@ struct traits > template struct match { enum { HasDirectAccess = internal::has_direct_access::ret, - StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)), + StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || Derived::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)), InnerStrideMatch = int(StrideType::InnerStrideAtCompileTime)==int(Dynamic) || int(StrideType::InnerStrideAtCompileTime)==int(Derived::InnerStrideAtCompileTime) || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1), @@ -172,8 +172,12 @@ protected: } else ::new (static_cast(this)) Base(expr.data(), expr.rows(), expr.cols()); - ::new (&m_stride) StrideBase(StrideType::OuterStrideAtCompileTime==0?0:expr.outerStride(), - StrideType::InnerStrideAtCompileTime==0?0:expr.innerStride()); + + if(Expression::IsVectorAtCompileTime && (!PlainObjectType::IsVectorAtCompileTime) && ((Expression::Flags&RowMajorBit)!=(PlainObjectType::Flags&RowMajorBit))) + ::new (&m_stride) StrideBase(expr.innerStride(), StrideType::InnerStrideAtCompileTime==0?0:1); + else + ::new (&m_stride) StrideBase(StrideType::OuterStrideAtCompileTime==0?0:expr.outerStride(), + StrideType::InnerStrideAtCompileTime==0?0:expr.innerStride()); } StrideBase m_stride; diff --git a/test/ref.cpp b/test/ref.cpp index f639d900b..19e81549c 100644 --- a/test/ref.cpp +++ b/test/ref.cpp @@ -154,59 +154,79 @@ template void check_const_correctness(const PlainObjec VERIFY( !(Ref::Flags & LvalueBit) ); } -EIGEN_DONT_INLINE void call_ref_1(Ref ) { } -EIGEN_DONT_INLINE void call_ref_2(const Ref& ) { } -EIGEN_DONT_INLINE void call_ref_3(Ref > ) { } -EIGEN_DONT_INLINE void call_ref_4(const Ref >& ) { } -EIGEN_DONT_INLINE void call_ref_5(Ref > ) { } -EIGEN_DONT_INLINE void call_ref_6(const Ref >& ) { } +template +EIGEN_DONT_INLINE void call_ref_1(Ref a, const B &b) { VERIFY_IS_EQUAL(a,b); } +template +EIGEN_DONT_INLINE void call_ref_2(const Ref& a, const B &b) { VERIFY_IS_EQUAL(a,b); } +template +EIGEN_DONT_INLINE void call_ref_3(Ref > a, const B &b) { VERIFY_IS_EQUAL(a,b); } +template +EIGEN_DONT_INLINE void call_ref_4(const Ref >& a, const B &b) { VERIFY_IS_EQUAL(a,b); } +template +EIGEN_DONT_INLINE void call_ref_5(Ref > a, const B &b) { VERIFY_IS_EQUAL(a,b); } +template +EIGEN_DONT_INLINE void call_ref_6(const Ref >& a, const B &b) { VERIFY_IS_EQUAL(a,b); } +template +EIGEN_DONT_INLINE void call_ref_7(Ref > a, const B &b) { VERIFY_IS_EQUAL(a,b); } void call_ref() { - VectorXcf ca(10); - VectorXf a(10); + VectorXcf ca = VectorXcf::Random(10); + VectorXf a = VectorXf::Random(10); + RowVectorXf b = RowVectorXf::Random(10); + MatrixXf A = MatrixXf::Random(10,10); + RowVector3f c = RowVector3f::Random(); const VectorXf& ac(a); VectorBlock ab(a,0,3); - MatrixXf A(10,10); const VectorBlock abc(a,0,3); + - VERIFY_EVALUATION_COUNT( call_ref_1(a), 0); - //call_ref_1(ac); // does not compile because ac is const - VERIFY_EVALUATION_COUNT( call_ref_1(ab), 0); - VERIFY_EVALUATION_COUNT( call_ref_1(a.head(4)), 0); - VERIFY_EVALUATION_COUNT( call_ref_1(abc), 0); - VERIFY_EVALUATION_COUNT( call_ref_1(A.col(3)), 0); - // call_ref_1(A.row(3)); // does not compile because innerstride!=1 - VERIFY_EVALUATION_COUNT( call_ref_3(A.row(3)), 0); - VERIFY_EVALUATION_COUNT( call_ref_4(A.row(3)), 0); - //call_ref_1(a+a); // does not compile for obvious reason + VERIFY_EVALUATION_COUNT( call_ref_1(a,a), 0); + VERIFY_EVALUATION_COUNT( call_ref_1(b,b.transpose()), 0); +// call_ref_1(ac); // does not compile because ac is const + VERIFY_EVALUATION_COUNT( call_ref_1(ab,ab), 0); + VERIFY_EVALUATION_COUNT( call_ref_1(a.head(4),a.head(4)), 0); + VERIFY_EVALUATION_COUNT( call_ref_1(abc,abc), 0); + VERIFY_EVALUATION_COUNT( call_ref_1(A.col(3),A.col(3)), 0); +// call_ref_1(A.row(3)); // does not compile because innerstride!=1 + VERIFY_EVALUATION_COUNT( call_ref_3(A.row(3),A.row(3).transpose()), 0); + VERIFY_EVALUATION_COUNT( call_ref_4(A.row(3),A.row(3).transpose()), 0); +// call_ref_1(a+a); // does not compile for obvious reason - VERIFY_EVALUATION_COUNT( call_ref_2(A*A.col(1)), 1); // evaluated into a temp - VERIFY_EVALUATION_COUNT( call_ref_2(ac.head(5)), 0); - VERIFY_EVALUATION_COUNT( call_ref_2(ac), 0); - VERIFY_EVALUATION_COUNT( call_ref_2(a), 0); - VERIFY_EVALUATION_COUNT( call_ref_2(ab), 0); - VERIFY_EVALUATION_COUNT( call_ref_2(a.head(4)), 0); - VERIFY_EVALUATION_COUNT( call_ref_2(a+a), 1); // evaluated into a temp - VERIFY_EVALUATION_COUNT( call_ref_2(ca.imag()), 1); // evaluated into a temp + MatrixXf tmp = A*A.col(1); + VERIFY_EVALUATION_COUNT( call_ref_2(A*A.col(1), tmp), 1); // evaluated into a temp + VERIFY_EVALUATION_COUNT( call_ref_2(ac.head(5),ac.head(5)), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(ac,ac), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(a,a), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(ab,ab), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(a.head(4),a.head(4)), 0); + tmp = a+a; + VERIFY_EVALUATION_COUNT( call_ref_2(a+a,tmp), 1); // evaluated into a temp + VERIFY_EVALUATION_COUNT( call_ref_2(ca.imag(),ca.imag()), 1); // evaluated into a temp - VERIFY_EVALUATION_COUNT( call_ref_4(ac.head(5)), 0); - VERIFY_EVALUATION_COUNT( call_ref_4(a+a), 1); // evaluated into a temp - VERIFY_EVALUATION_COUNT( call_ref_4(ca.imag()), 0); + VERIFY_EVALUATION_COUNT( call_ref_4(ac.head(5),ac.head(5)), 0); + tmp = a+a; + VERIFY_EVALUATION_COUNT( call_ref_4(a+a,tmp), 1); // evaluated into a temp + VERIFY_EVALUATION_COUNT( call_ref_4(ca.imag(),ca.imag()), 0); - VERIFY_EVALUATION_COUNT( call_ref_5(a), 0); - VERIFY_EVALUATION_COUNT( call_ref_5(a.head(3)), 0); - VERIFY_EVALUATION_COUNT( call_ref_5(A), 0); - // call_ref_5(A.transpose()); // does not compile - VERIFY_EVALUATION_COUNT( call_ref_5(A.block(1,1,2,2)), 0); + VERIFY_EVALUATION_COUNT( call_ref_5(a,a), 0); + VERIFY_EVALUATION_COUNT( call_ref_5(a.head(3),a.head(3)), 0); + VERIFY_EVALUATION_COUNT( call_ref_5(A,A), 0); +// call_ref_5(A.transpose()); // does not compile + VERIFY_EVALUATION_COUNT( call_ref_5(A.block(1,1,2,2),A.block(1,1,2,2)), 0); + VERIFY_EVALUATION_COUNT( call_ref_5(b,b), 0); // storage order do not match, but this is a degenerate case that should work + VERIFY_EVALUATION_COUNT( call_ref_5(a.row(3),a.row(3)), 0); - VERIFY_EVALUATION_COUNT( call_ref_6(a), 0); - VERIFY_EVALUATION_COUNT( call_ref_6(a.head(3)), 0); - VERIFY_EVALUATION_COUNT( call_ref_6(A.row(3)), 1); // evaluated into a temp thouth it could be avoided by viewing it as a 1xn matrix - VERIFY_EVALUATION_COUNT( call_ref_6(A+A), 1); // evaluated into a temp - VERIFY_EVALUATION_COUNT( call_ref_6(A), 0); - VERIFY_EVALUATION_COUNT( call_ref_6(A.transpose()), 1); // evaluated into a temp because the storage orders do not match - VERIFY_EVALUATION_COUNT( call_ref_6(A.block(1,1,2,2)), 0); + VERIFY_EVALUATION_COUNT( call_ref_6(a,a), 0); + VERIFY_EVALUATION_COUNT( call_ref_6(a.head(3),a.head(3)), 0); + VERIFY_EVALUATION_COUNT( call_ref_6(A.row(3),A.row(3)), 1); // evaluated into a temp thouth it could be avoided by viewing it as a 1xn matrix + tmp = A+A; + VERIFY_EVALUATION_COUNT( call_ref_6(A+A,tmp), 1); // evaluated into a temp + VERIFY_EVALUATION_COUNT( call_ref_6(A,A), 0); + VERIFY_EVALUATION_COUNT( call_ref_6(A.transpose(),A.transpose()), 1); // evaluated into a temp because the storage orders do not match + VERIFY_EVALUATION_COUNT( call_ref_6(A.block(1,1,2,2),A.block(1,1,2,2)), 0); + + VERIFY_EVALUATION_COUNT( call_ref_7(c,c), 0); } void test_ref() From 35a2c9cde7e26ef958883297cefd44db57b1a0bb Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 14 Mar 2014 16:48:29 +0100 Subject: [PATCH 0369/1103] clang does not accept this without template keyword --- Eigen/src/Core/AssignEvaluator.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 5b5d29ca9..5451a138f 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -574,13 +574,13 @@ public: template void assignPacket(Index row, Index col) { - m_functor.assignPacket(&m_dst.coeffRef(row,col), m_src.template packet(row,col)); + m_functor.template assignPacket(&m_dst.coeffRef(row,col), m_src.template packet(row,col)); } template void assignPacket(Index index) { - m_functor.assignPacket(&m_dst.coeffRef(index), m_src.template packet(index)); + m_functor.template assignPacket(&m_dst.coeffRef(index), m_src.template packet(index)); } template From 4fe56a0e0207e69e652dded5263bd0066fc0e4e8 Mon Sep 17 00:00:00 2001 From: Bo Li Date: Sat, 15 Mar 2014 08:42:20 +0800 Subject: [PATCH 0370/1103] fix Spline constructor --- unsupported/Eigen/src/Splines/Spline.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/Splines/Spline.h b/unsupported/Eigen/src/Splines/Spline.h index 771f10432..1b47992d6 100644 --- a/unsupported/Eigen/src/Splines/Spline.h +++ b/unsupported/Eigen/src/Splines/Spline.h @@ -57,7 +57,7 @@ namespace Eigen **/ Spline() : m_knots(1, (Degree==Dynamic ? 2 : 2*Degree+2)) - , m_ctrls(ControlPointVectorType::Zero(2,(Degree==Dynamic ? 1 : Degree+1))) + , m_ctrls(ControlPointVectorType::Zero(Dimension,(Degree==Dynamic ? 1 : Degree+1))) { // in theory this code can go to the initializer list but it will get pretty // much unreadable ... From 3e42b775ead02a2b389840b1b8fd7c28121fb387 Mon Sep 17 00:00:00 2001 From: giacomo po Date: Mon, 17 Mar 2014 16:33:52 -0700 Subject: [PATCH 0371/1103] MINRES, bug #715: add support for zero rhs, and remove square test. --- .../Eigen/src/IterativeSolvers/MINRES.h | 49 ++++++++++++------- unsupported/test/minres.cpp | 30 ++++++++---- 2 files changed, 52 insertions(+), 27 deletions(-) diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index 0e56342a8..98f9ecc17 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -37,22 +37,31 @@ namespace Eigen { typedef typename Dest::Scalar Scalar; typedef Matrix VectorType; + // Check for zero rhs + const RealScalar rhsNorm2(rhs.squaredNorm()); + if(rhsNorm2 == 0) + { + x.setZero(); + iters = 0; + tol_error = 0; + return; + } + // initialize const int maxIters(iters); // initialize maxIters to iters const int N(mat.cols()); // the size of the matrix - const RealScalar rhsNorm2(rhs.squaredNorm()); const RealScalar threshold2(tol_error*tol_error*rhsNorm2); // convergence threshold (compared to residualNorm2) // Initialize preconditioned Lanczos -// VectorType v_old(N); // will be initialized inside loop + VectorType v_old(N); // will be initialized inside loop VectorType v( VectorType::Zero(N) ); //initialize v VectorType v_new(rhs-mat*x); //initialize v_new RealScalar residualNorm2(v_new.squaredNorm()); -// VectorType w(N); // will be initialized inside loop + VectorType w(N); // will be initialized inside loop VectorType w_new(precond.solve(v_new)); // initialize w_new // RealScalar beta; // will be initialized inside loop RealScalar beta_new2(v_new.dot(w_new)); - eigen_assert(beta_new2 >= 0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE"); + eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE"); RealScalar beta_new(sqrt(beta_new2)); const RealScalar beta_one(beta_new); v_new /= beta_new; @@ -62,14 +71,14 @@ namespace Eigen { RealScalar c_old(1.0); RealScalar s(0.0); // the sine of the Givens rotation RealScalar s_old(0.0); // the sine of the Givens rotation -// VectorType p_oold(N); // will be initialized in loop + VectorType p_oold(N); // will be initialized in loop VectorType p_old(VectorType::Zero(N)); // initialize p_old=0 VectorType p(p_old); // initialize p=0 RealScalar eta(1.0); iters = 0; // reset iters - while ( iters < maxIters ){ - + while ( iters < maxIters ) + { // Preconditioned Lanczos /* Note that there are 4 variants on the Lanczos algorithm. These are * described in Paige, C. C. (1972). Computational variants of @@ -81,17 +90,17 @@ namespace Eigen { * A. Greenbaum, Iterative Methods for Solving Linear Systems, SIAM (1987). */ const RealScalar beta(beta_new); -// v_old = v; // update: at first time step, this makes v_old = 0 so value of beta doesn't matter - const VectorType v_old(v); // NOT SURE IF CREATING v_old EVERY ITERATION IS EFFICIENT + v_old = v; // update: at first time step, this makes v_old = 0 so value of beta doesn't matter +// const VectorType v_old(v); // NOT SURE IF CREATING v_old EVERY ITERATION IS EFFICIENT v = v_new; // update -// w = w_new; // update - const VectorType w(w_new); // NOT SURE IF CREATING w EVERY ITERATION IS EFFICIENT + w = w_new; // update +// const VectorType w(w_new); // NOT SURE IF CREATING w EVERY ITERATION IS EFFICIENT v_new.noalias() = mat*w - beta*v_old; // compute v_new const RealScalar alpha = v_new.dot(w); v_new -= alpha*v; // overwrite v_new w_new = precond.solve(v_new); // overwrite w_new beta_new2 = v_new.dot(w_new); // compute beta_new - eigen_assert(beta_new2 >= 0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE"); + eigen_assert(beta_new2 >= 0.0 && "PRECONDITIONER IS NOT POSITIVE DEFINITE"); beta_new = sqrt(beta_new2); // compute beta_new v_new /= beta_new; // overwrite v_new for next iteration w_new /= beta_new; // overwrite w_new for next iteration @@ -107,28 +116,34 @@ namespace Eigen { s=beta_new/r1; // new sine // Update solution -// p_oold = p_old; - const VectorType p_oold(p_old); // NOT SURE IF CREATING p_oold EVERY ITERATION IS EFFICIENT + p_oold = p_old; +// const VectorType p_oold(p_old); // NOT SURE IF CREATING p_oold EVERY ITERATION IS EFFICIENT p_old = p; p.noalias()=(w-r2*p_old-r3*p_oold) /r1; // IS NOALIAS REQUIRED? x += beta_one*c*eta*p; + + /* Update the squared residual. Note that this is the estimated residual. + The real residual |Ax-b|^2 may be slightly larger */ residualNorm2 *= s*s; - if ( residualNorm2 < threshold2){ + if ( residualNorm2 < threshold2) + { break; } eta=-s*eta; // update eta iters++; // increment iteration number (for output purposes) } - tol_error = std::sqrt(residualNorm2 / rhsNorm2); // return error. Note that this is the estimated error. The real error |Ax-b|/|b| may be slightly larger + + /* Compute error. Note that this is the estimated error. The real + error |Ax-b|/|b| may be slightly larger */ + tol_error = std::sqrt(residualNorm2 / rhsNorm2); } } template< typename _MatrixType, int _UpLo=Lower, typename _Preconditioner = IdentityPreconditioner> -// typename _Preconditioner = IdentityPreconditioner > // preconditioner must be positive definite class MINRES; namespace internal { diff --git a/unsupported/test/minres.cpp b/unsupported/test/minres.cpp index fd12da548..81b762c37 100644 --- a/unsupported/test/minres.cpp +++ b/unsupported/test/minres.cpp @@ -1,8 +1,8 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Gael Guennebaud // Copyright (C) 2012 Giacomo Po +// Copyright (C) 2011 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -14,19 +14,29 @@ template void test_minres_T() { - MINRES, Lower, DiagonalPreconditioner > minres_colmajor_diag; - MINRES, Lower, IdentityPreconditioner > minres_colmajor_I; -// MINRES, Lower, IncompleteLUT > minres_colmajor_ilut; - //minres, SSORPreconditioner > minres_colmajor_ssor; + // Identity preconditioner + MINRES, Lower, IdentityPreconditioner > minres_colmajor_lower_I; + MINRES, Upper, IdentityPreconditioner > minres_colmajor_upper_I; + + // Diagonal preconditioner + MINRES, Lower, DiagonalPreconditioner > minres_colmajor_lower_diag; + MINRES, Upper, DiagonalPreconditioner > minres_colmajor_upper_diag; + + // call tests for SPD matrix + CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_lower_I) ); + CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_upper_I) ); + + CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_lower_diag) ); + CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_upper_diag) ); + + // TO DO: symmetric semi-definite matrix + // TO DO: symmetric indefinite matrix - CALL_SUBTEST( check_sparse_square_solving(minres_colmajor_diag) ); - CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_I) ); - // CALL_SUBTEST( check_sparse_square_solving(minres_colmajor_ilut) ); - //CALL_SUBTEST( check_sparse_square_solving(minres_colmajor_ssor) ); } void test_minres() { CALL_SUBTEST_1(test_minres_T()); -// CALL_SUBTEST_2(test_minres_T >()); +// CALL_SUBTEST_2(test_minres_T >()); + } From 72707a86641ad0bd4c4e5cc45c4b8ced64b499ef Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 21 Mar 2014 11:40:29 -0700 Subject: [PATCH 0372/1103] Made sure that EIGEN_ALIGN is defined when EIGEN_DONT_VECTORIZE is set to true to prevent build failures when vectorization is disabled. --- Eigen/src/Core/util/Macros.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 733d3403e..bfd6ba7de 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -82,6 +82,8 @@ #define EIGEN_ALIGN_BYTES 32 #endif #define EIGEN_ALIGN 1 +#else + #define EIGEN_ALIGN 0 #endif // EIGEN_ALIGN_STATICALLY is the true test whether we want to align arrays on the stack or not. It takes into account both the user choice to explicitly disable From 08f7b3221d58e480d8ede105bd70b09a2104c5fb Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 24 Mar 2014 09:52:45 -0700 Subject: [PATCH 0373/1103] Added proper support for AVX and FMA in the makefiles. --- CMakeLists.txt | 8 +++++++- cmake/EigenTesting.cmake | 10 ++++++++++ 2 files changed, 17 insertions(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 16a9b5bcb..fb13769f4 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -196,12 +196,18 @@ if(NOT MSVC) message(STATUS "Enabling SSE4.2 in tests/examples") endif() - option(EIGEN_TEST_AVX "Enable/Disable AVX in tests/examples" ON) + option(EIGEN_TEST_AVX "Enable/Disable AVX in tests/examples" OFF) if(EIGEN_TEST_AVX) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx") message(STATUS "Enabling AVX in tests/examples") endif() + option(EIGEN_TEST_FMA "Enable/Disable FMA in tests/examples" OFF) + if(EIGEN_TEST_FMA) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfma") + message(STATUS "Enabling FMA in tests/examples") + endif() + option(EIGEN_TEST_ALTIVEC "Enable/Disable AltiVec in tests/examples" OFF) if(EIGEN_TEST_ALTIVEC) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -maltivec -mabi=altivec") diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index d1ea4b9d2..9b9776894 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -270,6 +270,12 @@ macro(ei_testing_print_summary) message(STATUS "AVX: Using architecture defaults") endif() + if(EIGEN_TEST_FMA) + message(STATUS "FMA: ON") + else() + message(STATUS "FMA: Using architecture defaults") + endif() + if(EIGEN_TEST_ALTIVEC) message(STATUS "Altivec: ON") else() @@ -414,6 +420,10 @@ macro(ei_get_cxxflags VAR) set(${VAR} NEON) elseif(EIGEN_TEST_ALTIVEC) set(${VAR} ALVEC) + elseif(EIGEN_TEST_FMA) + set(${VAR} FMA) + elseif(EIGEN_TEST_AVX) + set(${VAR} AVX) elseif(EIGEN_TEST_SSE4_2) set(${VAR} SSE42) elseif(EIGEN_TEST_SSE4_1) From 7ae9b0805dbb218506a462c06263bd67f046366b Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 24 Mar 2014 13:33:40 -0700 Subject: [PATCH 0374/1103] Used AVX instructions to vectorize the predux_min, predux_min, predux_max, and predux_max packet primitives. --- Eigen/src/Core/arch/AVX/PacketMath.h | 30 ++++++++++------------------ 1 file changed, 10 insertions(+), 20 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index d1a134087..26cc996db 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -304,37 +304,27 @@ template<> EIGEN_STRONG_INLINE double predux_mul(const Packet4d& a) template<> EIGEN_STRONG_INLINE float predux_min(const Packet8f& a) { - float result = a[0]; - for (int i = 1; i < 8; ++i) { - if (a[i] < result) result = a[i]; - } - return result; + Packet8f tmp = _mm256_min_ps(a, _mm256_permute2f128_ps(a,a,1)); + tmp = _mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); + return pfirst(_mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); } template<> EIGEN_STRONG_INLINE double predux_min(const Packet4d& a) { - double result = a[0]; - for (int i = 1; i < 4; ++i) { - if (a[i] < result) result = a[i]; - } - return result; + Packet4d tmp = _mm256_min_pd(a, _mm256_permute2f128_pd(a,a,1)); + return pfirst(_mm256_min_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); } template<> EIGEN_STRONG_INLINE float predux_max(const Packet8f& a) { - float result = a[0]; - for (int i = 1; i < 8; ++i) { - if (a[i] > result) result = a[i]; - } - return result; + Packet8f tmp = _mm256_max_ps(a, _mm256_permute2f128_ps(a,a,1)); + tmp = _mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); + return pfirst(_mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); } template<> EIGEN_STRONG_INLINE double predux_max(const Packet4d& a) { - double result = a[0]; - for (int i = 1; i < 4; ++i) { - if (a[i] > result) result = a[i]; - } - return result; + Packet4d tmp = _mm256_max_pd(a, _mm256_permute2f128_pd(a,a,1)); + return pfirst(_mm256_max_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); } From 6bf3cc2732eebff73dd7fadcd8ac421f22381baf Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 25 Mar 2014 09:00:43 -0700 Subject: [PATCH 0375/1103] Use AVX instructions to vectorize pset1, pset1, preverse, and preverse --- Eigen/src/Core/arch/AVX/Complex.h | 40 +++++++++++++------------------ 1 file changed, 16 insertions(+), 24 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 9fb44ecab..17c32d79c 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -76,11 +76,9 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploadu(const std::complex EIGEN_STRONG_INLINE Packet4cf pset1(const std::complex& from) { - __m256 result; - for (int i = 0; i < 8; i+=2) { - result[i] = std::real(from); - result[i+1] = std::imag(from); - } + const float r = std::real(from); + const float i = std::imag(from); + const __m256 result = _mm256_set_ps(i, r, i, r, i, r, i, r); return Packet4cf(result); } @@ -108,15 +106,15 @@ template<> EIGEN_STRONG_INLINE std::complex pfirst(const Pack } template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { + __m128 low = _mm256_extractf128_ps(a.v, 0); + __m128 high = _mm256_extractf128_ps(a.v, 1); + __m128d lowd = _mm_castps_pd(low); + __m128d highd = _mm_castps_pd(high); + low = _mm_castpd_ps(_mm_shuffle_pd(lowd,lowd,0x1)); + high = _mm_castpd_ps(_mm_shuffle_pd(highd,highd,0x1)); __m256 result; - result[0] = a.v[6]; - result[1] = a.v[7]; - result[2] = a.v[4]; - result[3] = a.v[5]; - result[4] = a.v[2]; - result[5] = a.v[3]; - result[6] = a.v[0]; - result[7] = a.v[1]; + result = _mm256_insertf128_ps(result, low, 1); + result = _mm256_insertf128_ps(result, high, 0); return Packet4cf(result); } @@ -298,13 +296,11 @@ template<> EIGEN_STRONG_INLINE Packet2cd pload (const std::complex EIGEN_STRONG_INLINE Packet2cd ploadu(const std::complex* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cd(ploadu((const double*)from)); } -template<> EIGEN_STRONG_INLINE Packet2cd pset1(const std::complex& from) +template<> EIGEN_STRONG_INLINE Packet2cd pset1(const std::complex& from) { - __m256d result; - for (int i = 0; i < 4; i+=2) { - result[i] = std::real(from); - result[i+1] = std::imag(from); - } + const double r = std::real(from); + const double i = std::imag(from); + const __m256d result = _mm256_set_pd(i, r, i, r); return Packet2cd(result); } @@ -321,11 +317,7 @@ template<> EIGEN_STRONG_INLINE std::complex pfirst(const Pac } template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) { - __m256d result; - result[0] = a.v[2]; - result[1] = a.v[3]; - result[2] = a.v[0]; - result[3] = a.v[1]; + __m256d result = _mm256_permute2f128_pd(a.v, a.v, 1); return Packet2cd(result); } From b286a1e75c6bd451c27da8c5ebed0b0fb86dfc2a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 26 Mar 2014 16:46:36 +0100 Subject: [PATCH 0376/1103] add pbroadcast2/4 generic intrinsics --- Eigen/src/Core/GenericPacketMath.h | 34 ++++++++++++++++++++++++++++ Eigen/src/Core/arch/SSE/PacketMath.h | 32 ++++++++++++++++++++++++++ 2 files changed, 66 insertions(+) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 538ab53b2..d07541285 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -173,6 +173,40 @@ pset1(const typename unpacket_traits::type& a) { return a; } template EIGEN_DEVICE_FUNC inline Packet pload1(const typename unpacket_traits::type *a) { return pset1(*a); } +/** \internal equivalent to + * \code + * a0 = pload1(a+0); + * a1 = pload1(a+1); + * a2 = pload1(a+2); + * a3 = pload1(a+3); + * \endcode + * \sa pset1, pload1, ploaddup, pbroadcast2 + */ +template EIGEN_DEVICE_FUNC +inline void pbroadcast4(const typename unpacket_traits::type *a, + Packet& a0, Packet& a1, Packet& a2, Packet& a3) +{ + a0 = pload1(a+0); + a1 = pload1(a+1); + a2 = pload1(a+2); + a3 = pload1(a+3); +} + +/** \internal equivalent to + * \code + * a0 = pload1(a+0); + * a1 = pload1(a+1); + * \endcode + * \sa pset1, pload1, ploaddup, pbroadcast4 + */ +template EIGEN_DEVICE_FUNC +inline void pbroadcast2(const typename unpacket_traits::type *a, + Packet& a0, Packet& a1) +{ + a0 = pload1(a+0); + a1 = pload1(a+1); +} + /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */ template inline typename packet_traits::type plset(const Scalar& a) { return a; } diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 293fb83e4..9f81a4623 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -391,6 +391,38 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) #endif } +// with AVX, the default implementations based on pload1 are faster +#ifndef __AVX__ +template<> EIGEN_STRONG_INLINE void +pbroadcast4(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = pload(a); + a0 = vec4f_swizzle1(a3, 0,0,0,0); + a1 = vec4f_swizzle1(a3, 1,1,1,1); + a2 = vec4f_swizzle1(a3, 2,2,2,2); + a3 = vec4f_swizzle1(a3, 3,3,3,3); +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4(const double *a, + Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) +{ +#ifdef EIGEN_VECTORIZE_SSE3 + a0 = _mm_loaddup_pd(a+0); + a1 = _mm_loaddup_pd(a+1); + a2 = _mm_loaddup_pd(a+2); + a3 = _mm_loaddup_pd(a+3); +#else + a1 = pload(a); + a0 = vec2d_swizzle1(a1, 0,0); + a1 = vec2d_swizzle1(a1, 1,1); + a3 = pload(a+2); + a2 = vec2d_swizzle1(a3, 0,0); + a3 = vec2d_swizzle1(a3, 1,1); +#endif +} +#endif + EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs) { vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55)); From bc401eb6fa9c4c14c7fb32acfe70b304c1850283 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 26 Mar 2014 18:53:00 +0100 Subject: [PATCH 0377/1103] Implement new 1 packet x 8 gebp kernel --- Eigen/src/Core/arch/SSE/PacketMath.h | 6 +- .../Core/products/GeneralBlockPanelKernel.h | 679 +++++++----------- .../Core/products/SelfadjointMatrixMatrix.h | 48 +- 3 files changed, 297 insertions(+), 436 deletions(-) diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 9f81a4623..b1f7b7717 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -397,7 +397,7 @@ template<> EIGEN_STRONG_INLINE void pbroadcast4(const float *a, Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) { - a3 = pload(a); + a3 = ploadu(a); a0 = vec4f_swizzle1(a3, 0,0,0,0); a1 = vec4f_swizzle1(a3, 1,1,1,1); a2 = vec4f_swizzle1(a3, 2,2,2,2); @@ -413,10 +413,10 @@ pbroadcast4(const double *a, a2 = _mm_loaddup_pd(a+2); a3 = _mm_loaddup_pd(a+3); #else - a1 = pload(a); + a1 = ploadu(a); a0 = vec2d_swizzle1(a1, 0,0); a1 = vec2d_swizzle1(a1, 1,1); - a3 = pload(a+2); + a3 = ploadu(a+2); a2 = vec2d_swizzle1(a3, 0,0); a3 = vec2d_swizzle1(a3, 1,1); #endif diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index ba6fad246..8a398d912 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -161,11 +161,11 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, - // register block size along the N direction (must be either 2 or 4) - nr = NumberOfRegisters/4, + // register block size along the N direction (must be either 4 or 8) + nr = NumberOfRegisters/2, // register block size along the M direction (currently, this one cannot be modified) - mr = 2 * LhsPacketSize, + mr = LhsPacketSize, WorkSpaceFactor = nr * RhsPacketSize, @@ -187,6 +187,16 @@ public: { p = pset1(ResScalar(0)); } + + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + { + pbroadcast4(b, b0, b1, b2, b3); + } + + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) + { + pbroadcast2(b, b0, b1); + } EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const { @@ -230,8 +240,8 @@ public: ResPacketSize = Vectorizable ? packet_traits::size : 1, NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, - nr = NumberOfRegisters/4, - mr = 2 * LhsPacketSize, + nr = NumberOfRegisters/2, + mr = LhsPacketSize, WorkSpaceFactor = nr*RhsPacketSize, LhsProgress = LhsPacketSize, @@ -262,6 +272,16 @@ public: { dest = pload(a); } + + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + { + pbroadcast4(b, b0, b1, b2, b3); + } + + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) + { + pbroadcast2(b, b0, b1); + } EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const { @@ -304,8 +324,9 @@ public: RealPacketSize = Vectorizable ? packet_traits::size : 1, ResPacketSize = Vectorizable ? packet_traits::size : 1, - nr = 2, - mr = 2 * ResPacketSize, + // FIXME: should depend on NumberOfRegisters + nr = 4, + mr = ResPacketSize, WorkSpaceFactor = Vectorizable ? 2*nr*RealPacketSize : nr, LhsProgress = ResPacketSize, @@ -333,16 +354,37 @@ public: p.second = pset1(RealScalar(0)); } + // Scalar path EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const { dest = pset1(*b); } + // Vectorized path EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacket& dest) const { dest.first = pset1(real(*b)); dest.second = pset1(imag(*b)); } + + // linking error if instantiated without being optimized out: + void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); + + // Vectorized path + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacket& b0, DoublePacket& b1) + { + // FIXME not sure that's the best way to implement it! + loadRhs(b+0, b0); + loadRhs(b+1, b1); + } + + // Scalar path + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsScalar& b0, RhsScalar& b1) + { + // FIXME not sure that's the best way to implement it! + loadRhs(b+0, b0); + loadRhs(b+1, b1); + } // nothing special here EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const @@ -414,8 +456,9 @@ public: ResPacketSize = Vectorizable ? packet_traits::size : 1, NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, + // FIXME: should depend on NumberOfRegisters nr = 4, - mr = 2*ResPacketSize, + mr = ResPacketSize, WorkSpaceFactor = nr*RhsPacketSize, LhsProgress = ResPacketSize, @@ -441,6 +484,16 @@ public: { dest = pset1(*b); } + + // linking error if instantiated without being optimized out: + void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); + + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) + { + // FIXME not sure that's the best way to implement it! + b0 = pload1(b+0); + b1 = pload1(b+1); + } EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const { @@ -511,11 +564,9 @@ void gebp_kernel if(strideA==-1) strideA = depth; if(strideB==-1) strideB = depth; conj_helper cj; -// conj_helper pcj; Index packet_cols = (cols/nr) * nr; + // Here we assume that mr==LhsProgress const Index peeled_mc = (rows/mr)*mr; - // FIXME: - const Index peeled_mc2 = peeled_mc + (rows-peeled_mc >= LhsProgress ? LhsProgress : 0); const Index peeled_kc = (depth/4)*4; // loops on each micro vertical panel of rhs (depth x nr) @@ -527,144 +578,88 @@ void gebp_kernel for(Index i=0; i(alpha); - R0 = ploadu(r0); - R1 = ploadu(r1); - R2 = ploadu(r2); - R3 = ploadu(r3); - R4 = ploadu(r0 + ResPacketSize); - R5 = ploadu(r1 + ResPacketSize); - R6 = ploadu(r2 + ResPacketSize); + R0 = ploadu(r0+0*resStride); + R1 = ploadu(r0+1*resStride); + R2 = ploadu(r0+2*resStride); + R3 = ploadu(r0+3*resStride); + R4 = ploadu(r0+4*resStride); + R5 = ploadu(r0+5*resStride); + R6 = ploadu(r0+6*resStride); traits.acc(C0, alphav, R0); - pstoreu(r0, R0); - R0 = ploadu(r3 + ResPacketSize); + pstoreu(r0+0*resStride, R0); + R0 = ploadu(r0+7*resStride); traits.acc(C1, alphav, R1); traits.acc(C2, alphav, R2); @@ -739,232 +711,107 @@ EIGEN_ASM_COMMENT("mybegin4"); traits.acc(C6, alphav, R6); traits.acc(C7, alphav, R0); - pstoreu(r1, R1); - pstoreu(r2, R2); - pstoreu(r3, R3); - pstoreu(r0 + ResPacketSize, R4); - pstoreu(r1 + ResPacketSize, R5); - pstoreu(r2 + ResPacketSize, R6); - pstoreu(r3 + ResPacketSize, R0); + pstoreu(r0+1*resStride, R1); + pstoreu(r0+2*resStride, R2); + pstoreu(r0+3*resStride, R3); + pstoreu(r0+4*resStride, R4); + pstoreu(r0+5*resStride, R5); + pstoreu(r0+6*resStride, R6); + pstoreu(r0+7*resStride, R0); } - else + else // nr==4 { - ResPacket R0, R1, R4; + ResPacket R0, R1, R2; ResPacket alphav = pset1(alpha); - R0 = ploadu(r0); - R1 = ploadu(r1); - R4 = ploadu(r0 + ResPacketSize); + R0 = ploadu(r0+0*resStride); + R1 = ploadu(r0+1*resStride); + R2 = ploadu(r0+2*resStride); traits.acc(C0, alphav, R0); - pstoreu(r0, R0); - R0 = ploadu(r1 + ResPacketSize); + pstoreu(r0+0*resStride, R0); + R0 = ploadu(r0+3*resStride); + traits.acc(C1, alphav, R1); - traits.acc(C4, alphav, R4); - traits.acc(C5, alphav, R0); - pstoreu(r1, R1); - pstoreu(r0 + ResPacketSize, R4); - pstoreu(r1 + ResPacketSize, R0); + traits.acc(C2, alphav, R2); + traits.acc(C3, alphav, R0); + + pstoreu(r0+1*resStride, R1); + pstoreu(r0+2*resStride, R2); + pstoreu(r0+3*resStride, R0); } } - if(rows-peeled_mc>=LhsProgress) - { - Index i = peeled_mc; - const LhsScalar* blA = &blockA[i*strideA+offsetA*LhsProgress]; - prefetch(&blA[0]); - - // gets res block as register - AccPacket C0, C1, C2, C3; - traits.initAcc(C0); - traits.initAcc(C1); - if(nr==4) traits.initAcc(C2); - if(nr==4) traits.initAcc(C3); - - // performs "inner" product - const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; - for(Index k=0; k(alpha); - - ResScalar* r0 = &res[(j2+0)*resStride + i]; - ResScalar* r1 = r0 + resStride; - ResScalar* r2 = r1 + resStride; - ResScalar* r3 = r2 + resStride; - - R0 = ploadu(r0); - R1 = ploadu(r1); - if(nr==4) R2 = ploadu(r2); - if(nr==4) R3 = ploadu(r3); - - traits.acc(C0, alphav, R0); - traits.acc(C1, alphav, R1); - if(nr==4) traits.acc(C2, alphav, R2); - if(nr==4) traits.acc(C3, alphav, R3); - - pstoreu(r0, R0); - pstoreu(r1, R1); - if(nr==4) pstoreu(r2, R2); - if(nr==4) pstoreu(r3, R3); - } - for(Index i=peeled_mc2; i do the same but with nr==1 for(Index j2=packet_cols; j2(alpha); - - ResScalar* r0 = &res[(j2+0)*resStride + i]; - - R0 = ploadu(r0); - R4 = ploadu(r0+ResPacketSize); - - traits.acc(C0, alphav, R0); - traits.acc(C4, alphav, R4); - - pstoreu(r0, R0); - pstoreu(r0+ResPacketSize, R4); - } - if(rows-peeled_mc>=LhsProgress) - { - Index i = peeled_mc; - const LhsScalar* blA = &blockA[i*strideA+offsetA*LhsProgress]; - prefetch(&blA[0]); - AccPacket C0; traits.initAcc(C0); @@ -1025,19 +832,23 @@ EIGEN_ASM_COMMENT("mybegin4"); { LhsPacket A0; RhsPacket B_0; - traits.loadLhs(blA, A0); - traits.loadRhs(blB, B_0); - traits.madd(A0, B_0, C0, B_0); + RhsPacket T0; + + traits.loadLhs(&blA[0*LhsProgress], A0); + traits.loadRhs(&blB[0*RhsProgress], B_0); + traits.madd(A0,B_0,C0,T0); + blB += RhsProgress; blA += LhsProgress; } - + ResPacket R0; ResPacket alphav = pset1(alpha); - ResPacket R0 = ploadu(&res[(j2+0)*resStride + i]); + ResScalar* r0 = &res[(j2+0)*resStride + i]; + R0 = ploadu(r0); traits.acc(C0, alphav, R0); - pstoreu(&res[(j2+0)*resStride + i], R0); + pstoreu(r0, R0); } - for(Index i=peeled_mc2; i=depth && offset<=stride)); - eigen_assert( (StorageOrder==RowMajor) || ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) ); + eigen_assert( (StorageOrder==RowMajor) || ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) || (Pack1<=4) ); conj_if::IsComplex && Conjugate> cj; const_blas_data_mapper lhs(_lhs,lhsStride); Index count = 0; @@ -1104,15 +915,25 @@ EIGEN_DONT_INLINE void gemm_pack_lhs=1*PacketSize) A = ploadu(&lhs(i+0*PacketSize, k)); - if(Pack1>=2*PacketSize) B = ploadu(&lhs(i+1*PacketSize, k)); - if(Pack1>=3*PacketSize) C = ploadu(&lhs(i+2*PacketSize, k)); - if(Pack1>=4*PacketSize) D = ploadu(&lhs(i+3*PacketSize, k)); - if(Pack1>=1*PacketSize) { pstore(blockA+count, cj.pconj(A)); count+=PacketSize; } - if(Pack1>=2*PacketSize) { pstore(blockA+count, cj.pconj(B)); count+=PacketSize; } - if(Pack1>=3*PacketSize) { pstore(blockA+count, cj.pconj(C)); count+=PacketSize; } - if(Pack1>=4*PacketSize) { pstore(blockA+count, cj.pconj(D)); count+=PacketSize; } + if((Pack1%PacketSize)==0) + { + Packet A, B, C, D; + if(Pack1>=1*PacketSize) A = ploadu(&lhs(i+0*PacketSize, k)); + if(Pack1>=2*PacketSize) B = ploadu(&lhs(i+1*PacketSize, k)); + if(Pack1>=3*PacketSize) C = ploadu(&lhs(i+2*PacketSize, k)); + if(Pack1>=4*PacketSize) D = ploadu(&lhs(i+3*PacketSize, k)); + if(Pack1>=1*PacketSize) { pstore(blockA+count, cj.pconj(A)); count+=PacketSize; } + if(Pack1>=2*PacketSize) { pstore(blockA+count, cj.pconj(B)); count+=PacketSize; } + if(Pack1>=3*PacketSize) { pstore(blockA+count, cj.pconj(C)); count+=PacketSize; } + if(Pack1>=4*PacketSize) { pstore(blockA+count, cj.pconj(D)); count+=PacketSize; } + } + else + { + if(Pack1>=1) blockA[count++] = cj(lhs(i+0, k)); + if(Pack1>=2) blockA[count++] = cj(lhs(i+1, k)); + if(Pack1>=3) blockA[count++] = cj(lhs(i+2, k)); + if(Pack1>=4) blockA[count++] = cj(lhs(i+3, k)); + } } } else @@ -1191,12 +1012,20 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=4) blockB[count+2] = cj(b2[k]); + if(nr>=4) blockB[count+3] = cj(b3[k]); + if(nr>=8) blockB[count+4] = cj(b4[k]); + if(nr>=8) blockB[count+5] = cj(b5[k]); + if(nr>=8) blockB[count+6] = cj(b6[k]); + if(nr>=8) blockB[count+7] = cj(b7[k]); count += nr; } // skip what we have after @@ -1251,8 +1080,12 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=4) blockB[count+2] = cj(b0[2]); + if(nr>=4) blockB[count+3] = cj(b0[3]); + if(nr>=8) blockB[count+4] = cj(b0[4]); + if(nr>=8) blockB[count+5] = cj(b0[5]); + if(nr>=8) blockB[count+6] = cj(b0[6]); + if(nr>=8) blockB[count+7] = cj(b0[7]); count += nr; } } diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index 99cf9e0ae..d9fd9f556 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -63,7 +63,7 @@ struct symm_pack_lhs for(Index i=peeled_mc; i=4) { blockB[count+2] = rhs(k,j2+2); blockB[count+3] = rhs(k,j2+3); } + if (nr>=8) + { + blockB[count+4] = rhs(k,j2+4); + blockB[count+5] = rhs(k,j2+5); + blockB[count+6] = rhs(k,j2+6); + blockB[count+7] = rhs(k,j2+7); + } count += nr; } } @@ -109,11 +116,18 @@ struct symm_pack_rhs { blockB[count+0] = numext::conj(rhs(j2+0,k)); blockB[count+1] = numext::conj(rhs(j2+1,k)); - if (nr==4) + if (nr>=4) { blockB[count+2] = numext::conj(rhs(j2+2,k)); blockB[count+3] = numext::conj(rhs(j2+3,k)); } + if (nr>=8) + { + blockB[count+4] = numext::conj(rhs(j2+4,k)); + blockB[count+5] = numext::conj(rhs(j2+5,k)); + blockB[count+6] = numext::conj(rhs(j2+6,k)); + blockB[count+7] = numext::conj(rhs(j2+7,k)); + } count += nr; } // symmetric @@ -137,11 +151,18 @@ struct symm_pack_rhs { blockB[count+0] = rhs(k,j2+0); blockB[count+1] = rhs(k,j2+1); - if (nr==4) + if (nr>=4) { blockB[count+2] = rhs(k,j2+2); blockB[count+3] = rhs(k,j2+3); } + if (nr>=8) + { + blockB[count+4] = rhs(k,j2+4); + blockB[count+5] = rhs(k,j2+5); + blockB[count+6] = rhs(k,j2+6); + blockB[count+7] = rhs(k,j2+7); + } count += nr; } } @@ -153,11 +174,18 @@ struct symm_pack_rhs { blockB[count+0] = numext::conj(rhs(j2+0,k)); blockB[count+1] = numext::conj(rhs(j2+1,k)); - if (nr==4) + if (nr>=4) { blockB[count+2] = numext::conj(rhs(j2+2,k)); blockB[count+3] = numext::conj(rhs(j2+3,k)); } + if (nr>=8) + { + blockB[count+4] = numext::conj(rhs(j2+4,k)); + blockB[count+5] = numext::conj(rhs(j2+5,k)); + blockB[count+6] = numext::conj(rhs(j2+6,k)); + blockB[count+7] = numext::conj(rhs(j2+7,k)); + } count += nr; } } @@ -422,11 +450,11 @@ struct SelfadjointProductMatrix NumTraits::IsComplex && EIGEN_LOGICAL_XOR(RhsIsUpper,bool(RhsBlasTraits::NeedToConjugate)), internal::traits::Flags&RowMajorBit ? RowMajor : ColMajor> ::run( - lhs.rows(), rhs.cols(), // sizes - &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info - &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info - &dst.coeffRef(0,0), dst.outerStride(), // result info - actualAlpha // alpha + lhs.rows(), rhs.cols(), // sizes + &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info + &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info + &dst.coeffRef(0,0), dst.outerStride(), // result info + actualAlpha // alpha ); } }; From cf1a7bfbe146e232bd091a0ff2f186ea8803564e Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 26 Mar 2014 12:03:31 -0700 Subject: [PATCH 0378/1103] Used AVX instructions to vectorize the complex version of the pfirst and ploaddup packet primitives. Silenced a few compilation warnings. --- Eigen/src/Core/arch/AVX/Complex.h | 29 +++++++++++++++++++---------- 1 file changed, 19 insertions(+), 10 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 17c32d79c..a7a1b7fda 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -78,19 +78,21 @@ template<> EIGEN_STRONG_INLINE Packet4cf pset1(const std::complex EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex* from) { - __m256 result; - for (int i = 0; i < 2; ++i) { - result[4*i] = std::real(from[i]); - result[4*i+1] = std::imag(from[i]); - result[4*i+2] = std::real(from[i]); - result[4*i+3] = std::imag(from[i]); - } + // This should be optimized. + __m128 complex1 = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)from); + complex1 = _mm_movelh_ps(complex1, complex1); + __m128 complex2 = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from+1)); + complex2 = _mm_movelh_ps(complex2, complex2); + __m256 result = _mm256_setzero_ps(); + result = _mm256_insertf128_ps(result, complex1, 0); + result = _mm256_insertf128_ps(result, complex2, 1); return Packet4cf(result); } @@ -102,7 +104,10 @@ template<> EIGEN_STRONG_INLINE void prefetch >(const std::co template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet4cf& a) { - return std::complex(a.v[0], a.v[1]); + __m128 low = _mm256_extractf128_ps(a.v, 0); + std::complex res; + _mm_storel_pi((__m64*)&res, low); + return res; } template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { @@ -112,7 +117,7 @@ template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { __m128d highd = _mm_castps_pd(high); low = _mm_castpd_ps(_mm_shuffle_pd(lowd,lowd,0x1)); high = _mm_castpd_ps(_mm_shuffle_pd(highd,highd,0x1)); - __m256 result; + __m256 result = _mm256_setzero_ps(); result = _mm256_insertf128_ps(result, low, 1); result = _mm256_insertf128_ps(result, high, 0); return Packet4cf(result); @@ -300,6 +305,7 @@ template<> EIGEN_STRONG_INLINE Packet2cd pset1(const std::complex EIGEN_STRONG_INLINE void prefetch >(const std::c template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cd& a) { - return std::complex(a.v[0],a.v[1]); + __m128d low = _mm256_extractf128_pd(a.v, 0); + EIGEN_ALIGN16 double res[2]; + _mm_store_pd(res, low); + return std::complex(res[0],res[1]); } template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) { From a078f442a382be64cc22f315ad300c353891a814 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 26 Mar 2014 15:11:18 -0700 Subject: [PATCH 0379/1103] Vectorized the multiplication and division of complex numbers using AVX instructions. --- Eigen/src/Core/arch/AVX/Complex.h | 36 +++++++++++++------------------ 1 file changed, 15 insertions(+), 21 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index a7a1b7fda..f63dfe6cf 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -221,13 +221,11 @@ template<> struct conj_helper template<> EIGEN_STRONG_INLINE Packet4cf pdiv(const Packet4cf& a, const Packet4cf& b) { - Packet4cf res; - for (int i = 0; i < 8; i+=2) { - std::complex result = std::complex(a.v[i], a.v[i+1]) / std::complex(b.v[i], b.v[i+1]); - res.v[i] = std::real(result); - res.v[i+1] = std::imag(result); - } - return res; + Packet4cf num = pmul(a, pconj(b)); + __m256 tmp = _mm256_mul_ps(b.v, b.v); + __m256 tmp2 = _mm256_shuffle_ps(tmp,tmp,0xB1); + __m256 denom = _mm256_add_ps(tmp, tmp2); + return Packet4cf(_mm256_div_ps(num.v, denom)); } template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip(const Packet4cf& x) @@ -282,13 +280,12 @@ template<> EIGEN_STRONG_INLINE Packet2cd pconj(const Packet2cd& a) template<> EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) { - __m256d tmp1 = _mm256_mul_pd(_mm256_permute_pd(a.v, 0), b.v); - // FIXME: _mm256_permute_pd(b.v, _MM_SHUFFLE2(1,0) won't work as expected, figure out an alternative. - __m256d op = {b.v[1], b.v[0], b.v[3], b.v[2]}; - __m256d tmp2 = _mm256_mul_pd(_mm256_permute_pd(a.v, 15), op); - __m256d result = _mm256_addsub_pd(tmp1, tmp2); - - return Packet2cd(result); + __m256d tmp1 = _mm256_shuffle_pd(a.v,a.v,0x0); + __m256d even = _mm256_mul_pd(tmp1, b.v); + __m256d tmp2 = _mm256_shuffle_pd(a.v,a.v,0xF); + __m256d tmp3 = _mm256_shuffle_pd(b.v,b.v,0x5); + __m256d odd = _mm256_mul_pd(tmp2, tmp3); + return Packet2cd(_mm256_addsub_pd(even, odd)); } template<> EIGEN_STRONG_INLINE Packet2cd pand (const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); } @@ -418,13 +415,10 @@ template<> struct conj_helper template<> EIGEN_STRONG_INLINE Packet2cd pdiv(const Packet2cd& a, const Packet2cd& b) { - Packet2cd res; - for (int i = 0; i < 4; i+=2) { - std::complex result = std::complex(a.v[i], a.v[i+1]) / std::complex(b.v[i], b.v[i+1]); - res.v[i] = std::real(result); - res.v[i+1] = std::imag(result); - } - return res; + Packet2cd num = pmul(a, pconj(b)); + __m256d tmp = _mm256_mul_pd(b.v, b.v); + __m256d denom = _mm256_hadd_pd(tmp, tmp); + return Packet2cd(_mm256_div_pd(num.v, denom)); } template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip(const Packet2cd& x) From 8be011e7765cf5d30cad86edd55850b0b1277741 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 26 Mar 2014 23:14:44 +0100 Subject: [PATCH 0380/1103] Remove remaining bits of the dead working buffer --- Eigen/src/Core/products/GeneralMatrixMatrix.h | 16 +--------------- 1 file changed, 1 insertion(+), 15 deletions(-) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 6776cf5a8..19991fa3f 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -278,13 +278,11 @@ class gemm_blocking_space Traits; enum { SizeA = ActualRows * MaxDepth, - SizeB = ActualCols * MaxDepth, - SizeW = MaxDepth * Traits::WorkSpaceFactor + SizeB = ActualCols * MaxDepth }; EIGEN_ALIGN16 LhsScalar m_staticA[SizeA]; EIGEN_ALIGN16 RhsScalar m_staticB[SizeB]; - EIGEN_ALIGN16 RhsScalar m_staticW[SizeW]; public: @@ -295,12 +293,10 @@ class gemm_blocking_spacem_kc = MaxDepth; this->m_blockA = m_staticA; this->m_blockB = m_staticB; - this->m_blockW = m_staticW; } inline void allocateA() {} inline void allocateB() {} - inline void allocateW() {} inline void allocateAll() {} }; @@ -319,7 +315,6 @@ class gemm_blocking_space(this->m_kc, this->m_mc, this->m_nc); m_sizeA = this->m_mc * this->m_kc; m_sizeB = this->m_kc * this->m_nc; - m_sizeW = this->m_kc*Traits::WorkSpaceFactor; } void allocateA() @@ -347,24 +341,16 @@ class gemm_blocking_spacem_blockB = aligned_new(m_sizeB); } - void allocateW() - { - if(this->m_blockW==0) - this->m_blockW = aligned_new(m_sizeW); - } - void allocateAll() { allocateA(); allocateB(); - allocateW(); } ~gemm_blocking_space() { aligned_delete(this->m_blockA, m_sizeA); aligned_delete(this->m_blockB, m_sizeB); - aligned_delete(this->m_blockW, m_sizeW); } }; From f0a4c9d5abe94703d8b4e5ec49cba20df014d0ca Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 26 Mar 2014 23:22:36 +0100 Subject: [PATCH 0381/1103] Update gebp kernel to process a panle of 4 columns at once for the remaining ones. --- .../Core/products/GeneralBlockPanelKernel.h | 464 +++++++++++------- .../Core/products/SelfadjointMatrixMatrix.h | 153 +++--- 2 files changed, 380 insertions(+), 237 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 8a398d912..2b520383b 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -167,8 +167,6 @@ public: // register block size along the M direction (currently, this one cannot be modified) mr = LhsPacketSize, - WorkSpaceFactor = nr * RhsPacketSize, - LhsProgress = LhsPacketSize, RhsProgress = 1 }; @@ -242,7 +240,6 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, nr = NumberOfRegisters/2, mr = LhsPacketSize, - WorkSpaceFactor = nr*RhsPacketSize, LhsProgress = LhsPacketSize, RhsProgress = 1 @@ -327,7 +324,6 @@ public: // FIXME: should depend on NumberOfRegisters nr = 4, mr = ResPacketSize, - WorkSpaceFactor = Vectorizable ? 2*nr*RealPacketSize : nr, LhsProgress = ResPacketSize, RhsProgress = 1 @@ -459,7 +455,6 @@ public: // FIXME: should depend on NumberOfRegisters nr = 4, mr = ResPacketSize, - WorkSpaceFactor = nr*RhsPacketSize, LhsProgress = ResPacketSize, RhsProgress = 1 @@ -564,63 +559,45 @@ void gebp_kernel if(strideA==-1) strideA = depth; if(strideB==-1) strideB = depth; conj_helper cj; - Index packet_cols = (cols/nr) * nr; + Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; + Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; // Here we assume that mr==LhsProgress const Index peeled_mc = (rows/mr)*mr; const Index peeled_kc = (depth/4)*4; // loops on each micro vertical panel of rhs (depth x nr) - for(Index j2=0; j2=8) { - // loops on each largest micro horizontal panel of lhs (mr x depth) - // => we select a mr x nr micro block of res which is entirely - // stored into mr/packet_size x nr registers. - for(Index i=0; i we select a mr x nr micro block of res which is entirely + // stored into mr/packet_size x nr registers. + for(Index i=0; i EIGEN_GEBGP_ONESTEP8(1,A0,A1); EIGEN_GEBGP_ONESTEP8(2,A1,A0); EIGEN_GEBGP_ONESTEP8(3,A0,A1); - } - blB += 4*nr*RhsProgress; - blA += 4*mr; - } - // process remaining peeled loop - for(Index k=peeled_kc; k(alpha); @@ -718,60 +683,20 @@ void gebp_kernel pstoreu(r0+5*resStride, R5); pstoreu(r0+6*resStride, R6); pstoreu(r0+7*resStride, R0); - } - else // nr==4 - { - ResPacket R0, R1, R2; - ResPacket alphav = pset1(alpha); - - R0 = ploadu(r0+0*resStride); - R1 = ploadu(r0+1*resStride); - R2 = ploadu(r0+2*resStride); - traits.acc(C0, alphav, R0); - pstoreu(r0+0*resStride, R0); - R0 = ploadu(r0+3*resStride); - - traits.acc(C1, alphav, R1); - traits.acc(C2, alphav, R2); - traits.acc(C3, alphav, R0); - pstoreu(r0+1*resStride, R1); - pstoreu(r0+2*resStride, R2); - pstoreu(r0+3*resStride, R0); } - } - - for(Index i=peeled_mc; i B_1 = blB[7]; MADD(cj,A0,B_0,C6, B_0); MADD(cj,A0,B_1,C7, B_1); - } - blB += nr; + blB += 8; + } + res[(j2+0)*resStride + i] += alpha*C0; + res[(j2+1)*resStride + i] += alpha*C1; + res[(j2+2)*resStride + i] += alpha*C2; + res[(j2+3)*resStride + i] += alpha*C3; + res[(j2+4)*resStride + i] += alpha*C4; + res[(j2+5)*resStride + i] += alpha*C5; + res[(j2+6)*resStride + i] += alpha*C6; + res[(j2+7)*resStride + i] += alpha*C7; + } + } + } + + // Second pass using depth x 4 panels + // If nr==8, then we have at most one such panel + if(nr>=4) + { + for(Index j2=packet_cols8; j2 we select a mr x 4 micro block of res which is entirely + // stored into mr/packet_size x 4 registers. + for(Index i=0; i(alpha); + + R0 = ploadu(r0+0*resStride); + R1 = ploadu(r0+1*resStride); + R2 = ploadu(r0+2*resStride); + traits.acc(C0, alphav, R0); + pstoreu(r0+0*resStride, R0); + R0 = ploadu(r0+3*resStride); + + traits.acc(C1, alphav, R1); + traits.acc(C2, alphav, R2); + traits.acc(C3, alphav, R0); + + pstoreu(r0+1*resStride, R1); + pstoreu(r0+2*resStride, R2); + pstoreu(r0+3*resStride, R0); + } + + for(Index i=peeled_mc; i do the same but with nr==1 - for(Index j2=packet_cols; j2=depth && offset<=stride)); conj_if::IsComplex && Conjugate> cj; - Index packet_cols = (cols/nr) * nr; + Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; + Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; Index count = 0; - for(Index j2=0; j2=8) { - // skip what we have before - if(PanelMode) count += nr * offset; - const Scalar* b0 = &rhs[(j2+0)*rhsStride]; - const Scalar* b1 = &rhs[(j2+1)*rhsStride]; - const Scalar* b2 = &rhs[(j2+2)*rhsStride]; - const Scalar* b3 = &rhs[(j2+3)*rhsStride]; - const Scalar* b4 = &rhs[(j2+4)*rhsStride]; - const Scalar* b5 = &rhs[(j2+5)*rhsStride]; - const Scalar* b6 = &rhs[(j2+6)*rhsStride]; - const Scalar* b7 = &rhs[(j2+7)*rhsStride]; - for(Index k=0; k=4) blockB[count+2] = cj(b2[k]); - if(nr>=4) blockB[count+3] = cj(b3[k]); - if(nr>=8) blockB[count+4] = cj(b4[k]); - if(nr>=8) blockB[count+5] = cj(b5[k]); - if(nr>=8) blockB[count+6] = cj(b6[k]); - if(nr>=8) blockB[count+7] = cj(b7[k]); - count += nr; + // skip what we have before + if(PanelMode) count += 8 * offset; + const Scalar* b0 = &rhs[(j2+0)*rhsStride]; + const Scalar* b1 = &rhs[(j2+1)*rhsStride]; + const Scalar* b2 = &rhs[(j2+2)*rhsStride]; + const Scalar* b3 = &rhs[(j2+3)*rhsStride]; + const Scalar* b4 = &rhs[(j2+4)*rhsStride]; + const Scalar* b5 = &rhs[(j2+5)*rhsStride]; + const Scalar* b6 = &rhs[(j2+6)*rhsStride]; + const Scalar* b7 = &rhs[(j2+7)*rhsStride]; + for(Index k=0; k=4) + { + for(Index j2=packet_cols8; j2=depth && offset<=stride)); conj_if::IsComplex && Conjugate> cj; - Index packet_cols = (cols/nr) * nr; + Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; + Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; Index count = 0; - for(Index j2=0; j2=8) { - // skip what we have before - if(PanelMode) count += nr * offset; - for(Index k=0; k(&rhs[k*rhsStride + j2]); - pstoreu(blockB+count, cj.pconj(A)); - count += PacketSize; - } else { - const Scalar* b0 = &rhs[k*rhsStride + j2]; - blockB[count+0] = cj(b0[0]); - blockB[count+1] = cj(b0[1]); - if(nr>=4) blockB[count+2] = cj(b0[2]); - if(nr>=4) blockB[count+3] = cj(b0[3]); - if(nr>=8) blockB[count+4] = cj(b0[4]); - if(nr>=8) blockB[count+5] = cj(b0[5]); - if(nr>=8) blockB[count+6] = cj(b0[6]); - if(nr>=8) blockB[count+7] = cj(b0[7]); - count += nr; + // skip what we have before + if(PanelMode) count += 8 * offset; + for(Index k=0; k(&rhs[k*rhsStride + j2]); + pstoreu(blockB+count, cj.pconj(A)); + count += PacketSize; + } else { + const Scalar* b0 = &rhs[k*rhsStride + j2]; + blockB[count+0] = cj(b0[0]); + blockB[count+1] = cj(b0[1]); + blockB[count+2] = cj(b0[2]); + blockB[count+3] = cj(b0[3]); + blockB[count+4] = cj(b0[4]); + blockB[count+5] = cj(b0[5]); + blockB[count+6] = cj(b0[6]); + blockB[count+7] = cj(b0[7]); + count += 8; + } } + // skip what we have after + if(PanelMode) count += 8 * (stride-offset-depth); + } + } + if(nr>=4) + { + for(Index j2=packet_cols8; j2(&rhs[k*rhsStride + j2]); + pstoreu(blockB+count, cj.pconj(A)); + count += PacketSize; + } else { + const Scalar* b0 = &rhs[k*rhsStride + j2]; + blockB[count+0] = cj(b0[0]); + blockB[count+1] = cj(b0[1]); + blockB[count+2] = cj(b0[2]); + blockB[count+3] = cj(b0[3]); + count += 4; + } + } + // skip what we have after + if(PanelMode) count += 4 * (stride-offset-depth); } - // skip what we have after - if(PanelMode) count += nr * (stride-offset-depth); } // copy the remaining columns one at a time (nr==1) - for(Index j2=packet_cols; j2 rhs(_rhs,rhsStride); - Index packet_cols = (cols/nr)*nr; + Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; + Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; // first part: normal case for(Index j2=0; j2=8 ? (std::min)(k2+rows,packet_cols8) : k2; + if(nr>=8) { - // again we can split vertically in three different parts (transpose, symmetric, normal) - // transpose - for(Index k=k2; k=4) + // again we can split vertically in three different parts (transpose, symmetric, normal) + // transpose + for(Index k=k2; k=8) - { blockB[count+4] = numext::conj(rhs(j2+4,k)); blockB[count+5] = numext::conj(rhs(j2+5,k)); blockB[count+6] = numext::conj(rhs(j2+6,k)); blockB[count+7] = numext::conj(rhs(j2+7,k)); + count += 8; } - count += nr; - } - // symmetric - Index h = 0; - for(Index k=j2; k=4) + // symmetric + Index h = 0; + for(Index k=j2; k=8) - { blockB[count+4] = rhs(k,j2+4); blockB[count+5] = rhs(k,j2+5); blockB[count+6] = rhs(k,j2+6); blockB[count+7] = rhs(k,j2+7); + count += 8; + } + } + } + if(nr>=4) + { + for(Index j2=end8; j2<(std::min)(k2+rows,packet_cols4); j2+=4) + { + // again we can split vertically in three different parts (transpose, symmetric, normal) + // transpose + for(Index k=k2; k=8) { - for(Index k=k2; k=4) + for(Index k=k2; k=8) - { blockB[count+4] = numext::conj(rhs(j2+4,k)); blockB[count+5] = numext::conj(rhs(j2+5,k)); blockB[count+6] = numext::conj(rhs(j2+6,k)); blockB[count+7] = numext::conj(rhs(j2+7,k)); + count += 8; + } + } + } + if(nr>=4) + { + for(Index j2=(std::max)(packet_cols8,k2+rows); j2 the same with nr==1) - for(Index j2=packet_cols; j2 gebp_kernel; symm_pack_lhs pack_lhs; @@ -376,11 +420,10 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix(kc, mc, nc); - std::size_t sizeW = kc*Traits::WorkSpaceFactor; - std::size_t sizeB = sizeW + kc*cols; + std::size_t sizeB = kc*cols; ei_declare_aligned_stack_constructed_variable(Scalar, blockA, kc*mc, 0); ei_declare_aligned_stack_constructed_variable(Scalar, allocatedBlockB, sizeB, 0); - Scalar* blockB = allocatedBlockB + sizeW; + Scalar* blockB = allocatedBlockB; gebp_kernel gebp_kernel; gemm_pack_lhs pack_lhs; From e45a6bed4548fb2821efd3d78471bf1bd2a7d4ef Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 26 Mar 2014 15:58:13 -0700 Subject: [PATCH 0382/1103] Specialized the pload1 packet primitive for Packet8f and Packet4d in order to take advantage of the vbroadcastss and vbroadcastsd instructions whenever possible. --- Eigen/src/Core/arch/AVX/PacketMath.h | 3 +++ 1 file changed, 3 insertions(+) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 26cc996db..98f8c5416 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -87,6 +87,9 @@ template<> EIGEN_STRONG_INLINE Packet8f pset1(const float& from) { re template<> EIGEN_STRONG_INLINE Packet4d pset1(const double& from) { return _mm256_set1_pd(from); } template<> EIGEN_STRONG_INLINE Packet8i pset1(const int& from) { return _mm256_set1_epi32(from); } +template<> EIGEN_STRONG_INLINE Packet8f pload1(const float* from) { return _mm256_broadcast_ss(from); } +template<> EIGEN_STRONG_INLINE Packet4d pload1(const double* from) { return _mm256_broadcast_sd(from); } + template<> EIGEN_STRONG_INLINE Packet8f plset(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); } template<> EIGEN_STRONG_INLINE Packet4d plset(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); } From 14bc4b9704b7e347ffcfe3c52588790e27e5118b Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 26 Mar 2014 17:35:18 -0700 Subject: [PATCH 0383/1103] Made sure that the version of gemm_pack_rhs specialized for row major matrices is vectorized when nr == 2*PacketSize (which is the case for SSE when compiling in 64bit mode). --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 569cfea71..d17752489 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -1091,7 +1091,11 @@ EIGEN_DONT_INLINE void gemm_pack_rhs(&rhs[k*rhsStride + j2]); pstoreu(blockB+count, cj.pconj(A)); - count += PacketSize; + } else if (nr == 2*PacketSize) { + Packet A = ploadu(&rhs[k*rhsStride + j2]); + Packet B = ploadu(&rhs[k*rhsStride + j2 + PacketSize]); + pstoreu(blockB+count, cj.pconj(A)); + pstoreu(blockB+count+PacketSize, cj.pconj(B)); } else { const Scalar* b0 = &rhs[k*rhsStride + j2]; blockB[count+0] = cj(b0[0]); @@ -1102,8 +1106,8 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=8) blockB[count+5] = cj(b0[5]); if(nr>=8) blockB[count+6] = cj(b0[6]); if(nr>=8) blockB[count+7] = cj(b0[7]); - count += nr; } + count += nr; } // skip what we have after if(PanelMode) count += nr * (stride-offset-depth); From ba3457cab2349dd03585435571ff1a8f90cf7403 Mon Sep 17 00:00:00 2001 From: Abhijit Kundu Date: Wed, 26 Mar 2014 22:02:48 -0400 Subject: [PATCH 0384/1103] Fixed compilation error due to obsolete internal::abs and internal::sqrt function calls --- demos/opengl/quaternion_demo.cpp | 2 +- demos/opengl/trackball.cpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/demos/opengl/quaternion_demo.cpp b/demos/opengl/quaternion_demo.cpp index 04165619b..dd323a4c9 100644 --- a/demos/opengl/quaternion_demo.cpp +++ b/demos/opengl/quaternion_demo.cpp @@ -234,7 +234,7 @@ void RenderingWidget::drawScene() gpu.drawVector(Vector3f::Zero(), length*Vector3f::UnitZ(), Color(0,0,1,1)); // draw the fractal object - float sqrt3 = internal::sqrt(3.); + float sqrt3 = std::sqrt(3.); glLightfv(GL_LIGHT0, GL_AMBIENT, Vector4f(0.5,0.5,0.5,1).data()); glLightfv(GL_LIGHT0, GL_DIFFUSE, Vector4f(0.5,1,0.5,1).data()); glLightfv(GL_LIGHT0, GL_SPECULAR, Vector4f(1,1,1,1).data()); diff --git a/demos/opengl/trackball.cpp b/demos/opengl/trackball.cpp index 77ac790c8..7c2da8e96 100644 --- a/demos/opengl/trackball.cpp +++ b/demos/opengl/trackball.cpp @@ -23,7 +23,7 @@ void Trackball::track(const Vector2i& point2D) { Vector3f axis = mLastPoint3D.cross(newPoint3D).normalized(); float cos_angle = mLastPoint3D.dot(newPoint3D); - if ( internal::abs(cos_angle) < 1.0 ) + if ( std::abs(cos_angle) < 1.0 ) { float angle = 2. * acos(cos_angle); if (mMode==Around) From a419cea4a0ff545f3221020119d5eb6ab4cd3e48 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 26 Mar 2014 19:03:07 -0700 Subject: [PATCH 0385/1103] Created the ptranspose packet primitive that can transpose an array of N packets, where N is the number of words in each packet. This primitive will be used to complete the vectorization of the gemm_pack_lhs and gemm_pack_rhs functions. Implemented the primitive using SSE instructions. --- Eigen/src/Core/GenericPacketMath.h | 15 ++++++++++++++- Eigen/src/Core/arch/SSE/Complex.h | 10 ++++++++++ Eigen/src/Core/arch/SSE/PacketMath.h | 25 +++++++++++++++++++++++++ test/packetmath.cpp | 12 ++++++++++++ 4 files changed, 61 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index d07541285..f9ddf4718 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -386,9 +386,22 @@ template<> inline std::complex pmul(const std::complex& a, const #endif + +/*************************************************************************** + * Kernel, that is a collection of N packets where N is the number of words + * in the packet. +***************************************************************************/ +template struct Kernel { + Packet packet[unpacket_traits::size]; +}; + +template EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& /*kernel*/) { + // Nothing to do in the scalar case, i.e. a 1x1 matrix. +} + } // end namespace internal } // end namespace Eigen #endif // EIGEN_GENERIC_PACKET_MATH_H - diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 91bba5e38..2dce66819 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -435,6 +435,16 @@ EIGEN_STRONG_INLINE Packet1cd pcplxflip/**/(const Packet1cd& x) return Packet1cd(preverse(x.v)); } +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + __m128d w1 = _mm_castps_pd(kernel.packet[0].v); + __m128d w2 = _mm_castps_pd(kernel.packet[1].v); + + __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2)); + kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2)); + kernel.packet[1].v = tmp; +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 9d8faa7d6..937f63f88 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -707,6 +707,31 @@ struct palign_impl }; #endif +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]); +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]); + kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]); + kernel.packet[1] = tmp; +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]); + __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]); + __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]); + __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]); + + kernel.packet[0] = _mm_unpacklo_epi64(T0, T1); + kernel.packet[1] = _mm_unpackhi_epi64(T0, T1); + kernel.packet[2] = _mm_unpacklo_epi64(T2, T3); + kernel.packet[3] = _mm_unpackhi_epi64(T2, T3); +} + } // end namespace internal } // end namespace Eigen diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 5a680d1ee..735af7017 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -208,6 +208,18 @@ template void packetmath() ref[i] = data1[PacketSize-i-1]; internal::pstore(data2, internal::preverse(internal::pload(data1))); VERIFY(areApprox(ref, data2, PacketSize) && "internal::preverse"); + + internal::Kernel kernel; + for (int i=0; i(data1+i*PacketSize); + } + ptranspose(kernel); + for (int i=0; i void packetmath_real() From 9ce0d785137c4034f19193c6a6781bbbb38e29fb Mon Sep 17 00:00:00 2001 From: Mark Borgerding Date: Wed, 26 Mar 2014 22:26:07 -0400 Subject: [PATCH 0386/1103] immintrin.h did not come until intel version 11 --- Eigen/Core | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/Core b/Eigen/Core index 468ae0c76..412409497 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -123,7 +123,7 @@ extern "C" { // In theory we should only include immintrin.h and not the other *mmintrin.h header files directly. // Doing so triggers some issues with ICC. However old gcc versions seems to not have this file, thus: - #ifdef __INTEL_COMPILER + #if defined(__INTEL_COMPILER) && __INTEL_COMPILER >= 1110 #include #else #include From fb03b56647e299428c03e3b1c3a01f3318e1af7e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 27 Mar 2014 11:38:35 +0100 Subject: [PATCH 0387/1103] Fix warning --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 2b520383b..cba76edfe 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -594,8 +594,9 @@ void gebp_kernel // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; - LhsPacket A0, A1; + LhsPacket A0; // uncomment for register prefetching + // LhsPacket A1; // traits.loadLhs(blA, A0); for(Index k=0; k // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*4]; - LhsPacket A0, A1; + LhsPacket A0; // uncomment for register prefetching + // LhsPacket A1; // traits.loadLhs(blA, A0); for(Index k=0; k Date: Thu, 27 Mar 2014 14:47:00 +0100 Subject: [PATCH 0388/1103] Implement pcplflip, palign, predux and the likes from AVC/complexes --- Eigen/Core | 3 +- Eigen/src/Core/arch/AVX/Complex.h | 81 +++++++++++-------------------- Eigen/src/Core/arch/SSE/Complex.h | 8 +++ 3 files changed, 37 insertions(+), 55 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index dbe68586d..39cae5d40 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -303,9 +303,10 @@ using std::ptrdiff_t; #if defined EIGEN_VECTORIZE_AVX // Use AVX for floats and doubles, SSE for integers + #include "src/Core/arch/SSE/PacketMath.h" + #include "src/Core/arch/SSE/Complex.h" #include "src/Core/arch/AVX/PacketMath.h" #include "src/Core/arch/AVX/Complex.h" - #include "src/Core/arch/SSE/PacketMath.h" // For integers #elif defined EIGEN_VECTORIZE_SSE #include "src/Core/arch/SSE/PacketMath.h" #include "src/Core/arch/SSE/MathFunctions.h" diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index f63dfe6cf..4f4f6eb52 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -99,8 +99,6 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } - template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet4cf& a) { @@ -125,28 +123,29 @@ template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { template<> EIGEN_STRONG_INLINE std::complex predux(const Packet4cf& a) { - return std::complex(a.v[0]+a.v[2]+a.v[4]+a.v[6], a.v[1]+a.v[3]+a.v[5]+a.v[7]); + return predux(padd(Packet2cf(_mm256_extractf128_ps(a.v,0)), + Packet2cf(_mm256_extractf128_ps(a.v,1)))); } template<> EIGEN_STRONG_INLINE Packet4cf preduxp(const Packet4cf* vecs) { - __m256 result = _mm256_setzero_ps(); - for (int i = 0; i < 4; ++i) { - for (int j = 0; j < 8; j+=2) { - result[2*i] += vecs[i].v[j]; - result[2*i+1] += vecs[i].v[j+1]; - } - } - return Packet4cf(result); + Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0)); + Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0)); + t0 = _mm256_hadd_ps(t0,t1); + Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0)); + Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0)); + t2 = _mm256_hadd_ps(t2,t3); + + t1 = _mm256_permute2f128_ps(t0,t2, 0 + (2<<4)); + t3 = _mm256_permute2f128_ps(t0,t2, 1 + (3<<4)); + + return Packet4cf(_mm256_add_ps(t1,t3)); } template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet4cf& a) { - std::complex result(a.v[0], a.v[1]); - for (int i = 2; i < 8; i+=2) { - result *= std::complex(a.v[i], a.v[i+1]); - } - return result; + return predux_mul(pmul(Packet2cf(_mm256_extractf128_ps(a.v, 0)), + Packet2cf(_mm256_extractf128_ps(a.v, 1)))); } template @@ -155,16 +154,7 @@ struct palign_impl static EIGEN_STRONG_INLINE void run(Packet4cf& first, const Packet4cf& second) { if (Offset==0) return; - for (int i = 0; i < 4-Offset; ++i) - { - first.v[2*i] = first.v[2*(i+Offset)]; - first.v[2*i+1] = first.v[2*(i+Offset)+1]; - } - for (int i = 4-Offset; i < 4; ++i) - { - first.v[2*i] = second.v[2*(i-4+Offset)]; - first.v[2*i+1] = second.v[2*(i-4+Offset)+1]; - } + palign_impl::run(first.v, second.v); } }; @@ -230,12 +220,7 @@ template<> EIGEN_STRONG_INLINE Packet4cf pdiv(const Packet4cf& a, con template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip(const Packet4cf& x) { - Packet4cf res; - for (int i = 0; i < 8; i+=2) { - res.v[i] = x.v[i+1]; - res.v[i+1] = x.v[i]; - } - return res; + return Packet4cf(_mm256_shuffle_ps(x.v, x.v, _MM_SHUFFLE(2, 3, 0 ,1))); } //---------- double ---------- @@ -312,8 +297,6 @@ template<> EIGEN_STRONG_INLINE Packet2cd ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } - template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cd& a) { __m128d low = _mm256_extractf128_pd(a.v, 0); @@ -329,24 +312,22 @@ template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) { template<> EIGEN_STRONG_INLINE std::complex predux(const Packet2cd& a) { - return std::complex(a.v[0]+a.v[2], a.v[1]+a.v[3]); + return predux(padd(Packet1cd(_mm256_extractf128_pd(a.v,0)), + Packet1cd(_mm256_extractf128_pd(a.v,1)))); } template<> EIGEN_STRONG_INLINE Packet2cd preduxp(const Packet2cd* vecs) { - __m256d result = _mm256_setzero_pd(); - for (int i = 0; i < 2; ++i) { - for (int j = 0; j < 4; j+=2) { - result[2*i] += vecs[i].v[j]; - result[2*i+1] += vecs[i].v[j+1]; - } - } - return Packet2cd(result); + Packet4d t0 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 0 + (2<<4)); + Packet4d t1 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 1 + (3<<4)); + + return Packet2cd(_mm256_add_pd(t0,t1)); } template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet2cd& a) { - return std::complex(a.v[0], a.v[1]) * std::complex(a.v[2], a.v[3]); + return predux(pmul(Packet1cd(_mm256_extractf128_pd(a.v,0)), + Packet1cd(_mm256_extractf128_pd(a.v,1)))); } template @@ -355,10 +336,7 @@ struct palign_impl static EIGEN_STRONG_INLINE void run(Packet2cd& first, const Packet2cd& second) { if (Offset==0) return; - first.v[0] = first.v[2]; - first.v[1] = first.v[3]; - first.v[2] = second.v[0]; - first.v[3] = second.v[1]; + palign_impl::run(first.v, second.v); } }; @@ -423,12 +401,7 @@ template<> EIGEN_STRONG_INLINE Packet2cd pdiv(const Packet2cd& a, con template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip(const Packet2cd& x) { - Packet2cd res; - for (int i = 0; i < 4; i+=2) { - res.v[i] = x.v[i+1]; - res.v[i+1] = x.v[i]; - } - return res; + return Packet2cd(_mm256_shuffle_pd(x.v, x.v, 0x5)); } } // end namespace internal diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 2dce66819..b92919b75 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -22,6 +22,9 @@ struct Packet2cf __m128 v; }; +// Use the packet_traits defined in AVX/PacketMath.h instead if we're going +// to leverage AVX instructions. +#ifndef EIGEN_VECTORIZE_AVX template<> struct packet_traits > : default_packet_traits { typedef Packet2cf type; @@ -42,6 +45,7 @@ template<> struct packet_traits > : default_packet_traits HasSetLinear = 0 }; }; +#endif template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; }; @@ -248,6 +252,9 @@ struct Packet1cd __m128d v; }; +// Use the packet_traits defined in AVX/PacketMath.h instead if we're going +// to leverage AVX instructions. +#ifndef EIGEN_VECTORIZE_AVX template<> struct packet_traits > : default_packet_traits { typedef Packet1cd type; @@ -268,6 +275,7 @@ template<> struct packet_traits > : default_packet_traits HasSetLinear = 0 }; }; +#endif template<> struct unpacket_traits { typedef std::complex type; enum {size=1}; }; From 6f123d209e65a8cf324df53a9a7e87d8e4c47e9f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 27 Mar 2014 15:29:56 +0100 Subject: [PATCH 0389/1103] Fix geo_* unit tests with respect to AVX --- test/geo_hyperplane.cpp | 8 ++++---- test/geo_parametrizedline.cpp | 2 +- test/geo_quaternion.cpp | 14 +++++++------- test/geo_transformations.cpp | 6 +++--- 4 files changed, 15 insertions(+), 15 deletions(-) diff --git a/test/geo_hyperplane.cpp b/test/geo_hyperplane.cpp index befd7d483..ed5928f10 100644 --- a/test/geo_hyperplane.cpp +++ b/test/geo_hyperplane.cpp @@ -129,9 +129,9 @@ template void hyperplane_alignment() typedef Hyperplane Plane3a; typedef Hyperplane Plane3u; - EIGEN_ALIGN16 Scalar array1[4]; - EIGEN_ALIGN16 Scalar array2[4]; - EIGEN_ALIGN16 Scalar array3[4+1]; + EIGEN_ALIGN_DEFAULT Scalar array1[4]; + EIGEN_ALIGN_DEFAULT Scalar array2[4]; + EIGEN_ALIGN_DEFAULT Scalar array3[4+1]; Scalar* array3u = array3+1; Plane3a *p1 = ::new(reinterpret_cast(array1)) Plane3a; @@ -146,7 +146,7 @@ template void hyperplane_alignment() VERIFY_IS_APPROX(p1->coeffs(), p3->coeffs()); #if defined(EIGEN_VECTORIZE) && EIGEN_ALIGN_STATICALLY - if(internal::packet_traits::Vectorizable) + if(internal::packet_traits::Vectorizable && internal::packet_traits::size<=4) VERIFY_RAISES_ASSERT((::new(reinterpret_cast(array3u)) Plane3a)); #endif } diff --git a/test/geo_parametrizedline.cpp b/test/geo_parametrizedline.cpp index 5a72b3575..427327cd7 100644 --- a/test/geo_parametrizedline.cpp +++ b/test/geo_parametrizedline.cpp @@ -86,7 +86,7 @@ template void parametrizedline_alignment() VERIFY_IS_APPROX(p1->direction(), p3->direction()); #if defined(EIGEN_VECTORIZE) && EIGEN_ALIGN_STATICALLY - if(internal::packet_traits::Vectorizable) + if(internal::packet_traits::Vectorizable && internal::packet_traits::size<=4) VERIFY_RAISES_ASSERT((::new(reinterpret_cast(array3u)) Line4a)); #endif } diff --git a/test/geo_quaternion.cpp b/test/geo_quaternion.cpp index 1694b32c7..de0f2aeda 100644 --- a/test/geo_quaternion.cpp +++ b/test/geo_quaternion.cpp @@ -181,9 +181,9 @@ template void mapQuaternion(void){ v1 = Vector3::Random(); Scalar a = internal::random(-Scalar(M_PI), Scalar(M_PI)); - EIGEN_ALIGN16 Scalar array1[4]; - EIGEN_ALIGN16 Scalar array2[4]; - EIGEN_ALIGN16 Scalar array3[4+1]; + EIGEN_ALIGN_DEFAULT Scalar array1[4]; + EIGEN_ALIGN_DEFAULT Scalar array2[4]; + EIGEN_ALIGN_DEFAULT Scalar array3[4+1]; Scalar* array3unaligned = array3+1; MQuaternionA mq1(array1); @@ -232,9 +232,9 @@ template void quaternionAlignment(void){ typedef Quaternion QuaternionA; typedef Quaternion QuaternionUA; - EIGEN_ALIGN16 Scalar array1[4]; - EIGEN_ALIGN16 Scalar array2[4]; - EIGEN_ALIGN16 Scalar array3[4+1]; + EIGEN_ALIGN_DEFAULT Scalar array1[4]; + EIGEN_ALIGN_DEFAULT Scalar array2[4]; + EIGEN_ALIGN_DEFAULT Scalar array3[4+1]; Scalar* arrayunaligned = array3+1; QuaternionA *q1 = ::new(reinterpret_cast(array1)) QuaternionA; @@ -248,7 +248,7 @@ template void quaternionAlignment(void){ VERIFY_IS_APPROX(q1->coeffs(), q2->coeffs()); VERIFY_IS_APPROX(q1->coeffs(), q3->coeffs()); #if defined(EIGEN_VECTORIZE) && EIGEN_ALIGN_STATICALLY - if(internal::packet_traits::Vectorizable) + if(internal::packet_traits::Vectorizable && internal::packet_traits::size<=4) VERIFY_RAISES_ASSERT((::new(reinterpret_cast(arrayunaligned)) QuaternionA)); #endif } diff --git a/test/geo_transformations.cpp b/test/geo_transformations.cpp index ee3030b5d..7d9080333 100644 --- a/test/geo_transformations.cpp +++ b/test/geo_transformations.cpp @@ -404,9 +404,9 @@ template void transform_alignment() typedef Transform Projective3a; typedef Transform Projective3u; - EIGEN_ALIGN16 Scalar array1[16]; - EIGEN_ALIGN16 Scalar array2[16]; - EIGEN_ALIGN16 Scalar array3[16+1]; + EIGEN_ALIGN_DEFAULT Scalar array1[16]; + EIGEN_ALIGN_DEFAULT Scalar array2[16]; + EIGEN_ALIGN_DEFAULT Scalar array3[16+1]; Scalar* array3u = array3+1; Projective3a *p1 = ::new(reinterpret_cast(array1)) Projective3a; From 7d73c7f18be20407ffd28f72b3275da233658f80 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 27 Mar 2014 15:38:40 +0100 Subject: [PATCH 0390/1103] Change abi version when enabling AVX with GCC --- CMakeLists.txt | 3 +++ 1 file changed, 3 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index fb13769f4..838a41b79 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -199,6 +199,9 @@ if(NOT MSVC) option(EIGEN_TEST_AVX "Enable/Disable AVX in tests/examples" OFF) if(EIGEN_TEST_AVX) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx") + if(CMAKE_COMPILER_IS_GNUCXX) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fabi-version=6") + endif() message(STATUS "Enabling AVX in tests/examples") endif() From c4902a3d0182dfc9ac02a24ec2a52cd567ac0104 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 27 Mar 2014 09:34:51 -0700 Subject: [PATCH 0391/1103] Implemented the AVX version of the ptranspose packet primitive. --- Eigen/src/Core/arch/AVX/Complex.h | 25 +++++++++++++++++ Eigen/src/Core/arch/AVX/PacketMath.h | 41 ++++++++++++++++++++++++++++ 2 files changed, 66 insertions(+) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 4f4f6eb52..0121cec86 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -404,6 +404,31 @@ template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip(const Packet2cd& x return Packet2cd(_mm256_shuffle_pd(x.v, x.v, 0x5)); } +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + __m256d P0 = _mm256_castps_pd(kernel.packet[0].v); + __m256d P1 = _mm256_castps_pd(kernel.packet[1].v); + __m256d P2 = _mm256_castps_pd(kernel.packet[2].v); + __m256d P3 = _mm256_castps_pd(kernel.packet[3].v); + + __m256d T0 = _mm256_shuffle_pd(P0, P1, 15); + __m256d T1 = _mm256_shuffle_pd(P0, P1, 0); + __m256d T2 = _mm256_shuffle_pd(P2, P3, 15); + __m256d T3 = _mm256_shuffle_pd(P2, P3, 0); + + kernel.packet[1].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 32)); + kernel.packet[3].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 49)); + kernel.packet[0].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 32)); + kernel.packet[2].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 49)); +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + __m256d tmp = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 0+(2<<4)); + kernel.packet[1].v = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 1+(3<<4)); + kernel.packet[0].v = tmp; +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 98f8c5416..96a4bc08c 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -414,6 +414,47 @@ struct palign_impl } }; +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); + __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); + __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); + __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); + __m256 T4 = _mm256_unpacklo_ps(kernel.packet[4], kernel.packet[5]); + __m256 T5 = _mm256_unpackhi_ps(kernel.packet[4], kernel.packet[5]); + __m256 T6 = _mm256_unpacklo_ps(kernel.packet[6], kernel.packet[7]); + __m256 T7 = _mm256_unpackhi_ps(kernel.packet[6], kernel.packet[7]); + __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); + __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); + __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); + __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); + __m256 S4 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(1,0,1,0)); + __m256 S5 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(3,2,3,2)); + __m256 S6 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(1,0,1,0)); + __m256 S7 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(3,2,3,2)); + kernel.packet[0] = _mm256_permute2f128_ps(S0, S4, 0x20); + kernel.packet[1] = _mm256_permute2f128_ps(S1, S5, 0x20); + kernel.packet[2] = _mm256_permute2f128_ps(S2, S6, 0x20); + kernel.packet[3] = _mm256_permute2f128_ps(S3, S7, 0x20); + kernel.packet[4] = _mm256_permute2f128_ps(S0, S4, 0x31); + kernel.packet[5] = _mm256_permute2f128_ps(S1, S5, 0x31); + kernel.packet[6] = _mm256_permute2f128_ps(S2, S6, 0x31); + kernel.packet[7] = _mm256_permute2f128_ps(S3, S7, 0x31); +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + __m256d T0 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 15); + __m256d T1 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 0); + __m256d T2 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 15); + __m256d T3 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 0); + + kernel.packet[1] = _mm256_permute2f128_pd(T0, T2, 32); + kernel.packet[3] = _mm256_permute2f128_pd(T0, T2, 49); + kernel.packet[0] = _mm256_permute2f128_pd(T1, T3, 32); + kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49); +} + } // end namespace internal } // end namespace Eigen From b776458ccbd0b8dbde56d0d2dd0a683c6b4b0692 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 27 Mar 2014 10:02:24 -0700 Subject: [PATCH 0392/1103] Vectorized the packing of a row-major matrix used as the left hand side argument in a matrix-matrix product. --- .../Core/products/GeneralBlockPanelKernel.h | 19 ++++++++++++++++--- 1 file changed, 16 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index d17752489..eeeb5290f 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -954,9 +954,22 @@ EIGEN_DONT_INLINE void gemm_pack_lhs kernel; + for (int p = 0; p < PacketSize; ++p) { + kernel.packet[p] = ploadu(&lhs(i+p+m, k)); + } + ptranspose(kernel); + for (int p = 0; p < PacketSize; ++p) { + pstore(blockA+count+m+Pack1*p, cj.pconj(kernel.packet[p])); + } + } + count += PacketSize*Pack1; + } + for(; k Date: Thu, 27 Mar 2014 10:38:41 -0700 Subject: [PATCH 0393/1103] Vectorized the packing of a col-major matrix used as the right hand side argument in a matrix-matrix product when AVX instructions are used. No vectorization takes place when SSE instructions are used, however this doesn't seem to impact performance. --- .../Core/products/GeneralBlockPanelKernel.h | 18 +++++++++++++++++- 1 file changed, 17 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index eeeb5290f..28c2a913e 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -1033,6 +1033,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs::IsComplex && Conjugate> cj; Index packet_cols = (cols/nr) * nr; Index count = 0; + const Index peeled_k = (depth/PacketSize)*PacketSize; for(Index j2=0; j2 kernel; + for (int p = 0; p < PacketSize; ++p) { + kernel.packet[p] = ploadu(&rhs[(j2+p)*rhsStride+k]); + } + ptranspose(kernel); + for (int p = 0; p < PacketSize; ++p) { + pstoreu(blockB+count, cj.pconj(kernel.packet[p])); + count+=PacketSize; + } + } + } + for(; k Date: Thu, 27 Mar 2014 11:00:47 -0700 Subject: [PATCH 0394/1103] Silenced "unused variable" warnings when compiling with FMA. --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 3 +++ 1 file changed, 3 insertions(+) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 28c2a913e..897b6a914 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -215,6 +215,7 @@ public: // (in the case where there is no FMA) gcc fails to figure out how to avoid // spilling register. #ifdef EIGEN_VECTORIZE_FMA + EIGEN_UNUSED_VARIABLE(tmp); c = pmadd(a,b,c); #else tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp); @@ -299,6 +300,7 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { #ifdef EIGEN_VECTORIZE_FMA + EIGEN_UNUSED_VARIABLE(tmp); c.v = pmadd(a.v,b,c.v); #else tmp = b; tmp = pmul(a.v,tmp); c.v = padd(c.v,tmp); @@ -520,6 +522,7 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { #ifdef EIGEN_VECTORIZE_FMA + EIGEN_UNUSED_VARIABLE(tmp); c = pmadd(a,b,c); #else tmp = b; tmp.v = pmul(a,tmp.v); c = padd(c,tmp); From 729363114f0c7ea8ff3f8ddd8f6a83335b0f3909 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 27 Mar 2014 11:20:41 -0700 Subject: [PATCH 0395/1103] Fixed compilation error when FMA instructions are enabled. --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 5 ----- 1 file changed, 5 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 897b6a914..cd89aed1f 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -521,12 +521,7 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { -#ifdef EIGEN_VECTORIZE_FMA - EIGEN_UNUSED_VARIABLE(tmp); - c = pmadd(a,b,c); -#else tmp = b; tmp.v = pmul(a,tmp.v); c = padd(c,tmp); -#endif } EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const From 58fe2fc2b21401365ace575738d878dad21eb184 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 27 Mar 2014 23:38:50 +0100 Subject: [PATCH 0396/1103] enforce the use of vfmadd231ps for pmadd (gcc and clang stupidely generates the other fmadd variants plus some register moves...) --- Eigen/src/Core/arch/AVX/PacketMath.h | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 96a4bc08c..dceddb518 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -124,7 +124,19 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv(const Packet8i& /*a*/, co } #ifdef EIGEN_VECTORIZE_FMA -template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { return _mm256_fmadd_ps(a,b,c); } +template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { +#if defined(__clang__) || defined(__GNUC__) + // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers, + // and gcc stupidly generates a vfmadd132ps instruction, + // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate + // the result of the product. + Packet8f res = c; + asm("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else + return _mm256_fmadd_ps(a,b,c); +#endif +} template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { return _mm256_fmadd_pd(a,b,c); } #endif From ee866790967ab4ab11a62987dd21bac66237cba9 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 27 Mar 2014 16:03:03 -0700 Subject: [PATCH 0397/1103] Introduced pscatter/pgather packet primitives. They will be used to optimize the loop peeling code of the block-panel matrix multiplication kernel. --- Eigen/src/Core/GenericPacketMath.h | 6 +++++ test/packetmath.cpp | 39 ++++++++++++++++++++++++++++-- 2 files changed, 43 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index f9ddf4718..03e7f410c 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -217,6 +217,12 @@ template EIGEN_DEVICE_FUNC inline void pstore( /** \internal copy the packet \a from to \a *to, (un-aligned store) */ template EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from) +{ (*to) = from; } + + template EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, int /*stride*/) + { return ploadu(from); } + + template EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, int /*stride*/) { (*to) = from; } /** \internal tries to do cache prefetching of \a addr */ diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 735af7017..7deefe890 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -356,8 +356,38 @@ template void packetmath_complex() internal::pstore(pval,internal::pcplxflip(internal::pload(data1))); VERIFY(areApprox(ref, pval, PacketSize) && "pcplxflip"); } - - +} + +template void packetmath_scatter_gather() { + typedef typename internal::packet_traits::type Packet; + typedef typename NumTraits::Real RealScalar; + const int PacketSize = internal::packet_traits::size; + Scalar data1[PacketSize]; + RealScalar refvalue = 0; + for (int i=0; i()/RealScalar(PacketSize); + } + Scalar buffer[PacketSize*11]; + memset(buffer, 0, 11*sizeof(Packet)); + Packet packet = internal::pload(data1); + internal::pscatter(buffer, packet, 11); + + for (int i = 0; i < PacketSize*11; ++i) { + if ((i%11) == 0) { + VERIFY(isApproxAbs(buffer[i], data1[i/11], refvalue)); + } else { + VERIFY(isApproxAbs(buffer[i], Scalar(0), refvalue)); + } + } + + for (int i=0; i()/RealScalar(PacketSize); + } + packet = internal::pgather(buffer, 7); + internal::pstore(data1, packet); + for (int i = 0; i < PacketSize; ++i) { + VERIFY(isApproxAbs(data1[i], buffer[i*7], refvalue)); + } } void test_packetmath() @@ -378,5 +408,10 @@ void test_packetmath() CALL_SUBTEST_1( packetmath_complex >() ); CALL_SUBTEST_2( packetmath_complex >() ); + + CALL_SUBTEST_1( packetmath_scatter_gather() ); + CALL_SUBTEST_2( packetmath_scatter_gather() ); + CALL_SUBTEST_3( packetmath_scatter_gather >() ); + CALL_SUBTEST_3( packetmath_scatter_gather >() ); } } From 7f3162f7071db63bdbdc21f4c101543df00e4661 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 27 Mar 2014 17:42:25 -0700 Subject: [PATCH 0398/1103] Implemented the AVX version of the gather and scatter packet primitives. --- Eigen/src/Core/arch/AVX/Complex.h | 39 +++++++++++++++++++++++++++- Eigen/src/Core/arch/AVX/PacketMath.h | 35 +++++++++++++++++++++++++ 2 files changed, 73 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 0121cec86..ec9c861f9 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -99,6 +99,29 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } +template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packet4cf>(const std::complex* from, int stride) +{ + return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]), + std::imag(from[2*stride]), std::real(from[2*stride]), + std::imag(from[1*stride]), std::real(from[1*stride]), + std::imag(from[0*stride]), std::real(from[0*stride]))); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet4cf>(std::complex* to, const Packet4cf& from, int stride) +{ + __m128 low = _mm256_extractf128_ps(from.v, 0); + to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)), + _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1))); + to[stride*1] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)), + _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3))); + + __m128 high = _mm256_extractf128_ps(from.v, 1); + to[stride*2] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 0)), + _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1))); + to[stride*3] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)), + _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3))); + +} template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet4cf& a) { @@ -297,7 +320,21 @@ template<> EIGEN_STRONG_INLINE Packet2cd ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } -template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cd& a) +template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather, Packet2cd>(const std::complex* from, int stride) +{ + return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]), + std::imag(from[0*stride]), std::real(from[0*stride]))); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cd>(std::complex* to, const Packet2cd& from, int stride) +{ + __m128d low = _mm256_extractf128_pd(from.v, 0); + to[stride*0] = std::complex(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1))); + __m128d high = _mm256_extractf128_pd(from.v, 1); + to[stride*1] = std::complex(_mm_cvtsd_f64(high), _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1))); +} + +template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cd& a) { __m128d low = _mm256_extractf128_pd(a.v, 0); EIGEN_ALIGN16 double res[2]; diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 96a4bc08c..aa2ac3b0b 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -179,6 +179,41 @@ template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet8f& template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } +// TODO: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available +template<> EIGEN_DEVICE_FUNC inline Packet8f pgather(const float* from, int stride) +{ + return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride], + from[3*stride], from[2*stride], from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet4d pgather(const double* from, int stride) +{ + return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet8f& from, int stride) +{ + __m128 low = _mm256_extractf128_ps(from, 0); + to[stride*0] = _mm_cvtss_f32(low); + to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1)); + to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)); + to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3)); + + __m128 high = _mm256_extractf128_ps(from, 1); + to[stride*4] = _mm_cvtss_f32(high); + to[stride*5] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1)); + to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)); + to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet4d& from, int stride) +{ + __m128d low = _mm256_extractf128_pd(from, 0); + to[stride*0] = _mm_cvtsd_f64(low); + to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1)); + __m128d high = _mm256_extractf128_pd(from, 1); + to[stride*2] = _mm_cvtsd_f64(high); + to[stride*3] = _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1)); +} + template<> EIGEN_STRONG_INLINE void pstore1(float* to, const float& a) { Packet8f pa = pset1(a); From 8a94cb3edde854b89031a0e985c524f2f6bf799d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 27 Mar 2014 18:29:01 -0700 Subject: [PATCH 0399/1103] Implemented the SSE version of the gather and scatter packet primitives. --- Eigen/src/Core/GenericPacketMath.h | 2 +- Eigen/src/Core/arch/SSE/Complex.h | 18 +++++++++++++++ Eigen/src/Core/arch/SSE/PacketMath.h | 33 ++++++++++++++++++++++++++++ test/packetmath.cpp | 1 + 4 files changed, 53 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 03e7f410c..82eeeed4a 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -223,7 +223,7 @@ template EIGEN_DEVICE_FUNC inline void pstoreu { return ploadu(from); } template EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, int /*stride*/) -{ (*to) = from; } + { pstore(to, from); } /** \internal tries to do cache prefetching of \a addr */ template inline void prefetch(const Scalar* addr) diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index b92919b75..86b90b2ee 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -111,6 +111,24 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } + +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) +{ + return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]), + std::imag(from[0*stride]), std::real(from[0*stride]))); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) +{ + /* for (int i = 0; i < 2; i+=2) { + to[stride*i] = std::complex(from.v[i], from.v[i+1]); + }*/ + to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)), + _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1))); + to[stride*1] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)), + _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3))); +} + template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cf& a) diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 937f63f88..a98ca6bea 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -339,6 +339,39 @@ template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), _mm_castps_pd(from)); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), _mm_castsi128_pd(from)); } +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) +{ + return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, int stride) +{ + return _mm_set_pd(from[1*stride], from[0*stride]); +} +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, int stride) +{ + return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); + } + +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, int stride) +{ + to[stride*0] = _mm_cvtss_f32(from); + to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1)); + to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2)); + to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, int stride) +{ + to[stride*0] = _mm_cvtsd_f64(from); + to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, int stride) +{ + to[stride*0] = _mm_cvtsi128_si32(from); + to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1)); + to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2)); + to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3)); +} + // some compilers might be tempted to perform multiple moves instead of using a vector path. template<> EIGEN_STRONG_INLINE void pstore1(float* to, const float& a) { diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 7deefe890..317b1c8fe 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -411,6 +411,7 @@ void test_packetmath() CALL_SUBTEST_1( packetmath_scatter_gather() ); CALL_SUBTEST_2( packetmath_scatter_gather() ); + CALL_SUBTEST_3( packetmath_scatter_gather() ); CALL_SUBTEST_3( packetmath_scatter_gather >() ); CALL_SUBTEST_3( packetmath_scatter_gather >() ); } From c94fde118a2b6b2b415a7860ba066ba240bc8ac3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 28 Mar 2014 09:11:06 +0100 Subject: [PATCH 0400/1103] Enable vectorization of gemv for PacketSize>4 through unaligned loads (still better than no vectorization) --- Eigen/src/Core/products/GeneralMatrixVector.h | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h index 59fc54c4e..340c51394 100644 --- a/Eigen/src/Core/products/GeneralMatrixVector.h +++ b/Eigen/src/Core/products/GeneralMatrixVector.h @@ -136,12 +136,17 @@ EIGEN_DONT_INLINE void general_matrix_vector_product 4. - if( (size_t(lhs)%sizeof(LhsScalar)) || (size_t(res)%sizeof(ResScalar)) || LhsPacketSize > 4) + if( (size_t(lhs)%sizeof(LhsScalar)) || (size_t(res)%sizeof(ResScalar)) ) { alignedSize = 0; alignedStart = 0; } + else if(LhsPacketSize > 4) + { + // TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4. + // Currently, it seems to be better to perform unaligned loads anyway + alignmentPattern = NoneAligned; + } else if (LhsPacketSize>1) { eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || size 4. - if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (size_t(lhs)%sizeof(LhsScalar)) || (size_t(rhs)%sizeof(RhsScalar)) || - (LhsPacketSize > 4)) + if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (size_t(lhs)%sizeof(LhsScalar)) || (size_t(rhs)%sizeof(RhsScalar)) ) { alignedSize = 0; alignedStart = 0; } + else if(LhsPacketSize > 4) + { + // TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4. + alignmentPattern = NoneAligned; + } else if (LhsPacketSize>1) { eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || depth Date: Fri, 28 Mar 2014 10:18:04 +0100 Subject: [PATCH 0401/1103] Add a mechanism to recursively access to half-size packet types --- Eigen/src/Core/GenericPacketMath.h | 6 +++++- Eigen/src/Core/arch/AVX/Complex.h | 4 ++++ Eigen/src/Core/arch/AVX/PacketMath.h | 4 ++++ Eigen/src/Core/arch/AltiVec/Complex.h | 1 + Eigen/src/Core/arch/AltiVec/PacketMath.h | 2 ++ Eigen/src/Core/arch/NEON/Complex.h | 1 + Eigen/src/Core/arch/NEON/PacketMath.h | 2 ++ Eigen/src/Core/arch/SSE/Complex.h | 4 ++++ Eigen/src/Core/arch/SSE/PacketMath.h | 5 +++++ 9 files changed, 28 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 82eeeed4a..3e5db1a88 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -42,6 +42,8 @@ namespace internal { struct default_packet_traits { enum { + HasHalfPacket = 0, + HasAdd = 1, HasSub = 1, HasMul = 1, @@ -71,10 +73,12 @@ struct default_packet_traits template struct packet_traits : default_packet_traits { typedef T type; + typedef T half; enum { Vectorizable = 0, size = 1, - AlignedOnScalar = 0 + AlignedOnScalar = 0, + HasHalfPacket = 0 }; enum { HasAdd = 0, diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index ec9c861f9..7c1947a4f 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -25,10 +25,12 @@ struct Packet4cf template<> struct packet_traits > : default_packet_traits { typedef Packet4cf type; + typedef Packet2cf half; enum { Vectorizable = 1, AlignedOnScalar = 1, size = 4, + HasHalfPacket = 1, HasAdd = 1, HasSub = 1, @@ -257,10 +259,12 @@ struct Packet2cd template<> struct packet_traits > : default_packet_traits { typedef Packet2cd type; + typedef Packet1cd half; enum { Vectorizable = 1, AlignedOnScalar = 0, size = 2, + HasHalfPacket = 1, HasAdd = 1, HasSub = 1, diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 6d5a5f53f..132c1abe3 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -40,10 +40,12 @@ template<> struct is_arithmetic<__m256d> { enum { value = true }; }; template<> struct packet_traits : default_packet_traits { typedef Packet8f type; + typedef Packet4f half; enum { Vectorizable = 1, AlignedOnScalar = 1, size=8, + HasHalfPacket = 1, HasDiv = 1, HasSin = 0, @@ -56,10 +58,12 @@ template<> struct packet_traits : default_packet_traits template<> struct packet_traits : default_packet_traits { typedef Packet4d type; + typedef Packet2d half; enum { Vectorizable = 1, AlignedOnScalar = 1, size=4, + HasHalfPacket = 1, HasDiv = 1, HasExp = 0 diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 68d9a2bff..db52074f4 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -33,6 +33,7 @@ struct Packet2cf template<> struct packet_traits > : default_packet_traits { typedef Packet2cf type; + typedef Packet2cf half; enum { Vectorizable = 1, AlignedOnScalar = 1, diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 45d1954f7..5d7a16f5c 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -73,6 +73,7 @@ static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui) template<> struct packet_traits : default_packet_traits { typedef Packet4f type; + typedef Packet4f half; enum { Vectorizable = 1, AlignedOnScalar = 1, @@ -89,6 +90,7 @@ template<> struct packet_traits : default_packet_traits template<> struct packet_traits : default_packet_traits { typedef Packet4i type; + typedef Packet4i half; enum { // FIXME check the Has* Vectorizable = 1, diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 8d9255eef..e49c1a873 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -28,6 +28,7 @@ struct Packet2cf template<> struct packet_traits > : default_packet_traits { typedef Packet2cf type; + typedef Packet2cf half; enum { Vectorizable = 1, AlignedOnScalar = 1, diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 05e891df2..fae7b55fc 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -66,6 +66,7 @@ typedef uint32x4_t Packet4ui; template<> struct packet_traits : default_packet_traits { typedef Packet4f type; + typedef Packet4f half; enum { Vectorizable = 1, AlignedOnScalar = 1, @@ -83,6 +84,7 @@ template<> struct packet_traits : default_packet_traits template<> struct packet_traits : default_packet_traits { typedef Packet4i type; + typedef Packet4i half; enum { Vectorizable = 1, AlignedOnScalar = 1, diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 86b90b2ee..694979e19 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -28,10 +28,12 @@ struct Packet2cf template<> struct packet_traits > : default_packet_traits { typedef Packet2cf type; + typedef Packet2cf half; enum { Vectorizable = 1, AlignedOnScalar = 1, size = 2, + HasHalfPacket = 0, HasAdd = 1, HasSub = 1, @@ -276,10 +278,12 @@ struct Packet1cd template<> struct packet_traits > : default_packet_traits { typedef Packet1cd type; + typedef Packet1cd half; enum { Vectorizable = 1, AlignedOnScalar = 0, size = 1, + HasHalfPacket = 0, HasAdd = 1, HasSub = 1, diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index a98ca6bea..ea05a3415 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -64,10 +64,12 @@ template<> struct is_arithmetic<__m128d> { enum { value = true }; }; template<> struct packet_traits : default_packet_traits { typedef Packet4f type; + typedef Packet4f half; enum { Vectorizable = 1, AlignedOnScalar = 1, size=4, + HasHalfPacket = 0, HasDiv = 1, HasSin = EIGEN_FAST_MATH, @@ -80,10 +82,12 @@ template<> struct packet_traits : default_packet_traits template<> struct packet_traits : default_packet_traits { typedef Packet2d type; + typedef Packet2d half; enum { Vectorizable = 1, AlignedOnScalar = 1, size=2, + HasHalfPacket = 0, HasDiv = 1, HasExp = 1, @@ -94,6 +98,7 @@ template<> struct packet_traits : default_packet_traits template<> struct packet_traits : default_packet_traits { typedef Packet4i type; + typedef Packet4i half; enum { // FIXME check the Has* Vectorizable = 1, From 39bfbd43f05691874a78a5a6bf4504cf0e6ff452 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 28 Mar 2014 12:00:08 -0700 Subject: [PATCH 0402/1103] Properly align the input data to prevent false failures of the packetmath.cpp test. --- test/packetmath.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 317b1c8fe..08566faf8 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -362,12 +362,12 @@ template void packetmath_scatter_gather() { typedef typename internal::packet_traits::type Packet; typedef typename NumTraits::Real RealScalar; const int PacketSize = internal::packet_traits::size; - Scalar data1[PacketSize]; + EIGEN_ALIGN_DEFAULT Scalar data1[PacketSize]; RealScalar refvalue = 0; for (int i=0; i()/RealScalar(PacketSize); } - Scalar buffer[PacketSize*11]; + EIGEN_ALIGN_DEFAULT Scalar buffer[PacketSize*11]; memset(buffer, 0, 11*sizeof(Packet)); Packet packet = internal::pload(data1); internal::pscatter(buffer, packet, 11); From ad59ade116969ca7b18409d690caf00c0b1c34c7 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 28 Mar 2014 12:11:23 -0700 Subject: [PATCH 0403/1103] Vectorized the loop peeling of the inner loop of the block-panel matrix multiplication code. This speeds up the multiplication of matrices which size is not a multiple of the packet size. --- .../Core/products/GeneralBlockPanelKernel.h | 222 ++++++++++++------ 1 file changed, 156 insertions(+), 66 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 0f47f6de5..3ed1fc5a3 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -206,6 +206,11 @@ public: dest = pload(a); } + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploadu(a); + } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, AccPacket& tmp) const { // It would be a lot cleaner to call pmadd all the time. Unfortunately if we @@ -278,7 +283,12 @@ public: { dest = pload(a); } - + + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploadu(a); + } + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) { pbroadcast4(b, b0, b1, b2, b3); @@ -334,7 +344,9 @@ public: && packet_traits::Vectorizable, RealPacketSize = Vectorizable ? packet_traits::size : 1, ResPacketSize = Vectorizable ? packet_traits::size : 1, - + LhsPacketSize = Vectorizable ? packet_traits::size : 1, + RhsPacketSize = Vectorizable ? packet_traits::size : 1, + // FIXME: should depend on NumberOfRegisters nr = 4, mr = ResPacketSize, @@ -402,6 +414,11 @@ public: dest = pload((const typename unpacket_traits::type*)(a)); } + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploadu((const typename unpacket_traits::type*)(a)); + } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacket& c, RhsPacket& /*tmp*/) const { c.first = padd(pmul(a,b.first), c.first); @@ -509,6 +526,11 @@ public: dest = ploaddup(a); } + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + { + dest = ploaddup(a); + } + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const { madd_impl(a, b, c, tmp, typename conditional::type()); @@ -706,49 +728,84 @@ void gebp_kernel const LhsScalar* blA = &blockA[i*strideA+offsetA]; prefetch(&blA[0]); - // gets a 1 x 8 res block as registers - ResScalar C0(0), C1(0), C2(0), C3(0), C4(0), C5(0), C6(0), C7(0); // FIXME directly use blockB ??? const RhsScalar* blB = &blockB[j2*strideB+offsetB*8]; - // TODO peel this loop - for(Index k=0; k SwappedTraits; + typedef typename SwappedTraits::ResScalar SResScalar; + typedef typename SwappedTraits::LhsPacket SLhsPacket; + typedef typename SwappedTraits::RhsPacket SRhsPacket; + typedef typename SwappedTraits::ResPacket SResPacket; + typedef typename SwappedTraits::AccPacket SAccPacket; + SwappedTraits straits; + + SAccPacket C0; + straits.initAcc(C0); + for(Index k=0; k(&res[j2*resStride + i], resStride); + SResPacket alphav = pset1(alpha); + straits.acc(C0, alphav, R); + pscatter(&res[j2*resStride + i], R, resStride); + + EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows"); + } + else + { + // gets a 1 x 8 res block as registers + ResScalar C0(0), C1(0), C2(0), C3(0), C4(0), C5(0), C6(0), C7(0); + + for(Index k=0; k const LhsScalar* blA = &blockA[i*strideA+offsetA]; prefetch(&blA[0]); - // gets a 1 x 4 res block as registers - ResScalar C0(0), C1(0), C2(0), C3(0); // FIXME directly use blockB ??? const RhsScalar* blB = &blockB[j2*strideB+offsetB*4]; - // TODO peel this loop - for(Index k=0; k SwappedTraits; + typedef typename SwappedTraits::ResScalar SResScalar; + typedef typename SwappedTraits::LhsPacket SLhsPacket; + typedef typename SwappedTraits::RhsPacket SRhsPacket; + typedef typename SwappedTraits::ResPacket SResPacket; + typedef typename SwappedTraits::AccPacket SAccPacket; + SwappedTraits straits; + + SAccPacket C0; + straits.initAcc(C0); + for(Index k=0; k(&res[j2*resStride + i], resStride); + SResPacket alphav = pset1(alpha); + straits.acc(C0, alphav, R); + pscatter(&res[j2*resStride + i], R, resStride); + + EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows"); + } else { + // gets a 1 x 4 res block as registers + ResScalar C0(0), C1(0), C2(0), C3(0); + + for(Index k=0; k Date: Sun, 30 Mar 2014 21:57:05 +0200 Subject: [PATCH 0404/1103] Optimize gebp kernel: 1 - increase peeling level along the depth dimention (+5% for large matrices, i.e., >1000) 2 - improve pipelining when dealing with latest rows of the lhs --- .../Core/products/GeneralBlockPanelKernel.h | 454 +++++++++++------- 1 file changed, 271 insertions(+), 183 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 3ed1fc5a3..a9e42c8aa 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -95,6 +95,9 @@ void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) k = std::min(k, l1/kdiv); SizeType _m = k>0 ? l2/(4 * sizeof(LhsScalar) * k) : 0; if(_m @@ -328,6 +331,22 @@ protected: conj_helper cj; }; +template +struct DoublePacket +{ + Packet first; + Packet second; +}; + +template +DoublePacket padd(const DoublePacket &a, const DoublePacket &b) +{ + DoublePacket res; + res.first = padd(a.first, b.first); + res.second = padd(a.second,b.second); + return res; +} + template class gebp_traits, std::complex, _ConjLhs, _ConjRhs > { @@ -357,20 +376,16 @@ public: typedef typename packet_traits::type RealPacket; typedef typename packet_traits::type ScalarPacket; - struct DoublePacket - { - RealPacket first; - RealPacket second; - }; + typedef DoublePacket DoublePacketType; typedef typename conditional::type LhsPacket; - typedef typename conditional::type RhsPacket; + typedef typename conditional::type RhsPacket; typedef typename conditional::type ResPacket; - typedef typename conditional::type AccPacket; + typedef typename conditional::type AccPacket; EIGEN_STRONG_INLINE void initAcc(Scalar& p) { p = Scalar(0); } - EIGEN_STRONG_INLINE void initAcc(DoublePacket& p) + EIGEN_STRONG_INLINE void initAcc(DoublePacketType& p) { p.first = pset1(RealScalar(0)); p.second = pset1(RealScalar(0)); @@ -383,7 +398,7 @@ public: } // Vectorized path - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacket& dest) const + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacketType& dest) const { dest.first = pset1(real(*b)); dest.second = pset1(imag(*b)); @@ -393,7 +408,7 @@ public: void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); // Vectorized path - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacket& b0, DoublePacket& b1) + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacketType& b0, DoublePacketType& b1) { // FIXME not sure that's the best way to implement it! loadRhs(b+0, b0); @@ -419,7 +434,7 @@ public: dest = ploadu((const typename unpacket_traits::type*)(a)); } - EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacket& c, RhsPacket& /*tmp*/) const + EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacketType& c, RhsPacket& /*tmp*/) const { c.first = padd(pmul(a,b.first), c.first); c.second = padd(pmul(a,b.second),c.second); @@ -432,7 +447,7 @@ public: EIGEN_STRONG_INLINE void acc(const Scalar& c, const Scalar& alpha, Scalar& r) const { r += alpha * c; } - EIGEN_STRONG_INLINE void acc(const DoublePacket& c, const ResPacket& alpha, ResPacket& r) const + EIGEN_STRONG_INLINE void acc(const DoublePacketType& c, const ResPacket& alpha, ResPacket& r) const { // assemble c ResPacket tmp; @@ -571,6 +586,14 @@ struct gebp_kernel typedef typename Traits::RhsPacket RhsPacket; typedef typename Traits::ResPacket ResPacket; typedef typename Traits::AccPacket AccPacket; + + typedef gebp_traits SwappedTraits; + typedef typename SwappedTraits::ResScalar SResScalar; + typedef typename SwappedTraits::LhsPacket SLhsPacket; + typedef typename SwappedTraits::RhsPacket SRhsPacket; + typedef typename SwappedTraits::ResPacket SResPacket; + typedef typename SwappedTraits::AccPacket SAccPacket; + enum { Vectorizable = Traits::Vectorizable, @@ -591,6 +614,7 @@ void gebp_kernel Index strideA, Index strideB, Index offsetA, Index offsetB) { Traits traits; + SwappedTraits straits; if(strideA==-1) strideA = depth; if(strideB==-1) strideB = depth; @@ -599,7 +623,9 @@ void gebp_kernel Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; // Here we assume that mr==LhsProgress const Index peeled_mc = (rows/mr)*mr; - const Index peeled_kc = (depth/4)*4; + enum { pk = 8 }; // NOTE Such a large peeling factor is important for large matrices (~ +5% when >1000 on Haswell) + const Index peeled_kc = depth & ~(pk-1); + const Index depth2 = depth & ~1; // loops on each micro vertical panel of rhs (depth x nr) // First pass using depth x 8 panels @@ -634,14 +660,14 @@ void gebp_kernel // uncomment for register prefetching // LhsPacket A1; // traits.loadLhs(blA, A0); - for(Index k=0; k EIGEN_GEBGP_ONESTEP8(1,A0,A1); EIGEN_GEBGP_ONESTEP8(2,A1,A0); EIGEN_GEBGP_ONESTEP8(3,A0,A1); + EIGEN_GEBGP_ONESTEP8(4,A1,A0); + EIGEN_GEBGP_ONESTEP8(5,A0,A1); + EIGEN_GEBGP_ONESTEP8(6,A1,A0); + EIGEN_GEBGP_ONESTEP8(7,A0,A1); - blB += 4*8*RhsProgress; - blA += 4*mr; + blB += pk*8*RhsProgress; + blA += pk*mr; } // process remaining peeled loop for(Index k=peeled_kc; k pstoreu(r0+5*resStride, R5); pstoreu(r0+6*resStride, R6); pstoreu(r0+7*resStride, R0); - } - for(Index i=peeled_mc; i SwappedTraits; - typedef typename SwappedTraits::ResScalar SResScalar; - typedef typename SwappedTraits::LhsPacket SLhsPacket; - typedef typename SwappedTraits::RhsPacket SRhsPacket; - typedef typename SwappedTraits::ResPacket SResPacket; - typedef typename SwappedTraits::AccPacket SAccPacket; - SwappedTraits straits; - - SAccPacket C0; - straits.initAcc(C0); - for(Index k=0; k(&res[j2*resStride + i], resStride); - SResPacket alphav = pset1(alpha); - straits.acc(C0, alphav, R); - pscatter(&res[j2*resStride + i], R, resStride); - - EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows"); - } - else - { - // gets a 1 x 8 res block as registers - ResScalar C0(0), C1(0), C2(0), C3(0), C4(0), C5(0), C6(0), C7(0); - - for(Index k=0; k SwappedTraits; + typedef typename SwappedTraits::ResScalar SResScalar; + typedef typename SwappedTraits::LhsPacket SLhsPacket; + typedef typename SwappedTraits::RhsPacket SRhsPacket; + typedef typename SwappedTraits::ResPacket SResPacket; + typedef typename SwappedTraits::AccPacket SAccPacket; + SwappedTraits straits; + + Index rows2 = (rows & ~1); + for(Index i=peeled_mc; i(&res[j2*resStride + i], resStride); + SResPacket alphav = pset1(alpha); + straits.acc(padd(C0,C1), alphav, R); + pscatter(&res[j2*resStride + i], R, resStride); + + R = pgather(&res[j2*resStride + i + 1], resStride); + straits.acc(padd(C2,C3), alphav, R); + pscatter(&res[j2*resStride + i + 1], R, resStride); + + EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows 8"); + } + if(rows2!=rows) + { + Index i = rows-1; + const LhsScalar* blA = &blockA[i*strideA+offsetA]; + const RhsScalar* blB = &blockB[j2*strideB+offsetB*8]; + + EIGEN_ASM_COMMENT("begin_vectorized_multiplication_of_last_rows 8"); + + SAccPacket C0,C1; + straits.initAcc(C0); // even + straits.initAcc(C1); // odd + + for(Index k=0; k(&res[j2*resStride + i], resStride); + SResPacket alphav = pset1(alpha); + straits.acc(padd(C0,C1), alphav, R); + pscatter(&res[j2*resStride + i], R, resStride); + } + } + else + { + // Pure scalar path + for(Index i=peeled_mc; i=4) { for(Index j2=packet_cols8; j2 for(Index i=0; i // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*4]; LhsPacket A0; - // uncomment for register prefetching - // LhsPacket A1; - // traits.loadLhs(blA, A0); - for(Index k=0; k traits.broadcastRhs(&blB[2+4*K*RhsProgress], B_0, B1); \ traits.madd(A0, B_0,C2, B_0); \ traits.madd(A0, B1, C3, B1) - + EIGEN_GEBGP_ONESTEP4(0); EIGEN_GEBGP_ONESTEP4(1); EIGEN_GEBGP_ONESTEP4(2); EIGEN_GEBGP_ONESTEP4(3); + EIGEN_GEBGP_ONESTEP4(4); + EIGEN_GEBGP_ONESTEP4(5); + EIGEN_GEBGP_ONESTEP4(6); + EIGEN_GEBGP_ONESTEP4(7); - blB += 4*4*RhsProgress; - blA += 4*mr; + blB += pk*4*RhsProgress; + blA += pk*mr; } - // process remaining peeled loop + // process remaining of peeled loop for(Index k=peeled_kc; k for(Index i=peeled_mc; i SwappedTraits; - typedef typename SwappedTraits::ResScalar SResScalar; - typedef typename SwappedTraits::LhsPacket SLhsPacket; - typedef typename SwappedTraits::RhsPacket SRhsPacket; - typedef typename SwappedTraits::ResPacket SResPacket; - typedef typename SwappedTraits::AccPacket SAccPacket; - SwappedTraits straits; - - SAccPacket C0; - straits.initAcc(C0); - for(Index k=0; k(&res[j2*resStride + i], resStride); - SResPacket alphav = pset1(alpha); - straits.acc(C0, alphav, R); - pscatter(&res[j2*resStride + i], R, resStride); - - EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows"); - } else { - // gets a 1 x 4 res block as registers - ResScalar C0(0), C1(0), C2(0), C3(0); + SLhsPacket A0; + straits.loadLhsUnaligned(blB, A0); + SRhsPacket B_0; + straits.loadRhs(&blA[k], B_0); + SRhsPacket T0; + straits.madd(A0,B_0,C0,T0); + blB += 4; + } + SResPacket R = pgather(&res[j2*resStride + i], resStride); + SResPacket alphav = pset1(alpha); + straits.acc(C0, alphav, R); + pscatter(&res[j2*resStride + i], R, resStride); + + EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows 1x4"); + } + else + { + // Pure scalar path + // gets a 1 x 4 res block as registers + ResScalar C0(0), C1(0), C2(0), C3(0); - for(Index k=0; k do the same but with nr==1 for(Index j2=packet_cols4; j2 traits.acc(C0, alphav, R0); pstoreu(r0, R0); } + // pure scalar path for(Index i=peeled_mc; i Date: Sun, 30 Mar 2014 22:43:47 +0200 Subject: [PATCH 0405/1103] Workaround alignment warnings --- Eigen/src/Core/arch/AVX/Complex.h | 32 +++++++++++----------------- Eigen/src/Core/arch/AVX/PacketMath.h | 7 ++++-- Eigen/src/Core/arch/SSE/Complex.h | 9 +++----- 3 files changed, 20 insertions(+), 28 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 7c1947a4f..cb16180c5 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -87,15 +87,10 @@ template<> EIGEN_STRONG_INLINE Packet4cf pset1(const std::complex EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex* from) { - // This should be optimized. - __m128 complex1 = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)from); - complex1 = _mm_movelh_ps(complex1, complex1); - __m128 complex2 = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from+1)); - complex2 = _mm_movelh_ps(complex2, complex2); - __m256 result = _mm256_setzero_ps(); - result = _mm256_insertf128_ps(result, complex1, 0); - result = _mm256_insertf128_ps(result, complex2, 1); - return Packet4cf(result); + // FIXME The following might be optimized using _mm256_movedup_pd + Packet2cf a = ploaddup(from); + Packet2cf b = ploaddup(from+1); + return Packet4cf(_mm256_insertf128_ps(_mm256_castps128_ps256(a.v), b.v, 1)); } template<> EIGEN_STRONG_INLINE void pstore >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } @@ -104,33 +99,30 @@ template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packet4cf>(const std::complex* from, int stride) { return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]), - std::imag(from[2*stride]), std::real(from[2*stride]), - std::imag(from[1*stride]), std::real(from[1*stride]), - std::imag(from[0*stride]), std::real(from[0*stride]))); + std::imag(from[2*stride]), std::real(from[2*stride]), + std::imag(from[1*stride]), std::real(from[1*stride]), + std::imag(from[0*stride]), std::real(from[0*stride]))); } template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet4cf>(std::complex* to, const Packet4cf& from, int stride) { __m128 low = _mm256_extractf128_ps(from.v, 0); to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)), - _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1))); + _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1))); to[stride*1] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)), - _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3))); + _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3))); __m128 high = _mm256_extractf128_ps(from.v, 1); to[stride*2] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 0)), - _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1))); + _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1))); to[stride*3] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)), - _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3))); + _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3))); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet4cf& a) { - __m128 low = _mm256_extractf128_ps(a.v, 0); - std::complex res; - _mm_storel_pi((__m64*)&res, low); - return res; + return pfirst(Packet2cf(_mm256_castps256_ps128(a.v))); } template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) { diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 132c1abe3..38f52ecc8 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -173,6 +173,7 @@ template<> EIGEN_STRONG_INLINE Packet8i ploadu(const int* from) { EIGE // Loads 4 floats from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, a3} template<> EIGEN_STRONG_INLINE Packet8f ploaddup(const float* from) { + // FIXME we should only load the first 128bits Packet8f tmp = ploadu(from); Packet8f tmp1 = _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2)); Packet8f tmp2 = _mm256_permute_ps(tmp, _MM_SHUFFLE(1,1,0,0)); @@ -181,6 +182,7 @@ template<> EIGEN_STRONG_INLINE Packet8f ploaddup(const float* from) // Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1} template<> EIGEN_STRONG_INLINE Packet4d ploaddup(const double* from) { + // FIXME we should only load the first 128bits Packet4d tmp = ploadu(from); Packet4d tmp1 = _mm256_permute_pd(tmp,0); Packet4d tmp2 = _mm256_permute_pd(tmp,3); @@ -195,11 +197,12 @@ template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet8f& template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } -// TODO: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available +// NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available +// NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4); template<> EIGEN_DEVICE_FUNC inline Packet8f pgather(const float* from, int stride) { return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride], - from[3*stride], from[2*stride], from[1*stride], from[0*stride]); + from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } template<> EIGEN_DEVICE_FUNC inline Packet4d pgather(const double* from, int stride) { diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 694979e19..e54ebbf90 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -117,18 +117,15 @@ template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) { return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]), - std::imag(from[0*stride]), std::real(from[0*stride]))); + std::imag(from[0*stride]), std::real(from[0*stride]))); } template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) { - /* for (int i = 0; i < 2; i+=2) { - to[stride*i] = std::complex(from.v[i], from.v[i+1]); - }*/ to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)), - _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1))); + _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1))); to[stride*1] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)), - _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3))); + _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3))); } template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } From 82c81630679b44bd1a3f7842152f12179428c9f7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 31 Mar 2014 10:41:40 +0200 Subject: [PATCH 0406/1103] Enable repetition in mixing type unit test --- test/mixingtypes.cpp | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/test/mixingtypes.cpp b/test/mixingtypes.cpp index 6c2f74875..ada2f69d3 100644 --- a/test/mixingtypes.cpp +++ b/test/mixingtypes.cpp @@ -126,7 +126,9 @@ template void mixingtypes(int size = SizeAtCompileType) void test_mixingtypes() { - CALL_SUBTEST_1(mixingtypes<3>()); - CALL_SUBTEST_2(mixingtypes<4>()); - CALL_SUBTEST_3(mixingtypes(internal::random(1,EIGEN_TEST_MAX_SIZE))); + for(int i = 0; i < g_repeat; i++) { + CALL_SUBTEST_1(mixingtypes<3>()); + CALL_SUBTEST_2(mixingtypes<4>()); + CALL_SUBTEST_3(mixingtypes(internal::random(1,EIGEN_TEST_MAX_SIZE))); + } } From 8d0441052e7fac530fad12016f53f5b234a68d47 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 31 Mar 2014 10:42:19 +0200 Subject: [PATCH 0407/1103] Finally, prefetching seems to help getting more stable performance --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index a9e42c8aa..d9e659c9a 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -639,7 +639,7 @@ void gebp_kernel for(Index i=0; i for(Index i=peeled_mc; i { Index i = rows-1; const LhsScalar* blA = &blockA[i*strideA+offsetA]; + prefetch(&blA[0]); const RhsScalar* blB = &blockB[j2*strideB+offsetB*8]; EIGEN_ASM_COMMENT("begin_vectorized_multiplication_of_last_rows 8"); @@ -863,6 +865,7 @@ void gebp_kernel for(Index i=peeled_mc; i for(Index i=0; i for(Index i=peeled_mc; i traits.initAcc(C0); const LhsScalar* blA = &blockA[i*strideA+offsetA*mr]; + prefetch(&blA[0]); const RhsScalar* blB = &blockB[j2*strideB+offsetB]; for(Index k=0; k for(Index i=peeled_mc; i Date: Mon, 31 Mar 2014 10:58:30 +0200 Subject: [PATCH 0408/1103] BTL: fix warnings and extend to 5k matrices, update GotoBlas to OpenBlas, etc. --- bench/bench_gemm.cpp | 16 ++++-- bench/btl/CMakeLists.txt | 1 + bench/btl/actions/action_axpby.hh | 2 +- bench/btl/actions/action_axpy.hh | 2 +- bench/btl/cmake/FindATLAS.cmake | 13 ++--- bench/btl/cmake/FindGOTO.cmake | 15 ------ bench/btl/cmake/FindGOTO2.cmake | 25 ---------- bench/btl/cmake/FindOpenBLAS.cmake | 17 +++++++ bench/btl/data/action_settings.txt | 32 ++++++------ bench/btl/data/perlib_plot_settings.txt | 4 +- bench/btl/generic_bench/bench_parameter.hh | 4 +- bench/btl/generic_bench/btl.hh | 4 +- bench/btl/generic_bench/init/init_function.hh | 8 +-- bench/btl/generic_bench/utils/size_lin_log.hh | 2 +- bench/btl/libs/BLAS/CMakeLists.txt | 29 +++-------- bench/btl/libs/BLAS/blas_interface_impl.hh | 6 +-- bench/btl/libs/BLAS/c_interface_base.h | 6 +-- bench/btl/libs/STL/STL_interface.hh | 4 +- bench/btl/libs/eigen3/eigen3_interface.hh | 50 +++++++++---------- 19 files changed, 104 insertions(+), 136 deletions(-) delete mode 100644 bench/btl/cmake/FindGOTO.cmake delete mode 100644 bench/btl/cmake/FindGOTO2.cmake create mode 100644 bench/btl/cmake/FindOpenBLAS.cmake diff --git a/bench/bench_gemm.cpp b/bench/bench_gemm.cpp index 41ca8b3b6..1ef2e72c2 100644 --- a/bench/bench_gemm.cpp +++ b/bench/bench_gemm.cpp @@ -129,13 +129,25 @@ int main(int argc, char ** argv) int tries = 2; // number of tries, we keep the best int s = 2048; + int m = s; + int n = s; + int p = s; int cache_size = -1; bool need_help = false; for (int i=1; i0) setCpuCacheSizes(cache_size,96*cache_size); - int m = s; - int n = s; - int p = s; + A a(m,p); a.setRandom(); B b(p,n); b.setRandom(); C c(m,n); c.setOnes(); diff --git a/bench/btl/CMakeLists.txt b/bench/btl/CMakeLists.txt index 0ac5c15fc..b299d9899 100644 --- a/bench/btl/CMakeLists.txt +++ b/bench/btl/CMakeLists.txt @@ -104,6 +104,7 @@ add_subdirectory(libs/mtl4) add_subdirectory(libs/blitz) add_subdirectory(libs/tvmet) add_subdirectory(libs/STL) +add_subdirectory(libs/blaze) add_subdirectory(data) diff --git a/bench/btl/actions/action_axpby.hh b/bench/btl/actions/action_axpby.hh index 98511ab6a..dadd0ccf3 100644 --- a/bench/btl/actions/action_axpby.hh +++ b/bench/btl/actions/action_axpby.hh @@ -33,7 +33,7 @@ class Action_axpby { public : // Ctor - Action_axpby( int size ):_size(size),_alpha(0.5),_beta(0.95) + Action_axpby( int size ):_alpha(0.5),_beta(0.95),_size(size) { MESSAGE("Action_axpby Ctor"); diff --git a/bench/btl/actions/action_axpy.hh b/bench/btl/actions/action_axpy.hh index e4cb3a5bd..261be4cb8 100644 --- a/bench/btl/actions/action_axpy.hh +++ b/bench/btl/actions/action_axpy.hh @@ -35,7 +35,7 @@ public : // Ctor - Action_axpy( int size ):_size(size),_coef(1.0) + Action_axpy( int size ):_coef(1.0),_size(size) { MESSAGE("Action_axpy Ctor"); diff --git a/bench/btl/cmake/FindATLAS.cmake b/bench/btl/cmake/FindATLAS.cmake index 6b9065206..14b1dee09 100644 --- a/bench/btl/cmake/FindATLAS.cmake +++ b/bench/btl/cmake/FindATLAS.cmake @@ -4,10 +4,7 @@ if (ATLAS_LIBRARIES) endif (ATLAS_LIBRARIES) find_file(ATLAS_LIB libatlas.so.3 PATHS /usr/lib $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) -find_library(ATLAS_LIB atlas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) - -find_file(ATLAS_CBLAS libcblas.so.3 PATHS /usr/lib $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) -find_library(ATLAS_CBLAS cblas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) +find_library(ATLAS_LIB satlas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) find_file(ATLAS_LAPACK liblapack_atlas.so.3 PATHS /usr/lib $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) find_library(ATLAS_LAPACK lapack_atlas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) @@ -22,14 +19,14 @@ find_library(ATLAS_F77BLAS f77blas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) if(ATLAS_LIB AND ATLAS_CBLAS AND ATLAS_LAPACK AND ATLAS_F77BLAS) - set(ATLAS_LIBRARIES ${ATLAS_LAPACK} ${ATLAS_CBLAS} ${ATLAS_F77BLAS} ${ATLAS_LIB}) + set(ATLAS_LIBRARIES ${ATLAS_LAPACK} ${ATLAS_LIB}) # search the default lapack lib link to it find_file(ATLAS_REFERENCE_LAPACK liblapack.so.3 PATHS /usr/lib /usr/lib64) find_library(ATLAS_REFERENCE_LAPACK NAMES lapack) - if(ATLAS_REFERENCE_LAPACK) - set(ATLAS_LIBRARIES ${ATLAS_LIBRARIES} ${ATLAS_REFERENCE_LAPACK}) - endif() +# if(ATLAS_REFERENCE_LAPACK) +# set(ATLAS_LIBRARIES ${ATLAS_LIBRARIES} ${ATLAS_REFERENCE_LAPACK}) +# endif() endif(ATLAS_LIB AND ATLAS_CBLAS AND ATLAS_LAPACK AND ATLAS_F77BLAS) diff --git a/bench/btl/cmake/FindGOTO.cmake b/bench/btl/cmake/FindGOTO.cmake deleted file mode 100644 index 67ea0934a..000000000 --- a/bench/btl/cmake/FindGOTO.cmake +++ /dev/null @@ -1,15 +0,0 @@ - -if (GOTO_LIBRARIES) - set(GOTO_FIND_QUIETLY TRUE) -endif (GOTO_LIBRARIES) - -find_library(GOTO_LIBRARIES goto PATHS $ENV{GOTODIR} ${LIB_INSTALL_DIR}) - -if(GOTO_LIBRARIES AND CMAKE_COMPILER_IS_GNUCXX) - set(GOTO_LIBRARIES ${GOTO_LIBRARIES} "-lpthread -lgfortran") -endif() - -include(FindPackageHandleStandardArgs) -find_package_handle_standard_args(GOTO DEFAULT_MSG GOTO_LIBRARIES) - -mark_as_advanced(GOTO_LIBRARIES) diff --git a/bench/btl/cmake/FindGOTO2.cmake b/bench/btl/cmake/FindGOTO2.cmake deleted file mode 100644 index baa68d213..000000000 --- a/bench/btl/cmake/FindGOTO2.cmake +++ /dev/null @@ -1,25 +0,0 @@ - -if (GOTO2_LIBRARIES) - set(GOTO2_FIND_QUIETLY TRUE) -endif (GOTO2_LIBRARIES) -# -# find_path(GOTO_INCLUDES -# NAMES -# cblas.h -# PATHS -# $ENV{GOTODIR}/include -# ${INCLUDE_INSTALL_DIR} -# ) - -find_file(GOTO2_LIBRARIES libgoto2.so PATHS /usr/lib $ENV{GOTO2DIR} ${LIB_INSTALL_DIR}) -find_library(GOTO2_LIBRARIES goto2 PATHS $ENV{GOTO2DIR} ${LIB_INSTALL_DIR}) - -if(GOTO2_LIBRARIES AND CMAKE_COMPILER_IS_GNUCXX) - set(GOTO2_LIBRARIES ${GOTO2_LIBRARIES} "-lpthread -lgfortran") -endif() - -include(FindPackageHandleStandardArgs) -find_package_handle_standard_args(GOTO2 DEFAULT_MSG - GOTO2_LIBRARIES) - -mark_as_advanced(GOTO2_LIBRARIES) diff --git a/bench/btl/cmake/FindOpenBLAS.cmake b/bench/btl/cmake/FindOpenBLAS.cmake new file mode 100644 index 000000000..c76fc251c --- /dev/null +++ b/bench/btl/cmake/FindOpenBLAS.cmake @@ -0,0 +1,17 @@ + +if (OPENBLAS_LIBRARIES) + set(OPENBLAS_FIND_QUIETLY TRUE) +endif (OPENBLAS_LIBRARIES) + +find_file(OPENBLAS_LIBRARIES libopenblas.so PATHS /usr/lib $ENV{OPENBLASDIR} ${LIB_INSTALL_DIR}) +find_library(OPENBLAS_LIBRARIES openblas PATHS $ENV{OPENBLASDIR} ${LIB_INSTALL_DIR}) + +if(OPENBLAS_LIBRARIES AND CMAKE_COMPILER_IS_GNUCXX) + set(OPENBLAS_LIBRARIES ${OPENBLAS_LIBRARIES} "-lpthread -lgfortran") +endif() + +include(FindPackageHandleStandardArgs) +find_package_handle_standard_args(OPENBLAS DEFAULT_MSG + OPENBLAS_LIBRARIES) + +mark_as_advanced(OPENBLAS_LIBRARIES) diff --git a/bench/btl/data/action_settings.txt b/bench/btl/data/action_settings.txt index e32213e22..674a738d9 100644 --- a/bench/btl/data/action_settings.txt +++ b/bench/btl/data/action_settings.txt @@ -1,19 +1,19 @@ -aat ; "{/*1.5 A x A^T}" ; "matrix size" ; 4:3000 -ata ; "{/*1.5 A^T x A}" ; "matrix size" ; 4:3000 -atv ; "{/*1.5 matrix^T x vector}" ; "matrix size" ; 4:3000 +aat ; "{/*1.5 A x A^T}" ; "matrix size" ; 4:5000 +ata ; "{/*1.5 A^T x A}" ; "matrix size" ; 4:5000 +atv ; "{/*1.5 matrix^T x vector}" ; "matrix size" ; 4:5000 axpby ; "{/*1.5 Y = alpha X + beta Y}" ; "vector size" ; 5:1000000 axpy ; "{/*1.5 Y += alpha X}" ; "vector size" ; 5:1000000 -matrix_matrix ; "{/*1.5 matrix matrix product}" ; "matrix size" ; 4:3000 -matrix_vector ; "{/*1.5 matrix vector product}" ; "matrix size" ; 4:3000 -trmm ; "{/*1.5 triangular matrix matrix product}" ; "matrix size" ; 4:3000 -trisolve_vector ; "{/*1.5 triangular solver - vector (X = inv(L) X)}" ; "size" ; 4:3000 -trisolve_matrix ; "{/*1.5 triangular solver - matrix (M = inv(L) M)}" ; "size" ; 4:3000 -cholesky ; "{/*1.5 Cholesky decomposition}" ; "matrix size" ; 4:3000 -complete_lu_decomp ; "{/*1.5 Complete LU decomposition}" ; "matrix size" ; 4:3000 -partial_lu_decomp ; "{/*1.5 Partial LU decomposition}" ; "matrix size" ; 4:3000 -tridiagonalization ; "{/*1.5 Tridiagonalization}" ; "matrix size" ; 4:3000 -hessenberg ; "{/*1.5 Hessenberg decomposition}" ; "matrix size" ; 4:3000 -symv ; "{/*1.5 symmetric matrix vector product}" ; "matrix size" ; 4:3000 -syr2 ; "{/*1.5 symmretric rank-2 update (A += u^T v + u v^T)}" ; "matrix size" ; 4:3000 -ger ; "{/*1.5 general rank-1 update (A += u v^T)}" ; "matrix size" ; 4:3000 +matrix_matrix ; "{/*1.5 matrix matrix product}" ; "matrix size" ; 4:5000 +matrix_vector ; "{/*1.5 matrix vector product}" ; "matrix size" ; 4:5000 +trmm ; "{/*1.5 triangular matrix matrix product}" ; "matrix size" ; 4:5000 +trisolve_vector ; "{/*1.5 triangular solver - vector (X = inv(L) X)}" ; "size" ; 4:5000 +trisolve_matrix ; "{/*1.5 triangular solver - matrix (M = inv(L) M)}" ; "size" ; 4:5000 +cholesky ; "{/*1.5 Cholesky decomposition}" ; "matrix size" ; 4:5000 +complete_lu_decomp ; "{/*1.5 Complete LU decomposition}" ; "matrix size" ; 4:5000 +partial_lu_decomp ; "{/*1.5 Partial LU decomposition}" ; "matrix size" ; 4:5000 +tridiagonalization ; "{/*1.5 Tridiagonalization}" ; "matrix size" ; 4:5000 +hessenberg ; "{/*1.5 Hessenberg decomposition}" ; "matrix size" ; 4:5000 +symv ; "{/*1.5 symmetric matrix vector product}" ; "matrix size" ; 4:5000 +syr2 ; "{/*1.5 symmretric rank-2 update (A += u^T v + u v^T)}" ; "matrix size" ; 4:5000 +ger ; "{/*1.5 general rank-1 update (A += u v^T)}" ; "matrix size" ; 4:5000 rot ; "{/*1.5 apply rotation in the plane}" ; "vector size" ; 4:1000000 \ No newline at end of file diff --git a/bench/btl/data/perlib_plot_settings.txt b/bench/btl/data/perlib_plot_settings.txt index 6844bab28..f023cfe02 100644 --- a/bench/btl/data/perlib_plot_settings.txt +++ b/bench/btl/data/perlib_plot_settings.txt @@ -10,7 +10,7 @@ ublas ; with lines lw 3 lt 1 lc rgbcolor "#00b7ff" mtl4 ; with lines lw 3 lt 1 lc rgbcolor "#d18847" blitz ; with lines lw 3 lt 1 lc rgbcolor "#ff00ff" F77 ; with lines lw 3 lt 3 lc rgbcolor "#e6e64c" -GOTO ; with lines lw 3 lt 3 lc rgbcolor "#C05600" -GOTO2 ; with lines lw 3 lt 1 lc rgbcolor "#C05600" +OPENBLAS ; with lines lw 3 lt 1 lc rgbcolor "#C05600" C ; with lines lw 3 lt 3 lc rgbcolor "#e6bd96" ACML ; with lines lw 2 lt 3 lc rgbcolor "#e6e64c" +blaze ; with lines lw 3 lt 1 lc rgbcolor "#ff00ff" diff --git a/bench/btl/generic_bench/bench_parameter.hh b/bench/btl/generic_bench/bench_parameter.hh index 4c355cd6e..5e341c14c 100644 --- a/bench/btl/generic_bench/bench_parameter.hh +++ b/bench/btl/generic_bench/bench_parameter.hh @@ -33,7 +33,7 @@ // min matrix size for matrix vector product bench #define MIN_MV 5 // max matrix size for matrix vector product bench -#define MAX_MV 3000 +#define MAX_MV 5000 // min matrix size for matrix matrix product bench #define MIN_MM 5 // max matrix size for matrix matrix product bench @@ -41,7 +41,7 @@ // min matrix size for LU bench #define MIN_LU 5 // max matrix size for LU bench -#define MAX_LU 3000 +#define MAX_LU 5000 // max size for tiny vector and matrix #define TINY_MV_MAX_SIZE 16 // default nb_sample for x86 timer diff --git a/bench/btl/generic_bench/btl.hh b/bench/btl/generic_bench/btl.hh index 4670f5113..92af1306a 100644 --- a/bench/btl/generic_bench/btl.hh +++ b/bench/btl/generic_bench/btl.hh @@ -176,7 +176,7 @@ public: if (_config!=NULL) { std::vector config = BtlString(_config).split(" \t\n"); - for (int i = 0; i -void size_lin_log(const int nb_point, const int size_min, const int size_max, Vector & X) +void size_lin_log(const int nb_point, const int /*size_min*/, const int size_max, Vector & X) { int ten=10; int nine=9; diff --git a/bench/btl/libs/BLAS/CMakeLists.txt b/bench/btl/libs/BLAS/CMakeLists.txt index de42fe047..22f09527d 100644 --- a/bench/btl/libs/BLAS/CMakeLists.txt +++ b/bench/btl/libs/BLAS/CMakeLists.txt @@ -18,27 +18,14 @@ if (MKL_FOUND) endif (MKL_FOUND) -find_package(GOTO2) -if (GOTO2_FOUND) - btl_add_bench(btl_goto2 main.cpp) - if(BUILD_btl_goto2) - target_link_libraries(btl_goto2 ${GOTO_LIBRARIES} ) - set_target_properties(btl_goto2 PROPERTIES COMPILE_FLAGS "-DCBLASNAME=GOTO2") - endif(BUILD_btl_goto2) -endif (GOTO2_FOUND) - -find_package(GOTO) -if (GOTO_FOUND) - if(GOTO2_FOUND) - btl_add_bench(btl_goto main.cpp OFF) - else() - btl_add_bench(btl_goto main.cpp) - endif() - if(BUILD_btl_goto) - target_link_libraries(btl_goto ${GOTO_LIBRARIES} ) - set_target_properties(btl_goto PROPERTIES COMPILE_FLAGS "-DCBLASNAME=GOTO") - endif(BUILD_btl_goto) -endif (GOTO_FOUND) +find_package(OPENBLAS) +if (OPENBLAS_FOUND) + btl_add_bench(btl_openblas main.cpp) + if(BUILD_btl_openblas) + target_link_libraries(btl_openblas ${GOTO_LIBRARIES} ) + set_target_properties(btl_openblas PROPERTIES COMPILE_FLAGS "-DCBLASNAME=OPENBLAS") + endif(BUILD_btl_openblas) +endif (OPENBLAS_FOUND) find_package(ACML) if (ACML_FOUND) diff --git a/bench/btl/libs/BLAS/blas_interface_impl.hh b/bench/btl/libs/BLAS/blas_interface_impl.hh index 0e84df038..fc4ba2a1f 100644 --- a/bench/btl/libs/BLAS/blas_interface_impl.hh +++ b/bench/btl/libs/BLAS/blas_interface_impl.hh @@ -75,7 +75,6 @@ public : static inline void partial_lu_decomp(const gene_matrix & X, gene_matrix & C, int N){ int N2 = N*N; BLAS_FUNC(copy)(&N2, X, &intone, C, &intone); - char uplo = 'L'; int info = 0; int * ipiv = (int*)alloca(sizeof(int)*N); BLAS_FUNC(getrf)(&N, &N, C, &N, ipiv, &info); @@ -92,7 +91,7 @@ public : BLAS_FUNC(trsm)(&right, &lower, ¬rans, &nonunit, &N, &N, &fone, L, &N, X, &N); } - static inline void trmm(gene_matrix & A, gene_matrix & B, gene_matrix & X, int N){ + static inline void trmm(gene_matrix & A, gene_matrix & B, gene_matrix & /*X*/, int N){ BLAS_FUNC(trmm)(&left, &lower, ¬rans,&nonunit, &N,&N,&fone,A,&N,B,&N); } @@ -101,7 +100,6 @@ public : static inline void lu_decomp(const gene_matrix & X, gene_matrix & C, int N){ int N2 = N*N; BLAS_FUNC(copy)(&N2, X, &intone, C, &intone); - char uplo = 'L'; int info = 0; int * ipiv = (int*)alloca(sizeof(int)*N); int * jpiv = (int*)alloca(sizeof(int)*N); @@ -134,8 +132,6 @@ public : } char uplo = 'U'; int info = 0; - int ilo = 1; - int ihi = N; int bsize = 64; int worksize = N*bsize; SCALAR* d = new SCALAR[3*N+worksize]; diff --git a/bench/btl/libs/BLAS/c_interface_base.h b/bench/btl/libs/BLAS/c_interface_base.h index 515d8dcfc..de613803b 100644 --- a/bench/btl/libs/BLAS/c_interface_base.h +++ b/bench/btl/libs/BLAS/c_interface_base.h @@ -17,12 +17,12 @@ public: typedef real* gene_matrix; typedef real* gene_vector; - static void free_matrix(gene_matrix & A, int N){ - delete A; + static void free_matrix(gene_matrix & A, int /*N*/){ + delete[] A; } static void free_vector(gene_vector & B){ - delete B; + delete[] B; } static inline void matrix_from_stl(gene_matrix & A, stl_matrix & A_stl){ diff --git a/bench/btl/libs/STL/STL_interface.hh b/bench/btl/libs/STL/STL_interface.hh index 93e76bd55..ef4cc9233 100644 --- a/bench/btl/libs/STL/STL_interface.hh +++ b/bench/btl/libs/STL/STL_interface.hh @@ -44,9 +44,9 @@ public : return "STL"; } - static void free_matrix(gene_matrix & A, int N){} + static void free_matrix(gene_matrix & /*A*/, int /*N*/){} - static void free_vector(gene_vector & B){} + static void free_vector(gene_vector & /*B*/){} static inline void matrix_from_stl(gene_matrix & A, stl_matrix & A_stl){ A = A_stl; diff --git a/bench/btl/libs/eigen3/eigen3_interface.hh b/bench/btl/libs/eigen3/eigen3_interface.hh index 0c8aa74da..b821fd721 100644 --- a/bench/btl/libs/eigen3/eigen3_interface.hh +++ b/bench/btl/libs/eigen3/eigen3_interface.hh @@ -45,9 +45,9 @@ public : return EIGEN_MAKESTRING(BTL_PREFIX); } - static void free_matrix(gene_matrix & A, int N) {} + static void free_matrix(gene_matrix & /*A*/, int /*N*/) {} - static void free_vector(gene_vector & B) {} + static void free_vector(gene_vector & /*B*/) {} static BTL_DONT_INLINE void matrix_from_stl(gene_matrix & A, stl_matrix & A_stl){ A.resize(A_stl[0].size(), A_stl.size()); @@ -74,7 +74,7 @@ public : } static BTL_DONT_INLINE void matrix_to_stl(gene_matrix & A, stl_matrix & A_stl){ - int N=A_stl.size(); + int N=A_stl.size(); for (int j=0;j().setZero(); X.template selfadjointView().rankUpdate(A); } - static inline void matrix_vector_product(const gene_matrix & A, const gene_vector & B, gene_vector & X, int N){ + static inline void matrix_vector_product(const gene_matrix & A, const gene_vector & B, gene_vector & X, int /*N*/){ X.noalias() = A*B; } - static inline void symv(const gene_matrix & A, const gene_vector & B, gene_vector & X, int N){ + static inline void symv(const gene_matrix & A, const gene_vector & B, gene_vector & X, int /*N*/){ X.noalias() = (A.template selfadjointView() * B); // internal::product_selfadjoint_vector(N,A.data(),N, B.data(), 1, X.data(), 1); } @@ -155,54 +155,54 @@ public : } } - static EIGEN_DONT_INLINE void syr2(gene_matrix & A, gene_vector & X, gene_vector & Y, int N){ + static EIGEN_DONT_INLINE void syr2(gene_matrix & A, gene_vector & X, gene_vector & Y, int N){ // internal::product_selfadjoint_rank2_update(N,A.data(),N, X.data(), 1, Y.data(), 1, -1); for(int j=0; j(c,s)); } - static inline void atv_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){ + static inline void atv_product(gene_matrix & A, gene_vector & B, gene_vector & X, int /*N*/){ X.noalias() = (A.transpose()*B); } - static inline void axpy(real coef, const gene_vector & X, gene_vector & Y, int N){ + static inline void axpy(real coef, const gene_vector & X, gene_vector & Y, int /*N*/){ Y += coef * X; } - static inline void axpby(real a, const gene_vector & X, real b, gene_vector & Y, int N){ + static inline void axpby(real a, const gene_vector & X, real b, gene_vector & Y, int /*N*/){ Y = a*X + b*Y; } - static EIGEN_DONT_INLINE void copy_matrix(const gene_matrix & source, gene_matrix & cible, int N){ + static EIGEN_DONT_INLINE void copy_matrix(const gene_matrix & source, gene_matrix & cible, int /*N*/){ cible = source; } - static EIGEN_DONT_INLINE void copy_vector(const gene_vector & source, gene_vector & cible, int N){ + static EIGEN_DONT_INLINE void copy_vector(const gene_vector & source, gene_vector & cible, int /*N*/){ cible = source; } - static inline void trisolve_lower(const gene_matrix & L, const gene_vector& B, gene_vector& X, int N){ + static inline void trisolve_lower(const gene_matrix & L, const gene_vector& B, gene_vector& X, int /*N*/){ X = L.template triangularView().solve(B); } - static inline void trisolve_lower_matrix(const gene_matrix & L, const gene_matrix& B, gene_matrix& X, int N){ + static inline void trisolve_lower_matrix(const gene_matrix & L, const gene_matrix& B, gene_matrix& X, int /*N*/){ X = L.template triangularView().solve(B); } - static inline void trmm(const gene_matrix & L, const gene_matrix& B, gene_matrix& X, int N){ + static inline void trmm(const gene_matrix & L, const gene_matrix& B, gene_matrix& X, int /*N*/){ X.noalias() = L.template triangularView() * B; } - static inline void cholesky(const gene_matrix & X, gene_matrix & C, int N){ + static inline void cholesky(const gene_matrix & X, gene_matrix & C, int /*N*/){ C = X; internal::llt_inplace::blocked(C); //C = X.llt().matrixL(); @@ -211,11 +211,11 @@ public : // Cholesky::computeInPlaceBlock(C); } - static inline void lu_decomp(const gene_matrix & X, gene_matrix & C, int N){ + static inline void lu_decomp(const gene_matrix & X, gene_matrix & C, int /*N*/){ C = X.fullPivLu().matrixLU(); } - static inline void partial_lu_decomp(const gene_matrix & X, gene_matrix & C, int N){ + static inline void partial_lu_decomp(const gene_matrix & X, gene_matrix & C, int N){ Matrix piv(N); DenseIndex nb; C = X; @@ -223,13 +223,13 @@ public : // C = X.partialPivLu().matrixLU(); } - static inline void tridiagonalization(const gene_matrix & X, gene_matrix & C, int N){ + static inline void tridiagonalization(const gene_matrix & X, gene_matrix & C, int N){ typename Tridiagonalization::CoeffVectorType aux(N-1); C = X; internal::tridiagonalization_inplace(C, aux); } - static inline void hessenberg(const gene_matrix & X, gene_matrix & C, int N){ + static inline void hessenberg(const gene_matrix & X, gene_matrix & C, int /*N*/){ C = HessenbergDecomposition(X).packedMatrix(); } From 93870d95b7b8a7716ab825d559e7af5f7f84308a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 31 Mar 2014 10:59:55 +0200 Subject: [PATCH 0409/1103] BTL: add blaze --- bench/btl/cmake/FindBLAZE.cmake | 31 ++++++ bench/btl/libs/blaze/CMakeLists.txt | 8 ++ bench/btl/libs/blaze/blaze_interface.hh | 140 ++++++++++++++++++++++++ bench/btl/libs/blaze/main.cpp | 46 ++++++++ 4 files changed, 225 insertions(+) create mode 100644 bench/btl/cmake/FindBLAZE.cmake create mode 100644 bench/btl/libs/blaze/CMakeLists.txt create mode 100644 bench/btl/libs/blaze/blaze_interface.hh create mode 100644 bench/btl/libs/blaze/main.cpp diff --git a/bench/btl/cmake/FindBLAZE.cmake b/bench/btl/cmake/FindBLAZE.cmake new file mode 100644 index 000000000..dba4c89f2 --- /dev/null +++ b/bench/btl/cmake/FindBLAZE.cmake @@ -0,0 +1,31 @@ +# - Try to find eigen2 headers +# Once done this will define +# +# BLAZE_FOUND - system has blaze lib +# BLAZE_INCLUDE_DIR - the blaze include directory +# +# Copyright (C) 2008 Gael Guennebaud +# Adapted from FindEigen.cmake: +# Copyright (c) 2006, 2007 Montel Laurent, +# Redistribution and use is allowed according to the terms of the BSD license. +# For details see the accompanying COPYING-CMAKE-SCRIPTS file. + +if (BLAZE_INCLUDE_DIR) + + # in cache already + set(BLAZE_FOUND TRUE) + +else (BLAZE_INCLUDE_DIR) + +find_path(BLAZE_INCLUDE_DIR NAMES blaze/Blaze.h + PATHS + ${INCLUDE_INSTALL_DIR} + ) + +include(FindPackageHandleStandardArgs) +find_package_handle_standard_args(BLAZE DEFAULT_MSG BLAZE_INCLUDE_DIR) + +mark_as_advanced(BLAZE_INCLUDE_DIR) + +endif(BLAZE_INCLUDE_DIR) + diff --git a/bench/btl/libs/blaze/CMakeLists.txt b/bench/btl/libs/blaze/CMakeLists.txt new file mode 100644 index 000000000..54ab929d8 --- /dev/null +++ b/bench/btl/libs/blaze/CMakeLists.txt @@ -0,0 +1,8 @@ + +find_package(BLAZE) +find_package(Boost) +if (BLAZE_FOUND AND Boost_FOUND) + include_directories(${BLAZE_INCLUDE_DIR} ${Boost_INCLUDE_DIRS}) + btl_add_bench(btl_blaze main.cpp) + target_link_libraries(btl_blaze ${Boost_LIBRARIES} ${Boost_system_LIBRARY} /opt/local/lib/libboost_system-mt.a ) +endif (BLAZE_FOUND) diff --git a/bench/btl/libs/blaze/blaze_interface.hh b/bench/btl/libs/blaze/blaze_interface.hh new file mode 100644 index 000000000..8020fef27 --- /dev/null +++ b/bench/btl/libs/blaze/blaze_interface.hh @@ -0,0 +1,140 @@ +//===================================================== +// Copyright (C) 2008 Gael Guennebaud +//===================================================== +// +// This program is free software; you can redistribute it and/or +// modify it under the terms of the GNU General Public License +// as published by the Free Software Foundation; either version 2 +// of the License, or (at your option) any later version. +// +// This program is distributed in the hope that it will be useful, +// but WITHOUT ANY WARRANTY; without even the implied warranty of +// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +// GNU General Public License for more details. +// You should have received a copy of the GNU General Public License +// along with this program; if not, write to the Free Software +// Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. +// +#ifndef BLAZE_INTERFACE_HH +#define BLAZE_INTERFACE_HH + +#include +#include +// using namespace blaze; + +#include + +template +class blaze_interface { + +public : + + typedef real real_type ; + + typedef std::vector stl_vector; + typedef std::vector stl_matrix; + + typedef blaze::DynamicMatrix gene_matrix; + typedef blaze::DynamicVector gene_vector; + + static inline std::string name() { return "blaze"; } + + static void free_matrix(gene_matrix & A, int N){ + return ; + } + + static void free_vector(gene_vector & B){ + return ; + } + + static inline void matrix_from_stl(gene_matrix & A, stl_matrix & A_stl){ + A.resize(A_stl[0].size(), A_stl.size()); + + for (int j=0; j ipvt(N); +// lu_factor(R, ipvt); +// } + +// static inline void trisolve_lower(const gene_matrix & L, const gene_vector& B, gene_vector & X, int N){ +// X = lower_trisolve(L, B); +// } + + static inline void copy_matrix(const gene_matrix & source, gene_matrix & cible, int N){ + cible = source; + } + + static inline void copy_vector(const gene_vector & source, gene_vector & cible, int N){ + cible = source; + } + +}; + +#endif diff --git a/bench/btl/libs/blaze/main.cpp b/bench/btl/libs/blaze/main.cpp new file mode 100644 index 000000000..b8508c8f3 --- /dev/null +++ b/bench/btl/libs/blaze/main.cpp @@ -0,0 +1,46 @@ +//===================================================== +// Copyright (C) 2008 Gael Guennebaud +//===================================================== +// +// This program is free software; you can redistribute it and/or +// modify it under the terms of the GNU General Public License +// as published by the Free Software Foundation; either version 2 +// of the License, or (at your option) any later version. +// +// This program is distributed in the hope that it will be useful, +// but WITHOUT ANY WARRANTY; without even the implied warranty of +// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +// GNU General Public License for more details. +// You should have received a copy of the GNU General Public License +// along with this program; if not, write to the Free Software +// Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. +// +#include "utilities.h" +#include "blaze_interface.hh" +#include "bench.hh" +#include "basic_actions.hh" +// #include "action_cholesky.hh" +// #include "action_lu_decomp.hh" + +BTL_MAIN; + +int main() +{ + + bench > >(MIN_AXPY,MAX_AXPY,NB_POINT); + bench > >(MIN_AXPY,MAX_AXPY,NB_POINT); + + bench > >(MIN_MV,MAX_MV,NB_POINT); +// bench > >(MIN_MV,MAX_MV,NB_POINT); + bench > >(MIN_MM,MAX_MM,NB_POINT); +// bench > >(MIN_MM,MAX_MM,NB_POINT); +// bench > >(MIN_MM,MAX_MM,NB_POINT); + +// bench > >(MIN_MM,MAX_MM,NB_POINT); +// bench > >(MIN_MM,MAX_MM,NB_POINT); +// bench > >(MIN_MM,MAX_MM,NB_POINT); + + return 0; +} + + From 1221dd90aad41fdd610bed108c5f2f7af283dba7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Apr 2014 11:21:14 +0200 Subject: [PATCH 0410/1103] Fix no newline at end of file warning --- bench/spbench/spbenchstyle.h | 3 ++- unsupported/Eigen/src/SparseExtra/MarketIO.h | 6 +++--- 2 files changed, 5 insertions(+), 4 deletions(-) diff --git a/bench/spbench/spbenchstyle.h b/bench/spbench/spbenchstyle.h index 17a05ce71..f6a981778 100644 --- a/bench/spbench/spbenchstyle.h +++ b/bench/spbench/spbenchstyle.h @@ -91,4 +91,5 @@ void printBenchStyle(std::ofstream& out) \n\n"; } -#endif \ No newline at end of file + +#endif diff --git a/unsupported/Eigen/src/SparseExtra/MarketIO.h b/unsupported/Eigen/src/SparseExtra/MarketIO.h index 7aafce928..1c40d3f7c 100644 --- a/unsupported/Eigen/src/SparseExtra/MarketIO.h +++ b/unsupported/Eigen/src/SparseExtra/MarketIO.h @@ -238,9 +238,9 @@ bool saveMarket(const SparseMatrixType& mat, const std::string& filename, int sy for(int j=0; j Date: Tue, 1 Apr 2014 11:23:28 +0200 Subject: [PATCH 0411/1103] Rename the vector() factories defined in blas/common.h into make_vector() to prevent a possible name conflict with std::vector. --- blas/common.h | 12 +++---- blas/double.cpp | 11 +++--- blas/level1_cplx_impl.h | 38 ++++++++++---------- blas/level1_impl.h | 39 ++++++++++----------- blas/level1_real_impl.h | 22 ++++++------ blas/level2_cplx_impl.h | 4 +-- blas/level2_impl.h | 77 ++++++++++++++++++++--------------------- blas/level2_real_impl.h | 8 ++--- 8 files changed, 103 insertions(+), 108 deletions(-) diff --git a/blas/common.h b/blas/common.h index 2bf642c6b..c39cc63a8 100644 --- a/blas/common.h +++ b/blas/common.h @@ -106,13 +106,13 @@ matrix(T* data, int rows, int cols, int stride) } template -Map, 0, InnerStride > vector(T* data, int size, int incr) +Map, 0, InnerStride > make_vector(T* data, int size, int incr) { return Map, 0, InnerStride >(data, size, InnerStride(incr)); } template -Map > vector(T* data, int size) +Map > make_vector(T* data, int size) { return Map >(data, size); } @@ -124,8 +124,8 @@ T* get_compact_vector(T* x, int n, int incx) return x; T* ret = new Scalar[n]; - if(incx<0) vector(ret,n) = vector(x,n,-incx).reverse(); - else vector(ret,n) = vector(x,n, incx); + if(incx<0) make_vector(ret,n) = make_vector(x,n,-incx).reverse(); + else make_vector(ret,n) = make_vector(x,n, incx); return ret; } @@ -135,8 +135,8 @@ T* copy_back(T* x_cpy, T* x, int n, int incx) if(x_cpy==x) return 0; - if(incx<0) vector(x,n,-incx).reverse() = vector(x_cpy,n); - else vector(x,n, incx) = vector(x_cpy,n); + if(incx<0) make_vector(x,n,-incx).reverse() = make_vector(x_cpy,n); + else make_vector(x,n, incx) = make_vector(x_cpy,n); return x_cpy; } diff --git a/blas/double.cpp b/blas/double.cpp index 8fd0709ba..295b1d1f2 100644 --- a/blas/double.cpp +++ b/blas/double.cpp @@ -23,11 +23,10 @@ double BLASFUNC(dsdot)(int* n, float* x, int* incx, float* y, int* incy) { if(*n<=0) return 0; - if(*incx==1 && *incy==1) return (vector(x,*n).cast().cwiseProduct(vector(y,*n).cast())).sum(); - else if(*incx>0 && *incy>0) return (vector(x,*n,*incx).cast().cwiseProduct(vector(y,*n,*incy).cast())).sum(); - else if(*incx<0 && *incy>0) return (vector(x,*n,-*incx).reverse().cast().cwiseProduct(vector(y,*n,*incy).cast())).sum(); - else if(*incx>0 && *incy<0) return (vector(x,*n,*incx).cast().cwiseProduct(vector(y,*n,-*incy).reverse().cast())).sum(); - else if(*incx<0 && *incy<0) return (vector(x,*n,-*incx).reverse().cast().cwiseProduct(vector(y,*n,-*incy).reverse().cast())).sum(); + if(*incx==1 && *incy==1) return (make_vector(x,*n).cast().cwiseProduct(make_vector(y,*n).cast())).sum(); + else if(*incx>0 && *incy>0) return (make_vector(x,*n,*incx).cast().cwiseProduct(make_vector(y,*n,*incy).cast())).sum(); + else if(*incx<0 && *incy>0) return (make_vector(x,*n,-*incx).reverse().cast().cwiseProduct(make_vector(y,*n,*incy).cast())).sum(); + else if(*incx>0 && *incy<0) return (make_vector(x,*n,*incx).cast().cwiseProduct(make_vector(y,*n,-*incy).reverse().cast())).sum(); + else if(*incx<0 && *incy<0) return (make_vector(x,*n,-*incx).reverse().cast().cwiseProduct(make_vector(y,*n,-*incy).reverse().cast())).sum(); else return 0; } - diff --git a/blas/level1_cplx_impl.h b/blas/level1_cplx_impl.h index ffe192481..719f5bac9 100644 --- a/blas/level1_cplx_impl.h +++ b/blas/level1_cplx_impl.h @@ -32,13 +32,14 @@ RealScalar EIGEN_CAT(EIGEN_CAT(REAL_SCALAR_SUFFIX,SCALAR_SUFFIX),asum_)(int *n, if(*n<=0) return 0; - if(*incx==1) return vector(x,*n).unaryExpr().sum(); - else return vector(x,*n,std::abs(*incx)).unaryExpr().sum(); + if(*incx==1) return make_vector(x,*n).unaryExpr().sum(); + else return make_vector(x,*n,std::abs(*incx)).unaryExpr().sum(); } // computes a dot product of a conjugated vector with another vector. int EIGEN_BLAS_FUNC(dotcw)(int *n, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar* pres) { +// std::cerr << "_dotc " << *n << " " << *incx << " " << *incy << "\n"; Scalar* res = reinterpret_cast(pres); if(*n<=0) @@ -50,11 +51,11 @@ int EIGEN_BLAS_FUNC(dotcw)(int *n, RealScalar *px, int *incx, RealScalar *py, in Scalar* x = reinterpret_cast(px); Scalar* y = reinterpret_cast(py); - if(*incx==1 && *incy==1) *res = (vector(x,*n).dot(vector(y,*n))); - else if(*incx>0 && *incy>0) *res = (vector(x,*n,*incx).dot(vector(y,*n,*incy))); - else if(*incx<0 && *incy>0) *res = (vector(x,*n,-*incx).reverse().dot(vector(y,*n,*incy))); - else if(*incx>0 && *incy<0) *res = (vector(x,*n,*incx).dot(vector(y,*n,-*incy).reverse())); - else if(*incx<0 && *incy<0) *res = (vector(x,*n,-*incx).reverse().dot(vector(y,*n,-*incy).reverse())); + if(*incx==1 && *incy==1) *res = (make_vector(x,*n).dot(make_vector(y,*n))); + else if(*incx>0 && *incy>0) *res = (make_vector(x,*n,*incx).dot(make_vector(y,*n,*incy))); + else if(*incx<0 && *incy>0) *res = (make_vector(x,*n,-*incx).reverse().dot(make_vector(y,*n,*incy))); + else if(*incx>0 && *incy<0) *res = (make_vector(x,*n,*incx).dot(make_vector(y,*n,-*incy).reverse())); + else if(*incx<0 && *incy<0) *res = (make_vector(x,*n,-*incx).reverse().dot(make_vector(y,*n,-*incy).reverse())); return 0; } @@ -72,11 +73,11 @@ int EIGEN_BLAS_FUNC(dotuw)(int *n, RealScalar *px, int *incx, RealScalar *py, in Scalar* x = reinterpret_cast(px); Scalar* y = reinterpret_cast(py); - if(*incx==1 && *incy==1) *res = (vector(x,*n).cwiseProduct(vector(y,*n))).sum(); - else if(*incx>0 && *incy>0) *res = (vector(x,*n,*incx).cwiseProduct(vector(y,*n,*incy))).sum(); - else if(*incx<0 && *incy>0) *res = (vector(x,*n,-*incx).reverse().cwiseProduct(vector(y,*n,*incy))).sum(); - else if(*incx>0 && *incy<0) *res = (vector(x,*n,*incx).cwiseProduct(vector(y,*n,-*incy).reverse())).sum(); - else if(*incx<0 && *incy<0) *res = (vector(x,*n,-*incx).reverse().cwiseProduct(vector(y,*n,-*incy).reverse())).sum(); + if(*incx==1 && *incy==1) *res = (make_vector(x,*n).cwiseProduct(make_vector(y,*n))).sum(); + else if(*incx>0 && *incy>0) *res = (make_vector(x,*n,*incx).cwiseProduct(make_vector(y,*n,*incy))).sum(); + else if(*incx<0 && *incy>0) *res = (make_vector(x,*n,-*incx).reverse().cwiseProduct(make_vector(y,*n,*incy))).sum(); + else if(*incx>0 && *incy<0) *res = (make_vector(x,*n,*incx).cwiseProduct(make_vector(y,*n,-*incy).reverse())).sum(); + else if(*incx<0 && *incy<0) *res = (make_vector(x,*n,-*incx).reverse().cwiseProduct(make_vector(y,*n,-*incy).reverse())).sum(); return 0; } @@ -88,9 +89,9 @@ RealScalar EIGEN_CAT(EIGEN_CAT(REAL_SCALAR_SUFFIX,SCALAR_SUFFIX),nrm2_)(int *n, Scalar* x = reinterpret_cast(px); if(*incx==1) - return vector(x,*n).stableNorm(); + return make_vector(x,*n).stableNorm(); - return vector(x,*n,*incx).stableNorm(); + return make_vector(x,*n,*incx).stableNorm(); } int EIGEN_CAT(EIGEN_CAT(SCALAR_SUFFIX,REAL_SCALAR_SUFFIX),rot_)(int *n, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pc, RealScalar *ps) @@ -102,8 +103,8 @@ int EIGEN_CAT(EIGEN_CAT(SCALAR_SUFFIX,REAL_SCALAR_SUFFIX),rot_)(int *n, RealScal RealScalar c = *pc; RealScalar s = *ps; - StridedVectorType vx(vector(x,*n,std::abs(*incx))); - StridedVectorType vy(vector(y,*n,std::abs(*incy))); + StridedVectorType vx(make_vector(x,*n,std::abs(*incx))); + StridedVectorType vy(make_vector(y,*n,std::abs(*incy))); Reverse rvx(vx); Reverse rvy(vy); @@ -125,9 +126,8 @@ int EIGEN_CAT(EIGEN_CAT(SCALAR_SUFFIX,REAL_SCALAR_SUFFIX),scal_)(int *n, RealSca // std::cerr << "__scal " << *n << " " << alpha << " " << *incx << "\n"; - if(*incx==1) vector(x,*n) *= alpha; - else vector(x,*n,std::abs(*incx)) *= alpha; + if(*incx==1) make_vector(x,*n) *= alpha; + else make_vector(x,*n,std::abs(*incx)) *= alpha; return 0; } - diff --git a/blas/level1_impl.h b/blas/level1_impl.h index b08c2f6be..98e4c0963 100644 --- a/blas/level1_impl.h +++ b/blas/level1_impl.h @@ -17,11 +17,11 @@ int EIGEN_BLAS_FUNC(axpy)(int *n, RealScalar *palpha, RealScalar *px, int *incx, if(*n<=0) return 0; - if(*incx==1 && *incy==1) vector(y,*n) += alpha * vector(x,*n); - else if(*incx>0 && *incy>0) vector(y,*n,*incy) += alpha * vector(x,*n,*incx); - else if(*incx>0 && *incy<0) vector(y,*n,-*incy).reverse() += alpha * vector(x,*n,*incx); - else if(*incx<0 && *incy>0) vector(y,*n,*incy) += alpha * vector(x,*n,-*incx).reverse(); - else if(*incx<0 && *incy<0) vector(y,*n,-*incy).reverse() += alpha * vector(x,*n,-*incx).reverse(); + if(*incx==1 && *incy==1) make_vector(y,*n) += alpha * make_vector(x,*n); + else if(*incx>0 && *incy>0) make_vector(y,*n,*incy) += alpha * make_vector(x,*n,*incx); + else if(*incx>0 && *incy<0) make_vector(y,*n,-*incy).reverse() += alpha * make_vector(x,*n,*incx); + else if(*incx<0 && *incy>0) make_vector(y,*n,*incy) += alpha * make_vector(x,*n,-*incx).reverse(); + else if(*incx<0 && *incy<0) make_vector(y,*n,-*incy).reverse() += alpha * make_vector(x,*n,-*incx).reverse(); return 0; } @@ -35,7 +35,7 @@ int EIGEN_BLAS_FUNC(copy)(int *n, RealScalar *px, int *incx, RealScalar *py, int // be carefull, *incx==0 is allowed !! if(*incx==1 && *incy==1) - vector(y,*n) = vector(x,*n); + make_vector(y,*n) = make_vector(x,*n); else { if(*incx<0) x = x - (*n-1)*(*incx); @@ -57,8 +57,8 @@ int EIGEN_CAT(EIGEN_CAT(i,SCALAR_SUFFIX),amax_)(int *n, RealScalar *px, int *inc Scalar* x = reinterpret_cast(px); DenseIndex ret; - if(*incx==1) vector(x,*n).cwiseAbs().maxCoeff(&ret); - else vector(x,*n,std::abs(*incx)).cwiseAbs().maxCoeff(&ret); + if(*incx==1) make_vector(x,*n).cwiseAbs().maxCoeff(&ret); + else make_vector(x,*n,std::abs(*incx)).cwiseAbs().maxCoeff(&ret); return ret+1; } @@ -66,10 +66,10 @@ int EIGEN_CAT(EIGEN_CAT(i,SCALAR_SUFFIX),amin_)(int *n, RealScalar *px, int *inc { if(*n<=0) return 0; Scalar* x = reinterpret_cast(px); - + DenseIndex ret; - if(*incx==1) vector(x,*n).cwiseAbs().minCoeff(&ret); - else vector(x,*n,std::abs(*incx)).cwiseAbs().minCoeff(&ret); + if(*incx==1) make_vector(x,*n).cwiseAbs().minCoeff(&ret); + else make_vector(x,*n,std::abs(*incx)).cwiseAbs().minCoeff(&ret); return ret+1; } @@ -77,7 +77,7 @@ int EIGEN_BLAS_FUNC(rotg)(RealScalar *pa, RealScalar *pb, RealScalar *pc, RealSc { using std::sqrt; using std::abs; - + Scalar& a = *reinterpret_cast(pa); Scalar& b = *reinterpret_cast(pb); RealScalar* c = pc; @@ -143,8 +143,8 @@ int EIGEN_BLAS_FUNC(scal)(int *n, RealScalar *palpha, RealScalar *px, int *incx) Scalar* x = reinterpret_cast(px); Scalar alpha = *reinterpret_cast(palpha); - if(*incx==1) vector(x,*n) *= alpha; - else vector(x,*n,std::abs(*incx)) *= alpha; + if(*incx==1) make_vector(x,*n) *= alpha; + else make_vector(x,*n,std::abs(*incx)) *= alpha; return 0; } @@ -156,12 +156,11 @@ int EIGEN_BLAS_FUNC(swap)(int *n, RealScalar *px, int *incx, RealScalar *py, int Scalar* x = reinterpret_cast(px); Scalar* y = reinterpret_cast(py); - if(*incx==1 && *incy==1) vector(y,*n).swap(vector(x,*n)); - else if(*incx>0 && *incy>0) vector(y,*n,*incy).swap(vector(x,*n,*incx)); - else if(*incx>0 && *incy<0) vector(y,*n,-*incy).reverse().swap(vector(x,*n,*incx)); - else if(*incx<0 && *incy>0) vector(y,*n,*incy).swap(vector(x,*n,-*incx).reverse()); - else if(*incx<0 && *incy<0) vector(y,*n,-*incy).reverse().swap(vector(x,*n,-*incx).reverse()); + if(*incx==1 && *incy==1) make_vector(y,*n).swap(make_vector(x,*n)); + else if(*incx>0 && *incy>0) make_vector(y,*n,*incy).swap(make_vector(x,*n,*incx)); + else if(*incx>0 && *incy<0) make_vector(y,*n,-*incy).reverse().swap(make_vector(x,*n,*incx)); + else if(*incx<0 && *incy>0) make_vector(y,*n,*incy).swap(make_vector(x,*n,-*incx).reverse()); + else if(*incx<0 && *incy<0) make_vector(y,*n,-*incy).reverse().swap(make_vector(x,*n,-*incx).reverse()); return 1; } - diff --git a/blas/level1_real_impl.h b/blas/level1_real_impl.h index 8acecdfc6..02586d519 100644 --- a/blas/level1_real_impl.h +++ b/blas/level1_real_impl.h @@ -19,8 +19,8 @@ RealScalar EIGEN_BLAS_FUNC(asum)(int *n, RealScalar *px, int *incx) if(*n<=0) return 0; - if(*incx==1) return vector(x,*n).cwiseAbs().sum(); - else return vector(x,*n,std::abs(*incx)).cwiseAbs().sum(); + if(*incx==1) return make_vector(x,*n).cwiseAbs().sum(); + else return make_vector(x,*n,std::abs(*incx)).cwiseAbs().sum(); } // computes a vector-vector dot product. @@ -33,11 +33,11 @@ Scalar EIGEN_BLAS_FUNC(dot)(int *n, RealScalar *px, int *incx, RealScalar *py, i Scalar* x = reinterpret_cast(px); Scalar* y = reinterpret_cast(py); - if(*incx==1 && *incy==1) return (vector(x,*n).cwiseProduct(vector(y,*n))).sum(); - else if(*incx>0 && *incy>0) return (vector(x,*n,*incx).cwiseProduct(vector(y,*n,*incy))).sum(); - else if(*incx<0 && *incy>0) return (vector(x,*n,-*incx).reverse().cwiseProduct(vector(y,*n,*incy))).sum(); - else if(*incx>0 && *incy<0) return (vector(x,*n,*incx).cwiseProduct(vector(y,*n,-*incy).reverse())).sum(); - else if(*incx<0 && *incy<0) return (vector(x,*n,-*incx).reverse().cwiseProduct(vector(y,*n,-*incy).reverse())).sum(); + if(*incx==1 && *incy==1) return (make_vector(x,*n).cwiseProduct(make_vector(y,*n))).sum(); + else if(*incx>0 && *incy>0) return (make_vector(x,*n,*incx).cwiseProduct(make_vector(y,*n,*incy))).sum(); + else if(*incx<0 && *incy>0) return (make_vector(x,*n,-*incx).reverse().cwiseProduct(make_vector(y,*n,*incy))).sum(); + else if(*incx>0 && *incy<0) return (make_vector(x,*n,*incx).cwiseProduct(make_vector(y,*n,-*incy).reverse())).sum(); + else if(*incx<0 && *incy<0) return (make_vector(x,*n,-*incx).reverse().cwiseProduct(make_vector(y,*n,-*incy).reverse())).sum(); else return 0; } @@ -50,8 +50,8 @@ Scalar EIGEN_BLAS_FUNC(nrm2)(int *n, RealScalar *px, int *incx) Scalar* x = reinterpret_cast(px); - if(*incx==1) return vector(x,*n).stableNorm(); - else return vector(x,*n,std::abs(*incx)).stableNorm(); + if(*incx==1) return make_vector(x,*n).stableNorm(); + else return make_vector(x,*n,std::abs(*incx)).stableNorm(); } int EIGEN_BLAS_FUNC(rot)(int *n, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pc, RealScalar *ps) @@ -64,8 +64,8 @@ int EIGEN_BLAS_FUNC(rot)(int *n, RealScalar *px, int *incx, RealScalar *py, int Scalar c = *reinterpret_cast(pc); Scalar s = *reinterpret_cast(ps); - StridedVectorType vx(vector(x,*n,std::abs(*incx))); - StridedVectorType vy(vector(y,*n,std::abs(*incy))); + StridedVectorType vx(make_vector(x,*n,std::abs(*incx))); + StridedVectorType vy(make_vector(y,*n,std::abs(*incy))); Reverse rvx(vx); Reverse rvy(vy); diff --git a/blas/level2_cplx_impl.h b/blas/level2_cplx_impl.h index b850b6cd1..afa9a7493 100644 --- a/blas/level2_cplx_impl.h +++ b/blas/level2_cplx_impl.h @@ -57,8 +57,8 @@ int EIGEN_BLAS_FUNC(hemv)(char *uplo, int *n, RealScalar *palpha, RealScalar *pa if(beta!=Scalar(1)) { - if(beta==Scalar(0)) vector(actual_y, *n).setZero(); - else vector(actual_y, *n) *= beta; + if(beta==Scalar(0)) make_vector(actual_y, *n).setZero(); + else make_vector(actual_y, *n) *= beta; } if(alpha!=Scalar(0)) diff --git a/blas/level2_impl.h b/blas/level2_impl.h index 5f3941975..233c7b753 100644 --- a/blas/level2_impl.h +++ b/blas/level2_impl.h @@ -58,8 +58,8 @@ int EIGEN_BLAS_FUNC(gemv)(char *opa, int *m, int *n, RealScalar *palpha, RealSca if(beta!=Scalar(1)) { - if(beta==Scalar(0)) vector(actual_c, actual_m).setZero(); - else vector(actual_c, actual_m) *= beta; + if(beta==Scalar(0)) make_vector(actual_c, actual_m).setZero(); + else make_vector(actual_c, actual_m) *= beta; } if(code>=4 || func[code]==0) @@ -206,7 +206,7 @@ int EIGEN_BLAS_FUNC(gbmv)(char *trans, int *m, int *n, int *kl, int *ku, RealSca Scalar alpha = *reinterpret_cast(palpha); Scalar beta = *reinterpret_cast(pbeta); int coeff_rows = *kl+*ku+1; - + int info = 0; if(OP(*trans)==INVALID) info = 1; else if(*m<0) info = 2; @@ -218,26 +218,26 @@ int EIGEN_BLAS_FUNC(gbmv)(char *trans, int *m, int *n, int *kl, int *ku, RealSca else if(*incy==0) info = 13; if(info) return xerbla_(SCALAR_SUFFIX_UP"GBMV ",&info,6); - + if(*m==0 || *n==0 || (alpha==Scalar(0) && beta==Scalar(1))) return 0; - + int actual_m = *m; int actual_n = *n; if(OP(*trans)!=NOTR) std::swap(actual_m,actual_n); - + Scalar* actual_x = get_compact_vector(x,actual_n,*incx); Scalar* actual_y = get_compact_vector(y,actual_m,*incy); - + if(beta!=Scalar(1)) { - if(beta==Scalar(0)) vector(actual_y, actual_m).setZero(); - else vector(actual_y, actual_m) *= beta; + if(beta==Scalar(0)) make_vector(actual_y, actual_m).setZero(); + else make_vector(actual_y, actual_m) *= beta; } - + MatrixType mat_coeffs(a,coeff_rows,*n,*lda); - + int nb = std::min(*n,(*m)+(*ku)); for(int j=0; j(pa); Scalar* x = reinterpret_cast(px); int coeff_rows = *k + 1; - + int info = 0; if(UPLO(*uplo)==INVALID) info = 1; else if(OP(*opa)==INVALID) info = 2; @@ -283,37 +283,37 @@ int EIGEN_BLAS_FUNC(tbmv)(char *uplo, char *opa, char *diag, int *n, int *k, Rea else if(*incx==0) info = 9; if(info) return xerbla_(SCALAR_SUFFIX_UP"TBMV ",&info,6); - + if(*n==0) return 0; - + int actual_n = *n; - + Scalar* actual_x = get_compact_vector(x,actual_n,*incx); - + MatrixType mat_coeffs(a,coeff_rows,*n,*lda); - + int ku = UPLO(*uplo)==UPPER ? *k : 0; int kl = UPLO(*uplo)==LOWER ? *k : 0; - + for(int j=0; j<*n; ++j) { int start = std::max(0,j - ku); int end = std::min((*m)-1,j + kl); int len = end - start + 1; int offset = (ku) - j + start; - + if(OP(*trans)==NOTR) - vector(actual_y+start,len) += (alpha*actual_x[j]) * mat_coeffs.col(j).segment(offset,len); + make_vector(actual_y+start,len) += (alpha*actual_x[j]) * mat_coeffs.col(j).segment(offset,len); else if(OP(*trans)==TR) - actual_y[j] += alpha * ( mat_coeffs.col(j).segment(offset,len).transpose() * vector(actual_x+start,len) ).value(); + actual_y[j] += alpha * ( mat_coeffs.col(j).segment(offset,len).transpose() * make_vector(actual_x+start,len) ).value(); else - actual_y[j] += alpha * ( mat_coeffs.col(j).segment(offset,len).adjoint() * vector(actual_x+start,len) ).value(); - } - + actual_y[j] += alpha * ( mat_coeffs.col(j).segment(offset,len).adjoint() * make_vector(actual_x+start,len) ).value(); + } + if(actual_x!=x) delete[] actual_x; if(actual_y!=y) delete[] copy_back(actual_y,y,actual_m,*incy); - + return 0; } #endif @@ -362,7 +362,7 @@ int EIGEN_BLAS_FUNC(tbsv)(char *uplo, char *op, char *diag, int *n, int *k, Real Scalar* a = reinterpret_cast(pa); Scalar* x = reinterpret_cast(px); int coeff_rows = *k+1; - + int info = 0; if(UPLO(*uplo)==INVALID) info = 1; else if(OP(*op)==INVALID) info = 2; @@ -373,22 +373,22 @@ int EIGEN_BLAS_FUNC(tbsv)(char *uplo, char *op, char *diag, int *n, int *k, Real else if(*incx==0) info = 9; if(info) return xerbla_(SCALAR_SUFFIX_UP"TBSV ",&info,6); - + if(*n==0 || (*k==0 && DIAG(*diag)==UNIT)) return 0; - + int actual_n = *n; - + Scalar* actual_x = get_compact_vector(x,actual_n,*incx); - + int code = OP(*op) | (UPLO(*uplo) << 2) | (DIAG(*diag) << 3); if(code>=16 || func[code]==0) return 0; func[code](*n, *k, a, *lda, actual_x); - + if(actual_x!=x) delete[] copy_back(actual_x,x,actual_n,*incx); - + return 0; } @@ -521,4 +521,3 @@ int EIGEN_BLAS_FUNC(tpsv)(char *uplo, char *opa, char *diag, int *n, RealScalar return 1; } - diff --git a/blas/level2_real_impl.h b/blas/level2_real_impl.h index 8d56eaaa1..9722a4674 100644 --- a/blas/level2_real_impl.h +++ b/blas/level2_real_impl.h @@ -51,8 +51,8 @@ int EIGEN_BLAS_FUNC(symv) (char *uplo, int *n, RealScalar *palpha, RealScalar *p if(beta!=Scalar(1)) { - if(beta==Scalar(0)) vector(actual_y, *n).setZero(); - else vector(actual_y, *n) *= beta; + if(beta==Scalar(0)) make_vector(actual_y, *n).setZero(); + else make_vector(actual_y, *n) *= beta; } int code = UPLO(*uplo); @@ -179,7 +179,7 @@ int EIGEN_BLAS_FUNC(syr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px Scalar* x_cpy = get_compact_vector(x,*n,*incx); Scalar* y_cpy = get_compact_vector(y,*n,*incy); - + int code = UPLO(*uplo); if(code>=2 || func[code]==0) return 0; @@ -366,5 +366,3 @@ int EIGEN_BLAS_FUNC(ger)(int *m, int *n, Scalar *palpha, Scalar *px, int *incx, return 1; } - - From d992634fbce54e2ad367e1f57834503b5a3b5eaf Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Apr 2014 11:31:21 +0200 Subject: [PATCH 0412/1103] Fix bug #776: it seems that mingw does not support weak linking --- blas/xerbla.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/blas/xerbla.cpp b/blas/xerbla.cpp index 0d57710fe..0422f79b7 100644 --- a/blas/xerbla.cpp +++ b/blas/xerbla.cpp @@ -1,7 +1,7 @@ #include -#if (defined __GNUC__) +#if (defined __GNUC__) && (!defined __MINGW32__) #define EIGEN_WEAK_LINKING __attribute__ ((weak)) #else #define EIGEN_WEAK_LINKING From ec65e6648cf2483bd5e10a3010395ffb42a82a77 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Apr 2014 11:45:43 +0200 Subject: [PATCH 0413/1103] bug #775: propagate generator when workingaround cmake bug #9220 --- cmake/language_support.cmake | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/cmake/language_support.cmake b/cmake/language_support.cmake index d687b71f6..93f8a8fd8 100644 --- a/cmake/language_support.cmake +++ b/cmake/language_support.cmake @@ -33,7 +33,7 @@ function(workaround_9220 language language_works) file(WRITE ${CMAKE_BINARY_DIR}/language_tests/${language}/CMakeLists.txt ${text}) execute_process( - COMMAND ${CMAKE_COMMAND} . + COMMAND ${CMAKE_COMMAND} . -G "${CMAKE_GENERATOR}" WORKING_DIRECTORY ${CMAKE_BINARY_DIR}/language_tests/${language} RESULT_VARIABLE return_code OUTPUT_QUIET @@ -64,3 +64,4 @@ endfunction(workaround_9220) #message("CXX_language_works = ${CXX_language_works}") #workaround_9220(CXXp CXXp_language_works) #message("CXXp_language_works = ${CXXp_language_works}") + From ceae5b4145ba2a70793c58cff05a43e95e721ff5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Apr 2014 11:52:23 +0200 Subject: [PATCH 0414/1103] Fix lapack build --- lapack/eigenvalues.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/lapack/eigenvalues.cpp b/lapack/eigenvalues.cpp index a1526ebcd..6141032ab 100644 --- a/lapack/eigenvalues.cpp +++ b/lapack/eigenvalues.cpp @@ -64,14 +64,14 @@ EIGEN_LAPACK_FUNC(syev,(char *jobz, char *uplo, int* n, Scalar* a, int *lda, Sca if(eig.info()==NoConvergence) { - vector(w,*n).setZero(); + make_vector(w,*n).setZero(); if(computeVectors) matrix(a,*n,*n,*lda).setIdentity(); //*info = 1; return 0; } - vector(w,*n) = eig.eigenvalues(); + make_vector(w,*n) = eig.eigenvalues(); if(computeVectors) matrix(a,*n,*n,*lda) = eig.eigenvectors(); From 56c4851323376876977a9d408cb5aec2a6159c54 Mon Sep 17 00:00:00 2001 From: Florian George Date: Tue, 1 Apr 2014 15:52:25 +0200 Subject: [PATCH 0415/1103] Fixed typo: symmretric -> symmetric --- bench/btl/data/action_settings.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/bench/btl/data/action_settings.txt b/bench/btl/data/action_settings.txt index e32213e22..0244fa59f 100644 --- a/bench/btl/data/action_settings.txt +++ b/bench/btl/data/action_settings.txt @@ -14,6 +14,6 @@ partial_lu_decomp ; "{/*1.5 Partial LU decomposition}" ; "matrix size" ; 4:3000 tridiagonalization ; "{/*1.5 Tridiagonalization}" ; "matrix size" ; 4:3000 hessenberg ; "{/*1.5 Hessenberg decomposition}" ; "matrix size" ; 4:3000 symv ; "{/*1.5 symmetric matrix vector product}" ; "matrix size" ; 4:3000 -syr2 ; "{/*1.5 symmretric rank-2 update (A += u^T v + u v^T)}" ; "matrix size" ; 4:3000 +syr2 ; "{/*1.5 symmetric rank-2 update (A += u^T v + u v^T)}" ; "matrix size" ; 4:3000 ger ; "{/*1.5 general rank-1 update (A += u v^T)}" ; "matrix size" ; 4:3000 rot ; "{/*1.5 apply rotation in the plane}" ; "vector size" ; 4:1000000 \ No newline at end of file From 1cb8de12501b3db3d12774774cbbe12983243cee Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 1 Apr 2014 17:44:48 +0200 Subject: [PATCH 0416/1103] Make some actual verifications inside the autodiff unit test --- unsupported/test/autodiff.cpp | 45 ++++++++++++++++++----------------- 1 file changed, 23 insertions(+), 22 deletions(-) diff --git a/unsupported/test/autodiff.cpp b/unsupported/test/autodiff.cpp index 6eb417e8d..087e7c542 100644 --- a/unsupported/test/autodiff.cpp +++ b/unsupported/test/autodiff.cpp @@ -127,46 +127,47 @@ template void forward_jacobian(const Func& f) VERIFY_IS_APPROX(j, jref); } + +// TODO also check actual derivatives! void test_autodiff_scalar() { - std::cerr << foo(1,2) << "\n"; + Vector2f p = Vector2f::Random(); typedef AutoDiffScalar AD; - AD ax(1,Vector2f::UnitX()); - AD ay(2,Vector2f::UnitY()); + AD ax(p.x(),Vector2f::UnitX()); + AD ay(p.y(),Vector2f::UnitY()); AD res = foo(ax,ay); - std::cerr << res.value() << " <> " - << res.derivatives().transpose() << "\n\n"; + VERIFY_IS_APPROX(res.value(), foo(p.x(),p.y())); } +// TODO also check actual derivatives! void test_autodiff_vector() { - std::cerr << foo(Vector2f(1,2)) << "\n"; + Vector2f p = Vector2f::Random(); typedef AutoDiffScalar AD; typedef Matrix VectorAD; - VectorAD p(AD(1),AD(-1)); - p.x().derivatives() = Vector2f::UnitX(); - p.y().derivatives() = Vector2f::UnitY(); + VectorAD ap = p.cast(); + ap.x().derivatives() = Vector2f::UnitX(); + ap.y().derivatives() = Vector2f::UnitY(); - AD res = foo(p); - std::cerr << res.value() << " <> " - << res.derivatives().transpose() << "\n\n"; + AD res = foo(ap); + VERIFY_IS_APPROX(res.value(), foo(p)); } void test_autodiff_jacobian() { - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); - CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); - CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); - CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); - CALL_SUBTEST(( forward_jacobian(TestFunc1(3,3)) )); - } + CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); + CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); + CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); + CALL_SUBTEST(( forward_jacobian(TestFunc1()) )); + CALL_SUBTEST(( forward_jacobian(TestFunc1(3,3)) )); } void test_autodiff() { - test_autodiff_scalar(); - test_autodiff_vector(); -// test_autodiff_jacobian(); + for(int i = 0; i < g_repeat; i++) { + CALL_SUBTEST_1( test_autodiff_scalar() ); + CALL_SUBTEST_2( test_autodiff_vector() ); + CALL_SUBTEST_3( test_autodiff_jacobian() ); + } } From 8044b00a7fa30af20cc184fa2991bd0acf0f9aa3 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 3 Apr 2014 23:41:47 +0200 Subject: [PATCH 0417/1103] bug #782: Workaround for gcc <= 4.4 compilation error on the NEON PacketMath code. --- Eigen/src/Core/arch/NEON/PacketMath.h | 1 + 1 file changed, 1 insertion(+) diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 05e891df2..c2f28e25c 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -95,6 +95,7 @@ template<> struct packet_traits : default_packet_traits // workaround gcc 4.2, 4.3 and 4.4 compilatin issue EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); } EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); } +EIGEN_STRONG_INLINE float32x2_t vld1_dup_f32 (const float* x) { return ::vld1_dup_f32 ((const float32_t*)x); } EIGEN_STRONG_INLINE void vst1q_f32(float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); } EIGEN_STRONG_INLINE void vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); } #endif From 096af597992609be62ec1104fdee476e3065f2e4 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 4 Apr 2014 17:48:37 +0200 Subject: [PATCH 0418/1103] Fix bug #784: Assert if assigning a product to a triangularView does not match the size. --- Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index ffa871cae..225b994d1 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -265,6 +265,8 @@ template template TriangularView& TriangularView::assignProduct(const ProductBase& prod, const Scalar& alpha) { + eigen_assert(m_matrix.rows() == prod.rows() && m_matrix.cols() == prod.cols()); + general_product_to_triangular_selector::run(m_matrix.const_cast_derived(), prod.derived(), alpha); return *this; From 5afcb4965c9ccdce05e1852dd22b552559696f61 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Fri, 4 Apr 2014 16:48:13 +0100 Subject: [PATCH 0419/1103] Remove out-dated comment in cholesky test. --- test/cholesky.cpp | 4 ---- 1 file changed, 4 deletions(-) diff --git a/test/cholesky.cpp b/test/cholesky.cpp index d4d90e467..569318f83 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -82,10 +82,6 @@ template void cholesky(const MatrixType& m) symm += a1 * a1.adjoint(); } - // to test if really Cholesky only uses the upper triangular part, uncomment the following - // FIXME: currently that fails !! - //symm.template part().setZero(); - { SquareMatrixType symmUp = symm.template triangularView(); SquareMatrixType symmLo = symm.template triangularView(); From b446ff037e9a8573da55c180b2e72e07dd3de4e6 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 4 Apr 2014 14:12:24 -0700 Subject: [PATCH 0420/1103] Deleted some dead code. --- Eigen/src/Core/products/GeneralMatrixMatrix.h | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index dd9d79657..b35625a11 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -241,7 +241,6 @@ class level3_blocking protected: LhsScalar* m_blockA; RhsScalar* m_blockB; - RhsScalar* m_blockW; DenseIndex m_mc; DenseIndex m_nc; @@ -250,7 +249,7 @@ class level3_blocking public: level3_blocking() - : m_blockA(0), m_blockB(0), m_blockW(0), m_mc(0), m_nc(0), m_kc(0) + : m_blockA(0), m_blockB(0), m_mc(0), m_nc(0), m_kc(0) {} inline DenseIndex mc() const { return m_mc; } @@ -259,7 +258,6 @@ class level3_blocking inline LhsScalar* blockA() { return m_blockA; } inline RhsScalar* blockB() { return m_blockB; } - inline RhsScalar* blockW() { return m_blockW; } }; template From 3b2321e3ab157b314286227e08d738d74bfbf5d1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 4 Apr 2014 17:08:47 -0700 Subject: [PATCH 0421/1103] Updated the geo_parametrizedline_2 test for AVX. --- test/geo_parametrizedline.cpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/test/geo_parametrizedline.cpp b/test/geo_parametrizedline.cpp index 427327cd7..58d8eb7f5 100644 --- a/test/geo_parametrizedline.cpp +++ b/test/geo_parametrizedline.cpp @@ -66,9 +66,9 @@ template void parametrizedline_alignment() typedef ParametrizedLine Line4a; typedef ParametrizedLine Line4u; - EIGEN_ALIGN_DEFAULT Scalar array1[8]; - EIGEN_ALIGN_DEFAULT Scalar array2[8]; - EIGEN_ALIGN_DEFAULT Scalar array3[8+1]; + EIGEN_ALIGN_DEFAULT Scalar array1[16]; + EIGEN_ALIGN_DEFAULT Scalar array2[16]; + EIGEN_ALIGN_DEFAULT Scalar array3[16+1]; Scalar* array3u = array3+1; Line4a *p1 = ::new(reinterpret_cast(array1)) Line4a; From a91a7a1964305311133858de96b845da49389922 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Mon, 7 Apr 2014 14:14:48 +0100 Subject: [PATCH 0422/1103] doc: Add references to Cholesky methods in SelfAdjointView. --- Eigen/Cholesky | 6 ++++-- Eigen/src/Cholesky/LDLT.h | 6 ++++-- Eigen/src/Cholesky/LLT.h | 6 ++++-- doc/Doxyfile.in | 3 ++- 4 files changed, 14 insertions(+), 7 deletions(-) diff --git a/Eigen/Cholesky b/Eigen/Cholesky index f727f5d89..7314d326c 100644 --- a/Eigen/Cholesky +++ b/Eigen/Cholesky @@ -10,9 +10,11 @@ * * * This module provides two variants of the Cholesky decomposition for selfadjoint (hermitian) matrices. - * Those decompositions are accessible via the following MatrixBase methods: - * - MatrixBase::llt(), + * Those decompositions are also accessible via the following methods: + * - MatrixBase::llt() * - MatrixBase::ldlt() + * - SelfAdjointView::llt() + * - SelfAdjointView::ldlt() * * \code * #include diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index b43e85e7f..efac7fe40 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -43,7 +43,7 @@ namespace internal { * Remember that Cholesky decompositions are not rank-revealing. Also, do not use a Cholesky * decomposition to determine whether a system of equations has a solution. * - * \sa MatrixBase::ldlt(), class LLT + * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt(), class LLT */ template class LDLT { @@ -179,7 +179,7 @@ template class LDLT * least-square solution of \f$ D y_3 = y_2 \f$ is computed. This does not mean that this function * computes the least-square solution of \f$ A x = b \f$ is \f$ A \f$ is singular. * - * \sa MatrixBase::ldlt() + * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt() */ template inline const internal::solve_retval @@ -582,6 +582,7 @@ MatrixType LDLT::reconstructedMatrix() const #ifndef __CUDACC__ /** \cholesky_module * \returns the Cholesky decomposition with full pivoting without square root of \c *this + * \sa MatrixBase::ldlt() */ template inline const LDLT::PlainObject, UpLo> @@ -592,6 +593,7 @@ SelfAdjointView::ldlt() const /** \cholesky_module * \returns the Cholesky decomposition with full pivoting without square root of \c *this + * \sa SelfAdjointView::ldlt() */ template inline const LDLT::PlainObject> diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index 2201c641e..45ed8438f 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -41,7 +41,7 @@ template struct LLT_Traits; * Example: \include LLT_example.cpp * Output: \verbinclude LLT_example.out * - * \sa MatrixBase::llt(), class LDLT + * \sa MatrixBase::llt(), SelfAdjointView::llt(), class LDLT */ /* HEY THIS DOX IS DISABLED BECAUSE THERE's A BUG EITHER HERE OR IN LDLT ABOUT THAT (OR BOTH) * Note that during the decomposition, only the upper triangular part of A is considered. Therefore, @@ -115,7 +115,7 @@ template class LLT * Example: \include LLT_solve.cpp * Output: \verbinclude LLT_solve.out * - * \sa solveInPlace(), MatrixBase::llt() + * \sa solveInPlace(), MatrixBase::llt(), SelfAdjointView::llt() */ template inline const internal::solve_retval @@ -468,6 +468,7 @@ MatrixType LLT::reconstructedMatrix() const #ifndef __CUDACC__ /** \cholesky_module * \returns the LLT decomposition of \c *this + * \sa SelfAdjointView::llt() */ template inline const LLT::PlainObject> @@ -478,6 +479,7 @@ MatrixBase::llt() const /** \cholesky_module * \returns the LLT decomposition of \c *this + * \sa SelfAdjointView::llt() */ template inline const LLT::PlainObject, UpLo> diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in index 85af9f1d4..7bbf693a0 100644 --- a/doc/Doxyfile.in +++ b/doc/Doxyfile.in @@ -223,7 +223,7 @@ ALIASES = "only_for_vectors=This is only for vectors (either row- "note_about_using_kernel_to_study_multiple_solutions=If you need a complete analysis of the space of solutions, take the one solution obtained by this method and add to it elements of the kernel, as determined by kernel()." \ "note_about_checking_solutions=This method just tries to find as good a solution as possible. If you want to check whether a solution exists or if it is accurate, just call this function to get a result and then compute the error of this result, or use MatrixBase::isApprox() directly, for instance like this: \code bool a_solution_exists = (A*result).isApprox(b, precision); \endcode This method avoids dividing by zero, so that the non-existence of a solution doesn't by itself mean that you'll get \c inf or \c nan values." \ "note_try_to_help_rvo=This function returns the result by value. In order to make that efficient, it is implemented as just a return statement using a special constructor, hopefully allowing the compiler to perform a RVO (return value optimization)." \ - "nonstableyet=\warning This is not considered to be part of the stable public API yet. Changes may happen in future releases. See \ref Experimental \"Experimental parts of Eigen\" + "nonstableyet=\warning This is not considered to be part of the stable public API yet. Changes may happen in future releases. See \ref Experimental \"Experimental parts of Eigen\"" ALIASES += "eigenAutoToc= " @@ -1583,6 +1583,7 @@ PREDEFINED = EIGEN_EMPTY_STRUCT \ EIGEN_VECTORIZE \ EIGEN_QT_SUPPORT \ EIGEN_STRONG_INLINE=inline \ + EIGEN_DEVICE_FUNC= \ "EIGEN2_SUPPORT_STAGE=99" \ "EIGEN_MAKE_CWISE_BINARY_OP(METHOD,FUNCTOR)=template const CwiseBinaryOp, const Derived, const OtherDerived> METHOD(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const;" \ "EIGEN_CWISE_PRODUCT_RETURN_TYPE(LHS,RHS)=CwiseBinaryOp, const LHS, const RHS>" From a1fcf599faa023d1752b37aef215d80286e76839 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 10 Apr 2014 11:19:37 -0700 Subject: [PATCH 0423/1103] Silenced a compilation warning produced by nvcc. --- Eigen/src/Core/util/Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 0f8ab065a..1a1d3e98d 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -89,7 +89,7 @@ inline void throw_std_bad_alloc() #ifdef EIGEN_EXCEPTIONS throw std::bad_alloc(); #else - std::size_t huge = -1; + std::size_t huge = -1ULL; new int[huge]; #endif } From 1b333c89c9762795fda69f29a1239541a712f4f1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 10 Apr 2014 17:43:13 -0700 Subject: [PATCH 0424/1103] Updated my previous fix to avoid introducing a compilation warning on ARM platforms. --- Eigen/src/Core/util/Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 1a1d3e98d..4988be5d9 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -89,7 +89,7 @@ inline void throw_std_bad_alloc() #ifdef EIGEN_EXCEPTIONS throw std::bad_alloc(); #else - std::size_t huge = -1ULL; + std::size_t huge = static_cast(-1); new int[huge]; #endif } From 91288e9bf9d6a9c3d63a9152e863b4390d0a40c7 Mon Sep 17 00:00:00 2001 From: Freddie Witherden Date: Sat, 12 Apr 2014 12:53:09 +0100 Subject: [PATCH 0425/1103] Add include LevenbergMarquardt in CMakeLists.txt. This fixes bug #768. --- unsupported/Eigen/src/CMakeLists.txt | 1 + unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt | 2 +- 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/CMakeLists.txt b/unsupported/Eigen/src/CMakeLists.txt index f3180b52b..8eb2808e3 100644 --- a/unsupported/Eigen/src/CMakeLists.txt +++ b/unsupported/Eigen/src/CMakeLists.txt @@ -2,6 +2,7 @@ ADD_SUBDIRECTORY(AutoDiff) ADD_SUBDIRECTORY(BVH) ADD_SUBDIRECTORY(FFT) ADD_SUBDIRECTORY(IterativeSolvers) +ADD_SUBDIRECTORY(LevenbergMarquardt) ADD_SUBDIRECTORY(MatrixFunctions) ADD_SUBDIRECTORY(MoreVectorization) ADD_SUBDIRECTORY(NonLinearOptimization) diff --git a/unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt b/unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt index 8513803ce..d9690854d 100644 --- a/unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt +++ b/unsupported/Eigen/src/LevenbergMarquardt/CMakeLists.txt @@ -2,5 +2,5 @@ FILE(GLOB Eigen_LevenbergMarquardt_SRCS "*.h") INSTALL(FILES ${Eigen_LevenbergMarquardt_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/LevenbergMarquardt COMPONENT Devel + DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/LevenbergMarquardt COMPONENT Devel ) From a803ff18a963fc23ddb3dccf33ed7058af415f39 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Sat, 12 Apr 2014 20:24:05 -0700 Subject: [PATCH 0426/1103] Fixed a typo in cuda_basic.cu --- test/cuda_basic.cu | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu index aa7f7a599..4c7e96c10 100644 --- a/test/cuda_basic.cu +++ b/test/cuda_basic.cu @@ -129,7 +129,7 @@ void test_cuda_basic() CALL_SUBTEST( run_and_compare_to_cuda(prod(), nthreads, in, out) ); CALL_SUBTEST( run_and_compare_to_cuda(diagonal(), nthreads, in, out) ); - CALL_SUBTEST( run_and_compare_to_c(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(diagonal(), nthreads, in, out) ); CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); CALL_SUBTEST( run_and_compare_to_cuda(eigenvalues(), nthreads, in, out) ); From 7903d3f27b275040702ce30eac8d329d6f571205 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Sat, 12 Apr 2014 23:39:37 -0700 Subject: [PATCH 0427/1103] Updated the compiler flags to enable nvcc to work with clang. --- test/CMakeLists.txt | 3 +++ 1 file changed, 3 insertions(+) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 62cbedae7..c2d827051 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -315,6 +315,9 @@ find_package(CUDA) if(CUDA_FOUND) set(CUDA_PROPAGATE_HOST_FLAGS OFF) + if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang") + set(CUDA_NVCC_FLAGS "-ccbin /usr/bin/clang" CACHE STRING "nvcc flags" FORCE) + endif() cuda_include_directories(${CMAKE_CURRENT_BINARY_DIR}) set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu") From 0587db8bf5ca5d5eb6fb8df1c02abddd5b5718ba Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 14 Apr 2014 11:43:08 +0200 Subject: [PATCH 0428/1103] bug #793: fix overflow in EigenSolver and add respective regression unit test --- Eigen/src/Eigenvalues/EigenSolver.h | 15 ++++++++++++++- test/eigensolver_generic.cpp | 13 +++++++++++++ 2 files changed, 27 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Eigenvalues/EigenSolver.h b/Eigen/src/Eigenvalues/EigenSolver.h index bf20e03ef..739466949 100644 --- a/Eigen/src/Eigenvalues/EigenSolver.h +++ b/Eigen/src/Eigenvalues/EigenSolver.h @@ -366,6 +366,7 @@ EigenSolver::compute(const MatrixType& matrix, bool computeEigenvect { using std::sqrt; using std::abs; + using std::max; eigen_assert(matrix.cols() == matrix.rows()); // Reduce to real Schur form. @@ -390,7 +391,19 @@ EigenSolver::compute(const MatrixType& matrix, bool computeEigenvect else { Scalar p = Scalar(0.5) * (m_matT.coeff(i, i) - m_matT.coeff(i+1, i+1)); - Scalar z = sqrt(abs(p * p + m_matT.coeff(i+1, i) * m_matT.coeff(i, i+1))); + Scalar z; + // Compute z = sqrt(abs(p * p + m_matT.coeff(i+1, i) * m_matT.coeff(i, i+1))); + // without overflow + { + Scalar t0 = m_matT.coeff(i+1, i); + Scalar t1 = m_matT.coeff(i, i+1); + Scalar maxval = (max)(abs(p),(max)(abs(t0),abs(t1))); + t0 /= maxval; + t1 /= maxval; + Scalar p0 = p/maxval; + z = maxval * sqrt(abs(p0 * p0 + t0 * t1)); + } + m_eivalues.coeffRef(i) = ComplexScalar(m_matT.coeff(i+1, i+1) + p, z); m_eivalues.coeffRef(i+1) = ComplexScalar(m_matT.coeff(i+1, i+1) + p, -z); i += 2; diff --git a/test/eigensolver_generic.cpp b/test/eigensolver_generic.cpp index 005af81eb..91383b5cf 100644 --- a/test/eigensolver_generic.cpp +++ b/test/eigensolver_generic.cpp @@ -121,5 +121,18 @@ void test_eigensolver_generic() } ); + // regression test for bug 793 +#ifdef EIGEN_TEST_PART_2 + { + MatrixXd a(3,3); + a << 0, 0, 1, + 1, 1, 1, + 1, 1e+200, 1; + Eigen::EigenSolver eig(a); + VERIFY_IS_APPROX(a * eig.pseudoEigenvectors(), eig.pseudoEigenvectors() * eig.pseudoEigenvalueMatrix()); + VERIFY_IS_APPROX(a * eig.eigenvectors(), eig.eigenvectors() * eig.eigenvalues().asDiagonal()); + } +#endif + TEST_SET_BUT_UNUSED_VARIABLE(s) } From 148acf8e4fb71294703d4d1deafaf52829535ab7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 14 Apr 2014 13:52:16 +0200 Subject: [PATCH 0429/1103] bug #790: fix overflow in real_2x2_jacobi_svd --- Eigen/src/SVD/JacobiSVD.h | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index eee31ca97..439eb5d29 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -415,6 +415,7 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, JacobiRotation *j_right) { using std::sqrt; + using std::abs; Matrix m; m << numext::real(matrix.coeff(p,p)), numext::real(matrix.coeff(p,q)), numext::real(matrix.coeff(q,p)), numext::real(matrix.coeff(q,q)); @@ -428,9 +429,11 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, } else { - RealScalar u = d / t; - rot1.c() = RealScalar(1) / sqrt(RealScalar(1) + numext::abs2(u)); - rot1.s() = rot1.c() * u; + RealScalar t2d2 = numext::hypot(t,d); + rot1.c() = abs(t)/t2d2; + rot1.s() = d/t2d2; + if(tmakeJacobi(m,0,1); From feaf7c7e6d01a4804cee5949a01ece1f8a46866f Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 14 Apr 2014 10:44:17 -0700 Subject: [PATCH 0430/1103] Optimized SSE unaligned loads and stores when compiling a 64bit target with a recent version of gcc (ie gcc 4.8). --- Eigen/src/Core/arch/SSE/PacketMath.h | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index ea05a3415..bc17726b4 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -271,14 +271,17 @@ template<> EIGEN_STRONG_INLINE Packet4i pload(const int* from) { E // TODO: do the same for MSVC (ICC is compatible) // NOTE: with the code below, MSVC's compiler crashes! -#if defined(__GNUC__) && defined(__i386__) +#if defined(__GNUC__) && (defined(__i386__) || (defined(__x86_64) && EIGEN_GNUC_AT_LEAST(4, 8))) // bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 + #define EIGEN_AVOID_CUSTOM_UNALIGNED_STORES 1 #elif defined(__clang__) // bug 201: Segfaults in __mm_loadh_pd with clang 2.8 #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 + #define EIGEN_AVOID_CUSTOM_UNALIGNED_STORES 0 #else #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0 + #define EIGEN_AVOID_CUSTOM_UNALIGNED_STORES 0 #endif template<> EIGEN_STRONG_INLINE Packet4f ploadu(const float* from) @@ -338,8 +341,12 @@ template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& f template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE +#if EIGEN_AVOID_CUSTOM_UNALIGNED_STORES + _mm_storeu_pd(to, from); +#else _mm_storel_pd((to), from); _mm_storeh_pd((to+1), from); +#endif } template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), _mm_castps_pd(from)); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), _mm_castsi128_pd(from)); } From 7098e6d976ee8d5b25776e749d3ef6e66a302829 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 14 Apr 2014 21:57:49 +0200 Subject: [PATCH 0431/1103] Add isfinite overload for complexes. --- Eigen/src/Core/MathFunctions.h | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 63fb92b75..20fc2be74 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -669,6 +669,15 @@ bool (isfinite)(const T& x) return x::highest() && x>NumTraits::lowest(); } +template +EIGEN_DEVICE_FUNC +bool (isfinite)(const std::complex& x) +{ + using std::real; + using std::imag; + return isfinite(real(x)) && isfinite(imag(x)); +} + } // end namespace numext namespace internal { From 3c66bb136bf2adcb9d73d3d66850a8b907bc9264 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 14 Apr 2014 22:00:27 +0200 Subject: [PATCH 0432/1103] bug #793: detect NaN and INF in EigenSolver instead of aborting with an assert. --- Eigen/src/Eigenvalues/EigenSolver.h | 25 ++++++++++++++++++++++--- 1 file changed, 22 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Eigenvalues/EigenSolver.h b/Eigen/src/Eigenvalues/EigenSolver.h index 739466949..d2563d470 100644 --- a/Eigen/src/Eigenvalues/EigenSolver.h +++ b/Eigen/src/Eigenvalues/EigenSolver.h @@ -275,10 +275,11 @@ template class EigenSolver */ EigenSolver& compute(const MatrixType& matrix, bool computeEigenvectors = true); + /** \returns NumericalIssue if the input contains INF or NaN values or overflow occured. Returns Success otherwise. */ ComputationInfo info() const { eigen_assert(m_isInitialized && "EigenSolver is not initialized."); - return m_realSchur.info(); + return m_info; } /** \brief Sets the maximum number of iterations allowed. */ @@ -302,6 +303,7 @@ template class EigenSolver EigenvalueType m_eivalues; bool m_isInitialized; bool m_eigenvectorsOk; + ComputationInfo m_info; RealSchur m_realSchur; MatrixType m_matT; @@ -367,12 +369,15 @@ EigenSolver::compute(const MatrixType& matrix, bool computeEigenvect using std::sqrt; using std::abs; using std::max; + using numext::isfinite; eigen_assert(matrix.cols() == matrix.rows()); // Reduce to real Schur form. m_realSchur.compute(matrix, computeEigenvectors); + + m_info = m_realSchur.info(); - if (m_realSchur.info() == Success) + if (m_info == Success) { m_matT = m_realSchur.matrixT(); if (computeEigenvectors) @@ -386,6 +391,13 @@ EigenSolver::compute(const MatrixType& matrix, bool computeEigenvect if (i == matrix.cols() - 1 || m_matT.coeff(i+1, i) == Scalar(0)) { m_eivalues.coeffRef(i) = m_matT.coeff(i, i); + if(!isfinite(m_eivalues.coeffRef(i))) + { + m_isInitialized = true; + m_eigenvectorsOk = false; + m_info = NumericalIssue; + return *this; + } ++i; } else @@ -406,6 +418,13 @@ EigenSolver::compute(const MatrixType& matrix, bool computeEigenvect m_eivalues.coeffRef(i) = ComplexScalar(m_matT.coeff(i+1, i+1) + p, z); m_eivalues.coeffRef(i+1) = ComplexScalar(m_matT.coeff(i+1, i+1) + p, -z); + if(!(isfinite(m_eivalues.coeffRef(i)) && isfinite(m_eivalues.coeffRef(i+1)))) + { + m_isInitialized = true; + m_eigenvectorsOk = false; + m_info = NumericalIssue; + return *this; + } i += 2; } } @@ -594,7 +613,7 @@ void EigenSolver::doComputeEigenvectors() } else { - eigen_assert(0 && "Internal bug in EigenSolver"); // this should not happen + eigen_assert(0 && "Internal bug in EigenSolver (INF or NaN has not been detected)"); // this should not happen } } From 1afd50e0f31d56f7ee228e915e15998422c3ea11 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 14 Apr 2014 14:26:30 -0700 Subject: [PATCH 0433/1103] Fixed a typo in CXX11Meta.h --- unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h index d6b5d75d9..618e2eb7b 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -285,7 +285,7 @@ struct equal_op { template constexpr static inli struct not_equal_op { template constexpr static inline auto run(A a, B b) -> decltype(a != b) { return a != b; } }; struct lesser_op { template constexpr static inline auto run(A a, B b) -> decltype(a < b) { return a < b; } }; struct lesser_equal_op { template constexpr static inline auto run(A a, B b) -> decltype(a <= b) { return a <= b; } }; -struct greater_op { template constexpr static inline auto run(A a, B b) -> decltype(a < b) { return a < b; } }; +struct greater_op { template constexpr static inline auto run(A a, B b) -> decltype(a > b) { return a > b; } }; struct greater_equal_op { template constexpr static inline auto run(A a, B b) -> decltype(a >= b) { return a >= b; } }; /* generic unary operations */ From 59f5f155c2d4c6069f61fc3df39c7f6264857488 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Tue, 15 Apr 2014 15:21:38 +0100 Subject: [PATCH 0434/1103] Port products with permutation matrices to evaluators. --- Eigen/src/Core/PermutationMatrix.h | 116 ++++++++++++++++++++++++++++- Eigen/src/Core/Product.h | 26 ++++++- Eigen/src/Core/ProductEvaluators.h | 87 ++++++++++++++++++++++ Eigen/src/Core/util/Constants.h | 1 + 4 files changed, 227 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 9add80c54..61aa0ce31 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -288,6 +288,10 @@ class PermutationMatrix : public PermutationBase Traits; public: +#ifdef EIGEN_TEST_EVALUATORS + typedef const PermutationMatrix& Nested; +#endif + #ifndef EIGEN_PARSED_BY_DOXYGEN typedef typename Traits::IndicesType IndicesType; #endif @@ -461,6 +465,22 @@ class Map, struct PermutationStorage {}; +#ifdef EIGEN_TEST_EVALUATORS + +// storage type of product of permutation wrapper with dense + +namespace internal { + +template<> struct promote_storage_type +{ typedef Dense ret; }; + +template<> struct promote_storage_type +{ typedef Dense ret; }; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS + template class TranspositionsWrapper; namespace internal { template @@ -473,8 +493,13 @@ struct traits > enum { RowsAtCompileTime = _IndicesType::SizeAtCompileTime, ColsAtCompileTime = _IndicesType::SizeAtCompileTime, - MaxRowsAtCompileTime = IndicesType::MaxRowsAtCompileTime, +#ifdef EIGEN_TEST_EVALUATORS + MaxRowsAtCompileTime = IndicesType::MaxSizeAtCompileTime, + MaxColsAtCompileTime = IndicesType::MaxSizeAtCompileTime, +#else + MaxRowsAtCompileTime = IndicesType::MaxRowsAtCompileTime, // is this a bug in Eigen 2.2 ? MaxColsAtCompileTime = IndicesType::MaxColsAtCompileTime, +#endif Flags = 0 #ifndef EIGEN_TEST_EVALUATORS , @@ -508,6 +533,37 @@ class PermutationWrapper : public PermutationBase +EIGEN_DEVICE_FUNC +const Product +operator*(const MatrixBase &matrix, + const PermutationBase& permutation) +{ + return Product + (matrix.derived(), permutation.derived()); +} + +/** \returns the matrix with the permutation applied to the rows. + */ +template +EIGEN_DEVICE_FUNC +const Product +operator*(const PermutationBase &permutation, + const MatrixBase& matrix) +{ + return Product + (permutation.derived(), matrix.derived()); +} + +#else // EIGEN_TEST_EVALUATORS + /** \returns the matrix with the permutation applied to the columns. */ template @@ -533,6 +589,8 @@ operator*(const PermutationBase &permutation, (permutation.derived(), matrix.derived()); } +#endif // EIGEN_TEST_EVALUATORS + namespace internal { template @@ -662,6 +720,28 @@ class Transpose > DenseMatrixType toDenseMatrix() const { return *this; } +#ifdef EIGEN_TEST_EVALUATORS + + /** \returns the matrix with the inverse permutation applied to the columns. + */ + template friend + const Product + operator*(const MatrixBase& matrix, const Transpose& trPerm) + { + return Product(matrix.derived(), trPerm.derived()); + } + + /** \returns the matrix with the inverse permutation applied to the rows. + */ + template + const Product + operator*(const MatrixBase& matrix) const + { + return Product(*this, matrix.derived()); + } + +#else // EIGEN_TEST_EVALUATORS + /** \returns the matrix with the inverse permutation applied to the columns. */ template friend @@ -680,6 +760,8 @@ class Transpose > return internal::permut_matrix_product_retval(m_permutation, matrix.derived()); } +#endif // EIGEN_TEST_EVALUATORS + const PermutationType& nestedPermutation() const { return m_permutation; } protected: @@ -692,6 +774,38 @@ const PermutationWrapper MatrixBase::asPermutation() con return derived(); } +#ifdef EIGEN_TEST_EVALUATORS +namespace internal { + +// TODO currently a permutation matrix expression has the form PermutationMatrix or PermutationWrapper +// or their transpose; in the future shape should be defined by the expression traits +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef PermutationShape Shape; + static const int AssumeAliasing = 0; +}; + +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef PermutationShape Shape; + static const int AssumeAliasing = 0; +}; + +template +struct evaluator_traits > > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef PermutationShape Shape; + static const int AssumeAliasing = 0; +}; + +} // end namespace internal +#endif // EIGEN_TEST_EVALUATORS + } // end namespace Eigen #endif // EIGEN_PERMUTATIONMATRIX_H diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 2785847a6..626b737c7 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -29,8 +29,30 @@ template class Pro * */ -// Use ProductReturnType to get correct traits, in particular vectorization flags + namespace internal { + +// Determine the scalar of Product. This is normally the same as Lhs::Scalar times +// Rhs::Scalar, but product with permutation matrices inherit the scalar of the other factor. +template::Shape, + typename RhsShape = typename evaluator_traits::Shape > +struct product_result_scalar +{ + typedef typename scalar_product_traits::ReturnType Scalar; +}; + +template +struct product_result_scalar +{ + typedef typename Rhs::Scalar Scalar; +}; + +template + struct product_result_scalar +{ + typedef typename Lhs::Scalar Scalar; +}; + template struct traits > { @@ -39,7 +61,7 @@ struct traits > typedef MatrixXpr XprKind; - typedef typename scalar_product_traits::ReturnType Scalar; + typedef typename product_result_scalar::Scalar Scalar; typedef typename promote_storage_type::StorageKind, typename traits::StorageKind>::ret StorageKind; typedef typename promote_index_type::Index, diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 1159c2f44..7298c51b1 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -885,6 +885,93 @@ struct product_evaluator, ProductTag, DenseShape, }; +/*************************************************************************** +* Products with permutation matrices +***************************************************************************/ + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + permut_matrix_product_retval pmpr(lhs, rhs); + pmpr.evalTo(dst); + } +}; + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + permut_matrix_product_retval pmpr(rhs, lhs); + pmpr.evalTo(dst); + } +}; + +template +struct generic_product_impl, Rhs, PermutationShape, DenseShape, ProductType> +{ + template + static void evalTo(Dest& dst, const Transpose& lhs, const Rhs& rhs) + { + permut_matrix_product_retval pmpr(lhs.nestedPermutation(), rhs); + pmpr.evalTo(dst); + } +}; + +template +struct generic_product_impl, DenseShape, PermutationShape, ProductType> +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Transpose& rhs) + { + permut_matrix_product_retval pmpr(rhs.nestedPermutation(), lhs); + pmpr.evalTo(dst); + } +}; + +// TODO: left/right and self-adj/symmetric/permutation look the same ... Too much boilerplate? +template +struct product_evaluator, ProductTag, PermutationShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +template +struct product_evaluator, ProductTag, DenseShape, PermutationShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 64efac8e9..aae8625d1 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -447,6 +447,7 @@ struct DiagonalShape { static std::string debugName() { return "DiagonalShape struct BandShape { static std::string debugName() { return "BandShape"; } }; struct TriangularShape { static std::string debugName() { return "TriangularShape"; } }; struct SelfAdjointShape { static std::string debugName() { return "SelfAdjointShape"; } }; +struct PermutationShape { static std::string debugName() { return "PermutationShape"; } }; struct SparseShape { static std::string debugName() { return "SparseShape"; } }; #endif From e0dbb68c2f17f3c8c6accc7dc0b2b8d544e2eebc Mon Sep 17 00:00:00 2001 From: Mark Borgerding Date: Tue, 15 Apr 2014 13:57:03 -0400 Subject: [PATCH 0435/1103] Check IMKL version for compatibility with Eigen --- Eigen/src/Core/util/MKL_support.h | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/MKL_support.h b/Eigen/src/Core/util/MKL_support.h index 1e6e355d6..8acca9c8c 100644 --- a/Eigen/src/Core/util/MKL_support.h +++ b/Eigen/src/Core/util/MKL_support.h @@ -54,8 +54,25 @@ #endif #if defined EIGEN_USE_MKL +# include +/*Check IMKL version for compatibility: < 10.3 is not usable with Eigen*/ +# ifndef INTEL_MKL_VERSION +# undef EIGEN_USE_MKL /* INTEL_MKL_VERSION is not even defined on older versions */ +# elif INTEL_MKL_VERSION < 100305 /* the intel-mkl-103-release-notes say this was when the lapacke.h interface was added*/ +# undef EIGEN_USE_MKL +# endif +# ifndef EIGEN_USE_MKL + /*If the MKL version is too old, undef everything*/ +# undef EIGEN_USE_MKL_ALL +# undef EIGEN_USE_BLAS +# undef EIGEN_USE_LAPACKE +# undef EIGEN_USE_MKL_VML +# undef EIGEN_USE_LAPACKE_STRICT +# undef EIGEN_USE_LAPACKE +# endif +#endif -#include +#if defined EIGEN_USE_MKL #include #define EIGEN_MKL_VML_THRESHOLD 128 From b30706bd5cfb9df8ab10f7861232623373aec626 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Tue, 15 Apr 2014 22:51:46 +0100 Subject: [PATCH 0436/1103] Fix typo in Inverse.h --- Eigen/src/Core/Inverse.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 57eaf99f1..51a39fe29 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -63,7 +63,7 @@ public: EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; } protected: - XprTypeNested &m_xpr; + XprTypeNested m_xpr; }; /** \internal From d5a795f67366db20a132cc70e4f0217f42372357 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 16 Apr 2014 17:05:11 +0200 Subject: [PATCH 0437/1103] New gebp kernel handling up to 3 packets x 4 register-level blocks. Huge speeup on Haswell. This changeset also introduce new vector functions: ploadquad and predux4. --- Eigen/src/Core/GenericPacketMath.h | 35 +- Eigen/src/Core/arch/AVX/Complex.h | 4 +- Eigen/src/Core/arch/AVX/PacketMath.h | 29 +- Eigen/src/Core/arch/AltiVec/PacketMath.h | 4 +- Eigen/src/Core/arch/NEON/Complex.h | 2 +- Eigen/src/Core/arch/NEON/PacketMath.h | 4 +- Eigen/src/Core/arch/SSE/Complex.h | 4 +- Eigen/src/Core/arch/SSE/PacketMath.h | 6 +- .../Core/products/GeneralBlockPanelKernel.h | 1430 +++++++++++------ Eigen/src/Core/products/GeneralMatrixMatrix.h | 121 +- Eigen/src/Core/products/Parallelizer.h | 25 +- .../Core/products/SelfadjointMatrixMatrix.h | 33 +- 12 files changed, 1064 insertions(+), 633 deletions(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 3e5db1a88..147298009 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -161,14 +161,6 @@ pload(const typename unpacket_traits::type* from) { return *from; } template EIGEN_DEVICE_FUNC inline Packet ploadu(const typename unpacket_traits::type* from) { return *from; } -/** \internal \returns a packet with elements of \a *from duplicated. - * For instance, for a packet of 8 elements, 4 scalar will be read from \a *from and - * duplicated to form: {from[0],from[0],from[1],from[1],,from[2],from[2],,from[3],from[3]} - * Currently, this function is only used for scalar * complex products. - */ -template EIGEN_DEVICE_FUNC inline Packet -ploaddup(const typename unpacket_traits::type* from) { return *from; } - /** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */ template EIGEN_DEVICE_FUNC inline Packet pset1(const typename unpacket_traits::type& a) { return a; } @@ -177,6 +169,24 @@ pset1(const typename unpacket_traits::type& a) { return a; } template EIGEN_DEVICE_FUNC inline Packet pload1(const typename unpacket_traits::type *a) { return pset1(*a); } +/** \internal \returns a packet with elements of \a *from duplicated. + * For instance, for a packet of 8 elements, 4 scalars will be read from \a *from and + * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]} + * Currently, this function is only used for scalar * complex products. + */ +template EIGEN_DEVICE_FUNC inline Packet +ploaddup(const typename unpacket_traits::type* from) { return *from; } + +/** \internal \returns a packet with elements of \a *from quadrupled. + * For instance, for a packet of 8 elements, 2 scalars will be read from \a *from and + * replicated to form: {from[0],from[0],from[0],from[0],from[1],from[1],from[1],from[1]} + * Currently, this function is only used in matrix products. + * For packet-size smaller or equal to 4, this function is equivalent to pload1 + */ +template EIGEN_DEVICE_FUNC inline Packet +ploadquad(const typename unpacket_traits::type* from) +{ return pload1(from); } + /** \internal equivalent to * \code * a0 = pload1(a+0); @@ -249,6 +259,15 @@ preduxp(const Packet* vecs) { return vecs[0]; } template EIGEN_DEVICE_FUNC inline typename unpacket_traits::type predux(const Packet& a) { return a; } +/** \internal \returns the sum of the elements of \a a by block of 4 elements. + * For a packet {a0, a1, a2, a3, a4, a5, a6, a7}, it returns a half packet {a0+a4, a1+a5, a2+a6, a3+a7} + * For packet-size smaller or equal to 4, this boils down to a noop. + */ +template EIGEN_DEVICE_FUNC inline +typename conditional<(unpacket_traits::size%8)==0,typename unpacket_traits::half,Packet>::type +predux4(const Packet& a) +{ return a; } + /** \internal \returns the product of the elements of \a a*/ template EIGEN_DEVICE_FUNC inline typename unpacket_traits::type predux_mul(const Packet& a) { return a; } diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index cb16180c5..8f95a7be7 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -45,7 +45,7 @@ template<> struct packet_traits > : default_packet_traits }; }; -template<> struct unpacket_traits { typedef std::complex type; enum {size=4}; }; +template<> struct unpacket_traits { typedef std::complex type; enum {size=4}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet4cf padd(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_add_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet4cf psub(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_sub_ps(a.v,b.v)); } @@ -271,7 +271,7 @@ template<> struct packet_traits > : default_packet_traits }; }; -template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; }; +template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; typedef Packet1cd half; }; template<> EIGEN_STRONG_INLINE Packet2cd padd(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_add_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cd psub(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_sub_pd(a.v,b.v)); } diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 38f52ecc8..47e10f6da 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -83,9 +83,9 @@ template<> struct packet_traits : default_packet_traits }; */ -template<> struct unpacket_traits { typedef float type; enum {size=8}; }; -template<> struct unpacket_traits { typedef double type; enum {size=4}; }; -template<> struct unpacket_traits { typedef int type; enum {size=8}; }; +template<> struct unpacket_traits { typedef float type; typedef Packet4f half; enum {size=8}; }; +template<> struct unpacket_traits { typedef double type; typedef Packet2d half; enum {size=4}; }; +template<> struct unpacket_traits { typedef int type; typedef Packet4i half; enum {size=8}; }; template<> EIGEN_STRONG_INLINE Packet8f pset1(const float& from) { return _mm256_set1_ps(from); } template<> EIGEN_STRONG_INLINE Packet4d pset1(const double& from) { return _mm256_set1_pd(from); } @@ -141,7 +141,16 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& return _mm256_fmadd_ps(a,b,c); #endif } -template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { return _mm256_fmadd_pd(a,b,c); } +template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { +#if defined(__clang__) || defined(__GNUC__) + // see above + Packet4d res = c; + asm("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); + return res; +#else + return _mm256_fmadd_pd(a,b,c); +#endif +} #endif template<> EIGEN_STRONG_INLINE Packet8f pmin(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); } @@ -189,6 +198,13 @@ template<> EIGEN_STRONG_INLINE Packet4d ploaddup(const double* from) return _mm256_blend_pd(tmp1,_mm256_permute2f128_pd(tmp2,tmp2,1),12); } +// Loads 2 floats from memory a returns the packet {a0, a0 a0, a0, a1, a1, a1, a1} +template<> EIGEN_STRONG_INLINE Packet8f ploadquad(const float* from) +{ + Packet8f tmp = _mm256_castps128_ps256(_mm_broadcast_ss(from)); + return _mm256_insertf128_ps(tmp, _mm_broadcast_ss(from+1), 1); +} + template<> EIGEN_STRONG_INLINE void pstore(float* to, const Packet8f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_ps(to, from); } template<> EIGEN_STRONG_INLINE void pstore(double* to, const Packet4d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_pd(to, from); } template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet8i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } @@ -345,6 +361,11 @@ template<> EIGEN_STRONG_INLINE double predux(const Packet4d& a) return pfirst(_mm256_hadd_pd(tmp0,tmp0)); } +template<> EIGEN_STRONG_INLINE Packet4f predux4(const Packet8f& a) +{ + return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)); +} + template<> EIGEN_STRONG_INLINE float predux_mul(const Packet8f& a) { Packet8f tmp; diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 5d7a16f5c..16948264f 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -99,8 +99,8 @@ template<> struct packet_traits : default_packet_traits }; }; -template<> struct unpacket_traits { typedef float type; enum {size=4}; }; -template<> struct unpacket_traits { typedef int type; enum {size=4}; }; +template<> struct unpacket_traits { typedef float type; enum {size=4}; typedef Packet4f half; }; +template<> struct unpacket_traits { typedef int type; enum {size=4}; typedef Packet4i half; }; /* inline std::ostream & operator <<(std::ostream & s, const Packet4f & v) { diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index e49c1a873..7ca76714f 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -47,7 +47,7 @@ template<> struct packet_traits > : default_packet_traits }; }; -template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; }; +template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex& from) { diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index fae7b55fc..83150507a 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -101,8 +101,8 @@ EIGEN_STRONG_INLINE void vst1q_f32(float* to, float32x4_t from) { ::vst1q EIGEN_STRONG_INLINE void vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); } #endif -template<> struct unpacket_traits { typedef float type; enum {size=4}; }; -template<> struct unpacket_traits { typedef int type; enum {size=4}; }; +template<> struct unpacket_traits { typedef float type; enum {size=4}; typedef Packet4f half; }; +template<> struct unpacket_traits { typedef int type; enum {size=4}; typedef Packet4i half; }; template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { return vdupq_n_f32(from); } template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return vdupq_n_s32(from); } diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index e54ebbf90..715e5a13c 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -49,7 +49,7 @@ template<> struct packet_traits > : default_packet_traits }; #endif -template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; }; +template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet2cf padd(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf psub(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); } @@ -296,7 +296,7 @@ template<> struct packet_traits > : default_packet_traits }; #endif -template<> struct unpacket_traits { typedef std::complex type; enum {size=1}; }; +template<> struct unpacket_traits { typedef std::complex type; enum {size=1}; typedef Packet1cd half; }; template<> EIGEN_STRONG_INLINE Packet1cd padd(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd psub(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); } diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index bc17726b4..89dfa6975 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -107,9 +107,9 @@ template<> struct packet_traits : default_packet_traits }; }; -template<> struct unpacket_traits { typedef float type; enum {size=4}; }; -template<> struct unpacket_traits { typedef double type; enum {size=2}; }; -template<> struct unpacket_traits { typedef int type; enum {size=4}; }; +template<> struct unpacket_traits { typedef float type; enum {size=4}; typedef Packet4f half; }; +template<> struct unpacket_traits { typedef double type; enum {size=2}; typedef Packet2d half; }; +template<> struct unpacket_traits { typedef int type; enum {size=4}; typedef Packet4i half; }; #if defined(_MSC_VER) && (_MSC_VER==1500) // Workaround MSVC 9 internal compiler error. diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index d9e659c9a..df30fdd3e 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -10,6 +10,12 @@ #ifndef EIGEN_GENERAL_BLOCK_PANEL_H #define EIGEN_GENERAL_BLOCK_PANEL_H +#ifdef USE_IACA +#include "iacaMarks.h" +#else +#define IACA_START +#define IACA_END +#endif namespace Eigen { @@ -92,12 +98,22 @@ void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) }; manage_caching_sizes(GetAction, &l1, &l2); - k = std::min(k, l1/kdiv); - SizeType _m = k>0 ? l2/(4 * sizeof(LhsScalar) * k) : 0; - if(_m(k, l1/kdiv); +// SizeType _m = k>0 ? l2/(4 * sizeof(LhsScalar) * k) : 0; +// if(_m(k,240); + n = std::min(n,3840/sizeof(RhsScalar)); + m = std::min(m,3840/sizeof(RhsScalar)); +#else + k = std::min(k,24); + n = std::min(n,384/sizeof(RhsScalar)); + m = std::min(m,384/sizeof(RhsScalar)); +#endif } template @@ -164,11 +180,15 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, - // register block size along the N direction (must be either 4 or 8) - nr = NumberOfRegisters/2, + // register block size along the N direction (must be either 2 or 4) + nr = 4,//NumberOfRegisters/4, // register block size along the M direction (currently, this one cannot be modified) - mr = LhsPacketSize, +#ifdef __FMA__ + mr = 3*LhsPacketSize, +#else + mr = 2*LhsPacketSize, +#endif LhsProgress = LhsPacketSize, RhsProgress = 1 @@ -198,23 +218,32 @@ public: { pbroadcast2(b, b0, b1); } - - EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const + + template + EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketType& dest) const { - dest = pset1(*b); + dest = pset1(*b); + } + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + dest = ploadquad(b); } - EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const + template + EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacketType& dest) const { - dest = pload(a); + dest = pload(a); } - EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const + template + EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacketType& dest) const { - dest = ploadu(a); + dest = ploadu(a); } - EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, AccPacket& tmp) const + template + EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, AccPacketType& tmp) const { // It would be a lot cleaner to call pmadd all the time. Unfortunately if we // let gcc allocate the register in which to store the result of the pmul @@ -232,6 +261,12 @@ public: { r = pmadd(c,alpha,r); } + + template + EIGEN_STRONG_INLINE void acc(const ResPacketHalf& c, const ResPacketHalf& alpha, ResPacketHalf& r) const + { + r = pmadd(c,alpha,r); + } protected: // conj_helper cj; @@ -281,6 +316,11 @@ public: { dest = pset1(*b); } + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + dest = pset1(*b); + } EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const { @@ -346,7 +386,23 @@ DoublePacket padd(const DoublePacket &a, const DoublePacket +const DoublePacket& predux4(const DoublePacket &a) +{ + return a; +} + +template struct unpacket_traits > { typedef DoublePacket half; }; +// template +// DoublePacket pmadd(const DoublePacket &a, const DoublePacket &b) +// { +// DoublePacket res; +// res.first = padd(a.first, b.first); +// res.second = padd(a.second,b.second); +// return res; +// } + template class gebp_traits, std::complex, _ConjLhs, _ConjRhs > { @@ -404,6 +460,16 @@ public: dest.second = pset1(imag(*b)); } + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, ResPacket& dest) const + { + loadRhs(b,dest); + } + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, DoublePacketType& dest) const + { + eigen_internal_assert(unpacket_traits::size<=4); + loadRhs(b,dest); + } + // linking error if instantiated without being optimized out: void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); @@ -619,26 +685,246 @@ void gebp_kernel if(strideA==-1) strideA = depth; if(strideB==-1) strideB = depth; conj_helper cj; - Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0; Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; - // Here we assume that mr==LhsProgress - const Index peeled_mc = (rows/mr)*mr; + const Index peeled_mc3 = mr>=3*Traits::LhsProgress ? (rows/(3*LhsProgress))*(3*LhsProgress) : 0; + const Index peeled_mc2 = mr>=2*Traits::LhsProgress ? peeled_mc3+((rows-peeled_mc3)/(2*LhsProgress))*(2*LhsProgress) : 0; + const Index peeled_mc1 = mr>=1*Traits::LhsProgress ? (rows/(1*LhsProgress))*(1*LhsProgress) : 0; enum { pk = 8 }; // NOTE Such a large peeling factor is important for large matrices (~ +5% when >1000 on Haswell) const Index peeled_kc = depth & ~(pk-1); - const Index depth2 = depth & ~1; - - // loops on each micro vertical panel of rhs (depth x nr) - // First pass using depth x 8 panels - if(nr>=8) +// const Index depth2 = depth & ~1; + +// std::cout << mr << " " << peeled_mc3 << " " << peeled_mc2 << " " << peeled_mc1 << "\n"; + + + //---------- Process 3 * LhsProgress rows at once ---------- + // This corresponds to 3*LhsProgress x nr register blocks. + // Usually, make sense only with FMA + if(mr>=3*Traits::LhsProgress) { - for(Index j2=0; j2 we select a mr x nr micro block of res which is entirely - // stored into mr/packet_size x nr registers. - for(Index i=0; i(alpha); + + R0 = ploadu(r0+0*Traits::ResPacketSize); + R1 = ploadu(r0+1*Traits::ResPacketSize); + R2 = ploadu(r0+2*Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + traits.acc(C8, alphav, R2); + pstoreu(r0+0*Traits::ResPacketSize, R0); + pstoreu(r0+1*Traits::ResPacketSize, R1); + pstoreu(r0+2*Traits::ResPacketSize, R2); + + R0 = ploadu(r1+0*Traits::ResPacketSize); + R1 = ploadu(r1+1*Traits::ResPacketSize); + R2 = ploadu(r1+2*Traits::ResPacketSize); + traits.acc(C1, alphav, R0); + traits.acc(C5, alphav, R1); + traits.acc(C9, alphav, R2); + pstoreu(r1+0*Traits::ResPacketSize, R0); + pstoreu(r1+1*Traits::ResPacketSize, R1); + pstoreu(r1+2*Traits::ResPacketSize, R2); + + R0 = ploadu(r2+0*Traits::ResPacketSize); + R1 = ploadu(r2+1*Traits::ResPacketSize); + R2 = ploadu(r2+2*Traits::ResPacketSize); + traits.acc(C2, alphav, R0); + traits.acc(C6, alphav, R1); + traits.acc(C10, alphav, R2); + pstoreu(r2+0*Traits::ResPacketSize, R0); + pstoreu(r2+1*Traits::ResPacketSize, R1); + pstoreu(r2+2*Traits::ResPacketSize, R2); + + R0 = ploadu(r3+0*Traits::ResPacketSize); + R1 = ploadu(r3+1*Traits::ResPacketSize); + R2 = ploadu(r3+2*Traits::ResPacketSize); + traits.acc(C3, alphav, R0); + traits.acc(C7, alphav, R1); + traits.acc(C11, alphav, R2); + pstoreu(r3+0*Traits::ResPacketSize, R0); + pstoreu(r3+1*Traits::ResPacketSize, R1); + pstoreu(r3+2*Traits::ResPacketSize, R2); + } + + // Deal with remaining columns of the rhs + for(Index j2=packet_cols4; j2(alpha); + + R0 = ploadu(r0+0*Traits::ResPacketSize); + R1 = ploadu(r0+1*Traits::ResPacketSize); + R2 = ploadu(r0+2*Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + traits.acc(C8 , alphav, R2); + pstoreu(r0+0*Traits::ResPacketSize, R0); + pstoreu(r0+1*Traits::ResPacketSize, R1); + pstoreu(r0+2*Traits::ResPacketSize, R2); + } + } + } + + //---------- Process 2 * LhsProgress rows at once ---------- + if(mr>=2*Traits::LhsProgress) + { + // loops on each largest micro horizontal panel of lhs (2*LhsProgress x depth) + for(Index i=peeled_mc3; i traits.initAcc(C7); ResScalar* r0 = &res[(j2+0)*resStride + i]; + ResScalar* r1 = &res[(j2+1)*resStride + i]; + ResScalar* r2 = &res[(j2+2)*resStride + i]; + ResScalar* r3 = &res[(j2+3)*resStride + i]; + + internal::prefetch(r0); + internal::prefetch(r1); + internal::prefetch(r2); + internal::prefetch(r3); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; - LhsPacket A0; - // uncomment for register prefetching - // LhsPacket A1; - // traits.loadLhs(blA, A0); + prefetch(&blB[0]); + LhsPacket A0, A1; + for(Index k=0; k(alpha); - - R0 = ploadu(r0+0*resStride); - R1 = ploadu(r0+1*resStride); - R2 = ploadu(r0+2*resStride); - R3 = ploadu(r0+3*resStride); - R4 = ploadu(r0+4*resStride); - R5 = ploadu(r0+5*resStride); - R6 = ploadu(r0+6*resStride); - traits.acc(C0, alphav, R0); - pstoreu(r0+0*resStride, R0); - R0 = ploadu(r0+7*resStride); - - traits.acc(C1, alphav, R1); - traits.acc(C2, alphav, R2); - traits.acc(C3, alphav, R3); - traits.acc(C4, alphav, R4); - traits.acc(C5, alphav, R5); - traits.acc(C6, alphav, R6); - traits.acc(C7, alphav, R0); - pstoreu(r0+1*resStride, R1); - pstoreu(r0+2*resStride, R2); - pstoreu(r0+3*resStride, R3); - pstoreu(r0+4*resStride, R4); - pstoreu(r0+5*resStride, R5); - pstoreu(r0+6*resStride, R6); - pstoreu(r0+7*resStride, R0); + R0 = ploadu(r0+0*Traits::ResPacketSize); + R1 = ploadu(r0+1*Traits::ResPacketSize); + R2 = ploadu(r1+0*Traits::ResPacketSize); + R3 = ploadu(r1+1*Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + traits.acc(C1, alphav, R2); + traits.acc(C5, alphav, R3); + pstoreu(r0+0*Traits::ResPacketSize, R0); + pstoreu(r0+1*Traits::ResPacketSize, R1); + pstoreu(r1+0*Traits::ResPacketSize, R2); + pstoreu(r1+1*Traits::ResPacketSize, R3); + + R0 = ploadu(r2+0*Traits::ResPacketSize); + R1 = ploadu(r2+1*Traits::ResPacketSize); + R2 = ploadu(r3+0*Traits::ResPacketSize); + R3 = ploadu(r3+1*Traits::ResPacketSize); + traits.acc(C2, alphav, R0); + traits.acc(C6, alphav, R1); + traits.acc(C3, alphav, R2); + traits.acc(C7, alphav, R3); + pstoreu(r2+0*Traits::ResPacketSize, R0); + pstoreu(r2+1*Traits::ResPacketSize, R1); + pstoreu(r3+0*Traits::ResPacketSize, R2); + pstoreu(r3+1*Traits::ResPacketSize, R3); } - // Deal with remaining rows of the lhs - // TODO we should vectorize if <= 8, and not strictly == - if(SwappedTraits::LhsProgress == 8) + // Deal with remaining columns of the rhs + for(Index j2=packet_cols4; j2 SwappedTraits; - typedef typename SwappedTraits::ResScalar SResScalar; - typedef typename SwappedTraits::LhsPacket SLhsPacket; - typedef typename SwappedTraits::RhsPacket SRhsPacket; - typedef typename SwappedTraits::ResPacket SResPacket; - typedef typename SwappedTraits::AccPacket SAccPacket; - SwappedTraits straits; - - Index rows2 = (rows & ~1); - for(Index i=peeled_mc; i(&res[j2*resStride + i], resStride); - SResPacket alphav = pset1(alpha); - straits.acc(padd(C0,C1), alphav, R); - pscatter(&res[j2*resStride + i], R, resStride); - - R = pgather(&res[j2*resStride + i + 1], resStride); - straits.acc(padd(C2,C3), alphav, R); - pscatter(&res[j2*resStride + i + 1], R, resStride); - - EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows 8"); - } - if(rows2!=rows) - { - Index i = rows-1; - const LhsScalar* blA = &blockA[i*strideA+offsetA]; - prefetch(&blA[0]); - const RhsScalar* blB = &blockB[j2*strideB+offsetB*8]; - - EIGEN_ASM_COMMENT("begin_vectorized_multiplication_of_last_rows 8"); - - SAccPacket C0,C1; - straits.initAcc(C0); // even - straits.initAcc(C1); // odd - - for(Index k=0; k(&res[j2*resStride + i], resStride); - SResPacket alphav = pset1(alpha); - straits.acc(padd(C0,C1), alphav, R); - pscatter(&res[j2*resStride + i], R, resStride); - } - } - else - { - // Pure scalar path - for(Index i=peeled_mc; i(alpha); + + R0 = ploadu(r0+0*Traits::ResPacketSize); + R1 = ploadu(r0+1*Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C4, alphav, R1); + pstoreu(r0+0*Traits::ResPacketSize, R0); + pstoreu(r0+1*Traits::ResPacketSize, R1); } } } - - // Second pass using depth x 4 panels - // If nr==8, then we have at most one such panel - // TODO: with 16 registers, we coud optimize this part to leverage more pipelinining, - // for instance, by using a 2 packet * 4 kernel. Useful when the rhs is thin - if(nr>=4) + //---------- Process 1 * LhsProgress rows at once ---------- + if(mr>=1*Traits::LhsProgress) { - for(Index j2=packet_cols8; j2 we select a mr x 4 micro block of res which is entirely - // stored into mr/packet_size x 4 registers. - for(Index i=0; i traits.initAcc(C3); ResScalar* r0 = &res[(j2+0)*resStride + i]; + ResScalar* r1 = &res[(j2+1)*resStride + i]; + ResScalar* r2 = &res[(j2+2)*resStride + i]; + ResScalar* r3 = &res[(j2+3)*resStride + i]; + + internal::prefetch(r0); + internal::prefetch(r1); + internal::prefetch(r2); + internal::prefetch(r3); // performs "inner" products - const RhsScalar* blB = &blockB[j2*strideB+offsetB*4]; + const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; + prefetch(&blB[0]); LhsPacket A0; + for(Index k=0; k(alpha); - - R0 = ploadu(r0+0*resStride); - R1 = ploadu(r0+1*resStride); - R2 = ploadu(r0+2*resStride); - traits.acc(C0, alphav, R0); - pstoreu(r0+0*resStride, R0); - R0 = ploadu(r0+3*resStride); - - traits.acc(C1, alphav, R1); - traits.acc(C2, alphav, R2); - traits.acc(C3, alphav, R0); - pstoreu(r0+1*resStride, R1); - pstoreu(r0+2*resStride, R2); - pstoreu(r0+3*resStride, R0); + R0 = ploadu(r0+0*Traits::ResPacketSize); + R1 = ploadu(r1+0*Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + traits.acc(C1, alphav, R1); + pstoreu(r0+0*Traits::ResPacketSize, R0); + pstoreu(r1+0*Traits::ResPacketSize, R1); + + R0 = ploadu(r2+0*Traits::ResPacketSize); + R1 = ploadu(r3+0*Traits::ResPacketSize); + traits.acc(C2, alphav, R0); + traits.acc(C3, alphav, R1); + pstoreu(r2+0*Traits::ResPacketSize, R0); + pstoreu(r3+0*Traits::ResPacketSize, R1); } - for(Index i=peeled_mc; i(alpha); + R0 = ploadu(r0+0*Traits::ResPacketSize); + traits.acc(C0, alphav, R0); + pstoreu(r0+0*Traits::ResPacketSize, R0); + } + } + } + //---------- Process remaining rows, 1 at once ---------- + { + // loop on each row of the lhs (1*LhsProgress x depth) + for(Index i=peeled_mc1; i::half,SResPacket>::type SResPacketHalf; + typedef typename conditional::half,SLhsPacket>::type SLhsPacketHalf; + typedef typename conditional::half,SRhsPacket>::type SRhsPacketHalf; + typedef typename conditional::half,SAccPacket>::type SAccPacketHalf; + + SResPacketHalf R = pgather(&res[j2*resStride + i], resStride); + SResPacketHalf alphav = pset1(alpha); + + if(depth-endk>0) + { + // We have to handle the last row of the rhs which corresponds to a half-packet + SLhsPacketHalf a0; + SRhsPacketHalf b0; + straits.loadLhsUnaligned(blB, a0); + straits.loadRhs(blA, b0); + SAccPacketHalf c0 = predux4(C0); + straits.madd(a0,b0,c0,b0); + straits.acc(c0, alphav, R); + } + else + { + straits.acc(predux4(C0), alphav, R); + } + pscatter(&res[j2*resStride + i], R, resStride); + } + else + { + SResPacket R = pgather(&res[j2*resStride + i], resStride); + SResPacket alphav = pset1(alpha); + straits.acc(C0, alphav, R); + pscatter(&res[j2*resStride + i], R, resStride); } - SResPacket R = pgather(&res[j2*resStride + i], resStride); - SResPacket alphav = pset1(alpha); - straits.acc(C0, alphav, R); - pscatter(&res[j2*resStride + i], R, resStride); - - EIGEN_ASM_COMMENT("end_vectorized_multiplication_of_last_rows 1x4"); } - else + else // scalar path { - // Pure scalar path - // gets a 1 x 4 res block as registers + // get a 1 x 4 res block as registers ResScalar C0(0), C1(0), C2(0), C3(0); for(Index k=0; k B_1 = blB[3]; MADD(cj,A0,B_0,C2, B_0); MADD(cj,A0,B_1,C3, B_1); - + blB += 4; } res[(j2+0)*resStride + i] += alpha*C0; @@ -1058,56 +1377,22 @@ void gebp_kernel res[(j2+3)*resStride + i] += alpha*C3; } } - } - } - - // process remaining rhs/res columns one at a time - for(Index j2=packet_cols4; j2(alpha); - ResScalar* r0 = &res[(j2+0)*resStride + i]; - R0 = ploadu(r0); - traits.acc(C0, alphav, R0); - pstoreu(r0, R0); - } - // pure scalar path - for(Index i=peeled_mc; i // // 32 33 34 35 ... // 36 36 38 39 ... -template -struct gemm_pack_lhs +template +struct gemm_pack_lhs { EIGEN_DONT_INLINE void operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride=0, Index offset=0); }; -template -EIGEN_DONT_INLINE void gemm_pack_lhs +template +EIGEN_DONT_INLINE void gemm_pack_lhs ::operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride, Index offset) { typedef typename packet_traits::type Packet; @@ -1146,87 +1431,174 @@ EIGEN_DONT_INLINE void gemm_pack_lhs=depth && offset<=stride)); - eigen_assert( (StorageOrder==RowMajor) || ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) || (Pack1<=4) ); + eigen_assert( ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) || (Pack1<=4) ); conj_if::IsComplex && Conjugate> cj; - const_blas_data_mapper lhs(_lhs,lhsStride); + const_blas_data_mapper lhs(_lhs,lhsStride); Index count = 0; - Index peeled_mc = (rows/Pack1)*Pack1; - for(Index i=0; i=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; + const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; + const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; + + // Pack 3 packets + if(Pack1>=3*PacketSize) { - if(PanelMode) count += Pack1 * offset; - - if(StorageOrder==ColMajor) + for(Index i=0; i=1*PacketSize) A = ploadu(&lhs(i+0*PacketSize, k)); - if(Pack1>=2*PacketSize) B = ploadu(&lhs(i+1*PacketSize, k)); - if(Pack1>=3*PacketSize) C = ploadu(&lhs(i+2*PacketSize, k)); - if(Pack1>=4*PacketSize) D = ploadu(&lhs(i+3*PacketSize, k)); - if(Pack1>=1*PacketSize) { pstore(blockA+count, cj.pconj(A)); count+=PacketSize; } - if(Pack1>=2*PacketSize) { pstore(blockA+count, cj.pconj(B)); count+=PacketSize; } - if(Pack1>=3*PacketSize) { pstore(blockA+count, cj.pconj(C)); count+=PacketSize; } - if(Pack1>=4*PacketSize) { pstore(blockA+count, cj.pconj(D)); count+=PacketSize; } - } - else - { - if(Pack1>=1) blockA[count++] = cj(lhs(i+0, k)); - if(Pack1>=2) blockA[count++] = cj(lhs(i+1, k)); - if(Pack1>=3) blockA[count++] = cj(lhs(i+2, k)); - if(Pack1>=4) blockA[count++] = cj(lhs(i+3, k)); - } + Packet A, B, C; + A = ploadu(&lhs(i+0*PacketSize, k)); + B = ploadu(&lhs(i+1*PacketSize, k)); + C = ploadu(&lhs(i+2*PacketSize, k)); + pstore(blockA+count, cj.pconj(A)); count+=PacketSize; + pstore(blockA+count, cj.pconj(B)); count+=PacketSize; + pstore(blockA+count, cj.pconj(C)); count+=PacketSize; } + if(PanelMode) count += (3*PacketSize) * (stride-offset-depth); } - else + } + // Pack 2 packets + if(Pack1>=2*PacketSize) + { + for(Index i=peeled_mc3; i(&lhs(i+0*PacketSize, k)); + B = ploadu(&lhs(i+1*PacketSize, k)); + pstore(blockA+count, cj.pconj(A)); count+=PacketSize; + pstore(blockA+count, cj.pconj(B)); count+=PacketSize; + } + if(PanelMode) count += (2*PacketSize) * (stride-offset-depth); + } + } + // Pack 1 packets + if(Pack1>=1*PacketSize) + { + for(Index i=peeled_mc2; i(&lhs(i+0*PacketSize, k)); + pstore(blockA+count, cj.pconj(A)); + count+=PacketSize; + } + if(PanelMode) count += (1*PacketSize) * (stride-offset-depth); + } + } + // Pack scalars +// if(rows-peeled_mc>=Pack2) +// { +// if(PanelMode) count += Pack2*offset; +// for(Index k=0; k +struct gemm_pack_lhs +{ + EIGEN_DONT_INLINE void operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride=0, Index offset=0); +}; + +template +EIGEN_DONT_INLINE void gemm_pack_lhs + ::operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride, Index offset) +{ + typedef typename packet_traits::type Packet; + enum { PacketSize = packet_traits::size }; + + EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS"); + EIGEN_UNUSED_VARIABLE(stride); + EIGEN_UNUSED_VARIABLE(offset); + eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride)); + conj_if::IsComplex && Conjugate> cj; + const_blas_data_mapper lhs(_lhs,lhsStride); + Index count = 0; + +// const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; +// const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; +// const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; + + int pack_packets = Pack1/PacketSize; + Index i = 0; + while(pack_packets>0) + { + + Index remaining_rows = rows-i; + Index peeled_mc = i+(remaining_rows/(pack_packets*PacketSize))*(pack_packets*PacketSize); +// std::cout << "pack_packets = " << pack_packets << " from " << i << " to " << peeled_mc << "\n"; + for(; i kernel; - for (int p = 0; p < PacketSize; ++p) { - kernel.packet[p] = ploadu(&lhs(i+p+m, k)); - } + for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = ploadu(&lhs(i+p+m, k)); ptranspose(kernel); - for (int p = 0; p < PacketSize; ++p) { - pstore(blockA+count+m+Pack1*p, cj.pconj(kernel.packet[p])); - } + for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack_packets*PacketSize)*p, cj.pconj(kernel.packet[p])); } - count += PacketSize*Pack1; + count += PacketSize*(pack_packets*PacketSize); } - for(; k=Pack2) - { - if(PanelMode) count += Pack2*offset; - for(Index k=0; k=Pack2) +// { +// if(PanelMode) count += Pack2*offset; +// for(Index k=0; k=4 ? (cols/4) * 4 : 0; Index count = 0; const Index peeled_k = (depth/PacketSize)*PacketSize; - if(nr>=8) - { - for(Index j2=0; j2 kernel; - for (int p = 0; p < PacketSize; ++p) { - kernel.packet[p] = ploadu(&rhs[(j2+p)*rhsStride+k]); - } - ptranspose(kernel); - for (int p = 0; p < PacketSize; ++p) { - pstoreu(blockB+count, cj.pconj(kernel.packet[p])); - count+=PacketSize; - } - } - } - for(; k=8) +// { +// for(Index j2=0; j2 kernel; +// for (int p = 0; p < PacketSize; ++p) { +// kernel.packet[p] = ploadu(&rhs[(j2+p)*rhsStride+k]); +// } +// ptranspose(kernel); +// for (int p = 0; p < PacketSize; ++p) { +// pstoreu(blockB+count, cj.pconj(kernel.packet[p])); +// count+=PacketSize; +// } +// } +// } +// for(; k=4) { @@ -1383,39 +1755,39 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=4 ? (cols/4) * 4 : 0; Index count = 0; - if(nr>=8) - { - for(Index j2=0; j2(&rhs[k*rhsStride + j2]); - pstoreu(blockB+count, cj.pconj(A)); - } else if (PacketSize==4) { - Packet A = ploadu(&rhs[k*rhsStride + j2]); - Packet B = ploadu(&rhs[k*rhsStride + j2 + PacketSize]); - pstoreu(blockB+count, cj.pconj(A)); - pstoreu(blockB+count+PacketSize, cj.pconj(B)); - } else { - const Scalar* b0 = &rhs[k*rhsStride + j2]; - blockB[count+0] = cj(b0[0]); - blockB[count+1] = cj(b0[1]); - blockB[count+2] = cj(b0[2]); - blockB[count+3] = cj(b0[3]); - blockB[count+4] = cj(b0[4]); - blockB[count+5] = cj(b0[5]); - blockB[count+6] = cj(b0[6]); - blockB[count+7] = cj(b0[7]); - } - count += 8; - } - // skip what we have after - if(PanelMode) count += 8 * (stride-offset-depth); - } - } +// if(nr>=8) +// { +// for(Index j2=0; j2(&rhs[k*rhsStride + j2]); +// pstoreu(blockB+count, cj.pconj(A)); +// } else if (PacketSize==4) { +// Packet A = ploadu(&rhs[k*rhsStride + j2]); +// Packet B = ploadu(&rhs[k*rhsStride + j2 + PacketSize]); +// pstoreu(blockB+count, cj.pconj(A)); +// pstoreu(blockB+count+PacketSize, cj.pconj(B)); +// } else { +// const Scalar* b0 = &rhs[k*rhsStride + j2]; +// blockB[count+0] = cj(b0[0]); +// blockB[count+1] = cj(b0[1]); +// blockB[count+2] = cj(b0[2]); +// blockB[count+3] = cj(b0[3]); +// blockB[count+4] = cj(b0[4]); +// blockB[count+5] = cj(b0[5]); +// blockB[count+6] = cj(b0[6]); +// blockB[count+7] = cj(b0[7]); +// } +// count += 8; +// } +// // skip what we have after +// if(PanelMode) count += 8 * (stride-offset-depth); +// } +// } if(nr>=4) { for(Index j2=packet_cols8; j2 struct general_matrix_matrix_product { + typedef gebp_traits Traits; + typedef typename scalar_product_traits::ReturnType ResScalar; static EIGEN_STRONG_INLINE void run( Index rows, Index cols, Index depth, @@ -51,6 +53,8 @@ template< struct general_matrix_matrix_product { +typedef gebp_traits Traits; + typedef typename scalar_product_traits::ReturnType ResScalar; static void run(Index rows, Index cols, Index depth, const LhsScalar* _lhs, Index lhsStride, @@ -63,11 +67,9 @@ static void run(Index rows, Index cols, Index depth, const_blas_data_mapper lhs(_lhs,lhsStride); const_blas_data_mapper rhs(_rhs,rhsStride); - typedef gebp_traits Traits; - Index kc = blocking.kc(); // cache block size along the K direction Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction - //Index nc = blocking.nc(); // cache block size along the N direction + Index nc = (std::min)(cols,blocking.nc()); // cache block size along the N direction gemm_pack_lhs pack_lhs; gemm_pack_rhs pack_rhs; @@ -80,66 +82,68 @@ static void run(Index rows, Index cols, Index depth, Index tid = omp_get_thread_num(); Index threads = omp_get_num_threads(); - std::size_t sizeA = kc*mc; - ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, 0); + LhsScalar* blockA = blocking.blockA(); + eigen_internal_assert(blockA!=0); - RhsScalar* blockB = blocking.blockB(); - eigen_internal_assert(blockB!=0); - + std::size_t sizeB = kc*nc; + ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, 0); + // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs... for(Index k=0; k rows of B', and cols of the A' // In order to reduce the chance that a thread has to wait for the other, - // let's start by packing A'. - pack_lhs(blockA, &lhs(0,k), lhsStride, actual_kc, mc); + // let's start by packing B'. + pack_rhs(blockB, &rhs(k,0), rhsStride, actual_kc, nc); - // Pack B_k to B' in a parallel fashion: - // each thread packs the sub block B_k,j to B'_j where j is the thread id. + // Pack A_k to A' in a parallel fashion: + // each thread packs the sub block A_k,i to A'_i where i is the thread id. - // However, before copying to B'_j, we have to make sure that no other thread is still using it, + // However, before copying to A'_i, we have to make sure that no other thread is still using it, // i.e., we test that info[tid].users equals 0. // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it. while(info[tid].users!=0) {} info[tid].users += threads; + + pack_lhs(blockA+info[tid].lhs_start*actual_kc, &lhs(info[tid].lhs_start,k), lhsStride, actual_kc, info[tid].lhs_length); - pack_rhs(blockB+info[tid].rhs_start*actual_kc, &rhs(k,info[tid].rhs_start), rhsStride, actual_kc, info[tid].rhs_length); - - // Notify the other threads that the part B'_j is ready to go. + // Notify the other threads that the part A'_i is ready to go. info[tid].sync = k; - - // Computes C_i += A' * B' per B'_j + + // Computes C_i += A' * B' per A'_i for(Index shift=0; shift0) - while(info[j].sync!=k) {} - - gebp(res+info[j].rhs_start*resStride, resStride, blockA, blockB+info[j].rhs_start*actual_kc, mc, actual_kc, info[j].rhs_length, alpha, -1,-1,0,0); + while(info[i].sync!=k) {} + gebp(res+info[i].lhs_start, resStride, blockA+info[i].lhs_start*actual_kc, blockB, info[i].lhs_length, actual_kc, nc, alpha); } - // Then keep going as usual with the remaining A' - for(Index i=mc; i Pack rhs's panel into a sequential chunk of memory (L2 caching) - // Note that this panel will be read as many times as the number of blocks in the lhs's - // vertical panel which is, in practice, a very low number. - pack_rhs(blockB, &rhs(k2,0), rhsStride, actual_kc, cols); + // => Pack lhs's panel into a sequential chunk of memory (L2/L3 caching) + // Note that this panel will be read as many times as the number of blocks in the rhs's + // horizontal panel which is, in practice, a very low number. + pack_lhs(blockA, &lhs(0,k2), lhsStride, actual_kc, rows); - // For each mc x kc block of the lhs's vertical panel... - // (==GEPP_VAR1) - for(Index i2=0; i2 > template struct gemm_functor { - gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha, - BlockingType& blocking) + gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha, BlockingType& blocking) : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking) {} void initParallelSession() const { - m_blocking.allocateB(); + m_blocking.allocateA(); } void operator() (Index row, Index rows, Index col=0, Index cols=-1, GemmParallelInfo* info=0) const @@ -220,6 +221,8 @@ struct gemm_functor (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(), m_actualAlpha, m_blocking, info); } + + typedef typename Gemm::Traits Traits; protected: const Lhs& m_lhs; @@ -316,13 +319,23 @@ class gemm_blocking_spacem_mc = Transpose ? cols : rows; this->m_nc = Transpose ? rows : cols; this->m_kc = depth; - computeProductBlockingSizes(this->m_kc, this->m_mc, this->m_nc); + if(full_rows) + { + DenseIndex m = this->m_mc; + computeProductBlockingSizes(this->m_kc, m, this->m_nc); + } + else // full columns + { + DenseIndex n = this->m_nc; + computeProductBlockingSizes(this->m_kc, this->m_mc, n); + } + m_sizeA = this->m_mc * this->m_kc; m_sizeB = this->m_kc * this->m_nc; } @@ -396,7 +409,7 @@ class GeneralProduct (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>, _ActualLhsType, _ActualRhsType, Dest, BlockingType> GemmFunctor; - BlockingType blocking(dst.rows(), dst.cols(), lhs.cols()); + BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), true); internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit); } diff --git a/Eigen/src/Core/products/Parallelizer.h b/Eigen/src/Core/products/Parallelizer.h index 5c3e9b7ac..4079063eb 100644 --- a/Eigen/src/Core/products/Parallelizer.h +++ b/Eigen/src/Core/products/Parallelizer.h @@ -73,13 +73,13 @@ namespace internal { template struct GemmParallelInfo { - GemmParallelInfo() : sync(-1), users(0), rhs_start(0), rhs_length(0) {} + GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {} int volatile sync; int volatile users; - Index rhs_start; - Index rhs_length; + Index lhs_start; + Index lhs_length; }; template @@ -107,7 +107,7 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos if((!Condition) || (omp_get_num_threads()>1)) return func(0,rows, 0,cols); - Index size = transpose ? cols : rows; + Index size = transpose ? rows : cols; // 2- compute the maximal number of threads from the size of the product: // FIXME this has to be fine tuned @@ -126,26 +126,25 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos std::swap(rows,cols); Index blockCols = (cols / threads) & ~Index(0x3); - Index blockRows = (rows / threads) & ~Index(0x7); + Index blockRows = (rows / threads); + blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr; GemmParallelInfo* info = new GemmParallelInfo[threads]; - #pragma omp parallel for schedule(static,1) num_threads(threads) - for(Index i=0; i +template struct symm_pack_lhs { template inline @@ -45,22 +45,29 @@ struct symm_pack_lhs } void operator()(Scalar* blockA, const Scalar* _lhs, Index lhsStride, Index cols, Index rows) { + enum { PacketSize = packet_traits::size }; const_blas_data_mapper lhs(_lhs,lhsStride); Index count = 0; - Index peeled_mc = (rows/Pack1)*Pack1; - for(Index i=0; i(blockA, lhs, cols, i, count); - } - - if(rows-peeled_mc>=Pack2) - { - pack(blockA, lhs, cols, peeled_mc, count); - peeled_mc += Pack2; - } + //Index peeled_mc3 = (rows/Pack1)*Pack1; + + const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; + const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; + const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; + + if(Pack1>=3*PacketSize) + for(Index i=0; i(blockA, lhs, cols, i, count); + + if(Pack1>=2*PacketSize) + for(Index i=peeled_mc3; i(blockA, lhs, cols, i, count); + + if(Pack1>=1*PacketSize) + for(Index i=peeled_mc2; i(blockA, lhs, cols, i, count); // do the same with mr==1 - for(Index i=peeled_mc; i Date: Wed, 16 Apr 2014 18:16:36 +0100 Subject: [PATCH 0438/1103] Implement evaluator. All supported tests pass apart from Sparse and Geometry, except test in adjoint_4 that a = a.transpose() raises an assert. --- Eigen/src/Core/CoreEvaluators.h | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 47f50d548..8df1b706d 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -1292,6 +1292,32 @@ private: EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; } }; +// -------------------- ReturnByValue -------------------- + +// Expression is evaluated in a temporary; default implementation of Assignment is bypassed so that +// when a ReturnByValue expression is assigned, the evaluator is not constructed. +// TODO: Finalize port to new regime; ReturnByValue should not exist in the expression world + +template +struct evaluator > + : evaluator::PlainObject>::type +{ + typedef typename ReturnByValue::PlainObject PlainObject; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const ReturnByValue& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + xpr.evalTo(m_result); + } + + PlainObject m_result; +}; + } // namespace internal } // end namespace Eigen From de8336a9bc6daf5dc47c3d711913497fb3a9f15d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 16 Apr 2014 23:14:58 +0200 Subject: [PATCH 0439/1103] Enable alloca on MAC OSX --- Eigen/src/Core/util/Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index a4d7e454d..34b387ded 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -552,7 +552,7 @@ template struct smart_memmove_helper { // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA // to the appropriate stack allocation function #ifndef EIGEN_ALLOCA - #if (defined __linux__) + #if (defined __linux__) || (defined __APPLE__) #define EIGEN_ALLOCA alloca #elif defined(_MSC_VER) #define EIGEN_ALLOCA _alloca From d936ddc3d129162f096d8eb39084ea1a47ff6f29 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 16 Apr 2014 23:15:42 +0200 Subject: [PATCH 0440/1103] Fallback to lazy products for very small ones. --- Eigen/src/Core/products/GeneralMatrixMatrix.h | 30 +++++++++++++++++++ 1 file changed, 30 insertions(+) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index d06e0f808..38ca7b242 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -386,7 +386,37 @@ class GeneralProduct typedef internal::scalar_product_op BinOp; EIGEN_CHECK_BINARY_COMPATIBILIY(BinOp,LhsScalar,RhsScalar); } + + template + inline void evalTo(Dest& dst) const + { + if((m_rhs.rows()+dst.rows()+dst.cols())<20) + dst.noalias() = m_lhs .lazyProduct( m_rhs ); + else + { + dst.setZero(); + scaleAndAddTo(dst,Scalar(1)); + } + } + template + inline void addTo(Dest& dst) const + { + if((m_rhs.rows()+dst.rows()+dst.cols())<20) + dst.noalias() += m_lhs .lazyProduct( m_rhs ); + else + scaleAndAddTo(dst,Scalar(1)); + } + + template + inline void subTo(Dest& dst) const + { + if((m_rhs.rows()+dst.rows()+dst.cols())<20) + dst.noalias() -= m_lhs .lazyProduct( m_rhs ); + else + scaleAndAddTo(dst,Scalar(-1)); + } + template void scaleAndAddTo(Dest& dst, const Scalar& alpha) const { eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols()); From 0fa8290366038f6b71494499aba163eb60698426 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 16:02:27 +0200 Subject: [PATCH 0441/1103] Optimize ploaddup for AVX --- Eigen/src/Core/arch/AVX/PacketMath.h | 19 +++++++++---------- 1 file changed, 9 insertions(+), 10 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 47e10f6da..8d2e88061 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -182,20 +182,19 @@ template<> EIGEN_STRONG_INLINE Packet8i ploadu(const int* from) { EIGE // Loads 4 floats from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, a3} template<> EIGEN_STRONG_INLINE Packet8f ploaddup(const float* from) { - // FIXME we should only load the first 128bits - Packet8f tmp = ploadu(from); - Packet8f tmp1 = _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2)); - Packet8f tmp2 = _mm256_permute_ps(tmp, _MM_SHUFFLE(1,1,0,0)); - return _mm256_blend_ps(_mm256_permute2f128_ps(tmp1,tmp1,1),tmp2,15); + // TODO try to find a way to avoid the need of a temporary register + Packet8f tmp = _mm256_castps128_ps256(_mm_loadu_ps(from)); + tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1); + return _mm256_unpacklo_ps(tmp,tmp); } // Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1} template<> EIGEN_STRONG_INLINE Packet4d ploaddup(const double* from) { - // FIXME we should only load the first 128bits - Packet4d tmp = ploadu(from); - Packet4d tmp1 = _mm256_permute_pd(tmp,0); - Packet4d tmp2 = _mm256_permute_pd(tmp,3); - return _mm256_blend_pd(tmp1,_mm256_permute2f128_pd(tmp2,tmp2,1),12); + // TODO try to find a way to avoid the need of a temporary register + Packet2d tmp0 = _mm_loadu_pd(from); + Packet2d tmp1 = _mm_permute_pd(tmp0,3); + tmp0 = _mm_permute_pd(tmp0,0); + return _mm256_insertf128_pd(_mm256_castpd128_pd256(tmp0), tmp1, 1); } // Loads 2 floats from memory a returns the packet {a0, a0 a0, a0, a1, a1, a1, a1} From 11fbdcbc385917f44b7b01671e158d07a695eb00 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 16:04:30 +0200 Subject: [PATCH 0442/1103] Fix and optimize mixed products --- .../Core/products/GeneralBlockPanelKernel.h | 166 ++++++++++-------- 1 file changed, 92 insertions(+), 74 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index df30fdd3e..dcc0b4a0d 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -180,14 +180,15 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, - // register block size along the N direction (must be either 2 or 4) - nr = 4,//NumberOfRegisters/4, + // register block size along the N direction must be 1 or 4 + nr = 4, // register block size along the M direction (currently, this one cannot be modified) #ifdef __FMA__ + // we assume 16 registers mr = 3*LhsPacketSize, #else - mr = 2*LhsPacketSize, + mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize, #endif LhsProgress = LhsPacketSize, @@ -209,15 +210,15 @@ public: p = pset1(ResScalar(0)); } - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) - { - pbroadcast4(b, b0, b1, b2, b3); - } - - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) - { - pbroadcast2(b, b0, b1); - } +// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) +// { +// pbroadcast4(b, b0, b1, b2, b3); +// } +// +// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) +// { +// pbroadcast2(b, b0, b1); +// } template EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketType& dest) const @@ -290,8 +291,13 @@ public: ResPacketSize = Vectorizable ? packet_traits::size : 1, NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, - nr = NumberOfRegisters/2, - mr = LhsPacketSize, + nr = 4, +#ifdef __FMA__ + // we assume 16 registers + mr = 3*LhsPacketSize, +#else + mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize, +#endif LhsProgress = LhsPacketSize, RhsProgress = 1 @@ -332,15 +338,15 @@ public: dest = ploadu(a); } - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) - { - pbroadcast4(b, b0, b1, b2, b3); - } - - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) - { - pbroadcast2(b, b0, b1); - } +// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) +// { +// pbroadcast4(b, b0, b1, b2, b3); +// } +// +// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) +// { +// pbroadcast2(b, b0, b1); +// } EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const { @@ -566,7 +572,7 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, // FIXME: should depend on NumberOfRegisters nr = 4, - mr = ResPacketSize, + mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*ResPacketSize, LhsProgress = ResPacketSize, RhsProgress = 1 @@ -593,19 +599,25 @@ public: } // linking error if instantiated without being optimized out: - void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); - - EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) - { - // FIXME not sure that's the best way to implement it! - b0 = pload1(b+0); - b1 = pload1(b+1); - } +// void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); +// +// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) +// { +// // FIXME not sure that's the best way to implement it! +// b0 = pload1(b+0); +// b1 = pload1(b+1); +// } EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const { dest = ploaddup(a); } + + EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const + { + eigen_internal_assert(unpacket_traits::size<=4); + loadRhs(b,dest); + } EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const { @@ -619,7 +631,13 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { +#ifdef EIGEN_VECTORIZE_FMA + EIGEN_UNUSED_VARIABLE(tmp); + c.v = pmadd(a,b.v,c.v); +#else tmp = b; tmp.v = pmul(a,tmp.v); c = padd(c,tmp); +#endif + } EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const @@ -956,7 +974,7 @@ void gebp_kernel for(Index k=0; k for(Index k=0; k EIGEN_GEBGP_ONESTEP(7); blB += pk*4*RhsProgress; - blA += pk*(1*Traits::LhsProgress); + blA += pk*1*LhsProgress; IACA_END } // process remaining peeled loop @@ -1169,7 +1187,7 @@ void gebp_kernel RhsPacket B_0, B1; EIGEN_GEBGP_ONESTEP(0); blB += 4*RhsProgress; - blA += 1*Traits::LhsProgress; + blA += 1*LhsProgress; } #undef EIGEN_GEBGP_ONESTEP @@ -1439,6 +1457,8 @@ EIGEN_DONT_INLINE void gemm_pack_lhs=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; + const Index peeled_mc0 = Pack2>=1*PacketSize ? peeled_mc1 + : Pack2>1 ? (rows/Pack2)*Pack2 : 0; // Pack 3 packets if(Pack1>=3*PacketSize) @@ -1496,16 +1516,20 @@ EIGEN_DONT_INLINE void gemm_pack_lhs=Pack2) -// { -// if(PanelMode) count += Pack2*offset; -// for(Index k=0; k1) + { + for(Index i=peeled_mc1; i=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; // const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; - int pack_packets = Pack1/PacketSize; + int pack = Pack1; Index i = 0; - while(pack_packets>0) + while(pack>0) { - Index remaining_rows = rows-i; - Index peeled_mc = i+(remaining_rows/(pack_packets*PacketSize))*(pack_packets*PacketSize); -// std::cout << "pack_packets = " << pack_packets << " from " << i << " to " << peeled_mc << "\n"; - for(; i=PacketSize) { - for (Index m = 0; m < (pack_packets*PacketSize); m += PacketSize) + for(; k kernel; - for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = ploadu(&lhs(i+p+m, k)); - ptranspose(kernel); - for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack_packets*PacketSize)*p, cj.pconj(kernel.packet[p])); + for (Index m = 0; m < pack; m += PacketSize) + { + Kernel kernel; + for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = ploadu(&lhs(i+p+m, k)); + ptranspose(kernel); + for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p])); + } + count += PacketSize*pack; } - count += PacketSize*(pack_packets*PacketSize); } for(; k=Pack2) -// { -// if(PanelMode) count += Pack2*offset; -// for(Index k=0; k Date: Thu, 17 Apr 2014 16:26:35 +0200 Subject: [PATCH 0443/1103] Extend mixingtype unit test to check transposed cases. --- test/mixingtypes.cpp | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-) diff --git a/test/mixingtypes.cpp b/test/mixingtypes.cpp index ada2f69d3..1e0e2d4c1 100644 --- a/test/mixingtypes.cpp +++ b/test/mixingtypes.cpp @@ -97,11 +97,25 @@ template void mixingtypes(int size = SizeAtCompileType) VERIFY_IS_APPROX(sd*mcd*md, sd*mcd*md.template cast()); VERIFY_IS_APPROX(scd*md*mcd, scd*md.template cast().eval()*mcd); VERIFY_IS_APPROX(scd*mcd*md, scd*mcd*md.template cast()); - + VERIFY_IS_APPROX(sf*mf*mcf, sf*mf.template cast()*mcf); VERIFY_IS_APPROX(sf*mcf*mf, sf*mcf*mf.template cast()); VERIFY_IS_APPROX(scf*mf*mcf, scf*mf.template cast()*mcf); VERIFY_IS_APPROX(scf*mcf*mf, scf*mcf*mf.template cast()); + + VERIFY_IS_APPROX(sd*md.adjoint()*mcd, (sd*md).template cast().eval().adjoint()*mcd); + VERIFY_IS_APPROX(sd*mcd.adjoint()*md, sd*mcd.adjoint()*md.template cast()); + VERIFY_IS_APPROX(sd*md.adjoint()*mcd.adjoint(), (sd*md).template cast().eval().adjoint()*mcd.adjoint()); + VERIFY_IS_APPROX(sd*mcd.adjoint()*md.adjoint(), sd*mcd.adjoint()*md.template cast().adjoint()); + VERIFY_IS_APPROX(sd*md*mcd.adjoint(), (sd*md).template cast().eval()*mcd.adjoint()); + VERIFY_IS_APPROX(sd*mcd*md.adjoint(), sd*mcd*md.template cast().adjoint()); + + VERIFY_IS_APPROX(sf*mf.adjoint()*mcf, (sf*mf).template cast().eval().adjoint()*mcf); + VERIFY_IS_APPROX(sf*mcf.adjoint()*mf, sf*mcf.adjoint()*mf.template cast()); + VERIFY_IS_APPROX(sf*mf.adjoint()*mcf.adjoint(), (sf*mf).template cast().eval().adjoint()*mcf.adjoint()); + VERIFY_IS_APPROX(sf*mcf.adjoint()*mf.adjoint(), sf*mcf.adjoint()*mf.template cast().adjoint()); + VERIFY_IS_APPROX(sf*mf*mcf.adjoint(), (sf*mf).template cast().eval()*mcf.adjoint()); + VERIFY_IS_APPROX(sf*mcf*mf.adjoint(), sf*mcf*mf.template cast().adjoint()); VERIFY_IS_APPROX(sf*mf*vcf, (sf*mf).template cast().eval()*vcf); VERIFY_IS_APPROX(scf*mf*vcf,(scf*mf.template cast()).eval()*vcf); From 45a4aad572605f7db3348b1089801e0b8ea7cfd8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 16:27:22 +0200 Subject: [PATCH 0444/1103] add unit tests for ploadquad and predux4, and split packetmath unit test wrt real/complex --- test/packetmath.cpp | 31 +++++++++++++++++++++++++------ 1 file changed, 25 insertions(+), 6 deletions(-) diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 08566faf8..9ee2cad21 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -183,11 +183,30 @@ template void packetmath() VERIFY(areApprox(ref, data2, PacketSize) && "ploaddup"); } } + if(PacketSize>2) + { + for(int offset=0;offset<4;++offset) + { + for(int i=0;i(data1+offset)); + VERIFY(areApprox(ref, data2, PacketSize) && "ploadquad"); + } + } ref[0] = 0; for (int i=0; i(data1)), refvalue) && "internal::predux"); + + { + for (int i=0; i<4; ++i) + ref[i] = 0; + for (int i=0; i(data1))); + VERIFY(areApprox(ref, data2, PacketSize>4?PacketSize/2:PacketSize) && "internal::predux4"); + } ref[0] = 1; for (int i=0; i() ); CALL_SUBTEST_2( packetmath() ); CALL_SUBTEST_3( packetmath() ); - CALL_SUBTEST_1( packetmath >() ); - CALL_SUBTEST_2( packetmath >() ); + CALL_SUBTEST_4( packetmath >() ); + CALL_SUBTEST_5( packetmath >() ); CALL_SUBTEST_1( packetmath_notcomplex() ); CALL_SUBTEST_2( packetmath_notcomplex() ); @@ -406,13 +425,13 @@ void test_packetmath() CALL_SUBTEST_1( packetmath_real() ); CALL_SUBTEST_2( packetmath_real() ); - CALL_SUBTEST_1( packetmath_complex >() ); - CALL_SUBTEST_2( packetmath_complex >() ); + CALL_SUBTEST_4( packetmath_complex >() ); + CALL_SUBTEST_5( packetmath_complex >() ); CALL_SUBTEST_1( packetmath_scatter_gather() ); CALL_SUBTEST_2( packetmath_scatter_gather() ); CALL_SUBTEST_3( packetmath_scatter_gather() ); - CALL_SUBTEST_3( packetmath_scatter_gather >() ); - CALL_SUBTEST_3( packetmath_scatter_gather >() ); + CALL_SUBTEST_4( packetmath_scatter_gather >() ); + CALL_SUBTEST_5( packetmath_scatter_gather >() ); } } From 1dd015fea64048219aa4c2d616fb56e0c37bad47 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 16:27:58 +0200 Subject: [PATCH 0445/1103] Reduce block sizes in unit tests. --- test/main.h | 1 + 1 file changed, 1 insertion(+) diff --git a/test/main.h b/test/main.h index 9dd8bc535..fcac0e3ab 100644 --- a/test/main.h +++ b/test/main.h @@ -31,6 +31,7 @@ // B0 is defined in POSIX header termios.h #define B0 FORBIDDEN_IDENTIFIER +#define EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS // shuts down ICC's remark #593: variable "XXX" was set but never used #define TEST_SET_BUT_UNUSED_VARIABLE(X) X = X + 0; From e5d0cb54a5f2a2200a4656d993c82a80f159a7c4 Mon Sep 17 00:00:00 2001 From: Benjamin Chretien Date: Thu, 17 Apr 2014 18:49:23 +0200 Subject: [PATCH 0446/1103] Fix typo in Reductions tutorial. --- doc/TutorialReductionsVisitorsBroadcasting.dox | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/doc/TutorialReductionsVisitorsBroadcasting.dox b/doc/TutorialReductionsVisitorsBroadcasting.dox index 992cf6f34..eb6787dbc 100644 --- a/doc/TutorialReductionsVisitorsBroadcasting.dox +++ b/doc/TutorialReductionsVisitorsBroadcasting.dox @@ -32,7 +32,7 @@ Eigen also provides the \link MatrixBase::norm() norm() \endlink method, which r These operations can also operate on matrices; in that case, a n-by-p matrix is seen as a vector of size (n*p), so for example the \link MatrixBase::norm() norm() \endlink method returns the "Frobenius" or "Hilbert-Schmidt" norm. We refrain from speaking of the \f$\ell^2\f$ norm of a matrix because that can mean different things. -If you want other \f$\ell^p\f$ norms, use the \link MatrixBase::lpNorm() lpNnorm

() \endlink method. The template parameter \a p can take the special value \a Infinity if you want the \f$\ell^\infty\f$ norm, which is the maximum of the absolute values of the coefficients. +If you want other \f$\ell^p\f$ norms, use the \link MatrixBase::lpNorm() lpNorm

() \endlink method. The template parameter \a p can take the special value \a Infinity if you want the \f$\ell^\infty\f$ norm, which is the maximum of the absolute values of the coefficients. The following example demonstrates these methods. From 9746396d1b8d039d3d0d6537ad477135e5e9d3f5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 20:51:04 +0200 Subject: [PATCH 0447/1103] Optimize AVX pset1 for complexes and ploaddup --- Eigen/src/Core/arch/AVX/Complex.h | 14 ++++---------- Eigen/src/Core/arch/AVX/PacketMath.h | 20 ++++++++++++-------- 2 files changed, 16 insertions(+), 18 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 8f95a7be7..d0646e77d 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -78,11 +78,7 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploadu(const std::complex EIGEN_STRONG_INLINE Packet4cf pset1(const std::complex& from) { - const float r = std::real(from); - const float i = std::imag(from); - // Beware, _mm256_set_ps expects the scalar values in reverse order (i.e. 7 to 0) - const __m256 result = _mm256_set_ps(i, r, i, r, i, r, i, r); - return Packet4cf(result); + return Packet4cf(_mm256_castps_pd(_mm256_broadcast_sd((const double*)(const void*)&from))); } template<> EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex* from) @@ -304,11 +300,9 @@ template<> EIGEN_STRONG_INLINE Packet2cd ploadu(const std::complex EIGEN_STRONG_INLINE Packet2cd pset1(const std::complex& from) { - const double r = std::real(from); - const double i = std::imag(from); - // Beware, _mm256_set_pd expects the scalar values in reverse order (i.e. 3 to 0) - const __m256d result = _mm256_set_pd(i, r, i, r); - return Packet2cd(result); + // in case casting to a __m128d* is really not safe, then we can still fallback to this version: (much slower though) +// return Packet2cd(_mm256_loadu2_m128d((const double*)&from,(const double*)&from)); + return Packet2cd(_mm256_broadcast_pd((const __m128d*)(const void*)&from)); } template<> EIGEN_STRONG_INLINE Packet2cd ploaddup(const std::complex* from) { return pset1(*from); } diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 8d2e88061..a8b94e191 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -183,18 +183,22 @@ template<> EIGEN_STRONG_INLINE Packet8i ploadu(const int* from) { EIGE template<> EIGEN_STRONG_INLINE Packet8f ploaddup(const float* from) { // TODO try to find a way to avoid the need of a temporary register - Packet8f tmp = _mm256_castps128_ps256(_mm_loadu_ps(from)); - tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1); - return _mm256_unpacklo_ps(tmp,tmp); +// Packet8f tmp = _mm256_castps128_ps256(_mm_loadu_ps(from)); +// tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1); +// return _mm256_unpacklo_ps(tmp,tmp); + + // _mm256_insertf128_ps is very slow on Haswell, thus: + Packet8f tmp = _mm256_broadcast_ps((const __m128*)(const void*)from); + // mimic an "inplace" permutation of the lower 128bits using a blend + tmp = _mm256_blend_ps(tmp,_mm256_castps128_ps256(_mm_permute_ps( _mm256_castps256_ps128(tmp), _MM_SHUFFLE(1,0,1,0))), 15); + // then we can perform a consistent permutation on the global register to get everything in shape: + return _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2)); } // Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1} template<> EIGEN_STRONG_INLINE Packet4d ploaddup(const double* from) { - // TODO try to find a way to avoid the need of a temporary register - Packet2d tmp0 = _mm_loadu_pd(from); - Packet2d tmp1 = _mm_permute_pd(tmp0,3); - tmp0 = _mm_permute_pd(tmp0,0); - return _mm256_insertf128_pd(_mm256_castpd128_pd256(tmp0), tmp1, 1); + Packet4d tmp = _mm256_broadcast_pd((const __m128d*)(const void*)from); + return _mm256_permute_pd(tmp, 3<<2); } // Loads 2 floats from memory a returns the packet {a0, a0 a0, a0, a1, a1, a1, a1} From 9777a5ca60f0a82bb789f55912fd046ab7f3d15d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 21:01:45 +0200 Subject: [PATCH 0448/1103] Various minor fixes in BTL --- ...{FindOpenBLAS.cmake => FindOPENBLAS.cmake} | 0 bench/btl/generic_bench/bench_parameter.hh | 2 +- bench/btl/libs/BLAS/CMakeLists.txt | 2 +- bench/btl/libs/BLAS/main.cpp | 10 ++++---- bench/btl/libs/blaze/CMakeLists.txt | 6 +++-- bench/btl/libs/blaze/blaze_interface.hh | 24 +++++++++---------- bench/btl/libs/blaze/main.cpp | 8 +------ bench/btl/libs/eigen3/main_adv.cpp | 14 +++++------ 8 files changed, 31 insertions(+), 35 deletions(-) rename bench/btl/cmake/{FindOpenBLAS.cmake => FindOPENBLAS.cmake} (100%) diff --git a/bench/btl/cmake/FindOpenBLAS.cmake b/bench/btl/cmake/FindOPENBLAS.cmake similarity index 100% rename from bench/btl/cmake/FindOpenBLAS.cmake rename to bench/btl/cmake/FindOPENBLAS.cmake diff --git a/bench/btl/generic_bench/bench_parameter.hh b/bench/btl/generic_bench/bench_parameter.hh index 5e341c14c..0f62bd421 100644 --- a/bench/btl/generic_bench/bench_parameter.hh +++ b/bench/btl/generic_bench/bench_parameter.hh @@ -41,7 +41,7 @@ // min matrix size for LU bench #define MIN_LU 5 // max matrix size for LU bench -#define MAX_LU 5000 +#define MAX_LU 3000 // max size for tiny vector and matrix #define TINY_MV_MAX_SIZE 16 // default nb_sample for x86 timer diff --git a/bench/btl/libs/BLAS/CMakeLists.txt b/bench/btl/libs/BLAS/CMakeLists.txt index 22f09527d..0272ccad0 100644 --- a/bench/btl/libs/BLAS/CMakeLists.txt +++ b/bench/btl/libs/BLAS/CMakeLists.txt @@ -22,7 +22,7 @@ find_package(OPENBLAS) if (OPENBLAS_FOUND) btl_add_bench(btl_openblas main.cpp) if(BUILD_btl_openblas) - target_link_libraries(btl_openblas ${GOTO_LIBRARIES} ) + target_link_libraries(btl_openblas ${OPENBLAS_LIBRARIES} ) set_target_properties(btl_openblas PROPERTIES COMPILE_FLAGS "-DCBLASNAME=OPENBLAS") endif(BUILD_btl_openblas) endif (OPENBLAS_FOUND) diff --git a/bench/btl/libs/BLAS/main.cpp b/bench/btl/libs/BLAS/main.cpp index 8347c9f0b..564d55ef2 100644 --- a/bench/btl/libs/BLAS/main.cpp +++ b/bench/btl/libs/BLAS/main.cpp @@ -56,13 +56,13 @@ int main() bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); #ifdef HAS_LAPACK - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); +// bench > >(MIN_LU,MAX_LU,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); #endif //bench > >(MIN_LU,MAX_LU,NB_POINT); diff --git a/bench/btl/libs/blaze/CMakeLists.txt b/bench/btl/libs/blaze/CMakeLists.txt index 54ab929d8..f8b1b2ec3 100644 --- a/bench/btl/libs/blaze/CMakeLists.txt +++ b/bench/btl/libs/blaze/CMakeLists.txt @@ -4,5 +4,7 @@ find_package(Boost) if (BLAZE_FOUND AND Boost_FOUND) include_directories(${BLAZE_INCLUDE_DIR} ${Boost_INCLUDE_DIRS}) btl_add_bench(btl_blaze main.cpp) - target_link_libraries(btl_blaze ${Boost_LIBRARIES} ${Boost_system_LIBRARY} /opt/local/lib/libboost_system-mt.a ) -endif (BLAZE_FOUND) + if(BUILD_btl_blaze) + target_link_libraries(btl_blaze ${Boost_LIBRARIES} ${Boost_system_LIBRARY} /opt/local/lib/libboost_system-mt.a ) + endif() +endif () diff --git a/bench/btl/libs/blaze/blaze_interface.hh b/bench/btl/libs/blaze/blaze_interface.hh index 8020fef27..ed43ecdd4 100644 --- a/bench/btl/libs/blaze/blaze_interface.hh +++ b/bench/btl/libs/blaze/blaze_interface.hh @@ -84,25 +84,25 @@ public : X = (A*B); } -// static inline void transposed_matrix_matrix_product(const gene_matrix & A, const gene_matrix & B, gene_matrix & X, int N){ -// X = (trans(A)*trans(B)); -// } + static inline void transposed_matrix_matrix_product(const gene_matrix & A, const gene_matrix & B, gene_matrix & X, int N){ + X = A.transpose()*B.transpose(); + } -// static inline void ata_product(const gene_matrix & A, gene_matrix & X, int N){ -// X = (trans(A)*A); -// } + static inline void ata_product(const gene_matrix & A, gene_matrix & X, int N){ + X = (A.transpose()*A); + } -// static inline void aat_product(const gene_matrix & A, gene_matrix & X, int N){ -// X = (A*trans(A)); -// } + static inline void aat_product(const gene_matrix & A, gene_matrix & X, int N){ + X = (A*A.transpose()); + } static inline void matrix_vector_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){ X = (A*B); } -// static inline void atv_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){ -// X = (trans(A)*B); -// } + static inline void atv_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){ + X = (A.transpose()*B); + } static inline void axpy(const real coef, const gene_vector & X, gene_vector & Y, int N){ Y += coef * X; diff --git a/bench/btl/libs/blaze/main.cpp b/bench/btl/libs/blaze/main.cpp index b8508c8f3..582a2956b 100644 --- a/bench/btl/libs/blaze/main.cpp +++ b/bench/btl/libs/blaze/main.cpp @@ -19,8 +19,6 @@ #include "blaze_interface.hh" #include "bench.hh" #include "basic_actions.hh" -// #include "action_cholesky.hh" -// #include "action_lu_decomp.hh" BTL_MAIN; @@ -32,14 +30,10 @@ int main() bench > >(MIN_MV,MAX_MV,NB_POINT); // bench > >(MIN_MV,MAX_MV,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); +// bench > >(MIN_MM,MAX_MM,NB_POINT); // bench > >(MIN_MM,MAX_MM,NB_POINT); // bench > >(MIN_MM,MAX_MM,NB_POINT); -// bench > >(MIN_MM,MAX_MM,NB_POINT); -// bench > >(MIN_MM,MAX_MM,NB_POINT); -// bench > >(MIN_MM,MAX_MM,NB_POINT); - return 0; } diff --git a/bench/btl/libs/eigen3/main_adv.cpp b/bench/btl/libs/eigen3/main_adv.cpp index efe5857e4..95865357e 100644 --- a/bench/btl/libs/eigen3/main_adv.cpp +++ b/bench/btl/libs/eigen3/main_adv.cpp @@ -29,14 +29,14 @@ BTL_MAIN; int main() { - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); +// bench > >(MIN_LU,MAX_LU,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); - bench > >(MIN_MM,MAX_MM,NB_POINT); +// bench > >(MIN_LU,MAX_LU,NB_POINT); + bench > >(MIN_LU,MAX_LU,NB_POINT); return 0; } From c354bd47f7001bd6b3c43fc4a4b5d27f764aa5c3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 21:03:26 +0200 Subject: [PATCH 0449/1103] Make our gemm bench a little more powerful. --- bench/bench_gemm.cpp | 100 ++++++++++++++++++++++++++++++++----------- 1 file changed, 76 insertions(+), 24 deletions(-) diff --git a/bench/bench_gemm.cpp b/bench/bench_gemm.cpp index 1ef2e72c2..8222271fb 100644 --- a/bench/bench_gemm.cpp +++ b/bench/bench_gemm.cpp @@ -2,6 +2,14 @@ // g++-4.4 bench_gemm.cpp -I .. -O2 -DNDEBUG -lrt -fopenmp && OMP_NUM_THREADS=2 ./a.out // icpc bench_gemm.cpp -I .. -O3 -DNDEBUG -lrt -openmp && OMP_NUM_THREADS=2 ./a.out +// Compilation options: +// +// -DSCALAR=std::complex +// -DSCALARA=double or -DSCALARB=double +// -DHAVE_BLAS +// -DDECOUPLED +// + #include #include #include @@ -14,10 +22,18 @@ using namespace Eigen; #define SCALAR float #endif +#ifndef SCALARA +#define SCALARA SCALAR +#endif + +#ifndef SCALARB +#define SCALARB SCALAR +#endif + typedef SCALAR Scalar; typedef NumTraits::Real RealScalar; -typedef Matrix A; -typedef Matrix B; +typedef Matrix A; +typedef Matrix B; typedef Matrix C; typedef Matrix M; @@ -135,32 +151,49 @@ int main(int argc, char ** argv) int cache_size = -1; bool need_help = false; - for (int i=1; i c t p\n"; + std::cout << argv[0] << " -s -c -t -p \n"; + std::cout << " : size\n"; + std::cout << " : rows columns depth\n"; return 1; } @@ -182,6 +215,7 @@ int main(int argc, char ** argv) // check the parallel product is correct #if defined EIGEN_HAS_OPENMP + Eigen::initParallel(); int procs = omp_get_max_threads(); if(procs>1) { @@ -198,11 +232,20 @@ int main(int argc, char ** argv) #elif defined HAVE_BLAS blas_gemm(a,b,r); c.noalias() += a * b; - if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n"; + if(!r.isApprox(c)) { + std::cout << r - c << "\n"; + std::cerr << "Warning, your product is crap!\n\n"; + } #else - gemm(a,b,c); - r.noalias() += a.cast() * b.cast(); - if(!r.isApprox(c)) std::cerr << "Warning, your product is crap!\n\n"; + if(1.*m*n*p<2000.*2000*2000) + { + gemm(a,b,c); + r.noalias() += a.cast() .lazyProduct( b.cast() ); + if(!r.isApprox(c)) { + std::cout << r - c << "\n"; + std::cerr << "Warning, your product is crap!\n\n"; + } + } #endif #ifdef HAVE_BLAS @@ -224,7 +267,7 @@ int main(int argc, char ** argv) { BenchTimer tmono; omp_set_num_threads(1); - Eigen::internal::setNbThreads(1); + Eigen::setNbThreads(1); c = rc; BENCH(tmono, tries, rep, gemm(a,b,c)); std::cout << "eigen mono cpu " << tmono.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmono.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmono.total(CPU_TIMER) << "s)\n"; @@ -233,6 +276,15 @@ int main(int argc, char ** argv) } #endif + if(1.*m*n*p<30*30*30) + { + BenchTimer tmt; + c = rc; + BENCH(tmt, tries, rep, c.noalias()+=a.lazyProduct(b)); + std::cout << "lazy cpu " << tmt.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(CPU_TIMER) << "s)\n"; + std::cout << "lazy real " << tmt.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(REAL_TIMER) << "s)\n"; + } + #ifdef DECOUPLED if((NumTraits::IsComplex) && (NumTraits::IsComplex)) { From 6d6df90c9a813287d18fe045b78a7f4e3996ee5f Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 17 Apr 2014 12:28:01 -0700 Subject: [PATCH 0450/1103] Implemented the pgather/pscatter packet primitives for the arm/NEON architecture --- Eigen/src/Core/arch/NEON/Complex.h | 16 +++++++++++++ Eigen/src/Core/arch/NEON/PacketMath.h | 34 +++++++++++++++++++++++++++ 2 files changed, 50 insertions(+) diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 7ca76714f..7b94733ab 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -111,6 +111,22 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) +{ + Packet4f res; + res = vsetq_lane_f32(std::real(from[0*stride]), res, 0); + res = vsetq_lane_f32(std::imag(from[0*stride]), res, 1); + res = vsetq_lane_f32(std::real(from[1*stride]), res, 2); + res = vsetq_lane_f32(std::imag(from[1*stride]), res, 3); + return Packet2cf(res); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) +{ + to[stride*0] = std::complex(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1)); + to[stride*1] = std::complex(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3)); +} + template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((float *)addr); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet2cf& a) diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 83150507a..7f3301f51 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -221,6 +221,40 @@ template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& f template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); } +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) +{ + Packet4f res; + res = vsetq_lane_f32(from[0*stride], res, 0); + res = vsetq_lane_f32(from[1*stride], res, 1); + res = vsetq_lane_f32(from[2*stride], res, 2); + res = vsetq_lane_f32(from[3*stride], res, 3); + return res; +} +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, int stride) +{ + Packet4i res; + res = vsetq_lane_s32(from[0*stride], res, 0); + res = vsetq_lane_s32(from[1*stride], res, 1); + res = vsetq_lane_s32(from[2*stride], res, 2); + res = vsetq_lane_s32(from[3*stride], res, 3); + return res; +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, int stride) +{ + to[stride*0] = vgetq_lane_f32(from, 0); + to[stride*1] = vgetq_lane_f32(from, 1); + to[stride*2] = vgetq_lane_f32(from, 2); + to[stride*3] = vgetq_lane_f32(from, 3); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, int stride) +{ + to[stride*0] = vgetq_lane_s32(from, 0); + to[stride*1] = vgetq_lane_s32(from, 1); + to[stride*2] = vgetq_lane_s32(from, 2); + to[stride*3] = vgetq_lane_s32(from, 3); +} + template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { EIGEN_ARM_PREFETCH(addr); } template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { EIGEN_ARM_PREFETCH(addr); } From 6d665d446be19f233a2d94e05db206cf29a12191 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Apr 2014 23:26:34 +0200 Subject: [PATCH 0451/1103] Fixes for fixed sizes and non vectorizable types --- Eigen/src/Core/products/GeneralMatrixMatrix.h | 2 +- Eigen/src/Core/util/XprHelper.h | 1 + 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 38ca7b242..b45797f09 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -287,7 +287,7 @@ class gemm_blocking_spacem_mc = ActualRows; this->m_nc = ActualCols; diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index a08538aff..1b3e122e1 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -101,6 +101,7 @@ template struct packet_traits; template struct unpacket_traits { typedef T type; + typedef T half; enum {size=1}; }; From 1388f4f9fd3f8804128a5f86ead7e478e59a4749 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Apr 2014 11:43:13 +0200 Subject: [PATCH 0452/1103] Fix typo (was working with clang\!) --- Eigen/src/Core/arch/AVX/Complex.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index d0646e77d..18be2f225 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -78,7 +78,7 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploadu(const std::complex EIGEN_STRONG_INLINE Packet4cf pset1(const std::complex& from) { - return Packet4cf(_mm256_castps_pd(_mm256_broadcast_sd((const double*)(const void*)&from))); + return Packet4cf(_mm256_castpd_ps(_mm256_broadcast_sd((const double*)(const void*)&from))); } template<> EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex* from) From 94684721bd2d10c0b67ef518ed599981603440bc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Apr 2014 15:35:34 +0200 Subject: [PATCH 0453/1103] Smarter block size computation --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index dcc0b4a0d..b1ed9264a 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -106,7 +106,11 @@ void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) // In unit tests we do not want to use extra large matrices, // so we reduce the block size to check the blocking strategy is not flawed #ifndef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS - k = std::min(k,240); +// k = std::min(k,240); +// n = std::min(n,3840/sizeof(RhsScalar)); +// m = std::min(m,3840/sizeof(RhsScalar)); + + k = std::min(k,sizeof(LhsScalar)<=4 ? 360 : 240); n = std::min(n,3840/sizeof(RhsScalar)); m = std::min(m,3840/sizeof(RhsScalar)); #else From 3454b4e5f1eb1a68e15415eeca827a30cb3bb58e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Apr 2014 17:06:03 +0200 Subject: [PATCH 0454/1103] Fix calls to lazy products (lazy product does not like matrices with 0 length) --- Eigen/src/Core/products/GeneralMatrixMatrix.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index b45797f09..6ad07eccb 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -390,7 +390,7 @@ class GeneralProduct template inline void evalTo(Dest& dst) const { - if((m_rhs.rows()+dst.rows()+dst.cols())<20) + if((m_rhs.rows()+dst.rows()+dst.cols())<20 && m_rhs.rows()>0) dst.noalias() = m_lhs .lazyProduct( m_rhs ); else { @@ -402,7 +402,7 @@ class GeneralProduct template inline void addTo(Dest& dst) const { - if((m_rhs.rows()+dst.rows()+dst.cols())<20) + if((m_rhs.rows()+dst.rows()+dst.cols())<20 && m_rhs.rows()>0) dst.noalias() += m_lhs .lazyProduct( m_rhs ); else scaleAndAddTo(dst,Scalar(1)); @@ -411,7 +411,7 @@ class GeneralProduct template inline void subTo(Dest& dst) const { - if((m_rhs.rows()+dst.rows()+dst.cols())<20) + if((m_rhs.rows()+dst.rows()+dst.cols())<20 && m_rhs.rows()>0) dst.noalias() -= m_lhs .lazyProduct( m_rhs ); else scaleAndAddTo(dst,Scalar(-1)); From a7d20038df1dc160e282e824052ff5bcf5ba2c9a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Apr 2014 17:06:31 +0200 Subject: [PATCH 0455/1103] Fix alignment assertion. --- Eigen/src/Core/DenseStorage.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 7264b44c7..79033a1b5 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -88,7 +88,7 @@ struct plain_array EIGEN_DEVICE_FUNC plain_array() { - EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(0xf); + EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(EIGEN_ALIGN_BYTES-1); check_static_allocation_size(); } From 2606abed535744fcaa41b923c71338a06b8ed3fa Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Apr 2014 21:14:40 +0200 Subject: [PATCH 0456/1103] Fix 128bit packet size assumptions in unit tests. --- Eigen/src/Core/Assign.h | 2 ++ Eigen/src/Core/GeneralProduct.h | 3 +-- test/dynalloc.cpp | 21 ++++++++++++------- test/unalignedassert.cpp | 37 ++++++++++++++++++++------------- test/vectorization_logic.cpp | 23 ++++++++++---------- 5 files changed, 52 insertions(+), 34 deletions(-) diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index cefa6f3cc..07da2fe31 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -105,6 +105,8 @@ public: EIGEN_DEBUG_VAR(DstIsAligned) EIGEN_DEBUG_VAR(SrcIsAligned) EIGEN_DEBUG_VAR(JointAlignment) + EIGEN_DEBUG_VAR(Derived::SizeAtCompileTime) + EIGEN_DEBUG_VAR(OtherDerived::CoeffReadCost) EIGEN_DEBUG_VAR(InnerSize) EIGEN_DEBUG_VAR(InnerMaxSize) EIGEN_DEBUG_VAR(PacketSize) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index adda6f784..229d12c3f 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -66,8 +66,7 @@ template struct product_type MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime, _Rhs::MaxRowsAtCompileTime), Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, - _Rhs::RowsAtCompileTime), - LargeThreshold = EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD + _Rhs::RowsAtCompileTime) }; // the splitting into different lines of code here, introducing the _select enums and the typedef below, diff --git a/test/dynalloc.cpp b/test/dynalloc.cpp index 8bbda1c94..c98cc80f0 100644 --- a/test/dynalloc.cpp +++ b/test/dynalloc.cpp @@ -10,11 +10,13 @@ #include "main.h" #if EIGEN_ALIGN -#define ALIGNMENT 16 +#define ALIGNMENT EIGEN_ALIGN_BYTES #else #define ALIGNMENT 1 #endif +typedef Matrix Vector8f; + void check_handmade_aligned_malloc() { for(int i = 1; i < 1000; i++) @@ -68,7 +70,7 @@ struct MyStruct { EIGEN_MAKE_ALIGNED_OPERATOR_NEW char dummychar; - Vector4f avec; + Vector8f avec; }; class MyClassA @@ -76,15 +78,19 @@ class MyClassA public: EIGEN_MAKE_ALIGNED_OPERATOR_NEW char dummychar; - Vector4f avec; + Vector8f avec; }; template void check_dynaligned() { - T* obj = new T; - VERIFY(T::NeedsToAlign==1); - VERIFY(size_t(obj)%ALIGNMENT==0); - delete obj; + // TODO have to be updated once we support multiple alignment values + if(T::SizeAtCompileTime % ALIGNMENT == 0) + { + T* obj = new T; + VERIFY(T::NeedsToAlign==1); + VERIFY(size_t(obj)%ALIGNMENT==0); + delete obj; + } } void test_dynalloc() @@ -102,6 +108,7 @@ void test_dynalloc() CALL_SUBTEST(check_dynaligned() ); CALL_SUBTEST(check_dynaligned() ); CALL_SUBTEST(check_dynaligned() ); + CALL_SUBTEST(check_dynaligned() ); } // check static allocation, who knows ? diff --git a/test/unalignedassert.cpp b/test/unalignedassert.cpp index 601dbf214..d8815263a 100644 --- a/test/unalignedassert.cpp +++ b/test/unalignedassert.cpp @@ -9,6 +9,8 @@ #include "main.h" +typedef Matrix Vector8f; + struct TestNew1 { MatrixXd m; // good: m will allocate its own array, taking care of alignment. @@ -69,7 +71,7 @@ void construct_at_boundary(int boundary) { char buf[sizeof(T)+256]; size_t _buf = reinterpret_cast(buf); - _buf += (16 - (_buf % 16)); // make 16-byte aligned + _buf += (EIGEN_ALIGN_BYTES - (_buf % EIGEN_ALIGN_BYTES)); // make 16/32-byte aligned _buf += boundary; // make exact boundary-aligned T *x = ::new(reinterpret_cast(_buf)) T; x[0].setZero(); // just in order to silence warnings @@ -85,18 +87,18 @@ void unalignedassert() construct_at_boundary(16); construct_at_boundary(16); construct_at_boundary(4); - construct_at_boundary(16); + construct_at_boundary(EIGEN_ALIGN_BYTES); construct_at_boundary(16); construct_at_boundary(4); - construct_at_boundary(16); - construct_at_boundary(16); + construct_at_boundary(EIGEN_ALIGN_BYTES); + construct_at_boundary(EIGEN_ALIGN_BYTES); construct_at_boundary(4); - construct_at_boundary(16); + construct_at_boundary(EIGEN_ALIGN_BYTES); construct_at_boundary(16); construct_at_boundary(4); - construct_at_boundary(16); + construct_at_boundary(EIGEN_ALIGN_BYTES); construct_at_boundary(16); #endif @@ -110,14 +112,21 @@ void unalignedassert() check_unalignedassert_good >(); #if EIGEN_ALIGN_STATICALLY - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); - VERIFY_RAISES_ASSERT(construct_at_boundary(8)); + if(EIGEN_ALIGN_BYTES==16) + { + VERIFY_RAISES_ASSERT(construct_at_boundary(8)); + VERIFY_RAISES_ASSERT(construct_at_boundary(8)); + VERIFY_RAISES_ASSERT(construct_at_boundary(8)); + } + for(int b=8; b(b)); + VERIFY_RAISES_ASSERT(construct_at_boundary(b)); + VERIFY_RAISES_ASSERT(construct_at_boundary(b)); + VERIFY_RAISES_ASSERT(construct_at_boundary(b)); + VERIFY_RAISES_ASSERT(construct_at_boundary(b)); + VERIFY_RAISES_ASSERT(construct_at_boundary(b)); + } #endif } diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp index aee68a87f..09b46660b 100644 --- a/test/vectorization_logic.cpp +++ b/test/vectorization_logic.cpp @@ -90,8 +90,8 @@ template::Vectori typedef Matrix Matrix22; typedef Matrix Matrix44; typedef Matrix Matrix44u; - typedef Matrix Matrix44c; - typedef Matrix Matrix44r; + typedef Matrix Matrix44c; + typedef Matrix Matrix44r; typedef Matrix::Vectori LinearTraversal,CompleteUnrolling)); VERIFY(test_assign(Matrix33c().col(0),Matrix33c().col(1)+Matrix33c().col(1), LinearTraversal,CompleteUnrolling)); - + VERIFY(test_assign(Matrix3(),Matrix3().cwiseQuotient(Matrix3()), LinearVectorizedTraversal,CompleteUnrolling)); - + VERIFY(test_assign(Matrix(),Matrix()+Matrix(), LinearTraversal,NoUnrolling)); - - VERIFY(test_assign(Matrix11(),Matrix().template block(2,3)+Matrix().template block(10,4), - DefaultTraversal,CompleteUnrolling)); + + VERIFY(test_assign(Matrix11(),Matrix().template block(2,3)+Matrix().template block(8,4), + DefaultTraversal,PacketSize>4?InnerUnrolling:CompleteUnrolling)); } VERIFY(test_redux(Matrix3(), @@ -174,18 +174,19 @@ template::Vectori VERIFY(test_redux(Matrix44r().template block<1,2*PacketSize>(2,1), LinearVectorizedTraversal,CompleteUnrolling)); - + VERIFY((test_assign< Map >, Matrix22 >(InnerVectorizedTraversal,CompleteUnrolling))); VERIFY((test_assign< - Map >, - Matrix22 + Map, Aligned, InnerStride<3*PacketSize> >, + Matrix >(DefaultTraversal,CompleteUnrolling))); - VERIFY((test_assign(Matrix11(), Matrix11()*Matrix11(), InnerVectorizedTraversal, CompleteUnrolling))); + VERIFY((test_assign(Matrix11(), Matrix()*Matrix(), + PacketSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD?DefaultTraversal:InnerVectorizedTraversal, CompleteUnrolling))); #endif VERIFY(test_assign(MatrixXX(10,10),MatrixXX(20,20).block(10,10,2,3), From 5c5231ab71ba51432a24fdfd7d36a835b44805ea Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Apr 2014 16:03:19 +0200 Subject: [PATCH 0457/1103] Workaround gcc's default ABI not being able to distinghish between vector types of different sizes. --- CMakeLists.txt | 3 -- Eigen/src/Core/arch/SSE/Complex.h | 22 ++++---- Eigen/src/Core/arch/SSE/MathFunctions.h | 10 ++-- Eigen/src/Core/arch/SSE/PacketMath.h | 68 +++++++++++++++++-------- 4 files changed, 63 insertions(+), 40 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 838a41b79..fb13769f4 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -199,9 +199,6 @@ if(NOT MSVC) option(EIGEN_TEST_AVX "Enable/Disable AVX in tests/examples" OFF) if(EIGEN_TEST_AVX) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx") - if(CMAKE_COMPILER_IS_GNUCXX) - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fabi-version=6") - endif() message(STATUS "Enabling AVX in tests/examples") endif() diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 715e5a13c..d0c080c4f 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -110,8 +110,8 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex EIGEN_STRONG_INLINE Packet2cf ploaddup(const std::complex* from) { return pset1(*from); } -template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } -template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } +template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), Packet4f(from.v)); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); } template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) @@ -145,7 +145,7 @@ template<> EIGEN_STRONG_INLINE std::complex pfirst(const Pack #endif } -template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(_mm_castps_pd(a.v)))); } +template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v))))); } template<> EIGEN_STRONG_INLINE std::complex predux(const Packet2cf& a) { @@ -235,7 +235,7 @@ template<> struct conj_helper { return padd(c, pmul(x,y)); } EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const - { return Packet2cf(Eigen::internal::pmul(x, y.v)); } + { return Packet2cf(Eigen::internal::pmul(x, y.v)); } }; template<> struct conj_helper @@ -244,7 +244,7 @@ template<> struct conj_helper { return padd(c, pmul(x,y)); } EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const - { return Packet2cf(Eigen::internal::pmul(x.v, y)); } + { return Packet2cf(Eigen::internal::pmul(x.v, y)); } }; template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, const Packet2cf& b) @@ -300,7 +300,7 @@ template<> struct unpacket_traits { typedef std::complex type template<> EIGEN_STRONG_INLINE Packet1cd padd(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd psub(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); } -template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); @@ -338,8 +338,8 @@ template<> EIGEN_STRONG_INLINE Packet1cd pset1(const std::complex EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex* from) { return pset1(*from); } // FIXME force unaligned store, this is a temporary fix -template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } -template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); } template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } @@ -437,7 +437,7 @@ template<> struct conj_helper { return padd(c, pmul(x,y)); } EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const - { return Packet1cd(Eigen::internal::pmul(x, y.v)); } + { return Packet1cd(Eigen::internal::pmul(x, y.v)); } }; template<> struct conj_helper @@ -446,7 +446,7 @@ template<> struct conj_helper { return padd(c, pmul(x,y)); } EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const - { return Packet1cd(Eigen::internal::pmul(x.v, y)); } + { return Packet1cd(Eigen::internal::pmul(x.v, y)); } }; template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) @@ -459,7 +459,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, con EIGEN_STRONG_INLINE Packet1cd pcplxflip/**/(const Packet1cd& x) { - return Packet1cd(preverse(x.v)); + return Packet1cd(preverse(Packet2d(x.v))); } template<> EIGEN_DEVICE_FUNC inline void diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index 09f74c651..8f78b3a6c 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -63,7 +63,7 @@ Packet4f plog(const Packet4f& _x) x = _mm_or_ps(x, p4f_half); emm0 = _mm_sub_epi32(emm0, p4i_0x7f); - Packet4f e = padd(_mm_cvtepi32_ps(emm0), p4f_1); + Packet4f e = padd(Packet4f(_mm_cvtepi32_ps(emm0)), p4f_1); /* part2: if( x < SQRTHF ) { @@ -72,9 +72,9 @@ Packet4f plog(const Packet4f& _x) } else { x = x - 1.0; } */ Packet4f mask = _mm_cmplt_ps(x, p4f_cephes_SQRTHF); - Packet4f tmp = _mm_and_ps(x, mask); + Packet4f tmp = pand(x, mask); x = psub(x, p4f_1); - e = psub(e, _mm_and_ps(p4f_1, mask)); + e = psub(e, pand(p4f_1, mask)); x = padd(x, tmp); Packet4f x2 = pmul(x,x); @@ -167,7 +167,7 @@ Packet4f pexp(const Packet4f& _x) emm0 = _mm_cvttps_epi32(fx); emm0 = _mm_add_epi32(emm0, p4i_0x7f); emm0 = _mm_slli_epi32(emm0, 23); - return pmul(y, _mm_castsi128_ps(emm0)); + return pmul(y, Packet4f(_mm_castsi128_ps(emm0))); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d pexp(const Packet2d& _x) @@ -241,7 +241,7 @@ Packet2d pexp(const Packet2d& _x) emm0 = _mm_add_epi32(emm0, p4i_1023_0); emm0 = _mm_slli_epi32(emm0, 20); emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3)); - return pmul(x, _mm_castsi128_pd(emm0)); + return pmul(x, Packet2d(_mm_castsi128_pd(emm0))); } /* evaluation of 4 sines at onces, using SSE2 intrinsics. diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 89dfa6975..e7e035f4e 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -22,9 +22,35 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) #endif +#if defined EIGEN_VECTORIZE_AVX && defined __GNUC__ && !(defined __clang__ || defined __INTEL_COMPILER) +// With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot +// have overloads for both types without linking error. +// One solution is to increase ABI version using -fabi-version=4 (or greater). +// To workaround this inconvenince, we rather wrap 128bit types into the following helper +// structure: +// TODO disable this wrapper if abi-versio>=4, but to detect that without asking the user to define a macro? +template +struct eigen_packet_wrapper +{ + EIGEN_ALWAYS_INLINE operator T&() { return m_val; } + EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; } + EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {} + EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {} + EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) { + m_val = v; + return *this; + } + + T m_val; +}; +typedef eigen_packet_wrapper<__m128> Packet4f; +typedef eigen_packet_wrapper<__m128i> Packet4i; +typedef eigen_packet_wrapper<__m128d> Packet2d; +#else typedef __m128 Packet4f; typedef __m128i Packet4i; typedef __m128d Packet2d; +#endif template<> struct is_arithmetic<__m128> { enum { value = true }; }; template<> struct is_arithmetic<__m128i> { enum { value = true }; }; @@ -161,7 +187,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) } template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { - return psub(_mm_setr_epi32(0,0,0,0), a); + return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a); } template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } @@ -244,7 +270,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pandnot(const Packet4i& a, con template<> EIGEN_STRONG_INLINE Packet4f pload(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); } template<> EIGEN_STRONG_INLINE Packet2d pload(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); } -template<> EIGEN_STRONG_INLINE Packet4i pload(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast(from)); } +template<> EIGEN_STRONG_INLINE Packet4i pload(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast(from)); } #if defined(_MSC_VER) template<> EIGEN_STRONG_INLINE Packet4f ploadu(const float* from) { @@ -262,7 +288,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pload(const int* from) { E #endif } template<> EIGEN_STRONG_INLINE Packet2d ploadu(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); } - template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast(from)); } + template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast(from)); } #else // Fast unaligned loads. Note that here we cannot directly use intrinsics: this would // require pointer casting to incompatible pointer types and leads to invalid code @@ -312,7 +338,7 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD #if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS - return _mm_loadu_si128(reinterpret_cast(from)); + return _mm_loadu_si128(reinterpret_cast(from)); #else __m128d res; res = _mm_load_sd((const double*)(from)) ; @@ -331,13 +357,13 @@ template<> EIGEN_STRONG_INLINE Packet2d ploaddup(const double* from) template<> EIGEN_STRONG_INLINE Packet4i ploaddup(const int* from) { Packet4i tmp; - tmp = _mm_loadl_epi64(reinterpret_cast(from)); + tmp = _mm_loadl_epi64(reinterpret_cast(from)); return vec4i_swizzle1(tmp, 0, 0, 1, 1); } template<> EIGEN_STRONG_INLINE void pstore(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); } template<> EIGEN_STRONG_INLINE void pstore(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); } -template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast(to), from); } +template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); } template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE @@ -348,8 +374,8 @@ template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& _mm_storeh_pd((to+1), from); #endif } -template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), _mm_castps_pd(from)); } -template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), _mm_castsi128_pd(from)); } +template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), Packet2d(_mm_castps_pd(from))); } +template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), Packet2d(_mm_castsi128_pd(from))); } template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) { @@ -388,13 +414,13 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const template<> EIGEN_STRONG_INLINE void pstore1(float* to, const float& a) { Packet4f pa = _mm_set_ss(a); - pstore(to, vec4f_swizzle1(pa,0,0,0,0)); + pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0))); } // some compilers might be tempted to perform multiple moves instead of using a vector path. template<> EIGEN_STRONG_INLINE void pstore1(double* to, const double& a) { Packet2d pa = _mm_set_sd(a); - pstore(to, vec2d_swizzle1(pa,0,0)); + pstore(to, Packet2d(vec2d_swizzle1(pa,0,0))); } #ifndef EIGEN_VECTORIZE_AVX @@ -507,10 +533,10 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) template<> EIGEN_STRONG_INLINE float predux(const Packet4f& a) { Packet4f tmp0 = _mm_hadd_ps(a,a); - return pfirst(_mm_hadd_ps(tmp0, tmp0)); + return pfirst(_mm_hadd_ps(tmp0, tmp0)); } -template<> EIGEN_STRONG_INLINE double predux(const Packet2d& a) { return pfirst(_mm_hadd_pd(a, a)); } +template<> EIGEN_STRONG_INLINE double predux(const Packet2d& a) { return pfirst(_mm_hadd_pd(a, a)); } // SSSE3 version: // EIGEN_STRONG_INLINE float predux(const Packet4i& a) @@ -553,7 +579,7 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) template<> EIGEN_STRONG_INLINE int predux(const Packet4i& a) { Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a)); - return pfirst(tmp) + pfirst(_mm_shuffle_epi32(tmp, 1)); + return pfirst(tmp) + pfirst(_mm_shuffle_epi32(tmp, 1)); } template<> EIGEN_STRONG_INLINE Packet4i preduxp(const Packet4i* vecs) @@ -576,11 +602,11 @@ template<> EIGEN_STRONG_INLINE Packet4i preduxp(const Packet4i* vecs) template<> EIGEN_STRONG_INLINE float predux_mul(const Packet4f& a) { Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a)); - return pfirst(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); + return pfirst(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); } template<> EIGEN_STRONG_INLINE double predux_mul(const Packet2d& a) { - return pfirst(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); + return pfirst(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); } template<> EIGEN_STRONG_INLINE int predux_mul(const Packet4i& a) { @@ -596,17 +622,17 @@ template<> EIGEN_STRONG_INLINE int predux_mul(const Packet4i& a) template<> EIGEN_STRONG_INLINE float predux_min(const Packet4f& a) { Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a)); - return pfirst(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); + return pfirst(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); } template<> EIGEN_STRONG_INLINE double predux_min(const Packet2d& a) { - return pfirst(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); + return pfirst(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); } template<> EIGEN_STRONG_INLINE int predux_min(const Packet4i& a) { #ifdef EIGEN_VECTORIZE_SSE4_1 Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); - return pfirst(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); + return pfirst(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); #else // after some experiments, it is seems this is the fastest way to implement it // for GCC (eg., it does not like using std::min after the pstore !!) @@ -622,17 +648,17 @@ template<> EIGEN_STRONG_INLINE int predux_min(const Packet4i& a) template<> EIGEN_STRONG_INLINE float predux_max(const Packet4f& a) { Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a)); - return pfirst(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); + return pfirst(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); } template<> EIGEN_STRONG_INLINE double predux_max(const Packet2d& a) { - return pfirst(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); + return pfirst(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); } template<> EIGEN_STRONG_INLINE int predux_max(const Packet4i& a) { #ifdef EIGEN_VECTORIZE_SSE4_1 Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); - return pfirst(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); + return pfirst(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); #else // after some experiments, it is seems this is the fastest way to implement it // for GCC (eg., it does not like using std::min after the pstore !!) From ecbd67a15a55372f189d3718ff8661791bfa1c9f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Apr 2014 17:03:57 +0200 Subject: [PATCH 0458/1103] Fix EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT macro --- Eigen/src/Core/DenseStorage.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 79033a1b5..94f796783 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -68,13 +68,13 @@ struct plain_array template EIGEN_ALWAYS_INLINE PtrType eigen_unaligned_array_assert_workaround_gcc47(PtrType array) { return array; } #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \ - eigen_assert((reinterpret_cast(eigen_unaligned_array_assert_workaround_gcc47(array)) & sizemask) == 0 \ + eigen_assert((reinterpret_cast(eigen_unaligned_array_assert_workaround_gcc47(array)) & (sizemask)) == 0 \ && "this assertion is explained here: " \ "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \ " **** READ THIS WEB PAGE !!! ****"); #else #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \ - eigen_assert((reinterpret_cast(array) & sizemask) == 0 \ + eigen_assert((reinterpret_cast(array) & (sizemask)) == 0 \ && "this assertion is explained here: " \ "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \ " **** READ THIS WEB PAGE !!! ****"); From 82b09fcb911c5a0734ccf441504269f33034194a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 23 Apr 2014 13:09:26 +0200 Subject: [PATCH 0459/1103] Add Altivec implementation of pgather/pscatter (not tested) --- Eigen/src/Core/arch/AltiVec/Complex.h | 16 ++++++++++ Eigen/src/Core/arch/AltiVec/PacketMath.h | 37 ++++++++++++++++++++++++ 2 files changed, 53 insertions(+) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index db52074f4..ec11cfaa0 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -66,6 +66,22 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) +{ + std::complex EIGEN_ALIGN16 af[2]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return Packet2cf(vec_ld(0, af)); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) +{ + std::complex EIGEN_ALIGN16 af[2]; + vec_st(from.v, 0, af); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + + template<> EIGEN_STRONG_INLINE Packet2cf padd(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_add(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf psub(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(vec_sub(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(a.v)); } diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 16948264f..80a99a004 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -163,6 +163,43 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return vc; } +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) +{ + float EIGEN_ALIGN16 af[4]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + af[2] = from[2*stride]; + af[3] = from[3*stride]; + return vec_ld(0, af); +} +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, int stride) +{ + int EIGEN_ALIGN16 ai[4]; + ai[0] = from[0*stride]; + ai[1] = from[1*stride]; + ai[2] = from[2*stride]; + ai[3] = from[3*stride]; + return vec_ld(0, ai); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, int stride) +{ + float EIGEN_ALIGN16 af[4]; + vec_st(from, 0, af); + to[0*stride] = af[0]; + to[1*stride] = af[1]; + to[2*stride] = af[2]; + to[3*stride] = af[3]; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, int stride) +{ + int EIGEN_ALIGN16 ai[4]; + vec_st(from, 0, ai); + to[0*stride] = ai[0]; + to[1*stride] = ai[1]; + to[2*stride] = ai[2]; + to[3*stride] = ai[3]; +} + template<> EIGEN_STRONG_INLINE Packet4f plset(const float& a) { return vec_add(pset1(a), p4f_COUNTDOWN); } template<> EIGEN_STRONG_INLINE Packet4i plset(const int& a) { return vec_add(pset1(a), p4i_COUNTDOWN); } From ccb4dec719c1dffbd3565497cd897abb1b9f730f Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 23 Apr 2014 18:22:10 -0700 Subject: [PATCH 0460/1103] Created a NEON version of the ptranspose packet primitives --- Eigen/src/Core/arch/NEON/Complex.h | 8 ++++++++ Eigen/src/Core/arch/NEON/PacketMath.h | 23 ++++++++++++++++++++++- 2 files changed, 30 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 7b94733ab..a668382e2 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -263,6 +263,14 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, con return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); } +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + float32x4_t tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v)); + kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v)); + kernel.packet[1].v = tmp; +} + + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index c9b9e5e9b..6a6876e69 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -447,9 +447,30 @@ PALIGN_NEON(0,Packet4i,vextq_s32) PALIGN_NEON(1,Packet4i,vextq_s32) PALIGN_NEON(2,Packet4i,vextq_s32) PALIGN_NEON(3,Packet4i,vextq_s32) - + #undef PALIGN_NEON +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]); + float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]); + + kernel.packet[0] = vcombine_f32(vget_low_f32(tmp1.val[0]), vget_low_f32(tmp2.val[0])); + kernel.packet[1] = vcombine_f32(vget_high_f32(tmp1.val[0]), vget_high_f32(tmp2.val[0])); + kernel.packet[2] = vcombine_f32(vget_low_f32(tmp1.val[1]), vget_low_f32(tmp2.val[1])); + kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1])); +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]); + int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]); + kernel.packet[0] = vcombine_s32(vget_low_s32(tmp1.val[0]), vget_low_s32(tmp2.val[0])); + kernel.packet[1] = vcombine_s32(vget_high_s32(tmp1.val[0]), vget_high_s32(tmp2.val[0])); + kernel.packet[2] = vcombine_s32(vget_low_s32(tmp1.val[1]), vget_low_s32(tmp2.val[1])); + kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1])); +} + } // end namespace internal } // end namespace Eigen From 4eb92e564710cd55323302d31e47aff74d770426 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 23 Apr 2014 18:23:07 -0700 Subject: [PATCH 0461/1103] Fixed the NEON implementation of predux_max. --- Eigen/src/Core/arch/NEON/PacketMath.h | 1 + 1 file changed, 1 insertion(+) diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 6a6876e69..6426623cf 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -422,6 +422,7 @@ template<> EIGEN_STRONG_INLINE int predux_max(const Packet4i& a) a_lo = vget_low_s32(a); a_hi = vget_high_s32(a); max = vpmax_s32(a_lo, a_hi); + max = vpmax_s32(max, max); return vget_lane_s32(max, 0); } From 8d85ce88e129d794d0700dd2c8eec2713449e54d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 24 Apr 2014 05:47:53 -0700 Subject: [PATCH 0462/1103] Implement ptranspose on altivec and fix pgather/pscatter --- Eigen/src/Core/arch/AltiVec/Complex.h | 15 ++++++++++--- Eigen/src/Core/arch/AltiVec/PacketMath.h | 27 ++++++++++++++++++++++++ 2 files changed, 39 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index ec11cfaa0..ee1f008b1 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -21,6 +21,8 @@ static Packet16uc p16uc_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8); static Packet16uc p16uc_COMPLEX_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);//{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_PSET_HI = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 1));//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 2), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 3));//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; +static Packet16uc p16uc_COMPLEX_TRANSPOSE_0 = { 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; +static Packet16uc p16uc_COMPLEX_TRANSPOSE_1 = { 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; //---------- float ---------- struct Packet2cf @@ -52,7 +54,7 @@ template<> struct packet_traits > : default_packet_traits }; }; -template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; }; +template<> struct unpacket_traits { typedef std::complex type; enum {size=2}; typedef Packet2cf half; }; template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex& from) { @@ -71,12 +73,12 @@ template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packe std::complex EIGEN_ALIGN16 af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; - return Packet2cf(vec_ld(0, af)); + return Packet2cf(vec_ld(0, (const float*)af)); } template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) { std::complex EIGEN_ALIGN16 af[2]; - vec_st(from.v, 0, af); + vec_st(from.v, 0, (float*)af); to[0*stride] = af[0]; to[1*stride] = af[1]; } @@ -227,6 +229,13 @@ template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip(const Packet2cf& x return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); } +template<> EIGEN_STRONG_INLINE void ptranspose(Kernel& kernel) +{ + Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); + kernel.packet[0].v = tmp; +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 80a99a004..618d95d85 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -146,6 +146,7 @@ inline std::ostream & operator <<(std::ostream & s, const Packetbi & v) return s; } */ + template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html float EIGEN_ALIGN16 af[4]; @@ -533,6 +534,32 @@ struct palign_impl } }; +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + Packet4f t0, t1, t2, t3; + t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); + t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); + t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); + t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); + kernel.packet[0] = vec_mergeh(t0, t2); + kernel.packet[1] = vec_mergel(t0, t2); + kernel.packet[2] = vec_mergeh(t1, t3); + kernel.packet[3] = vec_mergel(t1, t3); +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(Kernel& kernel) { + Packet4i t0, t1, t2, t3; + t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); + t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); + t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]); + t3 = vec_mergel(kernel.packet[1], kernel.packet[3]); + kernel.packet[0] = vec_mergeh(t0, t2); + kernel.packet[1] = vec_mergel(t0, t2); + kernel.packet[2] = vec_mergeh(t1, t3); + kernel.packet[3] = vec_mergel(t1, t3); +} + } // end namespace internal } // end namespace Eigen From b0e19db1cf462a07e25429d4f04f7d8e858f670f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 24 Apr 2014 23:17:18 +0200 Subject: [PATCH 0463/1103] Enable fused madd for Altivec --- Eigen/src/Core/arch/AVX/PacketMath.h | 6 ++++++ Eigen/src/Core/arch/AltiVec/PacketMath.h | 4 ++++ Eigen/src/Core/arch/SSE/PacketMath.h | 6 ++++++ Eigen/src/Core/products/GeneralBlockPanelKernel.h | 10 +++++----- 4 files changed, 21 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index a8b94e191..e6f540430 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -22,6 +22,12 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) #endif +#ifdef EIGEN_VECTORIZE_FMA +#ifndef EIGEN_HAS_FUSED_MADD +#define EIGEN_HAS_FUSED_MADD 1 +#endif +#endif + typedef __m256 Packet8f; typedef __m256i Packet8i; typedef __m256d Packet4d; diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 618d95d85..e26d88382 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -18,6 +18,10 @@ namespace internal { #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4 #endif +#ifndef EIGEN_HAS_FUSED_MADD +#define EIGEN_HAS_FUSED_MADD 1 +#endif + #ifndef EIGEN_HAS_FUSE_CJMADD #define EIGEN_HAS_FUSE_CJMADD 1 #endif diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index e7e035f4e..4f9d8c4fd 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -22,6 +22,12 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) #endif +#ifdef EIGEN_VECTORIZE_FMA +#ifndef EIGEN_HAS_FUSED_MADD +#define EIGEN_HAS_FUSED_MADD 1 +#endif +#endif + #if defined EIGEN_VECTORIZE_AVX && defined __GNUC__ && !(defined __clang__ || defined __INTEL_COMPILER) // With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot // have overloads for both types without linking error. diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index b1ed9264a..60251f624 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -188,7 +188,7 @@ public: nr = 4, // register block size along the M direction (currently, this one cannot be modified) -#ifdef __FMA__ +#ifdef EIGEN_HAS_FUSED_MADD // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -254,7 +254,7 @@ public: // let gcc allocate the register in which to store the result of the pmul // (in the case where there is no FMA) gcc fails to figure out how to avoid // spilling register. -#ifdef EIGEN_VECTORIZE_FMA +#ifdef EIGEN_HAS_FUSED_MADD EIGEN_UNUSED_VARIABLE(tmp); c = pmadd(a,b,c); #else @@ -296,7 +296,7 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, nr = 4, -#ifdef __FMA__ +#ifdef EIGEN_HAS_FUSED_MADD // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -359,7 +359,7 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { -#ifdef EIGEN_VECTORIZE_FMA +#ifdef EIGEN_HAS_FUSED_MADD EIGEN_UNUSED_VARIABLE(tmp); c.v = pmadd(a.v,b,c.v); #else @@ -635,7 +635,7 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { -#ifdef EIGEN_VECTORIZE_FMA +#ifdef EIGEN_HAS_FUSED_MADD EIGEN_UNUSED_VARIABLE(tmp); c.v = pmadd(a,b.v,c.v); #else From 3d8d0f6269478a06f4fcbd4b838c8e9b9d7e9d62 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 10:56:18 +0200 Subject: [PATCH 0464/1103] Enable vectorization of pack_rhs with a column-major RHS. Rename and generalize Kernel<*> to PacketBlock<*,N>. --- Eigen/src/Core/GenericPacketMath.h | 10 +++---- Eigen/src/Core/arch/AVX/Complex.h | 8 +++--- Eigen/src/Core/arch/AVX/PacketMath.h | 26 ++++++++++++++++--- Eigen/src/Core/arch/AltiVec/Complex.h | 2 +- Eigen/src/Core/arch/AltiVec/PacketMath.h | 4 +-- Eigen/src/Core/arch/NEON/Complex.h | 2 +- Eigen/src/Core/arch/NEON/PacketMath.h | 4 +-- Eigen/src/Core/arch/SSE/Complex.h | 4 +-- Eigen/src/Core/arch/SSE/PacketMath.h | 12 ++++----- .../Core/products/GeneralBlockPanelKernel.h | 25 ++++++++++-------- test/packetmath.cpp | 10 +++---- 11 files changed, 64 insertions(+), 43 deletions(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 147298009..98313c68f 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -417,15 +417,15 @@ template<> inline std::complex pmul(const std::complex& a, const /*************************************************************************** - * Kernel, that is a collection of N packets where N is the number of words - * in the packet. + * PacketBlock, that is a collection of N packets where the number of words + * in the packet is a multiple of N. ***************************************************************************/ -template struct Kernel { - Packet packet[unpacket_traits::size]; +template ::size> struct PacketBlock { + Packet packet[N]; }; template EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& /*kernel*/) { +ptranspose(PacketBlock& /*kernel*/) { // Nothing to do in the scalar case, i.e. a 1x1 matrix. } diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 18be2f225..9ced85132 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -431,8 +431,8 @@ template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip(const Packet2cd& x return Packet2cd(_mm256_shuffle_pd(x.v, x.v, 0x5)); } -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { __m256d P0 = _mm256_castps_pd(kernel.packet[0].v); __m256d P1 = _mm256_castps_pd(kernel.packet[1].v); __m256d P2 = _mm256_castps_pd(kernel.packet[2].v); @@ -449,8 +449,8 @@ ptranspose(Kernel& kernel) { kernel.packet[2].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 49)); } -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { __m256d tmp = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 0+(2<<4)); kernel.packet[1].v = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 1+(3<<4)); kernel.packet[0].v = tmp; diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index e6f540430..8b8307d75 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -498,8 +498,8 @@ struct palign_impl } }; -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); @@ -526,8 +526,26 @@ ptranspose(Kernel& kernel) { kernel.packet[7] = _mm256_permute2f128_ps(S3, S7, 0x31); } -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { + __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); + __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); + __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); + __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); + + __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); + __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); + __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); + __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); + + kernel.packet[0] = _mm256_permute2f128_ps(S0, S1, 0x20); + kernel.packet[1] = _mm256_permute2f128_ps(S2, S3, 0x20); + kernel.packet[2] = _mm256_permute2f128_ps(S0, S1, 0x31); + kernel.packet[3] = _mm256_permute2f128_ps(S2, S3, 0x31); +} + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { __m256d T0 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 15); __m256d T1 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 0); __m256d T2 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 15); diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index ee1f008b1..8fdffad5e 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -229,7 +229,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip(const Packet2cf& x return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); } -template<> EIGEN_STRONG_INLINE void ptranspose(Kernel& kernel) +template<> EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) { Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index e26d88382..8a67354e4 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -539,7 +539,7 @@ struct palign_impl }; template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +ptranspose(PacketBlock& kernel) { Packet4f t0, t1, t2, t3; t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); @@ -552,7 +552,7 @@ ptranspose(Kernel& kernel) { } template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +ptranspose(PacketBlock& kernel) { Packet4i t0, t1, t2, t3; t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); t1 = vec_mergel(kernel.packet[0], kernel.packet[2]); diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index a668382e2..3c60735a9 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -264,7 +264,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, con } template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +ptranspose(PacketBlock& kernel) { float32x4_t tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v)); kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v)); kernel.packet[1].v = tmp; diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 6426623cf..37fb840d4 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -452,7 +452,7 @@ PALIGN_NEON(3,Packet4i,vextq_s32) #undef PALIGN_NEON template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +ptranspose(PacketBlock& kernel) { float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]); float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]); @@ -463,7 +463,7 @@ ptranspose(Kernel& kernel) { } template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +ptranspose(PacketBlock& kernel) { int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]); int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]); kernel.packet[0] = vcombine_s32(vget_low_s32(tmp1.val[0]), vget_low_s32(tmp2.val[0])); diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index d0c080c4f..758183c18 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -462,8 +462,8 @@ EIGEN_STRONG_INLINE Packet1cd pcplxflip/**/(const Packet1cd& x) return Packet1cd(preverse(Packet2d(x.v))); } -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { __m128d w1 = _mm_castps_pd(kernel.packet[0].v); __m128d w2 = _mm_castps_pd(kernel.packet[1].v); diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 4f9d8c4fd..ad935d5f1 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -784,20 +784,20 @@ struct palign_impl }; #endif -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]); } -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]); kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]); kernel.packet[1] = tmp; } -template<> EIGEN_DEVICE_FUNC inline void -ptranspose(Kernel& kernel) { +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]); __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]); __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]); diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 60251f624..0a94f25e4 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -1585,7 +1585,7 @@ EIGEN_DONT_INLINE void gemm_pack_lhs kernel; + PacketBlock kernel; for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = ploadu(&lhs(i+p+m, k)); ptranspose(kernel); for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p])); @@ -1675,7 +1675,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs kernel; +// PacketBlock kernel; // for (int p = 0; p < PacketSize; ++p) { // kernel.packet[p] = ploadu(&rhs[(j2+p)*rhsStride+k]); // } @@ -1713,19 +1713,22 @@ EIGEN_DONT_INLINE void gemm_pack_rhs kernel; - for (int p = 0; p < PacketSize; ++p) { - kernel.packet[p] = ploadu(&rhs[(j2+p)*rhsStride+k]); - } + PacketBlock kernel; + kernel.packet[0] = ploadu(&b0[k]); + kernel.packet[1] = ploadu(&b1[k]); + kernel.packet[2] = ploadu(&b2[k]); + kernel.packet[3] = ploadu(&b3[k]); ptranspose(kernel); - for (int p = 0; p < PacketSize; ++p) { - pstoreu(blockB+count, cj.pconj(kernel.packet[p])); - count+=PacketSize; - } + pstoreu(blockB+count+0*PacketSize, cj.pconj(kernel.packet[0])); + pstoreu(blockB+count+1*PacketSize, cj.pconj(kernel.packet[1])); + pstoreu(blockB+count+2*PacketSize, cj.pconj(kernel.packet[2])); + pstoreu(blockB+count+3*PacketSize, cj.pconj(kernel.packet[3])); + count+=4*PacketSize; } } for(; k void packetmath() internal::pstore(data2, internal::preverse(internal::pload(data1))); VERIFY(areApprox(ref, data2, PacketSize) && "internal::preverse"); - internal::Kernel kernel; + internal::PacketBlock kernel; for (int i=0; i(data1+i*PacketSize); } @@ -236,7 +236,7 @@ template void packetmath() for (int i=0; i void packetmath_scatter_gather() { for (int i = 0; i < PacketSize*11; ++i) { if ((i%11) == 0) { - VERIFY(isApproxAbs(buffer[i], data1[i/11], refvalue)); + VERIFY(isApproxAbs(buffer[i], data1[i/11], refvalue) && "pscatter"); } else { - VERIFY(isApproxAbs(buffer[i], Scalar(0), refvalue)); + VERIFY(isApproxAbs(buffer[i], Scalar(0), refvalue) && "pscatter"); } } @@ -405,7 +405,7 @@ template void packetmath_scatter_gather() { packet = internal::pgather(buffer, 7); internal::pstore(data1, packet); for (int i = 0; i < PacketSize; ++i) { - VERIFY(isApproxAbs(data1[i], buffer[i*7], refvalue)); + VERIFY(isApproxAbs(data1[i], buffer[i*7], refvalue) && "pgather"); } } From cf7eaed38db25c5205f6aabfd4e96e774fd98af5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 11:04:02 +0200 Subject: [PATCH 0465/1103] Avoid blocking-size mismatch in unit tests calling Eigen's blas interface. --- test/cholmod_support.cpp | 1 + test/main.h | 4 ++++ test/pastix_support.cpp | 2 ++ test/spqr_support.cpp | 2 ++ test/superlu_support.cpp | 1 + test/umfpack_support.cpp | 1 + 6 files changed, 11 insertions(+) diff --git a/test/cholmod_support.cpp b/test/cholmod_support.cpp index 8f8be3c0e..87f119b1e 100644 --- a/test/cholmod_support.cpp +++ b/test/cholmod_support.cpp @@ -7,6 +7,7 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +#define EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS #include "sparse_solver.h" #include diff --git a/test/main.h b/test/main.h index fcac0e3ab..3ccc2ae88 100644 --- a/test/main.h +++ b/test/main.h @@ -31,7 +31,11 @@ // B0 is defined in POSIX header termios.h #define B0 FORBIDDEN_IDENTIFIER +// Unit tests calling Eigen's blas library must preserve the default blocking size +// to avoid troubles. +#ifndef EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS #define EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS +#endif // shuts down ICC's remark #593: variable "XXX" was set but never used #define TEST_SET_BUT_UNUSED_VARIABLE(X) X = X + 0; diff --git a/test/pastix_support.cpp b/test/pastix_support.cpp index 14da0944b..49239e3a5 100644 --- a/test/pastix_support.cpp +++ b/test/pastix_support.cpp @@ -7,6 +7,8 @@ // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#define EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS #include "sparse_solver.h" #include #include diff --git a/test/spqr_support.cpp b/test/spqr_support.cpp index b8980e081..901c42c40 100644 --- a/test/spqr_support.cpp +++ b/test/spqr_support.cpp @@ -5,6 +5,8 @@ // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed + +#define EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS #include "sparse.h" #include diff --git a/test/superlu_support.cpp b/test/superlu_support.cpp index 3b16135bc..98a7bc5c8 100644 --- a/test/superlu_support.cpp +++ b/test/superlu_support.cpp @@ -7,6 +7,7 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +#define EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS #include "sparse_solver.h" #include diff --git a/test/umfpack_support.cpp b/test/umfpack_support.cpp index 9eb84c14b..37ab11f0b 100644 --- a/test/umfpack_support.cpp +++ b/test/umfpack_support.cpp @@ -7,6 +7,7 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +#define EIGEN_NO_DEBUG_SMALL_PRODUCT_BLOCKS #include "sparse_solver.h" #include From c79bd4b64bcc7c66d2b53a9a668a3310a3ae2998 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 11:06:03 +0200 Subject: [PATCH 0466/1103] Minor optimizations in product kernel: - use pbroadcast4 (helpful when AVX is not available) - process all remaining rows at once (significant speedup for small matrices) --- .../Core/products/GeneralBlockPanelKernel.h | 166 ++++++++---------- 1 file changed, 75 insertions(+), 91 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 0a94f25e4..e76c12c39 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -214,11 +214,11 @@ public: p = pset1(ResScalar(0)); } -// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) -// { -// pbroadcast4(b, b0, b1, b2, b3); -// } -// + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + { + pbroadcast4(b, b0, b1, b2, b3); + } + // EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) // { // pbroadcast2(b, b0, b1); @@ -342,11 +342,11 @@ public: dest = ploadu(a); } -// EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) -// { -// pbroadcast4(b, b0, b1, b2, b3); -// } -// + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + { + pbroadcast4(b, b0, b1, b2, b3); + } + // EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) // { // pbroadcast2(b, b0, b1); @@ -713,11 +713,9 @@ void gebp_kernel const Index peeled_mc1 = mr>=1*Traits::LhsProgress ? (rows/(1*LhsProgress))*(1*LhsProgress) : 0; enum { pk = 8 }; // NOTE Such a large peeling factor is important for large matrices (~ +5% when >1000 on Haswell) const Index peeled_kc = depth & ~(pk-1); + const Index prefetch_res_offset = 32/sizeof(ResScalar); // const Index depth2 = depth & ~1; -// std::cout << mr << " " << peeled_mc3 << " " << peeled_mc2 << " " << peeled_mc1 << "\n"; - - //---------- Process 3 * LhsProgress rows at once ---------- // This corresponds to 3*LhsProgress x nr register blocks. // Usually, make sense only with FMA @@ -736,19 +734,12 @@ void gebp_kernel prefetch(&blA[0]); // gets res block as register - AccPacket C0, C1, C2, C3, C4, C5, C6, C7, C8, C9, C10, C11; - traits.initAcc(C0); - traits.initAcc(C1); - traits.initAcc(C2); - traits.initAcc(C3); - traits.initAcc(C4); - traits.initAcc(C5); - traits.initAcc(C6); - traits.initAcc(C7); - traits.initAcc(C8); - traits.initAcc(C9); - traits.initAcc(C10); - traits.initAcc(C11); + AccPacket C0, C1, C2, C3, + C4, C5, C6, C7, + C8, C9, C10, C11; + traits.initAcc(C0); traits.initAcc(C1); traits.initAcc(C2); traits.initAcc(C3); + traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7); + traits.initAcc(C8); traits.initAcc(C9); traits.initAcc(C10); traits.initAcc(C11); ResScalar* r0 = &res[(j2+0)*resStride + i]; ResScalar* r1 = &res[(j2+1)*resStride + i]; @@ -767,7 +758,6 @@ void gebp_kernel for(Index k=0; k prefetch(&blA[0]); // gets res block as register - AccPacket C0, C1, C2, C3, C4, C5, C6, C7; - traits.initAcc(C0); - traits.initAcc(C1); - traits.initAcc(C2); - traits.initAcc(C3); - traits.initAcc(C4); - traits.initAcc(C5); - traits.initAcc(C6); - traits.initAcc(C7); + AccPacket C0, C1, C2, C3, + C4, C5, C6, C7; + traits.initAcc(C0); traits.initAcc(C1); traits.initAcc(C2); traits.initAcc(C3); + traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7); ResScalar* r0 = &res[(j2+0)*resStride + i]; ResScalar* r1 = &res[(j2+1)*resStride + i]; ResScalar* r2 = &res[(j2+2)*resStride + i]; ResScalar* r3 = &res[(j2+3)*resStride + i]; - internal::prefetch(r0); - internal::prefetch(r1); - internal::prefetch(r2); - internal::prefetch(r3); + internal::prefetch(r0+prefetch_res_offset); + internal::prefetch(r1+prefetch_res_offset); + internal::prefetch(r2+prefetch_res_offset); + internal::prefetch(r3+prefetch_res_offset); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; @@ -977,26 +962,22 @@ void gebp_kernel for(Index k=0; k blB += pk*4*RhsProgress; blA += pk*(2*Traits::LhsProgress); - IACA_END } // process remaining peeled loop for(Index k=peeled_kc; k traits.initAcc(C4); ResScalar* r0 = &res[(j2+0)*resStride + i]; + internal::prefetch(r0+prefetch_res_offset); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB]; @@ -1079,7 +1060,7 @@ void gebp_kernel traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0); \ traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1); \ traits.loadRhs(&blB[(0+K)*RhsProgress], B_0); \ - traits.madd(A0, B_0, C0, B1); \ + traits.madd(A0, B_0, C0, B1); \ traits.madd(A1, B_0, C4, B_0) EIGEN_GEBGP_ONESTEP(0); @@ -1143,10 +1124,10 @@ void gebp_kernel ResScalar* r2 = &res[(j2+2)*resStride + i]; ResScalar* r3 = &res[(j2+3)*resStride + i]; - internal::prefetch(r0); - internal::prefetch(r1); - internal::prefetch(r2); - internal::prefetch(r3); + internal::prefetch(r0+prefetch_res_offset); + internal::prefetch(r1+prefetch_res_offset); + internal::prefetch(r2+prefetch_res_offset); + internal::prefetch(r3+prefetch_res_offset); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; @@ -1155,19 +1136,18 @@ void gebp_kernel for(Index k=0; k blB += pk*4*RhsProgress; blA += pk*1*LhsProgress; - IACA_END } // process remaining peeled loop for(Index k=peeled_kc; k } //---------- Process remaining rows, 1 at once ---------- { - // loop on each row of the lhs (1*LhsProgress x depth) - for(Index i=peeled_mc1; i Index k=0; for(; k res[(j2+3)*resStride + i] += alpha*C3; } } - // remaining columns - for(Index j2=packet_cols4; j2 Date: Fri, 25 Apr 2014 11:15:13 +0200 Subject: [PATCH 0467/1103] Fix ptranspose overload prototypes for NEON --- Eigen/src/Core/arch/NEON/Complex.h | 2 +- Eigen/src/Core/arch/NEON/PacketMath.h | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 3c60735a9..259f2e7b8 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -263,7 +263,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, con return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); } -template<> EIGEN_DEVICE_FUNC inline void +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { float32x4_t tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v)); kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v)); diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 37fb840d4..e5eb06f36 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -451,7 +451,7 @@ PALIGN_NEON(3,Packet4i,vextq_s32) #undef PALIGN_NEON -template<> EIGEN_DEVICE_FUNC inline void +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]); float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]); @@ -462,7 +462,7 @@ ptranspose(PacketBlock& kernel) { kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1])); } -template<> EIGEN_DEVICE_FUNC inline void +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]); int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]); From ae4d9434e23e62d9403e570c20aeb3b8b44a2dd3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 11:21:18 +0200 Subject: [PATCH 0468/1103] Add unit test for pbroadcast4/2 --- test/packetmath.cpp | 24 ++++++++++++++++++++++++ 1 file changed, 24 insertions(+) diff --git a/test/packetmath.cpp b/test/packetmath.cpp index b4815629e..eb2cf7ebe 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -171,6 +171,30 @@ template void packetmath() VERIFY(areApprox(ref, data2, PacketSize) && "internal::pset1"); } + for(int offset=0;offset<3;++offset) + { + for (int i=0; i(&data1[offset], A0, A1, A2, A3); + internal::pstore(data2+0*PacketSize, A0); + internal::pstore(data2+1*PacketSize, A1); + internal::pstore(data2+2*PacketSize, A2); + internal::pstore(data2+3*PacketSize, A3); + VERIFY(areApprox(ref, data2, 4*PacketSize) && "internal::pbroadcast4"); + } + + for(int offset=0;offset<3;++offset) + { + for (int i=0; i(&data1[offset], A0, A1); + internal::pstore(data2+0*PacketSize, A0); + internal::pstore(data2+1*PacketSize, A1); + VERIFY(areApprox(ref, data2, 2*PacketSize) && "internal::pbroadcast2"); + } + VERIFY(internal::isApprox(data1[0], internal::pfirst(internal::pload(data1))) && "internal::pfirst"); if(PacketSize>1) From c9788d55b910c4bd8fb6cdc7978695f6b5d97e4c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 11:48:22 +0200 Subject: [PATCH 0469/1103] Disable 3pX4 kernel on Altivec: despite this platform has 32 registers, this version seems significantly slower. --- .../Core/products/GeneralBlockPanelKernel.h | 26 +++++++++---------- 1 file changed, 13 insertions(+), 13 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index e76c12c39..41c46c67a 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -188,7 +188,7 @@ public: nr = 4, // register block size along the M direction (currently, this one cannot be modified) -#ifdef EIGEN_HAS_FUSED_MADD +#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -296,7 +296,7 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, nr = 4, -#ifdef EIGEN_HAS_FUSED_MADD +#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -759,29 +759,29 @@ void gebp_kernel for(Index k=0; k // process remaining peeled loop for(Index k=peeled_kc; k Date: Fri, 25 Apr 2014 02:46:22 -0700 Subject: [PATCH 0470/1103] pbroadcast4/2 assume aligned memory --- test/packetmath.cpp | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-) diff --git a/test/packetmath.cpp b/test/packetmath.cpp index eb2cf7ebe..9dab07522 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -171,12 +171,11 @@ template void packetmath() VERIFY(areApprox(ref, data2, PacketSize) && "internal::pset1"); } - for(int offset=0;offset<3;++offset) { for (int i=0; i(&data1[offset], A0, A1, A2, A3); + internal::pbroadcast4(data1, A0, A1, A2, A3); internal::pstore(data2+0*PacketSize, A0); internal::pstore(data2+1*PacketSize, A1); internal::pstore(data2+2*PacketSize, A2); @@ -184,12 +183,11 @@ template void packetmath() VERIFY(areApprox(ref, data2, 4*PacketSize) && "internal::pbroadcast4"); } - for(int offset=0;offset<3;++offset) { for (int i=0; i(&data1[offset], A0, A1); + internal::pbroadcast2(data1, A0, A1); internal::pstore(data2+0*PacketSize, A0); internal::pstore(data2+1*PacketSize, A1); VERIFY(areApprox(ref, data2, 2*PacketSize) && "internal::pbroadcast2"); From 2dbfd83424cd0d30dac3b42b27b970b44a4e4541 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 02:46:57 -0700 Subject: [PATCH 0471/1103] Implement pbroadcast4 on altivec --- Eigen/src/Core/arch/AltiVec/Complex.h | 2 +- Eigen/src/Core/arch/AltiVec/PacketMath.h | 26 ++++++++++++++++++++++-- 2 files changed, 25 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 8fdffad5e..5409ddedd 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -229,7 +229,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip(const Packet2cf& x return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); } -template<> EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) +EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) { Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 8a67354e4..0e9adf450 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -168,6 +168,28 @@ template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { return vc; } + +template<> EIGEN_STRONG_INLINE void +pbroadcast4(const float *a, + Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) +{ + a3 = vec_ld(0,a); + a0 = vec_splat(a3, 0); + a1 = vec_splat(a3, 1); + a2 = vec_splat(a3, 2); + a3 = vec_splat(a3, 3); +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4(const int *a, + Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3) +{ + a3 = vec_ld(0,a); + a0 = vec_splat(a3, 0); + a1 = vec_splat(a3, 1); + a2 = vec_splat(a3, 2); + a3 = vec_splat(a3, 3); +} + template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) { float EIGEN_ALIGN16 af[4]; @@ -538,7 +560,7 @@ struct palign_impl } }; -template<> EIGEN_DEVICE_FUNC inline void +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { Packet4f t0, t1, t2, t3; t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); @@ -551,7 +573,7 @@ ptranspose(PacketBlock& kernel) { kernel.packet[3] = vec_mergel(t1, t3); } -template<> EIGEN_DEVICE_FUNC inline void +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { Packet4i t0, t1, t2, t3; t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]); From c20e3641de5b6d56f5496fef2619a1f53f8a1835 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 13:22:34 +0200 Subject: [PATCH 0472/1103] Fix for mixed products --- .../Core/products/GeneralBlockPanelKernel.h | 53 +++++++++++-------- 1 file changed, 30 insertions(+), 23 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 41c46c67a..ebf438d57 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -480,8 +480,14 @@ public: loadRhs(b,dest); } - // linking error if instantiated without being optimized out: - void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); + EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + { + // FIXME not sure that's the best way to implement it! + loadRhs(b+0, b0); + loadRhs(b+1, b1); + loadRhs(b+2, b2); + loadRhs(b+3, b3); + } // Vectorized path EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacketType& b0, DoublePacketType& b1) @@ -602,9 +608,11 @@ public: dest = pset1(*b); } - // linking error if instantiated without being optimized out: -// void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3); -// + void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3) + { + pbroadcast4(b, b0, b1, b2, b3); + } + // EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1) // { // // FIXME not sure that's the best way to implement it! @@ -1137,19 +1145,16 @@ void gebp_kernel for(Index k=0; k // process remaining peeled loop for(Index k=peeled_kc; k=1*PacketSize ? peeled_mc1 : Pack2>1 ? (rows/Pack2)*Pack2 : 0; + Index i=0; + // Pack 3 packets if(Pack1>=3*PacketSize) { - for(Index i=0; i=2*PacketSize) { - for(Index i=peeled_mc3; i=1*PacketSize) { - for(Index i=peeled_mc2; i1) { - for(Index i=peeled_mc1; i Date: Fri, 25 Apr 2014 14:05:54 +0200 Subject: [PATCH 0473/1103] Fix sizeof unit test --- Eigen/src/Core/DenseStorage.h | 4 ++-- test/sizeof.cpp | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 94f796783..59f515495 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -83,7 +83,7 @@ struct plain_array template struct plain_array { - EIGEN_USER_ALIGN32 T array[Size]; + EIGEN_USER_ALIGN_DEFAULT T array[Size]; EIGEN_DEVICE_FUNC plain_array() @@ -102,7 +102,7 @@ struct plain_array template struct plain_array { - EIGEN_USER_ALIGN32 T array[1]; + EIGEN_USER_ALIGN_DEFAULT T array[1]; EIGEN_DEVICE_FUNC plain_array() {} EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {} }; diff --git a/test/sizeof.cpp b/test/sizeof.cpp index 7044d2062..7763e51bd 100644 --- a/test/sizeof.cpp +++ b/test/sizeof.cpp @@ -13,9 +13,9 @@ template void verifySizeOf(const MatrixType&) { typedef typename MatrixType::Scalar Scalar; if (MatrixType::RowsAtCompileTime!=Dynamic && MatrixType::ColsAtCompileTime!=Dynamic) - VERIFY(std::ptrdiff_t(sizeof(MatrixType))==std::ptrdiff_t(sizeof(Scalar))*std::ptrdiff_t(MatrixType::SizeAtCompileTime)); + VERIFY_IS_EQUAL(std::ptrdiff_t(sizeof(MatrixType)),std::ptrdiff_t(sizeof(Scalar))*std::ptrdiff_t(MatrixType::SizeAtCompileTime)); else - VERIFY(sizeof(MatrixType)==sizeof(Scalar*) + 2 * sizeof(typename MatrixType::Index)); + VERIFY_IS_EQUAL(sizeof(MatrixType),sizeof(Scalar*) + 2 * sizeof(typename MatrixType::Index)); } void test_sizeof() From f9d2f3903eea91b79a1f0a4ff96d5df544402dcc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 16:54:30 +0200 Subject: [PATCH 0474/1103] Product kernel: skip loop on columns if there is no remaining rows --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index ebf438d57..cc4a9c485 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -1255,6 +1255,7 @@ void gebp_kernel } } //---------- Process remaining rows, 1 at once ---------- + if(peeled_mc1 straits.loadLhsUnaligned(blB+0*SwappedTraits::LhsProgress, A0); straits.loadLhsUnaligned(blB+1*SwappedTraits::LhsProgress, A1); - straits.loadRhsQuad(blA+0*spk, B_0); straits.loadRhsQuad(blA+1*spk, B_1); straits.madd(A0,B_0,C0,B_0); straits.madd(A1,B_1,C1,B_1); - straits.loadLhsUnaligned(blB+2*SwappedTraits::LhsProgress, A0); straits.loadLhsUnaligned(blB+3*SwappedTraits::LhsProgress, A1); straits.loadRhsQuad(blA+2*spk, B_0); From 450d0c3de044c9f32fa2f37fee821f6e390df382 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 22:25:48 +0200 Subject: [PATCH 0475/1103] Make sure that calls to broadcast4 are 16 bytes aligned --- Eigen/src/Core/arch/SSE/PacketMath.h | 6 +++--- Eigen/src/Core/products/TriangularMatrixMatrix.h | 1 + 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index ad935d5f1..6912f3bc3 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -486,7 +486,7 @@ template<> EIGEN_STRONG_INLINE void pbroadcast4(const float *a, Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) { - a3 = ploadu(a); + a3 = pload(a); a0 = vec4f_swizzle1(a3, 0,0,0,0); a1 = vec4f_swizzle1(a3, 1,1,1,1); a2 = vec4f_swizzle1(a3, 2,2,2,2); @@ -502,10 +502,10 @@ pbroadcast4(const double *a, a2 = _mm_loaddup_pd(a+2); a3 = _mm_loaddup_pd(a+3); #else - a1 = ploadu(a); + a1 = pload(a); a0 = vec2d_swizzle1(a1, 0,0); a1 = vec2d_swizzle1(a1, 1,1); - a3 = ploadu(a+2); + a3 = pload(a+2); a2 = vec2d_swizzle1(a3, 0,0); a3 = vec2d_swizzle1(a3, 1,1); #endif diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index 62575aff4..8088aa691 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -300,6 +300,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix=cols) ? 0 : actual_kc; Scalar* geb = blockB+ts*ts; + geb = geb + internal::first_aligned(geb,EIGEN_ALIGN_BYTES/sizeof(Scalar)); pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, actual_kc, rs); From e7ef26fa44999b054cbf36fb909f9737a111c4fb Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 25 Apr 2014 23:36:22 +0200 Subject: [PATCH 0476/1103] TRMM: Make sure we have enough memory in rhs block to enforce alignment. --- Eigen/src/Core/products/TriangularMatrixMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index 8088aa691..db7b27f8e 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -263,7 +263,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix Date: Mon, 28 Apr 2014 14:10:22 +0100 Subject: [PATCH 0477/1103] Make gdb pretty printer Python3-compatible (bug #800). --- debug/gdb/printers.py | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/debug/gdb/printers.py b/debug/gdb/printers.py index 86996a4f9..0d67a5f99 100644 --- a/debug/gdb/printers.py +++ b/debug/gdb/printers.py @@ -49,7 +49,7 @@ class EigenMatrixPrinter: regex = re.compile('\<.*\>') m = regex.findall(tag)[0][1:-1] template_params = m.split(',') - template_params = map(lambda x:x.replace(" ", ""), template_params) + template_params = [x.replace(" ", "") for x in template_params] if template_params[1] == '-0x00000000000000001' or template_params[1] == '-0x000000001' or template_params[1] == '-1': self.rows = val['m_storage']['m_rows'] @@ -88,8 +88,11 @@ class EigenMatrixPrinter: def __iter__ (self): return self - + def next(self): + return self.__next__() # Python 2.x compatibility + + def __next__(self): row = self.currentRow col = self.currentCol @@ -151,8 +154,11 @@ class EigenQuaternionPrinter: def __iter__ (self): return self - + def next(self): + return self.__next__() # Python 2.x compatibility + + def __next__(self): element = self.currentElement if self.currentElement >= 4: #there are 4 elements in a quanternion From 2fb64578aa31d35e26c18f8c44e041c5daac9982 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 28 Apr 2014 16:16:29 +0200 Subject: [PATCH 0478/1103] Add a small benchmark to compare dense solvers for small to large problems. --- bench/dense_solvers.cpp | 76 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 76 insertions(+) create mode 100644 bench/dense_solvers.cpp diff --git a/bench/dense_solvers.cpp b/bench/dense_solvers.cpp new file mode 100644 index 000000000..f37a8bb5f --- /dev/null +++ b/bench/dense_solvers.cpp @@ -0,0 +1,76 @@ +#include +#include "BenchTimer.h" +#include +#include +#include +using namespace Eigen; + +std::map > results; + +template +void bench(int id, int size = Size) +{ + typedef Matrix Mat; + Mat A(size,size); + A.setRandom(); + A = A*A.adjoint(); + BenchTimer t_llt, t_ldlt, t_lu, t_fplu, t_qr, t_cpqr, t_fpqr, t_jsvd; + + int tries = 3; + int rep = 1000/size; + if(rep==0) rep = 1; + rep = rep*rep; + + LLT llt(A); + LDLT ldlt(A); + PartialPivLU lu(A); + FullPivLU fplu(A); + HouseholderQR qr(A); + ColPivHouseholderQR cpqr(A); + FullPivHouseholderQR fpqr(A); + JacobiSVD jsvd(A.rows(),A.cols()); + + BENCH(t_llt, tries, rep, llt.compute(A)); + BENCH(t_ldlt, tries, rep, ldlt.compute(A)); + BENCH(t_lu, tries, rep, lu.compute(A)); + BENCH(t_fplu, tries, rep, fplu.compute(A)); + BENCH(t_qr, tries, rep, qr.compute(A)); + BENCH(t_cpqr, tries, rep, cpqr.compute(A)); + BENCH(t_fpqr, tries, rep, fpqr.compute(A)); + if(size<500) // JacobiSVD is really too slow for too large matrices + BENCH(t_jsvd, tries, rep, jsvd.compute(A,ComputeFullU|ComputeFullV)); + + results["LLT"][id] = t_llt.best(); + results["LDLT"][id] = t_ldlt.best(); + results["PartialPivLU"][id] = t_lu.best(); + results["FullPivLU"][id] = t_fplu.best(); + results["HouseholderQR"][id] = t_qr.best(); + results["ColPivHouseholderQR"][id] = t_cpqr.best(); + results["FullPivHouseholderQR"][id] = t_fpqr.best(); + results["JacobiSVD"][id] = size<500 ? t_jsvd.best() : 0; +} + +int main() +{ + const int small = 8; + const int medium = 100; + const int large = 1000; + const int xl = 4000; + + bench(0); + bench(1,medium); + bench(2,large); + bench(3,xl); + + IOFormat fmt(3, 0, " \t", "\n", "", ""); + + std::cout << "solver/size " << small << "\t" << medium << "\t" << large << "\t" << xl << "\n"; + std::cout << "LLT (ms) " << (results["LLT"]/1000.).format(fmt) << "\n"; + std::cout << "LDLT (%) " << (results["LDLT"]/results["LLT"]).format(fmt) << "\n"; + std::cout << "PartialPivLU (%) " << (results["PartialPivLU"]/results["LLT"]).format(fmt) << "\n"; + std::cout << "FullPivLU (%) " << (results["FullPivLU"]/results["LLT"]).format(fmt) << "\n"; + std::cout << "HouseholderQR (%) " << (results["HouseholderQR"]/results["LLT"]).format(fmt) << "\n"; + std::cout << "ColPivHouseholderQR (%) " << (results["ColPivHouseholderQR"]/results["LLT"]).format(fmt) << "\n"; + std::cout << "FullPivHouseholderQR (%) " << (results["FullPivHouseholderQR"]/results["LLT"]).format(fmt) << "\n"; + std::cout << "JacobiSVD (%) " << (results["JacobiSVD"]/results["LLT"]).format(fmt) << "\n"; +} From c0f2cb016e60b7dbde1d5946f42234a709a711f9 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 28 Apr 2014 10:32:27 -0700 Subject: [PATCH 0479/1103] Extended support for Tensors: * Added ability to map a region of the memory to a tensor * Added basic support for unary and binary coefficient wise expressions, such as addition or square root * Provided an emulation layer to make it possible to compile the code with compilers (such as nvcc) that don't support cxx11. --- Eigen/src/Core/util/Macros.h | 5 + unsupported/Eigen/CXX11/Core | 14 +- unsupported/Eigen/CXX11/Tensor | 27 ++- .../Eigen/CXX11/src/Core/util/CXX11Meta.h | 24 +-- .../CXX11/src/Core/util/CXX11Workarounds.h | 16 +- .../CXX11/src/Core/util/EmulateCXX11Meta.h | 184 ++++++++++++++++++ unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 156 ++++++++------- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 52 +++++ .../Eigen/CXX11/src/Tensor/TensorBase.h | 82 ++++++++ .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 127 ++++++++++++ .../Eigen/CXX11/src/Tensor/TensorExpr.h | 161 +++++++++++++++ .../src/Tensor/TensorForwardDeclarations.h | 27 +++ .../Eigen/CXX11/src/Tensor/TensorMap.h | 101 ++++++++++ .../Eigen/CXX11/src/Tensor/TensorStorage.h | 52 ++--- .../Eigen/CXX11/src/Tensor/TensorTraits.h | 122 ++++++++++++ unsupported/test/CMakeLists.txt | 5 +- unsupported/test/cxx11_tensor_simple.cpp | 2 +- 17 files changed, 1028 insertions(+), 129 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorBase.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorMap.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index bfd6ba7de..3a928001e 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -121,6 +121,11 @@ #define EIGEN_HAVE_RVALUE_REFERENCES #endif +// Does the compiler support variadic templates? +#if __cplusplus > 199711L +#define EIGEN_HAS_VARIADIC_TEMPLATES 1 +#endif + /** Allows to disable some optimizations which might affect the accuracy of the result. * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them. * They currently include: diff --git a/unsupported/Eigen/CXX11/Core b/unsupported/Eigen/CXX11/Core index 4dc4ab224..bba3d578d 100644 --- a/unsupported/Eigen/CXX11/Core +++ b/unsupported/Eigen/CXX11/Core @@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2013 Christian Seiler +// Copyright (C) 2014 Benoit Steiner // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -21,20 +22,23 @@ * module. Note that at this stage, you should not need to include * this module directly. * + * It also provides a limited fallback for compilers that don't support + * CXX11 yet, such as nvcc. + * * \code * #include * \endcode */ -#include - +// Emulate the cxx11 functionality that we need if the compiler doesn't support it. +#if __cplusplus <= 199711L +#include "src/Core/util/EmulateCXX11Meta.h" +#else #include "src/Core/util/CXX11Workarounds.h" #include "src/Core/util/CXX11Meta.h" +#endif #include #endif // EIGEN_CXX11_CORE_MODULE -/* - * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; - */ diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index f2c5129b3..f554c204a 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -10,9 +10,10 @@ #ifndef EIGEN_CXX11_TENSOR_MODULE #define EIGEN_CXX11_TENSOR_MODULE -#include +#include "Eigen/src/Core/util/StaticAssert.h" +#include "unsupported/Eigen/CXX11/Core" -#include +#include "Eigen/src/Core/util/DisableStupidWarnings.h" /** \defgroup CXX11_Tensor_Module Tensor Module * @@ -27,13 +28,21 @@ #include #include -#include "src/Tensor/TensorStorage.h" -#include "src/Tensor/Tensor.h" +#include "Eigen/Core" -#include +#include "unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h" + +#include "unsupported/Eigen/CXX11/src/Tensor/TensorBase.h" + +#include "unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h" + +#include "unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h" +#include "unsupported/Eigen/CXX11/src/Tensor/Tensor.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorMap.h" + +#include "Eigen/src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_CXX11_TENSOR_MODULE - -/* - * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; - */ diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h index 618e2eb7b..47f06b1b5 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -317,7 +317,7 @@ constexpr inline decltype(reduce::run((*((Ts*)0))...)) arg_sum(Ts template constexpr inline Array h_array_reverse(Array arr, numeric_list) { - return {{std_array_get(arr)...}}; + return {{array_get(arr)...}}; } template @@ -335,9 +335,9 @@ constexpr inline std::array array_reverse(std::array arr) // an infinite loop) template struct h_array_reduce { - constexpr static inline auto run(std::array arr) -> decltype(Reducer::run(h_array_reduce::run(arr), std_array_get(arr))) + constexpr static inline auto run(std::array arr) -> decltype(Reducer::run(h_array_reduce::run(arr), array_get(arr))) { - return Reducer::run(h_array_reduce::run(arr), std_array_get(arr)); + return Reducer::run(h_array_reduce::run(arr), array_get(arr)); } }; @@ -346,7 +346,7 @@ struct h_array_reduce { constexpr static inline T run(std::array arr) { - return std_array_get<0>(arr); + return array_get<0>(arr); } }; @@ -375,7 +375,7 @@ constexpr inline auto array_prod(std::array arr) -> decltype(array_reduce< template constexpr inline std::array h_array_zip(std::array a, std::array b, numeric_list) { - return std::array{{ Op::run(std_array_get(a), std_array_get(b))... }}; + return std::array{{ Op::run(array_get(a), array_get(b))... }}; } template @@ -387,9 +387,9 @@ constexpr inline std::array array_zip(std::array< /* zip an array and reduce the result */ template -constexpr inline auto h_array_zip_and_reduce(std::array a, std::array b, numeric_list) -> decltype(reduce::type...>::run(Op::run(std_array_get(a), std_array_get(b))...)) +constexpr inline auto h_array_zip_and_reduce(std::array a, std::array b, numeric_list) -> decltype(reduce::type...>::run(Op::run(array_get(a), array_get(b))...)) { - return reduce::type...>::run(Op::run(std_array_get(a), std_array_get(b))...); + return reduce::type...>::run(Op::run(array_get(a), array_get(b))...); } template @@ -403,7 +403,7 @@ constexpr inline auto array_zip_and_reduce(std::array a, std::array template constexpr inline std::array h_array_apply(std::array a, numeric_list) { - return std::array{{ Op::run(std_array_get(a))... }}; + return std::array{{ Op::run(array_get(a))... }}; } template @@ -415,9 +415,9 @@ constexpr inline std::array array_apply(std::array -constexpr inline auto h_array_apply_and_reduce(std::array arr, numeric_list) -> decltype(reduce::type...>::run(Op::run(std_array_get(arr))...)) +constexpr inline auto h_array_apply_and_reduce(std::array arr, numeric_list) -> decltype(reduce::type...>::run(Op::run(array_get(arr))...)) { - return reduce::type...>::run(Op::run(std_array_get(arr))...); + return reduce::type...>::run(Op::run(array_get(arr))...); } template @@ -497,7 +497,3 @@ InstType instantiate_by_c_array(ArrType* arr) } // end namespace Eigen #endif // EIGEN_CXX11META_H - -/* - * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; - */ diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index 356ae10cf..77207f453 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -40,8 +40,18 @@ #error This library needs at least a C++11 compliant compiler. If you use g++/clang, please enable the -std=c++11 compiler flag. (-std=c++0x on older versions.) #endif +using std::array; + namespace Eigen { +// Use std::array as Eigen array +/*template +struct array : public std::array { + array() = default; + array(const std::initializer_list& a);// : std::array(a) {}; + array(const std::array& a); +};*/ + namespace internal { /* std::get is only constexpr in C++14, not yet in C++11 @@ -60,9 +70,9 @@ namespace internal { #define STD_GET_ARR_HACK std::template get(a) #endif -template constexpr inline T& std_array_get(std::array& a) { return (T&) STD_GET_ARR_HACK; } -template constexpr inline T&& std_array_get(std::array&& a) { return (T&&) STD_GET_ARR_HACK; } -template constexpr inline T const& std_array_get(std::array const& a) { return (T const&) STD_GET_ARR_HACK; } +template constexpr inline T& array_get(std::array& a) { return (T&) STD_GET_ARR_HACK; } +template constexpr inline T&& array_get(std::array&& a) { return (T&&) STD_GET_ARR_HACK; } +template constexpr inline T const& array_get(std::array const& a) { return (T const&) STD_GET_ARR_HACK; } #undef STD_GET_ARR_HACK diff --git a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h new file mode 100644 index 000000000..76fcba5b4 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h @@ -0,0 +1,184 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_EMULATE_CXX11_META_H +#define EIGEN_EMULATE_CXX11_META_H + + +namespace Eigen { + +// The array class is only available starting with cxx11. Emulate our own here +// if needed +template class array { + public: + T& operator[] (size_t index) { return values[index]; } + const T& operator[] (size_t index) const { return values[index]; } + + T values[n]; +}; + + +namespace internal { + +/** \internal + * \file CXX11/Core/util/EmulateCXX11Meta.h + * This file emulates a subset of the functionality provided by CXXMeta.h for + * compilers that don't yet support cxx11 such as nvcc. + */ + +struct empty_list { static const std::size_t count = 0; }; + +template struct type_list { + T head; + Tail tail; + static const std::size_t count = 1 + Tail::count; +}; + +struct null_type { }; + +template +struct make_type_list { + typedef typename make_type_list::type tailresult; + + typedef type_list type; +}; + +template<> struct make_type_list<> { + typedef empty_list type; +}; + + + +template +struct type2val { + static const T value = n; +}; + + +template struct gen_numeric_list_repeated; + +template struct gen_numeric_list_repeated { + typedef typename make_type_list >::type type; +}; + +template struct gen_numeric_list_repeated { + typedef typename make_type_list, type2val >::type type; +}; + +template struct gen_numeric_list_repeated { + typedef typename make_type_list, type2val, type2val >::type type; +}; + +template struct gen_numeric_list_repeated { + typedef typename make_type_list, type2val, type2val, type2val >::type type; +}; + +template struct gen_numeric_list_repeated { + typedef typename make_type_list, type2val, type2val, type2val, type2val >::type type; +}; + + + +template +array repeat(t v) { + array array; + array.fill(v); + return array; +} + +template +t array_prod(const array& a) { + t prod = 1; + for (size_t i = 0; i < n; ++i) { prod *= a[i]; } + return prod; +} +template +t array_prod(const array& /*a*/) { + return 0; +} + +template inline T& array_get(array& a) { + return a[I]; +} +template inline const T& array_get(const array& a) { + return a[I]; +} + +struct sum_op { + template static inline bool run(A a, B b) { return a + b; } +}; +struct product_op { + template static inline bool run(A a, B b) { return a * b; } +}; + +struct logical_and_op { + template static inline bool run(A a, B b) { return a && b; } +}; +struct logical_or_op { + template static inline bool run(A a, B b) { return a || b; } +}; + +struct equal_op { + template static inline bool run(A a, B b) { return a == b; } +}; +struct not_equal_op { + template static inline bool run(A a, B b) { return a != b; } +}; +struct lesser_op { + template static inline bool run(A a, B b) { return a < b; } +}; +struct lesser_equal_op { + template static inline bool run(A a, B b) { return a <= b; } +}; + +struct greater_op { + template static inline bool run(A a, B b) { return a > b; } +}; +struct greater_equal_op { + template static inline bool run(A a, B b) { return a >= b; } +}; + +struct not_op { + template static inline bool run(A a) { return !a; } +}; +struct negation_op { + template static inline bool run(A a) { return -a; } +}; +struct greater_equal_zero_op { + template static inline bool run(A a) { return a >= 0; } +}; + + +template +inline bool array_apply_and_reduce(const array& a) { + EIGEN_STATIC_ASSERT(N >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE) + bool result = Reducer::run(Op::run(a[0]), Op::run(a[1])); + for (size_t i = 2; i < N; ++i) { + result = Reducer::run(result, Op::run(a[i])); + } + return result; +} + +template +inline bool array_zip_and_reduce(const array& a, const array& b) { + EIGEN_STATIC_ASSERT(N >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE) + bool result = Reducer::run(Op::run(a[0], b[0]), Op::run(a[1], b[1])); + for (size_t i = 2; i < N; ++i) { + result = Reducer::run(result, Op::run(a[i], b[i])); + } + return result; +} + +} // end namespace internal + +} // end namespace Eigen + + + +#endif // EIGEN_EMULATE_CXX11_META_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index c6216e14c..7b8f14c6d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -57,28 +57,16 @@ namespace Eigen { * * \ref TopicStorageOrders */ -template -class Tensor; namespace internal { -template -struct traits> -{ - typedef Scalar_ Scalar; - typedef Dense StorageKind; - typedef DenseIndex Index; - enum { - Options = Options_ - }; -}; template struct tensor_index_linearization_helper { - constexpr static inline Index run(std::array const& indices, std::array const& dimensions) + static inline Index run(array const& indices, array const& dimensions) { - return std_array_get(indices) + - std_array_get(dimensions) * + return array_get(indices) + + array_get(dimensions) * tensor_index_linearization_helper::run(indices, dimensions); } }; @@ -86,39 +74,40 @@ struct tensor_index_linearization_helper template struct tensor_index_linearization_helper { - constexpr static inline Index run(std::array const& indices, std::array const&) + static inline Index run(array const& indices, array const&) { - return std_array_get(indices); + return array_get(indices); } }; /* Forward-declaration required for the symmetry support. */ template class tensor_symmetry_value_setter; + } // end namespace internal template -class Tensor +class Tensor : public TensorBase > { - static_assert(NumIndices_ >= 1, "A tensor must have at least one index."); - public: typedef Tensor Self; + typedef TensorBase > Base; + typedef typename Eigen::internal::nested::type Nested; typedef typename internal::traits::StorageKind StorageKind; typedef typename internal::traits::Index Index; - typedef typename internal::traits::Scalar Scalar; + typedef Scalar_ Scalar; typedef typename internal::packet_traits::type PacketScalar; typedef typename NumTraits::Real RealScalar; - typedef Self DenseType; + typedef typename Base::CoeffReturnType CoeffReturnType; - constexpr static int Options = Options_; - constexpr static std::size_t NumIndices = NumIndices_; + static const int Options = Options_; + static const std::size_t NumIndices = NumIndices_; protected: TensorStorage m_storage; public: EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return m_storage.dimensions()[n]; } - EIGEN_STRONG_INLINE std::array dimensions() const { return m_storage.dimensions(); } + EIGEN_STRONG_INLINE array dimensions() const { return m_storage.dimensions(); } EIGEN_STRONG_INLINE Index size() const { return internal::array_prod(m_storage.dimensions()); } EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } @@ -129,29 +118,17 @@ class Tensor inline Self& base() { return *this; } inline const Self& base() const { return *this; } - void setZero() - { - // FIXME: until we have implemented packet access and the - // expression engine w.r.t. nullary ops, use this - // as a kludge. Only works with POD types, but for - // any standard usage, this shouldn't be a problem - memset((void *)data(), 0, size() * sizeof(Scalar)); - } - - inline Self& operator=(Self const& other) - { - m_storage = other.m_storage; - return *this; - } - +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES template inline const Scalar& coeff(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const { - static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); - return coeff(std::array{{firstIndex, secondIndex, otherIndices...}}); + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return coeff(array{{firstIndex, secondIndex, otherIndices...}}); } +#endif - inline const Scalar& coeff(const std::array& indices) const + inline const Scalar& coeff(const array& indices) const { eigen_internal_assert(checkIndexRange(indices)); return m_storage.data()[linearizedIndex(indices)]; @@ -163,14 +140,17 @@ class Tensor return m_storage.data()[index]; } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES template inline Scalar& coeffRef(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) { - static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); - return coeffRef(std::array{{firstIndex, secondIndex, otherIndices...}}); + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return coeffRef(array{{firstIndex, secondIndex, otherIndices...}}); } +#endif - inline Scalar& coeffRef(const std::array& indices) + inline Scalar& coeffRef(const array& indices) { eigen_internal_assert(checkIndexRange(indices)); return m_storage.data()[linearizedIndex(indices)]; @@ -182,14 +162,17 @@ class Tensor return m_storage.data()[index]; } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES template inline const Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) const { - static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); - return this->operator()(std::array{{firstIndex, secondIndex, otherIndices...}}); + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return this->operator()(array{{firstIndex, secondIndex, otherIndices...}}); } +#endif - inline const Scalar& operator()(const std::array& indices) const + inline const Scalar& operator()(const array& indices) const { eigen_assert(checkIndexRange(indices)); return coeff(indices); @@ -203,18 +186,22 @@ class Tensor inline const Scalar& operator[](Index index) const { - static_assert(NumIndices == 1, "The bracket operator is only for vectors, use the parenthesis operator instead."); + // The bracket operator is only for vectors, use the parenthesis operator instead. + EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE); return coeff(index); } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES template inline Scalar& operator()(Index firstIndex, Index secondIndex, IndexTypes... otherIndices) { - static_assert(sizeof...(otherIndices) + 2 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); - return operator()(std::array{{firstIndex, secondIndex, otherIndices...}}); + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 2 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return operator()(array{{firstIndex, secondIndex, otherIndices...}}); } +#endif - inline Scalar& operator()(const std::array& indices) + inline Scalar& operator()(const array& indices) { eigen_assert(checkIndexRange(indices)); return coeffRef(indices); @@ -228,47 +215,70 @@ class Tensor inline Scalar& operator[](Index index) { - static_assert(NumIndices == 1, "The bracket operator is only for vectors, use the parenthesis operator instead."); + // The bracket operator is only for vectors, use the parenthesis operator instead + EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE) return coeffRef(index); } - inline Tensor() + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Tensor() : m_storage() { } - inline Tensor(const Self& other) + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Tensor(const Self& other) : m_storage(other.m_storage) { } - inline Tensor(Self&& other) - : m_storage(other.m_storage) - { - } +#ifdef EIGEN_HAVE_RVALUE_REFERENCES +// inline Tensor(Self&& other) +// : m_storage(other.m_storage) +// { +// } +#endif +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES template inline Tensor(Index firstDimension, IndexTypes... otherDimensions) : m_storage() { - static_assert(sizeof...(otherDimensions) + 1 == NumIndices, "Number of dimensions used to construct a tensor must be equal to the rank of the tensor."); - resize(std::array{{firstDimension, otherDimensions...}}); + // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + resize(array{{firstDimension, otherDimensions...}}); } +#endif - inline Tensor(std::array dimensions) + inline Tensor(const array& dimensions) : m_storage(internal::array_prod(dimensions), dimensions) { EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Tensor& operator=(const OtherDerived& other) + { + // FIXME: we need to resize the tensor to fix the dimensions of the other. + // Unfortunately this isn't possible yet when the rhs is an expression. + // resize(other.dimensions()); + internal::TensorAssign::run(*this, other); + return *this; + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES template void resize(Index firstDimension, IndexTypes... otherDimensions) { - static_assert(sizeof...(otherDimensions) + 1 == NumIndices, "Number of dimensions used to resize a tensor must be equal to the rank of the tensor."); - resize(std::array{{firstDimension, otherDimensions...}}); + // The number of dimensions used to resize a tensor must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + resize(array{{firstDimension, otherDimensions...}}); } +#endif - void resize(const std::array& dimensions) + void resize(const array& dimensions) { std::size_t i; Index size = Index(1); @@ -285,20 +295,22 @@ class Tensor #endif } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES template internal::tensor_symmetry_value_setter symCoeff(const Symmetry_& symmetry, Index firstIndex, IndexTypes... otherIndices) { - return symCoeff(symmetry, std::array{{firstIndex, otherIndices...}}); + return symCoeff(symmetry, array{{firstIndex, otherIndices...}}); } template - internal::tensor_symmetry_value_setter symCoeff(const Symmetry_& symmetry, std::array const& indices) + internal::tensor_symmetry_value_setter symCoeff(const Symmetry_& symmetry, array const& indices) { return internal::tensor_symmetry_value_setter(*this, symmetry, indices); } +#endif protected: - bool checkIndexRange(const std::array& indices) const + bool checkIndexRange(const array& indices) const { using internal::array_apply_and_reduce; using internal::array_zip_and_reduce; @@ -313,7 +325,7 @@ class Tensor array_zip_and_reduce(indices, m_storage.dimensions()); } - inline Index linearizedIndex(const std::array& indices) const + inline Index linearizedIndex(const array& indices) const { return internal::tensor_index_linearization_helper::run(indices, m_storage.dimensions()); } @@ -322,7 +334,3 @@ class Tensor } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_H - -/* - * kate: space-indent on; indent-width 2; mixedindent off; indent-mode cstyle; - */ diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h new file mode 100644 index 000000000..f1df827f9 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -0,0 +1,52 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H +#define EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H + + +namespace Eigen { + +/** \class TensorAssign + * \ingroup CXX11_Tensor_Module + * + * \brief The tensor assignment class. + * + * This class is responsible for triggering the evaluation of the expressions + * used on the lhs and rhs of an assignment operator and copy the result of + * the evaluation of the rhs expression at the address computed during the + * evaluation lhs expression. + * + * TODO: vectorization. For now the code only uses scalars + * TODO: parallelisation using multithreading on cpu, or kernels on gpu. + */ +namespace internal { + +template +struct TensorAssign +{ + typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC + static inline void run(Derived1& dst, const Derived2& src) + { + TensorEvaluator evalDst(dst); + TensorEvaluator evalSrc(src); + const Index size = dst.size(); + for(Index i = 0; i < size; ++i) { + evalDst.coeffRef(i) = evalSrc.coeff(i); + } + } +}; + + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h new file mode 100644 index 000000000..0b9f32f7f --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -0,0 +1,82 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_BASE_H +#define EIGEN_CXX11_TENSOR_TENSOR_BASE_H + +namespace Eigen { + +/** \class TensorBase + * \ingroup CXX11_Tensor_Module + * + * \brief The tensor base class. + * + * This class is the common parent of the Tensor and TensorMap class, thus + * making it possible to use either class interchangably in expressions. + */ + +template +class TensorBase +{ + public: + typedef typename internal::traits::Scalar Scalar; + typedef typename internal::traits::Index Index; + typedef Scalar CoeffReturnType; + + Derived& setZero() { + return setConstant(Scalar(0)); + } + + Derived& setConstant(const Scalar& val) { + Scalar* data = derived().data(); + for (int i = 0; i < derived().size(); ++i) { + data[i] = val; + } + return derived(); + } + + Derived& setRandom() { + Scalar* data = derived().data(); + for (int i = 0; i < derived().size(); ++i) { + data[i] = internal::random_default_impl::run(); + } + return derived(); + } + + // Coefficient-wise unary operators + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + operator-() const { return derived(); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + cwiseSqrt() const { return derived(); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + cwiseAbs() const { return derived(); } + + // Coefficient-wise binary operators. + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator+(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + protected: + template friend class TensorBase; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& derived() { return *static_cast(this); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Derived& derived() const { return *static_cast(this); } +}; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_BASE_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h new file mode 100644 index 000000000..f4f10eff5 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -0,0 +1,127 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_EVALUATOR_H +#define EIGEN_CXX11_TENSOR_TENSOR_EVALUATOR_H + +namespace Eigen { + +/** \class TensorEvaluator + * \ingroup CXX11_Tensor_Module + * + * \brief The tensor evaluator classes. + * + * These classes are responsible for the evaluation of the tensor expression. + * + * TODO: add support for more types of expressions, in particular expressions + * leading to lvalues (slicing, reshaping, etc...) + * TODO: add support for vectorization + */ + + +template +struct TensorEvaluator +{ + typedef typename Derived::Index Index; + typedef typename Derived::Scalar Scalar; + typedef typename Derived::Scalar& CoeffReturnType; + //typedef typename Derived::PacketScalar PacketScalar; + typedef TensorEvaluator nestedType; + + TensorEvaluator(Derived& m) + : m_data(const_cast(m.data())) + { } + + CoeffReturnType coeff(Index index) const { + return m_data[index]; + } + + Scalar& coeffRef(Index index) { + return m_data[index]; + } + + // to do: vectorized evaluation. + /* template + PacketReturnType packet(Index index) const + { + return ploadt(m_data + index); + } + + template + void writePacket(Index index, const PacketScalar& x) + { + return pstoret(const_cast(m_data) + index, x); + }*/ + + protected: + Scalar* m_data; +}; + + + + +// -------------------- CwiseUnaryOp -------------------- + +template +struct TensorEvaluator > +{ + typedef TensorCwiseUnaryOp XprType; + typedef TensorEvaluator nestedType; + + TensorEvaluator(const XprType& op) + : m_functor(op.functor()), + m_argImpl(op.nestedExpression()) + { } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + CoeffReturnType coeff(Index index) const + { + return m_functor(m_argImpl.coeff(index)); + } + + private: + const UnaryOp m_functor; + typename TensorEvaluator::nestedType m_argImpl; +}; + + +// -------------------- CwiseBinaryOp -------------------- + +template +struct TensorEvaluator > +{ + typedef TensorCwiseBinaryOp XprType; + typedef TensorEvaluator leftType; + typedef TensorEvaluator rightType; + + TensorEvaluator(const XprType& op) + : m_functor(op.functor()), + m_leftImpl(op.lhsExpression()), + m_rightImpl(op.rhsExpression()) + { } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + + CoeffReturnType coeff(Index index) const + { + return m_functor(m_leftImpl.coeff(index), m_rightImpl.coeff(index)); + } + + private: + const BinaryOp m_functor; + typename TensorEvaluator::nestedType m_leftImpl; + typename TensorEvaluator::nestedType m_rightImpl; +}; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_EVALUATOR_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h new file mode 100644 index 000000000..5a45cec31 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h @@ -0,0 +1,161 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_EXPR_H +#define EIGEN_CXX11_TENSOR_TENSOR_EXPR_H + +namespace Eigen { + +/** \class TensorExpr + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor expression classes. + * + * The TensorCwiseUnaryOp class represents an expression where a unary operator + * (e.g. cwiseSqrt) is applied to an expression. + * + * The TensorCwiseBinaryOp class represents an expression where a binary operator + * (e.g. addition) is applied to a lhs and a rhs expression. + * + */ + +namespace internal { +template +struct traits > + : traits +{ + typedef typename result_of< + UnaryOp(typename XprType::Scalar) + >::type Scalar; + typedef typename XprType::Nested XprTypeNested; + typedef typename remove_reference::type _XprTypeNested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorCwiseUnaryOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorCwiseUnaryOp type; +}; + +} // end namespace internal + + + +template +class TensorCwiseUnaryOp +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + inline TensorCwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) + : m_xpr(xpr), m_functor(func) {} + + EIGEN_DEVICE_FUNC + const UnaryOp& functor() const { return m_functor; } + + /** \returns the nested expression */ + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + nestedExpression() const { return m_xpr; } + + /** \returns the nested expression */ + EIGEN_DEVICE_FUNC + typename internal::remove_all::type& + nestedExpression() { return m_xpr.const_cast_derived(); } + + protected: + typename XprType::Nested m_xpr; + const UnaryOp m_functor; +}; + + +namespace internal { +template +struct traits > +{ + // Type promotion to handle the case where the types of the lhs and the rhs are different. + typedef typename result_of< + BinaryOp( + typename LhsXprType::Scalar, + typename RhsXprType::Scalar + ) + >::type Scalar; + typedef typename promote_storage_type::StorageKind, + typename traits::StorageKind>::ret StorageKind; + typedef typename promote_index_type::Index, + typename traits::Index>::type Index; + typedef typename LhsXprType::Nested LhsNested; + typedef typename RhsXprType::Nested RhsNested; + typedef typename remove_reference::type _LhsNested; + typedef typename remove_reference::type _RhsNested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorCwiseBinaryOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorCwiseBinaryOp type; +}; + +} // end namespace internal + + + +template +class TensorCwiseBinaryOp +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename internal::promote_storage_type::ret CoeffReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + inline TensorCwiseBinaryOp(const LhsXprType& lhs, const RhsXprType& rhs, const BinaryOp& func = BinaryOp()) + : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_functor(func) {} + + EIGEN_DEVICE_FUNC + const BinaryOp& functor() const { return m_functor; } + + /** \returns the nested expressions */ + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + lhsExpression() const { return m_lhs_xpr; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + rhsExpression() const { return m_rhs_xpr; } + + protected: + typename LhsXprType::Nested m_lhs_xpr; + typename RhsXprType::Nested m_rhs_xpr; + const BinaryOp m_functor; +}; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_EXPR_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h new file mode 100644 index 000000000..dc97764f0 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -0,0 +1,27 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_FORWARD_DECLARATIONS_H +#define EIGEN_CXX11_TENSOR_TENSOR_FORWARD_DECLARATIONS_H + +namespace Eigen { + +template class Tensor; +template class TensorMap; +template class TensorBase; + +template class TensorCwiseUnaryOp; +template class TensorCwiseBinaryOp; + +// Move to internal? +template struct TensorEvaluator; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_FORWARD_DECLARATIONS_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h new file mode 100644 index 000000000..7dec1e08d --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -0,0 +1,101 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_MAP_H +#define EIGEN_CXX11_TENSOR_TENSOR_MAP_H + +namespace Eigen { + +template class Stride; + + +/** \class TensorMap + * \ingroup CXX11_Tensor_Module + * + * \brief A tensor expression mapping an existing array of data. + * + */ + +template class TensorMap : public TensorBase > +{ + public: + typedef TensorMap Self; + typedef typename PlainObjectType::Base Base; + typedef typename Eigen::internal::nested::type Nested; + typedef typename internal::traits::StorageKind StorageKind; + typedef typename internal::traits::Index Index; + typedef typename internal::traits::Scalar Scalar; + typedef typename internal::packet_traits::type PacketScalar; + typedef typename NumTraits::Real RealScalar; + typedef typename Base::CoeffReturnType CoeffReturnType; + + /* typedef typename internal::conditional< + bool(internal::is_lvalue::value), + Scalar *, + const Scalar *>::type + PointerType;*/ + typedef Scalar* PointerType; + typedef PointerType PointerArgType; + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions({{firstDimension}}) { + // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(1 == PlainObjectType::NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions({{firstDimension, otherDimensions...}}) { + // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == PlainObjectType::NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + } +#endif + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index dimension(Index n) const { return m_dimensions[n]; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index size() const { return internal::array_prod(m_dimensions); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar* data() { return m_data; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar* data() const { return m_data; } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(Index index) const + { + eigen_internal_assert(index >= 0 && index < size()); + return m_data[index]; + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) + { + static_assert(sizeof...(otherIndices) + 1 == PlainObjectType::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + const Index index = internal::tensor_index_linearization_helper::run(array{{firstIndex, otherIndices...}}, m_dimensions); + return m_data[index]; + } +#endif + + template + EIGEN_DEVICE_FUNC + Self& operator=(const OtherDerived& other) + { + internal::TensorAssign::run(*this, other); + return *this; + } + + private: + typename PlainObjectType::Scalar* m_data; + array m_dimensions; +}; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_MAP_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index a34600ee6..503d7cfd6 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -37,14 +37,19 @@ template class TensorStorage : public TensorStorage::type> { - typedef TensorStorage::type> Base_; + typedef TensorStorage::type> Base_; + public: - TensorStorage() = default; - TensorStorage(const TensorStorage&) = default; - TensorStorage(TensorStorage&&) = default; + TensorStorage() { } + TensorStorage(const TensorStorage& other) : Base_(other) { } + +#ifdef EIGEN_HAVE_RVALUE_REFERENCES +// TensorStorage(TensorStorage&&) = default; +#endif TensorStorage(internal::constructor_without_unaligned_array_assert) : Base_(internal::constructor_without_unaligned_array_assert()) {} - TensorStorage(DenseIndex size, const std::array& dimensions) : Base_(size, dimensions) {} - TensorStorage& operator=(const TensorStorage&) = default; + TensorStorage(DenseIndex size, const array& dimensions) : Base_(size, dimensions) {} + + // TensorStorage& operator=(const TensorStorage&) = default; }; // pure dynamic @@ -52,17 +57,17 @@ template class TensorStorage::type> { T *m_data; - std::array m_dimensions; + array m_dimensions; typedef TensorStorage::type> Self_; public: - TensorStorage() : m_data(0), m_dimensions(internal::template repeat(0)) {} + TensorStorage() : m_data(0), m_dimensions() {} TensorStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_dimensions(internal::template repeat(0)) {} - TensorStorage(DenseIndex size, const std::array& dimensions) - : m_data(internal::conditional_aligned_new_auto(size)), m_dimensions(dimensions) - { EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN } - TensorStorage(const Self_& other) + TensorStorage(DenseIndex size, const array& dimensions) + : m_data(internal::conditional_aligned_new_auto(size)), m_dimensions(dimensions) + { EIGEN_INTERNAL_TENSOR_STORAGE_CTOR_PLUGIN } + TensorStorage(const Self_& other) : m_data(internal::conditional_aligned_new_auto(internal::array_prod(other.m_dimensions))) , m_dimensions(other.m_dimensions) { @@ -76,28 +81,34 @@ class TensorStorage(m_data, internal::array_prod(m_dimensions)); } void swap(Self_& other) { std::swap(m_data,other.m_data); std::swap(m_dimensions,other.m_dimensions); } - std::array dimensions(void) const {return m_dimensions;} - void conservativeResize(DenseIndex size, const std::array& nbDimensions) + const array& dimensions() const {return m_dimensions;} + + void conservativeResize(DenseIndex size, const array& nbDimensions) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, internal::array_prod(m_dimensions)); m_dimensions = nbDimensions; } - void resize(DenseIndex size, const std::array& nbDimensions) + void resize(DenseIndex size, const array& nbDimensions) { if(size != internal::array_prod(m_dimensions)) { @@ -110,8 +121,9 @@ class TensorStorage +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_TRAITS_H +#define EIGEN_CXX11_TENSOR_TENSOR_TRAITS_H + +namespace Eigen { +namespace internal { + + +template +class compute_tensor_flags +{ + enum { + is_dynamic_size_storage = 1, + + aligned_bit = + ( + ((Options&DontAlign)==0) && ( +#if EIGEN_ALIGN_STATICALLY + (!is_dynamic_size_storage) +#else + 0 +#endif + || +#if EIGEN_ALIGN + is_dynamic_size_storage +#else + 0 +#endif + ) + ) ? AlignedBit : 0, + packet_access_bit = packet_traits::Vectorizable && aligned_bit ? PacketAccessBit : 0 + }; + + public: + enum { ret = packet_access_bit | aligned_bit}; +}; + + +template +struct traits > +{ + typedef Scalar_ Scalar; + typedef Dense StorageKind; + typedef DenseIndex Index; + enum { + Options = Options_, + Flags = compute_tensor_flags::ret, + }; +}; + + +template +struct traits > + : public traits +{ + typedef traits BaseTraits; + typedef typename BaseTraits::Scalar Scalar; + typedef typename BaseTraits::StorageKind StorageKind; + typedef typename BaseTraits::Index Index; +}; + + +template +struct eval, Eigen::Dense> +{ + typedef const Tensor<_Scalar, NumIndices_, Options_>& type; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const Tensor<_Scalar, NumIndices_, Options_>& type; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorMap& type; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorMap& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef const Tensor& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef const Tensor& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef const TensorMap& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef const TensorMap& type; +}; + +} // end namespace internal +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_TRAITS_H diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 0a6c56c19..31583d3ca 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -93,7 +93,7 @@ ei_add_test(minres) ei_add_test(levenberg_marquardt) ei_add_test(bdcsvd) -option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." OFF) +option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." ON) if(EIGEN_TEST_CXX11) # FIXME: add C++11 compiler switch in some portable way # (MSVC doesn't need any for example, so this will @@ -101,4 +101,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_meta "-std=c++0x") ei_add_test(cxx11_tensor_simple "-std=c++0x") ei_add_test(cxx11_tensor_symmetry "-std=c++0x") + ei_add_test(cxx11_tensor_assign "-std=c++0x") + ei_add_test(cxx11_tensor_expr "-std=c++0x") + ei_add_test(cxx11_tensor_map "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_simple.cpp b/unsupported/test/cxx11_tensor_simple.cpp index ea512c9cc..1f76033ea 100644 --- a/unsupported/test/cxx11_tensor_simple.cpp +++ b/unsupported/test/cxx11_tensor_simple.cpp @@ -163,7 +163,7 @@ static void test_3d() VERIFY_IS_EQUAL((epsilon(0,2,1)), -1); VERIFY_IS_EQUAL((epsilon(1,0,2)), -1); - std::array dims{{2,3,4}}; + array dims{{2,3,4}}; Tensor t1(dims); Tensor t2(dims); From 07986189b76001c0c29b1583592b00340b3343ba Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 1 May 2014 23:03:54 +0200 Subject: [PATCH 0480/1103] Fix bug #803: avoid char* to int* conversion --- Eigen/src/Core/util/Memory.h | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index e1a12aef1..390b60c74 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -767,9 +767,9 @@ namespace internal { #ifdef EIGEN_CPUID -inline bool cpuid_is_vendor(int abcd[4], const char* vendor) +inline bool cpuid_is_vendor(int abcd[4], const int vendor[3]) { - return abcd[1]==(reinterpret_cast(vendor))[0] && abcd[3]==(reinterpret_cast(vendor))[1] && abcd[2]==(reinterpret_cast(vendor))[2]; + return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2]; } inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3) @@ -911,13 +911,16 @@ inline void queryCacheSizes(int& l1, int& l2, int& l3) { #ifdef EIGEN_CPUID int abcd[4]; + const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e}; + const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163}; + const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!" // identify the CPU vendor EIGEN_CPUID(abcd,0x0,0); int max_std_funcs = abcd[1]; - if(cpuid_is_vendor(abcd,"GenuineIntel")) + if(cpuid_is_vendor(abcd,GenuineIntel)) queryCacheSizes_intel(l1,l2,l3,max_std_funcs); - else if(cpuid_is_vendor(abcd,"AuthenticAMD") || cpuid_is_vendor(abcd,"AMDisbetter!")) + else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_)) queryCacheSizes_amd(l1,l2,l3); else // by default let's use Intel's API From d67aa1549b5675bb6089e9df0c299dab7d0bb80d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 3 May 2014 10:46:11 +0200 Subject: [PATCH 0481/1103] Add missing add_subdirectory directive --- Eigen/src/Core/arch/CMakeLists.txt | 1 + 1 file changed, 1 insertion(+) diff --git a/Eigen/src/Core/arch/CMakeLists.txt b/Eigen/src/Core/arch/CMakeLists.txt index 8456dec15..0db8c558d 100644 --- a/Eigen/src/Core/arch/CMakeLists.txt +++ b/Eigen/src/Core/arch/CMakeLists.txt @@ -1,4 +1,5 @@ ADD_SUBDIRECTORY(SSE) ADD_SUBDIRECTORY(AltiVec) ADD_SUBDIRECTORY(NEON) +ADD_SUBDIRECTORY(AVX) ADD_SUBDIRECTORY(Default) From 0b7f95a03f9678b4cbdbb0bb19227e54b6d05718 Mon Sep 17 00:00:00 2001 From: Benjamin Chretien Date: Sat, 3 May 2014 12:41:37 +0200 Subject: [PATCH 0482/1103] Fix typo in SparseMatrix assert. --- Eigen/src/SparseCore/SparseMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 9ac18bcf6..e0b7494c1 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -1139,7 +1139,7 @@ EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& Sparse m_data.value(p) = m_data.value(p-1); --p; } - eigen_assert((p<=startId || m_data.index(p-1)!=inner) && "you cannot insert an element that already exist, you must call coeffRef to this end"); + eigen_assert((p<=startId || m_data.index(p-1)!=inner) && "you cannot insert an element that already exists, you must call coeffRef to this end"); m_innerNonZeros[outer]++; From b5e3d76aa50dd4adc63ebb1e20e6693e261aa7dc Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 5 May 2014 14:22:27 +0200 Subject: [PATCH 0483/1103] Fixed bug #806: Missing scalar type cast in Quaternion::setFromTwoVectors() --- Eigen/src/Geometry/Quaternion.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index a712692e5..11e5398d4 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -587,7 +587,7 @@ inline Derived& QuaternionBase::setFromTwoVectors(const MatrixBase::dummy_precision()) { - c = max(c,-1); + c = (max)(c,Scalar(-1)); Matrix m; m << v0.transpose(), v1.transpose(); JacobiSVD > svd(m, ComputeFullV); Vector3 axis = svd.matrixV().col(2); From b4beba72a2d31934d3a2a49401c792f9a8cd49e0 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 5 May 2014 14:23:52 +0200 Subject: [PATCH 0484/1103] Fix bug #807: Missing scalar type cast in umeyama() --- Eigen/src/Geometry/Umeyama.h | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Geometry/Umeyama.h b/Eigen/src/Geometry/Umeyama.h index 345b47e0c..5e20662f8 100644 --- a/Eigen/src/Geometry/Umeyama.h +++ b/Eigen/src/Geometry/Umeyama.h @@ -113,7 +113,7 @@ umeyama(const MatrixBase& src, const MatrixBase& dst, boo const Index n = src.cols(); // number of measurements // required for demeaning ... - const RealScalar one_over_n = 1 / static_cast(n); + const RealScalar one_over_n = RealScalar(1) / static_cast(n); // computation of mean const VectorType src_mean = src.rowwise().sum() * one_over_n; @@ -136,16 +136,16 @@ umeyama(const MatrixBase& src, const MatrixBase& dst, boo // Eq. (39) VectorType S = VectorType::Ones(m); - if (sigma.determinant()<0) S(m-1) = -1; + if (sigma.determinant() 0 ) { + if ( svd.matrixU().determinant() * svd.matrixV().determinant() > Scalar(0) ) { Rt.block(0,0,m,m).noalias() = svd.matrixU()*svd.matrixV().transpose(); } else { - const Scalar s = S(m-1); S(m-1) = -1; + const Scalar s = S(m-1); S(m-1) = Scalar(-1); Rt.block(0,0,m,m).noalias() = svd.matrixU() * S.asDiagonal() * svd.matrixV().transpose(); S(m-1) = s; } @@ -156,7 +156,7 @@ umeyama(const MatrixBase& src, const MatrixBase& dst, boo if (with_scaling) { // Eq. (42) - const Scalar c = 1/src_var * svd.singularValues().dot(S); + const Scalar c = Scalar(1)/src_var * svd.singularValues().dot(S); // Eq. (41) Rt.col(m).head(m) = dst_mean; From 56de8d38161f1f190688d9550bff7afcf4e9dc3f Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 5 May 2014 15:03:29 +0200 Subject: [PATCH 0485/1103] Fixed unused variable warnings --- test/packetmath.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 9dab07522..a51d31dbd 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -186,7 +186,7 @@ template void packetmath() { for (int i=0; i(data1, A0, A1); internal::pstore(data2+0*PacketSize, A0); internal::pstore(data2+1*PacketSize, A1); From 84cb1d72b82589f99dd41a15485a01c3c70858d9 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 5 May 2014 15:06:37 +0200 Subject: [PATCH 0486/1103] Removed IACA-defines This caused redefinition warnings if IACA headers were included from elsewhere. For a clean solution we should define our own EIGEN_IACA_* macros --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 6 ------ 1 file changed, 6 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index cc4a9c485..de97f2b65 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -10,12 +10,6 @@ #ifndef EIGEN_GENERAL_BLOCK_PANEL_H #define EIGEN_GENERAL_BLOCK_PANEL_H -#ifdef USE_IACA -#include "iacaMarks.h" -#else -#define IACA_START -#define IACA_END -#endif namespace Eigen { From 9217de8bf2308b60a9bd410a09338301dea30134 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 5 May 2014 15:10:18 +0200 Subject: [PATCH 0487/1103] Missed to remove IACA_END in previous commit --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 1 - 1 file changed, 1 deletion(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index de97f2b65..7da52c2e8 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -799,7 +799,6 @@ void gebp_kernel blB += pk*4*RhsProgress; blA += pk*3*Traits::LhsProgress; - IACA_END } // process remaining peeled loop for(Index k=peeled_kc; k Date: Tue, 6 May 2014 12:53:18 +0200 Subject: [PATCH 0488/1103] Disabled unused warnings in Eigen2-tests --- test/eigen2/CMakeLists.txt | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/test/eigen2/CMakeLists.txt b/test/eigen2/CMakeLists.txt index 84931e037..41a02f4ad 100644 --- a/test/eigen2/CMakeLists.txt +++ b/test/eigen2/CMakeLists.txt @@ -5,6 +5,11 @@ add_dependencies(buildtests eigen2_buildtests) add_definitions("-DEIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API") +# Disable unused warnings for this module +# As EIGEN2 support is deprecated, it is not really worth fixing them +ei_add_cxx_compiler_flag("-Wno-unused-local-typedefs") +ei_add_cxx_compiler_flag("-Wno-unused-but-set-variable") + ei_add_test(eigen2_meta) ei_add_test(eigen2_sizeof) ei_add_test(eigen2_dynalloc) From 0320f7e3a71406b9a03d1bab0d168fd76e63d457 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 6 May 2014 11:18:37 -0700 Subject: [PATCH 0489/1103] Added support for fixed sized tensors. Improved support for tensor expressions. --- unsupported/Eigen/CXX11/Tensor | 2 + .../Eigen/CXX11/src/Core/util/CXX11Meta.h | 2 +- .../CXX11/src/Core/util/CXX11Workarounds.h | 12 +- .../CXX11/src/Core/util/EmulateCXX11Meta.h | 95 ++++++- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 39 +-- .../Eigen/CXX11/src/Tensor/TensorBase.h | 14 ++ .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 212 ++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 12 +- .../Eigen/CXX11/src/Tensor/TensorExpr.h | 9 +- .../Eigen/CXX11/src/Tensor/TensorFixedSize.h | 232 ++++++++++++++++++ .../src/Tensor/TensorForwardDeclarations.h | 1 + .../Eigen/CXX11/src/Tensor/TensorMap.h | 31 ++- .../Eigen/CXX11/src/Tensor/TensorStorage.h | 46 +++- .../Eigen/CXX11/src/Tensor/TensorTraits.h | 45 +++- unsupported/test/cxx11_tensor_assign.cpp | 195 +++++++++++++++ unsupported/test/cxx11_tensor_expr.cpp | 145 +++++++++++ unsupported/test/cxx11_tensor_fixed_size.cpp | 167 +++++++++++++ unsupported/test/cxx11_tensor_map.cpp | 142 +++++++++++ 18 files changed, 1319 insertions(+), 82 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h create mode 100644 unsupported/test/cxx11_tensor_assign.cpp create mode 100644 unsupported/test/cxx11_tensor_expr.cpp create mode 100644 unsupported/test/cxx11_tensor_fixed_size.cpp create mode 100644 unsupported/test/cxx11_tensor_map.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index f554c204a..f2b18ef31 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -31,6 +31,7 @@ #include "Eigen/Core" #include "unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBase.h" @@ -41,6 +42,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h" #include "unsupported/Eigen/CXX11/src/Tensor/Tensor.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMap.h" #include "Eigen/src/Core/util/ReenableStupidWarnings.h" diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h index 47f06b1b5..accaa94e7 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -112,7 +112,7 @@ template struct get<0, type_lis template struct get> { static_assert((n - n) < 0, "meta-template get: The element to extract from a list must be smaller than the size of the list."); }; template struct get> : get> {}; -template struct get<0, numeric_list> { constexpr static int value = a; }; +template struct get<0, numeric_list> { constexpr static T value = a; }; template struct get> { static_assert((n - n) < 0, "meta-template get: The element to extract from a list must be smaller than the size of the list."); }; /* always get type, regardless of dummy; good for parameter pack expansion */ diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index 77207f453..f102872ae 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -17,9 +17,6 @@ #error Intel Compiler only supports required C++ features since version 13.1. // note that most stuff in principle works with 13.0 but when combining // some features, at some point 13.0 will just fail with an internal assertion -#elif defined(__clang__) && (__clang_major__ < 3 || (__clang_major__ == 3 && __clang_minor__ < 1)) -// note that it _should_ work with 3.1 but it was only tested with 3.2 -#error Clang C++ Compiler (clang++) only supports required C++ features since version 3.1. #elif defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && (__GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 6)) // G++ < 4.6 by default will continue processing the source files - even if we use #error to make // it error out. For this reason, we use the pragma to make sure G++ aborts at the first error @@ -40,17 +37,10 @@ #error This library needs at least a C++11 compliant compiler. If you use g++/clang, please enable the -std=c++11 compiler flag. (-std=c++0x on older versions.) #endif -using std::array; - namespace Eigen { // Use std::array as Eigen array -/*template -struct array : public std::array { - array() = default; - array(const std::initializer_list& a);// : std::array(a) {}; - array(const std::array& a); -};*/ +template using array = std::array; namespace internal { diff --git a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h index 76fcba5b4..ab869177c 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h @@ -11,16 +11,63 @@ #define EIGEN_EMULATE_CXX11_META_H + namespace Eigen { // The array class is only available starting with cxx11. Emulate our own here // if needed template class array { public: - T& operator[] (size_t index) { return values[index]; } - const T& operator[] (size_t index) const { return values[index]; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE T& operator[] (size_t index) { return values[index]; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { return values[index]; } T values[n]; + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array() { } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array(const T& v) { + EIGEN_STATIC_ASSERT(n==1, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v; + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array(const T& v1, const T& v2) { + EIGEN_STATIC_ASSERT(n==2, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v1; + values[1] = v2; + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3) { + EIGEN_STATIC_ASSERT(n==3, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v1; + values[1] = v2; + values[2] = v3; + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4) { + EIGEN_STATIC_ASSERT(n==4, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v1; + values[1] = v2; + values[2] = v3; + values[3] = v4; + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4, const T& v5) { + EIGEN_STATIC_ASSERT(n==5, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v1; + values[1] = v2; + values[2] = v3; + values[3] = v4; + values[4] = v5; + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + array(std::initializer_list l) { + std::copy(l.begin(), l.end(), values); + } +#endif }; @@ -35,8 +82,10 @@ namespace internal { struct empty_list { static const std::size_t count = 0; }; template struct type_list { - T head; - Tail tail; + typedef T HeadType; + typedef Tail TailType; + static const T head; + static const Tail tail; static const std::size_t count = 1 + Tail::count; }; @@ -54,9 +103,25 @@ template<> struct make_type_list<> { }; +template struct get_type; +template +struct get_type<0, type_list > +{ + typedef Head type; +}; + +template +struct get_type > +{ + typedef typename get_type::type type; +}; + + +/* numeric list */ template struct type2val { + typedef T type; static const T value = n; }; @@ -84,6 +149,28 @@ template struct gen_numeric_list_repeated { }; +template struct get; + +template +struct get<0, type_list > +{ + typedef typename Head::type type; + static const type value = Head::value; +}; + +template +struct get > +{ + typedef typename get::type type; + static const type value = get::value; +}; + +template struct arg_prod { + static const typename NList::HeadType::type value = get<0, NList>::value * arg_prod::value; +}; +template <> struct arg_prod { + static const int value = 1; +}; template array repeat(t v) { diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index 7b8f14c6d..f5c027d1c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -60,26 +60,6 @@ namespace Eigen { namespace internal { -template -struct tensor_index_linearization_helper -{ - static inline Index run(array const& indices, array const& dimensions) - { - return array_get(indices) + - array_get(dimensions) * - tensor_index_linearization_helper::run(indices, dimensions); - } -}; - -template -struct tensor_index_linearization_helper -{ - static inline Index run(array const& indices, array const&) - { - return array_get(indices); - } -}; - /* Forward-declaration required for the symmetry support. */ template class tensor_symmetry_value_setter; @@ -102,13 +82,15 @@ class Tensor : public TensorBase > static const int Options = Options_; static const std::size_t NumIndices = NumIndices_; + typedef DSizes Dimensions; + protected: TensorStorage m_storage; public: EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return m_storage.dimensions()[n]; } - EIGEN_STRONG_INLINE array dimensions() const { return m_storage.dimensions(); } - EIGEN_STRONG_INLINE Index size() const { return internal::array_prod(m_storage.dimensions()); } + EIGEN_STRONG_INLINE const DSizes& dimensions() const { return m_storage.dimensions(); } + EIGEN_STRONG_INLINE Index size() const { return m_storage.size(); } EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } @@ -232,13 +214,6 @@ class Tensor : public TensorBase > { } -#ifdef EIGEN_HAVE_RVALUE_REFERENCES -// inline Tensor(Self&& other) -// : m_storage(other.m_storage) -// { -// } -#endif - #ifdef EIGEN_HAS_VARIADIC_TEMPLATES template inline Tensor(Index firstDimension, IndexTypes... otherDimensions) @@ -327,7 +302,11 @@ class Tensor : public TensorBase > inline Index linearizedIndex(const array& indices) const { - return internal::tensor_index_linearization_helper::run(indices, m_storage.dimensions()); + if (Options&RowMajor) { + return m_storage.dimensions().IndexOfRowMajor(indices); + } else { + return m_storage.dimensions().IndexOfColMajor(indices); + } } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 0b9f32f7f..9c7783aaf 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -62,6 +62,20 @@ class TensorBase EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> cwiseAbs() const { return derived(); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + cwisePow(Scalar exponent) const { + return TensorCwiseUnaryOp, const Derived> + (derived(), internal::scalar_pow_op(exponent)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + operator * (Scalar scale) const { + return TensorCwiseUnaryOp, const Derived> + (derived(), internal::scalar_multiple_op(scale)); + } + // Coefficient-wise binary operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h new file mode 100644 index 000000000..bd3bd5aca --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -0,0 +1,212 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H +#define EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H + + +namespace Eigen { + +/** \internal + * + * \class TensorDimensions + * \ingroup CXX11_Tensor_Module + * + * \brief Set of classes used to encode and store the dimensions of a Tensor. + * + * The Sizes class encodes as part of the type the number of dimensions and the + * sizes corresponding to each dimension. It uses no storage space since it is + * entirely known at compile time. + * The DSizes class is its dynamic sibling: the number of dimensions is known + * at compile time but the sizes are set during execution. + * + * \sa Tensor + */ + + + +// Boiler plate code +namespace internal { + +template struct dget { + static const std::size_t value = internal::get::value; + }; + + +template +struct fixed_size_tensor_index_linearization_helper +{ + template + static inline Index run(array const& indices, + const Dimensions& dimensions) + { + return array_get(indices) + + dget::value * + fixed_size_tensor_index_linearization_helper::run(indices, dimensions); + } +}; + +template +struct fixed_size_tensor_index_linearization_helper +{ + template + static inline Index run(array const& indices, + const Dimensions&) + { + return array_get(indices); + } +}; + +} // end namespace internal + + +// Fixed size +#ifndef EIGEN_EMULATE_CXX11_META_H +template +struct Sizes : internal::numeric_list { + typedef internal::numeric_list Base; + static const std::size_t total_size = internal::arg_prod(Indices...); + + static std::size_t TotalSize() { + return internal::arg_prod(Indices...); + } + + Sizes() { } + template + explicit Sizes(const array&/* indices*/) { + // todo: add assertion + } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + explicit Sizes(std::initializer_list/* l*/) { + // todo: add assertion + } +#endif + + template Sizes& operator = (const T&/* other*/) { + // add assertion failure if the size of other is different + return *this; + } + + template + size_t IndexOfColMajor(const array& indices) const { + return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this)); + } + template + size_t IndexOfRowMajor(const array& indices) const { + return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this)); + } +}; + +#else + +template +struct non_zero_size { + typedef internal::type2val type; +}; +template <> +struct non_zero_size<0> { + typedef internal::null_type type; +}; + +template struct Sizes { + typedef typename internal::make_type_list::type, typename non_zero_size::type, typename non_zero_size::type, typename non_zero_size::type, typename non_zero_size::type >::type Base; + static const size_t count = Base::count; + static const std::size_t total_size = internal::arg_prod::value; + + static const size_t TotalSize() { + return internal::arg_prod::value; + } + + Sizes() { } + template + explicit Sizes(const array& indices) { + // todo: add assertion + } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + explicit Sizes(std::initializer_list l) { + // todo: add assertion + } +#endif + + template Sizes& operator = (const T& other) { + // to do: check the size of other + return *this; + } + + template + size_t IndexOfColMajor(const array& indices) const { + return internal::fixed_size_tensor_index_linearization_helper::run(indices, *this); + } + template + size_t IndexOfRowMajor(const array& indices) const { + return internal::fixed_size_tensor_index_linearization_helper::run(indices, *this); + } +}; + +#endif + +// Boiler plate +namespace internal { +template +struct tensor_index_linearization_helper +{ + static inline Index run(array const& indices, array const& dimensions) + { + return array_get(indices) + + array_get(dimensions) * + tensor_index_linearization_helper::run(indices, dimensions); + } +}; + +template +struct tensor_index_linearization_helper +{ + static inline Index run(array const& indices, array const&) + { + return array_get(indices); + } +}; +} // end namespace internal + + + +// Dynamic size +template +struct DSizes : array { + typedef array Base; + + size_t TotalSize() const { + return internal::array_prod(*static_cast(this)); + } + + DSizes() { } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + // explicit DSizes(std::initializer_list l) : Base(l) { } +#endif + explicit DSizes(const array& a) : Base(a) { } + + DSizes& operator = (const array& other) { + *static_cast(this) = other; + return *this; + } + + // A constexpr would be so much better here + size_t IndexOfColMajor(const array& indices) const { + return internal::tensor_index_linearization_helper::run(indices, *static_cast(this)); + } + size_t IndexOfRowMajor(const array& indices) const { + return internal::tensor_index_linearization_helper::run(indices, *static_cast(this)); + } + +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index f4f10eff5..b0dbca041 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -24,15 +24,12 @@ namespace Eigen { * TODO: add support for vectorization */ - template struct TensorEvaluator { typedef typename Derived::Index Index; typedef typename Derived::Scalar Scalar; typedef typename Derived::Scalar& CoeffReturnType; - //typedef typename Derived::PacketScalar PacketScalar; - typedef TensorEvaluator nestedType; TensorEvaluator(Derived& m) : m_data(const_cast(m.data())) @@ -72,7 +69,6 @@ template struct TensorEvaluator > { typedef TensorCwiseUnaryOp XprType; - typedef TensorEvaluator nestedType; TensorEvaluator(const XprType& op) : m_functor(op.functor()), @@ -89,7 +85,7 @@ struct TensorEvaluator > private: const UnaryOp m_functor; - typename TensorEvaluator::nestedType m_argImpl; + TensorEvaluator m_argImpl; }; @@ -99,8 +95,6 @@ template struct TensorEvaluator > { typedef TensorCwiseBinaryOp XprType; - typedef TensorEvaluator leftType; - typedef TensorEvaluator rightType; TensorEvaluator(const XprType& op) : m_functor(op.functor()), @@ -118,8 +112,8 @@ struct TensorEvaluator::nestedType m_leftImpl; - typename TensorEvaluator::nestedType m_rightImpl; + TensorEvaluator m_leftImpl; + TensorEvaluator m_rightImpl; }; } // end namespace Eigen diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h index 5a45cec31..aa875dc31 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h @@ -54,7 +54,7 @@ struct nested, 1, typename eval -class TensorCwiseUnaryOp +class TensorCwiseUnaryOp : public TensorBase > { public: typedef typename Eigen::internal::traits::Scalar Scalar; @@ -75,11 +75,6 @@ class TensorCwiseUnaryOp const typename internal::remove_all::type& nestedExpression() const { return m_xpr; } - /** \returns the nested expression */ - EIGEN_DEVICE_FUNC - typename internal::remove_all::type& - nestedExpression() { return m_xpr.const_cast_derived(); } - protected: typename XprType::Nested m_xpr; const UnaryOp m_functor; @@ -124,7 +119,7 @@ struct nested, 1, typename template -class TensorCwiseBinaryOp +class TensorCwiseBinaryOp : public TensorBase > { public: typedef typename Eigen::internal::traits::Scalar Scalar; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h new file mode 100644 index 000000000..953880123 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h @@ -0,0 +1,232 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_FIXED_SIZE_H +#define EIGEN_CXX11_TENSOR_TENSOR_FIXED_SIZE_H + +namespace Eigen { + +/** \class TensorFixedSize + * \ingroup CXX11_Tensor_Module + * + * \brief The fixed sized version of the tensor class. + * + * The fixes sized equivalent of + * Eigen::Tensor t(3, 5, 7); + * is + * Eigen::TensorFixedSize> t; + */ + +template +class TensorFixedSize : public TensorBase > +{ + public: + typedef TensorFixedSize Self; + typedef TensorBase > Base; + typedef typename Eigen::internal::nested::type Nested; + typedef typename internal::traits::StorageKind StorageKind; + typedef typename internal::traits::Index Index; + typedef Scalar_ Scalar; + typedef typename internal::packet_traits::type PacketScalar; + typedef typename NumTraits::Real RealScalar; + typedef typename Base::CoeffReturnType CoeffReturnType; + + static const int Options = Options_; + typedef Dimensions_ Dimensions; + static const std::size_t NumIndices = Dimensions::count; + + protected: + TensorStorage m_storage; + + public: + EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return m_storage.dimensions()[n]; } + EIGEN_STRONG_INLINE array dimensions() const { return m_storage.dimensions(); } + EIGEN_STRONG_INLINE Index size() const { return m_storage.size(); } + EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } + EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } + + // This makes EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + // work, because that uses base().coeffRef() - and we don't yet + // implement a similar class hierarchy + inline Self& base() { return *this; } + inline const Self& base() const { return *this; } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template + inline const Scalar& coeff(Index firstIndex, IndexTypes... otherIndices) const + { + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return coeff(array{{firstIndex, otherIndices...}}); + } +#endif + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& coeff(const array& indices) const + { + eigen_internal_assert(checkIndexRange(indices)); + return m_storage.data()[linearizedIndex(indices)]; + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const + { + eigen_internal_assert(index >= 0 && index < size()); + return m_storage.data()[index]; + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template + inline Scalar& coeffRef(Index firstIndex, IndexTypes... otherIndices) + { + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return coeffRef(array{{firstIndex, otherIndices...}}); + } +#endif + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& coeffRef(const array& indices) + { + eigen_internal_assert(checkIndexRange(indices)); + return m_storage.data()[linearizedIndex(indices)]; + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) + { + eigen_internal_assert(index >= 0 && index < size()); + return m_storage.data()[index]; + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template + inline const Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) const + { + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return this->operator()(array{{firstIndex, otherIndices...}}); + } +#endif + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(const array& indices) const + { + eigen_assert(checkIndexRange(indices)); + return coeff(indices); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(Index index) const + { + eigen_internal_assert(index >= 0 && index < size()); + return coeff(index); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator[](Index index) const + { + // The bracket operator is only for vectors, use the parenthesis operator instead. + EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE); + return coeff(index); + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template + inline Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) + { + // The number of indices used to access a tensor coefficient must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(sizeof...(otherIndices) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + return operator()(array{{firstIndex, otherIndices...}}); + } +#endif + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(const array& indices) + { + eigen_assert(checkIndexRange(indices)); + return coeffRef(indices); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(Index index) + { + eigen_assert(index >= 0 && index < size()); + return coeffRef(index); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator[](Index index) + { + // The bracket operator is only for vectors, use the parenthesis operator instead + EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + return coeffRef(index); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorFixedSize() + : m_storage() + { + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorFixedSize(const Self& other) + : m_storage(other.m_storage) + { + } + +#ifdef EIGEN_HAVE_RVALUE_REFERENCES + inline TensorFixedSize(Self&& other) + : m_storage(other.m_storage) + { + } +#endif + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorFixedSize& operator=(const OtherDerived& other) + { + // FIXME: check that the dimensions of other match the dimensions of *this. + // Unfortunately this isn't possible yet when the rhs is an expression. + internal::TensorAssign::run(*this, other); + return *this; + } + + protected: + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE bool checkIndexRange(const array& /*indices*/) const + { + using internal::array_apply_and_reduce; + using internal::array_zip_and_reduce; + using internal::greater_equal_zero_op; + using internal::logical_and_op; + using internal::lesser_op; + + return true; + // check whether the indices are all >= 0 + /* array_apply_and_reduce(indices) && + // check whether the indices fit in the dimensions + array_zip_and_reduce(indices, m_storage.dimensions());*/ + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index linearizedIndex(const array& indices) const + { + if (Options&RowMajor) { + return m_storage.dimensions().IndexOfRowMajor(indices); + } else { + return m_storage.dimensions().IndexOfColMajor(indices); + } + } +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_FIXED_SIZE_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index dc97764f0..e8a2125c4 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -13,6 +13,7 @@ namespace Eigen { template class Tensor; +template class TensorFixedSize; template class TensorMap; template class TensorBase; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index 7dec1e08d..bb0b39c5a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -43,24 +43,38 @@ template class TensorMap : public TensorBase({{firstDimension}})) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. EIGEN_STATIC_ASSERT(1 == PlainObjectType::NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) } #ifdef EIGEN_HAS_VARIADIC_TEMPLATES template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions({{firstDimension, otherDimensions...}}) { + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions(array({{firstDimension, otherDimensions...}})) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == PlainObjectType::NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) } #endif + inline TensorMap(PointerArgType dataPtr, const array& dimensions) + : m_data(dataPtr), m_dimensions(dimensions) + { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index dimension(Index n) const { return m_dimensions[n]; } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Index size() const { return internal::array_prod(m_dimensions); } + EIGEN_STRONG_INLINE const typename PlainObjectType::Dimensions& dimensions() const { return m_dimensions; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index size() const { return m_dimensions.TotalSize(); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() { return m_data; } EIGEN_DEVICE_FUNC @@ -78,8 +92,13 @@ template class TensorMap : public TensorBase::run(array{{firstIndex, otherIndices...}}, m_dimensions); - return m_data[index]; + if (PlainObjectType::Options&RowMajor) { + const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); + return m_data[index]; + } else { + const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); + return m_data[index]; + } } #endif @@ -93,7 +112,7 @@ template class TensorMap : public TensorBase m_dimensions; + typename PlainObjectType::Dimensions m_dimensions; }; } // end namespace Eigen diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index 503d7cfd6..efcb39559 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -32,6 +32,35 @@ namespace Eigen { */ template class TensorStorage; + +// Pure fixed-size storage +template +class TensorStorage +{ + private: + T m_data[Size]; + FixedDimensions m_dimensions; + + public: + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorStorage() { + EIGEN_STATIC_ASSERT(Size == FixedDimensions::total_size, YOU_MADE_A_PROGRAMMING_MISTAKE) + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE T *data() { return m_data; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const T *data() const { return m_data; } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const FixedDimensions dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE DenseIndex size() const { return m_dimensions.TotalSize(); } +}; + + + // pure-dynamic, but without specification of all dimensions explicitly template class TensorStorage @@ -44,7 +73,7 @@ class TensorStorage TensorStorage(const TensorStorage& other) : Base_(other) { } #ifdef EIGEN_HAVE_RVALUE_REFERENCES -// TensorStorage(TensorStorage&&) = default; + // TensorStorage(TensorStorage&&) = default; #endif TensorStorage(internal::constructor_without_unaligned_array_assert) : Base_(internal::constructor_without_unaligned_array_assert()) {} TensorStorage(DenseIndex size, const array& dimensions) : Base_(size, dimensions) {} @@ -57,11 +86,11 @@ template class TensorStorage::type> { T *m_data; - array m_dimensions; + DSizes m_dimensions; typedef TensorStorage::type> Self_; public: - TensorStorage() : m_data(0), m_dimensions() {} + TensorStorage() : m_data(0), m_dimensions() {} TensorStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_dimensions(internal::template repeat(0)) {} TensorStorage(DenseIndex size, const array& dimensions) @@ -83,25 +112,25 @@ class TensorStorage(m_data, internal::array_prod(m_dimensions)); } void swap(Self_& other) { std::swap(m_data,other.m_data); std::swap(m_dimensions,other.m_dimensions); } - const array& dimensions() const {return m_dimensions;} + const DSizes& dimensions() const {return m_dimensions;} void conservativeResize(DenseIndex size, const array& nbDimensions) { @@ -124,9 +153,10 @@ class TensorStorage > }; +template +struct traits > +{ + typedef Scalar_ Scalar; + typedef Dense StorageKind; + typedef DenseIndex Index; +}; + + template struct traits > : public traits @@ -68,16 +77,28 @@ struct traits > }; -template -struct eval, Eigen::Dense> +template +struct eval, Eigen::Dense> { - typedef const Tensor<_Scalar, NumIndices_, Options_>& type; + typedef const Tensor<_Scalar, NumIndices_, Options>& type; }; -template -struct eval, Eigen::Dense> +template +struct eval, Eigen::Dense> { - typedef const Tensor<_Scalar, NumIndices_, Options_>& type; + typedef const Tensor<_Scalar, NumIndices_, Options>& type; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorFixedSize& type; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorFixedSize& type; }; template @@ -104,6 +125,18 @@ struct nested, 1, typename eval& type; }; +template +struct nested, 1, typename eval >::type> +{ + typedef const TensorFixedSize& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef const TensorFixedSize& type; +}; + template struct nested, 1, typename eval >::type> { diff --git a/unsupported/test/cxx11_tensor_assign.cpp b/unsupported/test/cxx11_tensor_assign.cpp new file mode 100644 index 000000000..c88872950 --- /dev/null +++ b/unsupported/test/cxx11_tensor_assign.cpp @@ -0,0 +1,195 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +static void test_1d() +{ + Tensor vec1(6); + Tensor vec2(6); + vec1(0) = 4; vec2(0) = 0; + vec1(1) = 8; vec2(1) = 1; + vec1(2) = 15; vec2(2) = 2; + vec1(3) = 16; vec2(3) = 3; + vec1(4) = 23; vec2(4) = 4; + vec1(5) = 42; vec2(5) = 5; + + int col_major[6]; + int row_major[6]; + memset(col_major, 0, 6*sizeof(int)); + memset(row_major, 0, 6*sizeof(int)); + TensorMap> vec3(col_major, 6); + TensorMap> vec4(row_major, 6); + + vec3 = vec1; + vec4 = vec2; + + VERIFY_IS_EQUAL(vec3(0), 4); + VERIFY_IS_EQUAL(vec3(1), 8); + VERIFY_IS_EQUAL(vec3(2), 15); + VERIFY_IS_EQUAL(vec3(3), 16); + VERIFY_IS_EQUAL(vec3(4), 23); + VERIFY_IS_EQUAL(vec3(5), 42); + + VERIFY_IS_EQUAL(vec4(0), 0); + VERIFY_IS_EQUAL(vec4(1), 1); + VERIFY_IS_EQUAL(vec4(2), 2); + VERIFY_IS_EQUAL(vec4(3), 3); + VERIFY_IS_EQUAL(vec4(4), 4); + VERIFY_IS_EQUAL(vec4(5), 5); + + vec1.setZero(); + vec2.setZero(); + vec1 = vec3; + vec2 = vec4; + + VERIFY_IS_EQUAL(vec1(0), 4); + VERIFY_IS_EQUAL(vec1(1), 8); + VERIFY_IS_EQUAL(vec1(2), 15); + VERIFY_IS_EQUAL(vec1(3), 16); + VERIFY_IS_EQUAL(vec1(4), 23); + VERIFY_IS_EQUAL(vec1(5), 42); + + VERIFY_IS_EQUAL(vec2(0), 0); + VERIFY_IS_EQUAL(vec2(1), 1); + VERIFY_IS_EQUAL(vec2(2), 2); + VERIFY_IS_EQUAL(vec2(3), 3); + VERIFY_IS_EQUAL(vec2(4), 4); + VERIFY_IS_EQUAL(vec2(5), 5); +} + +static void test_2d() +{ + Tensor mat1(2,3); + Tensor mat2(2,3); + + mat1(0,0) = 0; + mat1(0,1) = 1; + mat1(0,2) = 2; + mat1(1,0) = 3; + mat1(1,1) = 4; + mat1(1,2) = 5; + + mat2(0,0) = 0; + mat2(0,1) = 1; + mat2(0,2) = 2; + mat2(1,0) = 3; + mat2(1,1) = 4; + mat2(1,2) = 5; + + int col_major[6]; + int row_major[6]; + memset(col_major, 0, 6*sizeof(int)); + memset(row_major, 0, 6*sizeof(int)); + TensorMap> mat3(row_major, 2, 3); + TensorMap> mat4(col_major, 2, 3); + + mat3 = mat1; + mat4 = mat2; + + VERIFY_IS_EQUAL(mat3(0,0), 0); + VERIFY_IS_EQUAL(mat3(0,1), 1); + VERIFY_IS_EQUAL(mat3(0,2), 2); + VERIFY_IS_EQUAL(mat3(1,0), 3); + VERIFY_IS_EQUAL(mat3(1,1), 4); + VERIFY_IS_EQUAL(mat3(1,2), 5); + + VERIFY_IS_EQUAL(mat4(0,0), 0); + VERIFY_IS_EQUAL(mat4(0,1), 1); + VERIFY_IS_EQUAL(mat4(0,2), 2); + VERIFY_IS_EQUAL(mat4(1,0), 3); + VERIFY_IS_EQUAL(mat4(1,1), 4); + VERIFY_IS_EQUAL(mat4(1,2), 5); + + mat1.setZero(); + mat2.setZero(); + mat1 = mat3; + mat2 = mat4; + + VERIFY_IS_EQUAL(mat1(0,0), 0); + VERIFY_IS_EQUAL(mat1(0,1), 1); + VERIFY_IS_EQUAL(mat1(0,2), 2); + VERIFY_IS_EQUAL(mat1(1,0), 3); + VERIFY_IS_EQUAL(mat1(1,1), 4); + VERIFY_IS_EQUAL(mat1(1,2), 5); + + VERIFY_IS_EQUAL(mat2(0,0), 0); + VERIFY_IS_EQUAL(mat2(0,1), 1); + VERIFY_IS_EQUAL(mat2(0,2), 2); + VERIFY_IS_EQUAL(mat2(1,0), 3); + VERIFY_IS_EQUAL(mat2(1,1), 4); + VERIFY_IS_EQUAL(mat2(1,2), 5); +} + +static void test_3d() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + + int val = 0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + mat1(i,j,k) = val; + mat2(i,j,k) = val; + val++; + } + } + } + + int col_major[2*3*7]; + int row_major[2*3*7]; + memset(col_major, 0, 2*3*7*sizeof(int)); + memset(row_major, 0, 2*3*7*sizeof(int)); + TensorMap> mat3(col_major, 2, 3, 7); + TensorMap> mat4(row_major, 2, 3, 7); + + mat3 = mat1; + mat4 = mat2; + + val = 0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(mat3(i,j,k), val); + VERIFY_IS_EQUAL(mat4(i,j,k), val); + val++; + } + } + } + + mat1.setZero(); + mat2.setZero(); + mat1 = mat3; + mat2 = mat4; + + val = 0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(mat1(i,j,k), val); + VERIFY_IS_EQUAL(mat2(i,j,k), val); + val++; + } + } + } +} + + +void test_cxx11_tensor_assign() +{ + CALL_SUBTEST(test_1d()); + CALL_SUBTEST(test_2d()); + CALL_SUBTEST(test_3d()); +} diff --git a/unsupported/test/cxx11_tensor_expr.cpp b/unsupported/test/cxx11_tensor_expr.cpp new file mode 100644 index 000000000..e0124da8c --- /dev/null +++ b/unsupported/test/cxx11_tensor_expr.cpp @@ -0,0 +1,145 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +static void test_1d() +{ + Tensor vec1({6}); + Tensor vec2({6}); + + vec1(0) = 4.0; vec2(0) = 0.0; + vec1(1) = 8.0; vec2(1) = 1.0; + vec1(2) = 15.0; vec2(2) = 2.0; + vec1(3) = 16.0; vec2(3) = 3.0; + vec1(4) = 23.0; vec2(4) = 4.0; + vec1(5) = 42.0; vec2(5) = 5.0; + + float data3[6]; + TensorMap> vec3(data3, 6); + vec3 = vec1.cwiseSqrt(); + float data4[6]; + TensorMap> vec4(data4, 6); + vec4 = vec2.cwiseSqrt(); + + VERIFY_IS_APPROX(vec3(0), sqrtf(4.0)); + VERIFY_IS_APPROX(vec3(1), sqrtf(8.0)); + VERIFY_IS_APPROX(vec3(2), sqrtf(15.0)); + VERIFY_IS_APPROX(vec3(3), sqrtf(16.0)); + VERIFY_IS_APPROX(vec3(4), sqrtf(23.0)); + VERIFY_IS_APPROX(vec3(5), sqrtf(42.0)); + + VERIFY_IS_APPROX(vec4(0), sqrtf(0.0)); + VERIFY_IS_APPROX(vec4(1), sqrtf(1.0)); + VERIFY_IS_APPROX(vec4(2), sqrtf(2.0)); + VERIFY_IS_APPROX(vec4(3), sqrtf(3.0)); + VERIFY_IS_APPROX(vec4(4), sqrtf(4.0)); + VERIFY_IS_APPROX(vec4(5), sqrtf(5.0)); + + vec3 = vec1 + vec2; + VERIFY_IS_APPROX(vec3(0), 4.0f + 0.0f); + VERIFY_IS_APPROX(vec3(1), 8.0f + 1.0f); + VERIFY_IS_APPROX(vec3(2), 15.0f + 2.0f); + VERIFY_IS_APPROX(vec3(3), 16.0f + 3.0f); + VERIFY_IS_APPROX(vec3(4), 23.0f + 4.0f); + VERIFY_IS_APPROX(vec3(5), 42.0f + 5.0f); +} + +static void test_2d() +{ + float data1[6]; + TensorMap> mat1(data1, 2, 3); + float data2[6]; + TensorMap> mat2(data2, 2, 3); + + mat1(0,0) = 0.0; + mat1(0,1) = 1.0; + mat1(0,2) = 2.0; + mat1(1,0) = 3.0; + mat1(1,1) = 4.0; + mat1(1,2) = 5.0; + + mat2(0,0) = -0.0; + mat2(0,1) = -1.0; + mat2(0,2) = -2.0; + mat2(1,0) = -3.0; + mat2(1,1) = -4.0; + mat2(1,2) = -5.0; + + Tensor mat3(2,3); + Tensor mat4(2,3); + mat3 = mat1.cwiseAbs(); + mat4 = mat2.cwiseAbs(); + + VERIFY_IS_APPROX(mat3(0,0), 0.0f); + VERIFY_IS_APPROX(mat3(0,1), 1.0f); + VERIFY_IS_APPROX(mat3(0,2), 2.0f); + VERIFY_IS_APPROX(mat3(1,0), 3.0f); + VERIFY_IS_APPROX(mat3(1,1), 4.0f); + VERIFY_IS_APPROX(mat3(1,2), 5.0f); + + VERIFY_IS_APPROX(mat4(0,0), 0.0f); + VERIFY_IS_APPROX(mat4(0,1), 1.0f); + VERIFY_IS_APPROX(mat4(0,2), 2.0f); + VERIFY_IS_APPROX(mat4(1,0), 3.0f); + VERIFY_IS_APPROX(mat4(1,1), 4.0f); + VERIFY_IS_APPROX(mat4(1,2), 5.0f); +} + +static void test_3d() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + + float val = 0.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + mat1(i,j,k) = val; + mat2(i,j,k) = val; + val += 1.0; + } + } + } + + Tensor mat3(2,3,7); + mat3 = mat1 + mat1; + Tensor mat4(2,3,7); + mat4 = mat2 * 3.14f; + Tensor mat5(2,3,7); + mat5 = mat1.cwiseSqrt().cwiseSqrt(); + Tensor mat6(2,3,7); + mat6 = mat2.cwiseSqrt() * 3.14f; + + val = 0.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat3(i,j,k), val + val); + VERIFY_IS_APPROX(mat4(i,j,k), val * 3.14f); + VERIFY_IS_APPROX(mat5(i,j,k), sqrtf(sqrtf(val))); + VERIFY_IS_APPROX(mat6(i,j,k), sqrtf(val) * 3.14f); + val += 1.0; + } + } + } +} + + +void test_cxx11_tensor_expr() +{ + CALL_SUBTEST(test_1d()); + CALL_SUBTEST(test_2d()); + CALL_SUBTEST(test_3d()); +} diff --git a/unsupported/test/cxx11_tensor_fixed_size.cpp b/unsupported/test/cxx11_tensor_fixed_size.cpp new file mode 100644 index 000000000..c1d74d881 --- /dev/null +++ b/unsupported/test/cxx11_tensor_fixed_size.cpp @@ -0,0 +1,167 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + + +static void test_1d() +{ + TensorFixedSize > vec1; + TensorFixedSize, RowMajor> vec2; + + VERIFY_IS_EQUAL((vec1.size()), 6); + // VERIFY_IS_EQUAL((vec1.dimensions()[0]), 6); + // VERIFY_IS_EQUAL((vec1.dimension(0)), 6); + + vec1(0) = 4.0; vec2(0) = 0.0; + vec1(1) = 8.0; vec2(1) = 1.0; + vec1(2) = 15.0; vec2(2) = 2.0; + vec1(3) = 16.0; vec2(3) = 3.0; + vec1(4) = 23.0; vec2(4) = 4.0; + vec1(5) = 42.0; vec2(5) = 5.0; + + float data3[6]; + TensorMap > > vec3(data3, 6); + vec3 = vec1.cwiseSqrt(); + float data4[6]; + TensorMap, RowMajor> > vec4(data4, 6); + vec4 = vec2.cwiseSqrt(); + + VERIFY_IS_EQUAL((vec3.size()), 6); + // VERIFY_IS_EQUAL((vec3.dimensions()[0]), 6); + // VERIFY_IS_EQUAL((vec3.dimension(0)), 6); + + VERIFY_IS_APPROX(vec3(0), sqrtf(4.0)); + VERIFY_IS_APPROX(vec3(1), sqrtf(8.0)); + VERIFY_IS_APPROX(vec3(2), sqrtf(15.0)); + VERIFY_IS_APPROX(vec3(3), sqrtf(16.0)); + VERIFY_IS_APPROX(vec3(4), sqrtf(23.0)); + VERIFY_IS_APPROX(vec3(5), sqrtf(42.0)); + + VERIFY_IS_APPROX(vec4(0), sqrtf(0.0)); + VERIFY_IS_APPROX(vec4(1), sqrtf(1.0)); + VERIFY_IS_APPROX(vec4(2), sqrtf(2.0)); + VERIFY_IS_APPROX(vec4(3), sqrtf(3.0)); + VERIFY_IS_APPROX(vec4(4), sqrtf(4.0)); + VERIFY_IS_APPROX(vec4(5), sqrtf(5.0)); + + vec3 = vec1 + vec2; + VERIFY_IS_APPROX(vec3(0), 4.0f + 0.0f); + VERIFY_IS_APPROX(vec3(1), 8.0f + 1.0f); + VERIFY_IS_APPROX(vec3(2), 15.0f + 2.0f); + VERIFY_IS_APPROX(vec3(3), 16.0f + 3.0f); + VERIFY_IS_APPROX(vec3(4), 23.0f + 4.0f); + VERIFY_IS_APPROX(vec3(5), 42.0f + 5.0f); +} + +static void test_2d() +{ + float data1[6]; + TensorMap >> mat1(data1,2,3); + float data2[6]; + TensorMap, RowMajor>> mat2(data2,2,3); + + VERIFY_IS_EQUAL((mat1.size()), 2*3); + // VERIFY_IS_EQUAL((mat1.dimension(0)), 2); + // VERIFY_IS_EQUAL((mat1.dimension(1)), 3); + + mat1(0,0) = 0.0; + mat1(0,1) = 1.0; + mat1(0,2) = 2.0; + mat1(1,0) = 3.0; + mat1(1,1) = 4.0; + mat1(1,2) = 5.0; + + mat2(0,0) = -0.0; + mat2(0,1) = -1.0; + mat2(0,2) = -2.0; + mat2(1,0) = -3.0; + mat2(1,1) = -4.0; + mat2(1,2) = -5.0; + + TensorFixedSize> mat3; + TensorFixedSize, RowMajor> mat4; + mat3 = mat1.cwiseAbs(); + mat4 = mat2.cwiseAbs(); + + VERIFY_IS_EQUAL((mat3.size()), 2*3); + // VERIFY_IS_EQUAL((mat3.dimension(0)), 2); + // VERIFY_IS_EQUAL((mat3.dimension(1)), 3); + + VERIFY_IS_APPROX(mat3(0,0), 0.0f); + VERIFY_IS_APPROX(mat3(0,1), 1.0f); + VERIFY_IS_APPROX(mat3(0,2), 2.0f); + VERIFY_IS_APPROX(mat3(1,0), 3.0f); + VERIFY_IS_APPROX(mat3(1,1), 4.0f); + VERIFY_IS_APPROX(mat3(1,2), 5.0f); + + VERIFY_IS_APPROX(mat4(0,0), 0.0f); + VERIFY_IS_APPROX(mat4(0,1), 1.0f); + VERIFY_IS_APPROX(mat4(0,2), 2.0f); + VERIFY_IS_APPROX(mat4(1,0), 3.0f); + VERIFY_IS_APPROX(mat4(1,1), 4.0f); + VERIFY_IS_APPROX(mat4(1,2), 5.0f); +} + +static void test_3d() +{ + TensorFixedSize > mat1; + TensorFixedSize, RowMajor> mat2; + + VERIFY_IS_EQUAL((mat1.size()), 2*3*7); + // VERIFY_IS_EQUAL((mat1.dimension(0)), 2); + // VERIFY_IS_EQUAL((mat1.dimension(1)), 3); + // VERIFY_IS_EQUAL((mat1.dimension(2)), 7); + + float val = 0.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + mat1(i,j,k) = val; + mat2(i,j,k) = val; + val += 1.0; + } + } + } + + TensorFixedSize > mat3; + mat3 = mat1.cwiseSqrt(); + TensorFixedSize, RowMajor> mat4; + mat4 = mat2.cwiseSqrt(); + + VERIFY_IS_EQUAL((mat3.size()), 2*3*7); + // VERIFY_IS_EQUAL((mat3.dimension(0)), 2); + // VERIFY_IS_EQUAL((mat3.dimension(1)), 3); + // VERIFY_IS_EQUAL((mat3.dimension(2)), 7); + + + val = 0.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat3(i,j,k), sqrtf(val)); + VERIFY_IS_APPROX(mat4(i,j,k), sqrtf(val)); + val += 1.0; + } + } + } +} + + +void test_cxx11_tensor_fixed_size() +{ + CALL_SUBTEST(test_1d()); + CALL_SUBTEST(test_2d()); + CALL_SUBTEST(test_3d()); +} diff --git a/unsupported/test/cxx11_tensor_map.cpp b/unsupported/test/cxx11_tensor_map.cpp new file mode 100644 index 000000000..478c20306 --- /dev/null +++ b/unsupported/test/cxx11_tensor_map.cpp @@ -0,0 +1,142 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +static void test_1d() +{ + Tensor vec1(6); + Tensor vec2(6); + + TensorMap> vec3(vec1.data(), 6); + TensorMap> vec4(vec2.data(), 6); + + vec1(0) = 4; vec2(0) = 0; + vec1(1) = 8; vec2(1) = 1; + vec1(2) = 15; vec2(2) = 2; + vec1(3) = 16; vec2(3) = 3; + vec1(4) = 23; vec2(4) = 4; + vec1(5) = 42; vec2(5) = 5; + + VERIFY_IS_EQUAL(vec1.size(), 6); + VERIFY_IS_EQUAL(vec1.dimension(0), 6); + + VERIFY_IS_EQUAL(vec3(0), 4); + VERIFY_IS_EQUAL(vec3(1), 8); + VERIFY_IS_EQUAL(vec3(2), 15); + VERIFY_IS_EQUAL(vec3(3), 16); + VERIFY_IS_EQUAL(vec3(4), 23); + VERIFY_IS_EQUAL(vec3(5), 42); + + VERIFY_IS_EQUAL(vec4(0), 0); + VERIFY_IS_EQUAL(vec4(1), 1); + VERIFY_IS_EQUAL(vec4(2), 2); + VERIFY_IS_EQUAL(vec4(3), 3); + VERIFY_IS_EQUAL(vec4(4), 4); + VERIFY_IS_EQUAL(vec4(5), 5); +} + +static void test_2d() +{ + Tensor mat1(2,3); + Tensor mat2(2,3); + + mat1(0,0) = 0; + mat1(0,1) = 1; + mat1(0,2) = 2; + mat1(1,0) = 3; + mat1(1,1) = 4; + mat1(1,2) = 5; + + mat2(0,0) = 0; + mat2(0,1) = 1; + mat2(0,2) = 2; + mat2(1,0) = 3; + mat2(1,1) = 4; + mat2(1,2) = 5; + + TensorMap> mat3(mat1.data(), 2, 3); + TensorMap> mat4(mat2.data(), 2, 3); + + VERIFY_IS_EQUAL(mat3.size(), 6); + VERIFY_IS_EQUAL(mat3.dimension(0), 2); + VERIFY_IS_EQUAL(mat3.dimension(1), 3); + + VERIFY_IS_EQUAL(mat4.size(), 6); + VERIFY_IS_EQUAL(mat4.dimension(0), 2); + VERIFY_IS_EQUAL(mat4.dimension(1), 3); + + VERIFY_IS_EQUAL(mat3(0,0), 0); + VERIFY_IS_EQUAL(mat3(0,1), 1); + VERIFY_IS_EQUAL(mat3(0,2), 2); + VERIFY_IS_EQUAL(mat3(1,0), 3); + VERIFY_IS_EQUAL(mat3(1,1), 4); + VERIFY_IS_EQUAL(mat3(1,2), 5); + + VERIFY_IS_EQUAL(mat4(0,0), 0); + VERIFY_IS_EQUAL(mat4(0,1), 1); + VERIFY_IS_EQUAL(mat4(0,2), 2); + VERIFY_IS_EQUAL(mat4(1,0), 3); + VERIFY_IS_EQUAL(mat4(1,1), 4); + VERIFY_IS_EQUAL(mat4(1,2), 5); +} + +static void test_3d() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + + int val = 0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + mat1(i,j,k) = val; + mat2(i,j,k) = val; + val++; + } + } + } + + TensorMap> mat3(mat1.data(), 2, 3, 7); + TensorMap> mat4(mat2.data(), 2, 3, 7); + + VERIFY_IS_EQUAL(mat3.size(), 2*3*7); + VERIFY_IS_EQUAL(mat3.dimension(0), 2); + VERIFY_IS_EQUAL(mat3.dimension(1), 3); + VERIFY_IS_EQUAL(mat3.dimension(2), 7); + + VERIFY_IS_EQUAL(mat4.size(), 2*3*7); + VERIFY_IS_EQUAL(mat4.dimension(0), 2); + VERIFY_IS_EQUAL(mat4.dimension(1), 3); + VERIFY_IS_EQUAL(mat4.dimension(2), 7); + + val = 0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(mat3(i,j,k), val); + VERIFY_IS_EQUAL(mat4(i,j,k), val); + val++; + } + } + } +} + + +void test_cxx11_tensor_map() +{ + CALL_SUBTEST(test_1d()); + CALL_SUBTEST(test_2d()); + CALL_SUBTEST(test_3d()); +} From 881aab14b47edd02ec0a99167fd19652202ec548 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 7 May 2014 13:34:46 -0700 Subject: [PATCH 0490/1103] Made it possible to call the assignment operator on an Eigen::Block from a CUDA kernel. --- Eigen/src/Core/util/Macros.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index f69fc5ec4..32790fddb 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -326,13 +326,13 @@ namespace Eigen { #elif defined(__clang__) // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653) #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \ using Base::operator =; \ - EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \ template \ - EIGEN_STRONG_INLINE Derived& operator=(const DenseBase& other) { Base::operator=(other.derived()); return *this; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const DenseBase& other) { Base::operator=(other.derived()); return *this; } #else #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \ using Base::operator =; \ - EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) \ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) \ { \ Base::operator=(other); \ return *this; \ From edebb152758447c34b9e816154a5236478dc5d3e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Mon, 19 May 2014 18:21:29 +0200 Subject: [PATCH 0491/1103] PolynomialSolver: add a test to reveal a bug. --- unsupported/test/polynomialsolver.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/unsupported/test/polynomialsolver.cpp b/unsupported/test/polynomialsolver.cpp index 680ff6d31..f62c98de2 100644 --- a/unsupported/test/polynomialsolver.cpp +++ b/unsupported/test/polynomialsolver.cpp @@ -210,5 +210,6 @@ void test_polynomialsolver() CALL_SUBTEST_10((polynomialsolver( internal::random(9,13) )) ); + CALL_SUBTEST_11((polynomialsolver(1)) ); } } From 0f946079471dc5d6d694d892a843dc1c61d6a859 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Mon, 19 May 2014 18:34:10 +0200 Subject: [PATCH 0492/1103] PolynomialSolver: test template constructor in test suite. --- unsupported/test/polynomialsolver.cpp | 3 +++ 1 file changed, 3 insertions(+) diff --git a/unsupported/test/polynomialsolver.cpp b/unsupported/test/polynomialsolver.cpp index f62c98de2..0c87478dd 100644 --- a/unsupported/test/polynomialsolver.cpp +++ b/unsupported/test/polynomialsolver.cpp @@ -38,6 +38,9 @@ bool aux_evalSolver( const POLYNOMIAL& pols, SOLVER& psolve ) const Index deg = pols.size()-1; + // Test template constructor from coefficient vector + SOLVER solve_constr (pols); + psolve.compute( pols ); const RootsType& roots( psolve.roots() ); EvalRootsType evr( deg ); From df92649379d9dded0e807fde758e7394d8c5ca4c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Mon, 19 May 2014 18:40:29 +0200 Subject: [PATCH 0493/1103] PolynomialSolver: add missing constructors. --- unsupported/Eigen/src/Polynomials/PolynomialSolver.h | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/unsupported/Eigen/src/Polynomials/PolynomialSolver.h b/unsupported/Eigen/src/Polynomials/PolynomialSolver.h index cd5c04bbf..5d00fbeac 100644 --- a/unsupported/Eigen/src/Polynomials/PolynomialSolver.h +++ b/unsupported/Eigen/src/Polynomials/PolynomialSolver.h @@ -380,6 +380,13 @@ class PolynomialSolver<_Scalar,1> : public PolynomialSolverBase<_Scalar,1> m_roots[0] = -poly[0]/poly[poly.size()-1]; } + public: + template< typename OtherPolynomial > + inline PolynomialSolver( const OtherPolynomial& poly ){ + compute( poly ); } + + inline PolynomialSolver(){} + protected: using PS_Base::m_roots; }; From eda79321becc9886e922654210d0e3ae28f64647 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Mon, 19 May 2014 19:08:51 +0200 Subject: [PATCH 0494/1103] PolynomialSolver: fix bugs related to linear polynomials. --- .../Eigen/src/Polynomials/PolynomialSolver.h | 24 +++++++++++++------ 1 file changed, 17 insertions(+), 7 deletions(-) diff --git a/unsupported/Eigen/src/Polynomials/PolynomialSolver.h b/unsupported/Eigen/src/Polynomials/PolynomialSolver.h index 5d00fbeac..66a91d1a2 100644 --- a/unsupported/Eigen/src/Polynomials/PolynomialSolver.h +++ b/unsupported/Eigen/src/Polynomials/PolynomialSolver.h @@ -41,7 +41,7 @@ class PolynomialSolverBase protected: template< typename OtherPolynomial > inline void setPolynomial( const OtherPolynomial& poly ){ - m_roots.resize(poly.size()); } + m_roots.resize(poly.size()-1); } public: template< typename OtherPolynomial > @@ -345,10 +345,19 @@ class PolynomialSolver : public PolynomialSolverBase<_Scalar,_Deg> void compute( const OtherPolynomial& poly ) { eigen_assert( Scalar(0) != poly[poly.size()-1] ); - internal::companion companion( poly ); - companion.balance(); - m_eigenSolver.compute( companion.denseMatrix() ); - m_roots = m_eigenSolver.eigenvalues(); + eigen_assert( poly.size() > 1 ); + if(poly.size() > 2 ) + { + internal::companion companion( poly ); + companion.balance(); + m_eigenSolver.compute( companion.denseMatrix() ); + m_roots = m_eigenSolver.eigenvalues(); + } + else if(poly.size () == 2) + { + m_roots.resize(1); + m_roots[0] = -poly[0]/poly[1]; + } } public: @@ -376,8 +385,9 @@ class PolynomialSolver<_Scalar,1> : public PolynomialSolverBase<_Scalar,1> template< typename OtherPolynomial > void compute( const OtherPolynomial& poly ) { - eigen_assert( Scalar(0) != poly[poly.size()-1] ); - m_roots[0] = -poly[0]/poly[poly.size()-1]; + eigen_assert( poly.size() == 2 ); + eigen_assert( Scalar(0) != poly[1] ); + m_roots[0] = -poly[0]/poly[1]; } public: From c55c5763fed4c67ede074cae13ce872f2412b2e3 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Mon, 19 May 2014 19:24:02 +0200 Subject: [PATCH 0495/1103] PolynomialSolver: fix typo. --- unsupported/Eigen/src/Polynomials/PolynomialSolver.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/Polynomials/PolynomialSolver.h b/unsupported/Eigen/src/Polynomials/PolynomialSolver.h index 66a91d1a2..03198ec8e 100644 --- a/unsupported/Eigen/src/Polynomials/PolynomialSolver.h +++ b/unsupported/Eigen/src/Polynomials/PolynomialSolver.h @@ -316,7 +316,7 @@ class PolynomialSolverBase * - real roots with greatest, smallest absolute real value. * - greatest, smallest real roots. * - * WARNING: this polynomial solver is experimental, part of the unsuported Eigen modules. + * WARNING: this polynomial solver is experimental, part of the unsupported Eigen modules. * * * Currently a QR algorithm is used to compute the eigenvalues of the companion matrix of From aa524604b7ba36c164df3a47a6b6ef12952686b9 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 21 May 2014 14:08:04 +0000 Subject: [PATCH 0496/1103] README.md edited online with Bitbucket --- README.md | 3 +++ 1 file changed, 3 insertions(+) create mode 100644 README.md diff --git a/README.md b/README.md new file mode 100644 index 000000000..04fa63fce --- /dev/null +++ b/README.md @@ -0,0 +1,3 @@ +**Eigen is a C++ template library for linear algebra: matrices, vectors, numerical solvers, and related algorithms.** + +For more information go to http://eigen.tuxfamily.org/. From c794099e69353d82cfb0601c680d15e5848c7bc0 Mon Sep 17 00:00:00 2001 From: Mark Borgerding Date: Thu, 8 May 2014 15:14:12 -0400 Subject: [PATCH 0497/1103] fixed AsciiQuickReference typo: LinSpace -> LinSpaced (transplanted from e66781905586e3c438031597fae07306d47fea60 ) --- doc/AsciiQuickReference.txt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/doc/AsciiQuickReference.txt b/doc/AsciiQuickReference.txt index 4c8fe2f47..f62c112cd 100644 --- a/doc/AsciiQuickReference.txt +++ b/doc/AsciiQuickReference.txt @@ -41,8 +41,8 @@ MatrixXd::Ones(rows,cols) // ones(rows,cols) C.setOnes(rows,cols) // C = ones(rows,cols) MatrixXd::Random(rows,cols) // rand(rows,cols)*2-1 // MatrixXd::Random returns uniform random numbers in (-1, 1). C.setRandom(rows,cols) // C = rand(rows,cols)*2-1 -VectorXd::LinSpace(size,low,high) // linspace(low,high,size)' -v.setLinSpace(size,low,high) // v = linspace(low,high,size)' +VectorXd::LinSpaced(size,low,high) // linspace(low,high,size)' +v.setLinSpaced(size,low,high) // v = linspace(low,high,size)' // Matrix slicing and blocks. All expressions listed here are read/write. From e3ab46b8c931b4435fe2283d591023de9e257b1a Mon Sep 17 00:00:00 2001 From: Mark Borgerding Date: Fri, 16 May 2014 13:45:35 -0400 Subject: [PATCH 0498/1103] AsciiQuickReference: added .real(), .imag() (transplanted from 11462c1a291bdb9c0ac27db25fef364e51632484 ) --- doc/AsciiQuickReference.txt | 2 ++ 1 file changed, 2 insertions(+) diff --git a/doc/AsciiQuickReference.txt b/doc/AsciiQuickReference.txt index f62c112cd..c4d021624 100644 --- a/doc/AsciiQuickReference.txt +++ b/doc/AsciiQuickReference.txt @@ -168,6 +168,8 @@ x.cross(y) // cross(x, y) Requires #include A.cast(); // double(A) A.cast(); // single(A) A.cast(); // int32(A) +A.real(); // real(A) +A.imag(); // imag(A) // if the original type equals destination type, no work is done // Note that for most operations Eigen requires all operands to have the same type: From 7402fea0a8e63e3ea248257047c584afee8f8bde Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 16 May 2014 15:08:05 -0700 Subject: [PATCH 0499/1103] Vectorized the evaluation of tensor expression (using SSE, AVX, NEON, ...) Added the ability to parallelize the evaluation of a tensor expression over multiple cpu cores. Added the ability to offload the evaluation of a tensor expression to a GPU. --- unsupported/Eigen/CXX11/Tensor | 2 + unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 8 +- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 145 +++++++++++++++- .../Eigen/CXX11/src/Tensor/TensorBase.h | 12 ++ .../Eigen/CXX11/src/Tensor/TensorDevice.h | 83 +++++++++ .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 56 +++++++ .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 14 +- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 54 ++++-- .../Eigen/CXX11/src/Tensor/TensorExpr.h | 25 ++- .../Eigen/CXX11/src/Tensor/TensorFixedSize.h | 10 +- .../src/Tensor/TensorForwardDeclarations.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorMap.h | 158 ++++++++++++++++-- .../Eigen/CXX11/src/Tensor/TensorStorage.h | 19 --- unsupported/test/CMakeLists.txt | 3 + unsupported/test/cxx11_tensor_device.cpp | 126 ++++++++++++++ unsupported/test/cxx11_tensor_fixed_size.cpp | 28 ++++ unsupported/test/cxx11_tensor_thread_pool.cpp | 37 ++++ 17 files changed, 719 insertions(+), 65 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h create mode 100644 unsupported/test/cxx11_tensor_device.cpp create mode 100644 unsupported/test/cxx11_tensor_thread_pool.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index f2b18ef31..323d9edff 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -31,6 +31,7 @@ #include "Eigen/Core" #include "unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h" @@ -39,6 +40,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h" #include "unsupported/Eigen/CXX11/src/Tensor/Tensor.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index f5c027d1c..d8ff3f584 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -75,9 +75,15 @@ class Tensor : public TensorBase > typedef typename internal::traits::StorageKind StorageKind; typedef typename internal::traits::Index Index; typedef Scalar_ Scalar; - typedef typename internal::packet_traits::type PacketScalar; + typedef typename internal::packet_traits::type Packet; typedef typename NumTraits::Real RealScalar; typedef typename Base::CoeffReturnType CoeffReturnType; + typedef typename Base::PacketReturnType PacketReturnType; + + enum { + IsAligned = bool(EIGEN_ALIGN), + PacketAccess = true, + }; static const int Options = Options_; static const std::size_t NumIndices = NumIndices_; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index f1df827f9..e69ff6188 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -10,6 +10,9 @@ #ifndef EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H #define EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H +#ifdef EIGEN_USE_THREADS +#include +#endif namespace Eigen { @@ -28,7 +31,8 @@ namespace Eigen { */ namespace internal { -template +// Default strategy: the expressions are evaluated with a single cpu thread. +template::PacketAccess & TensorEvaluator::PacketAccess> struct TensorAssign { typedef typename Derived1::Index Index; @@ -38,13 +42,150 @@ struct TensorAssign TensorEvaluator evalDst(dst); TensorEvaluator evalSrc(src); const Index size = dst.size(); - for(Index i = 0; i < size; ++i) { + for (Index i = 0; i < size; ++i) { evalDst.coeffRef(i) = evalSrc.coeff(i); } } }; +template +struct TensorAssign +{ + typedef typename Derived1::Index Index; + EIGEN_DEVICE_FUNC + static inline void run(Derived1& dst, const Derived2& src) + { + TensorEvaluator evalDst(dst); + TensorEvaluator evalSrc(src); + const Index size = dst.size(); + + static const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + static const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + static const int PacketSize = unpacket_traits::PacketReturnType>::size; + static const int VectorizedSize = (size / PacketSize) * PacketSize; + + for (Index i = 0; i < VectorizedSize; i += PacketSize) { + evalDst.template writePacket(i, evalSrc.template packet(i)); + } + for (Index i = VectorizedSize; i < size; ++i) { + evalDst.coeffRef(i) = evalSrc.coeff(i); + } + } +}; + + + +// Multicore strategy: the index space is partitioned and each core is assigned to a partition +#ifdef EIGEN_USE_THREADS +template +struct EvalRange { + static void run(LhsEval& dst, const RhsEval& src, const Index first, const Index last) { + eigen_assert(last > first); + for (Index i = first; i < last; ++i) { + dst.coeffRef(i) = src.coeff(i); + } + } +}; + +template +struct EvalRange { + static void run(LhsEval& dst, const RhsEval& src, const Index first, const Index last) { + eigen_assert(last > first); + + Index i = first; + static const int PacketSize = unpacket_traits::size; + if (last - first > PacketSize) { + static const int LhsStoreMode = LhsEval::IsAligned ? Aligned : Unaligned; + static const int RhsLoadMode = RhsEval::IsAligned ? Aligned : Unaligned; + eigen_assert(first % PacketSize == 0); + Index lastPacket = last - (last % PacketSize); + for (; i < lastPacket; i += PacketSize) { + dst.template writePacket(i, src.template packet(i)); + } + } + + for (; i < last; ++i) { + dst.coeffRef(i) = src.coeff(i); + } + } +}; + +template +struct TensorAssignMultiThreaded +{ + typedef typename Derived1::Index Index; + static inline void run(Derived1& dst, const Derived2& src, const ThreadPoolDevice& device) + { + TensorEvaluator evalDst(dst); + TensorEvaluator evalSrc(src); + const Index size = dst.size(); + + static const bool Vectorizable = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess; + static const int PacketSize = Vectorizable ? unpacket_traits::PacketReturnType>::size : 1; + + int blocksz = static_cast(ceil(static_cast(size)/device.numThreads()) + PacketSize - 1); + const Index blocksize = std::max(PacketSize, (blocksz - (blocksz % PacketSize))); + const Index numblocks = size / blocksize; + + Index i = 0; + vector > results; + results.reserve(numblocks); + for (int i = 0; i < numblocks; ++i) { + results.push_back(std::async(std::launch::async, &EvalRange, TensorEvaluator, Index>::run, evalDst, evalSrc, i*blocksize, (i+1)*blocksize)); + } + + for (int i = 0; i < numblocks; ++i) { + results[i].get(); + } + + if (numblocks * blocksize < size) { + EvalRange, TensorEvaluator, Index>::run(evalDst, evalSrc, numblocks * blocksize, size); + } + } +}; +#endif + + +// GPU: the evaluation of the expressions is offloaded to a GPU. +#ifdef EIGEN_USE_GPU +template +__global__ void EigenMetaKernelNoCheck(LhsEvaluator evalDst, const RhsEvaluator evalSrc) { + const int index = blockIdx.x * blockDim.x + threadIdx.x; + evalDst.coeffRef(index) = evalSrc.coeff(index); +} +template +__global__ void EigenMetaKernelPeel(LhsEvaluator evalDst, const RhsEvaluator evalSrc, int peel_start_offset, int size) { + const int index = peel_start_offset + blockIdx.x * blockDim.x + threadIdx.x; + if (index < size) { + evalDst.coeffRef(index) = evalSrc.coeff(index); + } +} + +template +struct TensorAssignGpu +{ + typedef typename Derived1::Index Index; + static inline void run(Derived1& dst, const Derived2& src, const GpuDevice& device) + { + TensorEvaluator evalDst(dst); + TensorEvaluator evalSrc(src); + const Index size = dst.size(); + const int block_size = std::min(size, 32*32); + const int num_blocks = size / block_size; + EigenMetaKernelNoCheck, TensorEvaluator > <<>>(evalDst, evalSrc); + + const int remaining_items = size % block_size; + if (remaining_items > 0) { + const int peel_start_offset = num_blocks * block_size; + const int peel_block_size = std::min(size, 32); + const int peel_num_blocks = (remaining_items + peel_block_size - 1) / peel_block_size; + EigenMetaKernelPeel, TensorEvaluator > <<>>(evalDst, evalSrc, peel_start_offset, size); + } + } +}; +#endif + } // end namespace internal } // end namespace Eigen diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 9c7783aaf..fa1bd3498 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -28,6 +28,7 @@ class TensorBase typedef typename internal::traits::Scalar Scalar; typedef typename internal::traits::Index Index; typedef Scalar CoeffReturnType; + typedef typename internal::packet_traits::type PacketReturnType; Derived& setZero() { return setConstant(Scalar(0)); @@ -83,6 +84,17 @@ class TensorBase return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator-(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + template + TensorDevice device(const DeviceType& device) { + return TensorDevice(device, derived()); + } + protected: template friend class TensorBase; EIGEN_DEVICE_FUNC diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h new file mode 100644 index 000000000..71890e187 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h @@ -0,0 +1,83 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H +#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H + +namespace Eigen { + +/** \class TensorDevice + * \ingroup CXX11_Tensor_Module + * + * \brief Pseudo expression providing an operator = that will evaluate its argument + * on the specified computing 'device' (GPU, thread pool, ...) + * + * Example: + * C.device(EIGEN_GPU) = A + B; + * + * Todo: thread pools. + * Todo: operator +=, -=, *= and so on. + */ + +template class TensorDevice { + public: + TensorDevice(const DeviceType& device, ExpressionType& expression) : m_device(device), m_expression(expression) {} + + template + EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) { + internal::TensorAssign::run(m_expression, other); + return *this; + } + + protected: + const DeviceType& m_device; + ExpressionType& m_expression; +}; + + +#ifdef EIGEN_USE_THREADS +template class TensorDevice { + public: + TensorDevice(const ThreadPoolDevice& device, ExpressionType& expression) : m_device(device), m_expression(expression) {} + + template + EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) { + internal::TensorAssignMultiThreaded::run(m_expression, other, m_device); + return *this; + } + + protected: + const ThreadPoolDevice& m_device; + ExpressionType& m_expression; +}; +#endif + + +#ifdef EIGEN_USE_GPU +template class TensorDevice +{ + public: + TensorDevice(const GpuDevice& device, ExpressionType& expression) : m_device(device), m_expression(expression) {} + + template + EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) { + internal::TensorAssignGpu::run(m_expression, other, m_device); + return *this; + } + + protected: + const GpuDevice& m_device; + ExpressionType& m_expression; +}; +#endif + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h new file mode 100644 index 000000000..ded6ca604 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -0,0 +1,56 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_DEVICE_TYPE_H +#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_TYPE_H + + +namespace Eigen { + +// Default device for the machine (typically a single cpu core) +struct DefaultDevice { +}; + + +// Multiple cpu cores +// We should really use a thread pool here but first we need to find a portable thread pool library. +#ifdef EIGEN_USE_THREADS +struct ThreadPoolDevice { + ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : /*pool_(pool), */num_threads_(num_cores) { } + size_t numThreads() const { return num_threads_; } + /*ThreadPool* threadPool() const { return pool_; }*/ + + private: + // todo: NUMA, ... + size_t num_threads_; + /*ThreadPool* pool_;*/ +}; +#endif + + +// GPU offloading +#ifdef EIGEN_USE_GPU +struct GpuDevice { + // todo: support for multiple gpu; + GpuDevice() { + cudaStreamCreate(&stream_); + } + ~GpuDevice() { + cudaStreamDestroy(stream_); + } + const cudaStream_t& stream() const { return stream_; } + + private: + cudaStream_t stream_; +}; +#endif + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_TYPE_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index bd3bd5aca..43e9d6550 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -79,16 +79,16 @@ struct Sizes : internal::numeric_list { Sizes() { } template - explicit Sizes(const array&/* indices*/) { + explicit Sizes(const array& /*indices*/) { // todo: add assertion } #ifdef EIGEN_HAS_VARIADIC_TEMPLATES - explicit Sizes(std::initializer_list/* l*/) { + explicit Sizes(std::initializer_list /*l*/) { // todo: add assertion } #endif - template Sizes& operator = (const T&/* other*/) { + template Sizes& operator = (const T& /*other*/) { // add assertion failure if the size of other is different return *this; } @@ -119,7 +119,7 @@ template ::value; - static const size_t TotalSize() { + static size_t TotalSize() { return internal::arg_prod::value; } @@ -181,14 +181,11 @@ template struct DSizes : array { typedef array Base; - size_t TotalSize() const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() const { return internal::array_prod(*static_cast(this)); } DSizes() { } -#ifdef EIGEN_HAS_VARIADIC_TEMPLATES - // explicit DSizes(std::initializer_list l) : Base(l) { } -#endif explicit DSizes(const array& a) : Base(a) { } DSizes& operator = (const array& other) { @@ -203,7 +200,6 @@ struct DSizes : array { size_t IndexOfRowMajor(const array& indices) const { return internal::tensor_index_linearization_helper::run(indices, *static_cast(this)); } - }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index b0dbca041..3ce924dc3 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -29,32 +29,38 @@ struct TensorEvaluator { typedef typename Derived::Index Index; typedef typename Derived::Scalar Scalar; - typedef typename Derived::Scalar& CoeffReturnType; + typedef typename Derived::Packet Packet; + typedef typename Derived::Scalar CoeffReturnType; + typedef typename Derived::Packet PacketReturnType; + + enum { + IsAligned = Derived::IsAligned, + PacketAccess = Derived::PacketAccess, + }; TensorEvaluator(Derived& m) : m_data(const_cast(m.data())) { } - CoeffReturnType coeff(Index index) const { + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_data[index]; } - Scalar& coeffRef(Index index) { + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return m_data[index]; } - // to do: vectorized evaluation. - /* template + template PacketReturnType packet(Index index) const { - return ploadt(m_data + index); + return internal::ploadt(m_data + index); } - template - void writePacket(Index index, const PacketScalar& x) + template + void writePacket(Index index, const Packet& x) { - return pstoret(const_cast(m_data) + index, x); - }*/ + return internal::pstoret(m_data + index, x); + } protected: Scalar* m_data; @@ -70,6 +76,11 @@ struct TensorEvaluator > { typedef TensorCwiseUnaryOp XprType; + enum { + IsAligned = TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess & internal::functor_traits::PacketAccess, + }; + TensorEvaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) @@ -77,12 +88,19 @@ struct TensorEvaluator > typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_functor(m_argImpl.coeff(index)); } + template + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + { + return m_functor.packetOp(m_argImpl.template packet(index)); + } + private: const UnaryOp m_functor; TensorEvaluator m_argImpl; @@ -96,6 +114,12 @@ struct TensorEvaluator XprType; + enum { + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess & + internal::functor_traits::PacketAccess, + }; + TensorEvaluator(const XprType& op) : m_functor(op.functor()), m_leftImpl(op.lhsExpression()), @@ -104,11 +128,17 @@ struct TensorEvaluator + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + { + return m_functor.packetOp(m_leftImpl.template packet(index), m_rightImpl.template packet(index)); + } private: const BinaryOp m_functor; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h index aa875dc31..e32077f6e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h @@ -33,6 +33,9 @@ struct traits > typedef typename result_of< UnaryOp(typename XprType::Scalar) >::type Scalar; + typedef typename result_of< + UnaryOp(typename XprType::Packet) + >::type Packet; typedef typename XprType::Nested XprTypeNested; typedef typename remove_reference::type _XprTypeNested; }; @@ -57,14 +60,16 @@ template class TensorCwiseUnaryOp : public TensorBase > { public: - typedef typename Eigen::internal::traits::Scalar Scalar; - typedef typename Eigen::NumTraits::Real RealScalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename Eigen::internal::nested::type Nested; - typedef typename Eigen::internal::traits::StorageKind StorageKind; - typedef typename Eigen::internal::traits::Index Index; + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; - inline TensorCwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) : m_xpr(xpr), m_functor(func) {} EIGEN_DEVICE_FUNC @@ -92,6 +97,7 @@ struct traits > typename RhsXprType::Scalar ) >::type Scalar; + typedef typename internal::packet_traits::type Packet; typedef typename promote_storage_type::StorageKind, typename traits::StorageKind>::ret StorageKind; typedef typename promote_index_type::Index, @@ -123,14 +129,17 @@ class TensorCwiseBinaryOp : public TensorBase::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; typedef typename Eigen::NumTraits::Real RealScalar; typedef typename internal::promote_storage_type::ret CoeffReturnType; + typedef typename internal::promote_storage_type::ret PacketReturnType; typedef typename Eigen::internal::nested::type Nested; typedef typename Eigen::internal::traits::StorageKind StorageKind; typedef typename Eigen::internal::traits::Index Index; - inline TensorCwiseBinaryOp(const LhsXprType& lhs, const RhsXprType& rhs, const BinaryOp& func = BinaryOp()) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseBinaryOp(const LhsXprType& lhs, const RhsXprType& rhs, const BinaryOp& func = BinaryOp()) : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_functor(func) {} EIGEN_DEVICE_FUNC diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h index 953880123..dcc7ccd65 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h @@ -33,11 +33,17 @@ class TensorFixedSize : public TensorBase::StorageKind StorageKind; typedef typename internal::traits::Index Index; typedef Scalar_ Scalar; - typedef typename internal::packet_traits::type PacketScalar; + typedef typename internal::packet_traits::type Packet; typedef typename NumTraits::Real RealScalar; typedef typename Base::CoeffReturnType CoeffReturnType; - static const int Options = Options_; + static const int Options = Options_; + + enum { + IsAligned = bool(EIGEN_ALIGN), + PacketAccess = true, + }; + typedef Dimensions_ Dimensions; static const std::size_t NumIndices = Dimensions::count; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index e8a2125c4..09b0fe66d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -14,12 +14,14 @@ namespace Eigen { template class Tensor; template class TensorFixedSize; -template class TensorMap; +template class TensorMap; template class TensorBase; template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; +template class TensorDevice; + // Move to internal? template struct TensorEvaluator; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index bb0b39c5a..3fc9c5335 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -22,16 +22,16 @@ template class Strid * */ -template class TensorMap : public TensorBase > +template class TensorMap : public TensorBase > { public: - typedef TensorMap Self; + typedef TensorMap Self; typedef typename PlainObjectType::Base Base; typedef typename Eigen::internal::nested::type Nested; typedef typename internal::traits::StorageKind StorageKind; typedef typename internal::traits::Index Index; typedef typename internal::traits::Scalar Scalar; - typedef typename internal::packet_traits::type PacketScalar; + typedef typename internal::packet_traits::type Packet; typedef typename NumTraits::Real RealScalar; typedef typename Base::CoeffReturnType CoeffReturnType; @@ -43,13 +43,12 @@ template class TensorMap : public TensorBase({{firstDimension}})) { @@ -65,7 +64,7 @@ template class TensorMap : public TensorBase& dimensions) + inline TensorMap(PointerArgType dataPtr, const array& dimensions) : m_data(dataPtr), m_dimensions(dimensions) { } @@ -80,12 +79,97 @@ template class TensorMap : public TensorBase& indices) const + { + // eigen_assert(checkIndexRange(indices)); + if (PlainObjectType::Options&RowMajor) { + const Index index = m_dimensions.IndexOfRowMajor(indices); + return m_data[index]; + } else { + const Index index = m_dimensions.IndexOfColMajor(indices); + return m_data[index]; + } + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) const + { + static_assert(sizeof...(otherIndices) + 1 == PlainObjectType::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + if (PlainObjectType::Options&RowMajor) { + const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); + return m_data[index]; + } else { + const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); + return m_data[index]; + } + } +#else EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(Index index) const { eigen_internal_assert(index >= 0 && index < size()); return m_data[index]; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1) const + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i1 + i0 * m_dimensions[0]; + return m_data[index]; + } else { + const Index index = i0 + i1 * m_dimensions[0]; + return m_data[index]; + } + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2) const + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i2 + m_dimensions[1] * (i1 + m_dimensions[0] * i0); + return m_data[index]; + } else { + const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * i2); + return m_data[index]; + } + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2, Index i3) const + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0)); + return m_data[index]; + } else { + const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * i3)); + return m_data[index]; + } + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar& operator()(Index i0, Index i1, Index i2, Index i3, Index i4) const + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i4 + m_dimensions[4] * (i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0))); + return m_data[index]; + } else { + const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * (i3 + m_dimensions[3] * i4))); + return m_data[index]; + } + } +#endif + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(const array& indices) + { + // eigen_assert(checkIndexRange(indices)); + if (PlainObjectType::Options&RowMajor) { + const Index index = m_dimensions.IndexOfRowMajor(indices); + return m_data[index]; + } else { + const Index index = m_dimensions.IndexOfColMajor(indices); + return m_data[index]; + } + } #ifdef EIGEN_HAS_VARIADIC_TEMPLATES template EIGEN_DEVICE_FUNC @@ -100,8 +184,60 @@ template class TensorMap : public TensorBase= 0 && index < size()); + return m_data[index]; + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1) + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i1 + i0 * m_dimensions[0]; + return m_data[index]; + } else { + const Index index = i0 + i1 * m_dimensions[0]; + return m_data[index]; + } + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2) + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i2 + m_dimensions[1] * (i1 + m_dimensions[0] * i0); + return m_data[index]; + } else { + const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * i2); + return m_data[index]; + } + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2, Index i3) + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0)); + return m_data[index]; + } else { + const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * i3)); + return m_data[index]; + } + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& operator()(Index i0, Index i1, Index i2, Index i3, Index i4) + { + if (PlainObjectType::Options&RowMajor) { + const Index index = i4 + m_dimensions[4] * (i3 + m_dimensions[3] * (i2 + m_dimensions[2] * (i1 + m_dimensions[1] * i0))); + return m_data[index]; + } else { + const Index index = i0 + m_dimensions[0] * (i1 + m_dimensions[1] * (i2 + m_dimensions[2] * (i3 + m_dimensions[3] * i4))); + return m_data[index]; + } + } #endif + template EIGEN_DEVICE_FUNC Self& operator=(const OtherDerived& other) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index efcb39559..64098343e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -72,9 +72,6 @@ class TensorStorage TensorStorage() { } TensorStorage(const TensorStorage& other) : Base_(other) { } -#ifdef EIGEN_HAVE_RVALUE_REFERENCES - // TensorStorage(TensorStorage&&) = default; -#endif TensorStorage(internal::constructor_without_unaligned_array_assert) : Base_(internal::constructor_without_unaligned_array_assert()) {} TensorStorage(DenseIndex size, const array& dimensions) : Base_(size, dimensions) {} @@ -111,22 +108,6 @@ class TensorStorage(m_data, internal::array_prod(m_dimensions)); } void swap(Self_& other) { std::swap(m_data,other.m_data); std::swap(m_dimensions,other.m_dimensions); } diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 31583d3ca..abc3375e5 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -104,4 +104,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_assign "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") + ei_add_test(cxx11_tensor_device "-std=c++0x") +# ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") + ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_device.cpp b/unsupported/test/cxx11_tensor_device.cpp new file mode 100644 index 000000000..9eb1d0420 --- /dev/null +++ b/unsupported/test/cxx11_tensor_device.cpp @@ -0,0 +1,126 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#define EIGEN_TEST_NO_LONGDOUBLE +#define EIGEN_TEST_NO_COMPLEX +#define EIGEN_TEST_FUNC cxx11_tensor_device +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int +#define EIGEN_USE_GPU + + +#include "main.h" +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +// Context for evaluation on cpu +struct CPUContext { + CPUContext(const Eigen::Tensor& in1, Eigen::Tensor& in2, Eigen::Tensor& out) : in1_(in1), in2_(in2), out_(out) { } + + const Eigen::Tensor& in1() const { return in1_; } + const Eigen::Tensor& in2() const { return in2_; } + Eigen::TensorDevice, Eigen::DefaultDevice> out() { return TensorDevice, Eigen::DefaultDevice>(cpu_device_, out_); } + + private: + const Eigen::Tensor& in1_; + const Eigen::Tensor& in2_; + Eigen::Tensor& out_; + + Eigen::DefaultDevice cpu_device_; +}; + + +// Context for evaluation on GPU +struct GPUContext { + GPUContext(const Eigen::TensorMap >& in1, Eigen::TensorMap >& in2, Eigen::TensorMap >& out) : in1_(in1), in2_(in2), out_(out) { } + + const Eigen::TensorMap >& in1() const { return in1_; } + const Eigen::TensorMap >& in2() const { return in2_; } + Eigen::TensorDevice >, Eigen::GpuDevice> out() { return TensorDevice >, Eigen::GpuDevice>(gpu_device_, out_); } + + private: + const Eigen::TensorMap >& in1_; + const Eigen::TensorMap >& in2_; + Eigen::TensorMap >& out_; + Eigen::GpuDevice gpu_device_; +}; + + +// The actual expression to evaluate +template +static void test_contextual_eval(Context* context) +{ + context->out() = context->in1() + context->in2() * 3.14f; +} + +static void test_cpu() { + Eigen::Tensor in1(Eigen::array(2,3,7)); + Eigen::Tensor in2(Eigen::array(2,3,7)); + Eigen::Tensor out(Eigen::array(2,3,7)); + + in1.setRandom(); + in2.setRandom(); + CPUContext context(in1, in2, out); + test_contextual_eval(&context); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f); + } + } + } +} + +static void test_gpu() { + Eigen::Tensor in1(Eigen::array(2,3,7)); + Eigen::Tensor in2(Eigen::array(2,3,7)); + Eigen::Tensor out(Eigen::array(2,3,7)); + in1.setRandom(); + in2.setRandom(); + + std::size_t in1_bytes = in1.size() * sizeof(float); + std::size_t in2_bytes = in2.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(float); + + float* d_in1; + float* d_in2; + float* d_out; + cudaMalloc((void**)(&d_in1), in1_bytes); + cudaMalloc((void**)(&d_in2), in2_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice); + + Eigen::TensorMap > gpu_in1(d_in1, Eigen::array(2,3,7)); + Eigen::TensorMap > gpu_in2(d_in2, Eigen::array(2,3,7)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(2,3,7)); + + GPUContext context(gpu_in1, gpu_in2, gpu_out); + test_contextual_eval(&context); + + cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f); + } + } + } +} + + + +void test_cxx11_tensor_device() +{ + CALL_SUBTEST(test_cpu()); + CALL_SUBTEST(test_gpu()); +} diff --git a/unsupported/test/cxx11_tensor_fixed_size.cpp b/unsupported/test/cxx11_tensor_fixed_size.cpp index c1d74d881..214f6951d 100644 --- a/unsupported/test/cxx11_tensor_fixed_size.cpp +++ b/unsupported/test/cxx11_tensor_fixed_size.cpp @@ -159,9 +159,37 @@ static void test_3d() } +static void test_array() +{ + TensorFixedSize > mat1; + float val = 0.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + mat1(array(i,j,k)) = val; + val += 1.0; + } + } + } + + TensorFixedSize > mat3; + mat3 = mat1.cwisePow(3.5f); + + val = 0.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat3(array(i,j,k)), powf(val, 3.5f)); + val += 1.0; + } + } + } +} + void test_cxx11_tensor_fixed_size() { CALL_SUBTEST(test_1d()); CALL_SUBTEST(test_2d()); CALL_SUBTEST(test_3d()); + CALL_SUBTEST(test_array()); } diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp new file mode 100644 index 000000000..c9de71da3 --- /dev/null +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -0,0 +1,37 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#define EIGEN_USE_THREADS + + +#include "main.h" +#include + +using Eigen::Tensor; + +void test_cxx11_tensor_thread_pool() +{ + Eigen::Tensor in1(Eigen::array(2,3,7)); + Eigen::Tensor in2(Eigen::array(2,3,7)); + Eigen::Tensor out(Eigen::array(2,3,7)); + + in1.setRandom(); + in2.setRandom(); + + Eigen::ThreadPoolDevice thread_pool_device(3); + out.device(thread_pool_device) = in1 + in2 * 3.14; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f); + } + } + } +} From 736267cf6b17832a571acf7e34ca07c7f55907ee Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 22 May 2014 16:22:35 -0700 Subject: [PATCH 0500/1103] Added support for additional tensor operations: * comparison (<, <=, ==, !=, ...) * selection * nullary ops such as random or constant generation * misc unary ops such as log(), exp(), or a user defined unaryExpr() Cleaned up the code a little. --- .../Eigen/CXX11/src/Tensor/TensorBase.h | 139 ++++++++++++++++-- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 84 +++++++++++ .../Eigen/CXX11/src/Tensor/TensorExpr.h | 109 ++++++++++++++ .../src/Tensor/TensorForwardDeclarations.h | 2 + .../Eigen/CXX11/src/Tensor/TensorMap.h | 36 +++-- 5 files changed, 339 insertions(+), 31 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index fa1bd3498..8a88ba806 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -33,21 +33,25 @@ class TensorBase Derived& setZero() { return setConstant(Scalar(0)); } - Derived& setConstant(const Scalar& val) { - Scalar* data = derived().data(); - for (int i = 0; i < derived().size(); ++i) { - data[i] = val; - } - return derived(); + return derived() = constant(val); + } + Derived& setRandom() { + return derived() = random(); } - Derived& setRandom() { - Scalar* data = derived().data(); - for (int i = 0; i < derived().size(); ++i) { - data[i] = internal::random_default_impl::run(); - } - return derived(); + // Nullary operators + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> + constant(const Scalar& value) const { + return TensorCwiseNullaryOp, const Derived> + (internal::scalar_constant_op(value)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> + random() const { + return TensorCwiseNullaryOp, const Derived>(); } // Coefficient-wise unary operators @@ -57,15 +61,31 @@ class TensorBase EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - cwiseSqrt() const { return derived(); } + sqrt() const { return derived(); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + square() const { return derived(); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + inverse() const { return derived(); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + exp() const { return derived(); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + log() const { return derived(); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - cwiseAbs() const { return derived(); } + abs() const { return derived(); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - cwisePow(Scalar exponent) const { + pow(Scalar exponent) const { return TensorCwiseUnaryOp, const Derived> (derived(), internal::scalar_pow_op(exponent)); } @@ -77,6 +97,30 @@ class TensorBase (derived(), internal::scalar_multiple_op(scale)); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const TensorCwiseNullaryOp, const Derived> > + cwiseMax(Scalar threshold) const { + return cwiseMax(constant(threshold)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const TensorCwiseNullaryOp, const Derived> > + cwiseMin(Scalar threshold) const { + return cwiseMin(constant(threshold)); + } + + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp + unaryExpr(const CustomUnaryOp& func) const { + return TensorCwiseUnaryOp(derived(), func); + } + + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + cast() const { + return derived(); + } + // Coefficient-wise binary operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> @@ -90,6 +134,71 @@ class TensorBase return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator*(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator/(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + cwiseMax(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + cwiseMin(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + // Comparisons and tests. + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator<(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator<=(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator>(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator>=(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator==(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp, const Derived, const OtherDerived> + operator!=(const OtherDerived& other) const { + return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + // Coefficient-wise ternary operators. + template + inline const TensorSelectOp + select(const ThenDerived& thenTensor, const ElseDerived& elseTensor) const{ + return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); + } + + // Select the device on which to evaluate the expression. template TensorDevice device(const DeviceType& device) { return TensorDevice(device, derived()); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index 3ce924dc3..e0c0863b7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -68,6 +68,42 @@ struct TensorEvaluator +// -------------------- CwiseNullaryOp -------------------- + +template +struct TensorEvaluator > +{ + typedef TensorCwiseNullaryOp XprType; + + enum { + IsAligned = true, + PacketAccess = internal::functor_traits::PacketAccess, + }; + + TensorEvaluator(const XprType& op) + : m_functor(op.functor()) + { } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const + { + return m_functor(index); + } + + template + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + { + return m_functor.packetOp(index); + } + + private: + const NullaryOp m_functor; +}; + + // -------------------- CwiseUnaryOp -------------------- @@ -146,6 +182,54 @@ struct TensorEvaluator m_rightImpl; }; + +// -------------------- SelectOp -------------------- + +template +struct TensorEvaluator > +{ + typedef TensorSelectOp XprType; + + enum { + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess/* & + TensorEvaluator::PacketAccess*/, + }; + + TensorEvaluator(const XprType& op) + : m_condImpl(op.ifExpression()), + m_thenImpl(op.thenExpression()), + m_elseImpl(op.elseExpression()) + { } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const + { + return m_condImpl.coeff(index) ? m_thenImpl.coeff(index) : m_elseImpl.coeff(index); + } + template + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + { + static const int PacketSize = internal::unpacket_traits::size; + internal::Selector select; + for (Index i = 0; i < PacketSize; ++i) { + select.select[i] = m_condImpl.coeff(index+i); + } + return internal::pblend(select, + m_thenImpl.template packet(index), + m_elseImpl.template packet(index)); + } + + private: + TensorEvaluator m_condImpl; + TensorEvaluator m_thenImpl; + TensorEvaluator m_elseImpl; +}; + + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_EVALUATOR_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h index e32077f6e..94cfae05c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h @@ -17,6 +17,9 @@ namespace Eigen { * * \brief Tensor expression classes. * + * The TensorCwiseNullaryOp class applies a nullary operators to an expression. This + * is typically used to generate constants. + * * The TensorCwiseUnaryOp class represents an expression where a unary operator * (e.g. cwiseSqrt) is applied to an expression. * @@ -24,6 +27,46 @@ namespace Eigen { * (e.g. addition) is applied to a lhs and a rhs expression. * */ +namespace internal { +template +struct traits > + : traits +{ + typedef typename PlainObjectType::Packet Packet; + typedef typename PlainObjectType::Scalar Scalar; + typedef typename PlainObjectType::Nested XprTypeNested; + typedef typename remove_reference::type _XprTypeNested; +}; + +} // end namespace internal + + + +template +class TensorCwiseNullaryOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename PlainObjectType::CoeffReturnType CoeffReturnType; + typedef typename PlainObjectType::PacketReturnType PacketReturnType; + typedef TensorCwiseNullaryOp Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseNullaryOp(const NullaryOp& func = NullaryOp()) + : m_functor(func) {} + + EIGEN_DEVICE_FUNC + const NullaryOp& functor() const { return m_functor; } + + protected: + // todo: add tensor dimension to be able to do some sanity checks + const NullaryOp m_functor; +}; + + namespace internal { template @@ -160,6 +203,72 @@ class TensorCwiseBinaryOp : public TensorBase +struct traits > + : traits +{ + typedef typename traits::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename promote_storage_type::StorageKind, + typename traits::StorageKind>::ret StorageKind; + typedef typename promote_index_type::Index, + typename traits::Index>::type Index; + typedef typename IfXprType::Nested IfNested; + typedef typename ThenXprType::Nested ThenNested; + typedef typename ElseXprType::Nested ElseNested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorSelectOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorSelectOp type; +}; + +} // end namespace internal + + +template +class TensorSelectOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename internal::promote_storage_type::ret CoeffReturnType; + typedef typename internal::promote_storage_type::ret PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + TensorSelectOp(const IfXprType& a_condition, + const ThenXprType& a_then, + const ElseXprType& a_else) + : m_condition(a_condition), m_then(a_then), m_else(a_else) + { } + + const IfXprType& ifExpression() const { return m_condition; } + + const ThenXprType& thenExpression() const { return m_then; } + + const ElseXprType& elseExpression() const { return m_else; } + + protected: + typename IfXprType::Nested m_condition; + typename ThenXprType::Nested m_then; + typename ElseXprType::Nested m_else; +}; + + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_EXPR_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 09b0fe66d..03ac8d516 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -17,8 +17,10 @@ template class TensorFi template class TensorMap; template class TensorBase; +template class TensorCwiseNullaryOp; template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; +template class TensorSelectOp; template class TensorDevice; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index 3fc9c5335..3a2ff5b30 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -45,33 +45,37 @@ template class TensorMap : public Tensor static const int Options = Options_; + static const std::size_t NumIndices = PlainObjectType::NumIndices; + typedef typename PlainObjectType::Dimensions Dimensions; + + enum { IsAligned = bool(EIGEN_ALIGN) && ((int(Options_)&Aligned)==Aligned), PacketAccess = true, }; EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions(array({{firstDimension}})) { + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions(array(firstDimension)) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. - EIGEN_STATIC_ASSERT(1 == PlainObjectType::NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + EIGEN_STATIC_ASSERT(1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) } #ifdef EIGEN_HAS_VARIADIC_TEMPLATES template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions(array({{firstDimension, otherDimensions...}})) { + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions(array({{firstDimension, otherDimensions...}})) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. - EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == PlainObjectType::NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) } #endif - inline TensorMap(PointerArgType dataPtr, const array& dimensions) + inline TensorMap(PointerArgType dataPtr, const array& dimensions) : m_data(dataPtr), m_dimensions(dimensions) { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index dimension(Index n) const { return m_dimensions[n]; } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const typename PlainObjectType::Dimensions& dimensions() const { return m_dimensions; } + EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index size() const { return m_dimensions.TotalSize(); } EIGEN_DEVICE_FUNC @@ -80,7 +84,7 @@ template class TensorMap : public Tensor EIGEN_STRONG_INLINE const Scalar* data() const { return m_data; } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar& operator()(const array& indices) const + EIGEN_STRONG_INLINE const Scalar& operator()(const array& indices) const { // eigen_assert(checkIndexRange(indices)); if (PlainObjectType::Options&RowMajor) { @@ -96,12 +100,12 @@ template class TensorMap : public Tensor template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) const { - static_assert(sizeof...(otherIndices) + 1 == PlainObjectType::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + static_assert(sizeof...(otherIndices) + 1 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); if (PlainObjectType::Options&RowMajor) { - const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); + const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); return m_data[index]; } else { - const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); + const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); return m_data[index]; } } @@ -159,7 +163,7 @@ template class TensorMap : public Tensor #endif EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Scalar& operator()(const array& indices) + EIGEN_STRONG_INLINE Scalar& operator()(const array& indices) { // eigen_assert(checkIndexRange(indices)); if (PlainObjectType::Options&RowMajor) { @@ -175,12 +179,12 @@ template class TensorMap : public Tensor template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) { - static_assert(sizeof...(otherIndices) + 1 == PlainObjectType::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + static_assert(sizeof...(otherIndices) + 1 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); if (PlainObjectType::Options&RowMajor) { - const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); + const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); return m_data[index]; } else { - const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); + const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); return m_data[index]; } } @@ -247,8 +251,8 @@ template class TensorMap : public Tensor } private: - typename PlainObjectType::Scalar* m_data; - typename PlainObjectType::Dimensions m_dimensions; + Scalar* m_data; + Dimensions m_dimensions; }; } // end namespace Eigen From eb56461ac2c51374aa7c21a7c62e3c695f4da76b Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sat, 31 May 2014 23:05:18 +0100 Subject: [PATCH 0501/1103] Fix doc'n of FullPivLU re permutation matrices (bug #815). --- Eigen/src/LU/FullPivLU.h | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index 44699b763..971b9da1d 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -20,10 +20,11 @@ namespace Eigen { * * \param MatrixType the type of the matrix of which we are computing the LU decomposition * - * This class represents a LU decomposition of any matrix, with complete pivoting: the matrix A - * is decomposed as A = PLUQ where L is unit-lower-triangular, U is upper-triangular, and P and Q - * are permutation matrices. This is a rank-revealing LU decomposition. The eigenvalues (diagonal - * coefficients) of U are sorted in such a way that any zeros are at the end. + * This class represents a LU decomposition of any matrix, with complete pivoting: the matrix A is + * decomposed as \f$ A = P^{-1} L U Q^{-1} \f$ where L is unit-lower-triangular, U is + * upper-triangular, and P and Q are permutation matrices. This is a rank-revealing LU + * decomposition. The eigenvalues (diagonal coefficients) of U are sorted in such a way that any + * zeros are at the end. * * This decomposition provides the generic approach to solving systems of linear equations, computing * the rank, invertibility, inverse, kernel, and determinant. @@ -511,8 +512,8 @@ typename internal::traits::Scalar FullPivLU::determinant } /** \returns the matrix represented by the decomposition, - * i.e., it returns the product: P^{-1} L U Q^{-1}. - * This function is provided for debug purpose. */ + * i.e., it returns the product: \f$ P^{-1} L U Q^{-1} \f$. + * This function is provided for debug purposes. */ template MatrixType FullPivLU::reconstructedMatrix() const { From 789674809fb039c7137f52bea36c8ac71bcc9b81 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Mon, 2 Jun 2014 11:42:42 +0100 Subject: [PATCH 0502/1103] Fix test: EigenSolver on 1x1 matrix with NaN sets info to NumericalIssue. This was changed in 3c66bb136bf2adcb9d73d3d66850a8b907bc9264 . --- test/eigensolver_generic.cpp | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/test/eigensolver_generic.cpp b/test/eigensolver_generic.cpp index 91383b5cf..92d33f66a 100644 --- a/test/eigensolver_generic.cpp +++ b/test/eigensolver_generic.cpp @@ -114,10 +114,9 @@ void test_eigensolver_generic() CALL_SUBTEST_2( { MatrixXd A(1,1); - A(0,0) = std::sqrt(-1.); + A(0,0) = std::sqrt(-1.); // is Not-a-Number Eigen::EigenSolver solver(A); - MatrixXd V(1, 1); - V(0,0) = solver.eigenvectors()(0,0).real(); + VERIFY_IS_EQUAL(solver.info(), NumericalIssue); } ); From 0f1e321dd4a1dec90f25aa248f77dca5e353f394 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 4 Jun 2014 11:58:01 +0200 Subject: [PATCH 0503/1103] Fic bug #819: include path of details.h --- Eigen/src/StlSupport/StdDeque.h | 2 +- Eigen/src/StlSupport/StdList.h | 2 +- Eigen/src/StlSupport/StdVector.h | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/StlSupport/StdDeque.h b/Eigen/src/StlSupport/StdDeque.h index 4ee8e5c10..aaf66330b 100644 --- a/Eigen/src/StlSupport/StdDeque.h +++ b/Eigen/src/StlSupport/StdDeque.h @@ -11,7 +11,7 @@ #ifndef EIGEN_STDDEQUE_H #define EIGEN_STDDEQUE_H -#include "Eigen/src/StlSupport/details.h" +#include "details.h" // Define the explicit instantiation (e.g. necessary for the Intel compiler) #if defined(__INTEL_COMPILER) || defined(__GNUC__) diff --git a/Eigen/src/StlSupport/StdList.h b/Eigen/src/StlSupport/StdList.h index 627381ece..3c742430c 100644 --- a/Eigen/src/StlSupport/StdList.h +++ b/Eigen/src/StlSupport/StdList.h @@ -10,7 +10,7 @@ #ifndef EIGEN_STDLIST_H #define EIGEN_STDLIST_H -#include "Eigen/src/StlSupport/details.h" +#include "details.h" // Define the explicit instantiation (e.g. necessary for the Intel compiler) #if defined(__INTEL_COMPILER) || defined(__GNUC__) diff --git a/Eigen/src/StlSupport/StdVector.h b/Eigen/src/StlSupport/StdVector.h index 40a9abefa..611664a2e 100644 --- a/Eigen/src/StlSupport/StdVector.h +++ b/Eigen/src/StlSupport/StdVector.h @@ -11,7 +11,7 @@ #ifndef EIGEN_STDVECTOR_H #define EIGEN_STDVECTOR_H -#include "Eigen/src/StlSupport/details.h" +#include "details.h" /** * This section contains a convenience MACRO which allows an easy specialization of From 6fa6cdd2b988da98cbdd2b1a5fd2fd3b9d56a4b1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 4 Jun 2014 09:21:48 -0700 Subject: [PATCH 0504/1103] Added support for tensor contractions Updated expression evaluation mechanism to also compute the size of the tensor result Misc fixes and improvements. --- unsupported/Eigen/CXX11/Tensor | 1 + .../CXX11/src/Core/util/EmulateCXX11Meta.h | 2 + unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 38 ++- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 5 +- .../Eigen/CXX11/src/Tensor/TensorBase.h | 50 ++-- .../CXX11/src/Tensor/TensorContraction.h | 229 ++++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorDevice.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 15 +- .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 29 ++- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 44 +++- .../Eigen/CXX11/src/Tensor/TensorExpr.h | 36 +-- .../Eigen/CXX11/src/Tensor/TensorFixedSize.h | 2 +- .../src/Tensor/TensorForwardDeclarations.h | 2 + .../Eigen/CXX11/src/Tensor/TensorStorage.h | 11 +- 14 files changed, 370 insertions(+), 96 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 323d9edff..d4e8d3a15 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -39,6 +39,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h" diff --git a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h index ab869177c..636063f9e 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h @@ -23,6 +23,8 @@ template class array { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T& operator[] (size_t index) const { return values[index]; } + static const std::size_t size = n; + T values[n]; EIGEN_DEVICE_FUNC diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index d8ff3f584..e034f8c03 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -81,7 +81,7 @@ class Tensor : public TensorBase > typedef typename Base::PacketReturnType PacketReturnType; enum { - IsAligned = bool(EIGEN_ALIGN), + IsAligned = bool(EIGEN_ALIGN) & !(Options_&DontAlign), PacketAccess = true, }; @@ -94,11 +94,11 @@ class Tensor : public TensorBase > TensorStorage m_storage; public: - EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return m_storage.dimensions()[n]; } - EIGEN_STRONG_INLINE const DSizes& dimensions() const { return m_storage.dimensions(); } - EIGEN_STRONG_INLINE Index size() const { return m_storage.size(); } - EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } - EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return m_storage.dimensions()[n]; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const DSizes& dimensions() const { return m_storage.dimensions(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index size() const { return m_storage.size(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } // This makes EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED // work, because that uses base().coeffRef() - and we don't yet @@ -116,13 +116,13 @@ class Tensor : public TensorBase > } #endif - inline const Scalar& coeff(const array& indices) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(const array& indices) const { eigen_internal_assert(checkIndexRange(indices)); return m_storage.data()[linearizedIndex(indices)]; } - inline const Scalar& coeff(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const { eigen_internal_assert(index >= 0 && index < size()); return m_storage.data()[index]; @@ -138,13 +138,13 @@ class Tensor : public TensorBase > } #endif - inline Scalar& coeffRef(const array& indices) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(const array& indices) { eigen_internal_assert(checkIndexRange(indices)); return m_storage.data()[linearizedIndex(indices)]; } - inline Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) { eigen_internal_assert(index >= 0 && index < size()); return m_storage.data()[index]; @@ -160,19 +160,19 @@ class Tensor : public TensorBase > } #endif - inline const Scalar& operator()(const array& indices) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(const array& indices) const { eigen_assert(checkIndexRange(indices)); return coeff(indices); } - inline const Scalar& operator()(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator()(Index index) const { eigen_internal_assert(index >= 0 && index < size()); return coeff(index); } - inline const Scalar& operator[](Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar& operator[](Index index) const { // The bracket operator is only for vectors, use the parenthesis operator instead. EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE); @@ -189,19 +189,19 @@ class Tensor : public TensorBase > } #endif - inline Scalar& operator()(const array& indices) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(const array& indices) { eigen_assert(checkIndexRange(indices)); return coeffRef(indices); } - inline Scalar& operator()(Index index) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(Index index) { eigen_assert(index >= 0 && index < size()); return coeffRef(index); } - inline Scalar& operator[](Index index) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator[](Index index) { // The bracket operator is only for vectors, use the parenthesis operator instead EIGEN_STATIC_ASSERT(NumIndices == 1, YOU_MADE_A_PROGRAMMING_MISTAKE) @@ -223,11 +223,10 @@ class Tensor : public TensorBase > #ifdef EIGEN_HAS_VARIADIC_TEMPLATES template inline Tensor(Index firstDimension, IndexTypes... otherDimensions) - : m_storage() + : m_storage(internal::array_prod(array{{firstDimension, otherDimensions...}}), array{{firstDimension, otherDimensions...}}) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) - resize(array{{firstDimension, otherDimensions...}}); } #endif @@ -237,7 +236,6 @@ class Tensor : public TensorBase > EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } - template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor& operator=(const OtherDerived& other) @@ -306,7 +304,7 @@ class Tensor : public TensorBase > array_zip_and_reduce(indices, m_storage.dimensions()); } - inline Index linearizedIndex(const array& indices) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index linearizedIndex(const array& indices) const { if (Options&RowMajor) { return m_storage.dimensions().IndexOfRowMajor(indices); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index e69ff6188..da1eb62cb 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -53,7 +53,6 @@ template struct TensorAssign { typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC static inline void run(Derived1& dst, const Derived2& src) { TensorEvaluator evalDst(dst); @@ -63,7 +62,7 @@ struct TensorAssign static const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; static const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; static const int PacketSize = unpacket_traits::PacketReturnType>::size; - static const int VectorizedSize = (size / PacketSize) * PacketSize; + const int VectorizedSize = (size / PacketSize) * PacketSize; for (Index i = 0; i < VectorizedSize; i += PacketSize) { evalDst.template writePacket(i, evalSrc.template packet(i)); @@ -148,7 +147,7 @@ struct TensorAssignMultiThreaded // GPU: the evaluation of the expressions is offloaded to a GPU. -#ifdef EIGEN_USE_GPU +#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) template __global__ void EigenMetaKernelNoCheck(LhsEvaluator evalDst, const RhsEvaluator evalSrc) { const int index = blockIdx.x * blockDim.x + threadIdx.x; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 8a88ba806..c5c711313 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -30,13 +30,16 @@ class TensorBase typedef Scalar CoeffReturnType; typedef typename internal::packet_traits::type PacketReturnType; - Derived& setZero() { + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setZero() { return setConstant(Scalar(0)); } - Derived& setConstant(const Scalar& val) { + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setConstant(const Scalar& val) { return derived() = constant(val); } - Derived& setRandom() { + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setRandom() { return derived() = random(); } @@ -45,13 +48,13 @@ class TensorBase EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> constant(const Scalar& value) const { return TensorCwiseNullaryOp, const Derived> - (internal::scalar_constant_op(value)); + (derived(), internal::scalar_constant_op(value)); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> random() const { - return TensorCwiseNullaryOp, const Derived>(); + return TensorCwiseNullaryOp, const Derived>(derived()); } // Coefficient-wise unary operators @@ -124,77 +127,86 @@ class TensorBase // Coefficient-wise binary operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator+(const OtherDerived& other) const { + operator+(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator-(const OtherDerived& other) const { + operator-(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator*(const OtherDerived& other) const { + operator*(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator/(const OtherDerived& other) const { + operator/(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - cwiseMax(const OtherDerived& other) const { + cwiseMax(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - cwiseMin(const OtherDerived& other) const { + cwiseMin(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } // Comparisons and tests. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator<(const OtherDerived& other) const { + operator<(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator<=(const OtherDerived& other) const { + operator<=(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator>(const OtherDerived& other) const { + operator>(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator>=(const OtherDerived& other) const { + operator>=(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator==(const OtherDerived& other) const { + operator==(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> - operator!=(const OtherDerived& other) const { + operator!=(const OtherDerived& other) const { return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } + // Contractions. + typedef std::pair DimensionPair; + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorContractionOp + contract(const OtherDerived& other, const Dimensions& dims) const { + return TensorContractionOp(derived(), other.derived(), dims); + } + // Coefficient-wise ternary operators. - template + template inline const TensorSelectOp - select(const ThenDerived& thenTensor, const ElseDerived& elseTensor) const{ + select(const ThenDerived& thenTensor, const ElseDerived& elseTensor) const { return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h new file mode 100644 index 000000000..d424df36e --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -0,0 +1,229 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H +#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H + +namespace Eigen { + +/** \class TensorContraction + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor contraction class. + * + * + */ +namespace internal { +template +struct traits > +{ + // Type promotion to handle the case where the types of the lhs and the rhs are different. + typedef typename internal::promote_storage_type::ret Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename promote_storage_type::StorageKind, + typename traits::StorageKind>::ret StorageKind; + typedef typename promote_index_type::Index, + typename traits::Index>::type Index; + typedef typename LhsXprType::Nested LhsNested; + typedef typename RhsXprType::Nested RhsNested; + typedef typename remove_reference::type _LhsNested; + typedef typename remove_reference::type _RhsNested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorContractionOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorContractionOp type; +}; + +} // end namespace internal + + + +template +class TensorContractionOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename internal::promote_storage_type::ret CoeffReturnType; + typedef typename internal::promote_storage_type::ret PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionOp(const LhsXprType& lhs, const RhsXprType& rhs, const Indices& dims) + : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_indices(dims) {} + + EIGEN_DEVICE_FUNC + const Indices& indices() const { return m_indices; } + + /** \returns the nested expressions */ + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + lhsExpression() const { return m_lhs_xpr; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + rhsExpression() const { return m_rhs_xpr; } + + protected: + typename LhsXprType::Nested m_lhs_xpr; + typename RhsXprType::Nested m_rhs_xpr; + const Indices m_indices; +}; + + +template struct max_n_1 { + static const size_t size = n; +}; +template <> struct max_n_1<0> { + static const size_t size = 1; +}; + + +template +struct TensorEvaluator > +{ + typedef TensorContractionOp XprType; + + static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * Indices::size>::size; + typedef typename XprType::Index Index; + typedef DSizes Dimensions; + + enum { + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = /*TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess */ + false, + }; + + TensorEvaluator(const XprType& op) + : m_leftImpl(op.lhsExpression()), m_rightImpl(op.rhsExpression()) + { + Index index = 0; + Index stride = 1; + m_shiftright = 1; + + int skipped = 0; + const typename TensorEvaluator::Dimensions& left_dims = m_leftImpl.dimensions(); + for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { + bool skip = false; + for (int j = 0; j < Indices::size; ++j) { + if (op.indices()[j].first == i) { + skip = true; + m_leftOffsets[2*skipped] = stride; + m_leftOffsets[2*skipped+1] = stride * left_dims[i]; + m_stitchsize[skipped] = left_dims[i]; + break; + } + } + if (!skip) { + m_dimensions[index++] = left_dims[i]; + m_shiftright *= left_dims[i]; + } else { + ++skipped; + } + stride *= left_dims[i]; + } + + stride = 1; + skipped = 0; + const typename TensorEvaluator::Dimensions& right_dims = m_rightImpl.dimensions(); + for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { + bool skip = false; + for (int j = 0; j < Indices::size; ++j) { + if (op.indices()[j].second == i) { + skip = true; + m_rightOffsets[2*skipped] = stride; + m_rightOffsets[2*skipped+1] = stride * right_dims[i]; + break; + } + } + if (!skip) { + m_dimensions[index++] = right_dims[i]; + } else { + ++skipped; + } + stride *= right_dims[i]; + } + + // Scalar case + if (TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * Indices::size) { + m_dimensions[0] = 1; + } + } + + // typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + const Dimensions& dimensions() const { return m_dimensions; } + + void evalTo(typename XprType::Scalar* buffer) const { + for (int i = 0; i < dimensions().TotalSize(); ++i) { + buffer[i] += coeff(i); + } + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + const Index startLeft = index % m_shiftright; + const Index startRight = index / m_shiftright; + CoeffReturnType result = CoeffReturnType(0); + partialStitch(startLeft, startRight, 0, result); + return result; + } + + /* TODO: vectorization + template + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + { + assert(false); + }*/ + + private: + EIGEN_DEVICE_FUNC void partialStitch(Index startLeft, Index startRight, int StitchIndex, CoeffReturnType& accum) const { + Index firstLeft = (startLeft / m_leftOffsets[2*StitchIndex]) * m_leftOffsets[2*StitchIndex+1] + (startLeft % m_leftOffsets[2*StitchIndex]); + Index firstRight = (startRight / m_rightOffsets[2*StitchIndex]) * m_rightOffsets[2*StitchIndex+1] + (startRight % m_rightOffsets[2*StitchIndex]); + + for (int j = 0; j < m_stitchsize[StitchIndex]; ++j) { + const Index left = firstLeft+j*m_leftOffsets[2*StitchIndex]; + const Index right = firstRight+j*m_rightOffsets[2*StitchIndex]; + if (StitchIndex < Indices::size-1) { + partialStitch(left, right, StitchIndex+1, accum); + } else { + accum += m_leftImpl.coeff(left) * m_rightImpl.coeff(right); + } + } + } + + private: + array m_leftOffsets; + array m_rightOffsets; + array m_stitchsize; + Index m_shiftright; + Dimensions m_dimensions; + TensorEvaluator m_leftImpl; + TensorEvaluator m_rightImpl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h index 71890e187..dbe60a165 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h @@ -59,7 +59,7 @@ template class TensorDevice class TensorDevice { public: diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h index ded6ca604..d7f5ab7c9 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -37,17 +37,14 @@ struct ThreadPoolDevice { // GPU offloading #ifdef EIGEN_USE_GPU struct GpuDevice { - // todo: support for multiple gpu; - GpuDevice() { - cudaStreamCreate(&stream_); - } - ~GpuDevice() { - cudaStreamDestroy(stream_); - } - const cudaStream_t& stream() const { return stream_; } + // The cudastream is not owned: the caller is responsible for its initialization and eventual destruction. + GpuDevice(const cudaStream_t* stream) : stream_(stream) { eigen_assert(stream); } + + const cudaStream_t& stream() const { return *stream_; } private: - cudaStream_t stream_; + // TODO: multigpu. + const cudaStream_t* stream_; }; #endif diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 43e9d6550..c92b8c679 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -35,14 +35,14 @@ namespace Eigen { namespace internal { template struct dget { - static const std::size_t value = internal::get::value; - }; + static const std::size_t value = get::value; +}; template struct fixed_size_tensor_index_linearization_helper { - template + template EIGEN_DEVICE_FUNC static inline Index run(array const& indices, const Dimensions& dimensions) { @@ -55,7 +55,7 @@ struct fixed_size_tensor_index_linearization_helper template struct fixed_size_tensor_index_linearization_helper { - template + template EIGEN_DEVICE_FUNC static inline Index run(array const& indices, const Dimensions&) { @@ -93,11 +93,11 @@ struct Sizes : internal::numeric_list { return *this; } - template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfColMajor(const array& indices) const { return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this)); } - template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfRowMajor(const array& indices) const { return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this)); } @@ -139,11 +139,11 @@ template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfColMajor(const array& indices) const { return internal::fixed_size_tensor_index_linearization_helper::run(indices, *this); } - template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfRowMajor(const array& indices) const { return internal::fixed_size_tensor_index_linearization_helper::run(indices, *this); } @@ -180,13 +180,18 @@ struct tensor_index_linearization_helper template struct DSizes : array { typedef array Base; + static const std::size_t count = NumDims; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() const { return internal::array_prod(*static_cast(this)); } - DSizes() { } - explicit DSizes(const array& a) : Base(a) { } + EIGEN_DEVICE_FUNC DSizes() { + for (int i = 0 ; i < NumDims; ++i) { + (*this)[i] = 0; + } + } + EIGEN_DEVICE_FUNC explicit DSizes(const array& a) : Base(a) { } DSizes& operator = (const array& other) { *static_cast(this) = other; @@ -194,10 +199,10 @@ struct DSizes : array { } // A constexpr would be so much better here - size_t IndexOfColMajor(const array& indices) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfColMajor(const array& indices) const { return internal::tensor_index_linearization_helper::run(indices, *static_cast(this)); } - size_t IndexOfRowMajor(const array& indices) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfRowMajor(const array& indices) const { return internal::tensor_index_linearization_helper::run(indices, *static_cast(this)); } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index e0c0863b7..ab2513cea 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -21,7 +21,6 @@ namespace Eigen { * * TODO: add support for more types of expressions, in particular expressions * leading to lvalues (slicing, reshaping, etc...) - * TODO: add support for vectorization */ template @@ -32,16 +31,19 @@ struct TensorEvaluator typedef typename Derived::Packet Packet; typedef typename Derived::Scalar CoeffReturnType; typedef typename Derived::Packet PacketReturnType; + typedef typename Derived::Dimensions Dimensions; enum { IsAligned = Derived::IsAligned, PacketAccess = Derived::PacketAccess, }; - TensorEvaluator(Derived& m) - : m_data(const_cast(m.data())) + EIGEN_DEVICE_FUNC TensorEvaluator(Derived& m) + : m_data(const_cast(m.data())), m_dims(m.dimensions()) { } + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_dims; } + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_data[index]; } @@ -64,29 +66,34 @@ struct TensorEvaluator protected: Scalar* m_data; + Dimensions m_dims; }; // -------------------- CwiseNullaryOp -------------------- -template -struct TensorEvaluator > +template +struct TensorEvaluator > { - typedef TensorCwiseNullaryOp XprType; + typedef TensorCwiseNullaryOp XprType; enum { IsAligned = true, PacketAccess = internal::functor_traits::PacketAccess, }; + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op) - : m_functor(op.functor()) + : m_functor(op.functor()), m_argImpl(op.nestedExpression()) { } typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename TensorEvaluator::Dimensions Dimensions; + + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { @@ -101,6 +108,7 @@ struct TensorEvaluator > private: const NullaryOp m_functor; + TensorEvaluator m_argImpl; }; @@ -117,7 +125,7 @@ struct TensorEvaluator > PacketAccess = TensorEvaluator::PacketAccess & internal::functor_traits::PacketAccess, }; - TensorEvaluator(const XprType& op) + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -125,6 +133,9 @@ struct TensorEvaluator > typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename TensorEvaluator::Dimensions Dimensions; + + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { @@ -156,7 +167,7 @@ struct TensorEvaluator::PacketAccess, }; - TensorEvaluator(const XprType& op) + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op) : m_functor(op.functor()), m_leftImpl(op.lhsExpression()), m_rightImpl(op.rhsExpression()) @@ -165,6 +176,13 @@ struct TensorEvaluator::Dimensions Dimensions; + + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const + { + // TODO: use right impl instead if right impl dimensions are known at compile time. + return m_leftImpl.dimensions(); + } EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { @@ -196,7 +214,7 @@ struct TensorEvaluator TensorEvaluator::PacketAccess*/, }; - TensorEvaluator(const XprType& op) + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op) : m_condImpl(op.ifExpression()), m_thenImpl(op.thenExpression()), m_elseImpl(op.elseExpression()) @@ -205,7 +223,13 @@ struct TensorEvaluator typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename TensorEvaluator::Dimensions Dimensions; + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const + { + // TODO: use then or else impl instead if they happen to be known at compile time. + return m_condImpl.dimensions(); + } EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_condImpl.coeff(index) ? m_thenImpl.coeff(index) : m_elseImpl.coeff(index); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h index 94cfae05c..60908ee94 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h @@ -28,13 +28,13 @@ namespace Eigen { * */ namespace internal { -template -struct traits > - : traits +template +struct traits > + : traits { - typedef typename PlainObjectType::Packet Packet; - typedef typename PlainObjectType::Scalar Scalar; - typedef typename PlainObjectType::Nested XprTypeNested; + typedef typename XprType::Packet Packet; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::Nested XprTypeNested; typedef typename remove_reference::type _XprTypeNested; }; @@ -42,27 +42,31 @@ struct traits > -template -class TensorCwiseNullaryOp : public TensorBase > +template +class TensorCwiseNullaryOp : public TensorBase > { public: typedef typename Eigen::internal::traits::Scalar Scalar; typedef typename Eigen::internal::traits::Packet Packet; typedef typename Eigen::NumTraits::Real RealScalar; - typedef typename PlainObjectType::CoeffReturnType CoeffReturnType; - typedef typename PlainObjectType::PacketReturnType PacketReturnType; - typedef TensorCwiseNullaryOp Nested; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef TensorCwiseNullaryOp Nested; typedef typename Eigen::internal::traits::StorageKind StorageKind; typedef typename Eigen::internal::traits::Index Index; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseNullaryOp(const NullaryOp& func = NullaryOp()) - : m_functor(func) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCwiseNullaryOp(const XprType& xpr, const NullaryOp& func = NullaryOp()) + : m_xpr(xpr), m_functor(func) {} + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + nestedExpression() const { return m_xpr; } EIGEN_DEVICE_FUNC const NullaryOp& functor() const { return m_functor; } protected: - // todo: add tensor dimension to be able to do some sanity checks + typename XprType::Nested m_xpr; const NullaryOp m_functor; }; @@ -71,7 +75,7 @@ class TensorCwiseNullaryOp : public TensorBase struct traits > - : traits + : traits { typedef typename result_of< UnaryOp(typename XprType::Scalar) @@ -207,7 +211,7 @@ class TensorCwiseBinaryOp : public TensorBase struct traits > - : traits + : traits { typedef typename traits::Scalar Scalar; typedef typename internal::packet_traits::type Packet; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h index dcc7ccd65..789c04238 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h @@ -52,7 +52,7 @@ class TensorFixedSize : public TensorBase dimensions() const { return m_storage.dimensions(); } + EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_storage.dimensions(); } EIGEN_STRONG_INLINE Index size() const { return m_storage.size(); } EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 03ac8d516..239b5cb67 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -21,6 +21,8 @@ template class TensorCwiseNullaryO template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; template class TensorSelectOp; +template class TensorReductionOp; +template class TensorContractionOp; template class TensorDevice; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index 64098343e..c9d6517eb 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -53,7 +53,7 @@ class TensorStorage EIGEN_STRONG_INLINE const T *data() const { return m_data; } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const FixedDimensions dimensions() const { return m_dimensions; } + EIGEN_STRONG_INLINE const FixedDimensions& dimensions() const { return m_dimensions; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex size() const { return m_dimensions.TotalSize(); } @@ -111,7 +111,8 @@ class TensorStorage(m_data, internal::array_prod(m_dimensions)); } void swap(Self_& other) { std::swap(m_data,other.m_data); std::swap(m_dimensions,other.m_dimensions); } - const DSizes& dimensions() const {return m_dimensions;} + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const DSizes& dimensions() const {return m_dimensions;} void conservativeResize(DenseIndex size, const array& nbDimensions) { @@ -132,10 +133,10 @@ class TensorStorage Date: Wed, 4 Jun 2014 18:31:02 +0200 Subject: [PATCH 0505/1103] Fix compilation for CXX11/Tensor module if unsupported is not in include path --- unsupported/Eigen/CXX11/Tensor | 2 +- unsupported/Eigen/CXX11/TensorSymmetry | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index f2c5129b3..049ce5596 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -10,7 +10,7 @@ #ifndef EIGEN_CXX11_TENSOR_MODULE #define EIGEN_CXX11_TENSOR_MODULE -#include +#include #include diff --git a/unsupported/Eigen/CXX11/TensorSymmetry b/unsupported/Eigen/CXX11/TensorSymmetry index 027c6087f..f1dc25fea 100644 --- a/unsupported/Eigen/CXX11/TensorSymmetry +++ b/unsupported/Eigen/CXX11/TensorSymmetry @@ -10,7 +10,7 @@ #ifndef EIGEN_CXX11_TENSORSYMMETRY_MODULE #define EIGEN_CXX11_TENSORSYMMETRY_MODULE -#include +#include #include From 58cfac9a1277f2b2198ed1e5c3190543d1ed95a6 Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Wed, 4 Jun 2014 18:47:42 +0200 Subject: [PATCH 0506/1103] unsupported/ C++11 workarounds: don't use hack for libc++ if not required libc++ from 3.4 onwards supports constexpr std::get, but only if compiled with -std=c++1y. Change the detection so that libc++'s internals are only used if either -std=c++1y is not specified or the library is too old, making the whole hack a bit more future-proof. --- unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index 356ae10cf..d71a67590 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -48,13 +48,15 @@ namespace internal { * - libstdc++ from version 4.7 onwards has it nevertheless, * so use that * - libstdc++ older versions: use _M_instance directly - * - libc++ all versions so far: use __elems_ directly + * - libc++ from version 3.4 onwards has it IF compiled with + * -std=c++1y + * - libc++ older versions or -std=c++11: use __elems_ directly * - all other libs: use std::get to be portable, but * this may not be constexpr */ #if defined(__GLIBCXX__) && __GLIBCXX__ < 20120322 #define STD_GET_ARR_HACK a._M_instance[I] -#elif defined(_LIBCPP_VERSION) +#elif defined(_LIBCPP_VERSION) && (!defined(_LIBCPP_STD_VER) || _LIBCPP_STD_VER <= 11) #define STD_GET_ARR_HACK a.__elems_[I] #else #define STD_GET_ARR_HACK std::template get(a) From cee62018fc38f5408e0afe497c37fade64ca15d0 Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Wed, 4 Jun 2014 19:54:22 +0200 Subject: [PATCH 0507/1103] unsupported/CXX11/Core: allow gen_numeric_list to have a starting point Add a template parameter to gen_numeric_list that acts as a starting point for the list, i.e. gen_numeric_list will generate a numeric_list. --- .../Eigen/CXX11/src/Core/util/CXX11Meta.h | 12 ++++++------ unsupported/test/cxx11_meta.cpp | 18 ++++++++++++++++++ 2 files changed, 24 insertions(+), 6 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h index 618e2eb7b..0e274b801 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -42,14 +42,14 @@ struct numeric_list { constexpr static std::size_t count = sizeof.. * typename gen_numeric_list_repeated::type numeric_list */ -template struct gen_numeric_list : gen_numeric_list {}; -template struct gen_numeric_list { typedef numeric_list type; }; +template struct gen_numeric_list : gen_numeric_list {}; +template struct gen_numeric_list { typedef numeric_list type; }; -template struct gen_numeric_list_reversed : gen_numeric_list_reversed {}; -template struct gen_numeric_list_reversed { typedef numeric_list type; }; +template struct gen_numeric_list_reversed : gen_numeric_list_reversed {}; +template struct gen_numeric_list_reversed { typedef numeric_list type; }; -template struct gen_numeric_list_swapped_pair : gen_numeric_list_swapped_pair {}; -template struct gen_numeric_list_swapped_pair { typedef numeric_list type; }; +template struct gen_numeric_list_swapped_pair : gen_numeric_list_swapped_pair {}; +template struct gen_numeric_list_swapped_pair { typedef numeric_list type; }; template struct gen_numeric_list_repeated : gen_numeric_list_repeated {}; template struct gen_numeric_list_repeated { typedef numeric_list type; }; diff --git a/unsupported/test/cxx11_meta.cpp b/unsupported/test/cxx11_meta.cpp index a9843e9a9..af5cadbf9 100644 --- a/unsupported/test/cxx11_meta.cpp +++ b/unsupported/test/cxx11_meta.cpp @@ -91,18 +91,36 @@ static void test_gen_numeric_list() VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); + VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); VERIFY((is_same::type, numeric_list>::value)); From ea9943352368b990d27ba22eb8670287cf96302d Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Wed, 4 Jun 2014 20:27:42 +0200 Subject: [PATCH 0508/1103] unsupported/TensorSymmetry: make symgroup construction autodetect number of indices When constructing a symmetry group, make the code automatically detect the number of indices required from the indices of the group's generators. Also, allow the symmetry group to be applied to lists of indices that are larger than the number of indices of the symmetry group. Before: SGroup<4, Symmetry<0, 1>, Symmetry<2,3>> group; group.apply(std::array{{0, 1, 2, 3}}, 0); After: SGroup, Symmetry<2,3>> group; group.apply(std::array{{0, 1, 2, 3}}, 0); group.apply(std::array{{0, 1, 2, 3, 4}}, 0); This should make the symmetry group easier to use - especially if one wants to reuse the same symmetry group for different tensors of maybe different rank. static/runtime asserts remain for the case where the length of the index list to which a symmetry group is to be applied is too small. --- .../src/TensorSymmetry/DynamicSymmetry.h | 42 +++++++++++----- .../CXX11/src/TensorSymmetry/StaticSymmetry.h | 50 +++++++++++-------- .../Eigen/CXX11/src/TensorSymmetry/Symmetry.h | 47 +++++++++++++---- unsupported/test/cxx11_tensor_symmetry.cpp | 26 +++++----- 4 files changed, 108 insertions(+), 57 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h index b5738b778..0329278a9 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h @@ -15,7 +15,7 @@ namespace Eigen { class DynamicSGroup { public: - inline explicit DynamicSGroup(std::size_t numIndices) : m_numIndices(numIndices), m_elements(), m_generators(), m_globalFlags(0) { m_elements.push_back(ge(Generator(0, 0, 0))); } + inline explicit DynamicSGroup() : m_numIndices(1), m_elements(), m_generators(), m_globalFlags(0) { m_elements.push_back(ge(Generator(0, 0, 0))); } inline DynamicSGroup(const DynamicSGroup& o) : m_numIndices(o.m_numIndices), m_elements(o.m_elements), m_generators(o.m_generators), m_globalFlags(o.m_globalFlags) { } inline DynamicSGroup(DynamicSGroup&& o) : m_numIndices(o.m_numIndices), m_elements(), m_generators(o.m_generators), m_globalFlags(o.m_globalFlags) { std::swap(m_elements, o.m_elements); } inline DynamicSGroup& operator=(const DynamicSGroup& o) { m_numIndices = o.m_numIndices; m_elements = o.m_elements; m_generators = o.m_generators; m_globalFlags = o.m_globalFlags; return *this; } @@ -33,7 +33,7 @@ class DynamicSGroup template inline RV apply(const std::array& idx, RV initial, Args&&... args) const { - eigen_assert(N == m_numIndices); + eigen_assert(N >= m_numIndices && "Can only apply symmetry group to objects that have at least the required amount of indices."); for (std::size_t i = 0; i < size(); i++) initial = Op::run(h_permute(i, idx, typename internal::gen_numeric_list::type()), m_elements[i].flags, initial, std::forward(args)...); return initial; @@ -42,7 +42,7 @@ class DynamicSGroup template inline RV apply(const std::vector& idx, RV initial, Args&&... args) const { - eigen_assert(idx.size() == m_numIndices); + eigen_assert(idx.size() >= m_numIndices && "Can only apply symmetry group to objects that have at least the required amount of indices."); for (std::size_t i = 0; i < size(); i++) initial = Op::run(h_permute(i, idx), m_elements[i].flags, initial, std::forward(args)...); return initial; @@ -77,7 +77,7 @@ class DynamicSGroup template inline std::array h_permute(std::size_t which, const std::array& idx, internal::numeric_list) const { - return std::array{{ idx[m_elements[which].representation[n]]... }}; + return std::array{{ idx[n >= m_numIndices ? n : m_elements[which].representation[n]]... }}; } template @@ -87,6 +87,8 @@ class DynamicSGroup result.reserve(idx.size()); for (auto k : m_elements[which].representation) result.push_back(idx[k]); + for (std::size_t i = m_numIndices; i < idx.size(); i++) + result.push_back(idx[i]); return result; } @@ -135,18 +137,18 @@ class DynamicSGroup }; // dynamic symmetry group that auto-adds the template parameters in the constructor -template +template class DynamicSGroupFromTemplateArgs : public DynamicSGroup { public: - inline DynamicSGroupFromTemplateArgs() : DynamicSGroup(NumIndices) + inline DynamicSGroupFromTemplateArgs() : DynamicSGroup() { add_all(internal::type_list()); } inline DynamicSGroupFromTemplateArgs(DynamicSGroupFromTemplateArgs const& other) : DynamicSGroup(other) { } inline DynamicSGroupFromTemplateArgs(DynamicSGroupFromTemplateArgs&& other) : DynamicSGroup(other) { } - inline DynamicSGroupFromTemplateArgs& operator=(const DynamicSGroupFromTemplateArgs& o) { DynamicSGroup::operator=(o); return *this; } - inline DynamicSGroupFromTemplateArgs& operator=(DynamicSGroupFromTemplateArgs&& o) { DynamicSGroup::operator=(o); return *this; } + inline DynamicSGroupFromTemplateArgs& operator=(const DynamicSGroupFromTemplateArgs& o) { DynamicSGroup::operator=(o); return *this; } + inline DynamicSGroupFromTemplateArgs& operator=(DynamicSGroupFromTemplateArgs&& o) { DynamicSGroup::operator=(o); return *this; } private: template @@ -168,18 +170,32 @@ inline DynamicSGroup::GroupElement DynamicSGroup::mul(GroupElement g1, GroupElem GroupElement result; result.representation.reserve(m_numIndices); - for (std::size_t i = 0; i < m_numIndices; i++) - result.representation.push_back(g2.representation[g1.representation[i]]); + for (std::size_t i = 0; i < m_numIndices; i++) { + int v = g2.representation[g1.representation[i]]; + eigen_assert(v >= 0); + result.representation.push_back(v); + } result.flags = g1.flags ^ g2.flags; return result; } inline void DynamicSGroup::add(int one, int two, int flags) { - eigen_assert(one >= 0 && (std::size_t)one < m_numIndices); - eigen_assert(two >= 0 && (std::size_t)two < m_numIndices); + eigen_assert(one >= 0); + eigen_assert(two >= 0); eigen_assert(one != two); - Generator g{one, two ,flags}; + + if ((std::size_t)one >= m_numIndices || (std::size_t)two >= m_numIndices) { + std::size_t newNumIndices = (one > two) ? one : two + 1; + for (auto& gelem : m_elements) { + gelem.representation.reserve(newNumIndices); + for (std::size_t i = m_numIndices; i < newNumIndices; i++) + gelem.representation.push_back(i); + } + m_numIndices = newNumIndices; + } + + Generator g{one, two, flags}; GroupElement e = ge(g); /* special case for first generator */ diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h index c5a630105..0eb468fc0 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h @@ -114,20 +114,24 @@ struct tensor_static_symgroup_equality template struct tensor_static_symgroup { - typedef StaticSGroup type; + typedef StaticSGroup type; constexpr static std::size_t size = type::static_size; }; -template -constexpr static inline std::array tensor_static_symgroup_index_permute(std::array idx, internal::numeric_list) +template +constexpr static inline std::array tensor_static_symgroup_index_permute(std::array idx, internal::numeric_list, internal::numeric_list) { - return {{ idx[ii]... }}; + return {{ idx[ii]..., idx[jj]... }}; } template static inline std::vector tensor_static_symgroup_index_permute(std::vector idx, internal::numeric_list) { - return {{ idx[ii]... }}; + std::vector result{{ idx[ii]... }}; + std::size_t target_size = idx.size(); + for (std::size_t i = result.size(); i < target_size; i++) + result.push_back(idx[i]); + return result; } template struct tensor_static_symgroup_do_apply; @@ -135,32 +139,35 @@ template struct tensor_static_symgroup_do_apply; template struct tensor_static_symgroup_do_apply> { - template - static inline RV run(const std::array& idx, RV initial, Args&&... args) + template + static inline RV run(const std::array& idx, RV initial, Args&&... args) { - initial = Op::run(tensor_static_symgroup_index_permute(idx, typename first::indices()), first::flags, initial, std::forward(args)...); - return tensor_static_symgroup_do_apply>::template run(idx, initial, args...); + static_assert(NumIndices >= SGNumIndices, "Can only apply symmetry group to objects that have at least the required amount of indices."); + typedef typename internal::gen_numeric_list::type remaining_indices; + initial = Op::run(tensor_static_symgroup_index_permute(idx, typename first::indices(), remaining_indices()), first::flags, initial, std::forward(args)...); + return tensor_static_symgroup_do_apply>::template run(idx, initial, args...); } - template + template static inline RV run(const std::vector& idx, RV initial, Args&&... args) { + eigen_assert(idx.size() >= SGNumIndices && "Can only apply symmetry group to objects that have at least the required amount of indices."); initial = Op::run(tensor_static_symgroup_index_permute(idx, typename first::indices()), first::flags, initial, std::forward(args)...); - return tensor_static_symgroup_do_apply>::template run(idx, initial, args...); + return tensor_static_symgroup_do_apply>::template run(idx, initial, args...); } }; template struct tensor_static_symgroup_do_apply> { - template - static inline RV run(const std::array&, RV initial, Args&&...) + template + static inline RV run(const std::array&, RV initial, Args&&...) { // do nothing return initial; } - template + template static inline RV run(const std::vector&, RV initial, Args&&...) { // do nothing @@ -170,9 +177,10 @@ struct tensor_static_symgroup_do_apply +template class StaticSGroup { + constexpr static std::size_t NumIndices = internal::tensor_symmetry_num_indices::value; typedef internal::group_theory::enumerate_group_elements< internal::tensor_static_symgroup_multiply, internal::tensor_static_symgroup_equality, @@ -182,20 +190,20 @@ class StaticSGroup typedef typename group_elements::type ge; public: constexpr inline StaticSGroup() {} - constexpr inline StaticSGroup(const StaticSGroup&) {} - constexpr inline StaticSGroup(StaticSGroup&&) {} + constexpr inline StaticSGroup(const StaticSGroup&) {} + constexpr inline StaticSGroup(StaticSGroup&&) {} - template - static inline RV apply(const std::array& idx, RV initial, Args&&... args) + template + static inline RV apply(const std::array& idx, RV initial, Args&&... args) { - return internal::tensor_static_symgroup_do_apply::template run(idx, initial, args...); + return internal::tensor_static_symgroup_do_apply::template run(idx, initial, args...); } template static inline RV apply(const std::vector& idx, RV initial, Args&&... args) { eigen_assert(idx.size() == NumIndices); - return internal::tensor_static_symgroup_do_apply::template run(idx, initial, args...); + return internal::tensor_static_symgroup_do_apply::template run(idx, initial, args...); } constexpr static std::size_t static_size = ge::count; diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h index f0813086a..f1ccc33ef 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h @@ -30,6 +30,7 @@ template struct tenso template struct tensor_static_symgroup_if; template struct tensor_symmetry_calculate_flags; template struct tensor_symmetry_assign_value; +template struct tensor_symmetry_num_indices; } // end namespace internal @@ -94,7 +95,7 @@ class DynamicSGroup; * This class is a child class of DynamicSGroup. It uses the template arguments * specified to initialize itself. */ -template +template class DynamicSGroupFromTemplateArgs; /** \class StaticSGroup @@ -116,7 +117,7 @@ class DynamicSGroupFromTemplateArgs; * group becomes too large. (In that case, unrolling may not even be * beneficial.) */ -template +template class StaticSGroup; /** \class SGroup @@ -131,24 +132,50 @@ class StaticSGroup; * \sa StaticSGroup * \sa DynamicSGroup */ -template -class SGroup : public internal::tensor_symmetry_pre_analysis::root_type +template +class SGroup : public internal::tensor_symmetry_pre_analysis::value, Gen...>::root_type { public: + constexpr static std::size_t NumIndices = internal::tensor_symmetry_num_indices::value; typedef typename internal::tensor_symmetry_pre_analysis::root_type Base; // make standard constructors + assignment operators public inline SGroup() : Base() { } - inline SGroup(const SGroup& other) : Base(other) { } - inline SGroup(SGroup&& other) : Base(other) { } - inline SGroup& operator=(const SGroup& other) { Base::operator=(other); return *this; } - inline SGroup& operator=(SGroup&& other) { Base::operator=(other); return *this; } + inline SGroup(const SGroup& other) : Base(other) { } + inline SGroup(SGroup&& other) : Base(other) { } + inline SGroup& operator=(const SGroup& other) { Base::operator=(other); return *this; } + inline SGroup& operator=(SGroup&& other) { Base::operator=(other); return *this; } // all else is defined in the base class }; namespace internal { +template struct tensor_symmetry_num_indices +{ + constexpr static std::size_t value = 1; +}; + +template struct tensor_symmetry_num_indices, Sym...> +{ +private: + constexpr static std::size_t One = static_cast(One_); + constexpr static std::size_t Two = static_cast(Two_); + constexpr static std::size_t Three = tensor_symmetry_num_indices::value; + + // don't use std::max, since it's not constexpr until C++14... + constexpr static std::size_t maxOneTwoPlusOne = ((One > Two) ? One : Two) + 1; +public: + constexpr static std::size_t value = (maxOneTwoPlusOne > Three) ? maxOneTwoPlusOne : Three; +}; + +template struct tensor_symmetry_num_indices, Sym...> + : public tensor_symmetry_num_indices, Sym...> {}; +template struct tensor_symmetry_num_indices, Sym...> + : public tensor_symmetry_num_indices, Sym...> {}; +template struct tensor_symmetry_num_indices, Sym...> + : public tensor_symmetry_num_indices, Sym...> {}; + /** \internal * * \class tensor_symmetry_pre_analysis @@ -199,7 +226,7 @@ namespace internal { template struct tensor_symmetry_pre_analysis { - typedef StaticSGroup root_type; + typedef StaticSGroup<> root_type; }; template @@ -212,7 +239,7 @@ struct tensor_symmetry_pre_analysis typedef typename conditional< possible_size == 0 || possible_size >= max_static_elements, - DynamicSGroupFromTemplateArgs, + DynamicSGroupFromTemplateArgs, typename helper::type >::type root_type; }; diff --git a/unsupported/test/cxx11_tensor_symmetry.cpp b/unsupported/test/cxx11_tensor_symmetry.cpp index e8dfffd92..2a1669995 100644 --- a/unsupported/test/cxx11_tensor_symmetry.cpp +++ b/unsupported/test/cxx11_tensor_symmetry.cpp @@ -32,8 +32,8 @@ using Eigen::GlobalImagFlag; // helper function to determine if the compiler intantiated a static // or dynamic symmetry group -template -bool isDynGroup(StaticSGroup const& dummy) +template +bool isDynGroup(StaticSGroup const& dummy) { (void)dummy; return false; @@ -86,7 +86,7 @@ static void test_symgroups_static() std::array identity{{0,1,2,3,4,5,6}}; // Simple static symmetry group - StaticSGroup<7, + StaticSGroup< AntiSymmetry<0,1>, Hermiticity<0,2> > group; @@ -113,7 +113,7 @@ static void test_symgroups_dynamic() identity.push_back(i); // Simple dynamic symmetry group - DynamicSGroup group(7); + DynamicSGroup group; group.add(0,1,NegationFlag); group.add(0,2,ConjugationFlag); @@ -143,7 +143,7 @@ static void test_symgroups_selection() { // Do the same test as in test_symgroups_static but // require selection via SGroup - SGroup<7, + SGroup< AntiSymmetry<0,1>, Hermiticity<0,2> > group; @@ -168,7 +168,7 @@ static void test_symgroups_selection() // simple factorizing group: 5 generators, 2^5 = 32 elements // selection should make this dynamic, although static group // can still be reasonably generated - SGroup<10, + SGroup< Symmetry<0,1>, Symmetry<2,3>, Symmetry<4,5>, @@ -196,7 +196,7 @@ static void test_symgroups_selection() // no verify that we could also generate a static group // with these generators found.clear(); - StaticSGroup<10, + StaticSGroup< Symmetry<0,1>, Symmetry<2,3>, Symmetry<4,5>, @@ -211,7 +211,7 @@ static void test_symgroups_selection() { // try to create a HUGE group - SGroup<7, + SGroup< Symmetry<0,1>, Symmetry<1,2>, Symmetry<2,3>, @@ -657,7 +657,7 @@ static void test_symgroups_selection() static void test_tensor_epsilon() { - SGroup<3, AntiSymmetry<0,1>, AntiSymmetry<1,2>> sym; + SGroup, AntiSymmetry<1,2>> sym; Tensor epsilon(3,3,3); epsilon.setZero(); @@ -674,7 +674,7 @@ static void test_tensor_epsilon() static void test_tensor_sym() { - SGroup<4, Symmetry<0,1>, Symmetry<2,3>> sym; + SGroup, Symmetry<2,3>> sym; Tensor t(10,10,10,10); t.setZero(); @@ -703,7 +703,7 @@ static void test_tensor_sym() static void test_tensor_asym() { - SGroup<4, AntiSymmetry<0,1>, AntiSymmetry<2,3>> sym; + SGroup, AntiSymmetry<2,3>> sym; Tensor t(10,10,10,10); t.setZero(); @@ -740,7 +740,7 @@ static void test_tensor_asym() static void test_tensor_dynsym() { - DynamicSGroup sym(4); + DynamicSGroup sym; sym.addSymmetry(0,1); sym.addSymmetry(2,3); Tensor t(10,10,10,10); @@ -770,7 +770,7 @@ static void test_tensor_dynsym() static void test_tensor_randacc() { - SGroup<4, Symmetry<0,1>, Symmetry<2,3>> sym; + SGroup, Symmetry<2,3>> sym; Tensor t(10,10,10,10); t.setZero(); From 96cb58fa3b83448fcb2af2d131434a7ac10b915c Mon Sep 17 00:00:00 2001 From: Christian Seiler Date: Wed, 4 Jun 2014 20:44:22 +0200 Subject: [PATCH 0509/1103] unsupported/TensorSymmetry: factor out completely from Tensor module Remove the symCoeff() method of the the Tensor module and move the functionality into a new operator() of the symmetry classes. This makes the Tensor module now completely self-contained without symmetry support (even though previously it was only a forward declaration and a otherwise harmless trivial templated method) and also removes the inconsistency with the rest of eigen w.r.t. the method's naming scheme. --- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 15 --------------- .../CXX11/src/TensorSymmetry/DynamicSymmetry.h | 13 +++++++++++++ .../CXX11/src/TensorSymmetry/StaticSymmetry.h | 13 +++++++++++++ .../Eigen/CXX11/src/TensorSymmetry/Symmetry.h | 2 +- unsupported/test/cxx11_tensor_symmetry.cpp | 10 +++++----- 5 files changed, 32 insertions(+), 21 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index c6216e14c..70ca1433f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -91,9 +91,6 @@ struct tensor_index_linearization_helper return std_array_get(indices); } }; - -/* Forward-declaration required for the symmetry support. */ -template class tensor_symmetry_value_setter; } // end namespace internal template @@ -285,18 +282,6 @@ class Tensor #endif } - template - internal::tensor_symmetry_value_setter symCoeff(const Symmetry_& symmetry, Index firstIndex, IndexTypes... otherIndices) - { - return symCoeff(symmetry, std::array{{firstIndex, otherIndices...}}); - } - - template - internal::tensor_symmetry_value_setter symCoeff(const Symmetry_& symmetry, std::array const& indices) - { - return internal::tensor_symmetry_value_setter(*this, symmetry, indices); - } - protected: bool checkIndexRange(const std::array& indices) const { diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h index 0329278a9..bc4f2025f 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/DynamicSymmetry.h @@ -50,6 +50,19 @@ class DynamicSGroup inline int globalFlags() const { return m_globalFlags; } inline std::size_t size() const { return m_elements.size(); } + + template + inline internal::tensor_symmetry_value_setter operator()(Tensor_& tensor, typename Tensor_::Index firstIndex, IndexTypes... otherIndices) const + { + static_assert(sizeof...(otherIndices) + 1 == Tensor_::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + return operator()(tensor, std::array{{firstIndex, otherIndices...}}); + } + + template + inline internal::tensor_symmetry_value_setter operator()(Tensor_& tensor, std::array const& indices) const + { + return internal::tensor_symmetry_value_setter(tensor, *this, indices); + } private: struct GroupElement { std::vector representation; diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h index 0eb468fc0..942293bd7 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/StaticSymmetry.h @@ -212,6 +212,19 @@ class StaticSGroup return ge::count; } constexpr static inline int globalFlags() { return group_elements::global_flags; } + + template + inline internal::tensor_symmetry_value_setter> operator()(Tensor_& tensor, typename Tensor_::Index firstIndex, IndexTypes... otherIndices) const + { + static_assert(sizeof...(otherIndices) + 1 == Tensor_::NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + return operator()(tensor, std::array{{firstIndex, otherIndices...}}); + } + + template + inline internal::tensor_symmetry_value_setter> operator()(Tensor_& tensor, std::array const& indices) const + { + return internal::tensor_symmetry_value_setter>(tensor, *this, indices); + } }; } // end namespace Eigen diff --git a/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h index f1ccc33ef..879d6cd77 100644 --- a/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h +++ b/unsupported/Eigen/CXX11/src/TensorSymmetry/Symmetry.h @@ -293,7 +293,7 @@ struct tensor_symmetry_calculate_flags } }; -template +template class tensor_symmetry_value_setter { public: diff --git a/unsupported/test/cxx11_tensor_symmetry.cpp b/unsupported/test/cxx11_tensor_symmetry.cpp index 2a1669995..d680e9b3b 100644 --- a/unsupported/test/cxx11_tensor_symmetry.cpp +++ b/unsupported/test/cxx11_tensor_symmetry.cpp @@ -661,7 +661,7 @@ static void test_tensor_epsilon() Tensor epsilon(3,3,3); epsilon.setZero(); - epsilon.symCoeff(sym, 0, 1, 2) = 1; + sym(epsilon, 0, 1, 2) = 1; for (int i = 0; i < 3; i++) { for (int j = 0; j < 3; j++) { @@ -683,7 +683,7 @@ static void test_tensor_sym() for (int k = l; k < 10; k++) { for (int j = 0; j < 10; j++) { for (int i = j; i < 10; i++) { - t.symCoeff(sym, i, j, k, l) = (i + j) * (k + l); + sym(t, i, j, k, l) = (i + j) * (k + l); } } } @@ -712,7 +712,7 @@ static void test_tensor_asym() for (int k = l + 1; k < 10; k++) { for (int j = 0; j < 10; j++) { for (int i = j + 1; i < 10; i++) { - t.symCoeff(sym, i, j, k, l) = ((i * j) + (k * l)); + sym(t, i, j, k, l) = ((i * j) + (k * l)); } } } @@ -751,7 +751,7 @@ static void test_tensor_dynsym() for (int k = l; k < 10; k++) { for (int j = 0; j < 10; j++) { for (int i = j; i < 10; i++) { - t.symCoeff(sym, i, j, k, l) = (i + j) * (k + l); + sym(t, i, j, k, l) = (i + j) * (k + l); } } } @@ -787,7 +787,7 @@ static void test_tensor_randacc() std::swap(i, j); if (k < l) std::swap(k, l); - t.symCoeff(sym, i, j, k, l) = (i + j) * (k + l); + sym(t, i, j, k, l) = (i + j) * (k + l); } for (int l = 0; l < 10; l++) { From 8998f4099e20ebc80db0aba2582301cd48d31c5a Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 5 Jun 2014 10:49:34 -0700 Subject: [PATCH 0510/1103] Created additional tests for the tensor code. --- unsupported/test/CMakeLists.txt | 2 + unsupported/test/cxx11_tensor_comparisons.cpp | 84 +++++++++ unsupported/test/cxx11_tensor_contraction.cpp | 163 ++++++++++++++++++ unsupported/test/cxx11_tensor_device.cpp | 17 +- unsupported/test/cxx11_tensor_expr.cpp | 149 ++++++++++++++-- unsupported/test/cxx11_tensor_fixed_size.cpp | 14 +- unsupported/test/cxx11_tensor_thread_pool.cpp | 7 +- 7 files changed, 406 insertions(+), 30 deletions(-) create mode 100644 unsupported/test/cxx11_tensor_comparisons.cpp create mode 100644 unsupported/test/cxx11_tensor_contraction.cpp diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index abc3375e5..d6072c9f3 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -102,6 +102,8 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_simple "-std=c++0x") ei_add_test(cxx11_tensor_symmetry "-std=c++0x") ei_add_test(cxx11_tensor_assign "-std=c++0x") + ei_add_test(cxx11_tensor_comparison "-std=c++0x") + ei_add_test(cxx11_tensor_contraction "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") ei_add_test(cxx11_tensor_device "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_comparisons.cpp b/unsupported/test/cxx11_tensor_comparisons.cpp new file mode 100644 index 000000000..186f56ac3 --- /dev/null +++ b/unsupported/test/cxx11_tensor_comparisons.cpp @@ -0,0 +1,84 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +static void test_orderings() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + Tensor lt(2,3,7); + Tensor le(2,3,7); + Tensor gt(2,3,7); + Tensor ge(2,3,7); + + mat1.setRandom(); + mat2.setRandom(); + + lt = mat1 < mat2; + le = mat1 <= mat2; + gt = mat1 > mat2; + ge = mat1 >= mat2; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(lt(i,j,k), mat1(i,j,k) < mat2(i,j,k)); + VERIFY_IS_EQUAL(le(i,j,k), mat1(i,j,k) <= mat2(i,j,k)); + VERIFY_IS_EQUAL(gt(i,j,k), mat1(i,j,k) > mat2(i,j,k)); + VERIFY_IS_EQUAL(ge(i,j,k), mat1(i,j,k) >= mat2(i,j,k)); + } + } + } +} + + +static void test_equality() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + + mat1.setRandom(); + mat2.setRandom(); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + if (random() < 0.5) { + mat2(i,j,k) = mat1(i,j,k); + } + } + } + } + + Tensor eq(2,3,7); + Tensor ne(2,3,7); + eq = (mat1 == mat2); + ne = (mat1 != mat2); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(eq(i,j,k), mat1(i,j,k) == mat2(i,j,k)); + VERIFY_IS_EQUAL(ne(i,j,k), mat1(i,j,k) != mat2(i,j,k)); + } + } + } +} + + +void test_cxx11_tensor_comparisons() +{ + CALL_SUBTEST(test_orderings()); + CALL_SUBTEST(test_equality()); +} diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp new file mode 100644 index 000000000..1c89dfdd1 --- /dev/null +++ b/unsupported/test/cxx11_tensor_contraction.cpp @@ -0,0 +1,163 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +typedef Tensor::DimensionPair DimPair; + + +static void test_evals() +{ + Tensor mat1(2, 3); + Tensor mat2(2, 3); + Tensor mat3(3, 2); + + mat1.setRandom(); + mat2.setRandom(); + mat3.setRandom(); + + Tensor mat4(3,3); + mat4.setZero(); + Eigen::array dims3({{DimPair(0, 0)}}); + TensorEvaluator eval(mat1.contract(mat2, dims3)); + eval.evalTo(mat4.data()); + EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + VERIFY_IS_EQUAL(eval.dimensions()[0], 3); + VERIFY_IS_EQUAL(eval.dimensions()[1], 3); + + VERIFY_IS_APPROX(mat4(0,0), mat1(0,0)*mat2(0,0) + mat1(1,0)*mat2(1,0)); + VERIFY_IS_APPROX(mat4(0,1), mat1(0,0)*mat2(0,1) + mat1(1,0)*mat2(1,1)); + VERIFY_IS_APPROX(mat4(0,2), mat1(0,0)*mat2(0,2) + mat1(1,0)*mat2(1,2)); + VERIFY_IS_APPROX(mat4(1,0), mat1(0,1)*mat2(0,0) + mat1(1,1)*mat2(1,0)); + VERIFY_IS_APPROX(mat4(1,1), mat1(0,1)*mat2(0,1) + mat1(1,1)*mat2(1,1)); + VERIFY_IS_APPROX(mat4(1,2), mat1(0,1)*mat2(0,2) + mat1(1,1)*mat2(1,2)); + VERIFY_IS_APPROX(mat4(2,0), mat1(0,2)*mat2(0,0) + mat1(1,2)*mat2(1,0)); + VERIFY_IS_APPROX(mat4(2,1), mat1(0,2)*mat2(0,1) + mat1(1,2)*mat2(1,1)); + VERIFY_IS_APPROX(mat4(2,2), mat1(0,2)*mat2(0,2) + mat1(1,2)*mat2(1,2)); + + Tensor mat5(2,2); + mat5.setZero(); + Eigen::array dims4({{DimPair(1, 1)}}); + TensorEvaluator eval2(mat1.contract(mat2, dims4)); + eval2.evalTo(mat5.data()); + EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + VERIFY_IS_EQUAL(eval2.dimensions()[0], 2); + VERIFY_IS_EQUAL(eval2.dimensions()[1], 2); + + VERIFY_IS_APPROX(mat5(0,0), mat1(0,0)*mat2(0,0) + mat1(0,1)*mat2(0,1) + mat1(0,2)*mat2(0,2)); + VERIFY_IS_APPROX(mat5(0,1), mat1(0,0)*mat2(1,0) + mat1(0,1)*mat2(1,1) + mat1(0,2)*mat2(1,2)); + VERIFY_IS_APPROX(mat5(1,0), mat1(1,0)*mat2(0,0) + mat1(1,1)*mat2(0,1) + mat1(1,2)*mat2(0,2)); + VERIFY_IS_APPROX(mat5(1,1), mat1(1,0)*mat2(1,0) + mat1(1,1)*mat2(1,1) + mat1(1,2)*mat2(1,2)); + + Tensor mat6(2,2); + mat6.setZero(); + Eigen::array dims6({{DimPair(1, 0)}}); + TensorEvaluator eval3(mat1.contract(mat3, dims6)); + eval3.evalTo(mat6.data()); + EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + VERIFY_IS_EQUAL(eval3.dimensions()[0], 2); + VERIFY_IS_EQUAL(eval3.dimensions()[1], 2); + + VERIFY_IS_APPROX(mat6(0,0), mat1(0,0)*mat3(0,0) + mat1(0,1)*mat3(1,0) + mat1(0,2)*mat3(2,0)); + VERIFY_IS_APPROX(mat6(0,1), mat1(0,0)*mat3(0,1) + mat1(0,1)*mat3(1,1) + mat1(0,2)*mat3(2,1)); + VERIFY_IS_APPROX(mat6(1,0), mat1(1,0)*mat3(0,0) + mat1(1,1)*mat3(1,0) + mat1(1,2)*mat3(2,0)); + VERIFY_IS_APPROX(mat6(1,1), mat1(1,0)*mat3(0,1) + mat1(1,1)*mat3(1,1) + mat1(1,2)*mat3(2,1)); +} + + +static void test_scalar() +{ + Tensor vec1({6}); + Tensor vec2({6}); + + vec1.setRandom(); + vec2.setRandom(); + + Tensor scalar(1); + scalar.setZero(); + Eigen::array dims({{DimPair(0, 0)}}); + TensorEvaluator eval(vec1.contract(vec2, dims)); + eval.evalTo(scalar.data()); + EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==1ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + + float expected = 0.0f; + for (int i = 0; i < 6; ++i) { + expected += vec1(i) * vec2(i); + } + VERIFY_IS_APPROX(scalar(0), expected); +} + + +static void test_multidims() +{ + Tensor mat1(2, 2, 2); + Tensor mat2(2, 2, 2, 2); + + mat1.setRandom(); + mat2.setRandom(); + + Tensor mat3(2, 2, 2); + mat3.setZero(); + Eigen::array dims({{DimPair(1, 2), DimPair(2, 3)}}); + TensorEvaluator eval(mat1.contract(mat2, dims)); + eval.evalTo(mat3.data()); + EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==3ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + VERIFY_IS_EQUAL(eval.dimensions()[0], 2); + VERIFY_IS_EQUAL(eval.dimensions()[1], 2); + VERIFY_IS_EQUAL(eval.dimensions()[2], 2); + + VERIFY_IS_APPROX(mat3(0,0,0), mat1(0,0,0)*mat2(0,0,0,0) + mat1(0,1,0)*mat2(0,0,1,0) + + mat1(0,0,1)*mat2(0,0,0,1) + mat1(0,1,1)*mat2(0,0,1,1)); + VERIFY_IS_APPROX(mat3(0,0,1), mat1(0,0,0)*mat2(0,1,0,0) + mat1(0,1,0)*mat2(0,1,1,0) + + mat1(0,0,1)*mat2(0,1,0,1) + mat1(0,1,1)*mat2(0,1,1,1)); + VERIFY_IS_APPROX(mat3(0,1,0), mat1(0,0,0)*mat2(1,0,0,0) + mat1(0,1,0)*mat2(1,0,1,0) + + mat1(0,0,1)*mat2(1,0,0,1) + mat1(0,1,1)*mat2(1,0,1,1)); + VERIFY_IS_APPROX(mat3(0,1,1), mat1(0,0,0)*mat2(1,1,0,0) + mat1(0,1,0)*mat2(1,1,1,0) + + mat1(0,0,1)*mat2(1,1,0,1) + mat1(0,1,1)*mat2(1,1,1,1)); + VERIFY_IS_APPROX(mat3(1,0,0), mat1(1,0,0)*mat2(0,0,0,0) + mat1(1,1,0)*mat2(0,0,1,0) + + mat1(1,0,1)*mat2(0,0,0,1) + mat1(1,1,1)*mat2(0,0,1,1)); + VERIFY_IS_APPROX(mat3(1,0,1), mat1(1,0,0)*mat2(0,1,0,0) + mat1(1,1,0)*mat2(0,1,1,0) + + mat1(1,0,1)*mat2(0,1,0,1) + mat1(1,1,1)*mat2(0,1,1,1)); + VERIFY_IS_APPROX(mat3(1,1,0), mat1(1,0,0)*mat2(1,0,0,0) + mat1(1,1,0)*mat2(1,0,1,0) + + mat1(1,0,1)*mat2(1,0,0,1) + mat1(1,1,1)*mat2(1,0,1,1)); + VERIFY_IS_APPROX(mat3(1,1,1), mat1(1,0,0)*mat2(1,1,0,0) + mat1(1,1,0)*mat2(1,1,1,0) + + mat1(1,0,1)*mat2(1,1,0,1) + mat1(1,1,1)*mat2(1,1,1,1)); +} + + +static void test_expr() +{ + Tensor mat1(2, 3); + Tensor mat2(3, 2); + mat1.setRandom(); + mat2.setRandom(); + + Tensor mat3(2,2); + + Eigen::array dims({{DimPair(1, 0)}}); + mat3 = mat1.contract(mat2, dims); + + VERIFY_IS_APPROX(mat3(0,0), mat1(0,0)*mat2(0,0) + mat1(0,1)*mat2(1,0) + mat1(0,2)*mat2(2,0)); + VERIFY_IS_APPROX(mat3(0,1), mat1(0,0)*mat2(0,1) + mat1(0,1)*mat2(1,1) + mat1(0,2)*mat2(2,1)); + VERIFY_IS_APPROX(mat3(1,0), mat1(1,0)*mat2(0,0) + mat1(1,1)*mat2(1,0) + mat1(1,2)*mat2(2,0)); + VERIFY_IS_APPROX(mat3(1,1), mat1(1,0)*mat2(0,1) + mat1(1,1)*mat2(1,1) + mat1(1,2)*mat2(2,1)); +} + + +void test_cxx11_tensor_contraction() +{ + CALL_SUBTEST(test_evals()); + CALL_SUBTEST(test_scalar()); + CALL_SUBTEST(test_multidims()); + CALL_SUBTEST(test_expr()); +} diff --git a/unsupported/test/cxx11_tensor_device.cpp b/unsupported/test/cxx11_tensor_device.cpp index 9eb1d0420..365b109c7 100644 --- a/unsupported/test/cxx11_tensor_device.cpp +++ b/unsupported/test/cxx11_tensor_device.cpp @@ -15,7 +15,7 @@ #include "main.h" -#include +#include using Eigen::Tensor; using Eigen::RowMajor; @@ -39,8 +39,12 @@ struct CPUContext { // Context for evaluation on GPU struct GPUContext { - GPUContext(const Eigen::TensorMap >& in1, Eigen::TensorMap >& in2, Eigen::TensorMap >& out) : in1_(in1), in2_(in2), out_(out) { } - + GPUContext(const Eigen::TensorMap >& in1, Eigen::TensorMap >& in2, Eigen::TensorMap >& out) : in1_(in1), in2_(in2), out_(out), gpu_device_(&stream_) { + cudaStreamCreate(&stream_); + } + ~GPUContext() { + cudaStreamDestroy(stream_); + } const Eigen::TensorMap >& in1() const { return in1_; } const Eigen::TensorMap >& in2() const { return in2_; } Eigen::TensorDevice >, Eigen::GpuDevice> out() { return TensorDevice >, Eigen::GpuDevice>(gpu_device_, out_); } @@ -49,6 +53,7 @@ struct GPUContext { const Eigen::TensorMap >& in1_; const Eigen::TensorMap >& in2_; Eigen::TensorMap >& out_; + cudaStream_t stream_; Eigen::GpuDevice gpu_device_; }; @@ -57,7 +62,7 @@ struct GPUContext { template static void test_contextual_eval(Context* context) { - context->out() = context->in1() + context->in2() * 3.14f; + context->out() = context->in1() + context->in2() * 3.14f + context->in1().constant(2.718f); } static void test_cpu() { @@ -73,7 +78,7 @@ static void test_cpu() { for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f); + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); } } } @@ -111,7 +116,7 @@ static void test_gpu() { for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f); + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); } } } diff --git a/unsupported/test/cxx11_tensor_expr.cpp b/unsupported/test/cxx11_tensor_expr.cpp index e0124da8c..e85fcbfa9 100644 --- a/unsupported/test/cxx11_tensor_expr.cpp +++ b/unsupported/test/cxx11_tensor_expr.cpp @@ -28,10 +28,10 @@ static void test_1d() float data3[6]; TensorMap> vec3(data3, 6); - vec3 = vec1.cwiseSqrt(); + vec3 = vec1.sqrt(); float data4[6]; TensorMap> vec4(data4, 6); - vec4 = vec2.cwiseSqrt(); + vec4 = vec2.square(); VERIFY_IS_APPROX(vec3(0), sqrtf(4.0)); VERIFY_IS_APPROX(vec3(1), sqrtf(8.0)); @@ -40,12 +40,12 @@ static void test_1d() VERIFY_IS_APPROX(vec3(4), sqrtf(23.0)); VERIFY_IS_APPROX(vec3(5), sqrtf(42.0)); - VERIFY_IS_APPROX(vec4(0), sqrtf(0.0)); - VERIFY_IS_APPROX(vec4(1), sqrtf(1.0)); - VERIFY_IS_APPROX(vec4(2), sqrtf(2.0)); - VERIFY_IS_APPROX(vec4(3), sqrtf(3.0)); - VERIFY_IS_APPROX(vec4(4), sqrtf(4.0)); - VERIFY_IS_APPROX(vec4(5), sqrtf(5.0)); + VERIFY_IS_APPROX(vec4(0), 0.0f); + VERIFY_IS_APPROX(vec4(1), 1.0f); + VERIFY_IS_APPROX(vec4(2), 2.0f * 2.0f); + VERIFY_IS_APPROX(vec4(3), 3.0f * 3.0f); + VERIFY_IS_APPROX(vec4(4), 4.0f * 4.0f); + VERIFY_IS_APPROX(vec4(5), 5.0f * 5.0f); vec3 = vec1 + vec2; VERIFY_IS_APPROX(vec3(0), 4.0f + 0.0f); @@ -79,8 +79,8 @@ static void test_2d() Tensor mat3(2,3); Tensor mat4(2,3); - mat3 = mat1.cwiseAbs(); - mat4 = mat2.cwiseAbs(); + mat3 = mat1.abs(); + mat4 = mat2.abs(); VERIFY_IS_APPROX(mat3(0,0), 0.0f); VERIFY_IS_APPROX(mat3(0,1), 1.0f); @@ -102,7 +102,7 @@ static void test_3d() Tensor mat1(2,3,7); Tensor mat2(2,3,7); - float val = 0.0; + float val = 1.0; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { @@ -118,28 +118,147 @@ static void test_3d() Tensor mat4(2,3,7); mat4 = mat2 * 3.14f; Tensor mat5(2,3,7); - mat5 = mat1.cwiseSqrt().cwiseSqrt(); + mat5 = mat1.inverse().log(); Tensor mat6(2,3,7); - mat6 = mat2.cwiseSqrt() * 3.14f; + mat6 = mat2.pow(0.5f) * 3.14f; + Tensor mat7(2,3,7); + mat7 = mat1.cwiseMax(mat5 * 2.0f).exp(); + Tensor mat8(2,3,7); + mat8 = (-mat2).exp() * 3.14f; - val = 0.0; + val = 1.0; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { VERIFY_IS_APPROX(mat3(i,j,k), val + val); VERIFY_IS_APPROX(mat4(i,j,k), val * 3.14f); - VERIFY_IS_APPROX(mat5(i,j,k), sqrtf(sqrtf(val))); + VERIFY_IS_APPROX(mat5(i,j,k), logf(1.0f/val)); VERIFY_IS_APPROX(mat6(i,j,k), sqrtf(val) * 3.14f); + VERIFY_IS_APPROX(mat7(i,j,k), expf((std::max)(val, mat5(i,j,k) * 2.0f))); + VERIFY_IS_APPROX(mat8(i,j,k), expf(-val) * 3.14f); val += 1.0; } } } } +static void test_constants() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + Tensor mat3(2,3,7); + + float val = 1.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + mat1(i,j,k) = val; + val += 1.0; + } + } + } + mat2 = mat1.constant(3.14f); + mat3 = mat1.cwiseMax(7.3f).exp(); + + val = 1.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat2(i,j,k), 3.14f); + VERIFY_IS_APPROX(mat3(i,j,k), expf((std::max)(val, 7.3f))); + val += 1.0; + } + } + } +} + + +static void test_functors() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + Tensor mat3(2,3,7); + + float val = 1.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + mat1(i,j,k) = val; + val += 1.0; + } + } + } + mat2 = mat1.inverse().unaryExpr(&asinf); + mat3 = mat1.unaryExpr(&tanhf); + + val = 1.0; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat2(i,j,k), asinf(1.0f / mat1(i,j,k))); + VERIFY_IS_APPROX(mat3(i,j,k), tanhf(mat1(i,j,k))); + val += 1.0; + } + } + } +} + +static void test_type_casting() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + Tensor mat3(2,3,7); + mat1.setRandom(); + mat2.setRandom(); + + mat3 = mat1.template cast(); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat3(i,j,k), mat1(i,j,k) ? 1.0 : 0.0); + } + } + } + + mat3 = mat2.template cast(); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat3(i,j,k), static_cast(mat2(i,j,k))); + } + } + } +} + +static void test_select() +{ + Tensor selector(2,3,7); + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + Tensor result(2,3,7); + + selector.setRandom(); + mat1.setRandom(); + mat2.setRandom(); + result = (selector > selector.constant(0.5f)).select(mat1, mat2); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(result(i,j,k), (selector(i,j,k) > 0.5f) ? mat1(i,j,k) : mat2(i,j,k)); + } + } + } +} + void test_cxx11_tensor_expr() { CALL_SUBTEST(test_1d()); CALL_SUBTEST(test_2d()); CALL_SUBTEST(test_3d()); + CALL_SUBTEST(test_constants()); + CALL_SUBTEST(test_functors()); + CALL_SUBTEST(test_type_casting()); + CALL_SUBTEST(test_select()); } diff --git a/unsupported/test/cxx11_tensor_fixed_size.cpp b/unsupported/test/cxx11_tensor_fixed_size.cpp index 214f6951d..d270486f2 100644 --- a/unsupported/test/cxx11_tensor_fixed_size.cpp +++ b/unsupported/test/cxx11_tensor_fixed_size.cpp @@ -33,10 +33,10 @@ static void test_1d() float data3[6]; TensorMap > > vec3(data3, 6); - vec3 = vec1.cwiseSqrt(); + vec3 = vec1.sqrt(); float data4[6]; TensorMap, RowMajor> > vec4(data4, 6); - vec4 = vec2.cwiseSqrt(); + vec4 = vec2.sqrt(); VERIFY_IS_EQUAL((vec3.size()), 6); // VERIFY_IS_EQUAL((vec3.dimensions()[0]), 6); @@ -92,8 +92,8 @@ static void test_2d() TensorFixedSize> mat3; TensorFixedSize, RowMajor> mat4; - mat3 = mat1.cwiseAbs(); - mat4 = mat2.cwiseAbs(); + mat3 = mat1.abs(); + mat4 = mat2.abs(); VERIFY_IS_EQUAL((mat3.size()), 2*3); // VERIFY_IS_EQUAL((mat3.dimension(0)), 2); @@ -136,9 +136,9 @@ static void test_3d() } TensorFixedSize > mat3; - mat3 = mat1.cwiseSqrt(); + mat3 = mat1.sqrt(); TensorFixedSize, RowMajor> mat4; - mat4 = mat2.cwiseSqrt(); + mat4 = mat2.sqrt(); VERIFY_IS_EQUAL((mat3.size()), 2*3*7); // VERIFY_IS_EQUAL((mat3.dimension(0)), 2); @@ -173,7 +173,7 @@ static void test_array() } TensorFixedSize > mat3; - mat3 = mat1.cwisePow(3.5f); + mat3 = mat1.pow(3.5f); val = 0.0; for (int i = 0; i < 2; ++i) { diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp index c9de71da3..b371e8a71 100644 --- a/unsupported/test/cxx11_tensor_thread_pool.cpp +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -12,6 +12,7 @@ #include "main.h" #include +#include "thread/threadpool.h" using Eigen::Tensor; @@ -24,8 +25,10 @@ void test_cxx11_tensor_thread_pool() in1.setRandom(); in2.setRandom(); - Eigen::ThreadPoolDevice thread_pool_device(3); - out.device(thread_pool_device) = in1 + in2 * 3.14; + ThreadPool thread_pool(2); + thread_pool.StartWorkers(); + Eigen::ThreadPoolDevice thread_pool_device(&thread_pool, 3); + out.device(thread_pool_device) = in1 + in2 * 3.14f; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { From ed37c44765b4401629281a1ea7ae223cddf91fde Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 6 Jun 2014 11:02:20 +0200 Subject: [PATCH 0511/1103] Enable LinearAccessBit in Block expression for inner-panels --- Eigen/src/Core/Block.h | 2 +- Eigen/src/Core/MapBase.h | 2 ++ Eigen/src/Geometry/Transform.h | 4 ++-- 3 files changed, 5 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index e948e14aa..da193d1a2 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -84,7 +84,7 @@ struct traits > : traits::Flags&LinearAccessBit))) ? LinearAccessBit : 0, FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, Flags0 = traits::Flags & ( (HereditaryBits & ~RowMajorBit) | diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index a45a0b374..e8ecb175b 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -250,6 +250,8 @@ template class MapBase using Base::Base::operator=; }; +#undef EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS + } // end namespace Eigen #endif // EIGEN_MAPBASE_H diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index f40644011..cb93acf6b 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -194,9 +194,9 @@ public: /** type of the matrix used to represent the linear part of the transformation */ typedef Matrix LinearMatrixType; /** type of read/write reference to the linear part of the transformation */ - typedef Block LinearPart; + typedef Block LinearPart; /** type of read reference to the linear part of the transformation */ - typedef const Block ConstLinearPart; + typedef const Block ConstLinearPart; /** type of read/write reference to the affine part of the transformation */ typedef typename internal::conditional Date: Fri, 6 Jun 2014 11:06:44 +0200 Subject: [PATCH 0512/1103] Fix bug #738: use the "current" version of cmake project directories to ease the inclusion of Eigen within other projects. --- CMakeLists.txt | 4 ++-- cmake/EigenConfigureTesting.cmake | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index fb13769f4..a719e47fd 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -449,12 +449,12 @@ set ( EIGEN_INCLUDE_DIR ${INCLUDE_INSTALL_DIR} ) set ( EIGEN_INCLUDE_DIRS ${EIGEN_INCLUDE_DIR} ) set ( EIGEN_ROOT_DIR ${CMAKE_INSTALL_PREFIX} ) -configure_file ( ${CMAKE_SOURCE_DIR}/cmake/Eigen3Config.cmake.in +configure_file ( ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Eigen3Config.cmake.in ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake @ONLY ESCAPE_QUOTES ) -install ( FILES ${CMAKE_SOURCE_DIR}/cmake/UseEigen3.cmake +install ( FILES ${CMAKE_CURRENT_SOURCE_DIR}/cmake/UseEigen3.cmake ${CMAKE_CURRENT_BINARY_DIR}/Eigen3Config.cmake DESTINATION ${EIGEN_CONFIG_CMAKE_PATH} ) diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake index 7844bf4d3..0b5de95bb 100644 --- a/cmake/EigenConfigureTesting.cmake +++ b/cmake/EigenConfigureTesting.cmake @@ -49,7 +49,7 @@ else() set(EIGEN_MAKECOMMAND_PLACEHOLDER "${EIGEN_BUILD_COMMAND}") endif() -configure_file(${CMAKE_BINARY_DIR}/DartConfiguration.tcl ${CMAKE_BINARY_DIR}/DartConfiguration.tcl) +configure_file(${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl ${CMAKE_BINARY_DIR}/DartConfiguration.tcl) # restore default CMAKE_MAKE_PROGRAM set(CMAKE_MAKE_PROGRAM ${CMAKE_MAKE_PROGRAM_SAVE}) @@ -58,7 +58,7 @@ set(CMAKE_MAKE_PROGRAM ${CMAKE_MAKE_PROGRAM_SAVE}) unset(CMAKE_MAKE_PROGRAM_SAVE) unset(EIGEN_MAKECOMMAND_PLACEHOLDER) -configure_file(${CMAKE_SOURCE_DIR}/CTestCustom.cmake.in ${CMAKE_BINARY_DIR}/CTestCustom.cmake) +configure_file(${CMAKE_CURRENT_SOURCE_DIR}/CTestCustom.cmake.in ${CMAKE_BINARY_DIR}/CTestCustom.cmake) # some documentation of this function would be nice ei_init_testing() From abc1ca0af14872fe44e583faa2b43e496b038f8a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 6 Jun 2014 11:21:19 +0200 Subject: [PATCH 0513/1103] The BLAS interface is complete. --- blas/README.txt | 3 --- 1 file changed, 3 deletions(-) diff --git a/blas/README.txt b/blas/README.txt index 07a8bd92a..63a5203b9 100644 --- a/blas/README.txt +++ b/blas/README.txt @@ -1,9 +1,6 @@ This directory contains a BLAS library built on top of Eigen. -This is currently a work in progress which is far to be ready for use, -but feel free to contribute to it if you wish. - This module is not built by default. In order to compile it, you need to type 'make blas' from within your build dir. From a961d72e65fc537fe571845407b4e2ee0554bd49 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 6 Jun 2014 16:25:16 -0700 Subject: [PATCH 0514/1103] Added support for convolution and reshaping of tensors. --- unsupported/Eigen/CXX11/Tensor | 2 + .../Eigen/CXX11/src/Tensor/TensorBase.h | 14 ++ .../CXX11/src/Tensor/TensorConvolution.h | 206 ++++++++++++++++++ .../src/Tensor/TensorForwardDeclarations.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 119 ++++++++++ unsupported/test/cxx11_tensor_convolution.cpp | 70 ++++++ 6 files changed, 413 insertions(+), 2 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h create mode 100644 unsupported/test/cxx11_tensor_convolution.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index d4e8d3a15..c67020581 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -40,6 +40,8 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index c5c711313..932e5c82d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -203,6 +203,13 @@ class TensorBase return TensorContractionOp(derived(), other.derived(), dims); } + // Convolutions. + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorConvolutionOp + convolve(const KernelDerived& kernel, const Dimensions& dims) const { + return TensorConvolutionOp(derived(), kernel.derived(), dims); + } + // Coefficient-wise ternary operators. template inline const TensorSelectOp @@ -210,6 +217,13 @@ class TensorBase return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); } + // Morphing operators (slicing tbd). + template + inline const TensorReshapingOp + reshape(const NewDimensions& newDimensions) const { + return TensorReshapingOp(derived(), newDimensions); + } + // Select the device on which to evaluate the expression. template TensorDevice device(const DeviceType& device) { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h new file mode 100644 index 000000000..ca2e0e562 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -0,0 +1,206 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H +#define EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H + +namespace Eigen { + +/** \class TensorConvolution + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor convolution class. + * + * + */ +namespace internal { +template +struct traits > +{ + // Type promotion to handle the case where the types of the lhs and the rhs are different. + typedef typename internal::promote_storage_type::ret Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename promote_storage_type::StorageKind, + typename traits::StorageKind>::ret StorageKind; + typedef typename promote_index_type::Index, + typename traits::Index>::type Index; + typedef typename InputXprType::Nested LhsNested; + typedef typename KernelXprType::Nested RhsNested; + typedef typename remove_reference::type _LhsNested; + typedef typename remove_reference::type _RhsNested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorConvolutionOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorConvolutionOp type; +}; + +} // end namespace internal + + + +template +class TensorConvolutionOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename internal::promote_storage_type::ret CoeffReturnType; + typedef typename internal::promote_storage_type::ret PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConvolutionOp(const InputXprType& input, const KernelXprType& kernel, const Indices& dims) + : m_input_xpr(input), m_kernel_xpr(kernel), m_indices(dims) {} + + EIGEN_DEVICE_FUNC + const Indices& indices() const { return m_indices; } + + /** \returns the nested expressions */ + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + inputExpression() const { return m_input_xpr; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + kernelExpression() const { return m_kernel_xpr; } + + protected: + typename InputXprType::Nested m_input_xpr; + typename KernelXprType::Nested m_kernel_xpr; + const Indices m_indices; +}; + + +template +struct TensorEvaluator > +{ + typedef TensorConvolutionOp XprType; + + static const int NumDims = TensorEvaluator::Dimensions::count; + static const int KernelDims = Indices::size; + typedef typename XprType::Index Index; + typedef DSizes Dimensions; + + enum { + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = /*TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess */ + false, + }; + + TensorEvaluator(const XprType& op) + : m_inputImpl(op.inputExpression()), m_kernelImpl(op.kernelExpression()), m_dimensions(op.inputExpression().dimensions()) + { + const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); + const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); + + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_inputStride[i] = m_inputStride[i-1] * input_dims[i-1]; + } else { + m_inputStride[0] = 1; + } + } + + for (int i = 0; i < KernelDims; ++i) { + const Index index = op.indices()[i]; + const Index input_dim = input_dims[index]; + const Index kernel_dim = kernel_dims[i]; + const Index result_dim = input_dim - kernel_dim + 1; + m_dimensions[index] = result_dim; + + if (i > 0) { + m_kernelStride[i] = m_kernelStride[i-1] * kernel_dims[i-1]; + } else { + m_kernelStride[0] = 1; + } + m_indexStride[i] = m_inputStride[index]; + } + + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_outputStride[i] = m_outputStride[i-1] * m_dimensions[i-1]; + } else { + m_outputStride[0] = 1; + } + } + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + const Dimensions& dimensions() const { return m_dimensions; } + + void evalTo(typename XprType::Scalar* buffer) const { + for (int i = 0; i < dimensions().TotalSize(); ++i) { + buffer[i] += coeff(i); + } + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Index startInput = 0; + for (int i = NumDims - 1; i >= 0; --i) { + const Index idx = index / m_outputStride[i]; + startInput += idx * m_inputStride[i]; + index -= idx * m_outputStride[i]; + } + + CoeffReturnType result = CoeffReturnType(0); + convolve(startInput, 0, 0, result); + return result; + } + + /* TODO: vectorization + template + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + { + assert(false); + }*/ + + EIGEN_DEVICE_FUNC void convolve(Index firstIndex, Index firstKernel, int DimIndex, CoeffReturnType& accum) const { + for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) { + const Index input = firstIndex + j * m_indexStride[DimIndex]; + const Index kernel = firstKernel + j * m_kernelStride[DimIndex]; + if (DimIndex < KernelDims-1) { + convolve(input, kernel, DimIndex+1, accum); + } else { + + accum += m_inputImpl.coeff(input) * m_kernelImpl.coeff(kernel); + } + } + } + + private: + array m_inputStride; + array m_outputStride; + + array m_indexStride; + array m_kernelStride; + Dimensions m_dimensions; + TensorEvaluator m_inputImpl; + TensorEvaluator m_kernelImpl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 239b5cb67..b8833362c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -21,9 +21,9 @@ template class TensorCwiseNullaryO template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; template class TensorSelectOp; -template class TensorReductionOp; template class TensorContractionOp; - +template class TensorConvolutionOp; +template class TensorReshapingOp; template class TensorDevice; // Move to internal? diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h new file mode 100644 index 000000000..3e089fe1e --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -0,0 +1,119 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H +#define EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H + +namespace Eigen { + +/** \class TensorReshaping + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor reshaping class. + * + * + */ +namespace internal { +template +struct traits > +{ + // Type promotion to handle the case where the types of the lhs and the rhs are different. + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorReshapingOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorReshapingOp type; +}; + +} // end namespace internal + + + +template +class TensorReshapingOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReshapingOp(const XprType& expr, const NewDimensions& dims) + : m_xpr(expr), m_dims(dims) {} + + EIGEN_DEVICE_FUNC + const NewDimensions& dimensions() const { return m_dims; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + protected: + typename XprType::Nested m_xpr; + const NewDimensions m_dims; +}; + + +template +struct TensorEvaluator > +{ + typedef TensorReshapingOp XprType; + + enum { + IsAligned = TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess, + }; + + TensorEvaluator(const XprType& op) + : m_impl(op.expression()), m_dimensions(op.dimensions()) + { } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + const NewDimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + return m_impl.coeff(index); + } + + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return m_impl.template packet(index); + } + + private: + NewDimensions m_dimensions; + TensorEvaluator m_impl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp new file mode 100644 index 000000000..95e40f64f --- /dev/null +++ b/unsupported/test/cxx11_tensor_convolution.cpp @@ -0,0 +1,70 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + + +static void test_evals() +{ + Tensor input(3, 3); + Tensor kernel(2); + + input.setRandom(); + kernel.setRandom(); + + Tensor result(2,3); + result.setZero(); + Eigen::array::Index, 1> dims3({0}); + + TensorEvaluator eval(input.convolve(kernel, dims3)); + eval.evalTo(result.data()); + EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + VERIFY_IS_EQUAL(eval.dimensions()[0], 2); + VERIFY_IS_EQUAL(eval.dimensions()[1], 3); + + VERIFY_IS_APPROX(result(0,0), input(0,0)*kernel(0) + input(1,0)*kernel(1)); // index 0 + VERIFY_IS_APPROX(result(0,1), input(0,1)*kernel(0) + input(1,1)*kernel(1)); // index 2 + VERIFY_IS_APPROX(result(0,2), input(0,2)*kernel(0) + input(1,2)*kernel(1)); // index 4 + VERIFY_IS_APPROX(result(1,0), input(1,0)*kernel(0) + input(2,0)*kernel(1)); // index 1 + VERIFY_IS_APPROX(result(1,1), input(1,1)*kernel(0) + input(2,1)*kernel(1)); // index 3 + VERIFY_IS_APPROX(result(1,2), input(1,2)*kernel(0) + input(2,2)*kernel(1)); // index 5 +} + + +static void test_expr() +{ + Tensor input(3, 3); + Tensor kernel(2, 2); + input.setRandom(); + kernel.setRandom(); + + Tensor result(2,2); + Eigen::array dims({0, 1}); + result = input.convolve(kernel, dims); + + VERIFY_IS_APPROX(result(0,0), input(0,0)*kernel(0,0) + input(0,1)*kernel(0,1) + + input(1,0)*kernel(1,0) + input(1,1)*kernel(1,1)); + VERIFY_IS_APPROX(result(0,1), input(0,1)*kernel(0,0) + input(0,2)*kernel(0,1) + + input(1,1)*kernel(1,0) + input(1,2)*kernel(1,1)); + VERIFY_IS_APPROX(result(1,0), input(1,0)*kernel(0,0) + input(1,1)*kernel(0,1) + + input(2,0)*kernel(1,0) + input(2,1)*kernel(1,1)); + VERIFY_IS_APPROX(result(1,1), input(1,1)*kernel(0,0) + input(1,2)*kernel(0,1) + + input(2,1)*kernel(1,0) + input(2,2)*kernel(1,1)); +} + + +void test_cxx11_tensor_convolution() +{ + CALL_SUBTEST(test_evals()); + CALL_SUBTEST(test_expr()); +} From 79085e08e9512f678b4584df49d1b2835b40117f Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 6 Jun 2014 20:16:13 -0700 Subject: [PATCH 0515/1103] Fixed a typo --- unsupported/test/CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index d6072c9f3..e67e61263 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -102,7 +102,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_simple "-std=c++0x") ei_add_test(cxx11_tensor_symmetry "-std=c++0x") ei_add_test(cxx11_tensor_assign "-std=c++0x") - ei_add_test(cxx11_tensor_comparison "-std=c++0x") + ei_add_test(cxx11_tensor_comparisons "-std=c++0x") ei_add_test(cxx11_tensor_contraction "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") From 29aebf96e62f4fb5e4b1f3fb475e299df2e7a02e Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 6 Jun 2014 20:18:44 -0700 Subject: [PATCH 0516/1103] Created the pblend packet primitive and implemented it using SSE and AVX instructions. --- Eigen/src/Core/GenericPacketMath.h | 14 ++++++++++ Eigen/src/Core/arch/AVX/PacketMath.h | 15 ++++++++++ Eigen/src/Core/arch/SSE/Complex.h | 8 +++++- Eigen/src/Core/arch/SSE/PacketMath.h | 41 ++++++++++++++++++++++++++-- test/packetmath.cpp | 16 +++++++++++ 5 files changed, 90 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 98313c68f..0869dd49f 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -54,6 +54,7 @@ struct default_packet_traits HasMax = 1, HasConj = 1, HasSetLinear = 1, + HasBlend = 0, HasDiv = 0, HasSqrt = 0, @@ -429,6 +430,19 @@ ptranspose(PacketBlock& /*kernel*/) { // Nothing to do in the scalar case, i.e. a 1x1 matrix. } +/*************************************************************************** + * Selector, i.e. vector of N boolean values used to select (i.e. blend) + * words from 2 packets. +***************************************************************************/ +template struct Selector { + bool select[N]; +}; + +template EIGEN_DEVICE_FUNC inline Packet +pblend(const Selector::size>& ifPacket, const Packet& thenPacket, const Packet& elsePacket) { + return ifPacket.select[0] ? thenPacket : elsePacket; +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 8b8307d75..688ff91e4 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -59,6 +59,7 @@ template<> struct packet_traits : default_packet_traits HasLog = 0, HasExp = 0, HasSqrt = 0 + HasBlend = 1, }; }; template<> struct packet_traits : default_packet_traits @@ -73,6 +74,7 @@ template<> struct packet_traits : default_packet_traits HasDiv = 1, HasExp = 0 + HasBlend = 1, }; }; @@ -557,6 +559,19 @@ ptranspose(PacketBlock& kernel) { kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49); } +template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) { + const __m256 zero = _mm256_setzero_ps(); + const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ); + return _mm256_blendv_ps(thenPacket, elsePacket, false_mask); +} +template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) { + const __m256d zero = _mm256_setzero_pd(); + const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ); + return _mm256_blendv_pd(thenPacket, elsePacket, false_mask); +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 758183c18..0bc03cf9e 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -44,7 +44,8 @@ template<> struct packet_traits > : default_packet_traits HasAbs2 = 0, HasMin = 0, HasMax = 0, - HasSetLinear = 0 + HasSetLinear = 0, + HasBlend = 1 }; }; #endif @@ -472,6 +473,11 @@ ptranspose(PacketBlock& kernel) { kernel.packet[1].v = tmp; } +template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) { + __m128d result = pblend(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v)); + return Packet2cf(_mm_castpd_ps(result)); +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 6912f3bc3..1124b24df 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -108,7 +108,8 @@ template<> struct packet_traits : default_packet_traits HasCos = EIGEN_FAST_MATH, HasLog = 1, HasExp = 1, - HasSqrt = 1 + HasSqrt = 1, + HasBlend = 1 }; }; template<> struct packet_traits : default_packet_traits @@ -123,7 +124,8 @@ template<> struct packet_traits : default_packet_traits HasDiv = 1, HasExp = 1, - HasSqrt = 1 + HasSqrt = 1, + HasBlend = 1 }; }; #endif @@ -135,7 +137,9 @@ template<> struct packet_traits : default_packet_traits // FIXME check the Has* Vectorizable = 1, AlignedOnScalar = 1, - size=4 + size=4, + + HasBlend = 1 }; }; @@ -809,6 +813,37 @@ ptranspose(PacketBlock& kernel) { kernel.packet[3] = _mm_unpackhi_epi64(T2, T3); } +template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { + const __m128i zero = _mm_setzero_si128(); + const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m128i false_mask = _mm_cmpeq_epi32(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_epi8(thenPacket, elsePacket, false_mask); +#else + return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket)); +#endif +} +template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { + const __m128 zero = _mm_setzero_ps(); + const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); + __m128 false_mask = _mm_cmpeq_ps(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_ps(thenPacket, elsePacket, false_mask); +#else + return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket)); +#endif +} +template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { + const __m128d zero = _mm_setzero_pd(); + const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]); + __m128d false_mask = _mm_cmpeq_pd(select, zero); +#ifdef EIGEN_VECTORIZE_SSE4_1 + return _mm_blendv_pd(thenPacket, elsePacket, false_mask); +#else + return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket)); +#endif +} + } // end namespace internal } // end namespace Eigen diff --git a/test/packetmath.cpp b/test/packetmath.cpp index 9dab07522..663ab886d 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -261,6 +261,22 @@ template void packetmath() VERIFY(isApproxAbs(data2[j], data1[i+j*PacketSize], refvalue) && "ptranspose"); } } + + if (internal::packet_traits::HasBlend) { + Packet thenPacket = internal::pload(data1); + Packet elsePacket = internal::pload(data2); + EIGEN_ALIGN_DEFAULT internal::Selector selector; + for (int i = 0; i < PacketSize; ++i) { + selector.select[i] = i; + } + + Packet blend = internal::pblend(selector, thenPacket, elsePacket); + EIGEN_ALIGN_DEFAULT Scalar result[size]; + internal::pstore(result, blend); + for (int i = 0; i < PacketSize; ++i) { + VERIFY(isApproxAbs(result[i], (selector.select[i] ? data1[i] : data2[i]), refvalue)); + } + } } template void packetmath_real() From 8c8ae2d8193809744f5952713287639817e2b442 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Sat, 7 Jun 2014 11:24:38 -0700 Subject: [PATCH 0517/1103] Fixed a typo --- Eigen/src/Core/arch/AVX/PacketMath.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 688ff91e4..74d3746d9 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -58,8 +58,8 @@ template<> struct packet_traits : default_packet_traits HasCos = 0, HasLog = 0, HasExp = 0, - HasSqrt = 0 - HasBlend = 1, + HasSqrt = 0, + HasBlend = 1 }; }; template<> struct packet_traits : default_packet_traits @@ -73,8 +73,8 @@ template<> struct packet_traits : default_packet_traits HasHalfPacket = 1, HasDiv = 1, - HasExp = 0 - HasBlend = 1, + HasExp = 0, + HasBlend = 1 }; }; From fe102248ac8f78e33064caeb5cdea6fc41af637c Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 9 Jun 2014 09:19:21 -0700 Subject: [PATCH 0518/1103] Fixed the threadpool test --- unsupported/test/cxx11_tensor_thread_pool.cpp | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp index b371e8a71..2e67b2064 100644 --- a/unsupported/test/cxx11_tensor_thread_pool.cpp +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -12,7 +12,6 @@ #include "main.h" #include -#include "thread/threadpool.h" using Eigen::Tensor; @@ -25,9 +24,7 @@ void test_cxx11_tensor_thread_pool() in1.setRandom(); in2.setRandom(); - ThreadPool thread_pool(2); - thread_pool.StartWorkers(); - Eigen::ThreadPoolDevice thread_pool_device(&thread_pool, 3); + Eigen::ThreadPoolDevice thread_pool_device(3); out.device(thread_pool_device) = in1 + in2 * 3.14f; for (int i = 0; i < 2; ++i) { From 2859a31ac80af86fa58e5347be50d32fd07bcd3c Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 9 Jun 2014 09:42:34 -0700 Subject: [PATCH 0519/1103] Fixed compilation error --- unsupported/Eigen/CXX11/src/Tensor/TensorMap.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index 3a2ff5b30..3a06170fa 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -54,18 +54,18 @@ template class TensorMap : public Tensor PacketAccess = true, }; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions(array(firstDimension)) { - // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. - EIGEN_STATIC_ASSERT(1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) - } - #ifdef EIGEN_HAS_VARIADIC_TEMPLATES template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions(array({{firstDimension, otherDimensions...}})) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) } +#else + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions(array(firstDimension)) { + // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. + EIGEN_STATIC_ASSERT(1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + } #endif inline TensorMap(PointerArgType dataPtr, const array& dimensions) From 36a2b2e9dc9368356b3f327a1fb00616397c1e0e Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 9 Jun 2014 09:43:51 -0700 Subject: [PATCH 0520/1103] Prevent the generation of unlaunchable cuda kernels when compiling in debug mode. --- unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index c92b8c679..3e5687915 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -73,7 +73,7 @@ struct Sizes : internal::numeric_list { typedef internal::numeric_list Base; static const std::size_t total_size = internal::arg_prod(Indices...); - static std::size_t TotalSize() { + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t TotalSize() { return internal::arg_prod(Indices...); } @@ -119,7 +119,7 @@ template ::value; - static size_t TotalSize() { + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() { return internal::arg_prod::value; } @@ -156,7 +156,8 @@ namespace internal { template struct tensor_index_linearization_helper { - static inline Index run(array const& indices, array const& dimensions) + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index run(array const& indices, array const& dimensions) { return array_get(indices) + array_get(dimensions) * @@ -167,7 +168,8 @@ struct tensor_index_linearization_helper template struct tensor_index_linearization_helper { - static inline Index run(array const& indices, array const&) + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index run(array const& indices, array const&) { return array_get(indices); } From a669052f12d6d71ba815764d6419726d64fef675 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 9 Jun 2014 09:45:30 -0700 Subject: [PATCH 0521/1103] Improved support for rvalues in tensor expressions. --- .../Eigen/CXX11/src/Tensor/TensorBase.h | 58 ++++++++++++++----- .../CXX11/src/Tensor/TensorContraction.h | 4 ++ .../CXX11/src/Tensor/TensorConvolution.h | 4 ++ .../Eigen/CXX11/src/Tensor/TensorExpr.h | 8 +++ .../src/Tensor/TensorForwardDeclarations.h | 6 +- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 5 +- .../Eigen/CXX11/src/Tensor/TensorTraits.h | 6 +- 7 files changed, 71 insertions(+), 20 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 932e5c82d..e447a5d40 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -22,7 +22,7 @@ namespace Eigen { */ template -class TensorBase +class TensorBase { public: typedef typename internal::traits::Scalar Scalar; @@ -30,19 +30,6 @@ class TensorBase typedef Scalar CoeffReturnType; typedef typename internal::packet_traits::type PacketReturnType; - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Derived& setZero() { - return setConstant(Scalar(0)); - } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Derived& setConstant(const Scalar& val) { - return derived() = constant(val); - } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Derived& setRandom() { - return derived() = random(); - } - // Nullary operators EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> @@ -224,14 +211,53 @@ class TensorBase return TensorReshapingOp(derived(), newDimensions); } + protected: + template friend class TensorBase; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Derived& derived() const { return *static_cast(this); } +}; + + +template +class TensorBase : public TensorBase { + public: + typedef typename internal::traits::Scalar Scalar; + typedef typename internal::traits::Index Index; + typedef Scalar CoeffReturnType; + typedef typename internal::packet_traits::type PacketReturnType; + + template friend class TensorBase; + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setZero() { + return setConstant(Scalar(0)); + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setConstant(const Scalar& val) { + return derived() = this->constant(val); + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setRandom() { + return derived() = this->random(); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator+=(const OtherDerived& other) { + return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator-=(const OtherDerived& other) { + return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + // Select the device on which to evaluate the expression. template TensorDevice device(const DeviceType& device) { return TensorDevice(device, derived()); } - protected: - template friend class TensorBase; + protected: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& derived() { return *static_cast(this); } EIGEN_DEVICE_FUNC diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index d424df36e..d371eb76d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -35,6 +35,10 @@ struct traits > typedef typename RhsXprType::Nested RhsNested; typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; + + enum { + Flags = 0, + }; }; template diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index ca2e0e562..501e9a522 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -35,6 +35,10 @@ struct traits > typedef typename KernelXprType::Nested RhsNested; typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; + + enum { + Flags = 0, + }; }; template diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h index 60908ee94..de66da13f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h @@ -36,6 +36,10 @@ struct traits > typedef typename XprType::Scalar Scalar; typedef typename XprType::Nested XprTypeNested; typedef typename remove_reference::type _XprTypeNested; + + enum { + Flags = 0, + }; }; } // end namespace internal @@ -153,6 +157,10 @@ struct traits > typedef typename RhsXprType::Nested RhsNested; typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; + + enum { + Flags = 0, + }; }; template diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index b8833362c..1fb90478f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -15,7 +15,7 @@ namespace Eigen { template class Tensor; template class TensorFixedSize; template class TensorMap; -template class TensorBase; +template::value> class TensorBase; template class TensorCwiseNullaryOp; template class TensorCwiseUnaryOp; @@ -29,6 +29,10 @@ template class TensorDevice; // Move to internal? template struct TensorEvaluator; +namespace internal { +template struct TensorAssign; +} // end namespace internal + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_FORWARD_DECLARATIONS_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 3e089fe1e..7d5f9271e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -21,7 +21,7 @@ namespace Eigen { */ namespace internal { template -struct traits > +struct traits > : public traits { // Type promotion to handle the case where the types of the lhs and the rhs are different. typedef typename XprType::Scalar Scalar; @@ -81,6 +81,7 @@ template struct TensorEvaluator > { typedef TensorReshapingOp XprType; + typedef NewDimensions Dimensions; enum { IsAligned = TensorEvaluator::IsAligned, @@ -95,7 +96,7 @@ struct TensorEvaluator > typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; - const NewDimensions& dimensions() const { return m_dimensions; } + const Dimensions& dimensions() const { return m_dimensions; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h index 2de698a57..40f805741 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h @@ -52,7 +52,7 @@ struct traits > typedef DenseIndex Index; enum { Options = Options_, - Flags = compute_tensor_flags::ret, + Flags = compute_tensor_flags::ret | LvalueBit, }; }; @@ -63,6 +63,10 @@ struct traits > typedef Scalar_ Scalar; typedef Dense StorageKind; typedef DenseIndex Index; + enum { + Options = Options_, + Flags = compute_tensor_flags::ret | LvalueBit, + }; }; From a77458a8ff2a83e716add62253eb50ef64980b21 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 9 Jun 2014 10:06:57 -0700 Subject: [PATCH 0522/1103] Fixes compilation errors triggered when compiling the tensor contraction code with cxx11 enabled. --- .../CXX11/src/Core/util/CXX11Workarounds.h | 6 ++++++ .../CXX11/src/Core/util/EmulateCXX11Meta.h | 17 +++++++++++++---- .../Eigen/CXX11/src/Tensor/TensorContraction.h | 16 ++++++++-------- 3 files changed, 27 insertions(+), 12 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index f102872ae..423ca4be4 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -66,6 +66,12 @@ template constexpr inline T const& array_ #undef STD_GET_ARR_HACK +template struct array_size; +template struct array_size > { + static const size_t value = N; +}; + + /* Suppose you have a template of the form * template struct X; * And you want to specialize it in such a way: diff --git a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h index 636063f9e..1d3164d6a 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h @@ -182,23 +182,32 @@ array repeat(t v) { } template -t array_prod(const array& a) { +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const array& a) { t prod = 1; for (size_t i = 0; i < n; ++i) { prod *= a[i]; } return prod; } template -t array_prod(const array& /*a*/) { +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const array& /*a*/) { return 0; } -template inline T& array_get(array& a) { +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T& array_get(array& a) { return a[I]; } -template inline const T& array_get(const array& a) { +template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE +const T& array_get(const array& a) { return a[I]; } + +template struct array_size; +template struct array_size > { + static const size_t value = N; +}; + + struct sum_op { template static inline bool run(A a, B b) { return a + b; } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index d371eb76d..5149de1bb 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -107,7 +107,7 @@ struct TensorEvaluator XprType; - static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * Indices::size>::size; + static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * internal::array_size::value>::size; typedef typename XprType::Index Index; typedef DSizes Dimensions; @@ -128,7 +128,7 @@ struct TensorEvaluator::Dimensions& left_dims = m_leftImpl.dimensions(); for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { bool skip = false; - for (int j = 0; j < Indices::size; ++j) { + for (int j = 0; j < internal::array_size::value; ++j) { if (op.indices()[j].first == i) { skip = true; m_leftOffsets[2*skipped] = stride; @@ -151,7 +151,7 @@ struct TensorEvaluator::Dimensions& right_dims = m_rightImpl.dimensions(); for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { bool skip = false; - for (int j = 0; j < Indices::size; ++j) { + for (int j = 0; j < internal::array_size::value; ++j) { if (op.indices()[j].second == i) { skip = true; m_rightOffsets[2*skipped] = stride; @@ -168,7 +168,7 @@ struct TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * Indices::size) { + if (TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * internal::array_size::value) { m_dimensions[0] = 1; } } @@ -209,7 +209,7 @@ struct TensorEvaluator::value-1) { partialStitch(left, right, StitchIndex+1, accum); } else { accum += m_leftImpl.coeff(left) * m_rightImpl.coeff(right); @@ -218,9 +218,9 @@ struct TensorEvaluator m_leftOffsets; - array m_rightOffsets; - array m_stitchsize; + array::value> m_leftOffsets; + array::value> m_rightOffsets; + array::value> m_stitchsize; Index m_shiftright; Dimensions m_dimensions; TensorEvaluator m_leftImpl; From 925fb6b93710b95082ba44d30405289dff3707eb Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Jun 2014 09:14:44 -0700 Subject: [PATCH 0523/1103] TensorEval are now typed on the device: this will make it possible to use partial template specialization to optimize the strategy of each evaluator for each device type. Started work on partial evaluations. --- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 42 +++++----- .../Eigen/CXX11/src/Tensor/TensorBase.h | 14 +++- .../CXX11/src/Tensor/TensorContraction.h | 26 +++---- .../CXX11/src/Tensor/TensorConvolution.h | 20 ++--- .../Eigen/CXX11/src/Tensor/TensorDevice.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 28 ++++++- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 76 +++++++++---------- .../src/Tensor/TensorForwardDeclarations.h | 9 ++- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 14 ++-- 9 files changed, 129 insertions(+), 102 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index da1eb62cb..633a7a31b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -32,15 +32,15 @@ namespace Eigen { namespace internal { // Default strategy: the expressions are evaluated with a single cpu thread. -template::PacketAccess & TensorEvaluator::PacketAccess> +template::PacketAccess & TensorEvaluator::PacketAccess> struct TensorAssign { typedef typename Derived1::Index Index; EIGEN_DEVICE_FUNC - static inline void run(Derived1& dst, const Derived2& src) + static inline void run(Derived1& dst, const Derived2& src, const Device& device = Device()) { - TensorEvaluator evalDst(dst); - TensorEvaluator evalSrc(src); + TensorEvaluator evalDst(dst, device); + TensorEvaluator evalSrc(src, device); const Index size = dst.size(); for (Index i = 0; i < size; ++i) { evalDst.coeffRef(i) = evalSrc.coeff(i); @@ -49,19 +49,19 @@ struct TensorAssign }; -template -struct TensorAssign +template +struct TensorAssign { typedef typename Derived1::Index Index; - static inline void run(Derived1& dst, const Derived2& src) + static inline void run(Derived1& dst, const Derived2& src, const Device& device = Device()) { - TensorEvaluator evalDst(dst); - TensorEvaluator evalSrc(src); + TensorEvaluator evalDst(dst, device); + TensorEvaluator evalSrc(src, device); const Index size = dst.size(); - static const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; - static const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; - static const int PacketSize = unpacket_traits::PacketReturnType>::size; + static const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + static const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + static const int PacketSize = unpacket_traits::PacketReturnType>::size; const int VectorizedSize = (size / PacketSize) * PacketSize; for (Index i = 0; i < VectorizedSize; i += PacketSize) { @@ -116,12 +116,12 @@ struct TensorAssignMultiThreaded typedef typename Derived1::Index Index; static inline void run(Derived1& dst, const Derived2& src, const ThreadPoolDevice& device) { - TensorEvaluator evalDst(dst); - TensorEvaluator evalSrc(src); + TensorEvaluator evalDst(dst, DefaultDevice()); + TensorEvaluator evalSrc(src, Defaultevice()); const Index size = dst.size(); - static const bool Vectorizable = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess; - static const int PacketSize = Vectorizable ? unpacket_traits::PacketReturnType>::size : 1; + static const bool Vectorizable = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess; + static const int PacketSize = Vectorizable ? unpacket_traits::PacketReturnType>::size : 1; int blocksz = static_cast(ceil(static_cast(size)/device.numThreads()) + PacketSize - 1); const Index blocksize = std::max(PacketSize, (blocksz - (blocksz % PacketSize))); @@ -131,7 +131,7 @@ struct TensorAssignMultiThreaded vector > results; results.reserve(numblocks); for (int i = 0; i < numblocks; ++i) { - results.push_back(std::async(std::launch::async, &EvalRange, TensorEvaluator, Index>::run, evalDst, evalSrc, i*blocksize, (i+1)*blocksize)); + results.push_back(std::async(std::launch::async, &EvalRange, TensorEvaluator, Index>::run, evalDst, evalSrc, i*blocksize, (i+1)*blocksize)); } for (int i = 0; i < numblocks; ++i) { @@ -167,19 +167,19 @@ struct TensorAssignGpu typedef typename Derived1::Index Index; static inline void run(Derived1& dst, const Derived2& src, const GpuDevice& device) { - TensorEvaluator evalDst(dst); - TensorEvaluator evalSrc(src); + TensorEvaluator evalDst(dst, device); + TensorEvaluator evalSrc(src, device); const Index size = dst.size(); const int block_size = std::min(size, 32*32); const int num_blocks = size / block_size; - EigenMetaKernelNoCheck, TensorEvaluator > <<>>(evalDst, evalSrc); + EigenMetaKernelNoCheck, TensorEvaluator > <<>>(evalDst, evalSrc); const int remaining_items = size % block_size; if (remaining_items > 0) { const int peel_start_offset = num_blocks * block_size; const int peel_block_size = std::min(size, 32); const int peel_num_blocks = (remaining_items + peel_block_size - 1) / peel_block_size; - EigenMetaKernelPeel, TensorEvaluator > <<>>(evalDst, evalSrc, peel_start_offset, size); + EigenMetaKernelPeel, TensorEvaluator > <<>>(evalDst, evalSrc, peel_start_offset, size); } } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index e447a5d40..6b53d2a3d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -198,19 +198,25 @@ class TensorBase } // Coefficient-wise ternary operators. - template - inline const TensorSelectOp + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorSelectOp select(const ThenDerived& thenTensor, const ElseDerived& elseTensor) const { return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); } // Morphing operators (slicing tbd). - template - inline const TensorReshapingOp + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReshapingOp reshape(const NewDimensions& newDimensions) const { return TensorReshapingOp(derived(), newDimensions); } + // Force the evaluation of the expression. + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorForcedEvalOp eval() const { + return TensorForcedEvalOp(derived()); + } + protected: template friend class TensorBase; EIGEN_DEVICE_FUNC diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index 5149de1bb..cadbabda2 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -102,31 +102,31 @@ template <> struct max_n_1<0> { }; -template -struct TensorEvaluator > +template +struct TensorEvaluator, Device> { typedef TensorContractionOp XprType; - static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * internal::array_size::value>::size; + static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * internal::array_size::value>::size; typedef typename XprType::Index Index; typedef DSizes Dimensions; enum { - IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, PacketAccess = /*TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess */ false, }; - TensorEvaluator(const XprType& op) - : m_leftImpl(op.lhsExpression()), m_rightImpl(op.rhsExpression()) + TensorEvaluator(const XprType& op, const Device& device) + : m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device) { Index index = 0; Index stride = 1; m_shiftright = 1; int skipped = 0; - const typename TensorEvaluator::Dimensions& left_dims = m_leftImpl.dimensions(); - for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { + const typename TensorEvaluator::Dimensions& left_dims = m_leftImpl.dimensions(); + for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { bool skip = false; for (int j = 0; j < internal::array_size::value; ++j) { if (op.indices()[j].first == i) { @@ -148,8 +148,8 @@ struct TensorEvaluator::Dimensions& right_dims = m_rightImpl.dimensions(); - for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { + const typename TensorEvaluator::Dimensions& right_dims = m_rightImpl.dimensions(); + for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { bool skip = false; for (int j = 0; j < internal::array_size::value; ++j) { if (op.indices()[j].second == i) { @@ -168,7 +168,7 @@ struct TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * internal::array_size::value) { + if (TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * internal::array_size::value) { m_dimensions[0] = 1; } } @@ -223,8 +223,8 @@ struct TensorEvaluator::value> m_stitchsize; Index m_shiftright; Dimensions m_dimensions; - TensorEvaluator m_leftImpl; - TensorEvaluator m_rightImpl; + TensorEvaluator m_leftImpl; + TensorEvaluator m_rightImpl; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 501e9a522..a554b8260 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -94,27 +94,27 @@ class TensorConvolutionOp : public TensorBase -struct TensorEvaluator > +template +struct TensorEvaluator, Device> { typedef TensorConvolutionOp XprType; - static const int NumDims = TensorEvaluator::Dimensions::count; + static const int NumDims = TensorEvaluator::Dimensions::count; static const int KernelDims = Indices::size; typedef typename XprType::Index Index; typedef DSizes Dimensions; enum { - IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, PacketAccess = /*TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess */ false, }; - TensorEvaluator(const XprType& op) - : m_inputImpl(op.inputExpression()), m_kernelImpl(op.kernelExpression()), m_dimensions(op.inputExpression().dimensions()) + TensorEvaluator(const XprType& op, const Device& device) + : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_dimensions(op.inputExpression().dimensions()) { - const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); - const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); + const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); + const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); for (int i = 0; i < NumDims; ++i) { if (i > 0) { @@ -200,8 +200,8 @@ struct TensorEvaluator m_indexStride; array m_kernelStride; Dimensions m_dimensions; - TensorEvaluator m_inputImpl; - TensorEvaluator m_kernelImpl; + TensorEvaluator m_inputImpl; + TensorEvaluator m_kernelImpl; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h index dbe60a165..ce524a818 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h @@ -31,7 +31,7 @@ template class TensorDevice { template EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) { - internal::TensorAssign::run(m_expression, other); + internal::TensorAssign::run(m_expression, other, m_device); return *this; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h index d7f5ab7c9..142edda14 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -15,6 +15,12 @@ namespace Eigen { // Default device for the machine (typically a single cpu core) struct DefaultDevice { + EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { + return internal::aligned_malloc(num_bytes); + } + EIGEN_STRONG_INLINE void deallocate(void* buffer) const { + internal::aligned_free(buffer); + } }; @@ -22,14 +28,19 @@ struct DefaultDevice { // We should really use a thread pool here but first we need to find a portable thread pool library. #ifdef EIGEN_USE_THREADS struct ThreadPoolDevice { - ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : /*pool_(pool), */num_threads_(num_cores) { } + ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : /*pool_(pool), */num_threads_(num_cores) { } size_t numThreads() const { return num_threads_; } - /*ThreadPool* threadPool() const { return pool_; }*/ + + EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { + return internal::aligned_malloc(num_bytes); + } + EIGEN_STRONG_INLINE void deallocate(void* buffer) const { + internal::aligned_free(buffer); + } private: // todo: NUMA, ... size_t num_threads_; - /*ThreadPool* pool_;*/ }; #endif @@ -40,7 +51,16 @@ struct GpuDevice { // The cudastream is not owned: the caller is responsible for its initialization and eventual destruction. GpuDevice(const cudaStream_t* stream) : stream_(stream) { eigen_assert(stream); } - const cudaStream_t& stream() const { return *stream_; } + EIGEN_STRONG_INLINE const cudaStream_t& stream() const { return *stream_; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { + void* result; + cudaMalloc(&result, num_bytes); + return result; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const { + cudaFree(buffer); + } private: // TODO: multigpu. diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index ab2513cea..80fe06957 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -23,7 +23,7 @@ namespace Eigen { * leading to lvalues (slicing, reshaping, etc...) */ -template +template struct TensorEvaluator { typedef typename Derived::Index Index; @@ -38,7 +38,7 @@ struct TensorEvaluator PacketAccess = Derived::PacketAccess, }; - EIGEN_DEVICE_FUNC TensorEvaluator(Derived& m) + EIGEN_DEVICE_FUNC TensorEvaluator(Derived& m, const Device&) : m_data(const_cast(m.data())), m_dims(m.dimensions()) { } @@ -73,8 +73,8 @@ struct TensorEvaluator // -------------------- CwiseNullaryOp -------------------- -template -struct TensorEvaluator > +template +struct TensorEvaluator, Device> { typedef TensorCwiseNullaryOp XprType; @@ -84,14 +84,14 @@ struct TensorEvaluator > }; EIGEN_DEVICE_FUNC - TensorEvaluator(const XprType& op) - : m_functor(op.functor()), m_argImpl(op.nestedExpression()) + TensorEvaluator(const XprType& op, const Device& device) + : m_functor(op.functor()), m_argImpl(op.nestedExpression(), device) { } typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; - typedef typename TensorEvaluator::Dimensions Dimensions; + typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } @@ -108,32 +108,32 @@ struct TensorEvaluator > private: const NullaryOp m_functor; - TensorEvaluator m_argImpl; + TensorEvaluator m_argImpl; }; // -------------------- CwiseUnaryOp -------------------- -template -struct TensorEvaluator > +template +struct TensorEvaluator, Device> { typedef TensorCwiseUnaryOp XprType; enum { - IsAligned = TensorEvaluator::IsAligned, - PacketAccess = TensorEvaluator::PacketAccess & internal::functor_traits::PacketAccess, + IsAligned = TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess & internal::functor_traits::PacketAccess, }; - EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op) + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) : m_functor(op.functor()), - m_argImpl(op.nestedExpression()) + m_argImpl(op.nestedExpression(), device) { } typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; - typedef typename TensorEvaluator::Dimensions Dimensions; + typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } @@ -150,33 +150,33 @@ struct TensorEvaluator > private: const UnaryOp m_functor; - TensorEvaluator m_argImpl; + TensorEvaluator m_argImpl; }; // -------------------- CwiseBinaryOp -------------------- -template -struct TensorEvaluator > +template +struct TensorEvaluator, Device> { typedef TensorCwiseBinaryOp XprType; enum { - IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, - PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess & + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess & internal::functor_traits::PacketAccess, }; - EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op) + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) : m_functor(op.functor()), - m_leftImpl(op.lhsExpression()), - m_rightImpl(op.rhsExpression()) + m_leftImpl(op.lhsExpression(), device), + m_rightImpl(op.rhsExpression(), device) { } typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; - typedef typename TensorEvaluator::Dimensions Dimensions; + typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { @@ -196,34 +196,34 @@ struct TensorEvaluator m_leftImpl; - TensorEvaluator m_rightImpl; + TensorEvaluator m_leftImpl; + TensorEvaluator m_rightImpl; }; // -------------------- SelectOp -------------------- -template -struct TensorEvaluator > +template +struct TensorEvaluator, Device> { typedef TensorSelectOp XprType; enum { - IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, - PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess/* & + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess/* & TensorEvaluator::PacketAccess*/, }; - EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op) - : m_condImpl(op.ifExpression()), - m_thenImpl(op.thenExpression()), - m_elseImpl(op.elseExpression()) + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) + : m_condImpl(op.ifExpression(), device), + m_thenImpl(op.thenExpression(), device), + m_elseImpl(op.elseExpression(), device) { } typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; - typedef typename TensorEvaluator::Dimensions Dimensions; + typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { @@ -248,9 +248,9 @@ struct TensorEvaluator } private: - TensorEvaluator m_condImpl; - TensorEvaluator m_thenImpl; - TensorEvaluator m_elseImpl; + TensorEvaluator m_condImpl; + TensorEvaluator m_thenImpl; + TensorEvaluator m_elseImpl; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 1fb90478f..27bfe1d73 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -21,16 +21,17 @@ template class TensorCwiseNullaryO template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; template class TensorSelectOp; +template class TensorReductionOp; template class TensorContractionOp; template class TensorConvolutionOp; template class TensorReshapingOp; -template class TensorDevice; +template class TensorForcedEvalOp; -// Move to internal? -template struct TensorEvaluator; +template class TensorDevice; +template struct TensorEvaluator; namespace internal { -template struct TensorAssign; +template struct TensorAssign; } // end namespace internal } // end namespace Eigen diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 7d5f9271e..e9e74581f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -77,19 +77,19 @@ class TensorReshapingOp : public TensorBase -struct TensorEvaluator > +template +struct TensorEvaluator, Device> { typedef TensorReshapingOp XprType; typedef NewDimensions Dimensions; enum { - IsAligned = TensorEvaluator::IsAligned, - PacketAccess = TensorEvaluator::PacketAccess, + IsAligned = TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess, }; - TensorEvaluator(const XprType& op) - : m_impl(op.expression()), m_dimensions(op.dimensions()) + TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_dimensions(op.dimensions()) { } typedef typename XprType::Index Index; @@ -111,7 +111,7 @@ struct TensorEvaluator > private: NewDimensions m_dimensions; - TensorEvaluator m_impl; + TensorEvaluator m_impl; }; From aa664eabb912a1b96e417e9a8d9c98f423b7fc23 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Jun 2014 10:31:29 -0700 Subject: [PATCH 0524/1103] Fixed a few compilation errors. --- .../CXX11/src/Tensor/TensorConvolution.h | 2 +- unsupported/test/CMakeLists.txt | 9 +++---- unsupported/test/cxx11_tensor_contraction.cpp | 26 ++++++++++++------- unsupported/test/cxx11_tensor_convolution.cpp | 7 ++--- 4 files changed, 25 insertions(+), 19 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index a554b8260..c4cfe0cd8 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -100,7 +100,7 @@ struct TensorEvaluator XprType; static const int NumDims = TensorEvaluator::Dimensions::count; - static const int KernelDims = Indices::size; + static const int KernelDims = internal::array_size::value; typedef typename XprType::Index Index; typedef DSizes Dimensions; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 4a151bfa7..34130a192 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -95,9 +95,8 @@ ei_add_test(bdcsvd) option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." ON) if(EIGEN_TEST_CXX11) - # FIXME: add C++11 compiler switch in some portable way - # (MSVC doesn't need any for example, so this will - # clash there) + # It should be safe to always run these tests as there is some fallback code for + # older compiler that don't support cxx11. ei_add_test(cxx11_meta "-std=c++0x") ei_add_test(cxx11_tensor_simple "-std=c++0x") ei_add_test(cxx11_tensor_symmetry "-std=c++0x") @@ -107,7 +106,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_convolution "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") - ei_add_test(cxx11_tensor_device "-std=c++0x") +# ei_add_test(cxx11_tensor_device "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") - ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") +# ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp index 1c89dfdd1..fc67d500b 100644 --- a/unsupported/test/cxx11_tensor_contraction.cpp +++ b/unsupported/test/cxx11_tensor_contraction.cpp @@ -11,6 +11,7 @@ #include +using Eigen::DefaultDevice; using Eigen::Tensor; typedef Tensor::DimensionPair DimPair; @@ -29,9 +30,10 @@ static void test_evals() Tensor mat4(3,3); mat4.setZero(); Eigen::array dims3({{DimPair(0, 0)}}); - TensorEvaluator eval(mat1.contract(mat2, dims3)); + typedef TensorEvaluator Evaluator; + Evaluator eval(mat1.contract(mat2, dims3), DefaultDevice()); eval.evalTo(mat4.data()); - EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(Evaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); VERIFY_IS_EQUAL(eval.dimensions()[0], 3); VERIFY_IS_EQUAL(eval.dimensions()[1], 3); @@ -48,9 +50,10 @@ static void test_evals() Tensor mat5(2,2); mat5.setZero(); Eigen::array dims4({{DimPair(1, 1)}}); - TensorEvaluator eval2(mat1.contract(mat2, dims4)); + typedef TensorEvaluator Evaluator2; + Evaluator2 eval2(mat1.contract(mat2, dims4), DefaultDevice()); eval2.evalTo(mat5.data()); - EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(Evaluator2::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); VERIFY_IS_EQUAL(eval2.dimensions()[0], 2); VERIFY_IS_EQUAL(eval2.dimensions()[1], 2); @@ -62,9 +65,10 @@ static void test_evals() Tensor mat6(2,2); mat6.setZero(); Eigen::array dims6({{DimPair(1, 0)}}); - TensorEvaluator eval3(mat1.contract(mat3, dims6)); + typedef TensorEvaluator Evaluator3; + Evaluator3 eval3(mat1.contract(mat3, dims6), DefaultDevice()); eval3.evalTo(mat6.data()); - EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(Evaluator3::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); VERIFY_IS_EQUAL(eval3.dimensions()[0], 2); VERIFY_IS_EQUAL(eval3.dimensions()[1], 2); @@ -86,9 +90,10 @@ static void test_scalar() Tensor scalar(1); scalar.setZero(); Eigen::array dims({{DimPair(0, 0)}}); - TensorEvaluator eval(vec1.contract(vec2, dims)); + typedef TensorEvaluator Evaluator; + Evaluator eval(vec1.contract(vec2, dims), DefaultDevice()); eval.evalTo(scalar.data()); - EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==1ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(Evaluator::NumDims==1ul, YOU_MADE_A_PROGRAMMING_MISTAKE); float expected = 0.0f; for (int i = 0; i < 6; ++i) { @@ -109,9 +114,10 @@ static void test_multidims() Tensor mat3(2, 2, 2); mat3.setZero(); Eigen::array dims({{DimPair(1, 2), DimPair(2, 3)}}); - TensorEvaluator eval(mat1.contract(mat2, dims)); + typedef TensorEvaluator Evaluator; + Evaluator eval(mat1.contract(mat2, dims), DefaultDevice()); eval.evalTo(mat3.data()); - EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==3ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(Evaluator::NumDims==3ul, YOU_MADE_A_PROGRAMMING_MISTAKE); VERIFY_IS_EQUAL(eval.dimensions()[0], 2); VERIFY_IS_EQUAL(eval.dimensions()[1], 2); VERIFY_IS_EQUAL(eval.dimensions()[2], 2); diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp index 95e40f64f..bafe73edd 100644 --- a/unsupported/test/cxx11_tensor_convolution.cpp +++ b/unsupported/test/cxx11_tensor_convolution.cpp @@ -12,7 +12,7 @@ #include using Eigen::Tensor; - +using Eigen::DefaultDevice; static void test_evals() { @@ -26,9 +26,10 @@ static void test_evals() result.setZero(); Eigen::array::Index, 1> dims3({0}); - TensorEvaluator eval(input.convolve(kernel, dims3)); + typedef TensorEvaluator Evaluator; + Evaluator eval(input.convolve(kernel, dims3), DefaultDevice()); eval.evalTo(result.data()); - EIGEN_STATIC_ASSERT(TensorEvaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(Evaluator::NumDims==2ul, YOU_MADE_A_PROGRAMMING_MISTAKE); VERIFY_IS_EQUAL(eval.dimensions()[0], 2); VERIFY_IS_EQUAL(eval.dimensions()[1], 3); From 1ee4e2db15caef88c281cfcecb44577b3e59a357 Mon Sep 17 00:00:00 2001 From: Vladimir Chalupecky Date: Thu, 12 Jun 2014 10:51:02 +0900 Subject: [PATCH 0525/1103] Change variable names in Eigen3Config.cmake to EIGEN3_* --- cmake/Eigen3Config.cmake.in | 40 ++++++++++++++++++------------------- 1 file changed, 20 insertions(+), 20 deletions(-) diff --git a/cmake/Eigen3Config.cmake.in b/cmake/Eigen3Config.cmake.in index 257c595ed..e50f6dbe0 100644 --- a/cmake/Eigen3Config.cmake.in +++ b/cmake/Eigen3Config.cmake.in @@ -3,26 +3,26 @@ # Eigen3Config.cmake(.in) # Use the following variables to compile and link against Eigen: -# EIGEN_FOUND - True if Eigen was found on your system -# EIGEN_USE_FILE - The file making Eigen usable -# EIGEN_DEFINITIONS - Definitions needed to build with Eigen -# EIGEN_INCLUDE_DIR - Directory where signature_of_eigen3_matrix_library can be found -# EIGEN_INCLUDE_DIRS - List of directories of Eigen and it's dependencies -# EIGEN_ROOT_DIR - The base directory of Eigen -# EIGEN_VERSION_STRING - A human-readable string containing the version -# EIGEN_VERSION_MAJOR - The major version of Eigen -# EIGEN_VERSION_MINOR - The minor version of Eigen -# EIGEN_VERSION_PATCH - The patch version of Eigen +# EIGEN3_FOUND - True if Eigen was found on your system +# EIGEN3_USE_FILE - The file making Eigen usable +# EIGEN3_DEFINITIONS - Definitions needed to build with Eigen +# EIGEN3_INCLUDE_DIR - Directory where signature_of_eigen3_matrix_library can be found +# EIGEN3_INCLUDE_DIRS - List of directories of Eigen and it's dependencies +# EIGEN3_ROOT_DIR - The base directory of Eigen +# EIGEN3_VERSION_STRING - A human-readable string containing the version +# EIGEN3_VERSION_MAJOR - The major version of Eigen +# EIGEN3_VERSION_MINOR - The minor version of Eigen +# EIGEN3_VERSION_PATCH - The patch version of Eigen -set ( EIGEN_FOUND 1 ) -set ( EIGEN_USE_FILE "@EIGEN_USE_FILE@" ) +set ( EIGEN3_FOUND 1 ) +set ( EIGEN3_USE_FILE "@EIGEN_USE_FILE@" ) -set ( EIGEN_DEFINITIONS "@EIGEN_DEFINITIONS@" ) -set ( EIGEN_INCLUDE_DIR "@EIGEN_INCLUDE_DIR@" ) -set ( EIGEN_INCLUDE_DIRS "@EIGEN_INCLUDE_DIRS@" ) -set ( EIGEN_ROOT_DIR "@EIGEN_ROOT_DIR@" ) +set ( EIGEN3_DEFINITIONS "@EIGEN_DEFINITIONS@" ) +set ( EIGEN3_INCLUDE_DIR "@EIGEN_INCLUDE_DIR@" ) +set ( EIGEN3_INCLUDE_DIRS "@EIGEN_INCLUDE_DIRS@" ) +set ( EIGEN3_ROOT_DIR "@EIGEN_ROOT_DIR@" ) -set ( EIGEN_VERSION_STRING "@EIGEN_VERSION_STRING@" ) -set ( EIGEN_VERSION_MAJOR "@EIGEN_VERSION_MAJOR@" ) -set ( EIGEN_VERSION_MINOR "@EIGEN_VERSION_MINOR@" ) -set ( EIGEN_VERSION_PATCH "@EIGEN_VERSION_PATCH@" ) +set ( EIGEN3_VERSION_STRING "@EIGEN_VERSION_STRING@" ) +set ( EIGEN3_VERSION_MAJOR "@EIGEN_VERSION_MAJOR@" ) +set ( EIGEN3_VERSION_MINOR "@EIGEN_VERSION_MINOR@" ) +set ( EIGEN3_VERSION_PATCH "@EIGEN_VERSION_PATCH@" ) From 38ab7e6ed0491bd5a0c639f218d5ea4728bf1e81 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 13 Jun 2014 09:56:51 -0700 Subject: [PATCH 0526/1103] Reworked the expression evaluation mechanism in order to make it possible to efficiently compute convolutions and contractions in the future: * The scheduling of computation is moved out the the assignment code and into a new TensorExecutor class * The assignment itself is now a regular node on the expression tree * The expression evaluators start by recursively evaluating all their subexpressions if needed --- unsupported/Eigen/CXX11/Tensor | 4 + unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 270 ++++++++---------- .../CXX11/src/Tensor/TensorContraction.h | 8 + .../CXX11/src/Tensor/TensorConvolution.h | 9 + .../Eigen/CXX11/src/Tensor/TensorDevice.h | 16 +- .../Eigen/CXX11/src/Tensor/TensorEvalTo.h | 146 ++++++++++ .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 56 +++- .../Eigen/CXX11/src/Tensor/TensorExecutor.h | 194 +++++++++++++ .../Eigen/CXX11/src/Tensor/TensorFixedSize.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorForcedEval.h | 142 +++++++++ .../src/Tensor/TensorForwardDeclarations.h | 5 +- .../Eigen/CXX11/src/Tensor/TensorMap.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 7 + 14 files changed, 695 insertions(+), 174 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index c67020581..7e504b302 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -42,8 +42,12 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h" + #include "unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h" #include "unsupported/Eigen/CXX11/src/Tensor/Tensor.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index 7f614bbe8..09601fc7d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -236,7 +236,9 @@ class Tensor : public TensorBase > // FIXME: we need to resize the tensor to fix the dimensions of the other. // Unfortunately this isn't possible yet when the rhs is an expression. // resize(other.dimensions()); - internal::TensorAssign::run(*this, other); + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); return *this; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index 633a7a31b..a2a925775 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -10,10 +10,6 @@ #ifndef EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H #define EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H -#ifdef EIGEN_USE_THREADS -#include -#endif - namespace Eigen { /** \class TensorAssign @@ -21,172 +17,134 @@ namespace Eigen { * * \brief The tensor assignment class. * - * This class is responsible for triggering the evaluation of the expressions - * used on the lhs and rhs of an assignment operator and copy the result of - * the evaluation of the rhs expression at the address computed during the - * evaluation lhs expression. - * - * TODO: vectorization. For now the code only uses scalars - * TODO: parallelisation using multithreading on cpu, or kernels on gpu. + * This class is represents the assignment of the values resulting from the evaluation of + * the rhs expression to the memory locations denoted by the lhs expression. */ namespace internal { - -// Default strategy: the expressions are evaluated with a single cpu thread. -template::PacketAccess & TensorEvaluator::PacketAccess> -struct TensorAssign +template +struct traits > { - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1& dst, const Derived2& src, const Device& device = Device()) - { - TensorEvaluator evalDst(dst, device); - TensorEvaluator evalSrc(src, device); - const Index size = dst.size(); - for (Index i = 0; i < size; ++i) { - evalDst.coeffRef(i) = evalSrc.coeff(i); - } - } + typedef typename LhsXprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename promote_index_type::Index, + typename traits::Index>::type Index; + typedef typename LhsXprType::Nested LhsNested; + typedef typename RhsXprType::Nested RhsNested; + typedef typename remove_reference::type _LhsNested; + typedef typename remove_reference::type _RhsNested; + + enum { + Flags = 0, + }; }; - -template -struct TensorAssign +template +struct eval, Eigen::Dense> { - typedef typename Derived1::Index Index; - static inline void run(Derived1& dst, const Derived2& src, const Device& device = Device()) - { - TensorEvaluator evalDst(dst, device); - TensorEvaluator evalSrc(src, device); - const Index size = dst.size(); - - static const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; - static const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; - static const int PacketSize = unpacket_traits::PacketReturnType>::size; - const int VectorizedSize = (size / PacketSize) * PacketSize; - - for (Index i = 0; i < VectorizedSize; i += PacketSize) { - evalDst.template writePacket(i, evalSrc.template packet(i)); - } - for (Index i = VectorizedSize; i < size; ++i) { - evalDst.coeffRef(i) = evalSrc.coeff(i); - } - } + typedef const TensorAssignOp& type; }; - - -// Multicore strategy: the index space is partitioned and each core is assigned to a partition -#ifdef EIGEN_USE_THREADS -template -struct EvalRange { - static void run(LhsEval& dst, const RhsEval& src, const Index first, const Index last) { - eigen_assert(last > first); - for (Index i = first; i < last; ++i) { - dst.coeffRef(i) = src.coeff(i); - } - } -}; - -template -struct EvalRange { - static void run(LhsEval& dst, const RhsEval& src, const Index first, const Index last) { - eigen_assert(last > first); - - Index i = first; - static const int PacketSize = unpacket_traits::size; - if (last - first > PacketSize) { - static const int LhsStoreMode = LhsEval::IsAligned ? Aligned : Unaligned; - static const int RhsLoadMode = RhsEval::IsAligned ? Aligned : Unaligned; - eigen_assert(first % PacketSize == 0); - Index lastPacket = last - (last % PacketSize); - for (; i < lastPacket; i += PacketSize) { - dst.template writePacket(i, src.template packet(i)); - } - } - - for (; i < last; ++i) { - dst.coeffRef(i) = src.coeff(i); - } - } -}; - -template -struct TensorAssignMultiThreaded +template +struct nested, 1, typename eval >::type> { - typedef typename Derived1::Index Index; - static inline void run(Derived1& dst, const Derived2& src, const ThreadPoolDevice& device) - { - TensorEvaluator evalDst(dst, DefaultDevice()); - TensorEvaluator evalSrc(src, Defaultevice()); - const Index size = dst.size(); - - static const bool Vectorizable = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess; - static const int PacketSize = Vectorizable ? unpacket_traits::PacketReturnType>::size : 1; - - int blocksz = static_cast(ceil(static_cast(size)/device.numThreads()) + PacketSize - 1); - const Index blocksize = std::max(PacketSize, (blocksz - (blocksz % PacketSize))); - const Index numblocks = size / blocksize; - - Index i = 0; - vector > results; - results.reserve(numblocks); - for (int i = 0; i < numblocks; ++i) { - results.push_back(std::async(std::launch::async, &EvalRange, TensorEvaluator, Index>::run, evalDst, evalSrc, i*blocksize, (i+1)*blocksize)); - } - - for (int i = 0; i < numblocks; ++i) { - results[i].get(); - } - - if (numblocks * blocksize < size) { - EvalRange, TensorEvaluator, Index>::run(evalDst, evalSrc, numblocks * blocksize, size); - } - } + typedef TensorAssignOp type; }; -#endif + +} // end namespace internal -// GPU: the evaluation of the expressions is offloaded to a GPU. -#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) -template -__global__ void EigenMetaKernelNoCheck(LhsEvaluator evalDst, const RhsEvaluator evalSrc) { - const int index = blockIdx.x * blockDim.x + threadIdx.x; - evalDst.coeffRef(index) = evalSrc.coeff(index); -} -template -__global__ void EigenMetaKernelPeel(LhsEvaluator evalDst, const RhsEvaluator evalSrc, int peel_start_offset, int size) { - const int index = peel_start_offset + blockIdx.x * blockDim.x + threadIdx.x; - if (index < size) { - evalDst.coeffRef(index) = evalSrc.coeff(index); + +template +class TensorAssignOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename LhsXprType::CoeffReturnType CoeffReturnType; + typedef typename LhsXprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorAssignOp(LhsXprType& lhs, const RhsXprType& rhs) + : m_lhs_xpr(lhs), m_rhs_xpr(rhs) {} + + /** \returns the nested expressions */ + EIGEN_DEVICE_FUNC + typename internal::remove_all::type& + lhsExpression() const { return *((typename internal::remove_all::type*)&m_lhs_xpr); } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + rhsExpression() const { return m_rhs_xpr; } + + protected: + typename internal::remove_all::type& m_lhs_xpr; + const typename internal::remove_all::type& m_rhs_xpr; +}; + + +template +struct TensorEvaluator, Device> +{ + typedef TensorAssignOp XprType; + + enum { + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess, + }; + + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) : + m_leftImpl(op.lhsExpression(), device), + m_rightImpl(op.rhsExpression(), device) + { } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename TensorEvaluator::Dimensions Dimensions; + + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const + { + // TODO: use left impl instead if right impl dimensions are known at compile time. + return m_rightImpl.dimensions(); } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_leftImpl.evalSubExprsIfNeeded(); + m_rightImpl.evalSubExprsIfNeeded(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_leftImpl.cleanup(); + m_rightImpl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalScalar(Index i) { + m_leftImpl.coeffRef(i) = m_rightImpl.coeff(i); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalPacket(Index i) { + static const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + static const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + m_leftImpl.template writePacket(i, m_rightImpl.template packet(i)); + } + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const + { + return m_leftImpl.coeff(index); + } + template + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + { + return m_leftImpl.template packet(index); + } + + private: + TensorEvaluator m_leftImpl; + TensorEvaluator m_rightImpl; +}; + } -template -struct TensorAssignGpu -{ - typedef typename Derived1::Index Index; - static inline void run(Derived1& dst, const Derived2& src, const GpuDevice& device) - { - TensorEvaluator evalDst(dst, device); - TensorEvaluator evalSrc(src, device); - const Index size = dst.size(); - const int block_size = std::min(size, 32*32); - const int num_blocks = size / block_size; - EigenMetaKernelNoCheck, TensorEvaluator > <<>>(evalDst, evalSrc); - - const int remaining_items = size % block_size; - if (remaining_items > 0) { - const int peel_start_offset = num_blocks * block_size; - const int peel_block_size = std::min(size, 32); - const int peel_num_blocks = (remaining_items + peel_block_size - 1) / peel_block_size; - EigenMetaKernelPeel, TensorEvaluator > <<>>(evalDst, evalSrc, peel_start_offset, size); - } - } -}; -#endif - -} // end namespace internal - -} // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_ASSIGN_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index cadbabda2..b2e12fd15 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -184,6 +184,14 @@ struct TensorEvaluator class TensorDevice { template EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) { - internal::TensorAssign::run(m_expression, other, m_device); + typedef TensorAssignOp Assign; + Assign assign(m_expression, other); + static const bool Vectorize = TensorEvaluator::PacketAccess; + internal::TensorExecutor::run(assign, m_device); return *this; } @@ -48,7 +51,10 @@ template class TensorDevice EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) { - internal::TensorAssignMultiThreaded::run(m_expression, other, m_device); + typedef TensorAssignOp Assign; + Assign assign(m_expression, other); + static const bool Vectorize = TensorEvaluator::PacketAccess; + internal::TensorExecutor::run(assign, m_device); return *this; } @@ -67,13 +73,15 @@ template class TensorDevice template EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) { - internal::TensorAssignGpu::run(m_expression, other, m_device); + typedef TensorAssignOp Assign; + Assign assign(m_expression, other); + internal::TensorExecutor::run(assign, m_device); return *this; } protected: const GpuDevice& m_device; - ExpressionType& m_expression; + ExpressionType m_expression; }; #endif diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h new file mode 100644 index 000000000..db716a80e --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h @@ -0,0 +1,146 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_EVAL_TO_H +#define EIGEN_CXX11_TENSOR_TENSOR_EVAL_TO_H + +namespace Eigen { + +/** \class TensorForcedEval + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor reshaping class. + * + * + */ +namespace internal { +template +struct traits > +{ + // Type promotion to handle the case where the types of the lhs and the rhs are different. + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; + + enum { + Flags = 0, + }; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorEvalToOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorEvalToOp type; +}; + +} // end namespace internal + + + + +template +class TensorEvalToOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvalToOp(Scalar* buffer, const XprType& expr) + : m_xpr(expr), m_buffer(buffer) {} + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + EIGEN_DEVICE_FUNC Scalar* buffer() const { return m_buffer; } + + protected: + typename XprType::Nested m_xpr; + Scalar* m_buffer; +}; + + + +template +struct TensorEvaluator, Device> +{ + typedef TensorEvalToOp XprType; + typedef typename ArgType::Scalar Scalar; + typedef typename ArgType::Packet Packet; + typedef typename TensorEvaluator::Dimensions Dimensions; + + enum { + IsAligned = true, + PacketAccess = true, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_device(device), m_buffer(op.buffer()) + { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ~TensorEvaluator() { + } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_impl.evalSubExprsIfNeeded(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalScalar(Index i) { + m_buffer[i] = m_impl.coeff(i); + } + EIGEN_STRONG_INLINE void evalPacket(Index i) { + internal::pstoret(m_buffer + i, m_impl.template packet::IsAligned ? Aligned : Unaligned>(i)); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + return m_buffer[index]; + } + + template + EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return internal::ploadt(m_buffer + index); + } + + private: + TensorEvaluator m_impl; + const Device& m_device; + Scalar* m_buffer; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_EVAL_TO_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index 80fe06957..5c8b079da 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -38,27 +38,32 @@ struct TensorEvaluator PacketAccess = Derived::PacketAccess, }; - EIGEN_DEVICE_FUNC TensorEvaluator(Derived& m, const Device&) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(Derived& m, const Device&) : m_data(const_cast(m.data())), m_dims(m.dimensions()) { } - EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_dims; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dims; } - EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { + eigen_assert(m_data); return m_data[index]; } - EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) { + eigen_assert(m_data); return m_data[index]; } - template + template EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return internal::ploadt(m_data + index); } - template + template EIGEN_STRONG_INLINE void writePacket(Index index, const Packet& x) { return internal::pstoret(m_data + index, x); @@ -95,13 +100,16 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { } + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_functor(index); } template - EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return m_functor.packetOp(index); } @@ -137,13 +145,20 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_argImpl.evalSubExprsIfNeeded(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_argImpl.cleanup(); + } + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_functor(m_argImpl.coeff(index)); } template - EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return m_functor.packetOp(m_argImpl.template packet(index)); } @@ -184,12 +199,21 @@ struct TensorEvaluator - EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return m_functor.packetOp(m_leftImpl.template packet(index), m_rightImpl.template packet(index)); } @@ -230,12 +254,24 @@ struct TensorEvaluator // TODO: use then or else impl instead if they happen to be known at compile time. return m_condImpl.dimensions(); } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_condImpl.evalSubExprsIfNeeded(); + m_thenImpl.evalSubExprsIfNeeded(); + m_elseImpl.evalSubExprsIfNeeded(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_condImpl.cleanup(); + m_thenImpl.cleanup(); + m_elseImpl.cleanup(); + } + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_condImpl.coeff(index) ? m_thenImpl.coeff(index) : m_elseImpl.coeff(index); } template - EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + PacketReturnType packet(Index index) const { static const int PacketSize = internal::unpacket_traits::size; internal::Selector select; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h new file mode 100644 index 000000000..3e41f3290 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -0,0 +1,194 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H +#define EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H + +#ifdef EIGEN_USE_THREADS +#include +#endif + +namespace Eigen { + +/** \class TensorExecutor + * \ingroup CXX11_Tensor_Module + * + * \brief The tensor executor class. + * + * This class is responsible for launch the evaluation of the expression on + * the specified computing device. + */ +namespace internal { + +// Default strategy: the expression is evaluated with a single cpu thread. +template::PacketAccess> +struct TensorExecutor +{ + typedef typename Expression::Index Index; + EIGEN_DEVICE_FUNC + static inline void run(const Expression& expr, const Device& device = Device()) + { + TensorEvaluator evaluator(expr, device); + evaluator.evalSubExprsIfNeeded(); + + const Index size = evaluator.dimensions().TotalSize(); + for (Index i = 0; i < size; ++i) { + evaluator.evalScalar(i); + } + + evaluator.cleanup(); + } +}; + + +template +struct TensorExecutor +{ + typedef typename Expression::Index Index; + static inline void run(const Expression& expr, const DefaultDevice& device = DefaultDevice()) + { + TensorEvaluator evaluator(expr, device); + evaluator.evalSubExprsIfNeeded(); + + const Index size = evaluator.dimensions().TotalSize(); + static const int PacketSize = unpacket_traits::PacketReturnType>::size; + const int VectorizedSize = (size / PacketSize) * PacketSize; + + for (Index i = 0; i < VectorizedSize; i += PacketSize) { + evaluator.evalPacket(i); + } + for (Index i = VectorizedSize; i < size; ++i) { + evaluator.evalScalar(i); + } + + evaluator.cleanup(); + } +}; + + + +// Multicore strategy: the index space is partitioned and each partition is executed on a single core +#ifdef EIGEN_USE_THREADS +template +struct EvalRange { + static void run(Evaluator& evaluator, const Index first, const Index last) { + eigen_assert(last > first); + for (Index i = first; i < last; ++i) { + evaluator.evalScalar(i); + } + } +}; + +template +struct EvalRange { + static void run(Evaluator& evaluator, const Index first, const Index last,) { + eigen_assert(last > first); + + Index i = first; + static const int PacketSize = unpacket_traits::size; + if (last - first > PacketSize) { + eigen_assert(first % PacketSize == 0); + Index lastPacket = last - (last % PacketSize); + for (; i < lastPacket; i += PacketSize) { + evaluator.evalPacket(i); + } + } + + for (; i < last; ++i) { + evaluator.evalScalar(i); + } + } +}; + +template +struct TensorExecutor +{ + typedef typename Expression::Index Index; + static inline void run(const Expression& expr, const ThreadPoolDevice& device) + { + TensorEvaluator evaluator(expr, device); + evaluator.evalSubExprsIfNeeded(); + + const Index size = evaluator.dimensions().TotalSize(); + + static const int PacketSize = Vectorizable ? unpacket_traits::PacketReturnType>::size : 1; + + int blocksz = std::ceil(static_cast(size)/device.numThreads()) + PacketSize - 1; + const Index blocksize = std::max(PacketSize, (blocksz - (blocksz % PacketSize))); + const Index numblocks = size / blocksize; + + TensorEvaluator single_threaded_eval(expr, DefaultDevice()); + + Index i = 0; + vector > results; + results.reserve(numblocks); + for (int i = 0; i < numblocks; ++i) { + results.push_back(std::async(std::launch::async, &EvalRange, Index>::run, single_threaded_eval, i*blocksize, (i+1)*blocksize)); + } + + for (int i = 0; i < numblocks; ++i) { + results[i].get(); + } + + if (numblocks * blocksize < size) { + EvalRange, Index>::run(single_threaded_eval, numblocks * blocksize, size, nullptr); + } + + evaluator.cleanup(); + } +}; +#endif + + +// GPU: the evaluation of the expression is offloaded to a GPU. +#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) +template +__global__ void EigenMetaKernelNoCheck(Evaluator eval) { + const int index = blockIdx.x * blockDim.x + threadIdx.x; + eval.evalScalar(index); +} +template +__global__ void EigenMetaKernelPeel(Evaluator eval, int peel_start_offset, int size) { + const int index = peel_start_offset + blockIdx.x * blockDim.x + threadIdx.x; + if (index < size) { + eval.evalScalar(index); + } +} + +template +struct TensorExecutor +{ + typedef typename Expression::Index Index; + static inline void run(const Expression& expr, const GpuDevice& device) + { + TensorEvaluator evaluator(expr, device); + evaluator.evalSubExprsIfNeeded(); + + const Index size = evaluator.dimensions().TotalSize(); + const int block_size = std::min(size, 32*32); + const int num_blocks = size / block_size; + EigenMetaKernelNoCheck > <<>>(evaluator); + + const int remaining_items = size % block_size; + if (remaining_items > 0) { + const int peel_start_offset = num_blocks * block_size; + const int peel_block_size = std::min(size, 32); + const int peel_num_blocks = (remaining_items + peel_block_size - 1) / peel_block_size; + EigenMetaKernelPeel > <<>>(evaluator, peel_start_offset, size); + } + evaluator.cleanup(); + } +}; +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h index 789c04238..d42167da9 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h @@ -200,7 +200,9 @@ class TensorFixedSize : public TensorBase::run(*this, other); + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); return *this; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h new file mode 100644 index 000000000..6f6641de6 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h @@ -0,0 +1,142 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H +#define EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H + +namespace Eigen { + +/** \class TensorForcedEval + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor reshaping class. + * + * + */ +namespace internal { +template +struct traits > +{ + // Type promotion to handle the case where the types of the lhs and the rhs are different. + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; + + enum { + Flags = 0, + }; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorForcedEvalOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorForcedEvalOp type; +}; + +} // end namespace internal + + + +template +class TensorForcedEvalOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorForcedEvalOp(const XprType& expr) + : m_xpr(expr) {} + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + protected: + typename XprType::Nested m_xpr; +}; + + +template +struct TensorEvaluator, Device> +{ + typedef TensorForcedEvalOp XprType; + typedef typename ArgType::Scalar Scalar; + typedef typename ArgType::Packet Packet; + typedef typename TensorEvaluator::Dimensions Dimensions; + + enum { + IsAligned = true, + PacketAccess = true, + }; + + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_op(op.expression()), m_device(device), m_buffer(NULL) + { } + + EIGEN_DEVICE_FUNC ~TensorEvaluator() { + eigen_assert(!m_buffer); + } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); } + + EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_impl.evalSubExprsIfNeeded(); + m_buffer = (Scalar*)m_device.allocate(m_impl.dimensions().TotalSize() * sizeof(Scalar)); + + typedef TensorEvalToOp EvalTo; + EvalTo evalToTmp(m_buffer, m_op); + internal::TensorExecutor::PacketAccess>::run(evalToTmp, m_device); + m_impl.cleanup(); + } + EIGEN_STRONG_INLINE void cleanup() { + m_device.deallocate(m_buffer); + m_buffer = NULL; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + return m_buffer[index]; + } + + template + EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return internal::ploadt(m_buffer + index); + } + + private: + TensorEvaluator m_impl; + const ArgType m_op; + const Device& m_device; + Scalar* m_buffer; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 27bfe1d73..c0dffbd0c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -25,13 +25,16 @@ template class TensorReductionOp; template class TensorContractionOp; template class TensorConvolutionOp; template class TensorReshapingOp; +template class TensorAssignOp; + +template class TensorEvalToOp; template class TensorForcedEvalOp; template class TensorDevice; template struct TensorEvaluator; namespace internal { -template struct TensorAssign; +template class TensorExecutor; } // end namespace internal } // end namespace Eigen diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index 3a06170fa..c97135b63 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -246,7 +246,9 @@ template class TensorMap : public Tensor EIGEN_DEVICE_FUNC Self& operator=(const OtherDerived& other) { - internal::TensorAssign::run(*this, other); + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); return *this; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index e9e74581f..764bba4e6 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -98,6 +98,13 @@ struct TensorEvaluator, Device> const Dimensions& dimensions() const { return m_dimensions; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_impl.evalSubExprsIfNeeded(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { return m_impl.coeff(index); From f80c8e17eb042fc95767417eeca26cd3fa0c6ad6 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 13 Jun 2014 10:12:12 -0700 Subject: [PATCH 0527/1103] Silenced a compilation warning --- unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 58b1808a3..4bdf74286 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -208,9 +208,9 @@ struct TensorEvaluator m_indexStride; array m_kernelStride; - Dimensions m_dimensions; TensorEvaluator m_inputImpl; TensorEvaluator m_kernelImpl; + Dimensions m_dimensions; }; From 774c3c1e0aca307e484b00997b735ee5964d96d4 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 13 Jun 2014 10:20:28 -0700 Subject: [PATCH 0528/1103] Created additional unit tests for the tensor code and improved existing ones. --- unsupported/test/CMakeLists.txt | 3 + unsupported/test/cxx11_tensor_device.cpp | 28 ++++++++- unsupported/test/cxx11_tensor_lvalue.cpp | 42 +++++++++++++ unsupported/test/cxx11_tensor_morphing.cpp | 72 ++++++++++++++++++++++ 4 files changed, 143 insertions(+), 2 deletions(-) create mode 100644 unsupported/test/cxx11_tensor_lvalue.cpp create mode 100644 unsupported/test/cxx11_tensor_morphing.cpp diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 34130a192..7458128fb 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -105,7 +105,10 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_contraction "-std=c++0x") ei_add_test(cxx11_tensor_convolution "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") +# ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") + ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") + ei_add_test(cxx11_tensor_morphing "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") # ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_device.cpp b/unsupported/test/cxx11_tensor_device.cpp index 365b109c7..caf2e9735 100644 --- a/unsupported/test/cxx11_tensor_device.cpp +++ b/unsupported/test/cxx11_tensor_device.cpp @@ -65,6 +65,12 @@ static void test_contextual_eval(Context* context) context->out() = context->in1() + context->in2() * 3.14f + context->in1().constant(2.718f); } +template +static void test_forced_contextual_eval(Context* context) +{ + context->out() = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f); +} + static void test_cpu() { Eigen::Tensor in1(Eigen::array(2,3,7)); Eigen::Tensor in2(Eigen::array(2,3,7)); @@ -72,9 +78,9 @@ static void test_cpu() { in1.setRandom(); in2.setRandom(); + CPUContext context(in1, in2, out); test_contextual_eval(&context); - for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { @@ -82,6 +88,15 @@ static void test_cpu() { } } } + + test_forced_contextual_eval(&context); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k))) * 3.14f + 2.718f); + } + } + } } static void test_gpu() { @@ -111,7 +126,6 @@ static void test_gpu() { GPUContext context(gpu_in1, gpu_in2, gpu_out); test_contextual_eval(&context); - cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { @@ -120,6 +134,16 @@ static void test_gpu() { } } } + + test_forced_contextual_eval(&context); + cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k))) * 3.14f + 2.718f); + } + } + } } diff --git a/unsupported/test/cxx11_tensor_lvalue.cpp b/unsupported/test/cxx11_tensor_lvalue.cpp new file mode 100644 index 000000000..071f5b406 --- /dev/null +++ b/unsupported/test/cxx11_tensor_lvalue.cpp @@ -0,0 +1,42 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + + +static void test_compound_assignment() +{ + Tensor mat1(2,3,7); + Tensor mat2(2,3,7); + Tensor mat3(2,3,7); + + mat1.setRandom(); + mat2.setRandom(); + mat3 = mat1; + mat3 += mat2; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(mat3(i,j,k), mat1(i,j,k) + mat2(i,j,k)); + } + } + } +} + + +void test_cxx11_tensor_lvalue() +{ + CALL_SUBTEST(test_compound_assignment()); +} diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp new file mode 100644 index 000000000..21af9e0b5 --- /dev/null +++ b/unsupported/test/cxx11_tensor_morphing.cpp @@ -0,0 +1,72 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +static void test_simple_reshape() +{ + Tensor tensor1(2,3,1,7,1); + tensor1.setRandom(); + + Tensor tensor2(2,3,7); + Tensor tensor3(6,7); + Tensor tensor4(2,21); + + Tensor::Dimensions dim1{{2,3,7}}; + tensor2 = tensor1.reshape(dim1); + Tensor::Dimensions dim2{{6,7}}; + tensor3 = tensor1.reshape(dim2); + Tensor::Dimensions dim3{{2,21}}; + tensor4 = tensor1.reshape(dim1).reshape(dim3); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(tensor1(i,j,0,k,0), tensor2(i,j,k)); + VERIFY_IS_EQUAL(tensor1(i,j,0,k,0), tensor3(i+2*j,k)); + VERIFY_IS_EQUAL(tensor1(i,j,0,k,0), tensor4(i,j+3*k)); + } + } + } +} + + +static void test_reshape_in_expr() { + MatrixXf m1(2,3*5*7*11); + MatrixXf m2(3*5*7*11,13); + m1.setRandom(); + m2.setRandom(); + MatrixXf m3 = m1 * m2; + + TensorMap> tensor1(m1.data(), 2,3,5,7,11); + TensorMap> tensor2(m2.data(), 3,5,7,11,13); + Tensor::Dimensions newDims1{{2,3*5*7*11}}; + Tensor::Dimensions newDims2{{3*5*7*11,13}}; + typedef Tensor::DimensionPair DimPair; + array contract_along{{DimPair(1, 0)}}; + Tensor tensor3(2,13); + tensor3 = tensor1.reshape(newDims1).contract(tensor2.reshape(newDims2), contract_along); + + Map res(tensor3.data(), 2, 13); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 13; ++j) { + VERIFY_IS_APPROX(res(i,j), m3(i,j)); + } + } +} + +void test_cxx11_tensor_morphing() +{ + CALL_SUBTEST(test_simple_reshape()); + CALL_SUBTEST(test_reshape_in_expr()); +} From 95ecd582a37c5e3b3df47392f6807280488852f8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 17 Jun 2014 09:37:07 +0200 Subject: [PATCH 0529/1103] Update decompositions tables --- doc/TopicLinearAlgebraDecompositions.dox | 4 ++-- doc/TutorialLinearAlgebra.dox | 26 +++++++++++++++++++----- 2 files changed, 23 insertions(+), 7 deletions(-) diff --git a/doc/TopicLinearAlgebraDecompositions.dox b/doc/TopicLinearAlgebraDecompositions.dox index 77f2c92ab..5bcff2c96 100644 --- a/doc/TopicLinearAlgebraDecompositions.dox +++ b/doc/TopicLinearAlgebraDecompositions.dox @@ -116,7 +116,7 @@ For an introduction on linear solvers and decompositions, check this \link Tutor JacobiSVD (two-sided) - Slow (but fast for small matrices) - Excellent-Proven3 + Proven3 Yes Singular values/vectors, least squares Yes (and does least squares) @@ -132,7 +132,7 @@ For an introduction on linear solvers and decompositions, check this \link Tutor Yes Eigenvalues/vectors - - Good + Excellent Closed forms for 2x2 and 3x3 diff --git a/doc/TutorialLinearAlgebra.dox b/doc/TutorialLinearAlgebra.dox index e6c41fd70..cb92ceeae 100644 --- a/doc/TutorialLinearAlgebra.dox +++ b/doc/TutorialLinearAlgebra.dox @@ -40,8 +40,9 @@ depending on your matrix and the trade-off you want to make: Decomposition Method - Requirements on the matrix - Speed + Requirements
on the matrix + Speed
(small-to-medium) + Speed
(large) Accuracy @@ -49,6 +50,7 @@ depending on your matrix and the trade-off you want to make: partialPivLu() Invertible ++ + ++ + @@ -56,6 +58,7 @@ depending on your matrix and the trade-off you want to make: fullPivLu() None - + - - +++ @@ -63,20 +66,23 @@ depending on your matrix and the trade-off you want to make: householderQr() None ++ + ++ + ColPivHouseholderQR colPivHouseholderQr() None - + ++ + - + +++ FullPivHouseholderQR fullPivHouseholderQr() None - + - - +++ @@ -84,21 +90,31 @@ depending on your matrix and the trade-off you want to make: llt() Positive definite +++ + +++ + LDLT ldlt() - Positive or negative semidefinite + Positive or negative
semidefinite +++ + + ++ + + JacobiSVD + jacobiSvd() + None + - - + - - - + +++ + All of these decompositions offer a solve() method that works as in the above example. For example, if your matrix is positive definite, the above table says that a very good -choice is then the LDLT decomposition. Here's an example, also demonstrating that using a general +choice is then the LLT or LDLT decomposition. Here's an example, also demonstrating that using a general matrix (not a vector) as right hand side is possible. From c06ec0f464a312dbce24edfde1de75bb1a69c4a6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 17 Jun 2014 23:47:30 +0200 Subject: [PATCH 0530/1103] Fix Jacobi preconditioner with zero diagonal entries --- Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h index 73ca9bfde..1f3c060d0 100644 --- a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h +++ b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h @@ -65,10 +65,10 @@ class DiagonalPreconditioner { typename MatType::InnerIterator it(mat,j); while(it && it.index()!=j) ++it; - if(it && it.index()==j) + if(it && it.index()==j && it.value()!=Scalar(0)) m_invdiag(j) = Scalar(1)/it.value(); else - m_invdiag(j) = 0; + m_invdiag(j) = Scalar(1); } m_isInitialized = true; return *this; From afb1a8c124c9ee52e027a3625b14ecad6be99fa4 Mon Sep 17 00:00:00 2001 From: Mark Borgerding Date: Tue, 17 Jun 2014 18:25:56 -0400 Subject: [PATCH 0531/1103] fixed warning: -Wunused-local-typedefs --- Eigen/src/Core/products/TriangularMatrixVector_MKL.h | 2 -- 1 file changed, 2 deletions(-) diff --git a/Eigen/src/Core/products/TriangularMatrixVector_MKL.h b/Eigen/src/Core/products/TriangularMatrixVector_MKL.h index 09f110da7..3672b1240 100644 --- a/Eigen/src/Core/products/TriangularMatrixVector_MKL.h +++ b/Eigen/src/Core/products/TriangularMatrixVector_MKL.h @@ -129,7 +129,6 @@ struct triangular_matrix_vector_product_trmv MatrixLhs; \ if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \ x = x_tmp.data(); \ if (size MatrixLhs; \ if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \ x = x_tmp.data(); \ if (size Date: Thu, 19 Jun 2014 14:55:14 +0100 Subject: [PATCH 0532/1103] Add component-wise atan() function (see bug #80). --- Eigen/src/Core/Assign_MKL.h | 1 + Eigen/src/Core/GenericPacketMath.h | 4 ++++ Eigen/src/Core/GlobalFunctions.h | 1 + Eigen/src/Core/functors/UnaryFunctors.h | 20 ++++++++++++++++++++ Eigen/src/plugins/ArrayCwiseUnaryOps.h | 9 +++++++++ 5 files changed, 35 insertions(+) diff --git a/Eigen/src/Core/Assign_MKL.h b/Eigen/src/Core/Assign_MKL.h index 7772951b9..97134ffd7 100644 --- a/Eigen/src/Core/Assign_MKL.h +++ b/Eigen/src/Core/Assign_MKL.h @@ -202,6 +202,7 @@ EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(asin, Asin) EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(cos, Cos) EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(acos, Acos) EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(tan, Tan) +EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(atan, Atan) //EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs, Abs) EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(exp, Exp) EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(log, Ln) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 98313c68f..4523d2263 100755 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -318,6 +318,10 @@ Packet pasin(const Packet& a) { using std::asin; return asin(a); } template EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet pacos(const Packet& a) { using std::acos; return acos(a); } +/** \internal \returns the atan of \a a (coeff-wise) */ +template EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet patan(const Packet& a) { using std::atan; return atan(a); } + /** \internal \returns the exp of \a a (coeff-wise) */ template EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet pexp(const Packet& a) { using std::exp; return exp(a); } diff --git a/Eigen/src/Core/GlobalFunctions.h b/Eigen/src/Core/GlobalFunctions.h index 2acf97723..2067a2a6e 100644 --- a/Eigen/src/Core/GlobalFunctions.h +++ b/Eigen/src/Core/GlobalFunctions.h @@ -45,6 +45,7 @@ namespace Eigen EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(atan,scalar_atan_op) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(exp,scalar_exp_op) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op) diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h index a0fcea3f9..ec42e6850 100644 --- a/Eigen/src/Core/functors/UnaryFunctors.h +++ b/Eigen/src/Core/functors/UnaryFunctors.h @@ -320,6 +320,26 @@ struct functor_traits > }; }; + +/** \internal + * \brief Template functor to compute the atan of a scalar + * \sa class CwiseUnaryOp, ArrayBase::atan() + */ +template struct scalar_atan_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_atan_op) + inline const Scalar operator() (const Scalar& a) const { using std::atan; return atan(a); } + typedef typename packet_traits::type Packet; + inline Packet packetOp(const Packet& a) const { return internal::patan(a); } +}; +template +struct functor_traits > +{ + enum { + Cost = 5 * NumTraits::MulCost, + PacketAccess = packet_traits::HasATan + }; +}; + /** \internal * \brief Template functor to compute the inverse of a scalar * \sa class CwiseUnaryOp, Cwise::inverse() diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index aea3375ed..d2a8ea75b 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -141,6 +141,15 @@ tan() const return derived(); } +/** \returns an expression of the coefficient-wise arc tan of *this. + * + * \sa cos(), sin(), tan() + */ +inline const CwiseUnaryOp, Derived> +atan() const +{ + return derived(); +} /** \returns an expression of the coefficient-wise power of *this to the given exponent. * From 55453c51e8c7d7d69ffa67777ad3355ab9c6f771 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Thu, 19 Jun 2014 15:07:42 +0100 Subject: [PATCH 0533/1103] Add documentation and very simple test for array atan(). --- Eigen/src/plugins/ArrayCwiseUnaryOps.h | 3 +++ 1 file changed, 3 insertions(+) diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index d2a8ea75b..ce462e951 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -142,6 +142,9 @@ tan() const } /** \returns an expression of the coefficient-wise arc tan of *this. + * + * Example: \include Cwise_atan.cpp + * Output: \verbinclude Cwise_atan.out * * \sa cos(), sin(), tan() */ From de150b1e14b5f7fc6f4831b6cd982d3272ca3aca Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Thu, 19 Jun 2014 15:12:33 +0100 Subject: [PATCH 0534/1103] Add documentation and very simple test for array atan(), part 2 (files I forget in the previous commit). --- doc/snippets/Cwise_atan.cpp | 2 ++ test/array.cpp | 1 + 2 files changed, 3 insertions(+) create mode 100644 doc/snippets/Cwise_atan.cpp diff --git a/doc/snippets/Cwise_atan.cpp b/doc/snippets/Cwise_atan.cpp new file mode 100644 index 000000000..446844726 --- /dev/null +++ b/doc/snippets/Cwise_atan.cpp @@ -0,0 +1,2 @@ +ArrayXd v = ArrayXd::LinSpaced(5,0,1); +cout << v.atan() << endl; diff --git a/test/array.cpp b/test/array.cpp index 5f49fc1ea..010fead2d 100644 --- a/test/array.cpp +++ b/test/array.cpp @@ -178,6 +178,7 @@ template void array_real(const ArrayType& m) VERIFY_IS_APPROX(m1.asin(), asin(m1)); VERIFY_IS_APPROX(m1.acos(), acos(m1)); VERIFY_IS_APPROX(m1.tan(), tan(m1)); + VERIFY_IS_APPROX(m1.atan(), atan(m1)); VERIFY_IS_APPROX(cos(m1+RealScalar(3)*m2), cos((m1+RealScalar(3)*m2).eval())); From 1fdef63d1f935283eca4d7735722832eca179a80 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 20 Jun 2014 13:23:33 +0200 Subject: [PATCH 0535/1103] Explain how to export sparse linear problems in matrix-market format. --- doc/SparseLinearSystems.dox | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/doc/SparseLinearSystems.dox b/doc/SparseLinearSystems.dox index c00be10d3..f0456ff52 100644 --- a/doc/SparseLinearSystems.dox +++ b/doc/SparseLinearSystems.dox @@ -140,7 +140,16 @@ x2 = solver.solve(b2); For direct methods, the solution are computed at the machine precision. Sometimes, the solution need not be too accurate. In this case, the iterative methods are more suitable and the desired accuracy can be set before the solve step using \b setTolerance(). For all the available functions, please, refer to the documentation of the \link IterativeLinearSolvers_Module Iterative solvers module \endlink. \section BenchmarkRoutine -Most of the time, all you need is to know how much time it will take to qolve your system, and hopefully, what is the most suitable solver. In Eigen, we provide a benchmark routine that can be used for this purpose. It is very easy to use. In the build directory, navigate to bench/spbench and compile the routine by typing \b make \e spbenchsolver. Run it with --help option to get the list of all available options. Basically, the matrices to test should be in MatrixMarket Coordinate format, and the routine returns the statistics from all available solvers in Eigen. +Most of the time, all you need is to know how much time it will take to qolve your system, and hopefully, what is the most suitable solver. In Eigen, we provide a benchmark routine that can be used for this purpose. It is very easy to use. In the build directory, navigate to bench/spbench and compile the routine by typing \b make \e spbenchsolver. Run it with --help option to get the list of all available options. Basically, the matrices to test should be in MatrixMarket Coordinate format, and the routine returns the statistics from all available solvers in Eigen. + +To export your matrices and right-hand-side vectors in the matrix-market format, you can the the unsupported SparseExtra module: +\code +#include +... +Eigen::saveMarket(A, "filename.mtx"); +Eigen::saveMarket(A, "filename_SPD.mtx", Eigen::Symmetric); // if A is symmetric-positive-definite +Eigen::saveMarketVector(B, "filename_b.mtx"); +\endcode The following table gives an example of XML statistics from several Eigen built-in and external solvers.
From 98ef44fe55925ba8f144889c0ec42be9bf572cc3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 20 Jun 2014 14:43:47 +0200 Subject: [PATCH 0536/1103] Add assertion and warning on the requirements of SparseQR and COLAMDOrdering --- Eigen/src/OrderingMethods/Ordering.h | 8 ++++++-- Eigen/src/SparseQR/SparseQR.h | 17 +++++++++++++++++ 2 files changed, 23 insertions(+), 2 deletions(-) diff --git a/Eigen/src/OrderingMethods/Ordering.h b/Eigen/src/OrderingMethods/Ordering.h index b4da6531a..4e0609784 100644 --- a/Eigen/src/OrderingMethods/Ordering.h +++ b/Eigen/src/OrderingMethods/Ordering.h @@ -109,7 +109,7 @@ class NaturalOrdering * \class COLAMDOrdering * * Functor computing the \em column \em approximate \em minimum \em degree ordering - * The matrix should be in column-major format + * The matrix should be in column-major and \b compressed format (see SparseMatrix::makeCompressed()). */ template class COLAMDOrdering @@ -118,10 +118,14 @@ class COLAMDOrdering typedef PermutationMatrix PermutationType; typedef Matrix IndexVector; - /** Compute the permutation vector form a sparse matrix */ + /** Compute the permutation vector \a perm form the sparse matrix \a mat + * \warning The input sparse matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()). + */ template void operator() (const MatrixType& mat, PermutationType& perm) { + eigen_assert(mat.isCompressed() && "COLAMDOrdering requires a sparse matrix in compressed mode. Call .makeCompressed() before passing it to COLAMDOrdering"); + Index m = mat.rows(); Index n = mat.cols(); Index nnz = mat.nonZeros(); diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 267c48bc3..5fb5bc203 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -58,6 +58,7 @@ namespace internal { * \tparam _OrderingType The fill-reducing ordering method. See the \link OrderingMethods_Module * OrderingMethods \endlink module for the list of built-in and external ordering methods. * + * \warning The input sparse matrix A must be in compressed mode (see SparseMatrix::makeCompressed()). * */ template @@ -77,10 +78,23 @@ class SparseQR SparseQR () : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false) { } + /** Construct a QR factorization of the matrix \a mat. + * + * \warning The matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()). + * + * \sa compute() + */ SparseQR(const MatrixType& mat) : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false) { compute(mat); } + + /** Computes the QR factorization of the sparse matrix \a mat. + * + * \warning The matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()). + * + * \sa analyzePattern(), factorize() + */ void compute(const MatrixType& mat) { analyzePattern(mat); @@ -255,6 +269,8 @@ class SparseQR }; /** \brief Preprocessing step of a QR factorization + * + * \warning The matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()). * * In this step, the fill-reducing permutation is computed and applied to the columns of A * and the column elimination tree is computed as well. Only the sparsity pattern of \a mat is exploited. @@ -264,6 +280,7 @@ class SparseQR template void SparseQR::analyzePattern(const MatrixType& mat) { + eigen_assert(mat.isCompressed() && "SparseQR requires a sparse matrix in compressed mode. Call .makeCompressed() before passing it to SparseQR"); // Compute the column fill reducing ordering OrderingType ord; ord(mat, m_perm_c); From 963d338922e9ef1addcd29c1b43e9b66243207c0 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 20 Jun 2014 15:09:42 +0200 Subject: [PATCH 0537/1103] Fix bug #827: improve accuracy of quaternion to angle-axis conversion --- Eigen/src/Geometry/AngleAxis.h | 28 ++++++++++++++-------------- 1 file changed, 14 insertions(+), 14 deletions(-) diff --git a/Eigen/src/Geometry/AngleAxis.h b/Eigen/src/Geometry/AngleAxis.h index b42048c55..636712c2b 100644 --- a/Eigen/src/Geometry/AngleAxis.h +++ b/Eigen/src/Geometry/AngleAxis.h @@ -77,7 +77,9 @@ public: * represents an invalid rotation. */ template inline AngleAxis(const Scalar& angle, const MatrixBase& axis) : m_axis(axis), m_angle(angle) {} - /** Constructs and initialize the angle-axis rotation from a quaternion \a q. */ + /** Constructs and initialize the angle-axis rotation from a quaternion \a q. + * This function implicitly normalizes the quaternion \a q. + */ template inline explicit AngleAxis(const QuaternionBase& q) { *this = q; } /** Constructs and initialize the angle-axis rotation from a 3x3 rotation matrix. */ template @@ -149,29 +151,27 @@ typedef AngleAxis AngleAxisf; typedef AngleAxis AngleAxisd; /** Set \c *this from a \b unit quaternion. - * The axis is normalized. + * The resulting axis is normalized. * - * \warning As any other method dealing with quaternion, if the input quaternion - * is not normalized then the result is undefined. + * This function implicitly normalizes the quaternion \a q. */ template template AngleAxis& AngleAxis::operator=(const QuaternionBase& q) { - using std::acos; - EIGEN_USING_STD_MATH(min); - EIGEN_USING_STD_MATH(max); - using std::sqrt; - Scalar n2 = q.vec().squaredNorm(); - if (n2 < NumTraits::dummy_precision()*NumTraits::dummy_precision()) + using std::atan2; + Scalar n = q.vec().norm(); + if(n::epsilon()) + n = q.vec().stableNorm(); + if (n > Scalar(0)) { - m_angle = Scalar(0); - m_axis << Scalar(1), Scalar(0), Scalar(0); + m_angle = Scalar(2)*atan2(n, q.w()); + m_axis = q.vec() / n; } else { - m_angle = Scalar(2)*acos((min)((max)(Scalar(-1),q.w()),Scalar(1))); - m_axis = q.vec() / sqrt(n2); + m_angle = 0; + m_axis << 1, 0, 0; } return *this; } From 78bb80833708615c330659d9b64870b19185df37 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 20 Jun 2014 15:39:38 +0200 Subject: [PATCH 0538/1103] 1- Introduce sub-evaluator types for unary, binary, product, and map expressions to ease specializing them. 2- Remove a lot of code which should not be there with evaluators, in particular coeff/packet methods implemented in the expressions. --- Eigen/src/Core/ArrayBase.h | 4 + Eigen/src/Core/ArrayWrapper.h | 2 + Eigen/src/Core/Block.h | 13 +- Eigen/src/Core/CoreEvaluators.h | 247 ++++++++++-------- Eigen/src/Core/CwiseBinaryOp.h | 30 ++- Eigen/src/Core/CwiseUnaryOp.h | 14 +- Eigen/src/Core/CwiseUnaryView.h | 9 +- Eigen/src/Core/DenseBase.h | 5 +- Eigen/src/Core/DenseCoeffsBase.h | 45 +++- Eigen/src/Core/Inverse.h | 9 +- Eigen/src/Core/Replicate.h | 1 + Eigen/src/Core/ReturnByValue.h | 28 ++ Eigen/src/Core/Reverse.h | 1 + Eigen/src/Core/Transpose.h | 30 ++- Eigen/src/Core/TriangularMatrix.h | 4 +- Eigen/src/Core/products/GeneralMatrixMatrix.h | 4 +- Eigen/src/Core/util/ForwardDeclarations.h | 4 +- Eigen/src/Core/util/XprHelper.h | 9 + 18 files changed, 300 insertions(+), 159 deletions(-) diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index d1c422836..f5bae6357 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -123,7 +123,11 @@ template class ArrayBase EIGEN_DEVICE_FUNC Derived& operator=(const ArrayBase& other) { +#ifndef EIGEN_TEST_EVALUATORS return internal::assign_selector::run(derived(), other.derived()); +#else + internal::call_assignment(derived(), other.derived()); +#endif } EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/ArrayWrapper.h b/Eigen/src/Core/ArrayWrapper.h index 4bb648024..599d87f64 100644 --- a/Eigen/src/Core/ArrayWrapper.h +++ b/Eigen/src/Core/ArrayWrapper.h @@ -39,6 +39,7 @@ class ArrayWrapper : public ArrayBase > typedef ArrayBase Base; EIGEN_DENSE_PUBLIC_INTERFACE(ArrayWrapper) EIGEN_INHERIT_ASSIGNMENT_OPERATORS(ArrayWrapper) + typedef typename internal::remove_all::type NestedExpression; typedef typename internal::conditional< internal::is_lvalue::value, @@ -176,6 +177,7 @@ class MatrixWrapper : public MatrixBase > typedef MatrixBase > Base; EIGEN_DENSE_PUBLIC_INTERFACE(MatrixWrapper) EIGEN_INHERIT_ASSIGNMENT_OPERATORS(MatrixWrapper) + typedef typename internal::remove_all::type NestedExpression; typedef typename internal::conditional< internal::is_lvalue::value, diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index e0b24e199..ef6c143d3 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -344,6 +344,9 @@ class BlockImpl_dense : public MapBase > { typedef Block BlockType; + enum { + XprTypeIsRowMajor = (int(traits::Flags)&RowMajorBit) != 0 + }; public: typedef MapBase Base; @@ -354,9 +357,8 @@ class BlockImpl_dense */ EIGEN_DEVICE_FUNC inline BlockImpl_dense(XprType& xpr, Index i) - : Base(internal::const_cast_ptr(&xpr.coeffRef( - (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0, - (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0)), + : Base(xpr.data() + i * ( ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && (!XprTypeIsRowMajor)) + || ((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && ( XprTypeIsRowMajor)) ? xpr.innerStride() : xpr.outerStride()), BlockRows==1 ? 1 : xpr.rows(), BlockCols==1 ? 1 : xpr.cols()), m_xpr(xpr) @@ -368,7 +370,8 @@ class BlockImpl_dense */ EIGEN_DEVICE_FUNC inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol) - : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol))), m_xpr(xpr) + : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol)), + m_xpr(xpr) { init(); } @@ -379,7 +382,7 @@ class BlockImpl_dense inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) - : Base(internal::const_cast_ptr(&xpr.coeffRef(startRow,startCol)), blockRows, blockCols), + : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol), blockRows, blockCols), m_xpr(xpr) { init(); diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 47f50d548..f872b41e1 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -27,14 +27,6 @@ struct storage_kind_to_evaluator_kind { typedef IndexBased Kind; }; -// TODO to be moved to SparseCore: -/* -template<> -struct storage_kind_to_evaluator_kind { - typedef IteratorBased Kind -}; -*/ - // This class returns the evaluator shape from the expression storage kind. // It can be Dense, Sparse, Triangular, Diagonal, SelfAdjoint, Band, etc. template struct storage_kind_to_shape; @@ -45,22 +37,28 @@ struct storage_kind_to_shape { typedef DenseShape Shape; }; -// TODO to be moved to SparseCore: -/* -template<> -struct storage_kind_to_shape { - typedef SparseSpape Shape; -}; -*/ +// Evaluators have to be specialized with respect to various criteria such as: +// - storage/structure/shape +// - scalar type +// - etc. +// Therefore, we need specialization of evaluator providing additional template arguments for each kind of evaluators. +// We currently distinguish the following kind of evaluators: +// - unary_evaluator for expressions taking only one arguments (CwiseUnaryOp, CwiseUnaryView, Transpose, MatrixWrapper, ArrayWrapper, Reverse, Replicate) +// - binary_evaluator for expression taking two arguments (CwiseBinaryOp) +// - product_evaluator for linear algebra products (Product); special case of binary_evaluator because it requires additional tags for dispatching. +// - mapbase_evaluator for Map, Block, Ref +// - block_evaluator for Block (special dispatching to a mapbase_evaluator or unary_evaluator) - - template< typename T, - typename LhsKind = typename evaluator_traits::Kind, - typename RhsKind = typename evaluator_traits::Kind, + typename LhsKind = typename evaluator_traits::Kind, + typename RhsKind = typename evaluator_traits::Kind, typename LhsScalar = typename T::Lhs::Scalar, typename RhsScalar = typename T::Rhs::Scalar> struct binary_evaluator; +template< typename T, + typename Kind = typename evaluator_traits::Kind, + typename Scalar = typename T::Scalar> struct unary_evaluator; + // evaluator_traits contains traits for evaluator template @@ -80,15 +78,20 @@ struct evaluator_traits_base static const int AssumeAliasing = 0; }; +// Default evaluator traits template struct evaluator_traits : public evaluator_traits_base { }; -// expression class for evaluating nested expression to a temporary - -template -class EvalToTemp; + +// By default, we assume a unary expression: +template +struct evaluator : public unary_evaluator +{ + typedef unary_evaluator Base; + evaluator(const T& xpr) : Base(xpr) {} +}; // TODO: Think about const-correctness @@ -118,6 +121,8 @@ struct evaluator_base // // evaluator is a common base class for the // Matrix and Array evaluators. +// Here we directly specialize evaluator. This is not really a unary expression, and it is, by definition, dense, +// so no need for more sophisticated dispatching. template struct evaluator > @@ -245,81 +250,10 @@ struct evaluator > { } }; -// -------------------- EvalToTemp -------------------- - -template -struct traits > - : public traits -{ }; - -template -class EvalToTemp - : public dense_xpr_base >::type -{ - public: - - typedef typename dense_xpr_base::type Base; - EIGEN_GENERIC_PUBLIC_INTERFACE(EvalToTemp) - - EvalToTemp(const ArgType& arg) - : m_arg(arg) - { } - - const ArgType& arg() const - { - return m_arg; - } - - Index rows() const - { - return m_arg.rows(); - } - - Index cols() const - { - return m_arg.cols(); - } - - private: - const ArgType& m_arg; -}; - -template -struct evaluator > - : public evaluator::type -{ - typedef EvalToTemp XprType; - typedef typename ArgType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - typedef evaluator type; - typedef evaluator nestedType; - - evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - // TODO we should simply do m_result(xpr.arg()); - call_dense_assignment_loop(m_result, xpr.arg()); - } - - // This constructor is used when nesting an EvalTo evaluator in another evaluator - evaluator(const ArgType& arg) - : m_result(arg.rows(), arg.cols()) - { - ::new (static_cast(this)) Base(m_result); - // TODO we should simply do m_result(xpr.arg()); - call_dense_assignment_loop(m_result, arg); - } - -protected: - PlainObject m_result; -}; - // -------------------- Transpose -------------------- template -struct evaluator > +struct unary_evaluator > : evaluator_base > { typedef Transpose XprType; @@ -329,7 +263,7 @@ struct evaluator > Flags = evaluator::Flags ^ RowMajorBit }; - evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} + unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -386,6 +320,8 @@ protected: }; // -------------------- CwiseNullaryOp -------------------- +// Like Matrix and Array, this is not really a unary expression, so we directly specialize evaluator. +// Likewise, there is not need to more sophisticated dispatching here. template struct evaluator > @@ -441,7 +377,7 @@ protected: // -------------------- CwiseUnaryOp -------------------- template -struct evaluator > +struct unary_evaluator, IndexBased > : evaluator_base > { typedef CwiseUnaryOp XprType; @@ -454,7 +390,7 @@ struct evaluator > | (functor_traits::PacketAccess ? PacketAccessBit : 0)) }; - evaluator(const XprType& op) + unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -492,8 +428,19 @@ protected: // -------------------- CwiseBinaryOp -------------------- +// this is a binary expression template struct evaluator > + : public binary_evaluator > +{ + typedef CwiseBinaryOp XprType; + typedef binary_evaluator > Base; + + evaluator(const XprType& xpr) : Base(xpr) {} +}; + +template +struct binary_evaluator > : evaluator_base > { typedef CwiseBinaryOp XprType; @@ -517,7 +464,7 @@ struct evaluator > Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit) }; - evaluator(const XprType& xpr) + binary_evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) @@ -560,7 +507,7 @@ protected: // -------------------- CwiseUnaryView -------------------- template -struct evaluator > +struct unary_evaluator > : evaluator_base > { typedef CwiseUnaryView XprType; @@ -571,7 +518,7 @@ struct evaluator > Flags = (evaluator::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit)) }; - evaluator(const XprType& op) + unary_evaluator(const XprType& op) : m_unaryOp(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -884,6 +831,7 @@ struct block_evaluator @@ -934,7 +882,7 @@ protected: // -------------------- Replicate -------------------- template -struct evaluator > +struct unary_evaluator > : evaluator_base > { typedef Replicate XprType; @@ -953,7 +901,7 @@ struct evaluator > Flags = (evaluator::Flags & HereditaryBits & ~RowMajorBit) | (traits::Flags & RowMajorBit) }; - evaluator(const XprType& replicate) + unary_evaluator(const XprType& replicate) : m_arg(replicate.nestedExpression()), m_argImpl(m_arg), m_rows(replicate.nestedExpression().rows()), @@ -1111,23 +1059,23 @@ protected: }; template -struct evaluator > +struct unary_evaluator > : evaluator_wrapper_base > { typedef MatrixWrapper XprType; - evaluator(const XprType& wrapper) + unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; template -struct evaluator > +struct unary_evaluator > : evaluator_wrapper_base > { typedef ArrayWrapper XprType; - evaluator(const XprType& wrapper) + unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; @@ -1139,7 +1087,7 @@ struct evaluator > template struct reverse_packet_cond; template -struct evaluator > +struct unary_evaluator > : evaluator_base > { typedef Reverse XprType; @@ -1173,7 +1121,7 @@ struct evaluator > }; typedef internal::reverse_packet_cond reverse_packet; - evaluator(const XprType& reverse) + unary_evaluator(const XprType& reverse) : m_argImpl(reverse.nestedExpression()), m_rows(ReverseRow ? reverse.nestedExpression().rows() : 0), m_cols(ReverseCol ? reverse.nestedExpression().cols() : 0) @@ -1292,6 +1240,87 @@ private: EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; } }; + + +//---------------------------------------------------------------------- +// deprecated code +//---------------------------------------------------------------------- + +// -------------------- EvalToTemp -------------------- + +// expression class for evaluating nested expression to a temporary + +template class EvalToTemp; + +template +struct traits > + : public traits +{ }; + +template +class EvalToTemp + : public dense_xpr_base >::type +{ + public: + + typedef typename dense_xpr_base::type Base; + EIGEN_GENERIC_PUBLIC_INTERFACE(EvalToTemp) + + EvalToTemp(const ArgType& arg) + : m_arg(arg) + { } + + const ArgType& arg() const + { + return m_arg; + } + + Index rows() const + { + return m_arg.rows(); + } + + Index cols() const + { + return m_arg.cols(); + } + + private: + const ArgType& m_arg; +}; + +template +struct evaluator > + : public evaluator::type +{ + typedef EvalToTemp XprType; + typedef typename ArgType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + // TODO we should simply do m_result(xpr.arg()); + call_dense_assignment_loop(m_result, xpr.arg()); + } + + // This constructor is used when nesting an EvalTo evaluator in another evaluator + evaluator(const ArgType& arg) + : m_result(arg.rows(), arg.cols()) + { + ::new (static_cast(this)) Base(m_result); + // TODO we should simply do m_result(xpr.arg()); + call_dense_assignment_loop(m_result, arg); + } + +protected: + PlainObject m_result; +}; + } // namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index 07861dbc9..5e4fb147b 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -94,23 +94,26 @@ struct traits > template class CwiseBinaryOpImpl; -template +template class CwiseBinaryOp : internal::no_assignment_operator, public CwiseBinaryOpImpl< - BinaryOp, Lhs, Rhs, - typename internal::promote_storage_type::StorageKind, - typename internal::traits::StorageKind>::ret> + BinaryOp, LhsType, RhsType, + typename internal::promote_storage_type::StorageKind, + typename internal::traits::StorageKind>::ret> { public: + + typedef typename internal::remove_all::type Lhs; + typedef typename internal::remove_all::type Rhs; typedef typename CwiseBinaryOpImpl< - BinaryOp, Lhs, Rhs, - typename internal::promote_storage_type::StorageKind, + BinaryOp, LhsType, RhsType, + typename internal::promote_storage_type::StorageKind, typename internal::traits::StorageKind>::ret>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseBinaryOp) - typedef typename internal::nested::type LhsNested; - typedef typename internal::nested::type RhsNested; + typedef typename internal::nested::type LhsNested; + typedef typename internal::nested::type RhsNested; typedef typename internal::remove_reference::type _LhsNested; typedef typename internal::remove_reference::type _RhsNested; @@ -157,6 +160,7 @@ class CwiseBinaryOp : internal::no_assignment_operator, const BinaryOp m_functor; }; +#ifndef EIGEN_TEST_EVALUATORS template class CwiseBinaryOpImpl : public internal::dense_xpr_base >::type @@ -195,6 +199,16 @@ class CwiseBinaryOpImpl derived().rhs().template packet(index)); } }; +#else +// Generic API dispatcher +template +class CwiseBinaryOpImpl + : public internal::generic_xpr_base >::type +{ +public: + typedef typename internal::generic_xpr_base >::type Base; +}; +#endif /** replaces \c *this by \c *this - \a other. * diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index af05a9108..098034a1e 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -67,6 +67,7 @@ class CwiseUnaryOp : internal::no_assignment_operator, typedef typename CwiseUnaryOpImpl::StorageKind>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryOp) + typedef typename internal::remove_all::type NestedExpression; EIGEN_DEVICE_FUNC inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) @@ -96,6 +97,7 @@ class CwiseUnaryOp : internal::no_assignment_operator, const UnaryOp m_functor; }; +#ifndef EIGEN_TEST_EVALUATORS // This is the generic implementation for dense storage. // It can be used for any expression types implementing the dense concept. template @@ -107,7 +109,7 @@ class CwiseUnaryOpImpl typedef CwiseUnaryOp Derived; typedef typename internal::dense_xpr_base >::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) - + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const { @@ -133,6 +135,16 @@ class CwiseUnaryOpImpl return derived().functor().packetOp(derived().nestedExpression().template packet(index)); } }; +#else +// Generic API dispatcher +template +class CwiseUnaryOpImpl + : public internal::generic_xpr_base >::type +{ +public: + typedef typename internal::generic_xpr_base >::type Base; +}; +#endif } // end namespace Eigen diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index 9cdebb8e7..92b031e19 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -66,6 +66,7 @@ class CwiseUnaryView : public CwiseUnaryViewImpl::StorageKind>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryView) + typedef typename internal::remove_all::type NestedExpression; inline CwiseUnaryView(const MatrixType& mat, const ViewOp& func = ViewOp()) : m_matrix(mat), m_functor(func) {} @@ -104,8 +105,8 @@ class CwiseUnaryViewImpl EIGEN_DENSE_PUBLIC_INTERFACE(Derived) EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl) - inline Scalar* data() { return &coeffRef(0); } - inline const Scalar* data() const { return &coeff(0); } + inline Scalar* data() { return &(this->coeffRef(0)); } + inline const Scalar* data() const { return &(this->coeff(0)); } inline Index innerStride() const { @@ -116,6 +117,8 @@ class CwiseUnaryViewImpl { return derived().nestedExpression().outerStride() * sizeof(typename internal::traits::Scalar) / sizeof(Scalar); } + +#ifndef EIGEN_TEST_EVALUATORS EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const { @@ -136,6 +139,8 @@ class CwiseUnaryViewImpl { return derived().functor()(const_cast_derived().nestedExpression().coeffRef(index)); } + +#endif }; } // end namespace Eigen diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index f3c79aa31..624276240 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -74,6 +74,7 @@ template class DenseBase using Base::colIndexByOuterInner; using Base::coeff; using Base::coeffByOuterInner; +#ifndef EIGEN_TEST_EVALUATORS using Base::packet; using Base::packetByOuterInner; using Base::writePacket; @@ -84,6 +85,7 @@ template class DenseBase using Base::copyCoeffByOuterInner; using Base::copyPacket; using Base::copyPacketByOuterInner; +#endif using Base::operator(); using Base::operator[]; using Base::x; @@ -280,7 +282,8 @@ template class DenseBase Derived& operator=(const ReturnByValue& func); #ifndef EIGEN_PARSED_BY_DOXYGEN - /** Copies \a other into *this without evaluating other. \returns a reference to *this. */ + /** Copies \a other into *this without evaluating other. \returns a reference to *this. + * \deprecated */ template EIGEN_DEVICE_FUNC Derived& lazyAssign(const DenseBase& other); diff --git a/Eigen/src/Core/DenseCoeffsBase.h b/Eigen/src/Core/DenseCoeffsBase.h index efabb5e67..6f35a67ca 100644 --- a/Eigen/src/Core/DenseCoeffsBase.h +++ b/Eigen/src/Core/DenseCoeffsBase.h @@ -97,8 +97,12 @@ class DenseCoeffsBase : public EigenBase EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const { eigen_internal_assert(row >= 0 && row < rows() - && col >= 0 && col < cols()); + && col >= 0 && col < cols()); +#ifndef EIGEN_TEST_EVALUATORS return derived().coeff(row, col); +#else + return typename internal::evaluator::type(derived()).coeff(row,col); +#endif } EIGEN_DEVICE_FUNC @@ -117,7 +121,7 @@ class DenseCoeffsBase : public EigenBase { eigen_assert(row >= 0 && row < rows() && col >= 0 && col < cols()); - return derived().coeff(row, col); + return coeff(row, col); } /** Short version: don't use this function, use @@ -140,7 +144,11 @@ class DenseCoeffsBase : public EigenBase coeff(Index index) const { eigen_internal_assert(index >= 0 && index < size()); +#ifndef EIGEN_TEST_EVALUATORS return derived().coeff(index); +#else + return typename internal::evaluator::type(derived()).coeff(index); +#endif } @@ -161,7 +169,7 @@ class DenseCoeffsBase : public EigenBase THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD) #endif eigen_assert(index >= 0 && index < size()); - return derived().coeff(index); + return coeff(index); } /** \returns the coefficient at given index. @@ -179,7 +187,7 @@ class DenseCoeffsBase : public EigenBase operator()(Index index) const { eigen_assert(index >= 0 && index < size()); - return derived().coeff(index); + return coeff(index); } /** equivalent to operator[](0). */ @@ -219,9 +227,12 @@ class DenseCoeffsBase : public EigenBase template EIGEN_STRONG_INLINE PacketReturnType packet(Index row, Index col) const { - eigen_internal_assert(row >= 0 && row < rows() - && col >= 0 && col < cols()); + eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols()); +#ifndef EIGEN_TEST_EVALUATORS return derived().template packet(row,col); +#else + return typename internal::evaluator::type(derived()).template packet(row,col); +#endif } @@ -247,7 +258,11 @@ class DenseCoeffsBase : public EigenBase EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { eigen_internal_assert(index >= 0 && index < size()); +#ifndef EIGEN_TEST_EVALUATORS return derived().template packet(index); +#else + return typename internal::evaluator::type(derived()).template packet(index); +#endif } protected: @@ -327,8 +342,12 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && row < rows() - && col >= 0 && col < cols()); + && col >= 0 && col < cols()); +#ifndef EIGEN_TEST_EVALUATORS return derived().coeffRef(row, col); +#else + return typename internal::evaluator::type(derived()).coeffRef(row,col); +#endif } EIGEN_DEVICE_FUNC @@ -350,7 +369,7 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && row < rows() && col >= 0 && col < cols()); - return derived().coeffRef(row, col); + return coeffRef(row, col); } @@ -374,7 +393,11 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && index < size()); +#ifndef EIGEN_TEST_EVALUATORS return derived().coeffRef(index); +#else + return typename internal::evaluator::type(derived()).coeffRef(index); +#endif } /** \returns a reference to the coefficient at given index. @@ -393,7 +416,7 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && index < size()); - return derived().coeffRef(index); + return coeffRef(index); } /** \returns a reference to the coefficient at given index. @@ -410,7 +433,7 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && index < size()); - return derived().coeffRef(index); + return coeffRef(index); } /** equivalent to operator[](0). */ @@ -437,6 +460,7 @@ class DenseCoeffsBase : public DenseCoeffsBase : public DenseCoeffsBase(row, col, other); } #endif +#endif // EIGEN_TEST_EVALUATORS }; diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 57eaf99f1..788d0664d 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -81,6 +81,7 @@ public: typedef MatrixBase Base; EIGEN_DENSE_PUBLIC_INTERFACE(Derived) + typedef typename internal::remove_all::type NestedExpression; private: @@ -101,19 +102,19 @@ namespace internal { * \sa class Inverse */ template -struct evaluator > +struct unary_evaluator > : public evaluator::PlainObject>::type { typedef Inverse InverseType; typedef typename InverseType::PlainObject PlainObject; typedef typename evaluator::type Base; - typedef evaluator type; - typedef evaluator nestedType; + typedef evaluator type; + typedef evaluator nestedType; enum { Flags = Base::Flags | EvalBeforeNestingBit }; - evaluator(const InverseType& inv_xpr) + unary_evaluator(const InverseType& inv_xpr) : m_result(inv_xpr.rows(), inv_xpr.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/Core/Replicate.h b/Eigen/src/Core/Replicate.h index 2dff03ea3..e63f0d421 100644 --- a/Eigen/src/Core/Replicate.h +++ b/Eigen/src/Core/Replicate.h @@ -74,6 +74,7 @@ template class Replicate typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Replicate) + typedef typename internal::remove_all::type NestedExpression; template inline explicit Replicate(const OriginalMatrixType& a_matrix) diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index d63b96138..30ade1b75 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -92,6 +92,34 @@ Derived& DenseBase::operator=(const ReturnByValue& other) return derived(); } +#ifdef EIGEN_TEST_EVALUATORS +namespace internal { + +template +struct evaluator > + : public evaluator::ReturnType>::type +{ + typedef ReturnByValue XprType; + typedef typename internal::traits::ReturnType PlainObject; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + xpr.evalTo(m_result); + } + +protected: + PlainObject m_result; +}; + +} // end namespace internal +#endif + } // end namespace Eigen #endif // EIGEN_RETURNBYVALUE_H diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h index 4969bb4fc..ceb6e4701 100644 --- a/Eigen/src/Core/Reverse.h +++ b/Eigen/src/Core/Reverse.h @@ -77,6 +77,7 @@ template class Reverse typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(Reverse) + typedef typename internal::remove_all::type NestedExpression; using Base::IsRowMajor; // next line is necessary because otherwise const version of operator() diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 079f21fd9..4b605dfd6 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -68,6 +68,7 @@ template class Transpose typedef typename TransposeImpl::StorageKind>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose) + typedef typename internal::remove_all::type NestedExpression; EIGEN_DEVICE_FUNC inline Transpose(MatrixType& a_matrix) : m_matrix(a_matrix) {} @@ -113,6 +114,7 @@ template class TransposeImpl public: typedef typename internal::TransposeImpl_base::type Base; + using Base::coeffRef; EIGEN_DENSE_PUBLIC_INTERFACE(Transpose) EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TransposeImpl) @@ -127,6 +129,8 @@ template class TransposeImpl inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } inline const Scalar* data() const { return derived().nestedExpression().data(); } + +#ifndef EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue& coeffRef(Index rowId, Index colId) @@ -142,18 +146,6 @@ template class TransposeImpl return derived().nestedExpression().const_cast_derived().coeffRef(index); } - EIGEN_DEVICE_FUNC - inline const Scalar& coeffRef(Index rowId, Index colId) const - { - return derived().nestedExpression().coeffRef(colId, rowId); - } - - EIGEN_DEVICE_FUNC - inline const Scalar& coeffRef(Index index) const - { - return derived().nestedExpression().coeffRef(index); - } - EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index rowId, Index colId) const { @@ -189,6 +181,20 @@ template class TransposeImpl { derived().nestedExpression().const_cast_derived().template writePacket(index, x); } +#endif + + // FIXME: shall we keep the const version of coeffRef? + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index rowId, Index colId) const + { + return derived().nestedExpression().coeffRef(colId, rowId); + } + + EIGEN_DEVICE_FUNC + inline const Scalar& coeffRef(Index index) const + { + return derived().nestedExpression().coeffRef(index); + } }; /** \returns an expression of the transpose of *this. diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 064e0e10d..180efdd62 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -1071,8 +1071,8 @@ struct evaluator_traits > static const int AssumeAliasing = 0; }; -template -struct evaluator, Kind, typename MatrixType::Scalar> +template +struct evaluator > : evaluator::type> { typedef TriangularView XprType; diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 1726f98ed..b0a09216d 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -219,8 +219,8 @@ struct gemm_functor cols = m_rhs.cols(); Gemm::run(rows, cols, m_lhs.cols(), - /*(const Scalar*)*/&m_lhs.coeffRef(row,0), m_lhs.outerStride(), - /*(const Scalar*)*/&m_rhs.coeffRef(0,col), m_rhs.outerStride(), + &m_lhs.coeffRef(row,0), m_lhs.outerStride(), + &m_rhs.coeffRef(0,col), m_rhs.outerStride(), (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(), m_actualAlpha, m_blocking, info); } diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 092ba758e..ead5de650 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -38,9 +38,7 @@ template struct accessors_level template struct evaluator_traits; -template< typename T, - typename Kind = typename evaluator_traits::Kind, - typename Scalar = typename T::Scalar> struct evaluator; +template< typename T> struct evaluator; } // end namespace internal diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 0be5029c9..5407645c9 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -465,6 +465,15 @@ struct dense_xpr_base typedef ArrayBase type; }; +template::XprKind, typename StorageKind = typename traits::StorageKind> +struct generic_xpr_base; + +template +struct generic_xpr_base +{ + typedef typename dense_xpr_base::type type; +}; + /** \internal Helper base class to add a scalar multiple operator * overloads for complex types */ template Date: Fri, 20 Jun 2014 15:42:13 +0200 Subject: [PATCH 0539/1103] Started to move the SparseCore module to evaluators: implemented assignment and cwise-unary evaluator --- Eigen/SparseCore | 5 +- Eigen/src/SparseCore/SparseAssign.h | 252 ++++++++++++++++++++++ Eigen/src/SparseCore/SparseCwiseUnaryOp.h | 150 ++++++++++++- Eigen/src/SparseCore/SparseMatrix.h | 93 ++++++++ Eigen/src/SparseCore/SparseMatrixBase.h | 83 +------ Eigen/src/SparseCore/SparseUtil.h | 6 + 6 files changed, 508 insertions(+), 81 deletions(-) create mode 100644 Eigen/src/SparseCore/SparseAssign.h diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 9b5be5e15..c0f0a4121 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -35,14 +35,16 @@ struct Sparse {}; #include "src/SparseCore/SparseUtil.h" #include "src/SparseCore/SparseMatrixBase.h" +#include "src/SparseCore/SparseAssign.h" #include "src/SparseCore/CompressedStorage.h" #include "src/SparseCore/AmbiVector.h" #include "src/SparseCore/SparseMatrix.h" #include "src/SparseCore/MappedSparseMatrix.h" #include "src/SparseCore/SparseVector.h" +#include "src/SparseCore/SparseCwiseUnaryOp.h" +#ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparseBlock.h" #include "src/SparseCore/SparseTranspose.h" -#include "src/SparseCore/SparseCwiseUnaryOp.h" #include "src/SparseCore/SparseCwiseBinaryOp.h" #include "src/SparseCore/SparseDot.h" #include "src/SparseCore/SparsePermutation.h" @@ -57,6 +59,7 @@ struct Sparse {}; #include "src/SparseCore/SparseSelfAdjointView.h" #include "src/SparseCore/TriangularSolver.h" #include "src/SparseCore/SparseView.h" +#endif #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h new file mode 100644 index 000000000..99a1a8712 --- /dev/null +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -0,0 +1,252 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2014 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SPARSEASSIGN_H +#define EIGEN_SPARSEASSIGN_H + +namespace Eigen { + +#ifndef EIGEN_TEST_EVALUATORS + +template +template +Derived& SparseMatrixBase::operator=(const EigenBase &other) +{ + other.derived().evalTo(derived()); + return derived(); +} + +template +template +Derived& SparseMatrixBase::operator=(const ReturnByValue& other) +{ + other.evalTo(derived()); + return derived(); +} + +template +template +inline Derived& SparseMatrixBase::operator=(const SparseMatrixBase& other) +{ + return assign(other.derived()); +} + +template +inline Derived& SparseMatrixBase::operator=(const Derived& other) +{ +// if (other.isRValue()) +// derived().swap(other.const_cast_derived()); +// else + return assign(other.derived()); +} + +template +template +inline Derived& SparseMatrixBase::assign(const OtherDerived& other) +{ + const bool transpose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); + const Index outerSize = (int(OtherDerived::Flags) & RowMajorBit) ? other.rows() : other.cols(); + if ((!transpose) && other.isRValue()) + { + // eval without temporary + derived().resize(other.rows(), other.cols()); + derived().setZero(); + derived().reserve((std::max)(this->rows(),this->cols())*2); + for (Index j=0; j +template +inline void SparseMatrixBase::assignGeneric(const OtherDerived& other) +{ + //const bool transpose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); + eigen_assert(( ((internal::traits::SupportedAccessPatterns&OuterRandomAccessPattern)==OuterRandomAccessPattern) || + (!((Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit)))) && + "the transpose operation is supposed to be handled in SparseMatrix::operator="); + + enum { Flip = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit) }; + + const Index outerSize = other.outerSize(); + //typedef typename internal::conditional, Derived>::type TempType; + // thanks to shallow copies, we always eval to a tempary + Derived temp(other.rows(), other.cols()); + + temp.reserve((std::max)(this->rows(),this->cols())*2); + for (Index j=0; j +// inline Derived& operator=(const SparseSparseProduct& product); +// +// template +// Derived& operator+=(const SparseMatrixBase& other); +// template +// Derived& operator-=(const SparseMatrixBase& other); +// +// Derived& operator*=(const Scalar& other); +// Derived& operator/=(const Scalar& other); +// +// template +// Derived& operator*=(const SparseMatrixBase& other); + +#else // EIGEN_TEST_EVALUATORS + +template +template +Derived& SparseMatrixBase::operator=(const EigenBase &other) +{ + other.derived().evalTo(derived()); + return derived(); +} + +template +template +Derived& SparseMatrixBase::operator=(const ReturnByValue& other) +{ + other.evalTo(derived()); + return derived(); +} + +template +template +inline Derived& SparseMatrixBase::operator=(const SparseMatrixBase& other) +{ + internal::call_assignment_no_alias(derived(), other.derived()); + return derived(); +} + +template +inline Derived& SparseMatrixBase::operator=(const Derived& other) +{ + internal::call_assignment_no_alias(derived(), other.derived()); + return derived(); +} + +namespace internal { + +template<> +struct storage_kind_to_evaluator_kind { + typedef IteratorBased Kind; +}; + +template<> +struct storage_kind_to_shape { + typedef SparseShape Shape; +}; + +struct Sparse2Sparse {}; + +template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; + + +template +void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src) +{ + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + typedef typename DstXprType::Index Index; + typedef typename DstXprType::Scalar Scalar; + typedef typename internal::evaluator::type DstEvaluatorType; + typedef typename internal::evaluator::type SrcEvaluatorType; + + DstEvaluatorType dstEvaluator(dst); + SrcEvaluatorType srcEvaluator(src); + + const bool transpose = (DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit); + const Index outerSize = (int(SrcEvaluatorType::Flags) & RowMajorBit) ? src.rows() : src.cols(); + if ((!transpose) && src.isRValue()) + { + // eval without temporary + dst.resize(src.rows(), src.cols()); + dst.setZero(); + dst.reserve((std::max)(src.rows(),src.cols())*2); + for (Index j=0; j::SupportedAccessPatterns & OuterRandomAccessPattern)==OuterRandomAccessPattern) || + (!((DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit)))) && + "the transpose operation is supposed to be handled in SparseMatrix::operator="); + + enum { Flip = (DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit) }; + + const Index outerSize = src.outerSize(); + DstXprType temp(src.rows(), src.cols()); + + temp.reserve((std::max)(src.rows(),src.cols())*2); + for (Index j=0; j +struct Assignment +{ + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) + { + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + assign_sparse_to_sparse(dst.derived(), src.derived()); + } +}; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS + +} // end namespace Eigen + +#endif // EIGEN_SPARSEASSIGN_H diff --git a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h index 5a50c7803..c5042504b 100644 --- a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h @@ -11,6 +11,8 @@ #define EIGEN_SPARSE_CWISE_UNARY_OP_H namespace Eigen { + +#ifndef EIGEN_TEST_EVALUATORS template class CwiseUnaryOpImpl @@ -18,12 +20,12 @@ class CwiseUnaryOpImpl { public: - class InnerIterator; - class ReverseInnerIterator; - typedef CwiseUnaryOp Derived; EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) + class InnerIterator; + class ReverseInnerIterator; + protected: typedef typename internal::traits::_XprTypeNested _MatrixTypeNested; typedef typename _MatrixTypeNested::InnerIterator MatrixTypeIterator; @@ -138,6 +140,148 @@ class CwiseUnaryViewImpl::ReverseInnerIterator const ViewOp m_functor; }; +#else // EIGEN_TEST_EVALUATORS + +namespace internal { + +template +struct unary_evaluator, IteratorBased> + : public evaluator_base > +{ + public: + typedef CwiseUnaryOp XprType; + + class InnerIterator; + class ReverseInnerIterator; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost, + Flags = XprType::Flags + }; + + unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) {} + + protected: + typedef typename evaluator::InnerIterator EvalIterator; + typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; + + const UnaryOp m_functor; + typename evaluator::nestedType m_argImpl; +}; + +template +class unary_evaluator, IteratorBased>::InnerIterator + : public unary_evaluator, IteratorBased>::EvalIterator +{ + typedef typename XprType::Scalar Scalar; + typedef typename unary_evaluator, IteratorBased>::EvalIterator Base; + public: + + EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& unaryOp, typename XprType::Index outer) + : Base(unaryOp.m_argImpl,outer), m_functor(unaryOp.m_functor) + {} + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { Base::operator++(); return *this; } + + EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } + + protected: + const UnaryOp m_functor; + private: + Scalar& valueRef(); +}; + +template +class unary_evaluator, IteratorBased>::ReverseInnerIterator + : public unary_evaluator, IteratorBased>::EvalReverseIterator +{ + typedef typename XprType::Scalar Scalar; + typedef typename unary_evaluator, IteratorBased>::EvalReverseIterator Base; + public: + + EIGEN_STRONG_INLINE ReverseInnerIterator(const XprType& unaryOp, typename XprType::Index outer) + : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) + {} + + EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() + { Base::operator--(); return *this; } + + EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } + + protected: + const UnaryOp m_functor; + private: + Scalar& valueRef(); +}; + + +// template +// class CwiseUnaryViewImpl +// : public SparseMatrixBase > +// { +// public: +// +// class InnerIterator; +// class ReverseInnerIterator; +// +// typedef CwiseUnaryView Derived; +// EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) +// +// protected: +// typedef typename internal::traits::_MatrixTypeNested _MatrixTypeNested; +// typedef typename _MatrixTypeNested::InnerIterator MatrixTypeIterator; +// typedef typename _MatrixTypeNested::ReverseInnerIterator MatrixTypeReverseIterator; +// }; +// +// template +// class CwiseUnaryViewImpl::InnerIterator +// : public CwiseUnaryViewImpl::MatrixTypeIterator +// { +// typedef typename CwiseUnaryViewImpl::Scalar Scalar; +// typedef typename CwiseUnaryViewImpl::MatrixTypeIterator Base; +// public: +// +// EIGEN_STRONG_INLINE InnerIterator(const CwiseUnaryViewImpl& unaryOp, typename CwiseUnaryViewImpl::Index outer) +// : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) +// {} +// +// EIGEN_STRONG_INLINE InnerIterator& operator++() +// { Base::operator++(); return *this; } +// +// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar value() const { return m_functor(Base::value()); } +// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar& valueRef() { return m_functor(Base::valueRef()); } +// +// protected: +// const ViewOp m_functor; +// }; +// +// template +// class CwiseUnaryViewImpl::ReverseInnerIterator +// : public CwiseUnaryViewImpl::MatrixTypeReverseIterator +// { +// typedef typename CwiseUnaryViewImpl::Scalar Scalar; +// typedef typename CwiseUnaryViewImpl::MatrixTypeReverseIterator Base; +// public: +// +// EIGEN_STRONG_INLINE ReverseInnerIterator(const CwiseUnaryViewImpl& unaryOp, typename CwiseUnaryViewImpl::Index outer) +// : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) +// {} +// +// EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() +// { Base::operator--(); return *this; } +// +// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar value() const { return m_functor(Base::value()); } +// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar& valueRef() { return m_functor(Base::valueRef()); } +// +// protected: +// const ViewOp m_functor; +// }; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS + template EIGEN_STRONG_INLINE Derived& SparseMatrixBase::operator*=(const Scalar& other) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 9ac18bcf6..a25bd83eb 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -1053,6 +1053,7 @@ void SparseMatrix::sumupDuplicates() m_data.resize(m_outerIndex[m_outerSize]); } +#ifndef EIGEN_TEST_EVALUATORS template template EIGEN_DONT_INLINE SparseMatrix& SparseMatrix::operator=(const SparseMatrixBase& other) @@ -1114,6 +1115,71 @@ EIGEN_DONT_INLINE SparseMatrix& SparseMatrix +template +EIGEN_DONT_INLINE SparseMatrix& SparseMatrix::operator=(const SparseMatrixBase& other) +{ + EIGEN_STATIC_ASSERT((internal::is_same::value), + YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + + const bool needToTranspose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); + if (needToTranspose) + { + // two passes algorithm: + // 1 - compute the number of coeffs per dest inner vector + // 2 - do the actual copy/eval + // Since each coeff of the rhs has to be evaluated twice, let's evaluate it if needed + typedef typename internal::nested_eval::type OtherCopy; + typedef typename internal::remove_all::type _OtherCopy; + typedef internal::evaluator<_OtherCopy> OtherCopyEval; + OtherCopy otherCopy(other.derived()); + OtherCopyEval otherCopyEval(otherCopy); + + SparseMatrix dest(other.rows(),other.cols()); + Eigen::Map > (dest.m_outerIndex,dest.outerSize()).setZero(); + + // pass 1 + // FIXME the above copy could be merged with that pass + for (Index j=0; j positions(dest.outerSize()); + for (Index j=0; jswap(dest); + return *this; + } + else + { + if(other.isRValue()) + initAssignment(other.derived()); + // there is no special optimization + return Base::operator=(other.derived()); + } +} +#endif template EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& SparseMatrix<_Scalar,_Options,_Index>::insertUncompressed(Index row, Index col) @@ -1254,6 +1320,33 @@ EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& Sparse return (m_data.value(p) = 0); } +#ifdef EIGEN_ENABLE_EVALUATORS +namespace internal { + +template +struct evaluator > + : evaluator_base > +{ + typedef SparseMatrix<_Scalar,_Options,_Index> SparseMatrixType; + typedef typename SparseMatrixType::InnerIterator InnerIterator; + typedef typename SparseMatrixType::ReverseInnerIterator ReverseInnerIterator; + + enum { + CoeffReadCost = NumTraits<_Scalar>::ReadCost, + Flags = SparseMatrixType::Flags + }; + + evaluator(const SparseMatrixType &mat) : m_matrix(mat) {} + + operator SparseMatrixType&() { return m_matrix.const_cast_derived(); } + operator const SparseMatrixType&() const { return m_matrix; } + + const SparseMatrixType &m_matrix; +}; + +} +#endif + } // end namespace Eigen #endif // EIGEN_SPARSEMATRIX_H diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index bbcf7fb1c..a87407fe9 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -39,11 +39,7 @@ template class SparseMatrixBase : public EigenBase typedef EigenBase Base; template - Derived& operator=(const EigenBase &other) - { - other.derived().evalTo(derived()); - return derived(); - } + Derived& operator=(const EigenBase &other); enum { @@ -175,87 +171,20 @@ template class SparseMatrixBase : public EigenBase template - Derived& operator=(const ReturnByValue& other) - { - other.evalTo(derived()); - return derived(); - } - + Derived& operator=(const ReturnByValue& other); template - inline Derived& operator=(const SparseMatrixBase& other) - { - return assign(other.derived()); - } + inline Derived& operator=(const SparseMatrixBase& other); - inline Derived& operator=(const Derived& other) - { -// if (other.isRValue()) -// derived().swap(other.const_cast_derived()); -// else - return assign(other.derived()); - } + inline Derived& operator=(const Derived& other); protected: template - inline Derived& assign(const OtherDerived& other) - { - const bool transpose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); - const Index outerSize = (int(OtherDerived::Flags) & RowMajorBit) ? other.rows() : other.cols(); - if ((!transpose) && other.isRValue()) - { - // eval without temporary - derived().resize(other.rows(), other.cols()); - derived().setZero(); - derived().reserve((std::max)(this->rows(),this->cols())*2); - for (Index j=0; j - inline void assignGeneric(const OtherDerived& other) - { - //const bool transpose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); - eigen_assert(( ((internal::traits::SupportedAccessPatterns&OuterRandomAccessPattern)==OuterRandomAccessPattern) || - (!((Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit)))) && - "the transpose operation is supposed to be handled in SparseMatrix::operator="); - - enum { Flip = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit) }; - - const Index outerSize = other.outerSize(); - //typedef typename internal::conditional, Derived>::type TempType; - // thanks to shallow copies, we always eval to a tempary - Derived temp(other.rows(), other.cols()); - - temp.reserve((std::max)(this->rows(),this->cols())*2); - for (Index j=0; j struct plain_matrix_type typedef SparseMatrix<_Scalar, _Options, _Index> type; }; +template +struct generic_xpr_base +{ + typedef SparseMatrixBase type; +}; + } // end namespace internal /** \ingroup SparseCore_Module From 7fa87a8b12fbc6a3e8af91fc773ba054e59a5ae5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 20 Jun 2014 16:17:57 +0200 Subject: [PATCH 0540/1103] Backport changes from old to new expression engines --- Eigen/src/Core/CoreEvaluators.h | 9 ++--- Eigen/src/Core/products/GeneralMatrixMatrix.h | 34 ++++++++++++++++++- .../products/GeneralMatrixMatrixTriangular.h | 2 ++ Eigen/src/Core/util/XprHelper.h | 2 +- 4 files changed, 41 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 95ea72680..5c4da9978 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -658,7 +658,7 @@ struct evaluator > && ( bool(IsDynamicSize) || HasNoOuterStride || ( OuterStrideAtCompileTime!=Dynamic - && ((static_cast(sizeof(Scalar))*OuterStrideAtCompileTime)%16)==0 ) ), + && ((static_cast(sizeof(Scalar))*OuterStrideAtCompileTime)%EIGEN_ALIGN_BYTES)==0 ) ), Flags0 = evaluator::Flags, Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit), Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime)) @@ -723,8 +723,9 @@ struct evaluator > MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % packet_traits::size) == 0) && (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0, - MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % 16) == 0)) ? AlignedBit : 0, - FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1) ? LinearAccessBit : 0, + + MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) ? AlignedBit : 0, + FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (traits::Flags&LinearAccessBit))) ? LinearAccessBit : 0, FlagsRowMajorBit = XprType::Flags&RowMajorBit, Flags0 = evaluator::Flags & ( (HereditaryBits & ~RowMajorBit) | DirectAccessBit | @@ -825,7 +826,7 @@ struct block_evaluator(block) { // FIXME this should be an internal assertion - eigen_assert(EIGEN_IMPLIES(evaluator::Flags&AlignedBit, (size_t(block.data()) % 16) == 0) && "data is not aligned"); + eigen_assert(EIGEN_IMPLIES(evaluator::Flags&AlignedBit, (size_t(block.data()) % EIGEN_ALIGN_BYTES) == 0) && "data is not aligned"); } }; diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 5fd9a98ac..66a4fe536 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -471,6 +471,38 @@ struct generic_product_impl MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime) }; + typedef generic_product_impl lazyproduct; + + template + static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0) + lazyproduct::evalTo(dst, lhs, rhs); + else + { + dst.setZero(); + scaleAndAddTo(dst, lhs, rhs, Scalar(1)); + } + } + + template + static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0) + lazyproduct::addTo(dst, lhs, rhs); + else + scaleAndAddTo(dst,lhs, rhs, Scalar(1)); + } + + template + static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs) + { + if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0) + lazyproduct::subTo(dst, lhs, rhs); + else + scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); + } + template static void scaleAndAddTo(Dest& dst, const Lhs& a_lhs, const Rhs& a_rhs, const Scalar& alpha) { @@ -494,7 +526,7 @@ struct generic_product_impl (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>, ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor; - BlockingType blocking(dst.rows(), dst.cols(), lhs.cols()); + BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), true); internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)> (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), Dest::Flags&RowMajorBit); diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 6070664d5..b28f07bdc 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -266,6 +266,8 @@ template template TriangularView& TriangularView::_assignProduct(const ProductType& prod, const Scalar& alpha) { + eigen_assert(m_matrix.rows() == prod.rows() && m_matrix.cols() == prod.cols()); + general_product_to_triangular_selector::InnerSize==1>::run(m_matrix.const_cast_derived(), prod, alpha); return *this; diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 14adb6795..5c4b9dbcd 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -187,7 +187,7 @@ class compute_matrix_evaluator_flags ((Options&DontAlign)==0) && ( #if EIGEN_ALIGN_STATICALLY - ((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % 16) == 0)) + ((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) #else 0 #endif From ec0a8b2e6dab9af4b4fc6c91d2ab7b01e47ea5ce Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 20 Jun 2014 16:30:34 +0200 Subject: [PATCH 0541/1103] rm conflict --- test/vectorization_logic.cpp | 4 ---- 1 file changed, 4 deletions(-) diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp index 1fb079c48..5c6e8042d 100644 --- a/test/vectorization_logic.cpp +++ b/test/vectorization_logic.cpp @@ -232,12 +232,8 @@ template::Vectori Matrix >(DefaultTraversal,CompleteUnrolling))); -<<<<<<< local - VERIFY((test_assign(Matrix11(), Matrix11().lazyProduct(Matrix11()), InnerVectorizedTraversal, CompleteUnrolling))); -======= VERIFY((test_assign(Matrix11(), Matrix()*Matrix(), PacketSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD?DefaultTraversal:InnerVectorizedTraversal, CompleteUnrolling))); ->>>>>>> other #endif VERIFY(test_assign(MatrixXX(10,10),MatrixXX(20,20).block(10,10,2,3), From 5dbbe6b40067381a10ea1ffbb4937f68b7f27603 Mon Sep 17 00:00:00 2001 From: Jeff Date: Fri, 20 Jun 2014 22:12:45 -0600 Subject: [PATCH 0542/1103] Added Spline interpolation with derivatives. --- unsupported/Eigen/src/Splines/Spline.h | 64 +++- unsupported/Eigen/src/Splines/SplineFitting.h | 273 ++++++++++++++++++ unsupported/Eigen/src/Splines/SplineFwd.h | 5 + unsupported/test/splines.cpp | 34 +++ 4 files changed, 363 insertions(+), 13 deletions(-) diff --git a/unsupported/Eigen/src/Splines/Spline.h b/unsupported/Eigen/src/Splines/Spline.h index 1b47992d6..c46f728bc 100644 --- a/unsupported/Eigen/src/Splines/Spline.h +++ b/unsupported/Eigen/src/Splines/Spline.h @@ -44,9 +44,15 @@ namespace Eigen /** \brief The data type used to store knot vectors. */ typedef typename SplineTraits::KnotVectorType KnotVectorType; + + /** \brief The data type used to store parameter vectors. */ + typedef typename SplineTraits::ParameterVectorType ParameterVectorType; /** \brief The data type used to store non-zero basis functions. */ typedef typename SplineTraits::BasisVectorType BasisVectorType; + + /** \brief The data type used to store the values of the basis function derivatives. */ + typedef typename SplineTraits::BasisDerivativeType BasisDerivativeType; /** \brief The data type representing the spline's control points. */ typedef typename SplineTraits::ControlPointVectorType ControlPointVectorType; @@ -203,10 +209,25 @@ namespace Eigen **/ static BasisVectorType BasisFunctions(Scalar u, DenseIndex degree, const KnotVectorType& knots); + /** + * \copydoc Spline::basisFunctionDerivatives + * \param degree The degree of the underlying spline + * \param knots The underlying spline's knot vector. + **/ + static BasisDerivativeType BasisFunctionDerivatives( + const Scalar u, const DenseIndex order, const DenseIndex degree, const KnotVectorType& knots); private: KnotVectorType m_knots; /*!< Knot vector. */ ControlPointVectorType m_ctrls; /*!< Control points. */ + + template + static void BasisFunctionDerivativesImpl( + const typename Spline<_Scalar, _Dim, _Degree>::Scalar u, + const DenseIndex order, + const DenseIndex p, + const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& U, + DerivativeType& N_); }; template @@ -345,20 +366,24 @@ namespace Eigen } /* --------------------------------------------------------------------------------------------- */ - - template - void basisFunctionDerivativesImpl(const SplineType& spline, typename SplineType::Scalar u, DenseIndex order, DerivativeType& N_) + + + template + template + void Spline<_Scalar, _Dim, _Degree>::BasisFunctionDerivativesImpl( + const typename Spline<_Scalar, _Dim, _Degree>::Scalar u, + const DenseIndex order, + const DenseIndex p, + const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& U, + DerivativeType& N_) { + typedef Spline<_Scalar, _Dim, _Degree> SplineType; enum { Order = SplineTraits::OrderAtCompileTime }; typedef typename SplineTraits::Scalar Scalar; typedef typename SplineTraits::BasisVectorType BasisVectorType; - typedef typename SplineTraits::KnotVectorType KnotVectorType; - - const KnotVectorType& U = spline.knots(); - - const DenseIndex p = spline.degree(); - const DenseIndex span = spline.span(u); + + const DenseIndex span = SplineType::Span(u, p, U); const DenseIndex n = (std::min)(p, order); @@ -455,8 +480,8 @@ namespace Eigen typename SplineTraits< Spline<_Scalar, _Dim, _Degree> >::BasisDerivativeType Spline<_Scalar, _Dim, _Degree>::basisFunctionDerivatives(Scalar u, DenseIndex order) const { - typename SplineTraits< Spline >::BasisDerivativeType der; - basisFunctionDerivativesImpl(*this, u, order, der); + typename SplineTraits >::BasisDerivativeType der; + BasisFunctionDerivativesImpl(u, order, degree(), knots(), der); return der; } @@ -465,8 +490,21 @@ namespace Eigen typename SplineTraits< Spline<_Scalar, _Dim, _Degree>, DerivativeOrder >::BasisDerivativeType Spline<_Scalar, _Dim, _Degree>::basisFunctionDerivatives(Scalar u, DenseIndex order) const { - typename SplineTraits< Spline, DerivativeOrder >::BasisDerivativeType der; - basisFunctionDerivativesImpl(*this, u, order, der); + typename SplineTraits< Spline<_Scalar, _Dim, _Degree>, DerivativeOrder >::BasisDerivativeType der; + BasisFunctionDerivativesImpl(u, order, degree(), knots(), der); + return der; + } + + template + typename SplineTraits >::BasisDerivativeType + Spline<_Scalar, _Dim, _Degree>::BasisFunctionDerivatives( + const typename Spline<_Scalar, _Dim, _Degree>::Scalar u, + const DenseIndex order, + const DenseIndex degree, + const typename Spline<_Scalar, _Dim, _Degree>::KnotVectorType& knots) + { + typename SplineTraits::BasisDerivativeType der; + BasisFunctionDerivativesImpl(u, order, degree, knots, der); return der; } } diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h index 0265d532c..7c2484e0d 100644 --- a/unsupported/Eigen/src/Splines/SplineFitting.h +++ b/unsupported/Eigen/src/Splines/SplineFitting.h @@ -10,7 +10,10 @@ #ifndef EIGEN_SPLINE_FITTING_H #define EIGEN_SPLINE_FITTING_H +#include +#include #include +#include #include "SplineFwd.h" @@ -49,6 +52,129 @@ namespace Eigen knots.segment(knots.size()-degree-1,degree+1) = KnotVectorType::Ones(degree+1); } + /** + * \brief Computes knot averages when derivative constraints are present. + * Note that this is a technical interpretation of the referenced article + * since the algorithm contained therein is incorrect as written. + * \ingroup Splines_Module + * + * \param[in] parameters The parameters at which the interpolation B-Spline + * will intersect the given interpolation points. The parameters + * are assumed to be a non-decreasing sequence. + * \param[in] degree The degree of the interpolating B-Spline. This must be + * greater than zero. + * \param[in] derivativeIndices The indices corresponding to parameters at + * which there are derivative constraints. The indices are assumed + * to be a non-decreasing sequence. + * \param[out] knots The calculated knot vector. These will be returned as a + * non-decreasing sequence + * + * \sa Les A. Piegl, Khairan Rajab, Volha Smarodzinana. 2008. + * Curve interpolation with directional constraints for engineering design. + * Engineering with Computers + **/ + template + void KnotAveragingWithDerivatives(const ParameterVectorType& parameters, + const unsigned int degree, + const IndexArray& derivativeIndices, + KnotVectorType& knots) + { + typedef typename ParameterVectorType::Scalar Scalar; + + const unsigned int numParameters = parameters.size(); + const unsigned int numDerivatives = derivativeIndices.size(); + + if (numDerivatives < 1) + { + KnotAveraging(parameters, degree, knots); + return; + } + + uint startIndex; + uint endIndex; + + unsigned int numInternalDerivatives = numDerivatives; + + if (derivativeIndices[0] == 0) + { + startIndex = 0; + --numInternalDerivatives; + } + else + { + startIndex = 1; + } + if (derivativeIndices[numDerivatives - 1] == numParameters - 1) + { + endIndex = numParameters - degree; + --numInternalDerivatives; + } + else + { + endIndex = numParameters - degree - 1; + } + + // There are (endIndex - startIndex + 1) knots obtained from the averaging + // and 2 for the first and last parameters. + unsigned int numAverageKnots = endIndex - startIndex + 3; + KnotVectorType averageKnots(numAverageKnots); + averageKnots[0] = parameters[0]; + + int newKnotIndex = 0; + for (unsigned int i = startIndex; i <= endIndex; ++i) + averageKnots[++newKnotIndex] = parameters.segment(i, degree).mean(); + averageKnots[++newKnotIndex] = parameters[numParameters - 1]; + + newKnotIndex = -1; + + ParameterVectorType temporaryParameters(numParameters); + KnotVectorType derivativeKnots(numInternalDerivatives); + for (unsigned int i = 0; i < numAverageKnots - 1; ++i) + { + temporaryParameters[0] = averageKnots[i]; + ParameterVectorType parameterIndices(numParameters); + int temporaryParameterIndex = 1; + for (size_t j = 0; j < numParameters; ++j) + { + Scalar parameter = parameters[j]; + if (parameter >= averageKnots[i] && parameter < averageKnots[i + 1]) + { + parameterIndices[temporaryParameterIndex] = j; + temporaryParameters[temporaryParameterIndex++] = parameter; + } + } + temporaryParameters[temporaryParameterIndex] = averageKnots[i + 1]; + + for (int j = 0; j <= temporaryParameterIndex - 2; ++j) + { + for (DenseIndex k = 0; k < derivativeIndices.size(); ++k) + { + if (parameterIndices[j + 1] == derivativeIndices[k] + && parameterIndices[j + 1] != 0 + && parameterIndices[j + 1] != numParameters - 1) + { + derivativeKnots[++newKnotIndex] = temporaryParameters.segment(j, 3).mean(); + break; + } + } + } + } + + KnotVectorType temporaryKnots(averageKnots.size() + derivativeKnots.size()); + + std::merge(averageKnots.data(), averageKnots.data() + averageKnots.size(), + derivativeKnots.data(), derivativeKnots.data() + derivativeKnots.size(), + temporaryKnots.data()); + + // Number of control points (one for each point and derivative) plus spline order. + DenseIndex numKnots = numParameters + numDerivatives + degree + 1; + knots.resize(numKnots); + + knots.head(degree).fill(temporaryKnots[0]); + knots.tail(degree).fill(temporaryKnots.template tail<1>()[0]); + knots.segment(degree, temporaryKnots.size()) = temporaryKnots; + } + /** * \brief Computes chord length parameters which are required for spline interpolation. * \ingroup Splines_Module @@ -86,6 +212,7 @@ namespace Eigen struct SplineFitting { typedef typename SplineType::KnotVectorType KnotVectorType; + typedef typename SplineType::ParameterVectorType ParameterVectorType; /** * \brief Fits an interpolating Spline to the given data points. @@ -109,6 +236,52 @@ namespace Eigen **/ template static SplineType Interpolate(const PointArrayType& pts, DenseIndex degree, const KnotVectorType& knot_parameters); + + /** + * \brief Fits an interpolating spline to the given data points and + * derivatives. + * + * \param points The points for which an interpolating spline will be computed. + * \param derivatives The desired derivatives of the interpolating spline at interpolation + * points. + * \param derivativeIndices An array indicating which point each derivative belongs to. This + * must be the same size as @a derivatives. + * \param degree The degree of the interpolating spline. + * + * \returns A spline interpolating @a points with @a derivatives at those points. + * + * \sa Les A. Piegl, Khairan Rajab, Volha Smarodzinana. 2008. + * Curve interpolation with directional constraints for engineering design. + * Engineering with Computers + **/ + template + static SplineType InterpolateWithDerivatives(const PointArrayType& points, + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree); + + /** + * \brief Fits an interpolating spline to the given data points and derivatives. + * + * \param points The points for which an interpolating spline will be computed. + * \param derivatives The desired derivatives of the interpolating spline at interpolation points. + * \param derivativeIndices An array indicating which point each derivative belongs to. This + * must be the same size as @a derivatives. + * \param degree The degree of the interpolating spline. + * \param parameters The parameters corresponding to the interpolation points. + * + * \returns A spline interpolating @a points with @a derivatives at those points. + * + * \sa Les A. Piegl, Khairan Rajab, Volha Smarodzinana. 2008. + * Curve interpolation with directional constraints for engineering design. + * Engineering with Computers + */ + template + static SplineType InterpolateWithDerivatives(const PointArrayType& points, + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree, + const ParameterVectorType& parameters); }; template @@ -151,6 +324,106 @@ namespace Eigen ChordLengths(pts, chord_lengths); return Interpolate(pts, degree, chord_lengths); } + + template + template + SplineType + SplineFitting::InterpolateWithDerivatives(const PointArrayType& points, + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree, + const ParameterVectorType& parameters) + { + typedef typename SplineType::KnotVectorType::Scalar Scalar; + typedef typename SplineType::ControlPointVectorType ControlPointVectorType; + + typedef Matrix MatrixType; + + const DenseIndex n = points.cols() + derivatives.cols(); + + KnotVectorType knots; + + KnotAveragingWithDerivatives(parameters, degree, derivativeIndices, knots); + + // fill matrix + MatrixType A = MatrixType::Zero(n, n); + + // Use these dimensions for quicker populating, then transpose for solving. + MatrixType b(points.rows(), n); + + DenseIndex startRow; + DenseIndex derivativeStart; + + // End derivatives. + if (derivativeIndices[0] == 0) + { + A.template block<1, 2>(1, 0) << -1, 1; + + Scalar y = (knots(degree + 1) - knots(0)) / degree; + b.col(1) = y*derivatives.col(0); + + startRow = 2; + derivativeStart = 1; + } + else + { + startRow = 1; + derivativeStart = 0; + } + if (derivativeIndices[derivatives.cols() - 1] == points.cols() - 1) + { + A.template block<1, 2>(n - 2, n - 2) << -1, 1; + + Scalar y = (knots(knots.size() - 1) - knots(knots.size() - (degree + 2))) / degree; + b.col(b.cols() - 2) = y*derivatives.col(derivatives.cols() - 1); + } + + DenseIndex row = startRow; + DenseIndex derivativeIndex = derivativeStart; + for (DenseIndex i = 1; i < parameters.size() - 1; ++i) + { + const DenseIndex span = SplineType::Span(parameters[i], degree, knots); + + if (derivativeIndices[derivativeIndex] == i) + { + A.block(row, span - degree, 2, degree + 1) + = SplineType::BasisFunctionDerivatives(parameters[i], 1, degree, knots); + + b.col(row++) = points.col(i); + b.col(row++) = derivatives.col(derivativeIndex++); + } + else + { + A.row(row++).segment(span - degree, degree + 1) + = SplineType::BasisFunctions(parameters[i], degree, knots); + } + } + b.col(0) = points.col(0); + b.col(b.cols() - 1) = points.col(points.cols() - 1); + A(0,0) = 1; + A(n - 1, n - 1) = 1; + + // Solve + HouseholderQR qr(A); + ControlPointVectorType controlPoints = qr.solve(MatrixType(b.transpose())).transpose(); + + SplineType spline(knots, controlPoints); + + return spline; + } + + template + template + SplineType + SplineFitting::InterpolateWithDerivatives(const PointArrayType& points, + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree) + { + ParameterVectorType parameters; + ChordLengths(points, parameters); + return InterpolateWithDerivatives(points, derivatives, derivativeIndices, degree, parameters); + } } #endif // EIGEN_SPLINE_FITTING_H diff --git a/unsupported/Eigen/src/Splines/SplineFwd.h b/unsupported/Eigen/src/Splines/SplineFwd.h index 9ea23a9a1..ec3dc0ff5 100644 --- a/unsupported/Eigen/src/Splines/SplineFwd.h +++ b/unsupported/Eigen/src/Splines/SplineFwd.h @@ -48,6 +48,9 @@ namespace Eigen /** \brief The data type used to store knot vectors. */ typedef Array KnotVectorType; + + /** \brief The data type used to store parameter vectors. */ + typedef Array ParameterVectorType; /** \brief The data type representing the spline's control points. */ typedef Array ControlPointVectorType; @@ -85,6 +88,8 @@ namespace Eigen /** \brief 3D double B-spline with dynamic degree. */ typedef Spline Spline3d; + + typedef Array IndexArray; } #endif // EIGEN_SPLINES_FWD_H diff --git a/unsupported/test/splines.cpp b/unsupported/test/splines.cpp index a7eb3e0c4..1f3856143 100644 --- a/unsupported/test/splines.cpp +++ b/unsupported/test/splines.cpp @@ -234,6 +234,39 @@ void check_global_interpolation2d() } } +void check_global_interpolation_with_derivatives2d() +{ + typedef Spline2d::PointType PointType; + typedef Spline2d::KnotVectorType KnotVectorType; + + const unsigned int numPoints = 100; + const unsigned int dimension = 2; + const unsigned int degree = 3; + + ArrayXXd points = ArrayXXd::Random(dimension, numPoints); + + KnotVectorType knots; + Eigen::ChordLengths(points, knots); + + ArrayXXd derivatives = ArrayXXd::Random(dimension, numPoints); + Eigen::IndexArray derivativeIndices(numPoints); + + for (Eigen::DenseIndex i = 0; i < numPoints; ++i) + derivativeIndices(i) = i; + + const Spline2d spline = SplineFitting::InterpolateWithDerivatives( + points, derivatives, derivativeIndices, degree); + + for (Eigen::DenseIndex i = 0; i < points.cols(); ++i) + { + PointType point = spline(knots(i)); + PointType referencePoint = points.col(i); + VERIFY((point - referencePoint).matrix().norm() < 1e-12); + PointType derivative = spline.derivatives(knots(i), 1).col(1); + PointType referenceDerivative = derivatives.col(i); + VERIFY((derivative - referenceDerivative).matrix().norm() < 1e-10); + } +} void test_splines() { @@ -241,4 +274,5 @@ void test_splines() CALL_SUBTEST( eval_spline3d_onbrks() ); CALL_SUBTEST( eval_closed_spline2d() ); CALL_SUBTEST( check_global_interpolation2d() ); + CALL_SUBTEST( check_global_interpolation_with_derivatives2d() ); } From 3849cc65ee2192ddd9d63cdc2e65cb74128bcbb3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 23 Jun 2014 10:40:03 +0200 Subject: [PATCH 0543/1103] Implement binaryop and transpose evaluators for sparse matrices --- Eigen/SparseCore | 4 +- Eigen/src/Core/AssignEvaluator.h | 10 +- Eigen/src/Core/CoreEvaluators.h | 6 +- Eigen/src/Core/CwiseBinaryOp.h | 2 +- Eigen/src/Core/CwiseUnaryOp.h | 6 +- Eigen/src/Core/Transpose.h | 13 +- Eigen/src/SparseCore/SparseCwiseBinaryOp.h | 312 ++++++++++++++++++++- Eigen/src/SparseCore/SparseCwiseUnaryOp.h | 2 +- Eigen/src/SparseCore/SparseMatrix.h | 2 +- Eigen/src/SparseCore/SparseMatrixBase.h | 4 +- Eigen/src/SparseCore/SparseTranspose.h | 54 +++- 11 files changed, 396 insertions(+), 19 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index c0f0a4121..799cb1ba1 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -42,10 +42,10 @@ struct Sparse {}; #include "src/SparseCore/MappedSparseMatrix.h" #include "src/SparseCore/SparseVector.h" #include "src/SparseCore/SparseCwiseUnaryOp.h" +#include "src/SparseCore/SparseCwiseBinaryOp.h" +#include "src/SparseCore/SparseTranspose.h" #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparseBlock.h" -#include "src/SparseCore/SparseTranspose.h" -#include "src/SparseCore/SparseCwiseBinaryOp.h" #include "src/SparseCore/SparseDot.h" #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseRedux.h" diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index a083e340e..77b4dd9bc 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2011 Benoit Jacob -// Copyright (C) 2011-2013 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // Copyright (C) 2011-2012 Jitse Niesen // // This Source Code Form is subject to the terms of the Mozilla @@ -738,8 +738,10 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) && int(Dst::SizeAtCompileTime) != 1 }; - dst.resize(NeedToTranspose ? src.cols() : src.rows(), - NeedToTranspose ? src.rows() : src.cols()); + typename Dst::Index dstRows = NeedToTranspose ? src.cols() : src.rows(); + typename Dst::Index dstCols = NeedToTranspose ? src.rows() : src.cols(); + if((dst.rows()!=dstRows) || (dst.cols()!=dstCols)) + dst.resize(dstRows, dstCols); typedef typename internal::conditional, Dst>::type ActualDstTypeCleaned; typedef typename internal::conditional, Dst&>::type ActualDstType; @@ -749,7 +751,7 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) EIGEN_STATIC_ASSERT_LVALUE(Dst) EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(ActualDstTypeCleaned,Src) EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar); - + Assignment::run(actualDst, src, func); } template diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 5c4da9978..b56c3c635 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2011 Benoit Jacob -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // Copyright (C) 2011-2012 Jitse Niesen // // This Source Code Form is subject to the terms of the Mozilla @@ -253,7 +253,7 @@ struct evaluator > // -------------------- Transpose -------------------- template -struct unary_evaluator > +struct unary_evaluator, IndexBased> : evaluator_base > { typedef Transpose XprType; @@ -440,7 +440,7 @@ struct evaluator > }; template -struct binary_evaluator > +struct binary_evaluator, IndexBased, IndexBased> : evaluator_base > { typedef CwiseBinaryOp XprType; diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index 5e4fb147b..355f3bdc8 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // Copyright (C) 2006-2008 Benoit Jacob // // This Source Code Form is subject to the terms of the Mozilla diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index 098034a1e..c2bc47c93 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2010 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // Copyright (C) 2006-2008 Benoit Jacob // // This Source Code Form is subject to the terms of the Mozilla @@ -135,7 +135,7 @@ class CwiseUnaryOpImpl return derived().functor().packetOp(derived().nestedExpression().template packet(index)); } }; -#else +#else // EIGEN_TEST_EVALUATORS // Generic API dispatcher template class CwiseUnaryOpImpl @@ -144,7 +144,7 @@ class CwiseUnaryOpImpl public: typedef typename internal::generic_xpr_base >::type Base; }; -#endif +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 4b605dfd6..f5148221d 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2006-2008 Benoit Jacob -// Copyright (C) 2009-2010 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -108,6 +108,17 @@ struct TransposeImpl_base } // end namespace internal +#ifdef EIGEN_TEST_EVALUATORS +// Generic API dispatcher +template +class TransposeImpl + : public internal::generic_xpr_base >::type +{ +public: + typedef typename internal::generic_xpr_base >::type Base; +}; +#endif + template class TransposeImpl : public internal::TransposeImpl_base::type { diff --git a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h index ec86ca933..f43976641 100644 --- a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Gael Guennebaud +// Copyright (C) 2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -44,6 +44,8 @@ class sparse_cwise_binary_op_inner_iterator_selector; } // end namespace internal +#ifndef EIGEN_TEST_EVALUATORS + template class CwiseBinaryOpImpl : public SparseMatrixBase > @@ -291,6 +293,314 @@ class sparse_cwise_binary_op_inner_iterator_selector, Lhs, } // end namespace internal +#else // EIGEN_TEST_EVALUATORS + +namespace internal { + + +// Generic "sparse OP sparse" +template +struct binary_evaluator, IteratorBased, IteratorBased> + : evaluator_base > +{ +protected: + typedef typename evaluator::InnerIterator LhsIterator; + typedef typename evaluator::InnerIterator RhsIterator; +public: + typedef CwiseBinaryOp XprType; + + class ReverseInnerIterator; + class InnerIterator + { + typedef typename traits::Scalar Scalar; + typedef typename XprType::Index Index; + + public: + + EIGEN_STRONG_INLINE InnerIterator(const binary_evaluator& aEval, Index outer) + : m_lhsIter(aEval.m_lhsImpl,outer), m_rhsIter(aEval.m_rhsImpl,outer), m_functor(aEval.m_functor) + { + this->operator++(); + } + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { + if (m_lhsIter && m_rhsIter && (m_lhsIter.index() == m_rhsIter.index())) + { + m_id = m_lhsIter.index(); + m_value = m_functor(m_lhsIter.value(), m_rhsIter.value()); + ++m_lhsIter; + ++m_rhsIter; + } + else if (m_lhsIter && (!m_rhsIter || (m_lhsIter.index() < m_rhsIter.index()))) + { + m_id = m_lhsIter.index(); + m_value = m_functor(m_lhsIter.value(), Scalar(0)); + ++m_lhsIter; + } + else if (m_rhsIter && (!m_lhsIter || (m_lhsIter.index() > m_rhsIter.index()))) + { + m_id = m_rhsIter.index(); + m_value = m_functor(Scalar(0), m_rhsIter.value()); + ++m_rhsIter; + } + else + { + m_value = 0; // this is to avoid a compilation warning + m_id = -1; + } + return *this; + } + + EIGEN_STRONG_INLINE Scalar value() const { return m_value; } + + EIGEN_STRONG_INLINE Index index() const { return m_id; } + EIGEN_STRONG_INLINE Index row() const { return Lhs::IsRowMajor ? m_lhsIter.row() : index(); } + EIGEN_STRONG_INLINE Index col() const { return Lhs::IsRowMajor ? index() : m_lhsIter.col(); } + + EIGEN_STRONG_INLINE operator bool() const { return m_id>=0; } + + protected: + LhsIterator m_lhsIter; + RhsIterator m_rhsIter; + const BinaryOp& m_functor; + Scalar m_value; + Index m_id; + }; + + + enum { + CoeffReadCost = evaluator::CoeffReadCost + evaluator::CoeffReadCost + functor_traits::Cost, + Flags = XprType::Flags + }; + + binary_evaluator(const XprType& xpr) + : m_functor(xpr.functor()), + m_lhsImpl(xpr.lhs()), + m_rhsImpl(xpr.rhs()) + { } + +protected: + const BinaryOp m_functor; + typename evaluator::nestedType m_lhsImpl; + typename evaluator::nestedType m_rhsImpl; +}; + +// "sparse .* sparse" +template +struct binary_evaluator, Lhs, Rhs>, IteratorBased, IteratorBased> + : evaluator_base, Lhs, Rhs> > +{ +protected: + typedef scalar_product_op BinaryOp; + typedef typename evaluator::InnerIterator LhsIterator; + typedef typename evaluator::InnerIterator RhsIterator; +public: + typedef CwiseBinaryOp XprType; + + class ReverseInnerIterator; + class InnerIterator + { + typedef typename traits::Scalar Scalar; + typedef typename XprType::Index Index; + + public: + + EIGEN_STRONG_INLINE InnerIterator(const binary_evaluator& aEval, Index outer) + : m_lhsIter(aEval.m_lhsImpl,outer), m_rhsIter(aEval.m_rhsImpl,outer), m_functor(aEval.m_functor) + { + while (m_lhsIter && m_rhsIter && (m_lhsIter.index() != m_rhsIter.index())) + { + if (m_lhsIter.index() < m_rhsIter.index()) + ++m_lhsIter; + else + ++m_rhsIter; + } + } + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { + ++m_lhsIter; + ++m_rhsIter; + while (m_lhsIter && m_rhsIter && (m_lhsIter.index() != m_rhsIter.index())) + { + if (m_lhsIter.index() < m_rhsIter.index()) + ++m_lhsIter; + else + ++m_rhsIter; + } + return *this; + } + + EIGEN_STRONG_INLINE Scalar value() const { return m_functor(m_lhsIter.value(), m_rhsIter.value()); } + + EIGEN_STRONG_INLINE Index index() const { return m_lhsIter.index(); } + EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); } + EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); } + + EIGEN_STRONG_INLINE operator bool() const { return (m_lhsIter && m_rhsIter); } + + protected: + LhsIterator m_lhsIter; + RhsIterator m_rhsIter; + const BinaryOp& m_functor; + }; + + + enum { + CoeffReadCost = evaluator::CoeffReadCost + evaluator::CoeffReadCost + functor_traits::Cost, + Flags = XprType::Flags + }; + + binary_evaluator(const XprType& xpr) + : m_functor(xpr.functor()), + m_lhsImpl(xpr.lhs()), + m_rhsImpl(xpr.rhs()) + { } + +protected: + const BinaryOp m_functor; + typename evaluator::nestedType m_lhsImpl; + typename evaluator::nestedType m_rhsImpl; +}; + +// "dense .* sparse" +template +struct binary_evaluator, Lhs, Rhs>, IndexBased, IteratorBased> + : evaluator_base, Lhs, Rhs> > +{ +protected: + typedef scalar_product_op BinaryOp; + typedef typename evaluator::type LhsEvaluator; + typedef typename evaluator::InnerIterator RhsIterator; +public: + typedef CwiseBinaryOp XprType; + + class ReverseInnerIterator; + class InnerIterator + { + typedef typename traits::Scalar Scalar; + typedef typename XprType::Index Index; + enum { IsRowMajor = (int(Rhs::Flags)&RowMajorBit)==RowMajorBit }; + + public: + + EIGEN_STRONG_INLINE InnerIterator(const binary_evaluator& aEval, Index outer) + : m_lhsEval(aEval.m_lhsImpl), m_rhsIter(aEval.m_rhsImpl,outer), m_functor(aEval.m_functor), m_outer(outer) + {} + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { + ++m_rhsIter; + return *this; + } + + EIGEN_STRONG_INLINE Scalar value() const + { return m_functor(m_lhsEval.coeff(IsRowMajor?m_outer:m_rhsIter.index(),IsRowMajor?m_rhsIter.index():m_outer), m_rhsIter.value()); } + + EIGEN_STRONG_INLINE Index index() const { return m_rhsIter.index(); } + EIGEN_STRONG_INLINE Index row() const { return m_rhsIter.row(); } + EIGEN_STRONG_INLINE Index col() const { return m_rhsIter.col(); } + + EIGEN_STRONG_INLINE operator bool() const { return m_rhsIter; } + + protected: + const LhsEvaluator &m_lhsEval; + RhsIterator m_rhsIter; + const BinaryOp& m_functor; + const Index m_outer; + }; + + + enum { + CoeffReadCost = evaluator::CoeffReadCost + evaluator::CoeffReadCost + functor_traits::Cost, + Flags = XprType::Flags + }; + + binary_evaluator(const XprType& xpr) + : m_functor(xpr.functor()), + m_lhsImpl(xpr.lhs()), + m_rhsImpl(xpr.rhs()) + { } + +protected: + const BinaryOp m_functor; + typename evaluator::nestedType m_lhsImpl; + typename evaluator::nestedType m_rhsImpl; +}; + +// "sparse .* dense" +template +struct binary_evaluator, Lhs, Rhs>, IteratorBased, IndexBased> + : evaluator_base, Lhs, Rhs> > +{ +protected: + typedef scalar_product_op BinaryOp; + typedef typename evaluator::InnerIterator LhsIterator; + typedef typename evaluator::type RhsEvaluator; +public: + typedef CwiseBinaryOp XprType; + + class ReverseInnerIterator; + class InnerIterator + { + typedef typename traits::Scalar Scalar; + typedef typename XprType::Index Index; + enum { IsRowMajor = (int(Lhs::Flags)&RowMajorBit)==RowMajorBit }; + + public: + + EIGEN_STRONG_INLINE InnerIterator(const binary_evaluator& aEval, Index outer) + : m_lhsIter(aEval.m_lhsImpl,outer), m_rhsEval(aEval.m_rhsImpl), m_functor(aEval.m_functor), m_outer(outer) + {} + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { + ++m_lhsIter; + return *this; + } + + EIGEN_STRONG_INLINE Scalar value() const + { return m_functor(m_lhsIter.value(), + m_rhsEval.coeff(IsRowMajor?m_outer:m_lhsIter.index(),IsRowMajor?m_lhsIter.index():m_outer)); } + + EIGEN_STRONG_INLINE Index index() const { return m_lhsIter.index(); } + EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); } + EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); } + + EIGEN_STRONG_INLINE operator bool() const { return m_lhsIter; } + + protected: + LhsIterator m_lhsIter; + const RhsEvaluator &m_rhsEval; + const BinaryOp& m_functor; + const Index m_outer; + }; + + + enum { + CoeffReadCost = evaluator::CoeffReadCost + evaluator::CoeffReadCost + functor_traits::Cost, + Flags = XprType::Flags + }; + + binary_evaluator(const XprType& xpr) + : m_functor(xpr.functor()), + m_lhsImpl(xpr.lhs()), + m_rhsImpl(xpr.rhs()) + { } + +protected: + const BinaryOp m_functor; + typename evaluator::nestedType m_lhsImpl; + typename evaluator::nestedType m_rhsImpl; +}; + +} + +#endif + + + /*************************************************************************** * Implementation of SparseMatrixBase and SparseCwise functions/operators ***************************************************************************/ diff --git a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h index c5042504b..ead3af59d 100644 --- a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2010 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index aa8bfc0f5..8aa0ed7e6 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2010 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index a87407fe9..e4960a445 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2011 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -79,10 +79,12 @@ template class SparseMatrixBase : public EigenBase * constructed from this one. See the \ref flags "list of flags". */ +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = internal::traits::CoeffReadCost, /**< This is a rough measure of how expensive it is to read one coefficient from * this expression. */ +#endif IsRowMajor = Flags&RowMajorBit ? 1 : 0, diff --git a/Eigen/src/SparseCore/SparseTranspose.h b/Eigen/src/SparseCore/SparseTranspose.h index 7c300ee8d..b20843dd3 100644 --- a/Eigen/src/SparseCore/SparseTranspose.h +++ b/Eigen/src/SparseCore/SparseTranspose.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -12,6 +12,7 @@ namespace Eigen { +#ifndef EIGEN_TEST_EVALUATORS template class TransposeImpl : public SparseMatrixBase > { @@ -58,6 +59,57 @@ template class TransposeImpl::ReverseInn Index col() const { return Base::row(); } }; +#else // EIGEN_TEST_EVALUATORS + +namespace internal { + +template +struct unary_evaluator, IteratorBased> + : public evaluator_base > +{ + typedef typename evaluator::InnerIterator EvalIterator; + typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; + public: + typedef Transpose XprType; + typedef typename XprType::Index Index; + + class InnerIterator : public EvalIterator + { + public: + EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& unaryOp, typename XprType::Index outer) + : EvalIterator(unaryOp.m_argImpl,outer) + {} + + Index row() const { return EvalIterator::col(); } + Index col() const { return EvalIterator::row(); } + }; + + class ReverseInnerIterator : public EvalReverseIterator + { + public: + EIGEN_STRONG_INLINE ReverseInnerIterator(const unary_evaluator& unaryOp, typename XprType::Index outer) + : EvalReverseIterator(unaryOp.m_argImpl,outer) + {} + + Index row() const { return EvalReverseIterator::col(); } + Index col() const { return EvalReverseIterator::row(); } + }; + + enum { + CoeffReadCost = evaluator::CoeffReadCost, + Flags = XprType::Flags + }; + + unary_evaluator(const XprType& op) :m_argImpl(op.nestedExpression()) {} + + protected: + typename evaluator::nestedType m_argImpl; +}; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS + } // end namespace Eigen #endif // EIGEN_SPARSETRANSPOSE_H From 0ddde223e93978f6307ded8b03404bf5a81b8d08 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 23 Jun 2014 11:00:52 +0200 Subject: [PATCH 0544/1103] Fixed typos --- doc/SparseLinearSystems.dox | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/doc/SparseLinearSystems.dox b/doc/SparseLinearSystems.dox index f0456ff52..147b55376 100644 --- a/doc/SparseLinearSystems.dox +++ b/doc/SparseLinearSystems.dox @@ -102,7 +102,7 @@ x2 = solver.solve(b2); The compute() method is equivalent to calling both analyzePattern() and factorize(). Finally, each solver provides some specific features, such as determinant, access to the factors, controls of the iterations, and so on. -More details are availble in the documentations of the respective classes. +More details are available in the documentations of the respective classes. \section TheSparseCompute The Compute Step In the compute() function, the matrix is generally factorized: LLT for self-adjoint matrices, LDLT for general hermitian matrices, LU for non hermitian matrices and QR for rectangular matrices. These are the results of using direct solvers. For this class of solvers precisely, the compute step is further subdivided into analyzePattern() and factorize(). @@ -140,7 +140,7 @@ x2 = solver.solve(b2); For direct methods, the solution are computed at the machine precision. Sometimes, the solution need not be too accurate. In this case, the iterative methods are more suitable and the desired accuracy can be set before the solve step using \b setTolerance(). For all the available functions, please, refer to the documentation of the \link IterativeLinearSolvers_Module Iterative solvers module \endlink. \section BenchmarkRoutine -Most of the time, all you need is to know how much time it will take to qolve your system, and hopefully, what is the most suitable solver. In Eigen, we provide a benchmark routine that can be used for this purpose. It is very easy to use. In the build directory, navigate to bench/spbench and compile the routine by typing \b make \e spbenchsolver. Run it with --help option to get the list of all available options. Basically, the matrices to test should be in MatrixMarket Coordinate format, and the routine returns the statistics from all available solvers in Eigen. +Most of the time, all you need is to know how much time it will take to solve your system, and hopefully, what is the most suitable solver. In Eigen, we provide a benchmark routine that can be used for this purpose. It is very easy to use. In the build directory, navigate to bench/spbench and compile the routine by typing \b make \e spbenchsolver. Run it with --help option to get the list of all available options. Basically, the matrices to test should be in MatrixMarket Coordinate format, and the routine returns the statistics from all available solvers in Eigen. To export your matrices and right-hand-side vectors in the matrix-market format, you can the the unsupported SparseExtra module: \code From 1c3843bf867742a40e0c6b83c1200247bef5cde5 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 23 Jun 2014 11:04:12 +0200 Subject: [PATCH 0545/1103] Fix bug #729: Use alloca if it is defined --- Eigen/src/Core/util/Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 390b60c74..9718ef6a8 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -552,7 +552,7 @@ template struct smart_memmove_helper { // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA // to the appropriate stack allocation function #ifndef EIGEN_ALLOCA - #if (defined __linux__) || (defined __APPLE__) + #if (defined __linux__) || (defined __APPLE__) || (defined alloca) #define EIGEN_ALLOCA alloca #elif defined(_MSC_VER) #define EIGEN_ALLOCA _alloca From 3117036b80075390dbc46f60aa0d595e5a44661b Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 23 Jun 2014 11:15:42 +0200 Subject: [PATCH 0546/1103] Fix bug #826: Allow initialization of 1x1 Arrays/Matrices by passing a value. --- Eigen/src/Core/Array.h | 44 ++++++++++++++-------------- Eigen/src/Core/Map.h | 13 --------- Eigen/src/Core/Matrix.h | 34 ++++++++++++---------- Eigen/src/Core/PlainObjectBase.h | 49 ++++++++++++++++++++++++++++++++ 4 files changed, 91 insertions(+), 49 deletions(-) diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h index 8d2906a10..28d6f1443 100644 --- a/Eigen/src/Core/Array.h +++ b/Eigen/src/Core/Array.h @@ -144,24 +144,16 @@ class Array } #endif - /** Constructs a vector or row-vector with given dimension. \only_for_vectors - * - * Note that this is only useful for dynamic-size vectors. For fixed-size vectors, - * it is redundant to pass the dimension here, so it makes more sense to use the default - * constructor Matrix() instead. - */ - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE explicit Array(Index dim) - : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim) - { - Base::_check_template_params(); - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Array) - eigen_assert(dim >= 0); - eigen_assert(SizeAtCompileTime == Dynamic || SizeAtCompileTime == dim); - EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED - } #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE explicit Array(const T& x) + { + Base::_check_template_params(); + Base::template _init1(x); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Array(const T0& val0, const T1& val1) @@ -170,11 +162,23 @@ class Array this->template _init2(val0, val1); } #else - /** constructs an uninitialized matrix with \a rows rows and \a cols columns. + /** \brief Constructs a fixed-sized array initialized with coefficients starting at \a data */ + EIGEN_DEVICE_FUNC explicit Array(const Scalar *data); + /** Constructs a vector or row-vector with given dimension. \only_for_vectors * - * This is useful for dynamic-size matrices. For fixed-size matrices, + * Note that this is only useful for dynamic-size vectors. For fixed-size vectors, + * it is redundant to pass the dimension here, so it makes more sense to use the default + * constructor Array() instead. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE explicit Array(Index dim); + /** constructs an initialized 1x1 Array with the given coefficient */ + Array(const Scalar& value); + /** constructs an uninitialized array with \a rows rows and \a cols columns. + * + * This is useful for dynamic-size arrays. For fixed-size arrays, * it is redundant to pass these parameters, so one should use the default constructor - * Matrix() instead. */ + * Array() instead. */ Array(Index rows, Index cols); /** constructs an initialized 2D vector with given coefficients */ Array(const Scalar& val0, const Scalar& val1); @@ -202,8 +206,6 @@ class Array m_storage.data()[3] = val3; } - EIGEN_DEVICE_FUNC explicit Array(const Scalar *data); - /** Constructor copying the value of the expression \a other */ template EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index c75a5e95f..0838d69e3 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -179,19 +179,6 @@ template class Ma StrideType m_stride; }; -template -inline Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> - ::Array(const Scalar *data) -{ - this->_set_noalias(Eigen::Map(data)); -} - -template -inline Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> - ::Matrix(const Scalar *data) -{ - this->_set_noalias(Eigen::Map(data)); -} } // end namespace Eigen diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index c2cedbf6a..782d67f54 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -232,24 +232,17 @@ class Matrix } #endif - /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors - * - * Note that this is only useful for dynamic-size vectors. For fixed-size vectors, - * it is redundant to pass the dimension here, so it makes more sense to use the default - * constructor Matrix() instead. - */ + #ifndef EIGEN_PARSED_BY_DOXYGEN + + // This constructor is for both 1x1 matrices and dynamic vectors + template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE explicit Matrix(Index dim) - : Base(dim, RowsAtCompileTime == 1 ? 1 : dim, ColsAtCompileTime == 1 ? 1 : dim) + EIGEN_STRONG_INLINE explicit Matrix(const T& x) { Base::_check_template_params(); - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Matrix) - eigen_assert(dim >= 0); - eigen_assert(SizeAtCompileTime == Dynamic || SizeAtCompileTime == dim); - EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED + Base::template _init1(x); } - #ifndef EIGEN_PARSED_BY_DOXYGEN template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(const T0& x, const T1& y) @@ -258,6 +251,19 @@ class Matrix Base::template _init2(x, y); } #else + /** \brief Constructs a fixed-sized matrix initialized with coefficients starting at \a data */ + EIGEN_DEVICE_FUNC + explicit Matrix(const Scalar *data); + + /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors + * + * Note that this is only useful for dynamic-size vectors. For fixed-size vectors, + * it is redundant to pass the dimension here, so it makes more sense to use the default + * constructor Matrix() instead. + */ + EIGEN_STRONG_INLINE explicit Matrix(Index dim); + /** \brief Constructs an initialized 1x1 matrix with the given coefficient */ + Matrix(const Scalar& x); /** \brief Constructs an uninitialized matrix with \a rows rows and \a cols columns. * * This is useful for dynamic-size matrices. For fixed-size matrices, @@ -291,8 +297,6 @@ class Matrix m_storage.data()[3] = w; } - EIGEN_DEVICE_FUNC - explicit Matrix(const Scalar *data); /** \brief Constructor copying the value of the expression \a other */ template diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 0305066ba..b5b25ef70 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -690,6 +690,55 @@ class PlainObjectBase : public internal::dense_xpr_base::type m_storage.data()[1] = val1; } + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if::type* = 0) + { + EIGEN_STATIC_ASSERT(bool(NumTraits::IsInteger), + FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) + resize(size); + } + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if::type* = 0) + { + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) + m_storage.data()[0] = val0; + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Scalar* data){ + this->_set_noalias(ConstMapType(data)); + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const DenseBase& other){ + this->_set_noalias(other); + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const EigenBase& other){ + this->derived() = other; + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const ReturnByValue& other) + { + resize(other.rows(), other.cols()); + other.evalTo(this->derived()); + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const RotationBase& r) + { + this->derived() = r; + } + template friend struct internal::matrix_swap_impl; From 15c2c083e818ee6ca533e2b99a661ec4a0bba88c Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 23 Jun 2014 11:21:40 +0200 Subject: [PATCH 0547/1103] Additional unit tests for bug #826 by Gael --- test/basicstuff.cpp | 46 ++++++++++++++++++++++++++++++++++++--------- 1 file changed, 37 insertions(+), 9 deletions(-) diff --git a/test/basicstuff.cpp b/test/basicstuff.cpp index 8c0621ecd..56370d591 100644 --- a/test/basicstuff.cpp +++ b/test/basicstuff.cpp @@ -180,15 +180,41 @@ void casting() template void fixedSizeMatrixConstruction() { - const Scalar raw[3] = {1,2,3}; - Matrix m(raw); - Array a(raw); - VERIFY(m(0) == 1); - VERIFY(m(1) == 2); - VERIFY(m(2) == 3); - VERIFY(a(0) == 1); - VERIFY(a(1) == 2); - VERIFY(a(2) == 3); + Scalar raw[4]; + for(int k=0; k<4; ++k) + raw[k] = internal::random(); + { + Matrix m(raw); + Array a(raw); + for(int k=0; k<4; ++k) VERIFY(m(k) == raw[k]); + for(int k=0; k<4; ++k) VERIFY(a(k) == raw[k]); + VERIFY_IS_EQUAL(m,(Matrix(raw[0],raw[1],raw[2],raw[3]))); + VERIFY((a==(Array(raw[0],raw[1],raw[2],raw[3]))).all()); + } + { + Matrix m(raw); + Array a(raw); + for(int k=0; k<3; ++k) VERIFY(m(k) == raw[k]); + for(int k=0; k<3; ++k) VERIFY(a(k) == raw[k]); + VERIFY_IS_EQUAL(m,(Matrix(raw[0],raw[1],raw[2]))); + VERIFY((a==Array(raw[0],raw[1],raw[2])).all()); + } + { + Matrix m(raw); + Array a(raw); + for(int k=0; k<2; ++k) VERIFY(m(k) == raw[k]); + for(int k=0; k<2; ++k) VERIFY(a(k) == raw[k]); + VERIFY_IS_EQUAL(m,(Matrix(raw[0],raw[1]))); + VERIFY((a==Array(raw[0],raw[1])).all()); + } + { + Matrix m(raw); + Array a(raw); + VERIFY(m(0) == raw[0]); + VERIFY(a(0) == raw[0]); + VERIFY_IS_EQUAL(m,(Matrix(raw[0]))); + VERIFY((a==Array(raw[0])).all()); + } } void test_basicstuff() @@ -209,6 +235,8 @@ void test_basicstuff() CALL_SUBTEST_1(fixedSizeMatrixConstruction()); CALL_SUBTEST_1(fixedSizeMatrixConstruction()); CALL_SUBTEST_1(fixedSizeMatrixConstruction()); + CALL_SUBTEST_1(fixedSizeMatrixConstruction()); + CALL_SUBTEST_1(fixedSizeMatrixConstruction()); CALL_SUBTEST_2(casting()); } From 957c2c291ba0e2d356584d833432a3314094dfee Mon Sep 17 00:00:00 2001 From: Jeff Date: Mon, 23 Jun 2014 08:34:11 -0600 Subject: [PATCH 0548/1103] Changed uint to unsigned int. --- unsupported/Eigen/src/Splines/SplineFitting.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h index 7c2484e0d..c30129983 100644 --- a/unsupported/Eigen/src/Splines/SplineFitting.h +++ b/unsupported/Eigen/src/Splines/SplineFitting.h @@ -90,8 +90,8 @@ namespace Eigen return; } - uint startIndex; - uint endIndex; + unsigned int startIndex; + unsigned int endIndex; unsigned int numInternalDerivatives = numDerivatives; From 755be9016ab8f8ddcbe96699d772e05de1248c2c Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 23 Jun 2014 22:33:36 +0200 Subject: [PATCH 0549/1103] Workaround clang error introduced by 3117036b80075390dbc46f60aa0d595e5a44661b : "template argument for non-type template parameter is treated as function type 'bool (bool)'" --- Eigen/src/Core/util/StaticAssert.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h index 8872c5b64..ac2c0bedd 100644 --- a/Eigen/src/Core/util/StaticAssert.h +++ b/Eigen/src/Core/util/StaticAssert.h @@ -107,9 +107,9 @@ {Eigen::internal::static_assertion::MSG;} #else - + // In some cases clang interprets bool(CONDITION) as function declaration #define EIGEN_STATIC_ASSERT(CONDITION,MSG) \ - if (Eigen::internal::static_assertion::MSG) {} + if (Eigen::internal::static_assertion(CONDITION)>::MSG) {} #endif From b59f045c07a41e79df9e27cd56ce06642a7012e7 Mon Sep 17 00:00:00 2001 From: Jeff Date: Mon, 23 Jun 2014 19:04:52 -0600 Subject: [PATCH 0550/1103] Using LU decomposition with complete pivoting for better accuracy. --- unsupported/Eigen/src/Splines/SplineFitting.h | 5 +++-- unsupported/test/splines.cpp | 2 +- 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h index c30129983..3b8e70a18 100644 --- a/unsupported/Eigen/src/Splines/SplineFitting.h +++ b/unsupported/Eigen/src/Splines/SplineFitting.h @@ -17,6 +17,7 @@ #include "SplineFwd.h" +#include #include namespace Eigen @@ -404,8 +405,8 @@ namespace Eigen A(n - 1, n - 1) = 1; // Solve - HouseholderQR qr(A); - ControlPointVectorType controlPoints = qr.solve(MatrixType(b.transpose())).transpose(); + FullPivLU lu(A); + ControlPointVectorType controlPoints = lu.solve(MatrixType(b.transpose())).transpose(); SplineType spline(knots, controlPoints); diff --git a/unsupported/test/splines.cpp b/unsupported/test/splines.cpp index 1f3856143..755a21ece 100644 --- a/unsupported/test/splines.cpp +++ b/unsupported/test/splines.cpp @@ -261,7 +261,7 @@ void check_global_interpolation_with_derivatives2d() { PointType point = spline(knots(i)); PointType referencePoint = points.col(i); - VERIFY((point - referencePoint).matrix().norm() < 1e-12); + VERIFY((point - referencePoint).matrix().norm() < 1e-10); PointType derivative = spline.derivatives(knots(i), 1).col(1); PointType referenceDerivative = derivatives.col(i); VERIFY((derivative - referenceDerivative).matrix().norm() < 1e-10); From e86adc87e96ddee689b70560ad3698ecb774408c Mon Sep 17 00:00:00 2001 From: Jeff Date: Mon, 23 Jun 2014 19:52:42 -0600 Subject: [PATCH 0551/1103] Fixed type mixing issues. --- unsupported/Eigen/src/Splines/SplineFitting.h | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h index 3b8e70a18..ed608d365 100644 --- a/unsupported/Eigen/src/Splines/SplineFitting.h +++ b/unsupported/Eigen/src/Splines/SplineFitting.h @@ -82,8 +82,8 @@ namespace Eigen { typedef typename ParameterVectorType::Scalar Scalar; - const unsigned int numParameters = parameters.size(); - const unsigned int numDerivatives = derivativeIndices.size(); + DenseIndex numParameters = parameters.size(); + DenseIndex numDerivatives = derivativeIndices.size(); if (numDerivatives < 1) { @@ -91,10 +91,10 @@ namespace Eigen return; } - unsigned int startIndex; - unsigned int endIndex; + DenseIndex startIndex; + DenseIndex endIndex; - unsigned int numInternalDerivatives = numDerivatives; + DenseIndex numInternalDerivatives = numDerivatives; if (derivativeIndices[0] == 0) { @@ -117,12 +117,12 @@ namespace Eigen // There are (endIndex - startIndex + 1) knots obtained from the averaging // and 2 for the first and last parameters. - unsigned int numAverageKnots = endIndex - startIndex + 3; + DenseIndex numAverageKnots = endIndex - startIndex + 3; KnotVectorType averageKnots(numAverageKnots); averageKnots[0] = parameters[0]; int newKnotIndex = 0; - for (unsigned int i = startIndex; i <= endIndex; ++i) + for (DenseIndex i = startIndex; i <= endIndex; ++i) averageKnots[++newKnotIndex] = parameters.segment(i, degree).mean(); averageKnots[++newKnotIndex] = parameters[numParameters - 1]; @@ -135,7 +135,7 @@ namespace Eigen temporaryParameters[0] = averageKnots[i]; ParameterVectorType parameterIndices(numParameters); int temporaryParameterIndex = 1; - for (size_t j = 0; j < numParameters; ++j) + for (int j = 0; j < numParameters; ++j) { Scalar parameter = parameters[j]; if (parameter >= averageKnots[i] && parameter < averageKnots[i + 1]) From e745a450deb7ff4b7433d1e86e0779b6aece0c46 Mon Sep 17 00:00:00 2001 From: Jeff Date: Mon, 23 Jun 2014 20:18:16 -0600 Subject: [PATCH 0552/1103] Removed tabs and fixed indentation. --- unsupported/Eigen/src/Splines/SplineFitting.h | 82 +++++++++---------- unsupported/test/splines.cpp | 28 +++---- 2 files changed, 55 insertions(+), 55 deletions(-) diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h index 3b8e70a18..7ca8dd9be 100644 --- a/unsupported/Eigen/src/Splines/SplineFitting.h +++ b/unsupported/Eigen/src/Splines/SplineFitting.h @@ -76,9 +76,9 @@ namespace Eigen **/ template void KnotAveragingWithDerivatives(const ParameterVectorType& parameters, - const unsigned int degree, - const IndexArray& derivativeIndices, - KnotVectorType& knots) + const unsigned int degree, + const IndexArray& derivativeIndices, + KnotVectorType& knots) { typedef typename ParameterVectorType::Scalar Scalar; @@ -137,35 +137,35 @@ namespace Eigen int temporaryParameterIndex = 1; for (size_t j = 0; j < numParameters; ++j) { - Scalar parameter = parameters[j]; - if (parameter >= averageKnots[i] && parameter < averageKnots[i + 1]) - { - parameterIndices[temporaryParameterIndex] = j; - temporaryParameters[temporaryParameterIndex++] = parameter; - } + Scalar parameter = parameters[j]; + if (parameter >= averageKnots[i] && parameter < averageKnots[i + 1]) + { + parameterIndices[temporaryParameterIndex] = j; + temporaryParameters[temporaryParameterIndex++] = parameter; + } } temporaryParameters[temporaryParameterIndex] = averageKnots[i + 1]; for (int j = 0; j <= temporaryParameterIndex - 2; ++j) { - for (DenseIndex k = 0; k < derivativeIndices.size(); ++k) - { - if (parameterIndices[j + 1] == derivativeIndices[k] - && parameterIndices[j + 1] != 0 - && parameterIndices[j + 1] != numParameters - 1) - { - derivativeKnots[++newKnotIndex] = temporaryParameters.segment(j, 3).mean(); - break; - } - } + for (DenseIndex k = 0; k < derivativeIndices.size(); ++k) + { + if (parameterIndices[j + 1] == derivativeIndices[k] + && parameterIndices[j + 1] != 0 + && parameterIndices[j + 1] != numParameters - 1) + { + derivativeKnots[++newKnotIndex] = temporaryParameters.segment(j, 3).mean(); + break; + } + } } } KnotVectorType temporaryKnots(averageKnots.size() + derivativeKnots.size()); std::merge(averageKnots.data(), averageKnots.data() + averageKnots.size(), - derivativeKnots.data(), derivativeKnots.data() + derivativeKnots.size(), - temporaryKnots.data()); + derivativeKnots.data(), derivativeKnots.data() + derivativeKnots.size(), + temporaryKnots.data()); // Number of control points (one for each point and derivative) plus spline order. DenseIndex numKnots = numParameters + numDerivatives + degree + 1; @@ -257,9 +257,9 @@ namespace Eigen **/ template static SplineType InterpolateWithDerivatives(const PointArrayType& points, - const PointArrayType& derivatives, - const IndexArray& derivativeIndices, - const unsigned int degree); + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree); /** * \brief Fits an interpolating spline to the given data points and derivatives. @@ -279,10 +279,10 @@ namespace Eigen */ template static SplineType InterpolateWithDerivatives(const PointArrayType& points, - const PointArrayType& derivatives, - const IndexArray& derivativeIndices, - const unsigned int degree, - const ParameterVectorType& parameters); + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree, + const ParameterVectorType& parameters); }; template @@ -330,10 +330,10 @@ namespace Eigen template SplineType SplineFitting::InterpolateWithDerivatives(const PointArrayType& points, - const PointArrayType& derivatives, - const IndexArray& derivativeIndices, - const unsigned int degree, - const ParameterVectorType& parameters) + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree, + const ParameterVectorType& parameters) { typedef typename SplineType::KnotVectorType::Scalar Scalar; typedef typename SplineType::ControlPointVectorType ControlPointVectorType; @@ -387,16 +387,16 @@ namespace Eigen if (derivativeIndices[derivativeIndex] == i) { - A.block(row, span - degree, 2, degree + 1) - = SplineType::BasisFunctionDerivatives(parameters[i], 1, degree, knots); + A.block(row, span - degree, 2, degree + 1) + = SplineType::BasisFunctionDerivatives(parameters[i], 1, degree, knots); - b.col(row++) = points.col(i); - b.col(row++) = derivatives.col(derivativeIndex++); + b.col(row++) = points.col(i); + b.col(row++) = derivatives.col(derivativeIndex++); } else { - A.row(row++).segment(span - degree, degree + 1) - = SplineType::BasisFunctions(parameters[i], degree, knots); + A.row(row++).segment(span - degree, degree + 1) + = SplineType::BasisFunctions(parameters[i], degree, knots); } } b.col(0) = points.col(0); @@ -417,9 +417,9 @@ namespace Eigen template SplineType SplineFitting::InterpolateWithDerivatives(const PointArrayType& points, - const PointArrayType& derivatives, - const IndexArray& derivativeIndices, - const unsigned int degree) + const PointArrayType& derivatives, + const IndexArray& derivativeIndices, + const unsigned int degree) { ParameterVectorType parameters; ChordLengths(points, parameters); diff --git a/unsupported/test/splines.cpp b/unsupported/test/splines.cpp index 755a21ece..ece23953a 100644 --- a/unsupported/test/splines.cpp +++ b/unsupported/test/splines.cpp @@ -13,23 +13,23 @@ namespace Eigen { -// lets do some explicit instantiations and thus -// force the compilation of all spline functions... -template class Spline; -template class Spline; + // lets do some explicit instantiations and thus + // force the compilation of all spline functions... + template class Spline; + template class Spline; -template class Spline; -template class Spline; -template class Spline; -template class Spline; + template class Spline; + template class Spline; + template class Spline; + template class Spline; -template class Spline; -template class Spline; + template class Spline; + template class Spline; -template class Spline; -template class Spline; -template class Spline; -template class Spline; + template class Spline; + template class Spline; + template class Spline; + template class Spline; } From 17f119689e8b0060c41ca6ab0d291bd7d09fe5ce Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 25 Jun 2014 09:58:03 +0200 Subject: [PATCH 0553/1103] implement evaluator for SparseVector --- Eigen/src/SparseCore/SparseVector.h | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index 7e15c814b..cd22dfd49 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -410,6 +410,7 @@ class SparseVector::ReverseInnerIterator namespace internal { +#ifndef EIGEN_TEST_EVALUATORS template< typename Dest, typename Src> struct sparse_vector_assign_selector { static void run(Dest& dst, const Src& src) { @@ -431,6 +432,33 @@ struct sparse_vector_assign_selector { } } }; +#else // EIGEN_TEST_EVALUATORS +template< typename Dest, typename Src> +struct sparse_vector_assign_selector { + static void run(Dest& dst, const Src& src) { + eigen_internal_assert(src.innerSize()==src.size()); + typedef typename internal::evaluator::type SrcEvaluatorType; + SrcEvaluatorType srcEval(src); + for(typename SrcEvaluatorType::InnerIterator it(srcEval, 0); it; ++it) + dst.insert(it.index()) = it.value(); + } +}; + +template< typename Dest, typename Src> +struct sparse_vector_assign_selector { + static void run(Dest& dst, const Src& src) { + eigen_internal_assert(src.outerSize()==src.size()); + typedef typename internal::evaluator::type SrcEvaluatorType; + SrcEvaluatorType srcEval(src); + for(typename Dest::Index i=0; i struct sparse_vector_assign_selector { From e3ba5329ff369401da836ac5e4a86cd059a51a69 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 25 Jun 2014 09:58:26 +0200 Subject: [PATCH 0554/1103] Implement evaluators for sparse Block. --- Eigen/SparseCore | 2 +- Eigen/src/Core/Block.h | 2 + Eigen/src/Core/CoreEvaluators.h | 14 ++- Eigen/src/SparseCore/SparseAssign.h | 1 - Eigen/src/SparseCore/SparseBlock.h | 163 +++++++++++++++++++++++++++- 5 files changed, 175 insertions(+), 7 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 799cb1ba1..4ad6f47a0 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -44,8 +44,8 @@ struct Sparse {}; #include "src/SparseCore/SparseCwiseUnaryOp.h" #include "src/SparseCore/SparseCwiseBinaryOp.h" #include "src/SparseCore/SparseTranspose.h" -#ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparseBlock.h" +#ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparseDot.h" #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseRedux.h" diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 70e457028..0f3ab304a 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -122,6 +122,8 @@ template class typedef Impl Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Block) EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Block) + + typedef typename internal::remove_all::type NestedExpression; /** Column or Row constructor */ diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index b56c3c635..81b4d63bf 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -737,13 +737,25 @@ struct evaluator > evaluator(const XprType& block) : block_evaluator_type(block) {} }; +// no direct-access => dispatch to a unary evaluator template struct block_evaluator - : evaluator_base > + : unary_evaluator > { typedef Block XprType; block_evaluator(const XprType& block) + : unary_evaluator(block) + {} +}; + +template +struct unary_evaluator, IndexBased> + : evaluator_base > +{ + typedef Block XprType; + + unary_evaluator(const XprType& block) : m_argImpl(block.nestedExpression()), m_startRow(block.startRow()), m_startCol(block.startCol()) diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index 99a1a8712..d4434e0ae 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -181,7 +181,6 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src) typedef typename internal::evaluator::type DstEvaluatorType; typedef typename internal::evaluator::type SrcEvaluatorType; - DstEvaluatorType dstEvaluator(dst); SrcEvaluatorType srcEvaluator(src); const bool transpose = (DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit); diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 5b95cc33f..4926f6923 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -12,6 +12,7 @@ namespace Eigen { +// Subset of columns or rows template class BlockImpl : public SparseMatrixBase > @@ -25,6 +26,7 @@ protected: public: EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType) +#ifndef EIGEN_TEST_EVALUATORS class InnerIterator: public XprType::InnerIterator { typedef typename BlockImpl::Index Index; @@ -49,6 +51,7 @@ public: protected: Index m_outer; }; +#endif // EIGEN_TEST_EVALUATORS inline BlockImpl(const XprType& xpr, int i) : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) @@ -63,13 +66,30 @@ public: Index nonZeros() const { - Index nnz = 0; +#ifndef EIGEN_TEST_EVALUATORS + typedef typename internal::evaluator::type EvaluatorType; Index end = m_outerStart + m_outerSize.value(); for(int j=m_outerStart; j::type EvaluatorType; + EvaluatorType matEval(m_matrix); + Index nnz = 0; + Index end = m_outerStart + m_outerSize.value(); + for(int j=m_outerStart; j InnerIterator; class ReverseInnerIterator : public _MatrixTypeNested::ReverseInnerIterator @@ -431,6 +458,14 @@ public: inline operator bool() const { return Base::operator bool() && Base::index() >= m_begin; } }; +#endif // EIGEN_TEST_EVALUATORS + + inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } + Index startRow() const { return m_startRow.value(); } + Index startCol() const { return m_startCol.value(); } + Index blockRows() const { return m_blockRows.value(); } + Index blockCols() const { return m_blockCols.value(); } + protected: friend class internal::GenericSparseBlockInnerIteratorImpl; friend class ReverseInnerIterator; @@ -535,7 +570,127 @@ namespace internal { inline operator bool() const { return m_outerPos < m_end; } }; + +#ifdef EIGEN_TEST_EVALUATORS + +// +template +struct unary_evaluator, IteratorBased > + : public evaluator_base > +{ + class InnerVectorInnerIterator; + class OuterVectorInnerIterator; + public: + typedef Block XprType; + typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + + class ReverseInnerIterator; + + enum { + IsRowMajor = XprType::IsRowMajor, + + OuterVector = (BlockCols==1 && ArgType::IsRowMajor) + | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". + // revert to || as soon as not needed anymore. + (BlockRows==1 && !ArgType::IsRowMajor), + + CoeffReadCost = evaluator::CoeffReadCost, + Flags = XprType::Flags + }; + + typedef typename internal::conditional::type InnerIterator; + + unary_evaluator(const XprType& op) + : m_argImpl(op.nestedExpression()), m_block(op) + {} + + protected: + typedef typename evaluator::InnerIterator EvalIterator; + typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; + + typename evaluator::nestedType m_argImpl; + const XprType &m_block; +}; + +template +class unary_evaluator, IteratorBased>::InnerVectorInnerIterator + : public EvalIterator +{ + const XprType& m_block; + Index m_end; +public: + + EIGEN_STRONG_INLINE InnerVectorInnerIterator(const unary_evaluator& aEval, Index outer) + : EvalIterator(aEval.m_argImpl, outer + (IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol())), + m_block(aEval.m_block), + m_end(IsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows()) + { + while( (EvalIterator::operator bool()) && (EvalIterator::index() < (IsRowMajor ? m_block.startCol() : m_block.startRow())) ) + EvalIterator::operator++(); + } + inline Index index() const { return EvalIterator::index() - (IsRowMajor ? m_block.startCol() : m_block.startRow()); } + inline Index outer() const { return EvalIterator::outer() - (IsRowMajor ? m_block.startRow() : m_block.startCol()); } + inline Index row() const { return EvalIterator::row() - m_block.startRow(); } + inline Index col() const { return EvalIterator::col() - m_block.startCol(); } + + inline operator bool() const { return EvalIterator::operator bool() && EvalIterator::index() < m_end; } +}; + +template +class unary_evaluator, IteratorBased>::OuterVectorInnerIterator +{ + const XprType& m_block; + Index m_outerPos; + Index m_innerIndex; + Scalar m_value; + Index m_end; +public: + + EIGEN_STRONG_INLINE OuterVectorInnerIterator(const unary_evaluator& aEval, Index outer) + : m_block(aEval.m_block), + m_outerPos( (IsRowMajor ? aEval.startCol() : aEval.startRow()) - 1), // -1 so that operator++ finds the first non-zero entry + m_innerIndex(IsRowMajor ? aEval.startRow() : aEval.startCol()), + m_end(IsRowMajor ? aEval.startCol()+aEval.blockCols() : aEval.startRow()+aEval.blockRows()) + { + EIGEN_UNUSED_VARIABLE(outer); + eigen_assert(outer==0); + + ++(*this); + } + + inline Index index() const { return m_outerPos - (IsRowMajor ? m_block.startCol() : m_block.startRow()); } + inline Index outer() const { return 0; } + inline Index row() const { return IsRowMajor ? 0 : index(); } + inline Index col() const { return IsRowMajor ? index() : 0; } + + inline Scalar value() const { return m_value; } + + inline OuterVectorInnerIterator& operator++() + { + // search next non-zero entry + while(m_outerPos Date: Wed, 25 Jun 2014 17:21:04 +0200 Subject: [PATCH 0555/1103] Implement evaluators for sparse coeff-wise views --- Eigen/src/Core/CwiseUnaryView.h | 11 ++ Eigen/src/SparseCore/SparseCwiseUnaryOp.h | 133 ++++++++++++---------- 2 files changed, 83 insertions(+), 61 deletions(-) diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index 92b031e19..99cc03ac1 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -93,6 +93,17 @@ class CwiseUnaryView : public CwiseUnaryViewImpl +class CwiseUnaryViewImpl + : public internal::generic_xpr_base >::type +{ +public: + typedef typename internal::generic_xpr_base >::type Base; +}; +#endif + template class CwiseUnaryViewImpl : public internal::dense_xpr_base< CwiseUnaryView >::type diff --git a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h index ead3af59d..7a849c822 100644 --- a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h @@ -216,67 +216,78 @@ class unary_evaluator, IteratorBased>::ReverseInne }; -// template -// class CwiseUnaryViewImpl -// : public SparseMatrixBase > -// { -// public: -// -// class InnerIterator; -// class ReverseInnerIterator; -// -// typedef CwiseUnaryView Derived; -// EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) -// -// protected: -// typedef typename internal::traits::_MatrixTypeNested _MatrixTypeNested; -// typedef typename _MatrixTypeNested::InnerIterator MatrixTypeIterator; -// typedef typename _MatrixTypeNested::ReverseInnerIterator MatrixTypeReverseIterator; -// }; -// -// template -// class CwiseUnaryViewImpl::InnerIterator -// : public CwiseUnaryViewImpl::MatrixTypeIterator -// { -// typedef typename CwiseUnaryViewImpl::Scalar Scalar; -// typedef typename CwiseUnaryViewImpl::MatrixTypeIterator Base; -// public: -// -// EIGEN_STRONG_INLINE InnerIterator(const CwiseUnaryViewImpl& unaryOp, typename CwiseUnaryViewImpl::Index outer) -// : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) -// {} -// -// EIGEN_STRONG_INLINE InnerIterator& operator++() -// { Base::operator++(); return *this; } -// -// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar value() const { return m_functor(Base::value()); } -// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar& valueRef() { return m_functor(Base::valueRef()); } -// -// protected: -// const ViewOp m_functor; -// }; -// -// template -// class CwiseUnaryViewImpl::ReverseInnerIterator -// : public CwiseUnaryViewImpl::MatrixTypeReverseIterator -// { -// typedef typename CwiseUnaryViewImpl::Scalar Scalar; -// typedef typename CwiseUnaryViewImpl::MatrixTypeReverseIterator Base; -// public: -// -// EIGEN_STRONG_INLINE ReverseInnerIterator(const CwiseUnaryViewImpl& unaryOp, typename CwiseUnaryViewImpl::Index outer) -// : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) -// {} -// -// EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() -// { Base::operator--(); return *this; } -// -// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar value() const { return m_functor(Base::value()); } -// EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar& valueRef() { return m_functor(Base::valueRef()); } -// -// protected: -// const ViewOp m_functor; -// }; + + + +template +struct unary_evaluator, IteratorBased> + : public evaluator_base > +{ + public: + typedef CwiseUnaryView XprType; + + class InnerIterator; + class ReverseInnerIterator; + + enum { + CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost, + Flags = XprType::Flags + }; + + unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) {} + + protected: + typedef typename evaluator::InnerIterator EvalIterator; + typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; + + const ViewOp m_functor; + typename evaluator::nestedType m_argImpl; +}; + +template +class unary_evaluator, IteratorBased>::InnerIterator + : public unary_evaluator, IteratorBased>::EvalIterator +{ + typedef typename XprType::Scalar Scalar; + typedef typename unary_evaluator, IteratorBased>::EvalIterator Base; + public: + + EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& unaryOp, typename XprType::Index outer) + : Base(unaryOp.m_argImpl,outer), m_functor(unaryOp.m_functor) + {} + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { Base::operator++(); return *this; } + + EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } + EIGEN_STRONG_INLINE Scalar& valueRef() { return m_functor(Base::valueRef()); } + + protected: + const ViewOp m_functor; +}; + +template +class unary_evaluator, IteratorBased>::ReverseInnerIterator + : public unary_evaluator, IteratorBased>::EvalReverseIterator +{ + typedef typename XprType::Scalar Scalar; + typedef typename unary_evaluator, IteratorBased>::EvalReverseIterator Base; + public: + + EIGEN_STRONG_INLINE ReverseInnerIterator(const XprType& unaryOp, typename XprType::Index outer) + : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) + {} + + EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() + { Base::operator--(); return *this; } + + EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } + EIGEN_STRONG_INLINE Scalar& valueRef() { return m_functor(Base::valueRef()); } + + protected: + const ViewOp m_functor; +}; + } // end namespace internal From 3b19b466a7609fd381db2c53569d4d1d1621dbfc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 25 Jun 2014 17:22:12 +0200 Subject: [PATCH 0556/1103] Generalize static assertions on matching sizes to avoid the need for SizeAtCompileTime --- Eigen/src/Core/util/StaticAssert.h | 2 +- Eigen/src/Core/util/XprHelper.h | 5 +++++ 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h index f5691826e..eff12486d 100644 --- a/Eigen/src/Core/util/StaticAssert.h +++ b/Eigen/src/Core/util/StaticAssert.h @@ -159,7 +159,7 @@ #define EIGEN_PREDICATE_SAME_MATRIX_SIZE(TYPE0,TYPE1) \ ( \ - (int(TYPE0::SizeAtCompileTime)==0 && int(TYPE1::SizeAtCompileTime)==0) \ + (int(internal::size_of_xpr_at_compile_time::ret)==0 && int(internal::size_of_xpr_at_compile_time::ret)==0) \ || (\ (int(TYPE0::RowsAtCompileTime)==Eigen::Dynamic \ || int(TYPE1::RowsAtCompileTime)==Eigen::Dynamic \ diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 5c4b9dbcd..19adbefd7 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -216,6 +216,11 @@ template struct size_at_compile_time enum { ret = (_Rows==Dynamic || _Cols==Dynamic) ? Dynamic : _Rows * _Cols }; }; +template struct size_of_xpr_at_compile_time +{ + enum { ret = size_at_compile_time::RowsAtCompileTime,traits::ColsAtCompileTime>::ret }; +}; + /* plain_matrix_type : the difference from eval is that plain_matrix_type is always a plain matrix type, * whereas eval is a const reference in the case of a matrix */ From b868bfb84a4a86231bb338e9bcf08afd882d32dc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 25 Jun 2014 17:23:52 +0200 Subject: [PATCH 0557/1103] Make operator=(EigenBase<>) uses the new assignment mechanism and introduce a generic EigenBase to EigenBase assignment kind based on the previous evalTo mechanism. --- Eigen/src/Core/Assign.h | 2 +- Eigen/src/Core/AssignEvaluator.h | 21 ++++++++++++++++++- Eigen/src/Core/BandMatrix.h | 23 +++++++++++++++++++++ Eigen/src/Core/CoreEvaluators.h | 7 ++----- Eigen/src/Core/DiagonalMatrix.h | 21 +++++++++++++++++++ Eigen/src/Core/PermutationMatrix.h | 2 ++ Eigen/src/Householder/HouseholderSequence.h | 12 +++++++++++ Eigen/src/SparseCore/SparseAssign.h | 20 ++++++++++++++++++ Eigen/src/SparseCore/SparseBlock.h | 16 +++++++------- Eigen/src/SparseCore/SparseMatrixBase.h | 2 ++ 10 files changed, 111 insertions(+), 15 deletions(-) diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index abbba84c3..071cfbf7e 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -611,7 +611,7 @@ template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const EigenBase& other) { - other.derived().evalTo(derived()); + internal::call_assignment(derived(), other.derived()); return derived(); } diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 77b4dd9bc..c45ac96f2 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -663,7 +663,9 @@ template struct AssignmentKind; // Assignement kind defined in this file: struct Dense2Dense {}; +struct EigenBase2EigenBase {}; +template struct AssignmentKind { typedef EigenBase2EigenBase Kind; }; template<> struct AssignmentKind { typedef Dense2Dense Kind; }; // This is the main assignment class @@ -750,8 +752,12 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) // TODO check whether this is the right place to perform these checks: EIGEN_STATIC_ASSERT_LVALUE(Dst) EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(ActualDstTypeCleaned,Src) - EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar); + // TODO this line is commented to allow matrix = permutation + // Actually, the "Scalar" type for a permutation matrix does not really make sense, + // perhaps it could be void, and EIGEN_CHECK_BINARY_COMPATIBILIY could allow micing void with anything...? +// EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar); + Assignment::run(actualDst, src, func); } template @@ -772,6 +778,19 @@ struct Assignment } }; +// Generic assignment through evalTo. +// TODO: not sure we have to keep that one, but it helps porting current code to new evaluator mechanism. +template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> +struct Assignment +{ + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) + { + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + + src.evalTo(dst); + } +}; + } // namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/BandMatrix.h b/Eigen/src/Core/BandMatrix.h index ffd7fe8b3..ba4749707 100644 --- a/Eigen/src/Core/BandMatrix.h +++ b/Eigen/src/Core/BandMatrix.h @@ -327,6 +327,29 @@ class TridiagonalMatrix : public BandMatrix +struct evaluator_traits > + : public evaluator_traits_base > +{ + typedef BandShape Shape; +}; + +template +struct evaluator_traits > + : public evaluator_traits_base > +{ + typedef BandShape Shape; +}; + +template<> struct AssignmentKind { typedef EigenBase2EigenBase Kind; }; +#endif + + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 81b4d63bf..f9f229b09 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -32,10 +32,7 @@ struct storage_kind_to_evaluator_kind { template struct storage_kind_to_shape; -template<> -struct storage_kind_to_shape { - typedef DenseShape Shape; -}; +template<> struct storage_kind_to_shape { typedef DenseShape Shape; }; // Evaluators have to be specialized with respect to various criteria such as: // - storage/structure/shape @@ -507,7 +504,7 @@ protected: // -------------------- CwiseUnaryView -------------------- template -struct unary_evaluator > +struct unary_evaluator, IndexBased> : evaluator_base > { typedef CwiseUnaryView XprType; diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index ba0042ba4..fc9ecf561 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -44,6 +44,8 @@ class DiagonalBase : public EigenBase EIGEN_DEVICE_FUNC DenseMatrixType toDenseMatrix() const { return derived(); } + +#ifndef EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC void evalTo(MatrixBase &other) const; @@ -55,6 +57,7 @@ class DiagonalBase : public EigenBase EIGEN_DEVICE_FUNC void subTo(MatrixBase &other) const { other.diagonal() -= diagonal(); } +#endif // EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); } @@ -122,6 +125,7 @@ class DiagonalBase : public EigenBase #endif }; +#ifndef EIGEN_TEST_EVALUATORS template template void DiagonalBase::evalTo(MatrixBase &other) const @@ -129,6 +133,8 @@ void DiagonalBase::evalTo(MatrixBase &other) const other.setZero(); other.diagonal() = diagonal(); } +#endif // EIGEN_TEST_EVALUATORS + #endif /** \class DiagonalMatrix @@ -376,6 +382,21 @@ struct evaluator_traits > static const int AssumeAliasing = 0; }; +struct Diagonal2Dense {}; + +template<> struct AssignmentKind { typedef Diagonal2Dense Kind; }; + +// Diagonal matrix to Dense assignment +template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> +struct Assignment +{ + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) + { + dst.setZero(); + dst.diagonal() = src.diagonal(); + } +}; + } // namespace internal #endif // EIGEN_ENABLE_EVALUATORS diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 61aa0ce31..4399e1d64 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -803,6 +803,8 @@ struct evaluator_traits > > static const int AssumeAliasing = 0; }; +template<> struct AssignmentKind { typedef EigenBase2EigenBase Kind; }; + } // end namespace internal #endif // EIGEN_TEST_EVALUATORS diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h index d800ca1fa..99d3eb21f 100644 --- a/Eigen/src/Householder/HouseholderSequence.h +++ b/Eigen/src/Householder/HouseholderSequence.h @@ -73,6 +73,18 @@ struct traits > }; }; +#ifdef EIGEN_TEST_EVALUATORS + +struct HouseholderSequenceShape {}; + +template +struct evaluator_traits > + : public evaluator_traits_base > +{ + typedef HouseholderSequenceShape Shape; +}; +#endif + template struct hseq_side_dependent_impl { diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index d4434e0ae..86a0b08c5 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -167,8 +167,10 @@ struct storage_kind_to_shape { }; struct Sparse2Sparse {}; +struct Sparse2Dense {}; template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; +template<> struct AssignmentKind { typedef Sparse2Dense Kind; }; template @@ -242,6 +244,24 @@ struct Assignment } }; +// Sparse to Dense assignment +template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> +struct Assignment +{ + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) + { + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + typedef typename SrcXprType::Index Index; + + dst.setZero(); + typename internal::evaluator::type srcEval(src); + typename internal::evaluator::type dstEval(dst); + for (Index j=0; j::InnerIterator i(srcEval,j); i; ++i) + dstEval.coeffRef(i.row(),i.col()) = i.value(); + } +}; + } // end namespace internal #endif // EIGEN_TEST_EVALUATORS diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 4926f6923..f562b5479 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -67,7 +67,7 @@ public: Index nonZeros() const { #ifndef EIGEN_TEST_EVALUATORS - typedef typename internal::evaluator::type EvaluatorType; + Index nnz = 0; Index end = m_outerStart + m_outerSize.value(); for(int j=m_outerStart; j class unary_evaluator, IteratorBased>::OuterVectorInnerIterator { - const XprType& m_block; + const unary_evaluator& m_eval; Index m_outerPos; Index m_innerIndex; Scalar m_value; @@ -649,10 +649,10 @@ class unary_evaluator, IteratorBas public: EIGEN_STRONG_INLINE OuterVectorInnerIterator(const unary_evaluator& aEval, Index outer) - : m_block(aEval.m_block), - m_outerPos( (IsRowMajor ? aEval.startCol() : aEval.startRow()) - 1), // -1 so that operator++ finds the first non-zero entry - m_innerIndex(IsRowMajor ? aEval.startRow() : aEval.startCol()), - m_end(IsRowMajor ? aEval.startCol()+aEval.blockCols() : aEval.startRow()+aEval.blockRows()) + : m_eval(aEval), + m_outerPos( (IsRowMajor ? aEval.m_block.startCol() : aEval.m_block.startRow()) - 1), // -1 so that operator++ finds the first non-zero entry + m_innerIndex(IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol()), + m_end(IsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows()) { EIGEN_UNUSED_VARIABLE(outer); eigen_assert(outer==0); @@ -660,7 +660,7 @@ public: ++(*this); } - inline Index index() const { return m_outerPos - (IsRowMajor ? m_block.startCol() : m_block.startRow()); } + inline Index index() const { return m_outerPos - (IsRowMajor ? m_eval.m_block.startCol() : m_eval.m_block.startRow()); } inline Index outer() const { return 0; } inline Index row() const { return IsRowMajor ? 0 : index(); } inline Index col() const { return IsRowMajor ? index() : 0; } @@ -673,7 +673,7 @@ public: while(m_outerPos class SparseMatrixBase : public EigenBase Block innerVectors(Index outerStart, Index outerSize); const Block innerVectors(Index outerStart, Index outerSize) const; +#ifndef EIGEN_TEST_EVALUATORS /** \internal use operator= */ template void evalTo(MatrixBase& dst) const @@ -346,6 +347,7 @@ template class SparseMatrixBase : public EigenBase for (typename Derived::InnerIterator i(derived(),j); i; ++i) dst.coeffRef(i.row(),i.col()) = i.value(); } +#endif Matrix toDense() const { From a7bd4c455acbf19166e91f2148c042c3f2ea946e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 25 Jun 2014 17:24:43 +0200 Subject: [PATCH 0558/1103] Update sparse reduxions and sparse-vectors to evaluators. --- Eigen/SparseCore | 4 ++-- Eigen/src/SparseCore/SparseDot.h | 13 +++++++++++++ Eigen/src/SparseCore/SparseRedux.h | 6 ++++++ Eigen/src/SparseCore/SparseVector.h | 22 ++++++++++++++++++++++ 4 files changed, 43 insertions(+), 2 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 4ad6f47a0..0433f4877 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -45,10 +45,10 @@ struct Sparse {}; #include "src/SparseCore/SparseCwiseBinaryOp.h" #include "src/SparseCore/SparseTranspose.h" #include "src/SparseCore/SparseBlock.h" -#ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparseDot.h" -#include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseRedux.h" +#ifndef EIGEN_TEST_EVALUATORS +#include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" #include "src/SparseCore/ConservativeSparseSparseProduct.h" #include "src/SparseCore/SparseSparseProductWithPruning.h" diff --git a/Eigen/src/SparseCore/SparseDot.h b/Eigen/src/SparseCore/SparseDot.h index db39c9aec..a63cb003c 100644 --- a/Eigen/src/SparseCore/SparseDot.h +++ b/Eigen/src/SparseCore/SparseDot.h @@ -26,7 +26,12 @@ SparseMatrixBase::dot(const MatrixBase& other) const eigen_assert(size() == other.size()); eigen_assert(other.size()>0 && "you are using a non initialized vector"); +#ifndef EIGEN_TEST_EVALUATORS typename Derived::InnerIterator i(derived(),0); +#else + typename internal::evaluator::type thisEval(derived()); + typename internal::evaluator::InnerIterator i(thisEval, 0); +#endif Scalar res(0); while (i) { @@ -49,6 +54,7 @@ SparseMatrixBase::dot(const SparseMatrixBase& other) cons eigen_assert(size() == other.size()); +#ifndef EIGEN_TEST_EVALUATORS typedef typename Derived::Nested Nested; typedef typename OtherDerived::Nested OtherNested; typedef typename internal::remove_all::type NestedCleaned; @@ -59,6 +65,13 @@ SparseMatrixBase::dot(const SparseMatrixBase& other) cons typename NestedCleaned::InnerIterator i(nthis,0); typename OtherNestedCleaned::InnerIterator j(nother,0); +#else + typename internal::evaluator::type thisEval(derived()); + typename internal::evaluator::InnerIterator i(thisEval, 0); + + typename internal::evaluator::type otherEval(other.derived()); + typename internal::evaluator::InnerIterator j(otherEval, 0); +#endif Scalar res(0); while (i && j) { diff --git a/Eigen/src/SparseCore/SparseRedux.h b/Eigen/src/SparseCore/SparseRedux.h index f3da93a71..c6d0182d5 100644 --- a/Eigen/src/SparseCore/SparseRedux.h +++ b/Eigen/src/SparseCore/SparseRedux.h @@ -18,8 +18,14 @@ SparseMatrixBase::sum() const { eigen_assert(rows()>0 && cols()>0 && "you are using a non initialized matrix"); Scalar res(0); +#ifndef EIGEN_TEST_EVALUATORS for (Index j=0; j::type thisEval(derived()); + for (Index j=0; j::InnerIterator iter(thisEval,j); iter; ++iter) +#endif res += iter.value(); return res; } diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index cd22dfd49..81f5eb41d 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -433,6 +433,28 @@ struct sparse_vector_assign_selector { } }; #else // EIGEN_TEST_EVALUATORS + +template +struct evaluator > + : evaluator_base > +{ + typedef SparseVector<_Scalar,_Options,_Index> SparseVectorType; + typedef typename SparseVectorType::InnerIterator InnerIterator; + typedef typename SparseVectorType::ReverseInnerIterator ReverseInnerIterator; + + enum { + CoeffReadCost = NumTraits<_Scalar>::ReadCost, + Flags = SparseVectorType::Flags + }; + + evaluator(const SparseVectorType &mat) : m_matrix(mat) {} + + operator SparseVectorType&() { return m_matrix.const_cast_derived(); } + operator const SparseVectorType&() const { return m_matrix; } + + const SparseVectorType &m_matrix; +}; + template< typename Dest, typename Src> struct sparse_vector_assign_selector { static void run(Dest& dst, const Src& src) { From d73ee84d37f4a507524348edf8362a035e802a33 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 25 Jun 2014 22:44:43 +0200 Subject: [PATCH 0559/1103] Disabled HIDE_SCOPE_NAMES (default doxygen setting). This might help to avoid API confusions as in bug #830. --- doc/Doxyfile.in | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in index 7bbf693a0..509bd357a 100644 --- a/doc/Doxyfile.in +++ b/doc/Doxyfile.in @@ -474,7 +474,7 @@ CASE_SENSE_NAMES = YES # will show members with their full class and namespace scopes in the # documentation. If set to YES the scope will be hidden. -HIDE_SCOPE_NAMES = YES +HIDE_SCOPE_NAMES = NO # If the SHOW_INCLUDE_FILES tag is set to YES (the default) then Doxygen # will put a list of the files that are included by a file in the documentation From 54607665ab1650cc1d63575b231fa15e514d8aa8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 25 Jun 2014 23:44:59 +0200 Subject: [PATCH 0560/1103] Fix inverse evaluator --- Eigen/src/Core/Inverse.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 2df36dbd5..1d167867f 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -109,8 +109,8 @@ struct unary_evaluator > typedef typename InverseType::PlainObject PlainObject; typedef typename evaluator::type Base; - typedef evaluator type; - typedef evaluator nestedType; + typedef evaluator type; + typedef evaluator nestedType; enum { Flags = Base::Flags | EvalBeforeNestingBit }; From 08c615f1e425d6749ef45ffcbf574e550ba02fea Mon Sep 17 00:00:00 2001 From: Jeff Date: Wed, 25 Jun 2014 19:02:57 -0600 Subject: [PATCH 0561/1103] IndexArray is now a typename. Changed interpolate with derivatives test to use VERIFY_IS_APPROX. --- unsupported/Eigen/src/Splines/SplineFitting.h | 10 +++++----- unsupported/Eigen/src/Splines/SplineFwd.h | 2 -- unsupported/test/splines.cpp | 19 +++++++++++-------- 3 files changed, 16 insertions(+), 15 deletions(-) diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h index 9567c69c2..85de38b25 100644 --- a/unsupported/Eigen/src/Splines/SplineFitting.h +++ b/unsupported/Eigen/src/Splines/SplineFitting.h @@ -74,7 +74,7 @@ namespace Eigen * Curve interpolation with directional constraints for engineering design. * Engineering with Computers **/ - template + template void KnotAveragingWithDerivatives(const ParameterVectorType& parameters, const unsigned int degree, const IndexArray& derivativeIndices, @@ -255,7 +255,7 @@ namespace Eigen * Curve interpolation with directional constraints for engineering design. * Engineering with Computers **/ - template + template static SplineType InterpolateWithDerivatives(const PointArrayType& points, const PointArrayType& derivatives, const IndexArray& derivativeIndices, @@ -277,7 +277,7 @@ namespace Eigen * Curve interpolation with directional constraints for engineering design. * Engineering with Computers */ - template + template static SplineType InterpolateWithDerivatives(const PointArrayType& points, const PointArrayType& derivatives, const IndexArray& derivativeIndices, @@ -327,7 +327,7 @@ namespace Eigen } template - template + template SplineType SplineFitting::InterpolateWithDerivatives(const PointArrayType& points, const PointArrayType& derivatives, @@ -414,7 +414,7 @@ namespace Eigen } template - template + template SplineType SplineFitting::InterpolateWithDerivatives(const PointArrayType& points, const PointArrayType& derivatives, diff --git a/unsupported/Eigen/src/Splines/SplineFwd.h b/unsupported/Eigen/src/Splines/SplineFwd.h index ec3dc0ff5..0a95fbf3e 100644 --- a/unsupported/Eigen/src/Splines/SplineFwd.h +++ b/unsupported/Eigen/src/Splines/SplineFwd.h @@ -88,8 +88,6 @@ namespace Eigen /** \brief 3D double B-spline with dynamic degree. */ typedef Spline Spline3d; - - typedef Array IndexArray; } #endif // EIGEN_SPLINES_FWD_H diff --git a/unsupported/test/splines.cpp b/unsupported/test/splines.cpp index ece23953a..0b43eefb8 100644 --- a/unsupported/test/splines.cpp +++ b/unsupported/test/splines.cpp @@ -249,7 +249,7 @@ void check_global_interpolation_with_derivatives2d() Eigen::ChordLengths(points, knots); ArrayXXd derivatives = ArrayXXd::Random(dimension, numPoints); - Eigen::IndexArray derivativeIndices(numPoints); + VectorXd derivativeIndices(numPoints); for (Eigen::DenseIndex i = 0; i < numPoints; ++i) derivativeIndices(i) = i; @@ -261,18 +261,21 @@ void check_global_interpolation_with_derivatives2d() { PointType point = spline(knots(i)); PointType referencePoint = points.col(i); - VERIFY((point - referencePoint).matrix().norm() < 1e-10); + VERIFY_IS_APPROX(point, referencePoint); PointType derivative = spline.derivatives(knots(i), 1).col(1); PointType referenceDerivative = derivatives.col(i); - VERIFY((derivative - referenceDerivative).matrix().norm() < 1e-10); + VERIFY_IS_APPROX(derivative, referenceDerivative); } } void test_splines() { - CALL_SUBTEST( eval_spline3d() ); - CALL_SUBTEST( eval_spline3d_onbrks() ); - CALL_SUBTEST( eval_closed_spline2d() ); - CALL_SUBTEST( check_global_interpolation2d() ); - CALL_SUBTEST( check_global_interpolation_with_derivatives2d() ); + for (int i = 0; i < g_repeat; ++i) + { + CALL_SUBTEST( eval_spline3d() ); + CALL_SUBTEST( eval_spline3d_onbrks() ); + CALL_SUBTEST( eval_closed_spline2d() ); + CALL_SUBTEST( check_global_interpolation2d() ); + CALL_SUBTEST( check_global_interpolation_with_derivatives2d() ); + } } From f0648f886058a85930186bd935ffb2c9fa87b1f3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 26 Jun 2014 13:52:19 +0200 Subject: [PATCH 0562/1103] Implement evaluator for sparse views. --- Eigen/SparseCore | 2 +- Eigen/src/SparseCore/SparseUtil.h | 21 ++++ Eigen/src/SparseCore/SparseView.h | 153 +++++++++++++++++++++++++++++- 3 files changed, 171 insertions(+), 5 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 0433f4877..340ff7f52 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -47,6 +47,7 @@ struct Sparse {}; #include "src/SparseCore/SparseBlock.h" #include "src/SparseCore/SparseDot.h" #include "src/SparseCore/SparseRedux.h" +#include "src/SparseCore/SparseView.h" #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" @@ -58,7 +59,6 @@ struct Sparse {}; #include "src/SparseCore/SparseTriangularView.h" #include "src/SparseCore/SparseSelfAdjointView.h" #include "src/SparseCore/TriangularSolver.h" -#include "src/SparseCore/SparseView.h" #endif #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index e23576572..2a67af291 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -43,6 +43,8 @@ EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(Derived, -=) \ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, *=) \ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) +#ifndef EIGEN_TEST_EVALUATORS + #define _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, BaseClass) \ typedef BaseClass Base; \ typedef typename Eigen::internal::traits::Scalar Scalar; \ @@ -59,6 +61,25 @@ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) using Base::derived; \ using Base::const_cast_derived; +#else // EIGEN_TEST_EVALUATORS + +#define _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, BaseClass) \ + typedef BaseClass Base; \ + typedef typename Eigen::internal::traits::Scalar Scalar; \ + typedef typename Eigen::NumTraits::Real RealScalar; \ + typedef typename Eigen::internal::nested::type Nested; \ + typedef typename Eigen::internal::traits::StorageKind StorageKind; \ + typedef typename Eigen::internal::traits::Index Index; \ + enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ + ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ + Flags = Eigen::internal::traits::Flags, \ + SizeAtCompileTime = Base::SizeAtCompileTime, \ + IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \ + using Base::derived; \ + using Base::const_cast_derived; + +#endif // EIGEN_TEST_EVALUATORS + #define EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) \ _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, Eigen::SparseMatrixBase) diff --git a/Eigen/src/SparseCore/SparseView.h b/Eigen/src/SparseCore/SparseView.h index fd8450463..96d0a849c 100644 --- a/Eigen/src/SparseCore/SparseView.h +++ b/Eigen/src/SparseCore/SparseView.h @@ -34,25 +34,36 @@ class SparseView : public SparseMatrixBase > typedef typename internal::remove_all::type _MatrixTypeNested; public: EIGEN_SPARSE_PUBLIC_INTERFACE(SparseView) + typedef typename internal::remove_all::type NestedExpression; SparseView(const MatrixType& mat, const Scalar& m_reference = Scalar(0), - typename NumTraits::Real m_epsilon = NumTraits::dummy_precision()) : + RealScalar m_epsilon = NumTraits::dummy_precision()) : m_matrix(mat), m_reference(m_reference), m_epsilon(m_epsilon) {} +#ifndef EIGEN_TEST_EVALUATORS class InnerIterator; +#endif // EIGEN_TEST_EVALUATORS inline Index rows() const { return m_matrix.rows(); } inline Index cols() const { return m_matrix.cols(); } inline Index innerSize() const { return m_matrix.innerSize(); } inline Index outerSize() const { return m_matrix.outerSize(); } - + + /** \returns the nested expression */ + const typename internal::remove_all::type& + nestedExpression() const { return m_matrix; } + + Scalar reference() const { return m_reference; } + RealScalar epsilon() const { return m_epsilon; } + protected: MatrixTypeNested m_matrix; Scalar m_reference; - typename NumTraits::Real m_epsilon; + RealScalar m_epsilon; }; +#ifndef EIGEN_TEST_EVALUATORS template class SparseView::InnerIterator : public _MatrixTypeNested::InnerIterator { @@ -80,13 +91,147 @@ protected: private: void incrementToNonZero() { - while((bool(*this)) && internal::isMuchSmallerThan(value(), m_view.m_reference, m_view.m_epsilon)) + while((bool(*this)) && internal::isMuchSmallerThan(value(), m_view.reference(), m_view.epsilon())) { IterBase::operator++(); } } }; +#else // EIGEN_TEST_EVALUATORS + +namespace internal { + +// TODO find a way to unify the two following variants +// This is tricky because implementing an inner iterator on top of an IndexBased evaluator is +// not easy because the evaluators do not expose the sizes of the underlying expression. + +template +struct unary_evaluator, IteratorBased> + : public evaluator_base > +{ + typedef typename evaluator::InnerIterator EvalIterator; + public: + typedef SparseView XprType; + + class InnerIterator : public EvalIterator + { + typedef typename XprType::Scalar Scalar; + public: + + EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& sve, typename XprType::Index outer) + : EvalIterator(sve.m_argImpl,outer), m_view(sve.m_view) + { + incrementToNonZero(); + } + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { + EvalIterator::operator++(); + incrementToNonZero(); + return *this; + } + + using EvalIterator::value; + + protected: + const XprType &m_view; + + private: + void incrementToNonZero() + { + while((bool(*this)) && internal::isMuchSmallerThan(value(), m_view.reference(), m_view.epsilon())) + { + EvalIterator::operator++(); + } + } + }; + + enum { + CoeffReadCost = evaluator::CoeffReadCost, + Flags = XprType::Flags + }; + + unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {} + + protected: + typename evaluator::nestedType m_argImpl; + const XprType &m_view; +}; + +template +struct unary_evaluator, IndexBased> + : public evaluator_base > +{ + public: + typedef SparseView XprType; + protected: + enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit }; + typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + public: + + class InnerIterator + { + public: + + EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& sve, typename XprType::Index outer) + : m_sve(sve), m_inner(0), m_outer(outer), m_end(sve.m_view.innerSize()) + { + incrementToNonZero(); + } + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { + m_inner++; + incrementToNonZero(); + return *this; + } + + EIGEN_STRONG_INLINE Scalar value() const + { + return (IsRowMajor) ? m_sve.m_argImpl.coeff(m_outer, m_inner) + : m_sve.m_argImpl.coeff(m_inner, m_outer); + } + + EIGEN_STRONG_INLINE Index index() const { return m_inner; } + inline Index row() const { return IsRowMajor ? m_outer : index(); } + inline Index col() const { return IsRowMajor ? index() : m_outer; } + + EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; } + + protected: + const unary_evaluator &m_sve; + Index m_inner; + const Index m_outer; + const Index m_end; + + private: + void incrementToNonZero() + { + while((bool(*this)) && internal::isMuchSmallerThan(value(), m_sve.m_view.reference(), m_sve.m_view.epsilon())) + { + m_inner++; + } + } + }; + + enum { + CoeffReadCost = evaluator::CoeffReadCost, + Flags = XprType::Flags + }; + + unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {} + + protected: + typename evaluator::nestedType m_argImpl; + const XprType &m_view; +}; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS + template const SparseView MatrixBase::sparseView(const Scalar& m_reference, const typename NumTraits::Real& m_epsilon) const From ae039dde135a6af852d7028abd772316613a5249 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 27 Jun 2014 15:53:51 +0200 Subject: [PATCH 0563/1103] Add a NoPreferredStorageOrderBit flag for expression having no preferred storage order. It is currently only used in Product. --- Eigen/src/Core/DiagonalMatrix.h | 6 +++--- Eigen/src/Core/Product.h | 15 ++++++++++++++- Eigen/src/Core/util/Constants.h | 10 ++++++++++ Eigen/src/SparseCore/SparseAssign.h | 2 +- Eigen/src/SparseCore/SparseMatrix.h | 2 +- Eigen/src/SparseCore/SparseUtil.h | 2 +- 6 files changed, 30 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index fc9ecf561..7a0b736f1 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -30,7 +30,7 @@ class DiagonalBase : public EigenBase MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, IsVectorAtCompileTime = 0, - Flags = 0 + Flags = NoPreferredStorageOrderBit }; typedef Matrix DenseMatrixType; @@ -159,7 +159,7 @@ struct traits > typedef Dense StorageKind; typedef DenseIndex Index; enum { - Flags = LvalueBit + Flags = LvalueBit | NoPreferredStorageOrderBit }; }; } @@ -287,7 +287,7 @@ struct traits > ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, MaxRowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, MaxColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, - Flags = traits::Flags & LvalueBit + Flags = (traits::Flags & LvalueBit) | NoPreferredStorageOrderBit #ifndef EIGEN_TEST_EVALUATORS , CoeffReadCost = traits<_DiagonalVectorType>::CoeffReadCost diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 626b737c7..e381ac46a 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -82,7 +82,10 @@ struct traits > #endif // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator. - Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0) + Flags = ( MaxRowsAtCompileTime==1 + || ((LhsCleaned::Flags&NoPreferredStorageOrderBit) && (RhsCleaned::Flags&RowMajorBit)) + || ((RhsCleaned::Flags&NoPreferredStorageOrderBit) && (LhsCleaned::Flags&RowMajorBit)) ) + ? RowMajorBit : (MaxColsAtCompileTime==1 ? 0 : NoPreferredStorageOrderBit) }; }; @@ -156,6 +159,16 @@ public: } // namespace internal +#ifdef EIGEN_TEST_EVALUATORS +// Generic API dispatcher +template +class ProductImpl : public internal::generic_xpr_base, MatrixXpr, StorageKind>::type +{ + public: + typedef typename internal::generic_xpr_base, MatrixXpr, StorageKind>::type Base; +}; +#endif + template class ProductImpl : public internal::dense_product_base diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index aae8625d1..131f5d793 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -154,6 +154,16 @@ const unsigned int AlignedBit = 0x80; const unsigned int NestByRefBit = 0x100; +/** \ingroup flags + * + * for an expression, this means that the storage order + * can be either row-major or column-major. + * The precise choice will be decided at evaluation time or when + * combined with other expressions. + * \sa \ref RowMajorBit, \ref TopicStorageOrders */ +const unsigned int NoPreferredStorageOrderBit = 0x200; + + // list of flags that are inherited by default const unsigned int HereditaryBits = RowMajorBit | EvalBeforeNestingBit diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index 86a0b08c5..68a6ca20e 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -220,7 +220,7 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src) for (Index j=0; j& SparseMatrix::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - const bool needToTranspose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); + const bool needToTranspose = (Flags & RowMajorBit) != (internal::evaluator::Flags & RowMajorBit); if (needToTranspose) { // two passes algorithm: diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 2a67af291..0dfa15f7c 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -149,7 +149,7 @@ template struct plain_matrix_type { typedef typename traits::Scalar _Scalar; typedef typename traits::Index _Index; - enum { _Options = ((traits::Flags&RowMajorBit)==RowMajorBit) ? RowMajor : ColMajor }; + enum { _Options = ((evaluator::Flags&RowMajorBit)==RowMajorBit) ? RowMajor : ColMajor }; public: typedef SparseMatrix<_Scalar, _Options, _Index> type; }; From 73e686c6a44adfd54fe75a0d7581fd14bfa58f54 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 27 Jun 2014 15:54:44 +0200 Subject: [PATCH 0564/1103] Implement evaluators for sparse times diagonal products. --- Eigen/SparseCore | 2 +- Eigen/src/SparseCore/SparseDiagonalProduct.h | 126 ++++++++++++++++++- Eigen/src/SparseCore/SparseMatrixBase.h | 14 +++ 3 files changed, 139 insertions(+), 3 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 340ff7f52..b338950ca 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -48,6 +48,7 @@ struct Sparse {}; #include "src/SparseCore/SparseDot.h" #include "src/SparseCore/SparseRedux.h" #include "src/SparseCore/SparseView.h" +#include "src/SparseCore/SparseDiagonalProduct.h" #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" @@ -55,7 +56,6 @@ struct Sparse {}; #include "src/SparseCore/SparseSparseProductWithPruning.h" #include "src/SparseCore/SparseProduct.h" #include "src/SparseCore/SparseDenseProduct.h" -#include "src/SparseCore/SparseDiagonalProduct.h" #include "src/SparseCore/SparseTriangularView.h" #include "src/SparseCore/SparseSelfAdjointView.h" #include "src/SparseCore/TriangularSolver.h" diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index 1bb590e64..cf0f77342 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -24,8 +24,10 @@ namespace Eigen { // for that particular case // The two other cases are symmetric. +#ifndef EIGEN_TEST_EVALUATORS + namespace internal { - + template struct traits > { @@ -100,9 +102,14 @@ class SparseDiagonalProduct LhsNested m_lhs; RhsNested m_rhs; }; +#endif namespace internal { +#ifndef EIGEN_TEST_EVALUATORS + + + template class sparse_diagonal_product_inner_iterator_selector @@ -179,10 +186,124 @@ class sparse_diagonal_product_inner_iterator_selector inline Index row() const { return m_outer; } }; +#else // EIGEN_TEST_EVALUATORS +enum { + SDP_AsScalarProduct, + SDP_AsCwiseProduct +}; + +template +struct sparse_diagonal_product_evaluator; + +template +struct product_evaluator, ProductTag, DiagonalShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public sparse_diagonal_product_evaluator +{ + typedef Product XprType; + typedef evaluator type; + typedef evaluator nestedType; + enum { CoeffReadCost = Dynamic, Flags = Rhs::Flags&RowMajorBit }; // FIXME CoeffReadCost & Flags + + typedef sparse_diagonal_product_evaluator Base; + product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) {} +}; + +template +struct product_evaluator, ProductTag, SparseShape, DiagonalShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public sparse_diagonal_product_evaluator, Lhs::Flags&RowMajorBit?SDP_AsCwiseProduct:SDP_AsScalarProduct> +{ + typedef Product XprType; + typedef evaluator type; + typedef evaluator nestedType; + enum { CoeffReadCost = Dynamic, Flags = Lhs::Flags&RowMajorBit }; // FIXME CoeffReadCost & Flags + + typedef sparse_diagonal_product_evaluator, Lhs::Flags&RowMajorBit?SDP_AsCwiseProduct:SDP_AsScalarProduct> Base; + product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal()) {} +}; + +template +struct sparse_diagonal_product_evaluator +{ +protected: + typedef typename evaluator::InnerIterator SparseXprInnerIterator; + typedef typename SparseXprType::Scalar Scalar; + typedef typename SparseXprType::Index Index; + +public: + class InnerIterator : public SparseXprInnerIterator + { + public: + InnerIterator(const sparse_diagonal_product_evaluator &xprEval, Index outer) + : SparseXprInnerIterator(xprEval.m_sparseXprImpl, outer), + m_coeff(xprEval.m_diagCoeffImpl.coeff(outer)) + {} + + EIGEN_STRONG_INLINE Scalar value() const { return m_coeff * SparseXprInnerIterator::value(); } + protected: + typename DiagonalCoeffType::Scalar m_coeff; + }; + + sparse_diagonal_product_evaluator(const SparseXprType &sparseXpr, const DiagonalCoeffType &diagCoeff) + : m_sparseXprImpl(sparseXpr), m_diagCoeffImpl(diagCoeff) + {} + +protected: + typename evaluator::nestedType m_sparseXprImpl; + typename evaluator::nestedType m_diagCoeffImpl; +}; + + +template +struct sparse_diagonal_product_evaluator +{ + typedef typename SparseXprType::Scalar Scalar; + typedef typename SparseXprType::Index Index; + + typedef CwiseBinaryOp, + const typename SparseXprType::ConstInnerVectorReturnType, + const DiagCoeffType> CwiseProductType; + + typedef typename evaluator::type CwiseProductEval; + typedef typename evaluator::InnerIterator CwiseProductIterator; + + class InnerIterator : public CwiseProductIterator + { + public: + InnerIterator(const sparse_diagonal_product_evaluator &xprEval, Index outer) + : CwiseProductIterator(CwiseProductEval(xprEval.m_sparseXprNested.innerVector(outer).cwiseProduct(xprEval.m_diagCoeffNested)),0), + m_cwiseEval(xprEval.m_sparseXprNested.innerVector(outer).cwiseProduct(xprEval.m_diagCoeffNested)), + m_outer(outer) + { + ::new (static_cast(this)) CwiseProductIterator(m_cwiseEval,0); + } + + inline Index outer() const { return m_outer; } + inline Index col() const { return SparseXprType::IsRowMajor ? CwiseProductIterator::index() : m_outer; } + inline Index row() const { return SparseXprType::IsRowMajor ? m_outer : CwiseProductIterator::index(); } + + protected: + Index m_outer; + CwiseProductEval m_cwiseEval; + }; + + sparse_diagonal_product_evaluator(const SparseXprType &sparseXpr, const DiagCoeffType &diagCoeff) + : m_sparseXprNested(sparseXpr), m_diagCoeffNested(diagCoeff) + {} + +protected: + typename nested_eval::type m_sparseXprNested; + typename nested_eval::type m_diagCoeffNested; +}; + +#endif // EIGEN_TEST_EVALUATORS + + } // end namespace internal -// SparseMatrixBase functions +#ifndef EIGEN_TEST_EVALUATORS +// SparseMatrixBase functions template template const SparseDiagonalProduct @@ -190,6 +311,7 @@ SparseMatrixBase::operator*(const DiagonalBase &other) co { return SparseDiagonalProduct(this->derived(), other.derived()); } +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index 4c46008ae..c71244d3e 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -269,6 +269,7 @@ template class SparseMatrixBase : public EigenBase const typename SparseSparseProductReturnType::Type operator*(const SparseMatrixBase &other) const; +#ifndef EIGEN_TEST_EVALUATORS // sparse * diagonal template const SparseDiagonalProduct @@ -279,6 +280,19 @@ template class SparseMatrixBase : public EigenBase const SparseDiagonalProduct operator*(const DiagonalBase &lhs, const SparseMatrixBase& rhs) { return SparseDiagonalProduct(lhs.derived(), rhs.derived()); } +#else // EIGEN_TEST_EVALUATORS + // sparse * diagonal + template + const Product + operator*(const DiagonalBase &other) const + { return Product(derived(), other.derived()); } + + // diagonal * sparse + template friend + const Product + operator*(const DiagonalBase &lhs, const SparseMatrixBase& rhs) + { return Product(lhs.derived(), rhs.derived()); } +#endif // EIGEN_TEST_EVALUATORS /** dense * sparse (return a dense object unless it is an outer product) */ template friend From c401167712e632a2739ceed26233a510238fc5fc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 27 Jun 2014 16:41:45 +0200 Subject: [PATCH 0565/1103] Fix double constructions of the nested CwiseBinaryOp evaluator in sparse*diagonal product iterator. --- Eigen/src/SparseCore/SparseBlock.h | 3 +-- Eigen/src/SparseCore/SparseDiagonalProduct.h | 23 ++++++++++++-------- 2 files changed, 15 insertions(+), 11 deletions(-) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index f562b5479..3cbbfb185 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -607,7 +607,6 @@ struct unary_evaluator, IteratorBa protected: typedef typename evaluator::InnerIterator EvalIterator; - typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; typename evaluator::nestedType m_argImpl; const XprType &m_block; @@ -620,7 +619,7 @@ class unary_evaluator, IteratorBas const XprType& m_block; Index m_end; public: - + EIGEN_STRONG_INLINE InnerVectorInnerIterator(const unary_evaluator& aEval, Index outer) : EvalIterator(aEval.m_argImpl, outer + (IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol())), m_block(aEval.m_block), diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index cf0f77342..fd7fd1c24 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -266,24 +266,29 @@ struct sparse_diagonal_product_evaluator::type CwiseProductEval; typedef typename evaluator::InnerIterator CwiseProductIterator; - class InnerIterator : public CwiseProductIterator + class InnerIterator { public: InnerIterator(const sparse_diagonal_product_evaluator &xprEval, Index outer) - : CwiseProductIterator(CwiseProductEval(xprEval.m_sparseXprNested.innerVector(outer).cwiseProduct(xprEval.m_diagCoeffNested)),0), - m_cwiseEval(xprEval.m_sparseXprNested.innerVector(outer).cwiseProduct(xprEval.m_diagCoeffNested)), + : m_cwiseEval(xprEval.m_sparseXprNested.innerVector(outer).cwiseProduct(xprEval.m_diagCoeffNested)), + m_cwiseIter(m_cwiseEval, 0), m_outer(outer) - { - ::new (static_cast(this)) CwiseProductIterator(m_cwiseEval,0); - } + {} + inline Scalar value() const { return m_cwiseIter.value(); } + inline Index index() const { return m_cwiseIter.index(); } inline Index outer() const { return m_outer; } - inline Index col() const { return SparseXprType::IsRowMajor ? CwiseProductIterator::index() : m_outer; } - inline Index row() const { return SparseXprType::IsRowMajor ? m_outer : CwiseProductIterator::index(); } + inline Index col() const { return SparseXprType::IsRowMajor ? m_cwiseIter.index() : m_outer; } + inline Index row() const { return SparseXprType::IsRowMajor ? m_outer : m_cwiseIter.index(); } + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { ++m_cwiseIter; return *this; } + inline operator bool() const { return m_cwiseIter; } protected: - Index m_outer; CwiseProductEval m_cwiseEval; + CwiseProductIterator m_cwiseIter; + Index m_outer; }; sparse_diagonal_product_evaluator(const SparseXprType &sparseXpr, const DiagCoeffType &diagCoeff) From 75e574275c97f8b2ab53c792c9fd886f32013b77 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 10:24:46 +0200 Subject: [PATCH 0566/1103] Fix bug #836: extend SparseQR to support more columns than rows. --- Eigen/src/SparseQR/SparseQR.h | 114 ++++++++++++++++++---------------- test/sparseqr.cpp | 35 +++++++---- 2 files changed, 84 insertions(+), 65 deletions(-) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 5fb5bc203..4c6553bf2 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2012-2013 Desire Nuentsa -// Copyright (C) 2012-2013 Gael Guennebaud +// Copyright (C) 2012-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -180,7 +180,7 @@ class SparseQR y.bottomRows(y.rows()-rank).setZero(); // Apply the column permutation - if (m_perm_c.size()) dest.topRows(cols()) = colsPermutation() * y.topRows(cols()); + if (m_perm_c.size()) dest = colsPermutation() * y.topRows(cols()); else dest = y.topRows(cols()); m_info = Success; @@ -286,6 +286,7 @@ void SparseQR::analyzePattern(const MatrixType& mat) ord(mat, m_perm_c); Index n = mat.cols(); Index m = mat.rows(); + Index diagSize = (std::min)(m,n); if (!m_perm_c.size()) { @@ -297,13 +298,13 @@ void SparseQR::analyzePattern(const MatrixType& mat) m_outputPerm_c = m_perm_c.inverse(); internal::coletree(mat, m_etree, m_firstRowElt, m_outputPerm_c.indices().data()); - m_R.resize(n, n); - m_Q.resize(m, n); + m_R.resize(m, n); + m_Q.resize(m, diagSize); // Allocate space for nonzero elements : rough estimation m_R.reserve(2*mat.nonZeros()); //FIXME Get a more accurate estimation through symbolic factorization with the etree m_Q.reserve(2*mat.nonZeros()); - m_hcoeffs.resize(n); + m_hcoeffs.resize(diagSize); m_analysisIsok = true; } @@ -323,11 +324,12 @@ void SparseQR::factorize(const MatrixType& mat) eigen_assert(m_analysisIsok && "analyzePattern() should be called before this step"); Index m = mat.rows(); Index n = mat.cols(); - IndexVector mark(m); mark.setConstant(-1); // Record the visited nodes - IndexVector Ridx(n), Qidx(m); // Store temporarily the row indexes for the current column of R and Q - Index nzcolR, nzcolQ; // Number of nonzero for the current column of R and Q - ScalarVector tval(m); // The dense vector used to compute the current column - bool found_diag; + Index diagSize = (std::min)(m,n); + IndexVector mark((std::max)(m,n)); mark.setConstant(-1); // Record the visited nodes + IndexVector Ridx(n), Qidx(m); // Store temporarily the row indexes for the current column of R and Q + Index nzcolR, nzcolQ; // Number of nonzero for the current column of R and Q + ScalarVector tval(m); // The dense vector used to compute the current column + RealScalar pivotThreshold = m_threshold; m_pmat = mat; m_pmat.uncompress(); // To have the innerNonZeroPtr allocated @@ -339,7 +341,7 @@ void SparseQR::factorize(const MatrixType& mat) m_pmat.innerNonZeroPtr()[p] = mat.outerIndexPtr()[i+1] - mat.outerIndexPtr()[i]; } - /* Compute the default threshold, see : + /* Compute the default threshold as in MatLab, see: * Tim Davis, "Algorithm 915, SuiteSparseQR: Multifrontal Multithreaded Rank-Revealing * Sparse QR Factorization, ACM Trans. on Math. Soft. 38(1), 2011, Page 8:3 */ @@ -347,24 +349,24 @@ void SparseQR::factorize(const MatrixType& mat) { RealScalar max2Norm = 0.0; for (int j = 0; j < n; j++) max2Norm = (max)(max2Norm, m_pmat.col(j).norm()); - m_threshold = 20 * (m + n) * max2Norm * NumTraits::epsilon(); + pivotThreshold = 20 * (m + n) * max2Norm * NumTraits::epsilon(); } // Initialize the numerical permutation m_pivotperm.setIdentity(n); Index nonzeroCol = 0; // Record the number of valid pivots + m_Q.startVec(0); // Left looking rank-revealing QR factorization: compute a column of R and Q at a time - for (Index col = 0; col < (std::min)(n,m); ++col) + for (Index col = 0; col < n; ++col) { mark.setConstant(-1); m_R.startVec(col); - m_Q.startVec(col); mark(nonzeroCol) = col; Qidx(0) = nonzeroCol; nzcolR = 0; nzcolQ = 1; - found_diag = col>=m; + bool found_diag = nonzeroCol>=m; tval.setZero(); // Symbolic factorization: find the nonzero locations of the column k of the factors R and Q, i.e., @@ -373,7 +375,7 @@ void SparseQR::factorize(const MatrixType& mat) // thus the trick with found_diag that permits to do one more iteration on the diagonal element if this one has not been found. for (typename MatrixType::InnerIterator itp(m_pmat, col); itp || !found_diag; ++itp) { - Index curIdx = nonzeroCol ; + Index curIdx = nonzeroCol; if(itp) curIdx = itp.row(); if(curIdx == nonzeroCol) found_diag = true; @@ -415,7 +417,7 @@ void SparseQR::factorize(const MatrixType& mat) // Browse all the indexes of R(:,col) in reverse order for (Index i = nzcolR-1; i >= 0; i--) { - Index curIdx = m_pivotperm.indices()(Ridx(i)); + Index curIdx = Ridx(i); // Apply the curIdx-th householder vector to the current column (temporarily stored into tval) Scalar tdot(0); @@ -444,34 +446,37 @@ void SparseQR::factorize(const MatrixType& mat) } } } // End update current column - - // Compute the Householder reflection that eliminate the current column - // FIXME this step should call the Householder module. + Scalar tau; - RealScalar beta; - Scalar c0 = nzcolQ ? tval(Qidx(0)) : Scalar(0); + RealScalar beta = 0; - // First, the squared norm of Q((col+1):m, col) - RealScalar sqrNorm = 0.; - for (Index itq = 1; itq < nzcolQ; ++itq) sqrNorm += numext::abs2(tval(Qidx(itq))); - - if(sqrNorm == RealScalar(0) && numext::imag(c0) == RealScalar(0)) + if(nonzeroCol < diagSize) { - tau = RealScalar(0); - beta = numext::real(c0); - tval(Qidx(0)) = 1; - } - else - { - using std::sqrt; - beta = sqrt(numext::abs2(c0) + sqrNorm); - if(numext::real(c0) >= RealScalar(0)) - beta = -beta; - tval(Qidx(0)) = 1; - for (Index itq = 1; itq < nzcolQ; ++itq) - tval(Qidx(itq)) /= (c0 - beta); - tau = numext::conj((beta-c0) / beta); - + // Compute the Householder reflection that eliminate the current column + // FIXME this step should call the Householder module. + Scalar c0 = nzcolQ ? tval(Qidx(0)) : Scalar(0); + + // First, the squared norm of Q((col+1):m, col) + RealScalar sqrNorm = 0.; + for (Index itq = 1; itq < nzcolQ; ++itq) sqrNorm += numext::abs2(tval(Qidx(itq))); + if(sqrNorm == RealScalar(0) && numext::imag(c0) == RealScalar(0)) + { + tau = RealScalar(0); + beta = numext::real(c0); + tval(Qidx(0)) = 1; + } + else + { + using std::sqrt; + beta = sqrt(numext::abs2(c0) + sqrNorm); + if(numext::real(c0) >= RealScalar(0)) + beta = -beta; + tval(Qidx(0)) = 1; + for (Index itq = 1; itq < nzcolQ; ++itq) + tval(Qidx(itq)) /= (c0 - beta); + tau = numext::conj((beta-c0) / beta); + + } } // Insert values in R @@ -485,24 +490,25 @@ void SparseQR::factorize(const MatrixType& mat) } } - if(abs(beta) >= m_threshold) + if(nonzeroCol < diagSize && abs(beta) >= pivotThreshold) { m_R.insertBackByOuterInner(col, nonzeroCol) = beta; - nonzeroCol++; // The householder coefficient - m_hcoeffs(col) = tau; + m_hcoeffs(nonzeroCol) = tau; // Record the householder reflections for (Index itq = 0; itq < nzcolQ; ++itq) { Index iQ = Qidx(itq); - m_Q.insertBackByOuterInnerUnordered(col,iQ) = tval(iQ); + m_Q.insertBackByOuterInnerUnordered(nonzeroCol,iQ) = tval(iQ); tval(iQ) = Scalar(0.); - } + } + nonzeroCol++; + if(nonzeroCol::factorize(const MatrixType& mat) } } + m_hcoeffs.tail(diagSize-nonzeroCol).setZero(); + // Finalize the column pointers of the sparse matrices R and Q m_Q.finalize(); m_Q.makeCompressed(); @@ -579,14 +587,16 @@ struct SparseQR_QProduct : ReturnByValue void evalTo(DesType& res) const { + Index m = m_qr.rows(); Index n = m_qr.cols(); + Index diagSize = (std::min)(m,n); res = m_other; if (m_transpose) { eigen_assert(m_qr.m_Q.rows() == m_other.rows() && "Non conforming object sizes"); //Compute res = Q' * other column by column for(Index j = 0; j < res.cols(); j++){ - for (Index k = 0; k < n; k++) + for (Index k = 0; k < diagSize; k++) { Scalar tau = Scalar(0); tau = m_qr.m_Q.col(k).dot(res.col(j)); @@ -599,10 +609,10 @@ struct SparseQR_QProduct : ReturnByValue=0; k--) + for (Index k = diagSize-1; k >=0; k--) { Scalar tau = Scalar(0); tau = m_qr.m_Q.col(k).dot(res.col(j)); @@ -636,7 +646,7 @@ struct SparseQRMatrixQReturnType : public EigenBase(m_qr); } inline Index rows() const { return m_qr.rows(); } - inline Index cols() const { return m_qr.cols(); } + inline Index cols() const { return (std::min)(m_qr.rows(),m_qr.cols()); } // To use for operations with the transpose of Q SparseQRMatrixQTransposeReturnType transpose() const { diff --git a/test/sparseqr.cpp b/test/sparseqr.cpp index 6edba30b2..1fe4a98ee 100644 --- a/test/sparseqr.cpp +++ b/test/sparseqr.cpp @@ -2,24 +2,24 @@ // for linear algebra. // // Copyright (C) 2012 Desire Nuentsa Wakam +// Copyright (C) 2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed #include "sparse.h" #include - template -int generate_sparse_rectangular_problem(MatrixType& A, DenseMat& dA, int maxRows = 300, int maxCols = 300) +int generate_sparse_rectangular_problem(MatrixType& A, DenseMat& dA, int maxRows = 300, int maxCols = 150) { eigen_assert(maxRows >= maxCols); typedef typename MatrixType::Scalar Scalar; int rows = internal::random(1,maxRows); - int cols = internal::random(1,rows); + int cols = internal::random(1,maxCols); double density = (std::max)(8./(rows*cols), 0.01); - A.resize(rows,rows); - dA.resize(rows,rows); + A.resize(rows,cols); + dA.resize(rows,cols); initSparse(density, dA, A,ForceNonZeroDiag); A.makeCompressed(); int nop = internal::random(0, internal::random(0,1) > 0.5 ? cols/2 : 0); @@ -31,6 +31,13 @@ int generate_sparse_rectangular_problem(MatrixType& A, DenseMat& dA, int maxRows A.col(j0) = s * A.col(j1); dA.col(j0) = s * dA.col(j1); } + +// if(rows void test_sparseqr_scalar() MatrixType A; DenseMat dA; DenseVector refX,x,b; - SparseQR > solver; + SparseQR > solver; generate_sparse_rectangular_problem(A,dA); - int n = A.cols(); - b = DenseVector::Random(n); + b = dA * DenseVector::Random(A.cols()); solver.compute(A); if (solver.info() != Success) { @@ -60,17 +66,19 @@ template void test_sparseqr_scalar() std::cerr << "sparse QR factorization failed\n"; exit(0); return; - } + } + + VERIFY_IS_APPROX(A * x, b); + //Compare with a dense QR solver ColPivHouseholderQR dqr(dA); refX = dqr.solve(b); VERIFY_IS_EQUAL(dqr.rank(), solver.rank()); - - if(solver.rank() (A * x - b).norm() ); - else + if(solver.rank()==A.cols()) // full rank VERIFY_IS_APPROX(x, refX); +// else +// VERIFY((dA * refX - b).norm() * 2 > (A * x - b).norm() ); // Compute explicitly the matrix Q MatrixType Q, QtQ, idM; @@ -88,3 +96,4 @@ void test_sparseqr() CALL_SUBTEST_2(test_sparseqr_scalar >()); } } + From 7ffd55c98034964f735e7d6f11c01dcc2cf883ee Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 11:48:49 +0200 Subject: [PATCH 0567/1103] Do not bypass aliasing in sparse e assignments --- Eigen/src/SparseCore/SparseAssign.h | 4 ++-- Eigen/src/SparseCore/SparseMatrix.h | 19 +++++++++++++++---- 2 files changed, 17 insertions(+), 6 deletions(-) diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index 68a6ca20e..c99f4b74e 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -143,7 +143,7 @@ template template inline Derived& SparseMatrixBase::operator=(const SparseMatrixBase& other) { - internal::call_assignment_no_alias(derived(), other.derived()); + internal::call_assignment/*_no_alias*/(derived(), other.derived()); return derived(); } @@ -184,7 +184,7 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src) typedef typename internal::evaluator::type SrcEvaluatorType; SrcEvaluatorType srcEvaluator(src); - + const bool transpose = (DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit); const Index outerSize = (int(SrcEvaluatorType::Flags) & RowMajorBit) ? src.rows() : src.cols(); if ((!transpose) && src.isRValue()) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 56bf5aaf1..6df4da51a 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -649,7 +649,13 @@ class SparseMatrix EIGEN_STATIC_ASSERT((internal::is_same::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) check_template_parameters(); +#ifndef EIGEN_TEST_EVALUATORS *this = other.derived(); +#else + const bool needToTranspose = (Flags & RowMajorBit) != (internal::evaluator::Flags & RowMajorBit); + if (needToTranspose) *this = other.derived(); + else internal::call_assignment_no_alias(*this, other.derived()); +#endif } /** Constructs a sparse matrix from the sparse selfadjoint view \a other */ @@ -722,6 +728,7 @@ class SparseMatrix return *this; } +#ifndef EIGEN_TEST_EVALUATORS #ifndef EIGEN_PARSED_BY_DOXYGEN template inline SparseMatrix& operator=(const SparseSparseProduct& product) @@ -738,6 +745,7 @@ class SparseMatrix inline SparseMatrix& operator=(const EigenBase& other) { return Base::operator=(other.derived()); } #endif +#endif // EIGEN_TEST_EVALUATORS template EIGEN_DONT_INLINE SparseMatrix& operator=(const SparseMatrixBase& other); @@ -1174,7 +1182,9 @@ EIGEN_DONT_INLINE SparseMatrix& SparseMatrix > Flags = SparseMatrixType::Flags }; - evaluator(const SparseMatrixType &mat) : m_matrix(mat) {} + evaluator() : m_matrix(0) {} + evaluator(const SparseMatrixType &mat) : m_matrix(&mat) {} - operator SparseMatrixType&() { return m_matrix.const_cast_derived(); } - operator const SparseMatrixType&() const { return m_matrix; } + operator SparseMatrixType&() { return m_matrix->const_cast_derived(); } + operator const SparseMatrixType&() const { return *m_matrix; } - const SparseMatrixType &m_matrix; + const SparseMatrixType *m_matrix; }; } From 0ad7a644df5c71b9a75c0300210ce17985b88044 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 11:49:46 +0200 Subject: [PATCH 0568/1103] Implement nonZeros() for Transpose --- Eigen/src/SparseCore/SparseTranspose.h | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/Eigen/src/SparseCore/SparseTranspose.h b/Eigen/src/SparseCore/SparseTranspose.h index b20843dd3..f5eff6133 100644 --- a/Eigen/src/SparseCore/SparseTranspose.h +++ b/Eigen/src/SparseCore/SparseTranspose.h @@ -61,6 +61,17 @@ template class TransposeImpl::ReverseInn #else // EIGEN_TEST_EVALUATORS +// Implement nonZeros() for transpose. I'm not sure that's the best approach for that. +// Perhaps it should be implemented in Transpose<> itself. +template class TransposeImpl + : public SparseMatrixBase > +{ + protected: + typedef SparseMatrixBase > Base; + public: + inline typename MatrixType::Index nonZeros() const { return Base::derived().nestedExpression().nonZeros(); } +}; + namespace internal { template From 441f97b2df8465cb8d5c601e9f1ed324af71491e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 11:50:20 +0200 Subject: [PATCH 0569/1103] Implement evaluators for sparse * sparse products --- Eigen/SparseCore | 4 +- .../ConservativeSparseSparseProduct.h | 13 ++++ Eigen/src/SparseCore/SparseMatrixBase.h | 11 +++- Eigen/src/SparseCore/SparseProduct.h | 64 +++++++++++++++++++ 4 files changed, 88 insertions(+), 4 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index b338950ca..0c91c3b59 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -49,12 +49,12 @@ struct Sparse {}; #include "src/SparseCore/SparseRedux.h" #include "src/SparseCore/SparseView.h" #include "src/SparseCore/SparseDiagonalProduct.h" +#include "src/SparseCore/ConservativeSparseSparseProduct.h" +#include "src/SparseCore/SparseProduct.h" #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" -#include "src/SparseCore/ConservativeSparseSparseProduct.h" #include "src/SparseCore/SparseSparseProductWithPruning.h" -#include "src/SparseCore/SparseProduct.h" #include "src/SparseCore/SparseDenseProduct.h" #include "src/SparseCore/SparseTriangularView.h" #include "src/SparseCore/SparseSelfAdjointView.h" diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 5c320e2d2..193d71ca9 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -36,6 +36,11 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r // per column of the lhs. // Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs) Index estimated_nnz_prod = lhs.nonZeros() + rhs.nonZeros(); + +#ifdef EIGEN_TEST_EVALUATORS + typename evaluator::type lhsEval(lhs); + typename evaluator::type rhsEval(rhs); +#endif res.setZero(); res.reserve(Index(estimated_nnz_prod)); @@ -45,11 +50,19 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r res.startVec(j); Index nnz = 0; +#ifndef EIGEN_TEST_EVALUATORS for (typename Rhs::InnerIterator rhsIt(rhs, j); rhsIt; ++rhsIt) +#else + for (typename evaluator::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt) +#endif { Scalar y = rhsIt.value(); Index k = rhsIt.index(); +#ifndef EIGEN_TEST_EVALUATORS for (typename Lhs::InnerIterator lhsIt(lhs, k); lhsIt; ++lhsIt) +#else + for (typename evaluator::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt) +#endif { Index i = lhsIt.index(); Scalar x = lhsIt.value(); diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index c71244d3e..cebf3990d 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -190,8 +190,10 @@ template class SparseMatrixBase : public EigenBase public: +#ifndef EIGEN_TEST_EVALUATORS template inline Derived& operator=(const SparseSparseProduct& product); +#endif friend std::ostream & operator << (std::ostream & s, const SparseMatrixBase& m) { @@ -264,12 +266,12 @@ template class SparseMatrixBase : public EigenBase EIGEN_STRONG_INLINE const EIGEN_SPARSE_CWISE_PRODUCT_RETURN_TYPE cwiseProduct(const MatrixBase &other) const; +#ifndef EIGEN_TEST_EVALUATORS // sparse * sparse template const typename SparseSparseProductReturnType::Type operator*(const SparseMatrixBase &other) const; - -#ifndef EIGEN_TEST_EVALUATORS + // sparse * diagonal template const SparseDiagonalProduct @@ -292,6 +294,11 @@ template class SparseMatrixBase : public EigenBase const Product operator*(const DiagonalBase &lhs, const SparseMatrixBase& rhs) { return Product(lhs.derived(), rhs.derived()); } + + // sparse * sparse + template + const Product + operator*(const SparseMatrixBase &other) const; #endif // EIGEN_TEST_EVALUATORS /** dense * sparse (return a dense object unless it is an outer product) */ diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index cf7663070..52c452f92 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -12,6 +12,8 @@ namespace Eigen { +#ifndef EIGEN_TEST_EVALUATORS + template struct SparseSparseProductReturnType { @@ -183,6 +185,68 @@ SparseMatrixBase::operator*(const SparseMatrixBase &other return typename SparseSparseProductReturnType::Type(derived(), other.derived()); } +#else // EIGEN_TEST_EVALUATORS + + +/** \returns an expression of the product of two sparse matrices. + * By default a conservative product preserving the symbolic non zeros is performed. + * The automatic pruning of the small values can be achieved by calling the pruned() function + * in which case a totally different product algorithm is employed: + * \code + * C = (A*B).pruned(); // supress numerical zeros (exact) + * C = (A*B).pruned(ref); + * C = (A*B).pruned(ref,epsilon); + * \endcode + * where \c ref is a meaningful non zero reference value. + * */ +template +template +inline const Product +SparseMatrixBase::operator*(const SparseMatrixBase &other) const +{ + return Product(derived(), other.derived()); +} + +namespace internal { + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + typedef typename nested_eval::type LhsNested; + typedef typename nested_eval::type RhsNested; + LhsNested lhsNested(lhs); + RhsNested rhsNested(rhs); + internal::conservative_sparse_sparse_product_selector::type, + typename remove_all::type, Dest>::run(lhsNested,rhsNested,dst); + } +}; + +template +struct product_evaluator, ProductTag, SparseShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS + } // end namespace Eigen #endif // EIGEN_SPARSEPRODUCT_H From 746d2db6ed5d1bb104757f2170e2018eda524a12 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 13:18:56 +0200 Subject: [PATCH 0570/1103] Implement evaluators for sparse * sparse with auto pruning. --- Eigen/SparseCore | 2 +- Eigen/src/SparseCore/SparseMatrixBase.h | 3 + Eigen/src/SparseCore/SparseProduct.h | 30 +++++++++ .../SparseSparseProductWithPruning.h | 63 +++++++++++++++++++ Eigen/src/SparseCore/SparseView.h | 26 +++++++- 5 files changed, 120 insertions(+), 4 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 0c91c3b59..f74df3038 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -50,11 +50,11 @@ struct Sparse {}; #include "src/SparseCore/SparseView.h" #include "src/SparseCore/SparseDiagonalProduct.h" #include "src/SparseCore/ConservativeSparseSparseProduct.h" +#include "src/SparseCore/SparseSparseProductWithPruning.h" #include "src/SparseCore/SparseProduct.h" #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" -#include "src/SparseCore/SparseSparseProductWithPruning.h" #include "src/SparseCore/SparseDenseProduct.h" #include "src/SparseCore/SparseTriangularView.h" #include "src/SparseCore/SparseSelfAdjointView.h" diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index cebf3990d..3a81916fb 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -394,6 +394,9 @@ template class SparseMatrixBase : public EigenBase { return typename internal::eval::type(derived()); } Scalar sum() const; + + inline const SparseView + pruned(const Scalar& reference = Scalar(0), const RealScalar& epsilon = NumTraits::dummy_precision()) const; protected: diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index 52c452f92..8b9578836 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -242,7 +242,37 @@ struct product_evaluator, ProductTag, SparseSh protected: PlainObject m_result; }; + +template +struct evaluator > > + : public evaluator::PlainObject>::type +{ + typedef SparseView > XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + using std::abs; + ::new (static_cast(this)) Base(m_result); + typedef typename nested_eval::type LhsNested; + typedef typename nested_eval::type RhsNested; + LhsNested lhsNested(xpr.nestedExpression().lhs()); + RhsNested rhsNested(xpr.nestedExpression().rhs()); + + internal::sparse_sparse_product_with_pruning_selector::type, + typename remove_all::type, PlainObject>::run(lhsNested,rhsNested,m_result, + abs(xpr.reference())*xpr.epsilon()); + } + +protected: + PlainObject m_result; +}; + } // end namespace internal #endif // EIGEN_TEST_EVALUATORS diff --git a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h index fcc18f5c9..c33ec6bfd 100644 --- a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h +++ b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h @@ -46,6 +46,11 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r res.resize(cols, rows); else res.resize(rows, cols); + + #ifdef EIGEN_TEST_EVALUATORS + typename evaluator::type lhsEval(lhs); + typename evaluator::type rhsEval(rhs); + #endif res.reserve(estimated_nnz_prod); double ratioColRes = double(estimated_nnz_prod)/double(lhs.rows()*rhs.cols()); @@ -56,12 +61,20 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r // let's do a more accurate determination of the nnz ratio for the current column j of res tempVector.init(ratioColRes); tempVector.setZero(); +#ifndef EIGEN_TEST_EVALUATORS for (typename Rhs::InnerIterator rhsIt(rhs, j); rhsIt; ++rhsIt) +#else + for (typename evaluator::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt) +#endif { // FIXME should be written like this: tmp += rhsIt.value() * lhs.col(rhsIt.index()) tempVector.restart(); Scalar x = rhsIt.value(); +#ifndef EIGEN_TEST_EVALUATORS for (typename Lhs::InnerIterator lhsIt(lhs, rhsIt.index()); lhsIt; ++lhsIt) +#else + for (typename evaluator::InnerIterator lhsIt(lhsEval, rhsIt.index()); lhsIt; ++lhsIt) +#endif { tempVector.coeffRef(lhsIt.index()) += lhsIt.value() * x; } @@ -140,8 +153,58 @@ struct sparse_sparse_product_with_pruning_selector +struct sparse_sparse_product_with_pruning_selector +{ + typedef typename ResultType::RealScalar RealScalar; + static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance) + { + typedef SparseMatrix RowMajorMatrixLhs; + RowMajorMatrixLhs rowLhs(lhs); + sparse_sparse_product_with_pruning_selector(rowLhs,rhs,res,tolerance); + } +}; + +template +struct sparse_sparse_product_with_pruning_selector +{ + typedef typename ResultType::RealScalar RealScalar; + static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance) + { + typedef SparseMatrix RowMajorMatrixRhs; + RowMajorMatrixRhs rowRhs(rhs); + sparse_sparse_product_with_pruning_selector(lhs,rowRhs,res,tolerance); + } +}; + +template +struct sparse_sparse_product_with_pruning_selector +{ + typedef typename ResultType::RealScalar RealScalar; + static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance) + { + typedef SparseMatrix ColMajorMatrixRhs; + ColMajorMatrixRhs colRhs(rhs); + internal::sparse_sparse_product_with_pruning_impl(lhs, colRhs, res, tolerance); + } +}; + +template +struct sparse_sparse_product_with_pruning_selector +{ + typedef typename ResultType::RealScalar RealScalar; + static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance) + { + typedef SparseMatrix ColMajorMatrixLhs; + ColMajorMatrixLhs colLhs(lhs); + internal::sparse_sparse_product_with_pruning_impl(colLhs, rhs, res, tolerance); + } +}; +#endif } // end namespace internal diff --git a/Eigen/src/SparseCore/SparseView.h b/Eigen/src/SparseCore/SparseView.h index 96d0a849c..7bffbb9cd 100644 --- a/Eigen/src/SparseCore/SparseView.h +++ b/Eigen/src/SparseCore/SparseView.h @@ -233,10 +233,30 @@ struct unary_evaluator, IndexBased> #endif // EIGEN_TEST_EVALUATORS template -const SparseView MatrixBase::sparseView(const Scalar& m_reference, - const typename NumTraits::Real& m_epsilon) const +const SparseView MatrixBase::sparseView(const Scalar& reference, + const typename NumTraits::Real& epsilon) const { - return SparseView(derived(), m_reference, m_epsilon); + return SparseView(derived(), reference, epsilon); +} + +/** \returns an expression of \c *this with values smaller than + * \a reference * \a epsilon are removed. + * + * This method is typically used in conjunction with the product of two sparse matrices + * to automatically prune the smallest values as follows: + * \code + * C = (A*B).pruned(); // suppress numerical zeros (exact) + * C = (A*B).pruned(ref); + * C = (A*B).pruned(ref,epsilon); + * \endcode + * where \c ref is a meaningful non zero reference value. + * */ +template +const SparseView +SparseMatrixBase::pruned(const Scalar& reference, + const RealScalar& epsilon) const +{ + return SparseView(derived(), reference, epsilon); } } // end namespace Eigen From 3c63446507dbbc891e44b58af07f12f6ef58a175 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 13:27:35 +0200 Subject: [PATCH 0571/1103] Update copyright dates --- Eigen/src/SparseCore/ConservativeSparseSparseProduct.h | 2 +- Eigen/src/SparseCore/SparseCwiseBinaryOp.h | 2 +- Eigen/src/SparseCore/SparseDiagonalProduct.h | 2 +- Eigen/src/SparseCore/SparseProduct.h | 2 +- Eigen/src/SparseCore/SparseRedux.h | 2 +- Eigen/src/SparseCore/SparseSparseProductWithPruning.h | 2 +- Eigen/src/SparseCore/SparseUtil.h | 2 +- Eigen/src/SparseCore/SparseVector.h | 2 +- Eigen/src/SparseCore/SparseView.h | 2 +- 9 files changed, 9 insertions(+), 9 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 193d71ca9..f6824a895 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2011 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h index f43976641..dcb095dcf 100644 --- a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2014 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index fd7fd1c24..0998feba3 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index 8b9578836..4e181d471 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2010 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseRedux.h b/Eigen/src/SparseCore/SparseRedux.h index c6d0182d5..cf78d0e91 100644 --- a/Eigen/src/SparseCore/SparseRedux.h +++ b/Eigen/src/SparseCore/SparseRedux.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h index c33ec6bfd..b42b33e55 100644 --- a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h +++ b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2011 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 0dfa15f7c..847dfe575 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index 81f5eb41d..ed78d7c1e 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2009 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/Eigen/src/SparseCore/SparseView.h b/Eigen/src/SparseCore/SparseView.h index 7bffbb9cd..c1c50f6e9 100644 --- a/Eigen/src/SparseCore/SparseView.h +++ b/Eigen/src/SparseCore/SparseView.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // Copyright (C) 2010 Daniel Lowengrub // // This Source Code Form is subject to the terms of the Mozilla From 324e7e8fc9a20503d3f7ee9969c886400bbf5786 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 1 Jul 2014 16:58:11 +0200 Subject: [PATCH 0572/1103] Removed the deprecated EIGEN2_SUPPORT, as previously announced. A compilation error is raised, if this compile-switch is defined. The documentation references to the corresponding pages from Eigen3.2 now. Also, the Eigen2 testsuite has been removed. --- Eigen/Array | 11 - Eigen/Core | 35 +- Eigen/Eigen2Support | 82 -- Eigen/Geometry | 38 +- Eigen/LU | 4 - Eigen/LeastSquares | 32 - Eigen/QR | 8 - Eigen/SVD | 4 - Eigen/src/Cholesky/LDLT.h | 16 - Eigen/src/Cholesky/LLT.h | 11 - Eigen/src/Core/DenseBase.h | 11 - Eigen/src/Core/DenseCoeffsBase.h | 4 - Eigen/src/Core/DiagonalMatrix.h | 15 - Eigen/src/Core/Dot.h | 28 - Eigen/src/Core/Map.h | 5 - Eigen/src/Core/Matrix.h | 7 - Eigen/src/Core/MatrixBase.h | 78 -- Eigen/src/Core/NumTraits.h | 7 - Eigen/src/Core/ProductBase.h | 6 +- Eigen/src/Core/SelfAdjointView.h | 25 - Eigen/src/Core/TriangularMatrix.h | 64 -- Eigen/src/Core/VectorwiseOp.h | 2 - Eigen/src/Core/util/ForwardDeclarations.h | 35 - Eigen/src/Core/util/StaticAssert.h | 7 +- Eigen/src/Eigen2Support/Block.h | 126 --- Eigen/src/Eigen2Support/CMakeLists.txt | 8 - Eigen/src/Eigen2Support/Cwise.h | 192 ----- Eigen/src/Eigen2Support/CwiseOperators.h | 298 ------- Eigen/src/Eigen2Support/Geometry/AlignedBox.h | 159 ---- Eigen/src/Eigen2Support/Geometry/All.h | 115 --- Eigen/src/Eigen2Support/Geometry/AngleAxis.h | 228 ----- .../src/Eigen2Support/Geometry/CMakeLists.txt | 6 - Eigen/src/Eigen2Support/Geometry/Hyperplane.h | 254 ------ .../Eigen2Support/Geometry/ParametrizedLine.h | 141 ---- Eigen/src/Eigen2Support/Geometry/Quaternion.h | 495 ----------- Eigen/src/Eigen2Support/Geometry/Rotation2D.h | 145 ---- .../src/Eigen2Support/Geometry/RotationBase.h | 123 --- Eigen/src/Eigen2Support/Geometry/Scaling.h | 167 ---- Eigen/src/Eigen2Support/Geometry/Transform.h | 786 ------------------ .../src/Eigen2Support/Geometry/Translation.h | 184 ---- Eigen/src/Eigen2Support/LU.h | 120 --- Eigen/src/Eigen2Support/Lazy.h | 71 -- Eigen/src/Eigen2Support/LeastSquares.h | 170 ---- Eigen/src/Eigen2Support/Macros.h | 20 - Eigen/src/Eigen2Support/MathFunctions.h | 57 -- Eigen/src/Eigen2Support/Memory.h | 45 - Eigen/src/Eigen2Support/Meta.h | 75 -- Eigen/src/Eigen2Support/Minor.h | 117 --- Eigen/src/Eigen2Support/QR.h | 67 -- Eigen/src/Eigen2Support/SVD.h | 637 -------------- Eigen/src/Eigen2Support/TriangularSolver.h | 42 - Eigen/src/Eigen2Support/VectorBlock.h | 94 --- .../src/Eigenvalues/SelfAdjointEigenSolver.h | 34 - Eigen/src/LU/PartialPivLU.h | 3 - Eigen/src/SparseCore/SparseMatrixBase.h | 11 - Eigen/src/SparseCore/TriangularSolver.h | 24 - doc/A05_PortingFrom2To3.dox | 7 +- doc/A10_Eigen2SupportModes.dox | 93 --- doc/Doxyfile.in | 4 +- doc/PreprocessorDirectives.dox | 11 +- doc/examples/MatrixBase_cwise_const.cpp | 18 - test/CMakeLists.txt | 5 +- test/cwiseop.cpp | 182 ---- test/eigen2/CMakeLists.txt | 65 -- test/eigen2/eigen2_adjoint.cpp | 101 --- test/eigen2/eigen2_alignedbox.cpp | 60 -- test/eigen2/eigen2_array.cpp | 142 ---- test/eigen2/eigen2_basicstuff.cpp | 108 --- test/eigen2/eigen2_bug_132.cpp | 26 - test/eigen2/eigen2_cholesky.cpp | 113 --- test/eigen2/eigen2_commainitializer.cpp | 46 - test/eigen2/eigen2_cwiseop.cpp | 158 ---- test/eigen2/eigen2_determinant.cpp | 61 -- test/eigen2/eigen2_dynalloc.cpp | 131 --- test/eigen2/eigen2_eigensolver.cpp | 146 ---- test/eigen2/eigen2_first_aligned.cpp | 49 -- test/eigen2/eigen2_geometry.cpp | 431 ---------- .../eigen2_geometry_with_eigen2_prefix.cpp | 434 ---------- test/eigen2/eigen2_hyperplane.cpp | 126 --- test/eigen2/eigen2_inverse.cpp | 63 -- test/eigen2/eigen2_linearstructure.cpp | 84 -- test/eigen2/eigen2_lu.cpp | 122 --- test/eigen2/eigen2_map.cpp | 114 --- test/eigen2/eigen2_meta.cpp | 60 -- test/eigen2/eigen2_miscmatrices.cpp | 48 -- test/eigen2/eigen2_mixingtypes.cpp | 77 -- test/eigen2/eigen2_newstdvector.cpp | 149 ---- test/eigen2/eigen2_nomalloc.cpp | 63 -- test/eigen2/eigen2_packetmath.cpp | 132 --- test/eigen2/eigen2_parametrizedline.cpp | 62 -- test/eigen2/eigen2_prec_inverse_4x4.cpp | 84 -- test/eigen2/eigen2_product_large.cpp | 45 - test/eigen2/eigen2_product_small.cpp | 22 - test/eigen2/eigen2_qr.cpp | 69 -- test/eigen2/eigen2_qtvector.cpp | 158 ---- test/eigen2/eigen2_regression.cpp | 136 --- test/eigen2/eigen2_sizeof.cpp | 31 - test/eigen2/eigen2_smallvectors.cpp | 42 - test/eigen2/eigen2_sparse_basic.cpp | 317 ------- test/eigen2/eigen2_sparse_product.cpp | 115 --- test/eigen2/eigen2_sparse_solvers.cpp | 200 ----- test/eigen2/eigen2_sparse_vector.cpp | 84 -- test/eigen2/eigen2_stdvector.cpp | 148 ---- test/eigen2/eigen2_submatrices.cpp | 148 ---- test/eigen2/eigen2_sum.cpp | 71 -- test/eigen2/eigen2_svd.cpp | 87 -- test/eigen2/eigen2_swap.cpp | 83 -- test/eigen2/eigen2_triangular.cpp | 158 ---- test/eigen2/eigen2_unalignedassert.cpp | 116 --- test/eigen2/eigen2_visitor.cpp | 116 --- test/eigen2/gsl_helper.h | 175 ---- test/eigen2/main.h | 399 --------- test/eigen2/product.h | 132 --- test/eigen2/sparse.h | 154 ---- unsupported/Eigen/SVD | 4 - 115 files changed, 30 insertions(+), 12104 deletions(-) delete mode 100644 Eigen/Array delete mode 100644 Eigen/Eigen2Support delete mode 100644 Eigen/LeastSquares delete mode 100644 Eigen/src/Eigen2Support/Block.h delete mode 100644 Eigen/src/Eigen2Support/CMakeLists.txt delete mode 100644 Eigen/src/Eigen2Support/Cwise.h delete mode 100644 Eigen/src/Eigen2Support/CwiseOperators.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/AlignedBox.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/All.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/AngleAxis.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/CMakeLists.txt delete mode 100644 Eigen/src/Eigen2Support/Geometry/Hyperplane.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/ParametrizedLine.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/Quaternion.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/Rotation2D.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/RotationBase.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/Scaling.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/Transform.h delete mode 100644 Eigen/src/Eigen2Support/Geometry/Translation.h delete mode 100644 Eigen/src/Eigen2Support/LU.h delete mode 100644 Eigen/src/Eigen2Support/Lazy.h delete mode 100644 Eigen/src/Eigen2Support/LeastSquares.h delete mode 100644 Eigen/src/Eigen2Support/Macros.h delete mode 100644 Eigen/src/Eigen2Support/MathFunctions.h delete mode 100644 Eigen/src/Eigen2Support/Memory.h delete mode 100644 Eigen/src/Eigen2Support/Meta.h delete mode 100644 Eigen/src/Eigen2Support/Minor.h delete mode 100644 Eigen/src/Eigen2Support/QR.h delete mode 100644 Eigen/src/Eigen2Support/SVD.h delete mode 100644 Eigen/src/Eigen2Support/TriangularSolver.h delete mode 100644 Eigen/src/Eigen2Support/VectorBlock.h delete mode 100644 doc/A10_Eigen2SupportModes.dox delete mode 100644 doc/examples/MatrixBase_cwise_const.cpp delete mode 100644 test/cwiseop.cpp delete mode 100644 test/eigen2/CMakeLists.txt delete mode 100644 test/eigen2/eigen2_adjoint.cpp delete mode 100644 test/eigen2/eigen2_alignedbox.cpp delete mode 100644 test/eigen2/eigen2_array.cpp delete mode 100644 test/eigen2/eigen2_basicstuff.cpp delete mode 100644 test/eigen2/eigen2_bug_132.cpp delete mode 100644 test/eigen2/eigen2_cholesky.cpp delete mode 100644 test/eigen2/eigen2_commainitializer.cpp delete mode 100644 test/eigen2/eigen2_cwiseop.cpp delete mode 100644 test/eigen2/eigen2_determinant.cpp delete mode 100644 test/eigen2/eigen2_dynalloc.cpp delete mode 100644 test/eigen2/eigen2_eigensolver.cpp delete mode 100644 test/eigen2/eigen2_first_aligned.cpp delete mode 100644 test/eigen2/eigen2_geometry.cpp delete mode 100644 test/eigen2/eigen2_geometry_with_eigen2_prefix.cpp delete mode 100644 test/eigen2/eigen2_hyperplane.cpp delete mode 100644 test/eigen2/eigen2_inverse.cpp delete mode 100644 test/eigen2/eigen2_linearstructure.cpp delete mode 100644 test/eigen2/eigen2_lu.cpp delete mode 100644 test/eigen2/eigen2_map.cpp delete mode 100644 test/eigen2/eigen2_meta.cpp delete mode 100644 test/eigen2/eigen2_miscmatrices.cpp delete mode 100644 test/eigen2/eigen2_mixingtypes.cpp delete mode 100644 test/eigen2/eigen2_newstdvector.cpp delete mode 100644 test/eigen2/eigen2_nomalloc.cpp delete mode 100644 test/eigen2/eigen2_packetmath.cpp delete mode 100644 test/eigen2/eigen2_parametrizedline.cpp delete mode 100644 test/eigen2/eigen2_prec_inverse_4x4.cpp delete mode 100644 test/eigen2/eigen2_product_large.cpp delete mode 100644 test/eigen2/eigen2_product_small.cpp delete mode 100644 test/eigen2/eigen2_qr.cpp delete mode 100644 test/eigen2/eigen2_qtvector.cpp delete mode 100644 test/eigen2/eigen2_regression.cpp delete mode 100644 test/eigen2/eigen2_sizeof.cpp delete mode 100644 test/eigen2/eigen2_smallvectors.cpp delete mode 100644 test/eigen2/eigen2_sparse_basic.cpp delete mode 100644 test/eigen2/eigen2_sparse_product.cpp delete mode 100644 test/eigen2/eigen2_sparse_solvers.cpp delete mode 100644 test/eigen2/eigen2_sparse_vector.cpp delete mode 100644 test/eigen2/eigen2_stdvector.cpp delete mode 100644 test/eigen2/eigen2_submatrices.cpp delete mode 100644 test/eigen2/eigen2_sum.cpp delete mode 100644 test/eigen2/eigen2_svd.cpp delete mode 100644 test/eigen2/eigen2_swap.cpp delete mode 100644 test/eigen2/eigen2_triangular.cpp delete mode 100644 test/eigen2/eigen2_unalignedassert.cpp delete mode 100644 test/eigen2/eigen2_visitor.cpp delete mode 100644 test/eigen2/gsl_helper.h delete mode 100644 test/eigen2/main.h delete mode 100644 test/eigen2/product.h delete mode 100644 test/eigen2/sparse.h diff --git a/Eigen/Array b/Eigen/Array deleted file mode 100644 index 3d004fb69..000000000 --- a/Eigen/Array +++ /dev/null @@ -1,11 +0,0 @@ -#ifndef EIGEN_ARRAY_MODULE_H -#define EIGEN_ARRAY_MODULE_H - -// include Core first to handle Eigen2 support macros -#include "Core" - -#ifndef EIGEN2_SUPPORT - #error The Eigen/Array header does no longer exist in Eigen3. All that functionality has moved to Eigen/Core. -#endif - -#endif // EIGEN_ARRAY_MODULE_H diff --git a/Eigen/Core b/Eigen/Core index 661c7812e..a135afc6a 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -244,34 +244,9 @@ inline static const char *SimdInstructionSetsInUse(void) { } // end namespace Eigen -#define STAGE10_FULL_EIGEN2_API 10 -#define STAGE20_RESOLVE_API_CONFLICTS 20 -#define STAGE30_FULL_EIGEN3_API 30 -#define STAGE40_FULL_EIGEN3_STRICTNESS 40 -#define STAGE99_NO_EIGEN2_SUPPORT 99 - -#if defined EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS - #define EIGEN2_SUPPORT - #define EIGEN2_SUPPORT_STAGE STAGE40_FULL_EIGEN3_STRICTNESS -#elif defined EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API - #define EIGEN2_SUPPORT - #define EIGEN2_SUPPORT_STAGE STAGE30_FULL_EIGEN3_API -#elif defined EIGEN2_SUPPORT_STAGE20_RESOLVE_API_CONFLICTS - #define EIGEN2_SUPPORT - #define EIGEN2_SUPPORT_STAGE STAGE20_RESOLVE_API_CONFLICTS -#elif defined EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API - #define EIGEN2_SUPPORT - #define EIGEN2_SUPPORT_STAGE STAGE10_FULL_EIGEN2_API -#elif defined EIGEN2_SUPPORT - // default to stage 3, that's what it's always meant - #define EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API - #define EIGEN2_SUPPORT_STAGE STAGE30_FULL_EIGEN3_API -#else - #define EIGEN2_SUPPORT_STAGE STAGE99_NO_EIGEN2_SUPPORT -#endif - -#ifdef EIGEN2_SUPPORT -#undef minor +#if defined EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS || defined EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API || defined EIGEN2_SUPPORT_STAGE20_RESOLVE_API_CONFLICTS || defined EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API || defined EIGEN2_SUPPORT +// This will generate an error message: +#error Eigen2-support is only available up to version 3.2. Please go to "http://eigen.tuxfamily.org/index.php?title=Eigen2" for further information #endif // we use size_t frequently and we'll never remember to prepend it with std:: everytime just to @@ -429,8 +404,4 @@ using std::ptrdiff_t; #include "src/Core/util/ReenableStupidWarnings.h" -#ifdef EIGEN2_SUPPORT -#include "Eigen2Support" -#endif - #endif // EIGEN_CORE_H diff --git a/Eigen/Eigen2Support b/Eigen/Eigen2Support deleted file mode 100644 index 36156d29a..000000000 --- a/Eigen/Eigen2Support +++ /dev/null @@ -1,82 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2009 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2SUPPORT_H -#define EIGEN2SUPPORT_H - -#if (!defined(EIGEN2_SUPPORT)) || (!defined(EIGEN_CORE_H)) -#error Eigen2 support must be enabled by defining EIGEN2_SUPPORT before including any Eigen header -#endif - -#include "src/Core/util/DisableStupidWarnings.h" - -/** \ingroup Support_modules - * \defgroup Eigen2Support_Module Eigen2 support module - * This module provides a couple of deprecated functions improving the compatibility with Eigen2. - * - * To use it, define EIGEN2_SUPPORT before including any Eigen header - * \code - * #define EIGEN2_SUPPORT - * \endcode - * - */ - -#include "src/Eigen2Support/Macros.h" -#include "src/Eigen2Support/Memory.h" -#include "src/Eigen2Support/Meta.h" -#include "src/Eigen2Support/Lazy.h" -#include "src/Eigen2Support/Cwise.h" -#include "src/Eigen2Support/CwiseOperators.h" -#include "src/Eigen2Support/TriangularSolver.h" -#include "src/Eigen2Support/Block.h" -#include "src/Eigen2Support/VectorBlock.h" -#include "src/Eigen2Support/Minor.h" -#include "src/Eigen2Support/MathFunctions.h" - - -#include "src/Core/util/ReenableStupidWarnings.h" - -// Eigen2 used to include iostream -#include - -#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, SizeSuffix) \ -using Eigen::Matrix##SizeSuffix##TypeSuffix; \ -using Eigen::Vector##SizeSuffix##TypeSuffix; \ -using Eigen::RowVector##SizeSuffix##TypeSuffix; - -#define EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(TypeSuffix) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 2) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 3) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 4) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, X) \ - -#define EIGEN_USING_MATRIX_TYPEDEFS \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(i) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(f) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(d) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cf) \ -EIGEN_USING_MATRIX_TYPEDEFS_FOR_TYPE(cd) - -#define USING_PART_OF_NAMESPACE_EIGEN \ -EIGEN_USING_MATRIX_TYPEDEFS \ -using Eigen::Matrix; \ -using Eigen::MatrixBase; \ -using Eigen::ei_random; \ -using Eigen::ei_real; \ -using Eigen::ei_imag; \ -using Eigen::ei_conj; \ -using Eigen::ei_abs; \ -using Eigen::ei_abs2; \ -using Eigen::ei_sqrt; \ -using Eigen::ei_exp; \ -using Eigen::ei_log; \ -using Eigen::ei_sin; \ -using Eigen::ei_cos; - -#endif // EIGEN2SUPPORT_H diff --git a/Eigen/Geometry b/Eigen/Geometry index f9bc6fc57..1c642b7ee 100644 --- a/Eigen/Geometry +++ b/Eigen/Geometry @@ -33,29 +33,23 @@ #include "src/Geometry/OrthoMethods.h" #include "src/Geometry/EulerAngles.h" -#if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS - #include "src/Geometry/Homogeneous.h" - #include "src/Geometry/RotationBase.h" - #include "src/Geometry/Rotation2D.h" - #include "src/Geometry/Quaternion.h" - #include "src/Geometry/AngleAxis.h" - #include "src/Geometry/Transform.h" - #include "src/Geometry/Translation.h" - #include "src/Geometry/Scaling.h" - #include "src/Geometry/Hyperplane.h" - #include "src/Geometry/ParametrizedLine.h" - #include "src/Geometry/AlignedBox.h" - #include "src/Geometry/Umeyama.h" +#include "src/Geometry/Homogeneous.h" +#include "src/Geometry/RotationBase.h" +#include "src/Geometry/Rotation2D.h" +#include "src/Geometry/Quaternion.h" +#include "src/Geometry/AngleAxis.h" +#include "src/Geometry/Transform.h" +#include "src/Geometry/Translation.h" +#include "src/Geometry/Scaling.h" +#include "src/Geometry/Hyperplane.h" +#include "src/Geometry/ParametrizedLine.h" +#include "src/Geometry/AlignedBox.h" +#include "src/Geometry/Umeyama.h" - // Use the SSE optimized version whenever possible. At the moment the - // SSE version doesn't compile when AVX is enabled - #if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX - #include "src/Geometry/arch/Geometry_SSE.h" - #endif -#endif - -#ifdef EIGEN2_SUPPORT -#include "src/Eigen2Support/Geometry/All.h" +// Use the SSE optimized version whenever possible. At the moment the +// SSE version doesn't compile when AVX is enabled +#if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX +#include "src/Geometry/arch/Geometry_SSE.h" #endif #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/LU b/Eigen/LU index e5c3f32f7..29a98cb9a 100644 --- a/Eigen/LU +++ b/Eigen/LU @@ -33,10 +33,6 @@ #include "src/LU/arch/Inverse_SSE.h" #endif -#ifdef EIGEN2_SUPPORT - #include "src/Eigen2Support/LU.h" -#endif - #include "src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_LU_MODULE_H diff --git a/Eigen/LeastSquares b/Eigen/LeastSquares deleted file mode 100644 index 35137c25d..000000000 --- a/Eigen/LeastSquares +++ /dev/null @@ -1,32 +0,0 @@ -#ifndef EIGEN_REGRESSION_MODULE_H -#define EIGEN_REGRESSION_MODULE_H - -#ifndef EIGEN2_SUPPORT -#error LeastSquares is only available in Eigen2 support mode (define EIGEN2_SUPPORT) -#endif - -// exclude from normal eigen3-only documentation -#ifdef EIGEN2_SUPPORT - -#include "Core" - -#include "src/Core/util/DisableStupidWarnings.h" - -#include "Eigenvalues" -#include "Geometry" - -/** \defgroup LeastSquares_Module LeastSquares module - * This module provides linear regression and related features. - * - * \code - * #include - * \endcode - */ - -#include "src/Eigen2Support/LeastSquares.h" - -#include "src/Core/util/ReenableStupidWarnings.h" - -#endif // EIGEN2_SUPPORT - -#endif // EIGEN_REGRESSION_MODULE_H diff --git a/Eigen/QR b/Eigen/QR index 8c7c6162e..4c2533610 100644 --- a/Eigen/QR +++ b/Eigen/QR @@ -33,15 +33,7 @@ #include "src/QR/ColPivHouseholderQR_MKL.h" #endif -#ifdef EIGEN2_SUPPORT -#include "src/Eigen2Support/QR.h" -#endif - #include "src/Core/util/ReenableStupidWarnings.h" -#ifdef EIGEN2_SUPPORT -#include "Eigenvalues" -#endif - #endif // EIGEN_QR_MODULE_H /* vim: set filetype=cpp et sw=2 ts=2 ai: */ diff --git a/Eigen/SVD b/Eigen/SVD index fd310017a..5eee46df5 100644 --- a/Eigen/SVD +++ b/Eigen/SVD @@ -27,10 +27,6 @@ #endif #include "src/SVD/UpperBidiagonalization.h" -#ifdef EIGEN2_SUPPORT -#include "src/Eigen2Support/SVD.h" -#endif - #include "src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_SVD_MODULE_H diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index efac7fe40..1d84438e2 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -151,13 +151,6 @@ template class LDLT eigen_assert(m_isInitialized && "LDLT is not initialized."); return m_sign == internal::PositiveSemiDef || m_sign == internal::ZeroSign; } - - #ifdef EIGEN2_SUPPORT - inline bool isPositiveDefinite() const - { - return isPositive(); - } - #endif /** \returns true if the matrix is negative (semidefinite) */ inline bool isNegative(void) const @@ -191,15 +184,6 @@ template class LDLT return internal::solve_retval(*this, b.derived()); } - #ifdef EIGEN2_SUPPORT - template - bool solve(const MatrixBase& b, ResultType *result) const - { - *result = this->solve(b); - return true; - } - #endif - template bool solveInPlace(MatrixBase &bAndX) const; diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index 45ed8438f..38e820165 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -127,17 +127,6 @@ template class LLT return internal::solve_retval(*this, b.derived()); } - #ifdef EIGEN2_SUPPORT - template - bool solve(const MatrixBase& b, ResultType *result) const - { - *result = this->solve(b); - return true; - } - - bool isPositiveDefinite() const { return true; } - #endif - template void solveInPlace(MatrixBase &bAndX) const; diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 4794c2f13..bd5dd14ed 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -506,17 +506,6 @@ template class DenseBase # endif #undef EIGEN_CURRENT_STORAGE_BASE_CLASS -#ifdef EIGEN2_SUPPORT - - Block corner(CornerType type, Index cRows, Index cCols); - const Block corner(CornerType type, Index cRows, Index cCols) const; - template - Block corner(CornerType type); - template - const Block corner(CornerType type) const; - -#endif // EIGEN2_SUPPORT - // disable the use of evalTo for dense objects with a nice compilation error template diff --git a/Eigen/src/Core/DenseCoeffsBase.h b/Eigen/src/Core/DenseCoeffsBase.h index efabb5e67..4e986e875 100644 --- a/Eigen/src/Core/DenseCoeffsBase.h +++ b/Eigen/src/Core/DenseCoeffsBase.h @@ -156,10 +156,8 @@ class DenseCoeffsBase : public EigenBase EIGEN_STRONG_INLINE CoeffReturnType operator[](Index index) const { - #ifndef EIGEN2_SUPPORT EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime, THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD) - #endif eigen_assert(index >= 0 && index < size()); return derived().coeff(index); } @@ -388,10 +386,8 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && index < size()); return derived().coeffRef(index); } diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index f7ac22f8b..96b65483d 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -95,21 +95,6 @@ class DiagonalBase : public EigenBase { return other.diagonal() * scalar; } - - #ifdef EIGEN2_SUPPORT - template - EIGEN_DEVICE_FUNC - bool isApprox(const DiagonalBase& other, typename NumTraits::Real precision = NumTraits::dummy_precision()) const - { - return diagonal().isApprox(other.diagonal(), precision); - } - template - EIGEN_DEVICE_FUNC - bool isApprox(const MatrixBase& other, typename NumTraits::Real precision = NumTraits::dummy_precision()) const - { - return toDenseMatrix().isApprox(other, precision); - } - #endif }; template diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h index 718de5d1a..db16e4acc 100644 --- a/Eigen/src/Core/Dot.h +++ b/Eigen/src/Core/Dot.h @@ -76,34 +76,6 @@ MatrixBase::dot(const MatrixBase& other) const return internal::dot_nocheck::run(*this, other); } -#ifdef EIGEN2_SUPPORT -/** \returns the dot product of *this with other, with the Eigen2 convention that the dot product is linear in the first variable - * (conjugating the second variable). Of course this only makes a difference in the complex case. - * - * This method is only available in EIGEN2_SUPPORT mode. - * - * \only_for_vectors - * - * \sa dot() - */ -template -template -typename internal::traits::Scalar -MatrixBase::eigen2_dot(const MatrixBase& other) const -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived) - EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived) - EIGEN_STATIC_ASSERT((internal::is_same::value), - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - - eigen_assert(size() == other.size()); - - return internal::dot_nocheck::run(other,*this); -} -#endif - - //---------- implementation of L2 norm and related functions ---------- /** \returns, for vectors, the squared \em l2 norm of \c *this, and for matrices the Frobenius norm. diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index 0838d69e3..ced1b76ba 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -110,14 +110,9 @@ template class Ma EIGEN_DENSE_PUBLIC_INTERFACE(Map) typedef typename Base::PointerType PointerType; -#if EIGEN2_SUPPORT_STAGE <= STAGE30_FULL_EIGEN3_API - typedef const Scalar* PointerArgType; - inline PointerType cast_to_pointer_type(PointerArgType ptr) { return const_cast(ptr); } -#else typedef PointerType PointerArgType; EIGEN_DEVICE_FUNC inline PointerType cast_to_pointer_type(PointerArgType ptr) { return ptr; } -#endif EIGEN_DEVICE_FUNC inline Index innerStride() const diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 782d67f54..cd2ff91e7 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -366,13 +366,6 @@ class Matrix EIGEN_DEVICE_FUNC Matrix& operator=(const RotationBase& r); - #ifdef EIGEN2_SUPPORT - template - explicit Matrix(const eigen2_RotationBase& r); - template - Matrix& operator=(const eigen2_RotationBase& r); - #endif - // allow to extend Matrix outside Eigen #ifdef EIGEN_MATRIX_PLUGIN #include EIGEN_MATRIX_PLUGIN diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 172929562..f5987d194 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -221,11 +221,6 @@ template class MatrixBase typename internal::scalar_product_traits::Scalar,typename internal::traits::Scalar>::ReturnType dot(const MatrixBase& other) const; - #ifdef EIGEN2_SUPPORT - template - Scalar eigen2_dot(const MatrixBase& other) const; - #endif - EIGEN_DEVICE_FUNC RealScalar squaredNorm() const; EIGEN_DEVICE_FUNC RealScalar norm() const; RealScalar stableNorm() const; @@ -269,17 +264,6 @@ template class MatrixBase typename ConstDiagonalIndexReturnType::Type diagonal(Index index) const; #endif - #ifdef EIGEN2_SUPPORT - template typename internal::eigen2_part_return_type::type part(); - template const typename internal::eigen2_part_return_type::type part() const; - - // huuuge hack. make Eigen2's matrix.part() work in eigen3. Problem: Diagonal is now a class template instead - // of an integer constant. Solution: overload the part() method template wrt template parameters list. - template class U> - const DiagonalWrapper part() const - { return diagonal().asDiagonal(); } - #endif // EIGEN2_SUPPORT - template struct TriangularViewReturnType { typedef TriangularView Type; }; template struct ConstTriangularViewReturnType { typedef const TriangularView Type; }; @@ -373,24 +357,7 @@ template class MatrixBase EIGEN_DEVICE_FUNC const FullPivLU fullPivLu() const; EIGEN_DEVICE_FUNC const PartialPivLU partialPivLu() const; - #if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS - const LU lu() const; - #endif - - #ifdef EIGEN2_SUPPORT - const LU eigen2_lu() const; - #endif - - #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS const PartialPivLU lu() const; - #endif - - #ifdef EIGEN2_SUPPORT - template - void computeInverse(MatrixBase *result) const { - *result = this->inverse(); - } - #endif EIGEN_DEVICE_FUNC const internal::inverse_impl inverse() const; @@ -419,10 +386,6 @@ template class MatrixBase const HouseholderQR householderQr() const; const ColPivHouseholderQR colPivHouseholderQr() const; const FullPivHouseholderQR fullPivHouseholderQr() const; - - #ifdef EIGEN2_SUPPORT - const QR qr() const; - #endif EigenvaluesReturnType eigenvalues() const; RealScalar operatorNorm() const; @@ -431,10 +394,6 @@ template class MatrixBase JacobiSVD jacobiSvd(unsigned int computationOptions = 0) const; - #ifdef EIGEN2_SUPPORT - SVD svd() const; - #endif - /////////// Geometry module /////////// #ifndef EIGEN_PARSED_BY_DOXYGEN @@ -458,13 +417,11 @@ template class MatrixBase Matrix eulerAngles(Index a0, Index a1, Index a2) const; - #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS ScalarMultipleReturnType operator*(const UniformScaling& s) const; // put this as separate enum value to work around possible GCC 4.3 bug (?) enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1?Vertical:Horizontal }; typedef Homogeneous HomogeneousReturnType; HomogeneousReturnType homogeneous() const; - #endif enum { SizeMinusOne = SizeAtCompileTime==Dynamic ? Dynamic : SizeAtCompileTime-1 @@ -513,41 +470,6 @@ template class MatrixBase const MatrixPowerReturnValue pow(const RealScalar& p) const; const MatrixComplexPowerReturnValue pow(const std::complex& p) const; -#ifdef EIGEN2_SUPPORT - template - Derived& operator+=(const Flagged, 0, - EvalBeforeAssigningBit>& other); - - template - Derived& operator-=(const Flagged, 0, - EvalBeforeAssigningBit>& other); - - /** \deprecated because .lazy() is deprecated - * Overloaded for cache friendly product evaluation */ - template - Derived& lazyAssign(const Flagged& other) - { return lazyAssign(other._expression()); } - - template - const Flagged marked() const; - const Flagged lazy() const; - - inline const Cwise cwise() const; - inline Cwise cwise(); - - VectorBlock start(Index size); - const VectorBlock start(Index size) const; - VectorBlock end(Index size); - const VectorBlock end(Index size) const; - template VectorBlock start(); - template const VectorBlock start() const; - template VectorBlock end(); - template const VectorBlock end() const; - - Minor minor(Index row, Index col); - const Minor minor(Index row, Index col) const; -#endif - protected: EIGEN_DEVICE_FUNC MatrixBase() : Base() {} diff --git a/Eigen/src/Core/NumTraits.h b/Eigen/src/Core/NumTraits.h index 2b6633c9c..a04227f57 100644 --- a/Eigen/src/Core/NumTraits.h +++ b/Eigen/src/Core/NumTraits.h @@ -85,13 +85,6 @@ template struct GenericNumTraits } static inline T highest() { return (std::numeric_limits::max)(); } static inline T lowest() { return IsInteger ? (std::numeric_limits::min)() : (-(std::numeric_limits::max)()); } - -#ifdef EIGEN2_SUPPORT - enum { - HasFloatingPoint = !IsInteger - }; - typedef NonInteger FloatingPoint; -#endif }; template struct NumTraits : GenericNumTraits diff --git a/Eigen/src/Core/ProductBase.h b/Eigen/src/Core/ProductBase.h index a494b5f87..483914a9b 100644 --- a/Eigen/src/Core/ProductBase.h +++ b/Eigen/src/Core/ProductBase.h @@ -131,17 +131,13 @@ class ProductBase : public MatrixBase const Diagonal diagonal(Index index) const { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs).diagonal(index); } - // restrict coeff accessors to 1x1 expressions. No need to care about mutators here since this isnt a Lvalue expression + // restrict coeff accessors to 1x1 expressions. No need to care about mutators here since this isn't an Lvalue expression typename Base::CoeffReturnType coeff(Index row, Index col) const { -#ifdef EIGEN2_SUPPORT - return lhs().row(row).cwiseProduct(rhs().col(col).transpose()).sum(); -#else EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) eigen_assert(this->rows() == 1 && this->cols() == 1); Matrix result = *this; return result.coeff(row,col); -#endif } typename Base::CoeffReturnType coeff(Index i) const diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 8231e3f5c..6c2733650 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -177,31 +177,6 @@ template class SelfAdjointView EigenvaluesReturnType eigenvalues() const; EIGEN_DEVICE_FUNC RealScalar operatorNorm() const; - - #ifdef EIGEN2_SUPPORT - template - EIGEN_DEVICE_FUNC - SelfAdjointView& operator=(const MatrixBase& other) - { - enum { - OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper - }; - m_matrix.const_cast_derived().template triangularView() = other; - m_matrix.const_cast_derived().template triangularView() = other.adjoint(); - return *this; - } - template - EIGEN_DEVICE_FUNC - SelfAdjointView& operator=(const TriangularView& other) - { - enum { - OtherPart = UpLo == Upper ? StrictlyLower : StrictlyUpper - }; - m_matrix.const_cast_derived().template triangularView() = other.toDenseMatrix(); - m_matrix.const_cast_derived().template triangularView() = other.toDenseMatrix().adjoint(); - return *this; - } - #endif protected: MatrixTypeNested m_matrix; diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 1d6e34650..cdd341965 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -342,35 +342,6 @@ template class TriangularView (lhs.derived(),rhs.m_matrix); } - #ifdef EIGEN2_SUPPORT - template - struct eigen2_product_return_type - { - typedef typename TriangularView::DenseMatrixType DenseMatrixType; - typedef typename OtherDerived::PlainObject::DenseType OtherPlainObject; - typedef typename ProductReturnType::Type ProdRetType; - typedef typename ProdRetType::PlainObject type; - }; - template - const typename eigen2_product_return_type::type - operator*(const EigenBase& rhs) const - { - typename OtherDerived::PlainObject::DenseType rhsPlainObject; - rhs.evalTo(rhsPlainObject); - return this->toDenseMatrix() * rhsPlainObject; - } - template - bool isApprox(const TriangularView& other, typename NumTraits::Real precision = NumTraits::dummy_precision()) const - { - return this->toDenseMatrix().isApprox(other.toDenseMatrix(), precision); - } - template - bool isApprox(const MatrixBase& other, typename NumTraits::Real precision = NumTraits::dummy_precision()) const - { - return this->toDenseMatrix().isApprox(other, precision); - } - #endif // EIGEN2_SUPPORT - template EIGEN_DEVICE_FUNC inline const internal::triangular_solve_retval @@ -780,41 +751,6 @@ void TriangularBase::evalToLazy(MatrixBase &other) const * Implementation of MatrixBase methods ***************************************************************************/ -#ifdef EIGEN2_SUPPORT - -// implementation of part<>(), including the SelfAdjoint case. - -namespace internal { -template -struct eigen2_part_return_type -{ - typedef TriangularView type; -}; - -template -struct eigen2_part_return_type -{ - typedef SelfAdjointView type; -}; -} - -/** \deprecated use MatrixBase::triangularView() */ -template -template -const typename internal::eigen2_part_return_type::type MatrixBase::part() const -{ - return derived(); -} - -/** \deprecated use MatrixBase::triangularView() */ -template -template -typename internal::eigen2_part_return_type::type MatrixBase::part() -{ - return derived(); -} -#endif - /** * \returns an expression of a triangular view extracted from the current matrix * diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index f25ddca17..52eb4f604 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -560,9 +560,7 @@ template class VectorwiseOp /////////// Geometry module /////////// - #if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS Homogeneous homogeneous() const; - #endif typedef typename ExpressionType::PlainObject CrossReturnType; template diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 0a2144c69..33deb88ec 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -236,35 +236,12 @@ template class Rotation2D; template class AngleAxis; template class Translation; -#ifdef EIGEN2_SUPPORT -template class eigen2_RotationBase; -template class eigen2_Cross; -template class eigen2_Quaternion; -template class eigen2_Rotation2D; -template class eigen2_AngleAxis; -template class eigen2_Transform; -template class eigen2_ParametrizedLine; -template class eigen2_Hyperplane; -template class eigen2_Translation; -template class eigen2_Scaling; -#endif - -#if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS -template class Quaternion; -template class Transform; -template class ParametrizedLine; -template class Hyperplane; -template class Scaling; -#endif - -#if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS template class Quaternion; template class Transform; template class ParametrizedLine; template class Hyperplane; template class UniformScaling; template class Homogeneous; -#endif // MatrixFunctions module template struct MatrixExponentialReturnValue; @@ -283,18 +260,6 @@ struct stem_function }; } - -#ifdef EIGEN2_SUPPORT -template class Cwise; -template class Minor; -template class LU; -template class QR; -template class SVD; -namespace internal { -template struct eigen2_part_return_type; -} -#endif - } // end namespace Eigen #endif // EIGEN_FORWARDDECLARATIONS_H diff --git a/Eigen/src/Core/util/StaticAssert.h b/Eigen/src/Core/util/StaticAssert.h index ac2c0bedd..59aa0811c 100644 --- a/Eigen/src/Core/util/StaticAssert.h +++ b/Eigen/src/Core/util/StaticAssert.h @@ -168,13 +168,8 @@ ) \ ) -#ifdef EIGEN2_SUPPORT - #define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \ - eigen_assert(!NumTraits::IsInteger); -#else - #define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \ +#define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \ EIGEN_STATIC_ASSERT(!NumTraits::IsInteger, THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES) -#endif // static assertion failing if it is guaranteed at compile-time that the two matrix expression types have different sizes diff --git a/Eigen/src/Eigen2Support/Block.h b/Eigen/src/Eigen2Support/Block.h deleted file mode 100644 index 604456f40..000000000 --- a/Eigen/src/Eigen2Support/Block.h +++ /dev/null @@ -1,126 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_BLOCK2_H -#define EIGEN_BLOCK2_H - -namespace Eigen { - -/** \returns a dynamic-size expression of a corner of *this. - * - * \param type the type of corner. Can be \a Eigen::TopLeft, \a Eigen::TopRight, - * \a Eigen::BottomLeft, \a Eigen::BottomRight. - * \param cRows the number of rows in the corner - * \param cCols the number of columns in the corner - * - * Example: \include MatrixBase_corner_enum_int_int.cpp - * Output: \verbinclude MatrixBase_corner_enum_int_int.out - * - * \note Even though the returned expression has dynamic size, in the case - * when it is applied to a fixed-size matrix, it inherits a fixed maximal size, - * which means that evaluating it does not cause a dynamic memory allocation. - * - * \sa class Block, block(Index,Index,Index,Index) - */ -template -inline Block DenseBase - ::corner(CornerType type, Index cRows, Index cCols) -{ - switch(type) - { - default: - eigen_assert(false && "Bad corner type."); - case TopLeft: - return Block(derived(), 0, 0, cRows, cCols); - case TopRight: - return Block(derived(), 0, cols() - cCols, cRows, cCols); - case BottomLeft: - return Block(derived(), rows() - cRows, 0, cRows, cCols); - case BottomRight: - return Block(derived(), rows() - cRows, cols() - cCols, cRows, cCols); - } -} - -/** This is the const version of corner(CornerType, Index, Index).*/ -template -inline const Block -DenseBase::corner(CornerType type, Index cRows, Index cCols) const -{ - switch(type) - { - default: - eigen_assert(false && "Bad corner type."); - case TopLeft: - return Block(derived(), 0, 0, cRows, cCols); - case TopRight: - return Block(derived(), 0, cols() - cCols, cRows, cCols); - case BottomLeft: - return Block(derived(), rows() - cRows, 0, cRows, cCols); - case BottomRight: - return Block(derived(), rows() - cRows, cols() - cCols, cRows, cCols); - } -} - -/** \returns a fixed-size expression of a corner of *this. - * - * \param type the type of corner. Can be \a Eigen::TopLeft, \a Eigen::TopRight, - * \a Eigen::BottomLeft, \a Eigen::BottomRight. - * - * The template parameters CRows and CCols arethe number of rows and columns in the corner. - * - * Example: \include MatrixBase_template_int_int_corner_enum.cpp - * Output: \verbinclude MatrixBase_template_int_int_corner_enum.out - * - * \sa class Block, block(Index,Index,Index,Index) - */ -template -template -inline Block -DenseBase::corner(CornerType type) -{ - switch(type) - { - default: - eigen_assert(false && "Bad corner type."); - case TopLeft: - return Block(derived(), 0, 0); - case TopRight: - return Block(derived(), 0, cols() - CCols); - case BottomLeft: - return Block(derived(), rows() - CRows, 0); - case BottomRight: - return Block(derived(), rows() - CRows, cols() - CCols); - } -} - -/** This is the const version of corner(CornerType).*/ -template -template -inline const Block -DenseBase::corner(CornerType type) const -{ - switch(type) - { - default: - eigen_assert(false && "Bad corner type."); - case TopLeft: - return Block(derived(), 0, 0); - case TopRight: - return Block(derived(), 0, cols() - CCols); - case BottomLeft: - return Block(derived(), rows() - CRows, 0); - case BottomRight: - return Block(derived(), rows() - CRows, cols() - CCols); - } -} - -} // end namespace Eigen - -#endif // EIGEN_BLOCK2_H diff --git a/Eigen/src/Eigen2Support/CMakeLists.txt b/Eigen/src/Eigen2Support/CMakeLists.txt deleted file mode 100644 index 7ae41b3cb..000000000 --- a/Eigen/src/Eigen2Support/CMakeLists.txt +++ /dev/null @@ -1,8 +0,0 @@ -FILE(GLOB Eigen_Eigen2Support_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Eigen2Support_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Eigen2Support COMPONENT Devel - ) - -ADD_SUBDIRECTORY(Geometry) \ No newline at end of file diff --git a/Eigen/src/Eigen2Support/Cwise.h b/Eigen/src/Eigen2Support/Cwise.h deleted file mode 100644 index d95009b6e..000000000 --- a/Eigen/src/Eigen2Support/Cwise.h +++ /dev/null @@ -1,192 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_CWISE_H -#define EIGEN_CWISE_H - -namespace Eigen { - -/** \internal - * convenient macro to defined the return type of a cwise binary operation */ -#define EIGEN_CWISE_BINOP_RETURN_TYPE(OP) \ - CwiseBinaryOp::Scalar>, ExpressionType, OtherDerived> - -/** \internal - * convenient macro to defined the return type of a cwise unary operation */ -#define EIGEN_CWISE_UNOP_RETURN_TYPE(OP) \ - CwiseUnaryOp::Scalar>, ExpressionType> - -/** \internal - * convenient macro to defined the return type of a cwise comparison to a scalar */ -#define EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(OP) \ - CwiseBinaryOp::Scalar>, ExpressionType, \ - typename ExpressionType::ConstantReturnType > - -/** \class Cwise - * - * \brief Pseudo expression providing additional coefficient-wise operations - * - * \param ExpressionType the type of the object on which to do coefficient-wise operations - * - * This class represents an expression with additional coefficient-wise features. - * It is the return type of MatrixBase::cwise() - * and most of the time this is the only way it is used. - * - * Example: \include MatrixBase_cwise_const.cpp - * Output: \verbinclude MatrixBase_cwise_const.out - * - * This class can be extended with the help of the plugin mechanism described on the page - * \ref TopicCustomizingEigen by defining the preprocessor symbol \c EIGEN_CWISE_PLUGIN. - * - * \sa MatrixBase::cwise() const, MatrixBase::cwise() - */ -template class Cwise -{ - public: - - typedef typename internal::traits::Scalar Scalar; - typedef typename internal::conditional::ret, - ExpressionType, const ExpressionType&>::type ExpressionTypeNested; - typedef CwiseUnaryOp, ExpressionType> ScalarAddReturnType; - - inline Cwise(const ExpressionType& matrix) : m_matrix(matrix) {} - - /** \internal */ - inline const ExpressionType& _expression() const { return m_matrix; } - - template - const EIGEN_CWISE_PRODUCT_RETURN_TYPE(ExpressionType,OtherDerived) - operator*(const MatrixBase &other) const; - - template - const EIGEN_CWISE_BINOP_RETURN_TYPE(internal::scalar_quotient_op) - operator/(const MatrixBase &other) const; - - /** \deprecated ArrayBase::min() */ - template - const EIGEN_CWISE_BINOP_RETURN_TYPE(internal::scalar_min_op) - (min)(const MatrixBase &other) const - { return EIGEN_CWISE_BINOP_RETURN_TYPE(internal::scalar_min_op)(_expression(), other.derived()); } - - /** \deprecated ArrayBase::max() */ - template - const EIGEN_CWISE_BINOP_RETURN_TYPE(internal::scalar_max_op) - (max)(const MatrixBase &other) const - { return EIGEN_CWISE_BINOP_RETURN_TYPE(internal::scalar_max_op)(_expression(), other.derived()); } - - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_abs_op) abs() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_abs2_op) abs2() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_square_op) square() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_cube_op) cube() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_inverse_op) inverse() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_sqrt_op) sqrt() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_exp_op) exp() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_log_op) log() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_cos_op) cos() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_sin_op) sin() const; - const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_pow_op) pow(const Scalar& exponent) const; - - const ScalarAddReturnType - operator+(const Scalar& scalar) const; - - /** \relates Cwise */ - friend const ScalarAddReturnType - operator+(const Scalar& scalar, const Cwise& mat) - { return mat + scalar; } - - ExpressionType& operator+=(const Scalar& scalar); - - const ScalarAddReturnType - operator-(const Scalar& scalar) const; - - ExpressionType& operator-=(const Scalar& scalar); - - template - inline ExpressionType& operator*=(const MatrixBase &other); - - template - inline ExpressionType& operator/=(const MatrixBase &other); - - template const EIGEN_CWISE_BINOP_RETURN_TYPE(std::less) - operator<(const MatrixBase& other) const; - - template const EIGEN_CWISE_BINOP_RETURN_TYPE(std::less_equal) - operator<=(const MatrixBase& other) const; - - template const EIGEN_CWISE_BINOP_RETURN_TYPE(std::greater) - operator>(const MatrixBase& other) const; - - template const EIGEN_CWISE_BINOP_RETURN_TYPE(std::greater_equal) - operator>=(const MatrixBase& other) const; - - template const EIGEN_CWISE_BINOP_RETURN_TYPE(std::equal_to) - operator==(const MatrixBase& other) const; - - template const EIGEN_CWISE_BINOP_RETURN_TYPE(std::not_equal_to) - operator!=(const MatrixBase& other) const; - - // comparisons to a scalar value - const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::less) - operator<(Scalar s) const; - - const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::less_equal) - operator<=(Scalar s) const; - - const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::greater) - operator>(Scalar s) const; - - const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::greater_equal) - operator>=(Scalar s) const; - - const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::equal_to) - operator==(Scalar s) const; - - const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::not_equal_to) - operator!=(Scalar s) const; - - // allow to extend Cwise outside Eigen - #ifdef EIGEN_CWISE_PLUGIN - #include EIGEN_CWISE_PLUGIN - #endif - - protected: - ExpressionTypeNested m_matrix; -}; - - -/** \returns a Cwise wrapper of *this providing additional coefficient-wise operations - * - * Example: \include MatrixBase_cwise_const.cpp - * Output: \verbinclude MatrixBase_cwise_const.out - * - * \sa class Cwise, cwise() - */ -template -inline const Cwise MatrixBase::cwise() const -{ - return derived(); -} - -/** \returns a Cwise wrapper of *this providing additional coefficient-wise operations - * - * Example: \include MatrixBase_cwise.cpp - * Output: \verbinclude MatrixBase_cwise.out - * - * \sa class Cwise, cwise() const - */ -template -inline Cwise MatrixBase::cwise() -{ - return derived(); -} - -} // end namespace Eigen - -#endif // EIGEN_CWISE_H diff --git a/Eigen/src/Eigen2Support/CwiseOperators.h b/Eigen/src/Eigen2Support/CwiseOperators.h deleted file mode 100644 index 482f30648..000000000 --- a/Eigen/src/Eigen2Support/CwiseOperators.h +++ /dev/null @@ -1,298 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_ARRAY_CWISE_OPERATORS_H -#define EIGEN_ARRAY_CWISE_OPERATORS_H - -namespace Eigen { - -/*************************************************************************** -* The following functions were defined in Core -***************************************************************************/ - - -/** \deprecated ArrayBase::abs() */ -template -EIGEN_STRONG_INLINE const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_abs_op) -Cwise::abs() const -{ - return _expression(); -} - -/** \deprecated ArrayBase::abs2() */ -template -EIGEN_STRONG_INLINE const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_abs2_op) -Cwise::abs2() const -{ - return _expression(); -} - -/** \deprecated ArrayBase::exp() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_exp_op) -Cwise::exp() const -{ - return _expression(); -} - -/** \deprecated ArrayBase::log() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_log_op) -Cwise::log() const -{ - return _expression(); -} - -/** \deprecated ArrayBase::operator*() */ -template -template -EIGEN_STRONG_INLINE const EIGEN_CWISE_PRODUCT_RETURN_TYPE(ExpressionType,OtherDerived) -Cwise::operator*(const MatrixBase &other) const -{ - return EIGEN_CWISE_PRODUCT_RETURN_TYPE(ExpressionType,OtherDerived)(_expression(), other.derived()); -} - -/** \deprecated ArrayBase::operator/() */ -template -template -EIGEN_STRONG_INLINE const EIGEN_CWISE_BINOP_RETURN_TYPE(internal::scalar_quotient_op) -Cwise::operator/(const MatrixBase &other) const -{ - return EIGEN_CWISE_BINOP_RETURN_TYPE(internal::scalar_quotient_op)(_expression(), other.derived()); -} - -/** \deprecated ArrayBase::operator*=() */ -template -template -inline ExpressionType& Cwise::operator*=(const MatrixBase &other) -{ - return m_matrix.const_cast_derived() = *this * other; -} - -/** \deprecated ArrayBase::operator/=() */ -template -template -inline ExpressionType& Cwise::operator/=(const MatrixBase &other) -{ - return m_matrix.const_cast_derived() = *this / other; -} - -/*************************************************************************** -* The following functions were defined in Array -***************************************************************************/ - -// -- unary operators -- - -/** \deprecated ArrayBase::sqrt() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_sqrt_op) -Cwise::sqrt() const -{ - return _expression(); -} - -/** \deprecated ArrayBase::cos() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_cos_op) -Cwise::cos() const -{ - return _expression(); -} - - -/** \deprecated ArrayBase::sin() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_sin_op) -Cwise::sin() const -{ - return _expression(); -} - - -/** \deprecated ArrayBase::log() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_pow_op) -Cwise::pow(const Scalar& exponent) const -{ - return EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_pow_op)(_expression(), internal::scalar_pow_op(exponent)); -} - - -/** \deprecated ArrayBase::inverse() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_inverse_op) -Cwise::inverse() const -{ - return _expression(); -} - -/** \deprecated ArrayBase::square() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_square_op) -Cwise::square() const -{ - return _expression(); -} - -/** \deprecated ArrayBase::cube() */ -template -inline const EIGEN_CWISE_UNOP_RETURN_TYPE(internal::scalar_cube_op) -Cwise::cube() const -{ - return _expression(); -} - - -// -- binary operators -- - -/** \deprecated ArrayBase::operator<() */ -template -template -inline const EIGEN_CWISE_BINOP_RETURN_TYPE(std::less) -Cwise::operator<(const MatrixBase &other) const -{ - return EIGEN_CWISE_BINOP_RETURN_TYPE(std::less)(_expression(), other.derived()); -} - -/** \deprecated ArrayBase::<=() */ -template -template -inline const EIGEN_CWISE_BINOP_RETURN_TYPE(std::less_equal) -Cwise::operator<=(const MatrixBase &other) const -{ - return EIGEN_CWISE_BINOP_RETURN_TYPE(std::less_equal)(_expression(), other.derived()); -} - -/** \deprecated ArrayBase::operator>() */ -template -template -inline const EIGEN_CWISE_BINOP_RETURN_TYPE(std::greater) -Cwise::operator>(const MatrixBase &other) const -{ - return EIGEN_CWISE_BINOP_RETURN_TYPE(std::greater)(_expression(), other.derived()); -} - -/** \deprecated ArrayBase::operator>=() */ -template -template -inline const EIGEN_CWISE_BINOP_RETURN_TYPE(std::greater_equal) -Cwise::operator>=(const MatrixBase &other) const -{ - return EIGEN_CWISE_BINOP_RETURN_TYPE(std::greater_equal)(_expression(), other.derived()); -} - -/** \deprecated ArrayBase::operator==() */ -template -template -inline const EIGEN_CWISE_BINOP_RETURN_TYPE(std::equal_to) -Cwise::operator==(const MatrixBase &other) const -{ - return EIGEN_CWISE_BINOP_RETURN_TYPE(std::equal_to)(_expression(), other.derived()); -} - -/** \deprecated ArrayBase::operator!=() */ -template -template -inline const EIGEN_CWISE_BINOP_RETURN_TYPE(std::not_equal_to) -Cwise::operator!=(const MatrixBase &other) const -{ - return EIGEN_CWISE_BINOP_RETURN_TYPE(std::not_equal_to)(_expression(), other.derived()); -} - -// comparisons to scalar value - -/** \deprecated ArrayBase::operator<(Scalar) */ -template -inline const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::less) -Cwise::operator<(Scalar s) const -{ - return EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::less)(_expression(), - typename ExpressionType::ConstantReturnType(_expression().rows(), _expression().cols(), s)); -} - -/** \deprecated ArrayBase::operator<=(Scalar) */ -template -inline const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::less_equal) -Cwise::operator<=(Scalar s) const -{ - return EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::less_equal)(_expression(), - typename ExpressionType::ConstantReturnType(_expression().rows(), _expression().cols(), s)); -} - -/** \deprecated ArrayBase::operator>(Scalar) */ -template -inline const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::greater) -Cwise::operator>(Scalar s) const -{ - return EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::greater)(_expression(), - typename ExpressionType::ConstantReturnType(_expression().rows(), _expression().cols(), s)); -} - -/** \deprecated ArrayBase::operator>=(Scalar) */ -template -inline const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::greater_equal) -Cwise::operator>=(Scalar s) const -{ - return EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::greater_equal)(_expression(), - typename ExpressionType::ConstantReturnType(_expression().rows(), _expression().cols(), s)); -} - -/** \deprecated ArrayBase::operator==(Scalar) */ -template -inline const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::equal_to) -Cwise::operator==(Scalar s) const -{ - return EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::equal_to)(_expression(), - typename ExpressionType::ConstantReturnType(_expression().rows(), _expression().cols(), s)); -} - -/** \deprecated ArrayBase::operator!=(Scalar) */ -template -inline const EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::not_equal_to) -Cwise::operator!=(Scalar s) const -{ - return EIGEN_CWISE_COMP_TO_SCALAR_RETURN_TYPE(std::not_equal_to)(_expression(), - typename ExpressionType::ConstantReturnType(_expression().rows(), _expression().cols(), s)); -} - -// scalar addition - -/** \deprecated ArrayBase::operator+(Scalar) */ -template -inline const typename Cwise::ScalarAddReturnType -Cwise::operator+(const Scalar& scalar) const -{ - return typename Cwise::ScalarAddReturnType(m_matrix, internal::scalar_add_op(scalar)); -} - -/** \deprecated ArrayBase::operator+=(Scalar) */ -template -inline ExpressionType& Cwise::operator+=(const Scalar& scalar) -{ - return m_matrix.const_cast_derived() = *this + scalar; -} - -/** \deprecated ArrayBase::operator-(Scalar) */ -template -inline const typename Cwise::ScalarAddReturnType -Cwise::operator-(const Scalar& scalar) const -{ - return *this + (-scalar); -} - -/** \deprecated ArrayBase::operator-=(Scalar) */ -template -inline ExpressionType& Cwise::operator-=(const Scalar& scalar) -{ - return m_matrix.const_cast_derived() = *this - scalar; -} - -} // end namespace Eigen - -#endif // EIGEN_ARRAY_CWISE_OPERATORS_H diff --git a/Eigen/src/Eigen2Support/Geometry/AlignedBox.h b/Eigen/src/Eigen2Support/Geometry/AlignedBox.h deleted file mode 100644 index 2e4309dd9..000000000 --- a/Eigen/src/Eigen2Support/Geometry/AlignedBox.h +++ /dev/null @@ -1,159 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -/** \geometry_module \ingroup Geometry_Module - * \nonstableyet - * - * \class AlignedBox - * - * \brief An axis aligned box - * - * \param _Scalar the type of the scalar coefficients - * \param _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic. - * - * This class represents an axis aligned box as a pair of the minimal and maximal corners. - */ -template -class AlignedBox -{ -public: -EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim==Dynamic ? Dynamic : _AmbientDim+1) - enum { AmbientDimAtCompileTime = _AmbientDim }; - typedef _Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - /** Default constructor initializing a null box. */ - inline AlignedBox() - { if (AmbientDimAtCompileTime!=Dynamic) setNull(); } - - /** Constructs a null box with \a _dim the dimension of the ambient space. */ - inline explicit AlignedBox(int _dim) : m_min(_dim), m_max(_dim) - { setNull(); } - - /** Constructs a box with extremities \a _min and \a _max. */ - inline AlignedBox(const VectorType& _min, const VectorType& _max) : m_min(_min), m_max(_max) {} - - /** Constructs a box containing a single point \a p. */ - inline explicit AlignedBox(const VectorType& p) : m_min(p), m_max(p) {} - - ~AlignedBox() {} - - /** \returns the dimension in which the box holds */ - inline int dim() const { return AmbientDimAtCompileTime==Dynamic ? m_min.size()-1 : AmbientDimAtCompileTime; } - - /** \returns true if the box is null, i.e, empty. */ - inline bool isNull() const { return (m_min.cwise() > m_max).any(); } - - /** Makes \c *this a null/empty box. */ - inline void setNull() - { - m_min.setConstant( (std::numeric_limits::max)()); - m_max.setConstant(-(std::numeric_limits::max)()); - } - - /** \returns the minimal corner */ - inline const VectorType& (min)() const { return m_min; } - /** \returns a non const reference to the minimal corner */ - inline VectorType& (min)() { return m_min; } - /** \returns the maximal corner */ - inline const VectorType& (max)() const { return m_max; } - /** \returns a non const reference to the maximal corner */ - inline VectorType& (max)() { return m_max; } - - /** \returns true if the point \a p is inside the box \c *this. */ - inline bool contains(const VectorType& p) const - { return (m_min.cwise()<=p).all() && (p.cwise()<=m_max).all(); } - - /** \returns true if the box \a b is entirely inside the box \c *this. */ - inline bool contains(const AlignedBox& b) const - { return (m_min.cwise()<=(b.min)()).all() && ((b.max)().cwise()<=m_max).all(); } - - /** Extends \c *this such that it contains the point \a p and returns a reference to \c *this. */ - inline AlignedBox& extend(const VectorType& p) - { m_min = (m_min.cwise().min)(p); m_max = (m_max.cwise().max)(p); return *this; } - - /** Extends \c *this such that it contains the box \a b and returns a reference to \c *this. */ - inline AlignedBox& extend(const AlignedBox& b) - { m_min = (m_min.cwise().min)(b.m_min); m_max = (m_max.cwise().max)(b.m_max); return *this; } - - /** Clamps \c *this by the box \a b and returns a reference to \c *this. */ - inline AlignedBox& clamp(const AlignedBox& b) - { m_min = (m_min.cwise().max)(b.m_min); m_max = (m_max.cwise().min)(b.m_max); return *this; } - - /** Translate \c *this by the vector \a t and returns a reference to \c *this. */ - inline AlignedBox& translate(const VectorType& t) - { m_min += t; m_max += t; return *this; } - - /** \returns the squared distance between the point \a p and the box \c *this, - * and zero if \a p is inside the box. - * \sa exteriorDistance() - */ - inline Scalar squaredExteriorDistance(const VectorType& p) const; - - /** \returns the distance between the point \a p and the box \c *this, - * and zero if \a p is inside the box. - * \sa squaredExteriorDistance() - */ - inline Scalar exteriorDistance(const VectorType& p) const - { return ei_sqrt(squaredExteriorDistance(p)); } - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { - return typename internal::cast_return_type >::type(*this); - } - - /** Copy constructor with scalar type conversion */ - template - inline explicit AlignedBox(const AlignedBox& other) - { - m_min = (other.min)().template cast(); - m_max = (other.max)().template cast(); - } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const AlignedBox& other, typename NumTraits::Real prec = precision()) const - { return m_min.isApprox(other.m_min, prec) && m_max.isApprox(other.m_max, prec); } - -protected: - - VectorType m_min, m_max; -}; - -template -inline Scalar AlignedBox::squaredExteriorDistance(const VectorType& p) const -{ - Scalar dist2(0); - Scalar aux; - for (int k=0; k - -#ifndef M_PI -#define M_PI 3.14159265358979323846 -#endif - -#if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS -#include "RotationBase.h" -#include "Rotation2D.h" -#include "Quaternion.h" -#include "AngleAxis.h" -#include "Transform.h" -#include "Translation.h" -#include "Scaling.h" -#include "AlignedBox.h" -#include "Hyperplane.h" -#include "ParametrizedLine.h" -#endif - - -#define RotationBase eigen2_RotationBase -#define Rotation2D eigen2_Rotation2D -#define Rotation2Df eigen2_Rotation2Df -#define Rotation2Dd eigen2_Rotation2Dd - -#define Quaternion eigen2_Quaternion -#define Quaternionf eigen2_Quaternionf -#define Quaterniond eigen2_Quaterniond - -#define AngleAxis eigen2_AngleAxis -#define AngleAxisf eigen2_AngleAxisf -#define AngleAxisd eigen2_AngleAxisd - -#define Transform eigen2_Transform -#define Transform2f eigen2_Transform2f -#define Transform2d eigen2_Transform2d -#define Transform3f eigen2_Transform3f -#define Transform3d eigen2_Transform3d - -#define Translation eigen2_Translation -#define Translation2f eigen2_Translation2f -#define Translation2d eigen2_Translation2d -#define Translation3f eigen2_Translation3f -#define Translation3d eigen2_Translation3d - -#define Scaling eigen2_Scaling -#define Scaling2f eigen2_Scaling2f -#define Scaling2d eigen2_Scaling2d -#define Scaling3f eigen2_Scaling3f -#define Scaling3d eigen2_Scaling3d - -#define AlignedBox eigen2_AlignedBox - -#define Hyperplane eigen2_Hyperplane -#define ParametrizedLine eigen2_ParametrizedLine - -#define ei_toRotationMatrix eigen2_ei_toRotationMatrix -#define ei_quaternion_assign_impl eigen2_ei_quaternion_assign_impl -#define ei_transform_product_impl eigen2_ei_transform_product_impl - -#include "RotationBase.h" -#include "Rotation2D.h" -#include "Quaternion.h" -#include "AngleAxis.h" -#include "Transform.h" -#include "Translation.h" -#include "Scaling.h" -#include "AlignedBox.h" -#include "Hyperplane.h" -#include "ParametrizedLine.h" - -#undef ei_toRotationMatrix -#undef ei_quaternion_assign_impl -#undef ei_transform_product_impl - -#undef RotationBase -#undef Rotation2D -#undef Rotation2Df -#undef Rotation2Dd - -#undef Quaternion -#undef Quaternionf -#undef Quaterniond - -#undef AngleAxis -#undef AngleAxisf -#undef AngleAxisd - -#undef Transform -#undef Transform2f -#undef Transform2d -#undef Transform3f -#undef Transform3d - -#undef Translation -#undef Translation2f -#undef Translation2d -#undef Translation3f -#undef Translation3d - -#undef Scaling -#undef Scaling2f -#undef Scaling2d -#undef Scaling3f -#undef Scaling3d - -#undef AlignedBox - -#undef Hyperplane -#undef ParametrizedLine - -#endif // EIGEN2_GEOMETRY_MODULE_H diff --git a/Eigen/src/Eigen2Support/Geometry/AngleAxis.h b/Eigen/src/Eigen2Support/Geometry/AngleAxis.h deleted file mode 100644 index a0b4ac44e..000000000 --- a/Eigen/src/Eigen2Support/Geometry/AngleAxis.h +++ /dev/null @@ -1,228 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -/** \geometry_module \ingroup Geometry_Module - * - * \class AngleAxis - * - * \brief Represents a 3D rotation as a rotation angle around an arbitrary 3D axis - * - * \param _Scalar the scalar type, i.e., the type of the coefficients. - * - * The following two typedefs are provided for convenience: - * \li \c AngleAxisf for \c float - * \li \c AngleAxisd for \c double - * - * \addexample AngleAxisForEuler \label How to define a rotation from Euler-angles - * - * Combined with MatrixBase::Unit{X,Y,Z}, AngleAxis can be used to easily - * mimic Euler-angles. Here is an example: - * \include AngleAxis_mimic_euler.cpp - * Output: \verbinclude AngleAxis_mimic_euler.out - * - * \note This class is not aimed to be used to store a rotation transformation, - * but rather to make easier the creation of other rotation (Quaternion, rotation Matrix) - * and transformation objects. - * - * \sa class Quaternion, class Transform, MatrixBase::UnitX() - */ - -template struct ei_traits > -{ - typedef _Scalar Scalar; -}; - -template -class AngleAxis : public RotationBase,3> -{ - typedef RotationBase,3> Base; - -public: - - using Base::operator*; - - enum { Dim = 3 }; - /** the scalar type of the coefficients */ - typedef _Scalar Scalar; - typedef Matrix Matrix3; - typedef Matrix Vector3; - typedef Quaternion QuaternionType; - -protected: - - Vector3 m_axis; - Scalar m_angle; - -public: - - /** Default constructor without initialization. */ - AngleAxis() {} - - /** Constructs and initialize the angle-axis rotation from an \a angle in radian - * and an \a axis which must be normalized. */ - template - inline AngleAxis(Scalar angle, const MatrixBase& axis) : m_axis(axis), m_angle(angle) - { - using std::sqrt; - using std::abs; - // since we compare against 1, this is equal to computing the relative error - eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( NumTraits::dummy_precision() ) ); - } - - /** Constructs and initialize the angle-axis rotation from a quaternion \a q. */ - inline AngleAxis(const QuaternionType& q) { *this = q; } - - /** Constructs and initialize the angle-axis rotation from a 3x3 rotation matrix. */ - template - inline explicit AngleAxis(const MatrixBase& m) { *this = m; } - - Scalar angle() const { return m_angle; } - Scalar& angle() { return m_angle; } - - const Vector3& axis() const { return m_axis; } - Vector3& axis() { return m_axis; } - - /** Concatenates two rotations */ - inline QuaternionType operator* (const AngleAxis& other) const - { return QuaternionType(*this) * QuaternionType(other); } - - /** Concatenates two rotations */ - inline QuaternionType operator* (const QuaternionType& other) const - { return QuaternionType(*this) * other; } - - /** Concatenates two rotations */ - friend inline QuaternionType operator* (const QuaternionType& a, const AngleAxis& b) - { return a * QuaternionType(b); } - - /** Concatenates two rotations */ - inline Matrix3 operator* (const Matrix3& other) const - { return toRotationMatrix() * other; } - - /** Concatenates two rotations */ - inline friend Matrix3 operator* (const Matrix3& a, const AngleAxis& b) - { return a * b.toRotationMatrix(); } - - /** Applies rotation to vector */ - inline Vector3 operator* (const Vector3& other) const - { return toRotationMatrix() * other; } - - /** \returns the inverse rotation, i.e., an angle-axis with opposite rotation angle */ - AngleAxis inverse() const - { return AngleAxis(-m_angle, m_axis); } - - AngleAxis& operator=(const QuaternionType& q); - template - AngleAxis& operator=(const MatrixBase& m); - - template - AngleAxis& fromRotationMatrix(const MatrixBase& m); - Matrix3 toRotationMatrix(void) const; - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { return typename internal::cast_return_type >::type(*this); } - - /** Copy constructor with scalar type conversion */ - template - inline explicit AngleAxis(const AngleAxis& other) - { - m_axis = other.axis().template cast(); - m_angle = Scalar(other.angle()); - } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const AngleAxis& other, typename NumTraits::Real prec = precision()) const - { return m_axis.isApprox(other.m_axis, prec) && ei_isApprox(m_angle,other.m_angle, prec); } -}; - -/** \ingroup Geometry_Module - * single precision angle-axis type */ -typedef AngleAxis AngleAxisf; -/** \ingroup Geometry_Module - * double precision angle-axis type */ -typedef AngleAxis AngleAxisd; - -/** Set \c *this from a quaternion. - * The axis is normalized. - */ -template -AngleAxis& AngleAxis::operator=(const QuaternionType& q) -{ - Scalar n2 = q.vec().squaredNorm(); - if (n2 < precision()*precision()) - { - m_angle = 0; - m_axis << 1, 0, 0; - } - else - { - m_angle = 2*std::acos(q.w()); - m_axis = q.vec() / ei_sqrt(n2); - - using std::sqrt; - using std::abs; - // since we compare against 1, this is equal to computing the relative error - eigen_assert( abs(m_axis.derived().squaredNorm() - 1) < sqrt( NumTraits::dummy_precision() ) ); - } - return *this; -} - -/** Set \c *this from a 3x3 rotation matrix \a mat. - */ -template -template -AngleAxis& AngleAxis::operator=(const MatrixBase& mat) -{ - // Since a direct conversion would not be really faster, - // let's use the robust Quaternion implementation: - return *this = QuaternionType(mat); -} - -/** Constructs and \returns an equivalent 3x3 rotation matrix. - */ -template -typename AngleAxis::Matrix3 -AngleAxis::toRotationMatrix(void) const -{ - Matrix3 res; - Vector3 sin_axis = ei_sin(m_angle) * m_axis; - Scalar c = ei_cos(m_angle); - Vector3 cos1_axis = (Scalar(1)-c) * m_axis; - - Scalar tmp; - tmp = cos1_axis.x() * m_axis.y(); - res.coeffRef(0,1) = tmp - sin_axis.z(); - res.coeffRef(1,0) = tmp + sin_axis.z(); - - tmp = cos1_axis.x() * m_axis.z(); - res.coeffRef(0,2) = tmp + sin_axis.y(); - res.coeffRef(2,0) = tmp - sin_axis.y(); - - tmp = cos1_axis.y() * m_axis.z(); - res.coeffRef(1,2) = tmp - sin_axis.x(); - res.coeffRef(2,1) = tmp + sin_axis.x(); - - res.diagonal() = (cos1_axis.cwise() * m_axis).cwise() + c; - - return res; -} - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/CMakeLists.txt b/Eigen/src/Eigen2Support/Geometry/CMakeLists.txt deleted file mode 100644 index c347a8f26..000000000 --- a/Eigen/src/Eigen2Support/Geometry/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_Eigen2Support_Geometry_SRCS "*.h") - -INSTALL(FILES - ${Eigen_Eigen2Support_Geometry_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen/src/Eigen2Support/Geometry - ) diff --git a/Eigen/src/Eigen2Support/Geometry/Hyperplane.h b/Eigen/src/Eigen2Support/Geometry/Hyperplane.h deleted file mode 100644 index b95bf00ec..000000000 --- a/Eigen/src/Eigen2Support/Geometry/Hyperplane.h +++ /dev/null @@ -1,254 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -/** \geometry_module \ingroup Geometry_Module - * - * \class Hyperplane - * - * \brief A hyperplane - * - * A hyperplane is an affine subspace of dimension n-1 in a space of dimension n. - * For example, a hyperplane in a plane is a line; a hyperplane in 3-space is a plane. - * - * \param _Scalar the scalar type, i.e., the type of the coefficients - * \param _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic. - * Notice that the dimension of the hyperplane is _AmbientDim-1. - * - * This class represents an hyperplane as the zero set of the implicit equation - * \f$ n \cdot x + d = 0 \f$ where \f$ n \f$ is a unit normal vector of the plane (linear part) - * and \f$ d \f$ is the distance (offset) to the origin. - */ -template -class Hyperplane -{ -public: - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim==Dynamic ? Dynamic : _AmbientDim+1) - enum { AmbientDimAtCompileTime = _AmbientDim }; - typedef _Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - typedef Matrix Coefficients; - typedef Block NormalReturnType; - - /** Default constructor without initialization */ - inline Hyperplane() {} - - /** Constructs a dynamic-size hyperplane with \a _dim the dimension - * of the ambient space */ - inline explicit Hyperplane(int _dim) : m_coeffs(_dim+1) {} - - /** Construct a plane from its normal \a n and a point \a e onto the plane. - * \warning the vector normal is assumed to be normalized. - */ - inline Hyperplane(const VectorType& n, const VectorType& e) - : m_coeffs(n.size()+1) - { - normal() = n; - offset() = -e.eigen2_dot(n); - } - - /** Constructs a plane from its normal \a n and distance to the origin \a d - * such that the algebraic equation of the plane is \f$ n \cdot x + d = 0 \f$. - * \warning the vector normal is assumed to be normalized. - */ - inline Hyperplane(const VectorType& n, Scalar d) - : m_coeffs(n.size()+1) - { - normal() = n; - offset() = d; - } - - /** Constructs a hyperplane passing through the two points. If the dimension of the ambient space - * is greater than 2, then there isn't uniqueness, so an arbitrary choice is made. - */ - static inline Hyperplane Through(const VectorType& p0, const VectorType& p1) - { - Hyperplane result(p0.size()); - result.normal() = (p1 - p0).unitOrthogonal(); - result.offset() = -result.normal().eigen2_dot(p0); - return result; - } - - /** Constructs a hyperplane passing through the three points. The dimension of the ambient space - * is required to be exactly 3. - */ - static inline Hyperplane Through(const VectorType& p0, const VectorType& p1, const VectorType& p2) - { - EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(VectorType, 3) - Hyperplane result(p0.size()); - result.normal() = (p2 - p0).cross(p1 - p0).normalized(); - result.offset() = -result.normal().eigen2_dot(p0); - return result; - } - - /** Constructs a hyperplane passing through the parametrized line \a parametrized. - * If the dimension of the ambient space is greater than 2, then there isn't uniqueness, - * so an arbitrary choice is made. - */ - // FIXME to be consitent with the rest this could be implemented as a static Through function ?? - explicit Hyperplane(const ParametrizedLine& parametrized) - { - normal() = parametrized.direction().unitOrthogonal(); - offset() = -normal().eigen2_dot(parametrized.origin()); - } - - ~Hyperplane() {} - - /** \returns the dimension in which the plane holds */ - inline int dim() const { return int(AmbientDimAtCompileTime)==Dynamic ? m_coeffs.size()-1 : int(AmbientDimAtCompileTime); } - - /** normalizes \c *this */ - void normalize(void) - { - m_coeffs /= normal().norm(); - } - - /** \returns the signed distance between the plane \c *this and a point \a p. - * \sa absDistance() - */ - inline Scalar signedDistance(const VectorType& p) const { return p.eigen2_dot(normal()) + offset(); } - - /** \returns the absolute distance between the plane \c *this and a point \a p. - * \sa signedDistance() - */ - inline Scalar absDistance(const VectorType& p) const { return ei_abs(signedDistance(p)); } - - /** \returns the projection of a point \a p onto the plane \c *this. - */ - inline VectorType projection(const VectorType& p) const { return p - signedDistance(p) * normal(); } - - /** \returns a constant reference to the unit normal vector of the plane, which corresponds - * to the linear part of the implicit equation. - */ - inline const NormalReturnType normal() const { return NormalReturnType(*const_cast(&m_coeffs),0,0,dim(),1); } - - /** \returns a non-constant reference to the unit normal vector of the plane, which corresponds - * to the linear part of the implicit equation. - */ - inline NormalReturnType normal() { return NormalReturnType(m_coeffs,0,0,dim(),1); } - - /** \returns the distance to the origin, which is also the "constant term" of the implicit equation - * \warning the vector normal is assumed to be normalized. - */ - inline const Scalar& offset() const { return m_coeffs.coeff(dim()); } - - /** \returns a non-constant reference to the distance to the origin, which is also the constant part - * of the implicit equation */ - inline Scalar& offset() { return m_coeffs(dim()); } - - /** \returns a constant reference to the coefficients c_i of the plane equation: - * \f$ c_0*x_0 + ... + c_{d-1}*x_{d-1} + c_d = 0 \f$ - */ - inline const Coefficients& coeffs() const { return m_coeffs; } - - /** \returns a non-constant reference to the coefficients c_i of the plane equation: - * \f$ c_0*x_0 + ... + c_{d-1}*x_{d-1} + c_d = 0 \f$ - */ - inline Coefficients& coeffs() { return m_coeffs; } - - /** \returns the intersection of *this with \a other. - * - * \warning The ambient space must be a plane, i.e. have dimension 2, so that \c *this and \a other are lines. - * - * \note If \a other is approximately parallel to *this, this method will return any point on *this. - */ - VectorType intersection(const Hyperplane& other) - { - EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(VectorType, 2) - Scalar det = coeffs().coeff(0) * other.coeffs().coeff(1) - coeffs().coeff(1) * other.coeffs().coeff(0); - // since the line equations ax+by=c are normalized with a^2+b^2=1, the following tests - // whether the two lines are approximately parallel. - if(ei_isMuchSmallerThan(det, Scalar(1))) - { // special case where the two lines are approximately parallel. Pick any point on the first line. - if(ei_abs(coeffs().coeff(1))>ei_abs(coeffs().coeff(0))) - return VectorType(coeffs().coeff(1), -coeffs().coeff(2)/coeffs().coeff(1)-coeffs().coeff(0)); - else - return VectorType(-coeffs().coeff(2)/coeffs().coeff(0)-coeffs().coeff(1), coeffs().coeff(0)); - } - else - { // general case - Scalar invdet = Scalar(1) / det; - return VectorType(invdet*(coeffs().coeff(1)*other.coeffs().coeff(2)-other.coeffs().coeff(1)*coeffs().coeff(2)), - invdet*(other.coeffs().coeff(0)*coeffs().coeff(2)-coeffs().coeff(0)*other.coeffs().coeff(2))); - } - } - - /** Applies the transformation matrix \a mat to \c *this and returns a reference to \c *this. - * - * \param mat the Dim x Dim transformation matrix - * \param traits specifies whether the matrix \a mat represents an Isometry - * or a more generic Affine transformation. The default is Affine. - */ - template - inline Hyperplane& transform(const MatrixBase& mat, TransformTraits traits = Affine) - { - if (traits==Affine) - normal() = mat.inverse().transpose() * normal(); - else if (traits==Isometry) - normal() = mat * normal(); - else - { - ei_assert("invalid traits value in Hyperplane::transform()"); - } - return *this; - } - - /** Applies the transformation \a t to \c *this and returns a reference to \c *this. - * - * \param t the transformation of dimension Dim - * \param traits specifies whether the transformation \a t represents an Isometry - * or a more generic Affine transformation. The default is Affine. - * Other kind of transformations are not supported. - */ - inline Hyperplane& transform(const Transform& t, - TransformTraits traits = Affine) - { - transform(t.linear(), traits); - offset() -= t.translation().eigen2_dot(normal()); - return *this; - } - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { - return typename internal::cast_return_type >::type(*this); - } - - /** Copy constructor with scalar type conversion */ - template - inline explicit Hyperplane(const Hyperplane& other) - { m_coeffs = other.coeffs().template cast(); } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const Hyperplane& other, typename NumTraits::Real prec = precision()) const - { return m_coeffs.isApprox(other.m_coeffs, prec); } - -protected: - - Coefficients m_coeffs; -}; - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/ParametrizedLine.h b/Eigen/src/Eigen2Support/Geometry/ParametrizedLine.h deleted file mode 100644 index 9b57b7e0b..000000000 --- a/Eigen/src/Eigen2Support/Geometry/ParametrizedLine.h +++ /dev/null @@ -1,141 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -/** \geometry_module \ingroup Geometry_Module - * - * \class ParametrizedLine - * - * \brief A parametrized line - * - * A parametrized line is defined by an origin point \f$ \mathbf{o} \f$ and a unit - * direction vector \f$ \mathbf{d} \f$ such that the line corresponds to - * the set \f$ l(t) = \mathbf{o} + t \mathbf{d} \f$, \f$ l \in \mathbf{R} \f$. - * - * \param _Scalar the scalar type, i.e., the type of the coefficients - * \param _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic. - */ -template -class ParametrizedLine -{ -public: - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim) - enum { AmbientDimAtCompileTime = _AmbientDim }; - typedef _Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - /** Default constructor without initialization */ - inline ParametrizedLine() {} - - /** Constructs a dynamic-size line with \a _dim the dimension - * of the ambient space */ - inline explicit ParametrizedLine(int _dim) : m_origin(_dim), m_direction(_dim) {} - - /** Initializes a parametrized line of direction \a direction and origin \a origin. - * \warning the vector direction is assumed to be normalized. - */ - ParametrizedLine(const VectorType& origin, const VectorType& direction) - : m_origin(origin), m_direction(direction) {} - - explicit ParametrizedLine(const Hyperplane<_Scalar, _AmbientDim>& hyperplane); - - /** Constructs a parametrized line going from \a p0 to \a p1. */ - static inline ParametrizedLine Through(const VectorType& p0, const VectorType& p1) - { return ParametrizedLine(p0, (p1-p0).normalized()); } - - ~ParametrizedLine() {} - - /** \returns the dimension in which the line holds */ - inline int dim() const { return m_direction.size(); } - - const VectorType& origin() const { return m_origin; } - VectorType& origin() { return m_origin; } - - const VectorType& direction() const { return m_direction; } - VectorType& direction() { return m_direction; } - - /** \returns the squared distance of a point \a p to its projection onto the line \c *this. - * \sa distance() - */ - RealScalar squaredDistance(const VectorType& p) const - { - VectorType diff = p-origin(); - return (diff - diff.eigen2_dot(direction())* direction()).squaredNorm(); - } - /** \returns the distance of a point \a p to its projection onto the line \c *this. - * \sa squaredDistance() - */ - RealScalar distance(const VectorType& p) const { return ei_sqrt(squaredDistance(p)); } - - /** \returns the projection of a point \a p onto the line \c *this. */ - VectorType projection(const VectorType& p) const - { return origin() + (p-origin()).eigen2_dot(direction()) * direction(); } - - Scalar intersection(const Hyperplane<_Scalar, _AmbientDim>& hyperplane); - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { - return typename internal::cast_return_type >::type(*this); - } - - /** Copy constructor with scalar type conversion */ - template - inline explicit ParametrizedLine(const ParametrizedLine& other) - { - m_origin = other.origin().template cast(); - m_direction = other.direction().template cast(); - } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const ParametrizedLine& other, typename NumTraits::Real prec = precision()) const - { return m_origin.isApprox(other.m_origin, prec) && m_direction.isApprox(other.m_direction, prec); } - -protected: - - VectorType m_origin, m_direction; -}; - -/** Constructs a parametrized line from a 2D hyperplane - * - * \warning the ambient space must have dimension 2 such that the hyperplane actually describes a line - */ -template -inline ParametrizedLine<_Scalar, _AmbientDim>::ParametrizedLine(const Hyperplane<_Scalar, _AmbientDim>& hyperplane) -{ - EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(VectorType, 2) - direction() = hyperplane.normal().unitOrthogonal(); - origin() = -hyperplane.normal()*hyperplane.offset(); -} - -/** \returns the parameter value of the intersection between \c *this and the given hyperplane - */ -template -inline _Scalar ParametrizedLine<_Scalar, _AmbientDim>::intersection(const Hyperplane<_Scalar, _AmbientDim>& hyperplane) -{ - return -(hyperplane.offset()+origin().eigen2_dot(hyperplane.normal())) - /(direction().eigen2_dot(hyperplane.normal())); -} - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/Quaternion.h b/Eigen/src/Eigen2Support/Geometry/Quaternion.h deleted file mode 100644 index 4b6390cf1..000000000 --- a/Eigen/src/Eigen2Support/Geometry/Quaternion.h +++ /dev/null @@ -1,495 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -template -struct ei_quaternion_assign_impl; - -/** \geometry_module \ingroup Geometry_Module - * - * \class Quaternion - * - * \brief The quaternion class used to represent 3D orientations and rotations - * - * \param _Scalar the scalar type, i.e., the type of the coefficients - * - * This class represents a quaternion \f$ w+xi+yj+zk \f$ that is a convenient representation of - * orientations and rotations of objects in three dimensions. Compared to other representations - * like Euler angles or 3x3 matrices, quatertions offer the following advantages: - * \li \b compact storage (4 scalars) - * \li \b efficient to compose (28 flops), - * \li \b stable spherical interpolation - * - * The following two typedefs are provided for convenience: - * \li \c Quaternionf for \c float - * \li \c Quaterniond for \c double - * - * \sa class AngleAxis, class Transform - */ - -template struct ei_traits > -{ - typedef _Scalar Scalar; -}; - -template -class Quaternion : public RotationBase,3> -{ - typedef RotationBase,3> Base; - -public: - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,4) - - using Base::operator*; - - /** the scalar type of the coefficients */ - typedef _Scalar Scalar; - - /** the type of the Coefficients 4-vector */ - typedef Matrix Coefficients; - /** the type of a 3D vector */ - typedef Matrix Vector3; - /** the equivalent rotation matrix type */ - typedef Matrix Matrix3; - /** the equivalent angle-axis type */ - typedef AngleAxis AngleAxisType; - - /** \returns the \c x coefficient */ - inline Scalar x() const { return m_coeffs.coeff(0); } - /** \returns the \c y coefficient */ - inline Scalar y() const { return m_coeffs.coeff(1); } - /** \returns the \c z coefficient */ - inline Scalar z() const { return m_coeffs.coeff(2); } - /** \returns the \c w coefficient */ - inline Scalar w() const { return m_coeffs.coeff(3); } - - /** \returns a reference to the \c x coefficient */ - inline Scalar& x() { return m_coeffs.coeffRef(0); } - /** \returns a reference to the \c y coefficient */ - inline Scalar& y() { return m_coeffs.coeffRef(1); } - /** \returns a reference to the \c z coefficient */ - inline Scalar& z() { return m_coeffs.coeffRef(2); } - /** \returns a reference to the \c w coefficient */ - inline Scalar& w() { return m_coeffs.coeffRef(3); } - - /** \returns a read-only vector expression of the imaginary part (x,y,z) */ - inline const Block vec() const { return m_coeffs.template start<3>(); } - - /** \returns a vector expression of the imaginary part (x,y,z) */ - inline Block vec() { return m_coeffs.template start<3>(); } - - /** \returns a read-only vector expression of the coefficients (x,y,z,w) */ - inline const Coefficients& coeffs() const { return m_coeffs; } - - /** \returns a vector expression of the coefficients (x,y,z,w) */ - inline Coefficients& coeffs() { return m_coeffs; } - - /** Default constructor leaving the quaternion uninitialized. */ - inline Quaternion() {} - - /** Constructs and initializes the quaternion \f$ w+xi+yj+zk \f$ from - * its four coefficients \a w, \a x, \a y and \a z. - * - * \warning Note the order of the arguments: the real \a w coefficient first, - * while internally the coefficients are stored in the following order: - * [\c x, \c y, \c z, \c w] - */ - inline Quaternion(Scalar w, Scalar x, Scalar y, Scalar z) - { m_coeffs << x, y, z, w; } - - /** Copy constructor */ - inline Quaternion(const Quaternion& other) { m_coeffs = other.m_coeffs; } - - /** Constructs and initializes a quaternion from the angle-axis \a aa */ - explicit inline Quaternion(const AngleAxisType& aa) { *this = aa; } - - /** Constructs and initializes a quaternion from either: - * - a rotation matrix expression, - * - a 4D vector expression representing quaternion coefficients. - * \sa operator=(MatrixBase) - */ - template - explicit inline Quaternion(const MatrixBase& other) { *this = other; } - - Quaternion& operator=(const Quaternion& other); - Quaternion& operator=(const AngleAxisType& aa); - template - Quaternion& operator=(const MatrixBase& m); - - /** \returns a quaternion representing an identity rotation - * \sa MatrixBase::Identity() - */ - static inline Quaternion Identity() { return Quaternion(1, 0, 0, 0); } - - /** \sa Quaternion::Identity(), MatrixBase::setIdentity() - */ - inline Quaternion& setIdentity() { m_coeffs << 0, 0, 0, 1; return *this; } - - /** \returns the squared norm of the quaternion's coefficients - * \sa Quaternion::norm(), MatrixBase::squaredNorm() - */ - inline Scalar squaredNorm() const { return m_coeffs.squaredNorm(); } - - /** \returns the norm of the quaternion's coefficients - * \sa Quaternion::squaredNorm(), MatrixBase::norm() - */ - inline Scalar norm() const { return m_coeffs.norm(); } - - /** Normalizes the quaternion \c *this - * \sa normalized(), MatrixBase::normalize() */ - inline void normalize() { m_coeffs.normalize(); } - /** \returns a normalized version of \c *this - * \sa normalize(), MatrixBase::normalized() */ - inline Quaternion normalized() const { return Quaternion(m_coeffs.normalized()); } - - /** \returns the dot product of \c *this and \a other - * Geometrically speaking, the dot product of two unit quaternions - * corresponds to the cosine of half the angle between the two rotations. - * \sa angularDistance() - */ - inline Scalar eigen2_dot(const Quaternion& other) const { return m_coeffs.eigen2_dot(other.m_coeffs); } - - inline Scalar angularDistance(const Quaternion& other) const; - - Matrix3 toRotationMatrix(void) const; - - template - Quaternion& setFromTwoVectors(const MatrixBase& a, const MatrixBase& b); - - inline Quaternion operator* (const Quaternion& q) const; - inline Quaternion& operator*= (const Quaternion& q); - - Quaternion inverse(void) const; - Quaternion conjugate(void) const; - - Quaternion slerp(Scalar t, const Quaternion& other) const; - - template - Vector3 operator* (const MatrixBase& vec) const; - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { return typename internal::cast_return_type >::type(*this); } - - /** Copy constructor with scalar type conversion */ - template - inline explicit Quaternion(const Quaternion& other) - { m_coeffs = other.coeffs().template cast(); } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const Quaternion& other, typename NumTraits::Real prec = precision()) const - { return m_coeffs.isApprox(other.m_coeffs, prec); } - -protected: - Coefficients m_coeffs; -}; - -/** \ingroup Geometry_Module - * single precision quaternion type */ -typedef Quaternion Quaternionf; -/** \ingroup Geometry_Module - * double precision quaternion type */ -typedef Quaternion Quaterniond; - -// Generic Quaternion * Quaternion product -template inline Quaternion -ei_quaternion_product(const Quaternion& a, const Quaternion& b) -{ - return Quaternion - ( - a.w() * b.w() - a.x() * b.x() - a.y() * b.y() - a.z() * b.z(), - a.w() * b.x() + a.x() * b.w() + a.y() * b.z() - a.z() * b.y(), - a.w() * b.y() + a.y() * b.w() + a.z() * b.x() - a.x() * b.z(), - a.w() * b.z() + a.z() * b.w() + a.x() * b.y() - a.y() * b.x() - ); -} - -/** \returns the concatenation of two rotations as a quaternion-quaternion product */ -template -inline Quaternion Quaternion::operator* (const Quaternion& other) const -{ - return ei_quaternion_product(*this,other); -} - -/** \sa operator*(Quaternion) */ -template -inline Quaternion& Quaternion::operator*= (const Quaternion& other) -{ - return (*this = *this * other); -} - -/** Rotation of a vector by a quaternion. - * \remarks If the quaternion is used to rotate several points (>1) - * then it is much more efficient to first convert it to a 3x3 Matrix. - * Comparison of the operation cost for n transformations: - * - Quaternion: 30n - * - Via a Matrix3: 24 + 15n - */ -template -template -inline typename Quaternion::Vector3 -Quaternion::operator* (const MatrixBase& v) const -{ - // Note that this algorithm comes from the optimization by hand - // of the conversion to a Matrix followed by a Matrix/Vector product. - // It appears to be much faster than the common algorithm found - // in the litterature (30 versus 39 flops). It also requires two - // Vector3 as temporaries. - Vector3 uv; - uv = 2 * this->vec().cross(v); - return v + this->w() * uv + this->vec().cross(uv); -} - -template -inline Quaternion& Quaternion::operator=(const Quaternion& other) -{ - m_coeffs = other.m_coeffs; - return *this; -} - -/** Set \c *this from an angle-axis \a aa and returns a reference to \c *this - */ -template -inline Quaternion& Quaternion::operator=(const AngleAxisType& aa) -{ - Scalar ha = Scalar(0.5)*aa.angle(); // Scalar(0.5) to suppress precision loss warnings - this->w() = ei_cos(ha); - this->vec() = ei_sin(ha) * aa.axis(); - return *this; -} - -/** Set \c *this from the expression \a xpr: - * - if \a xpr is a 4x1 vector, then \a xpr is assumed to be a quaternion - * - if \a xpr is a 3x3 matrix, then \a xpr is assumed to be rotation matrix - * and \a xpr is converted to a quaternion - */ -template -template -inline Quaternion& Quaternion::operator=(const MatrixBase& xpr) -{ - ei_quaternion_assign_impl::run(*this, xpr.derived()); - return *this; -} - -/** Convert the quaternion to a 3x3 rotation matrix */ -template -inline typename Quaternion::Matrix3 -Quaternion::toRotationMatrix(void) const -{ - // NOTE if inlined, then gcc 4.2 and 4.4 get rid of the temporary (not gcc 4.3 !!) - // if not inlined then the cost of the return by value is huge ~ +35%, - // however, not inlining this function is an order of magnitude slower, so - // it has to be inlined, and so the return by value is not an issue - Matrix3 res; - - const Scalar tx = Scalar(2)*this->x(); - const Scalar ty = Scalar(2)*this->y(); - const Scalar tz = Scalar(2)*this->z(); - const Scalar twx = tx*this->w(); - const Scalar twy = ty*this->w(); - const Scalar twz = tz*this->w(); - const Scalar txx = tx*this->x(); - const Scalar txy = ty*this->x(); - const Scalar txz = tz*this->x(); - const Scalar tyy = ty*this->y(); - const Scalar tyz = tz*this->y(); - const Scalar tzz = tz*this->z(); - - res.coeffRef(0,0) = Scalar(1)-(tyy+tzz); - res.coeffRef(0,1) = txy-twz; - res.coeffRef(0,2) = txz+twy; - res.coeffRef(1,0) = txy+twz; - res.coeffRef(1,1) = Scalar(1)-(txx+tzz); - res.coeffRef(1,2) = tyz-twx; - res.coeffRef(2,0) = txz-twy; - res.coeffRef(2,1) = tyz+twx; - res.coeffRef(2,2) = Scalar(1)-(txx+tyy); - - return res; -} - -/** Sets *this to be a quaternion representing a rotation sending the vector \a a to the vector \a b. - * - * \returns a reference to *this. - * - * Note that the two input vectors do \b not have to be normalized. - */ -template -template -inline Quaternion& Quaternion::setFromTwoVectors(const MatrixBase& a, const MatrixBase& b) -{ - Vector3 v0 = a.normalized(); - Vector3 v1 = b.normalized(); - Scalar c = v0.eigen2_dot(v1); - - // if dot == 1, vectors are the same - if (ei_isApprox(c,Scalar(1))) - { - // set to identity - this->w() = 1; this->vec().setZero(); - return *this; - } - // if dot == -1, vectors are opposites - if (ei_isApprox(c,Scalar(-1))) - { - this->vec() = v0.unitOrthogonal(); - this->w() = 0; - return *this; - } - - Vector3 axis = v0.cross(v1); - Scalar s = ei_sqrt((Scalar(1)+c)*Scalar(2)); - Scalar invs = Scalar(1)/s; - this->vec() = axis * invs; - this->w() = s * Scalar(0.5); - - return *this; -} - -/** \returns the multiplicative inverse of \c *this - * Note that in most cases, i.e., if you simply want the opposite rotation, - * and/or the quaternion is normalized, then it is enough to use the conjugate. - * - * \sa Quaternion::conjugate() - */ -template -inline Quaternion Quaternion::inverse() const -{ - // FIXME should this function be called multiplicativeInverse and conjugate() be called inverse() or opposite() ?? - Scalar n2 = this->squaredNorm(); - if (n2 > 0) - return Quaternion(conjugate().coeffs() / n2); - else - { - // return an invalid result to flag the error - return Quaternion(Coefficients::Zero()); - } -} - -/** \returns the conjugate of the \c *this which is equal to the multiplicative inverse - * if the quaternion is normalized. - * The conjugate of a quaternion represents the opposite rotation. - * - * \sa Quaternion::inverse() - */ -template -inline Quaternion Quaternion::conjugate() const -{ - return Quaternion(this->w(),-this->x(),-this->y(),-this->z()); -} - -/** \returns the angle (in radian) between two rotations - * \sa eigen2_dot() - */ -template -inline Scalar Quaternion::angularDistance(const Quaternion& other) const -{ - double d = ei_abs(this->eigen2_dot(other)); - if (d>=1.0) - return 0; - return Scalar(2) * std::acos(d); -} - -/** \returns the spherical linear interpolation between the two quaternions - * \c *this and \a other at the parameter \a t - */ -template -Quaternion Quaternion::slerp(Scalar t, const Quaternion& other) const -{ - static const Scalar one = Scalar(1) - machine_epsilon(); - Scalar d = this->eigen2_dot(other); - Scalar absD = ei_abs(d); - - Scalar scale0; - Scalar scale1; - - if (absD>=one) - { - scale0 = Scalar(1) - t; - scale1 = t; - } - else - { - // theta is the angle between the 2 quaternions - Scalar theta = std::acos(absD); - Scalar sinTheta = ei_sin(theta); - - scale0 = ei_sin( ( Scalar(1) - t ) * theta) / sinTheta; - scale1 = ei_sin( ( t * theta) ) / sinTheta; - if (d<0) - scale1 = -scale1; - } - - return Quaternion(scale0 * coeffs() + scale1 * other.coeffs()); -} - -// set from a rotation matrix -template -struct ei_quaternion_assign_impl -{ - typedef typename Other::Scalar Scalar; - static inline void run(Quaternion& q, const Other& mat) - { - // This algorithm comes from "Quaternion Calculus and Fast Animation", - // Ken Shoemake, 1987 SIGGRAPH course notes - Scalar t = mat.trace(); - if (t > 0) - { - t = ei_sqrt(t + Scalar(1.0)); - q.w() = Scalar(0.5)*t; - t = Scalar(0.5)/t; - q.x() = (mat.coeff(2,1) - mat.coeff(1,2)) * t; - q.y() = (mat.coeff(0,2) - mat.coeff(2,0)) * t; - q.z() = (mat.coeff(1,0) - mat.coeff(0,1)) * t; - } - else - { - int i = 0; - if (mat.coeff(1,1) > mat.coeff(0,0)) - i = 1; - if (mat.coeff(2,2) > mat.coeff(i,i)) - i = 2; - int j = (i+1)%3; - int k = (j+1)%3; - - t = ei_sqrt(mat.coeff(i,i)-mat.coeff(j,j)-mat.coeff(k,k) + Scalar(1.0)); - q.coeffs().coeffRef(i) = Scalar(0.5) * t; - t = Scalar(0.5)/t; - q.w() = (mat.coeff(k,j)-mat.coeff(j,k))*t; - q.coeffs().coeffRef(j) = (mat.coeff(j,i)+mat.coeff(i,j))*t; - q.coeffs().coeffRef(k) = (mat.coeff(k,i)+mat.coeff(i,k))*t; - } - } -}; - -// set from a vector of coefficients assumed to be a quaternion -template -struct ei_quaternion_assign_impl -{ - typedef typename Other::Scalar Scalar; - static inline void run(Quaternion& q, const Other& vec) - { - q.coeffs() = vec; - } -}; - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/Rotation2D.h b/Eigen/src/Eigen2Support/Geometry/Rotation2D.h deleted file mode 100644 index 19b8582a1..000000000 --- a/Eigen/src/Eigen2Support/Geometry/Rotation2D.h +++ /dev/null @@ -1,145 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -/** \geometry_module \ingroup Geometry_Module - * - * \class Rotation2D - * - * \brief Represents a rotation/orientation in a 2 dimensional space. - * - * \param _Scalar the scalar type, i.e., the type of the coefficients - * - * This class is equivalent to a single scalar representing a counter clock wise rotation - * as a single angle in radian. It provides some additional features such as the automatic - * conversion from/to a 2x2 rotation matrix. Moreover this class aims to provide a similar - * interface to Quaternion in order to facilitate the writing of generic algorithms - * dealing with rotations. - * - * \sa class Quaternion, class Transform - */ -template struct ei_traits > -{ - typedef _Scalar Scalar; -}; - -template -class Rotation2D : public RotationBase,2> -{ - typedef RotationBase,2> Base; - -public: - - using Base::operator*; - - enum { Dim = 2 }; - /** the scalar type of the coefficients */ - typedef _Scalar Scalar; - typedef Matrix Vector2; - typedef Matrix Matrix2; - -protected: - - Scalar m_angle; - -public: - - /** Construct a 2D counter clock wise rotation from the angle \a a in radian. */ - inline Rotation2D(Scalar a) : m_angle(a) {} - - /** \returns the rotation angle */ - inline Scalar angle() const { return m_angle; } - - /** \returns a read-write reference to the rotation angle */ - inline Scalar& angle() { return m_angle; } - - /** \returns the inverse rotation */ - inline Rotation2D inverse() const { return -m_angle; } - - /** Concatenates two rotations */ - inline Rotation2D operator*(const Rotation2D& other) const - { return m_angle + other.m_angle; } - - /** Concatenates two rotations */ - inline Rotation2D& operator*=(const Rotation2D& other) - { return m_angle += other.m_angle; return *this; } - - /** Applies the rotation to a 2D vector */ - Vector2 operator* (const Vector2& vec) const - { return toRotationMatrix() * vec; } - - template - Rotation2D& fromRotationMatrix(const MatrixBase& m); - Matrix2 toRotationMatrix(void) const; - - /** \returns the spherical interpolation between \c *this and \a other using - * parameter \a t. It is in fact equivalent to a linear interpolation. - */ - inline Rotation2D slerp(Scalar t, const Rotation2D& other) const - { return m_angle * (1-t) + other.angle() * t; } - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { return typename internal::cast_return_type >::type(*this); } - - /** Copy constructor with scalar type conversion */ - template - inline explicit Rotation2D(const Rotation2D& other) - { - m_angle = Scalar(other.angle()); - } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const Rotation2D& other, typename NumTraits::Real prec = precision()) const - { return ei_isApprox(m_angle,other.m_angle, prec); } -}; - -/** \ingroup Geometry_Module - * single precision 2D rotation type */ -typedef Rotation2D Rotation2Df; -/** \ingroup Geometry_Module - * double precision 2D rotation type */ -typedef Rotation2D Rotation2Dd; - -/** Set \c *this from a 2x2 rotation matrix \a mat. - * In other words, this function extract the rotation angle - * from the rotation matrix. - */ -template -template -Rotation2D& Rotation2D::fromRotationMatrix(const MatrixBase& mat) -{ - EIGEN_STATIC_ASSERT(Derived::RowsAtCompileTime==2 && Derived::ColsAtCompileTime==2,YOU_MADE_A_PROGRAMMING_MISTAKE) - m_angle = ei_atan2(mat.coeff(1,0), mat.coeff(0,0)); - return *this; -} - -/** Constructs and \returns an equivalent 2x2 rotation matrix. - */ -template -typename Rotation2D::Matrix2 -Rotation2D::toRotationMatrix(void) const -{ - Scalar sinA = ei_sin(m_angle); - Scalar cosA = ei_cos(m_angle); - return (Matrix2() << cosA, -sinA, sinA, cosA).finished(); -} - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/RotationBase.h b/Eigen/src/Eigen2Support/Geometry/RotationBase.h deleted file mode 100644 index b1c8f38da..000000000 --- a/Eigen/src/Eigen2Support/Geometry/RotationBase.h +++ /dev/null @@ -1,123 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -// this file aims to contains the various representations of rotation/orientation -// in 2D and 3D space excepted Matrix and Quaternion. - -/** \class RotationBase - * - * \brief Common base class for compact rotation representations - * - * \param Derived is the derived type, i.e., a rotation type - * \param _Dim the dimension of the space - */ -template -class RotationBase -{ - public: - enum { Dim = _Dim }; - /** the scalar type of the coefficients */ - typedef typename ei_traits::Scalar Scalar; - - /** corresponding linear transformation matrix type */ - typedef Matrix RotationMatrixType; - - inline const Derived& derived() const { return *static_cast(this); } - inline Derived& derived() { return *static_cast(this); } - - /** \returns an equivalent rotation matrix */ - inline RotationMatrixType toRotationMatrix() const { return derived().toRotationMatrix(); } - - /** \returns the inverse rotation */ - inline Derived inverse() const { return derived().inverse(); } - - /** \returns the concatenation of the rotation \c *this with a translation \a t */ - inline Transform operator*(const Translation& t) const - { return toRotationMatrix() * t; } - - /** \returns the concatenation of the rotation \c *this with a scaling \a s */ - inline RotationMatrixType operator*(const Scaling& s) const - { return toRotationMatrix() * s; } - - /** \returns the concatenation of the rotation \c *this with an affine transformation \a t */ - inline Transform operator*(const Transform& t) const - { return toRotationMatrix() * t; } -}; - -/** \geometry_module - * - * Constructs a Dim x Dim rotation matrix from the rotation \a r - */ -template -template -Matrix<_Scalar, _Rows, _Cols, _Storage, _MaxRows, _MaxCols> -::Matrix(const RotationBase& r) -{ - EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(Matrix,int(OtherDerived::Dim),int(OtherDerived::Dim)) - *this = r.toRotationMatrix(); -} - -/** \geometry_module - * - * Set a Dim x Dim rotation matrix from the rotation \a r - */ -template -template -Matrix<_Scalar, _Rows, _Cols, _Storage, _MaxRows, _MaxCols>& -Matrix<_Scalar, _Rows, _Cols, _Storage, _MaxRows, _MaxCols> -::operator=(const RotationBase& r) -{ - EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(Matrix,int(OtherDerived::Dim),int(OtherDerived::Dim)) - return *this = r.toRotationMatrix(); -} - -/** \internal - * - * Helper function to return an arbitrary rotation object to a rotation matrix. - * - * \param Scalar the numeric type of the matrix coefficients - * \param Dim the dimension of the current space - * - * It returns a Dim x Dim fixed size matrix. - * - * Default specializations are provided for: - * - any scalar type (2D), - * - any matrix expression, - * - any type based on RotationBase (e.g., Quaternion, AngleAxis, Rotation2D) - * - * Currently ei_toRotationMatrix is only used by Transform. - * - * \sa class Transform, class Rotation2D, class Quaternion, class AngleAxis - */ -template -static inline Matrix ei_toRotationMatrix(const Scalar& s) -{ - EIGEN_STATIC_ASSERT(Dim==2,YOU_MADE_A_PROGRAMMING_MISTAKE) - return Rotation2D(s).toRotationMatrix(); -} - -template -static inline Matrix ei_toRotationMatrix(const RotationBase& r) -{ - return r.toRotationMatrix(); -} - -template -static inline const MatrixBase& ei_toRotationMatrix(const MatrixBase& mat) -{ - EIGEN_STATIC_ASSERT(OtherDerived::RowsAtCompileTime==Dim && OtherDerived::ColsAtCompileTime==Dim, - YOU_MADE_A_PROGRAMMING_MISTAKE) - return mat; -} - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/Scaling.h b/Eigen/src/Eigen2Support/Geometry/Scaling.h deleted file mode 100644 index b8fa6cd3f..000000000 --- a/Eigen/src/Eigen2Support/Geometry/Scaling.h +++ /dev/null @@ -1,167 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -/** \geometry_module \ingroup Geometry_Module - * - * \class Scaling - * - * \brief Represents a possibly non uniform scaling transformation - * - * \param _Scalar the scalar type, i.e., the type of the coefficients. - * \param _Dim the dimension of the space, can be a compile time value or Dynamic - * - * \note This class is not aimed to be used to store a scaling transformation, - * but rather to make easier the constructions and updates of Transform objects. - * - * \sa class Translation, class Transform - */ -template -class Scaling -{ -public: - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Dim) - /** dimension of the space */ - enum { Dim = _Dim }; - /** the scalar type of the coefficients */ - typedef _Scalar Scalar; - /** corresponding vector type */ - typedef Matrix VectorType; - /** corresponding linear transformation matrix type */ - typedef Matrix LinearMatrixType; - /** corresponding translation type */ - typedef Translation TranslationType; - /** corresponding affine transformation type */ - typedef Transform TransformType; - -protected: - - VectorType m_coeffs; - -public: - - /** Default constructor without initialization. */ - Scaling() {} - /** Constructs and initialize a uniform scaling transformation */ - explicit inline Scaling(const Scalar& s) { m_coeffs.setConstant(s); } - /** 2D only */ - inline Scaling(const Scalar& sx, const Scalar& sy) - { - ei_assert(Dim==2); - m_coeffs.x() = sx; - m_coeffs.y() = sy; - } - /** 3D only */ - inline Scaling(const Scalar& sx, const Scalar& sy, const Scalar& sz) - { - ei_assert(Dim==3); - m_coeffs.x() = sx; - m_coeffs.y() = sy; - m_coeffs.z() = sz; - } - /** Constructs and initialize the scaling transformation from a vector of scaling coefficients */ - explicit inline Scaling(const VectorType& coeffs) : m_coeffs(coeffs) {} - - const VectorType& coeffs() const { return m_coeffs; } - VectorType& coeffs() { return m_coeffs; } - - /** Concatenates two scaling */ - inline Scaling operator* (const Scaling& other) const - { return Scaling(coeffs().cwise() * other.coeffs()); } - - /** Concatenates a scaling and a translation */ - inline TransformType operator* (const TranslationType& t) const; - - /** Concatenates a scaling and an affine transformation */ - inline TransformType operator* (const TransformType& t) const; - - /** Concatenates a scaling and a linear transformation matrix */ - // TODO returns an expression - inline LinearMatrixType operator* (const LinearMatrixType& other) const - { return coeffs().asDiagonal() * other; } - - /** Concatenates a linear transformation matrix and a scaling */ - // TODO returns an expression - friend inline LinearMatrixType operator* (const LinearMatrixType& other, const Scaling& s) - { return other * s.coeffs().asDiagonal(); } - - template - inline LinearMatrixType operator*(const RotationBase& r) const - { return *this * r.toRotationMatrix(); } - - /** Applies scaling to vector */ - inline VectorType operator* (const VectorType& other) const - { return coeffs().asDiagonal() * other; } - - /** \returns the inverse scaling */ - inline Scaling inverse() const - { return Scaling(coeffs().cwise().inverse()); } - - inline Scaling& operator=(const Scaling& other) - { - m_coeffs = other.m_coeffs; - return *this; - } - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { return typename internal::cast_return_type >::type(*this); } - - /** Copy constructor with scalar type conversion */ - template - inline explicit Scaling(const Scaling& other) - { m_coeffs = other.coeffs().template cast(); } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const Scaling& other, typename NumTraits::Real prec = precision()) const - { return m_coeffs.isApprox(other.m_coeffs, prec); } - -}; - -/** \addtogroup Geometry_Module */ -//@{ -typedef Scaling Scaling2f; -typedef Scaling Scaling2d; -typedef Scaling Scaling3f; -typedef Scaling Scaling3d; -//@} - -template -inline typename Scaling::TransformType -Scaling::operator* (const TranslationType& t) const -{ - TransformType res; - res.matrix().setZero(); - res.linear().diagonal() = coeffs(); - res.translation() = m_coeffs.cwise() * t.vector(); - res(Dim,Dim) = Scalar(1); - return res; -} - -template -inline typename Scaling::TransformType -Scaling::operator* (const TransformType& t) const -{ - TransformType res = t; - res.prescale(m_coeffs); - return res; -} - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/Transform.h b/Eigen/src/Eigen2Support/Geometry/Transform.h deleted file mode 100644 index fab60b251..000000000 --- a/Eigen/src/Eigen2Support/Geometry/Transform.h +++ /dev/null @@ -1,786 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -// Note that we have to pass Dim and HDim because it is not allowed to use a template -// parameter to define a template specialization. To be more precise, in the following -// specializations, it is not allowed to use Dim+1 instead of HDim. -template< typename Other, - int Dim, - int HDim, - int OtherRows=Other::RowsAtCompileTime, - int OtherCols=Other::ColsAtCompileTime> -struct ei_transform_product_impl; - -/** \geometry_module \ingroup Geometry_Module - * - * \class Transform - * - * \brief Represents an homogeneous transformation in a N dimensional space - * - * \param _Scalar the scalar type, i.e., the type of the coefficients - * \param _Dim the dimension of the space - * - * The homography is internally represented and stored as a (Dim+1)^2 matrix which - * is available through the matrix() method. - * - * Conversion methods from/to Qt's QMatrix and QTransform are available if the - * preprocessor token EIGEN_QT_SUPPORT is defined. - * - * \sa class Matrix, class Quaternion - */ -template -class Transform -{ -public: - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Dim==Dynamic ? Dynamic : (_Dim+1)*(_Dim+1)) - enum { - Dim = _Dim, ///< space dimension in which the transformation holds - HDim = _Dim+1 ///< size of a respective homogeneous vector - }; - /** the scalar type of the coefficients */ - typedef _Scalar Scalar; - /** type of the matrix used to represent the transformation */ - typedef Matrix MatrixType; - /** type of the matrix used to represent the linear part of the transformation */ - typedef Matrix LinearMatrixType; - /** type of read/write reference to the linear part of the transformation */ - typedef Block LinearPart; - /** type of read/write reference to the linear part of the transformation */ - typedef const Block ConstLinearPart; - /** type of a vector */ - typedef Matrix VectorType; - /** type of a read/write reference to the translation part of the rotation */ - typedef Block TranslationPart; - /** type of a read/write reference to the translation part of the rotation */ - typedef const Block ConstTranslationPart; - /** corresponding translation type */ - typedef Translation TranslationType; - /** corresponding scaling transformation type */ - typedef Scaling ScalingType; - -protected: - - MatrixType m_matrix; - -public: - - /** Default constructor without initialization of the coefficients. */ - inline Transform() { } - - inline Transform(const Transform& other) - { - m_matrix = other.m_matrix; - } - - inline explicit Transform(const TranslationType& t) { *this = t; } - inline explicit Transform(const ScalingType& s) { *this = s; } - template - inline explicit Transform(const RotationBase& r) { *this = r; } - - inline Transform& operator=(const Transform& other) - { m_matrix = other.m_matrix; return *this; } - - template // MSVC 2005 will commit suicide if BigMatrix has a default value - struct construct_from_matrix - { - static inline void run(Transform *transform, const MatrixBase& other) - { - transform->matrix() = other; - } - }; - - template struct construct_from_matrix - { - static inline void run(Transform *transform, const MatrixBase& other) - { - transform->linear() = other; - transform->translation().setZero(); - transform->matrix()(Dim,Dim) = Scalar(1); - transform->matrix().template block<1,Dim>(Dim,0).setZero(); - } - }; - - /** Constructs and initializes a transformation from a Dim^2 or a (Dim+1)^2 matrix. */ - template - inline explicit Transform(const MatrixBase& other) - { - construct_from_matrix::run(this, other); - } - - /** Set \c *this from a (Dim+1)^2 matrix. */ - template - inline Transform& operator=(const MatrixBase& other) - { m_matrix = other; return *this; } - - #ifdef EIGEN_QT_SUPPORT - inline Transform(const QMatrix& other); - inline Transform& operator=(const QMatrix& other); - inline QMatrix toQMatrix(void) const; - inline Transform(const QTransform& other); - inline Transform& operator=(const QTransform& other); - inline QTransform toQTransform(void) const; - #endif - - /** shortcut for m_matrix(row,col); - * \sa MatrixBase::operaror(int,int) const */ - inline Scalar operator() (int row, int col) const { return m_matrix(row,col); } - /** shortcut for m_matrix(row,col); - * \sa MatrixBase::operaror(int,int) */ - inline Scalar& operator() (int row, int col) { return m_matrix(row,col); } - - /** \returns a read-only expression of the transformation matrix */ - inline const MatrixType& matrix() const { return m_matrix; } - /** \returns a writable expression of the transformation matrix */ - inline MatrixType& matrix() { return m_matrix; } - - /** \returns a read-only expression of the linear (linear) part of the transformation */ - inline ConstLinearPart linear() const { return m_matrix.template block(0,0); } - /** \returns a writable expression of the linear (linear) part of the transformation */ - inline LinearPart linear() { return m_matrix.template block(0,0); } - - /** \returns a read-only expression of the translation vector of the transformation */ - inline ConstTranslationPart translation() const { return m_matrix.template block(0,Dim); } - /** \returns a writable expression of the translation vector of the transformation */ - inline TranslationPart translation() { return m_matrix.template block(0,Dim); } - - /** \returns an expression of the product between the transform \c *this and a matrix expression \a other - * - * The right hand side \a other might be either: - * \li a vector of size Dim, - * \li an homogeneous vector of size Dim+1, - * \li a transformation matrix of size Dim+1 x Dim+1. - */ - // note: this function is defined here because some compilers cannot find the respective declaration - template - inline const typename ei_transform_product_impl::ResultType - operator * (const MatrixBase &other) const - { return ei_transform_product_impl::run(*this,other.derived()); } - - /** \returns the product expression of a transformation matrix \a a times a transform \a b - * The transformation matrix \a a must have a Dim+1 x Dim+1 sizes. */ - template - friend inline const typename ProductReturnType::Type - operator * (const MatrixBase &a, const Transform &b) - { return a.derived() * b.matrix(); } - - /** Contatenates two transformations */ - inline const Transform - operator * (const Transform& other) const - { return Transform(m_matrix * other.matrix()); } - - /** \sa MatrixBase::setIdentity() */ - void setIdentity() { m_matrix.setIdentity(); } - static const typename MatrixType::IdentityReturnType Identity() - { - return MatrixType::Identity(); - } - - template - inline Transform& scale(const MatrixBase &other); - - template - inline Transform& prescale(const MatrixBase &other); - - inline Transform& scale(Scalar s); - inline Transform& prescale(Scalar s); - - template - inline Transform& translate(const MatrixBase &other); - - template - inline Transform& pretranslate(const MatrixBase &other); - - template - inline Transform& rotate(const RotationType& rotation); - - template - inline Transform& prerotate(const RotationType& rotation); - - Transform& shear(Scalar sx, Scalar sy); - Transform& preshear(Scalar sx, Scalar sy); - - inline Transform& operator=(const TranslationType& t); - inline Transform& operator*=(const TranslationType& t) { return translate(t.vector()); } - inline Transform operator*(const TranslationType& t) const; - - inline Transform& operator=(const ScalingType& t); - inline Transform& operator*=(const ScalingType& s) { return scale(s.coeffs()); } - inline Transform operator*(const ScalingType& s) const; - friend inline Transform operator*(const LinearMatrixType& mat, const Transform& t) - { - Transform res = t; - res.matrix().row(Dim) = t.matrix().row(Dim); - res.matrix().template block(0,0) = (mat * t.matrix().template block(0,0)).lazy(); - return res; - } - - template - inline Transform& operator=(const RotationBase& r); - template - inline Transform& operator*=(const RotationBase& r) { return rotate(r.toRotationMatrix()); } - template - inline Transform operator*(const RotationBase& r) const; - - LinearMatrixType rotation() const; - template - void computeRotationScaling(RotationMatrixType *rotation, ScalingMatrixType *scaling) const; - template - void computeScalingRotation(ScalingMatrixType *scaling, RotationMatrixType *rotation) const; - - template - Transform& fromPositionOrientationScale(const MatrixBase &position, - const OrientationType& orientation, const MatrixBase &scale); - - inline const MatrixType inverse(TransformTraits traits = Affine) const; - - /** \returns a const pointer to the column major internal matrix */ - const Scalar* data() const { return m_matrix.data(); } - /** \returns a non-const pointer to the column major internal matrix */ - Scalar* data() { return m_matrix.data(); } - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { return typename internal::cast_return_type >::type(*this); } - - /** Copy constructor with scalar type conversion */ - template - inline explicit Transform(const Transform& other) - { m_matrix = other.matrix().template cast(); } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const Transform& other, typename NumTraits::Real prec = precision()) const - { return m_matrix.isApprox(other.m_matrix, prec); } - - #ifdef EIGEN_TRANSFORM_PLUGIN - #include EIGEN_TRANSFORM_PLUGIN - #endif - -protected: - -}; - -/** \ingroup Geometry_Module */ -typedef Transform Transform2f; -/** \ingroup Geometry_Module */ -typedef Transform Transform3f; -/** \ingroup Geometry_Module */ -typedef Transform Transform2d; -/** \ingroup Geometry_Module */ -typedef Transform Transform3d; - -/************************** -*** Optional QT support *** -**************************/ - -#ifdef EIGEN_QT_SUPPORT -/** Initialises \c *this from a QMatrix assuming the dimension is 2. - * - * This function is available only if the token EIGEN_QT_SUPPORT is defined. - */ -template -Transform::Transform(const QMatrix& other) -{ - *this = other; -} - -/** Set \c *this from a QMatrix assuming the dimension is 2. - * - * This function is available only if the token EIGEN_QT_SUPPORT is defined. - */ -template -Transform& Transform::operator=(const QMatrix& other) -{ - EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE) - m_matrix << other.m11(), other.m21(), other.dx(), - other.m12(), other.m22(), other.dy(), - 0, 0, 1; - return *this; -} - -/** \returns a QMatrix from \c *this assuming the dimension is 2. - * - * \warning this convertion might loss data if \c *this is not affine - * - * This function is available only if the token EIGEN_QT_SUPPORT is defined. - */ -template -QMatrix Transform::toQMatrix(void) const -{ - EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE) - return QMatrix(m_matrix.coeff(0,0), m_matrix.coeff(1,0), - m_matrix.coeff(0,1), m_matrix.coeff(1,1), - m_matrix.coeff(0,2), m_matrix.coeff(1,2)); -} - -/** Initialises \c *this from a QTransform assuming the dimension is 2. - * - * This function is available only if the token EIGEN_QT_SUPPORT is defined. - */ -template -Transform::Transform(const QTransform& other) -{ - *this = other; -} - -/** Set \c *this from a QTransform assuming the dimension is 2. - * - * This function is available only if the token EIGEN_QT_SUPPORT is defined. - */ -template -Transform& Transform::operator=(const QTransform& other) -{ - EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE) - m_matrix << other.m11(), other.m21(), other.dx(), - other.m12(), other.m22(), other.dy(), - other.m13(), other.m23(), other.m33(); - return *this; -} - -/** \returns a QTransform from \c *this assuming the dimension is 2. - * - * This function is available only if the token EIGEN_QT_SUPPORT is defined. - */ -template -QTransform Transform::toQTransform(void) const -{ - EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE) - return QTransform(m_matrix.coeff(0,0), m_matrix.coeff(1,0), m_matrix.coeff(2,0), - m_matrix.coeff(0,1), m_matrix.coeff(1,1), m_matrix.coeff(2,1), - m_matrix.coeff(0,2), m_matrix.coeff(1,2), m_matrix.coeff(2,2)); -} -#endif - -/********************* -*** Procedural API *** -*********************/ - -/** Applies on the right the non uniform scale transformation represented - * by the vector \a other to \c *this and returns a reference to \c *this. - * \sa prescale() - */ -template -template -Transform& -Transform::scale(const MatrixBase &other) -{ - EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim)) - linear() = (linear() * other.asDiagonal()).lazy(); - return *this; -} - -/** Applies on the right a uniform scale of a factor \a c to \c *this - * and returns a reference to \c *this. - * \sa prescale(Scalar) - */ -template -inline Transform& Transform::scale(Scalar s) -{ - linear() *= s; - return *this; -} - -/** Applies on the left the non uniform scale transformation represented - * by the vector \a other to \c *this and returns a reference to \c *this. - * \sa scale() - */ -template -template -Transform& -Transform::prescale(const MatrixBase &other) -{ - EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim)) - m_matrix.template block(0,0) = (other.asDiagonal() * m_matrix.template block(0,0)).lazy(); - return *this; -} - -/** Applies on the left a uniform scale of a factor \a c to \c *this - * and returns a reference to \c *this. - * \sa scale(Scalar) - */ -template -inline Transform& Transform::prescale(Scalar s) -{ - m_matrix.template corner(TopLeft) *= s; - return *this; -} - -/** Applies on the right the translation matrix represented by the vector \a other - * to \c *this and returns a reference to \c *this. - * \sa pretranslate() - */ -template -template -Transform& -Transform::translate(const MatrixBase &other) -{ - EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim)) - translation() += linear() * other; - return *this; -} - -/** Applies on the left the translation matrix represented by the vector \a other - * to \c *this and returns a reference to \c *this. - * \sa translate() - */ -template -template -Transform& -Transform::pretranslate(const MatrixBase &other) -{ - EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim)) - translation() += other; - return *this; -} - -/** Applies on the right the rotation represented by the rotation \a rotation - * to \c *this and returns a reference to \c *this. - * - * The template parameter \a RotationType is the type of the rotation which - * must be known by ei_toRotationMatrix<>. - * - * Natively supported types includes: - * - any scalar (2D), - * - a Dim x Dim matrix expression, - * - a Quaternion (3D), - * - a AngleAxis (3D) - * - * This mechanism is easily extendable to support user types such as Euler angles, - * or a pair of Quaternion for 4D rotations. - * - * \sa rotate(Scalar), class Quaternion, class AngleAxis, prerotate(RotationType) - */ -template -template -Transform& -Transform::rotate(const RotationType& rotation) -{ - linear() *= ei_toRotationMatrix(rotation); - return *this; -} - -/** Applies on the left the rotation represented by the rotation \a rotation - * to \c *this and returns a reference to \c *this. - * - * See rotate() for further details. - * - * \sa rotate() - */ -template -template -Transform& -Transform::prerotate(const RotationType& rotation) -{ - m_matrix.template block(0,0) = ei_toRotationMatrix(rotation) - * m_matrix.template block(0,0); - return *this; -} - -/** Applies on the right the shear transformation represented - * by the vector \a other to \c *this and returns a reference to \c *this. - * \warning 2D only. - * \sa preshear() - */ -template -Transform& -Transform::shear(Scalar sx, Scalar sy) -{ - EIGEN_STATIC_ASSERT(int(Dim)==2, YOU_MADE_A_PROGRAMMING_MISTAKE) - VectorType tmp = linear().col(0)*sy + linear().col(1); - linear() << linear().col(0) + linear().col(1)*sx, tmp; - return *this; -} - -/** Applies on the left the shear transformation represented - * by the vector \a other to \c *this and returns a reference to \c *this. - * \warning 2D only. - * \sa shear() - */ -template -Transform& -Transform::preshear(Scalar sx, Scalar sy) -{ - EIGEN_STATIC_ASSERT(int(Dim)==2, YOU_MADE_A_PROGRAMMING_MISTAKE) - m_matrix.template block(0,0) = LinearMatrixType(1, sx, sy, 1) * m_matrix.template block(0,0); - return *this; -} - -/****************************************************** -*** Scaling, Translation and Rotation compatibility *** -******************************************************/ - -template -inline Transform& Transform::operator=(const TranslationType& t) -{ - linear().setIdentity(); - translation() = t.vector(); - m_matrix.template block<1,Dim>(Dim,0).setZero(); - m_matrix(Dim,Dim) = Scalar(1); - return *this; -} - -template -inline Transform Transform::operator*(const TranslationType& t) const -{ - Transform res = *this; - res.translate(t.vector()); - return res; -} - -template -inline Transform& Transform::operator=(const ScalingType& s) -{ - m_matrix.setZero(); - linear().diagonal() = s.coeffs(); - m_matrix.coeffRef(Dim,Dim) = Scalar(1); - return *this; -} - -template -inline Transform Transform::operator*(const ScalingType& s) const -{ - Transform res = *this; - res.scale(s.coeffs()); - return res; -} - -template -template -inline Transform& Transform::operator=(const RotationBase& r) -{ - linear() = ei_toRotationMatrix(r); - translation().setZero(); - m_matrix.template block<1,Dim>(Dim,0).setZero(); - m_matrix.coeffRef(Dim,Dim) = Scalar(1); - return *this; -} - -template -template -inline Transform Transform::operator*(const RotationBase& r) const -{ - Transform res = *this; - res.rotate(r.derived()); - return res; -} - -/************************ -*** Special functions *** -************************/ - -/** \returns the rotation part of the transformation - * \nonstableyet - * - * \svd_module - * - * \sa computeRotationScaling(), computeScalingRotation(), class SVD - */ -template -typename Transform::LinearMatrixType -Transform::rotation() const -{ - LinearMatrixType result; - computeRotationScaling(&result, (LinearMatrixType*)0); - return result; -} - - -/** decomposes the linear part of the transformation as a product rotation x scaling, the scaling being - * not necessarily positive. - * - * If either pointer is zero, the corresponding computation is skipped. - * - * \nonstableyet - * - * \svd_module - * - * \sa computeScalingRotation(), rotation(), class SVD - */ -template -template -void Transform::computeRotationScaling(RotationMatrixType *rotation, ScalingMatrixType *scaling) const -{ - JacobiSVD svd(linear(), ComputeFullU|ComputeFullV); - Scalar x = (svd.matrixU() * svd.matrixV().adjoint()).determinant(); // so x has absolute value 1 - Matrix sv(svd.singularValues()); - sv.coeffRef(0) *= x; - if(scaling) - { - scaling->noalias() = svd.matrixV() * sv.asDiagonal() * svd.matrixV().adjoint(); - } - if(rotation) - { - LinearMatrixType m(svd.matrixU()); - m.col(0) /= x; - rotation->noalias() = m * svd.matrixV().adjoint(); - } -} - -/** decomposes the linear part of the transformation as a product rotation x scaling, the scaling being - * not necessarily positive. - * - * If either pointer is zero, the corresponding computation is skipped. - * - * \nonstableyet - * - * \svd_module - * - * \sa computeRotationScaling(), rotation(), class SVD - */ -template -template -void Transform::computeScalingRotation(ScalingMatrixType *scaling, RotationMatrixType *rotation) const -{ - JacobiSVD svd(linear(), ComputeFullU|ComputeFullV); - Scalar x = (svd.matrixU() * svd.matrixV().adjoint()).determinant(); // so x has absolute value 1 - Matrix sv(svd.singularValues()); - sv.coeffRef(0) *= x; - if(scaling) - { - scaling->noalias() = svd.matrixU() * sv.asDiagonal() * svd.matrixU().adjoint(); - } - if(rotation) - { - LinearMatrixType m(svd.matrixU()); - m.col(0) /= x; - rotation->noalias() = m * svd.matrixV().adjoint(); - } -} - -/** Convenient method to set \c *this from a position, orientation and scale - * of a 3D object. - */ -template -template -Transform& -Transform::fromPositionOrientationScale(const MatrixBase &position, - const OrientationType& orientation, const MatrixBase &scale) -{ - linear() = ei_toRotationMatrix(orientation); - linear() *= scale.asDiagonal(); - translation() = position; - m_matrix.template block<1,Dim>(Dim,0).setZero(); - m_matrix(Dim,Dim) = Scalar(1); - return *this; -} - -/** \nonstableyet - * - * \returns the inverse transformation matrix according to some given knowledge - * on \c *this. - * - * \param traits allows to optimize the inversion process when the transformion - * is known to be not a general transformation. The possible values are: - * - Projective if the transformation is not necessarily affine, i.e., if the - * last row is not guaranteed to be [0 ... 0 1] - * - Affine is the default, the last row is assumed to be [0 ... 0 1] - * - Isometry if the transformation is only a concatenations of translations - * and rotations. - * - * \warning unless \a traits is always set to NoShear or NoScaling, this function - * requires the generic inverse method of MatrixBase defined in the LU module. If - * you forget to include this module, then you will get hard to debug linking errors. - * - * \sa MatrixBase::inverse() - */ -template -inline const typename Transform::MatrixType -Transform::inverse(TransformTraits traits) const -{ - if (traits == Projective) - { - return m_matrix.inverse(); - } - else - { - MatrixType res; - if (traits == Affine) - { - res.template corner(TopLeft) = linear().inverse(); - } - else if (traits == Isometry) - { - res.template corner(TopLeft) = linear().transpose(); - } - else - { - ei_assert("invalid traits value in Transform::inverse()"); - } - // translation and remaining parts - res.template corner(TopRight) = - res.template corner(TopLeft) * translation(); - res.template corner<1,Dim>(BottomLeft).setZero(); - res.coeffRef(Dim,Dim) = Scalar(1); - return res; - } -} - -/***************************************************** -*** Specializations of operator* with a MatrixBase *** -*****************************************************/ - -template -struct ei_transform_product_impl -{ - typedef Transform TransformType; - typedef typename TransformType::MatrixType MatrixType; - typedef typename ProductReturnType::Type ResultType; - static ResultType run(const TransformType& tr, const Other& other) - { return tr.matrix() * other; } -}; - -template -struct ei_transform_product_impl -{ - typedef Transform TransformType; - typedef typename TransformType::MatrixType MatrixType; - typedef TransformType ResultType; - static ResultType run(const TransformType& tr, const Other& other) - { - TransformType res; - res.translation() = tr.translation(); - res.matrix().row(Dim) = tr.matrix().row(Dim); - res.linear() = (tr.linear() * other).lazy(); - return res; - } -}; - -template -struct ei_transform_product_impl -{ - typedef Transform TransformType; - typedef typename TransformType::MatrixType MatrixType; - typedef typename ProductReturnType::Type ResultType; - static ResultType run(const TransformType& tr, const Other& other) - { return tr.matrix() * other; } -}; - -template -struct ei_transform_product_impl -{ - typedef typename Other::Scalar Scalar; - typedef Transform TransformType; - typedef Matrix ResultType; - static ResultType run(const TransformType& tr, const Other& other) - { return ((tr.linear() * other) + tr.translation()) - * (Scalar(1) / ( (tr.matrix().template block<1,Dim>(Dim,0) * other).coeff(0) + tr.matrix().coeff(Dim,Dim))); } -}; - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/Geometry/Translation.h b/Eigen/src/Eigen2Support/Geometry/Translation.h deleted file mode 100644 index 2b9859f6f..000000000 --- a/Eigen/src/Eigen2Support/Geometry/Translation.h +++ /dev/null @@ -1,184 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// no include guard, we'll include this twice from All.h from Eigen2Support, and it's internal anyway - -namespace Eigen { - -/** \geometry_module \ingroup Geometry_Module - * - * \class Translation - * - * \brief Represents a translation transformation - * - * \param _Scalar the scalar type, i.e., the type of the coefficients. - * \param _Dim the dimension of the space, can be a compile time value or Dynamic - * - * \note This class is not aimed to be used to store a translation transformation, - * but rather to make easier the constructions and updates of Transform objects. - * - * \sa class Scaling, class Transform - */ -template -class Translation -{ -public: - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Dim) - /** dimension of the space */ - enum { Dim = _Dim }; - /** the scalar type of the coefficients */ - typedef _Scalar Scalar; - /** corresponding vector type */ - typedef Matrix VectorType; - /** corresponding linear transformation matrix type */ - typedef Matrix LinearMatrixType; - /** corresponding scaling transformation type */ - typedef Scaling ScalingType; - /** corresponding affine transformation type */ - typedef Transform TransformType; - -protected: - - VectorType m_coeffs; - -public: - - /** Default constructor without initialization. */ - Translation() {} - /** */ - inline Translation(const Scalar& sx, const Scalar& sy) - { - ei_assert(Dim==2); - m_coeffs.x() = sx; - m_coeffs.y() = sy; - } - /** */ - inline Translation(const Scalar& sx, const Scalar& sy, const Scalar& sz) - { - ei_assert(Dim==3); - m_coeffs.x() = sx; - m_coeffs.y() = sy; - m_coeffs.z() = sz; - } - /** Constructs and initialize the scaling transformation from a vector of scaling coefficients */ - explicit inline Translation(const VectorType& vector) : m_coeffs(vector) {} - - const VectorType& vector() const { return m_coeffs; } - VectorType& vector() { return m_coeffs; } - - /** Concatenates two translation */ - inline Translation operator* (const Translation& other) const - { return Translation(m_coeffs + other.m_coeffs); } - - /** Concatenates a translation and a scaling */ - inline TransformType operator* (const ScalingType& other) const; - - /** Concatenates a translation and a linear transformation */ - inline TransformType operator* (const LinearMatrixType& linear) const; - - template - inline TransformType operator*(const RotationBase& r) const - { return *this * r.toRotationMatrix(); } - - /** Concatenates a linear transformation and a translation */ - // its a nightmare to define a templated friend function outside its declaration - friend inline TransformType operator* (const LinearMatrixType& linear, const Translation& t) - { - TransformType res; - res.matrix().setZero(); - res.linear() = linear; - res.translation() = linear * t.m_coeffs; - res.matrix().row(Dim).setZero(); - res(Dim,Dim) = Scalar(1); - return res; - } - - /** Concatenates a translation and an affine transformation */ - inline TransformType operator* (const TransformType& t) const; - - /** Applies translation to vector */ - inline VectorType operator* (const VectorType& other) const - { return m_coeffs + other; } - - /** \returns the inverse translation (opposite) */ - Translation inverse() const { return Translation(-m_coeffs); } - - Translation& operator=(const Translation& other) - { - m_coeffs = other.m_coeffs; - return *this; - } - - /** \returns \c *this with scalar type casted to \a NewScalarType - * - * Note that if \a NewScalarType is equal to the current scalar type of \c *this - * then this function smartly returns a const reference to \c *this. - */ - template - inline typename internal::cast_return_type >::type cast() const - { return typename internal::cast_return_type >::type(*this); } - - /** Copy constructor with scalar type conversion */ - template - inline explicit Translation(const Translation& other) - { m_coeffs = other.vector().template cast(); } - - /** \returns \c true if \c *this is approximately equal to \a other, within the precision - * determined by \a prec. - * - * \sa MatrixBase::isApprox() */ - bool isApprox(const Translation& other, typename NumTraits::Real prec = precision()) const - { return m_coeffs.isApprox(other.m_coeffs, prec); } - -}; - -/** \addtogroup Geometry_Module */ -//@{ -typedef Translation Translation2f; -typedef Translation Translation2d; -typedef Translation Translation3f; -typedef Translation Translation3d; -//@} - - -template -inline typename Translation::TransformType -Translation::operator* (const ScalingType& other) const -{ - TransformType res; - res.matrix().setZero(); - res.linear().diagonal() = other.coeffs(); - res.translation() = m_coeffs; - res(Dim,Dim) = Scalar(1); - return res; -} - -template -inline typename Translation::TransformType -Translation::operator* (const LinearMatrixType& linear) const -{ - TransformType res; - res.matrix().setZero(); - res.linear() = linear; - res.translation() = m_coeffs; - res.matrix().row(Dim).setZero(); - res(Dim,Dim) = Scalar(1); - return res; -} - -template -inline typename Translation::TransformType -Translation::operator* (const TransformType& t) const -{ - TransformType res = t; - res.pretranslate(m_coeffs); - return res; -} - -} // end namespace Eigen diff --git a/Eigen/src/Eigen2Support/LU.h b/Eigen/src/Eigen2Support/LU.h deleted file mode 100644 index 49f19ad76..000000000 --- a/Eigen/src/Eigen2Support/LU.h +++ /dev/null @@ -1,120 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2011 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_LU_H -#define EIGEN2_LU_H - -namespace Eigen { - -template -class LU : public FullPivLU -{ - public: - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix IntRowVectorType; - typedef Matrix IntColVectorType; - typedef Matrix RowVectorType; - typedef Matrix ColVectorType; - - typedef Matrix KernelResultType; - - typedef Matrix ImageResultType; - - typedef FullPivLU Base; - - template - explicit LU(const T& t) : Base(t), m_originalMatrix(t) {} - - template - bool solve(const MatrixBase& b, ResultType *result) const - { - *result = static_cast(this)->solve(b); - return true; - } - - template - inline void computeInverse(ResultType *result) const - { - solve(MatrixType::Identity(this->rows(), this->cols()), result); - } - - template - void computeKernel(KernelMatrixType *result) const - { - *result = static_cast(this)->kernel(); - } - - template - void computeImage(ImageMatrixType *result) const - { - *result = static_cast(this)->image(m_originalMatrix); - } - - const ImageResultType image() const - { - return static_cast(this)->image(m_originalMatrix); - } - - const MatrixType& m_originalMatrix; -}; - -#if EIGEN2_SUPPORT_STAGE < STAGE20_RESOLVE_API_CONFLICTS -/** \lu_module - * - * Synonym of partialPivLu(). - * - * \return the partial-pivoting LU decomposition of \c *this. - * - * \sa class PartialPivLU - */ -template -inline const LU::PlainObject> -MatrixBase::lu() const -{ - return LU(eval()); -} -#endif - -#ifdef EIGEN2_SUPPORT -/** \lu_module - * - * Synonym of partialPivLu(). - * - * \return the partial-pivoting LU decomposition of \c *this. - * - * \sa class PartialPivLU - */ -template -inline const LU::PlainObject> -MatrixBase::eigen2_lu() const -{ - return LU(eval()); -} -#endif - -} // end namespace Eigen - -#endif // EIGEN2_LU_H diff --git a/Eigen/src/Eigen2Support/Lazy.h b/Eigen/src/Eigen2Support/Lazy.h deleted file mode 100644 index 593fc78e6..000000000 --- a/Eigen/src/Eigen2Support/Lazy.h +++ /dev/null @@ -1,71 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_LAZY_H -#define EIGEN_LAZY_H - -namespace Eigen { - -/** \deprecated it is only used by lazy() which is deprecated - * - * \returns an expression of *this with added flags - * - * Example: \include MatrixBase_marked.cpp - * Output: \verbinclude MatrixBase_marked.out - * - * \sa class Flagged, extract(), part() - */ -template -template -inline const Flagged -MatrixBase::marked() const -{ - return derived(); -} - -/** \deprecated use MatrixBase::noalias() - * - * \returns an expression of *this with the EvalBeforeAssigningBit flag removed. - * - * Example: \include MatrixBase_lazy.cpp - * Output: \verbinclude MatrixBase_lazy.out - * - * \sa class Flagged, marked() - */ -template -inline const Flagged -MatrixBase::lazy() const -{ - return derived(); -} - - -/** \internal - * Overloaded to perform an efficient C += (A*B).lazy() */ -template -template -Derived& MatrixBase::operator+=(const Flagged, 0, - EvalBeforeAssigningBit>& other) -{ - other._expression().derived().addTo(derived()); return derived(); -} - -/** \internal - * Overloaded to perform an efficient C -= (A*B).lazy() */ -template -template -Derived& MatrixBase::operator-=(const Flagged, 0, - EvalBeforeAssigningBit>& other) -{ - other._expression().derived().subTo(derived()); return derived(); -} - -} // end namespace Eigen - -#endif // EIGEN_LAZY_H diff --git a/Eigen/src/Eigen2Support/LeastSquares.h b/Eigen/src/Eigen2Support/LeastSquares.h deleted file mode 100644 index 0e6fdb488..000000000 --- a/Eigen/src/Eigen2Support/LeastSquares.h +++ /dev/null @@ -1,170 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2006-2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_LEASTSQUARES_H -#define EIGEN2_LEASTSQUARES_H - -namespace Eigen { - -/** \ingroup LeastSquares_Module - * - * \leastsquares_module - * - * For a set of points, this function tries to express - * one of the coords as a linear (affine) function of the other coords. - * - * This is best explained by an example. This function works in full - * generality, for points in a space of arbitrary dimension, and also over - * the complex numbers, but for this example we will work in dimension 3 - * over the real numbers (doubles). - * - * So let us work with the following set of 5 points given by their - * \f$(x,y,z)\f$ coordinates: - * @code - Vector3d points[5]; - points[0] = Vector3d( 3.02, 6.89, -4.32 ); - points[1] = Vector3d( 2.01, 5.39, -3.79 ); - points[2] = Vector3d( 2.41, 6.01, -4.01 ); - points[3] = Vector3d( 2.09, 5.55, -3.86 ); - points[4] = Vector3d( 2.58, 6.32, -4.10 ); - * @endcode - * Suppose that we want to express the second coordinate (\f$y\f$) as a linear - * expression in \f$x\f$ and \f$z\f$, that is, - * \f[ y=ax+bz+c \f] - * for some constants \f$a,b,c\f$. Thus, we want to find the best possible - * constants \f$a,b,c\f$ so that the plane of equation \f$y=ax+bz+c\f$ fits - * best the five above points. To do that, call this function as follows: - * @code - Vector3d coeffs; // will store the coefficients a, b, c - linearRegression( - 5, - &points, - &coeffs, - 1 // the coord to express as a function of - // the other ones. 0 means x, 1 means y, 2 means z. - ); - * @endcode - * Now the vector \a coeffs is approximately - * \f$( 0.495 , -1.927 , -2.906 )\f$. - * Thus, we get \f$a=0.495, b = -1.927, c = -2.906\f$. Let us check for - * instance how near points[0] is from the plane of equation \f$y=ax+bz+c\f$. - * Looking at the coords of points[0], we see that: - * \f[ax+bz+c = 0.495 * 3.02 + (-1.927) * (-4.32) + (-2.906) = 6.91.\f] - * On the other hand, we have \f$y=6.89\f$. We see that the values - * \f$6.91\f$ and \f$6.89\f$ - * are near, so points[0] is very near the plane of equation \f$y=ax+bz+c\f$. - * - * Let's now describe precisely the parameters: - * @param numPoints the number of points - * @param points the array of pointers to the points on which to perform the linear regression - * @param result pointer to the vector in which to store the result. - This vector must be of the same type and size as the - data points. The meaning of its coords is as follows. - For brevity, let \f$n=Size\f$, - \f$r_i=result[i]\f$, - and \f$f=funcOfOthers\f$. Denote by - \f$x_0,\ldots,x_{n-1}\f$ - the n coordinates in the n-dimensional space. - Then the resulting equation is: - \f[ x_f = r_0 x_0 + \cdots + r_{f-1}x_{f-1} - + r_{f+1}x_{f+1} + \cdots + r_{n-1}x_{n-1} + r_n. \f] - * @param funcOfOthers Determines which coord to express as a function of the - others. Coords are numbered starting from 0, so that a - value of 0 means \f$x\f$, 1 means \f$y\f$, - 2 means \f$z\f$, ... - * - * \sa fitHyperplane() - */ -template -void linearRegression(int numPoints, - VectorType **points, - VectorType *result, - int funcOfOthers ) -{ - typedef typename VectorType::Scalar Scalar; - typedef Hyperplane HyperplaneType; - const int size = points[0]->size(); - result->resize(size); - HyperplaneType h(size); - fitHyperplane(numPoints, points, &h); - for(int i = 0; i < funcOfOthers; i++) - result->coeffRef(i) = - h.coeffs()[i] / h.coeffs()[funcOfOthers]; - for(int i = funcOfOthers; i < size; i++) - result->coeffRef(i) = - h.coeffs()[i+1] / h.coeffs()[funcOfOthers]; -} - -/** \ingroup LeastSquares_Module - * - * \leastsquares_module - * - * This function is quite similar to linearRegression(), so we refer to the - * documentation of this function and only list here the differences. - * - * The main difference from linearRegression() is that this function doesn't - * take a \a funcOfOthers argument. Instead, it finds a general equation - * of the form - * \f[ r_0 x_0 + \cdots + r_{n-1}x_{n-1} + r_n = 0, \f] - * where \f$n=Size\f$, \f$r_i=retCoefficients[i]\f$, and we denote by - * \f$x_0,\ldots,x_{n-1}\f$ the n coordinates in the n-dimensional space. - * - * Thus, the vector \a retCoefficients has size \f$n+1\f$, which is another - * difference from linearRegression(). - * - * In practice, this function performs an hyper-plane fit in a total least square sense - * via the following steps: - * 1 - center the data to the mean - * 2 - compute the covariance matrix - * 3 - pick the eigenvector corresponding to the smallest eigenvalue of the covariance matrix - * The ratio of the smallest eigenvalue and the second one gives us a hint about the relevance - * of the solution. This value is optionally returned in \a soundness. - * - * \sa linearRegression() - */ -template -void fitHyperplane(int numPoints, - VectorType **points, - HyperplaneType *result, - typename NumTraits::Real* soundness = 0) -{ - typedef typename VectorType::Scalar Scalar; - typedef Matrix CovMatrixType; - EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorType) - ei_assert(numPoints >= 1); - int size = points[0]->size(); - ei_assert(size+1 == result->coeffs().size()); - - // compute the mean of the data - VectorType mean = VectorType::Zero(size); - for(int i = 0; i < numPoints; ++i) - mean += *(points[i]); - mean /= numPoints; - - // compute the covariance matrix - CovMatrixType covMat = CovMatrixType::Zero(size, size); - VectorType remean = VectorType::Zero(size); - for(int i = 0; i < numPoints; ++i) - { - VectorType diff = (*(points[i]) - mean).conjugate(); - covMat += diff * diff.adjoint(); - } - - // now we just have to pick the eigen vector with smallest eigen value - SelfAdjointEigenSolver eig(covMat); - result->normal() = eig.eigenvectors().col(0); - if (soundness) - *soundness = eig.eigenvalues().coeff(0)/eig.eigenvalues().coeff(1); - - // let's compute the constant coefficient such that the - // plane pass trough the mean point: - result->offset() = - (result->normal().cwise()* mean).sum(); -} - -} // end namespace Eigen - -#endif // EIGEN2_LEASTSQUARES_H diff --git a/Eigen/src/Eigen2Support/Macros.h b/Eigen/src/Eigen2Support/Macros.h deleted file mode 100644 index 351c32afb..000000000 --- a/Eigen/src/Eigen2Support/Macros.h +++ /dev/null @@ -1,20 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2011 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_MACROS_H -#define EIGEN2_MACROS_H - -#define ei_assert eigen_assert -#define ei_internal_assert eigen_internal_assert - -#define EIGEN_ALIGN_128 EIGEN_ALIGN16 - -#define EIGEN_ARCH_WANTS_ALIGNMENT EIGEN_ALIGN_STATICALLY - -#endif // EIGEN2_MACROS_H diff --git a/Eigen/src/Eigen2Support/MathFunctions.h b/Eigen/src/Eigen2Support/MathFunctions.h deleted file mode 100644 index 3544af253..000000000 --- a/Eigen/src/Eigen2Support/MathFunctions.h +++ /dev/null @@ -1,57 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2010 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_MATH_FUNCTIONS_H -#define EIGEN2_MATH_FUNCTIONS_H - -namespace Eigen { - -template inline typename NumTraits::Real ei_real(const T& x) { return numext::real(x); } -template inline typename NumTraits::Real ei_imag(const T& x) { return numext::imag(x); } -template inline T ei_conj(const T& x) { return numext::conj(x); } -template inline typename NumTraits::Real ei_abs (const T& x) { using std::abs; return abs(x); } -template inline typename NumTraits::Real ei_abs2(const T& x) { return numext::abs2(x); } -template inline T ei_sqrt(const T& x) { using std::sqrt; return sqrt(x); } -template inline T ei_exp (const T& x) { using std::exp; return exp(x); } -template inline T ei_log (const T& x) { using std::log; return log(x); } -template inline T ei_sin (const T& x) { using std::sin; return sin(x); } -template inline T ei_cos (const T& x) { using std::cos; return cos(x); } -template inline T ei_atan2(const T& x,const T& y) { using std::atan2; return atan2(x,y); } -template inline T ei_pow (const T& x,const T& y) { return numext::pow(x,y); } -template inline T ei_random () { return internal::random(); } -template inline T ei_random (const T& x, const T& y) { return internal::random(x, y); } - -template inline T precision () { return NumTraits::dummy_precision(); } -template inline T machine_epsilon () { return NumTraits::epsilon(); } - - -template -inline bool ei_isMuchSmallerThan(const Scalar& x, const OtherScalar& y, - typename NumTraits::Real precision = NumTraits::dummy_precision()) -{ - return internal::isMuchSmallerThan(x, y, precision); -} - -template -inline bool ei_isApprox(const Scalar& x, const Scalar& y, - typename NumTraits::Real precision = NumTraits::dummy_precision()) -{ - return internal::isApprox(x, y, precision); -} - -template -inline bool ei_isApproxOrLessThan(const Scalar& x, const Scalar& y, - typename NumTraits::Real precision = NumTraits::dummy_precision()) -{ - return internal::isApproxOrLessThan(x, y, precision); -} - -} // end namespace Eigen - -#endif // EIGEN2_MATH_FUNCTIONS_H diff --git a/Eigen/src/Eigen2Support/Memory.h b/Eigen/src/Eigen2Support/Memory.h deleted file mode 100644 index f86372b6b..000000000 --- a/Eigen/src/Eigen2Support/Memory.h +++ /dev/null @@ -1,45 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2011 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_MEMORY_H -#define EIGEN2_MEMORY_H - -namespace Eigen { - -inline void* ei_aligned_malloc(size_t size) { return internal::aligned_malloc(size); } -inline void ei_aligned_free(void *ptr) { internal::aligned_free(ptr); } -inline void* ei_aligned_realloc(void *ptr, size_t new_size, size_t old_size) { return internal::aligned_realloc(ptr, new_size, old_size); } -inline void* ei_handmade_aligned_malloc(size_t size) { return internal::handmade_aligned_malloc(size); } -inline void ei_handmade_aligned_free(void *ptr) { internal::handmade_aligned_free(ptr); } - -template inline void* ei_conditional_aligned_malloc(size_t size) -{ - return internal::conditional_aligned_malloc(size); -} -template inline void ei_conditional_aligned_free(void *ptr) -{ - internal::conditional_aligned_free(ptr); -} -template inline void* ei_conditional_aligned_realloc(void* ptr, size_t new_size, size_t old_size) -{ - return internal::conditional_aligned_realloc(ptr, new_size, old_size); -} - -template inline T* ei_aligned_new(size_t size) -{ - return internal::aligned_new(size); -} -template inline void ei_aligned_delete(T *ptr, size_t size) -{ - return internal::aligned_delete(ptr, size); -} - -} // end namespace Eigen - -#endif // EIGEN2_MACROS_H diff --git a/Eigen/src/Eigen2Support/Meta.h b/Eigen/src/Eigen2Support/Meta.h deleted file mode 100644 index fa37cfc96..000000000 --- a/Eigen/src/Eigen2Support/Meta.h +++ /dev/null @@ -1,75 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2011 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_META_H -#define EIGEN2_META_H - -namespace Eigen { - -template -struct ei_traits : internal::traits -{}; - -struct ei_meta_true { enum { ret = 1 }; }; -struct ei_meta_false { enum { ret = 0 }; }; - -template -struct ei_meta_if { typedef Then ret; }; - -template -struct ei_meta_if { typedef Else ret; }; - -template struct ei_is_same_type { enum { ret = 0 }; }; -template struct ei_is_same_type { enum { ret = 1 }; }; - -template struct ei_unref { typedef T type; }; -template struct ei_unref { typedef T type; }; - -template struct ei_unpointer { typedef T type; }; -template struct ei_unpointer { typedef T type; }; -template struct ei_unpointer { typedef T type; }; - -template struct ei_unconst { typedef T type; }; -template struct ei_unconst { typedef T type; }; -template struct ei_unconst { typedef T & type; }; -template struct ei_unconst { typedef T * type; }; - -template struct ei_cleantype { typedef T type; }; -template struct ei_cleantype { typedef typename ei_cleantype::type type; }; -template struct ei_cleantype { typedef typename ei_cleantype::type type; }; -template struct ei_cleantype { typedef typename ei_cleantype::type type; }; -template struct ei_cleantype { typedef typename ei_cleantype::type type; }; -template struct ei_cleantype { typedef typename ei_cleantype::type type; }; - -/** \internal In short, it computes int(sqrt(\a Y)) with \a Y an integer. - * Usage example: \code ei_meta_sqrt<1023>::ret \endcode - */ -template Y))) > - // use ?: instead of || just to shut up a stupid gcc 4.3 warning -class ei_meta_sqrt -{ - enum { - MidX = (InfX+SupX)/2, - TakeInf = MidX*MidX > Y ? 1 : 0, - NewInf = int(TakeInf) ? InfX : int(MidX), - NewSup = int(TakeInf) ? int(MidX) : SupX - }; - public: - enum { ret = ei_meta_sqrt::ret }; -}; - -template -class ei_meta_sqrt { public: enum { ret = (SupX*SupX <= Y) ? SupX : InfX }; }; - -} // end namespace Eigen - -#endif // EIGEN2_META_H diff --git a/Eigen/src/Eigen2Support/Minor.h b/Eigen/src/Eigen2Support/Minor.h deleted file mode 100644 index 4cded5734..000000000 --- a/Eigen/src/Eigen2Support/Minor.h +++ /dev/null @@ -1,117 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2006-2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_MINOR_H -#define EIGEN_MINOR_H - -namespace Eigen { - -/** - * \class Minor - * - * \brief Expression of a minor - * - * \param MatrixType the type of the object in which we are taking a minor - * - * This class represents an expression of a minor. It is the return - * type of MatrixBase::minor() and most of the time this is the only way it - * is used. - * - * \sa MatrixBase::minor() - */ - -namespace internal { -template -struct traits > - : traits -{ - typedef typename nested::type MatrixTypeNested; - typedef typename remove_reference::type _MatrixTypeNested; - typedef typename MatrixType::StorageKind StorageKind; - enum { - RowsAtCompileTime = (MatrixType::RowsAtCompileTime != Dynamic) ? - int(MatrixType::RowsAtCompileTime) - 1 : Dynamic, - ColsAtCompileTime = (MatrixType::ColsAtCompileTime != Dynamic) ? - int(MatrixType::ColsAtCompileTime) - 1 : Dynamic, - MaxRowsAtCompileTime = (MatrixType::MaxRowsAtCompileTime != Dynamic) ? - int(MatrixType::MaxRowsAtCompileTime) - 1 : Dynamic, - MaxColsAtCompileTime = (MatrixType::MaxColsAtCompileTime != Dynamic) ? - int(MatrixType::MaxColsAtCompileTime) - 1 : Dynamic, - Flags = _MatrixTypeNested::Flags & (HereditaryBits | LvalueBit), - CoeffReadCost = _MatrixTypeNested::CoeffReadCost // minor is used typically on tiny matrices, - // where loops are unrolled and the 'if' evaluates at compile time - }; -}; -} - -template class Minor - : public MatrixBase > -{ - public: - - typedef MatrixBase Base; - EIGEN_DENSE_PUBLIC_INTERFACE(Minor) - - inline Minor(const MatrixType& matrix, - Index row, Index col) - : m_matrix(matrix), m_row(row), m_col(col) - { - eigen_assert(row >= 0 && row < matrix.rows() - && col >= 0 && col < matrix.cols()); - } - - EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Minor) - - inline Index rows() const { return m_matrix.rows() - 1; } - inline Index cols() const { return m_matrix.cols() - 1; } - - inline Scalar& coeffRef(Index row, Index col) - { - return m_matrix.const_cast_derived().coeffRef(row + (row >= m_row), col + (col >= m_col)); - } - - inline const Scalar coeff(Index row, Index col) const - { - return m_matrix.coeff(row + (row >= m_row), col + (col >= m_col)); - } - - protected: - const typename MatrixType::Nested m_matrix; - const Index m_row, m_col; -}; - -/** - * \return an expression of the (\a row, \a col)-minor of *this, - * i.e. an expression constructed from *this by removing the specified - * row and column. - * - * Example: \include MatrixBase_minor.cpp - * Output: \verbinclude MatrixBase_minor.out - * - * \sa class Minor - */ -template -inline Minor -MatrixBase::minor(Index row, Index col) -{ - return Minor(derived(), row, col); -} - -/** - * This is the const version of minor(). */ -template -inline const Minor -MatrixBase::minor(Index row, Index col) const -{ - return Minor(derived(), row, col); -} - -} // end namespace Eigen - -#endif // EIGEN_MINOR_H diff --git a/Eigen/src/Eigen2Support/QR.h b/Eigen/src/Eigen2Support/QR.h deleted file mode 100644 index 2042c9851..000000000 --- a/Eigen/src/Eigen2Support/QR.h +++ /dev/null @@ -1,67 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2011 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_QR_H -#define EIGEN2_QR_H - -namespace Eigen { - -template -class QR : public HouseholderQR -{ - public: - - typedef HouseholderQR Base; - typedef Block MatrixRBlockType; - - QR() : Base() {} - - template - explicit QR(const T& t) : Base(t) {} - - template - bool solve(const MatrixBase& b, ResultType *result) const - { - *result = static_cast(this)->solve(b); - return true; - } - - MatrixType matrixQ(void) const { - MatrixType ret = MatrixType::Identity(this->rows(), this->cols()); - ret = this->householderQ() * ret; - return ret; - } - - bool isFullRank() const { - return true; - } - - const TriangularView - matrixR(void) const - { - int cols = this->cols(); - return MatrixRBlockType(this->matrixQR(), 0, 0, cols, cols).template triangularView(); - } -}; - -/** \return the QR decomposition of \c *this. - * - * \sa class QR - */ -template -const QR::PlainObject> -MatrixBase::qr() const -{ - return QR(eval()); -} - -} // end namespace Eigen - -#endif // EIGEN2_QR_H diff --git a/Eigen/src/Eigen2Support/SVD.h b/Eigen/src/Eigen2Support/SVD.h deleted file mode 100644 index 3d03d2288..000000000 --- a/Eigen/src/Eigen2Support/SVD.h +++ /dev/null @@ -1,637 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_SVD_H -#define EIGEN2_SVD_H - -namespace Eigen { - -/** \ingroup SVD_Module - * \nonstableyet - * - * \class SVD - * - * \brief Standard SVD decomposition of a matrix and associated features - * - * \param MatrixType the type of the matrix of which we are computing the SVD decomposition - * - * This class performs a standard SVD decomposition of a real matrix A of size \c M x \c N - * with \c M \>= \c N. - * - * - * \sa MatrixBase::SVD() - */ -template class SVD -{ - private: - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - - enum { - PacketSize = internal::packet_traits::size, - AlignmentMask = int(PacketSize)-1, - MinSize = EIGEN_SIZE_MIN_PREFER_DYNAMIC(MatrixType::RowsAtCompileTime, MatrixType::ColsAtCompileTime) - }; - - typedef Matrix ColVector; - typedef Matrix RowVector; - - typedef Matrix MatrixUType; - typedef Matrix MatrixVType; - typedef Matrix SingularValuesType; - - public: - - SVD() {} // a user who relied on compiler-generated default compiler reported problems with MSVC in 2.0.7 - - SVD(const MatrixType& matrix) - : m_matU(matrix.rows(), (std::min)(matrix.rows(), matrix.cols())), - m_matV(matrix.cols(),matrix.cols()), - m_sigma((std::min)(matrix.rows(),matrix.cols())) - { - compute(matrix); - } - - template - bool solve(const MatrixBase &b, ResultType* result) const; - - const MatrixUType& matrixU() const { return m_matU; } - const SingularValuesType& singularValues() const { return m_sigma; } - const MatrixVType& matrixV() const { return m_matV; } - - void compute(const MatrixType& matrix); - SVD& sort(); - - template - void computeUnitaryPositive(UnitaryType *unitary, PositiveType *positive) const; - template - void computePositiveUnitary(PositiveType *positive, UnitaryType *unitary) const; - template - void computeRotationScaling(RotationType *unitary, ScalingType *positive) const; - template - void computeScalingRotation(ScalingType *positive, RotationType *unitary) const; - - protected: - /** \internal */ - MatrixUType m_matU; - /** \internal */ - MatrixVType m_matV; - /** \internal */ - SingularValuesType m_sigma; -}; - -/** Computes / recomputes the SVD decomposition A = U S V^* of \a matrix - * - * \note this code has been adapted from JAMA (public domain) - */ -template -void SVD::compute(const MatrixType& matrix) -{ - const int m = matrix.rows(); - const int n = matrix.cols(); - const int nu = (std::min)(m,n); - ei_assert(m>=n && "In Eigen 2.0, SVD only works for MxN matrices with M>=N. Sorry!"); - ei_assert(m>1 && "In Eigen 2.0, SVD doesn't work on 1x1 matrices"); - - m_matU.resize(m, nu); - m_matU.setZero(); - m_sigma.resize((std::min)(m,n)); - m_matV.resize(n,n); - - RowVector e(n); - ColVector work(m); - MatrixType matA(matrix); - const bool wantu = true; - const bool wantv = true; - int i=0, j=0, k=0; - - // Reduce A to bidiagonal form, storing the diagonal elements - // in s and the super-diagonal elements in e. - int nct = (std::min)(m-1,n); - int nrt = (std::max)(0,(std::min)(n-2,m)); - for (k = 0; k < (std::max)(nct,nrt); ++k) - { - if (k < nct) - { - // Compute the transformation for the k-th column and - // place the k-th diagonal in m_sigma[k]. - m_sigma[k] = matA.col(k).end(m-k).norm(); - if (m_sigma[k] != 0.0) // FIXME - { - if (matA(k,k) < 0.0) - m_sigma[k] = -m_sigma[k]; - matA.col(k).end(m-k) /= m_sigma[k]; - matA(k,k) += 1.0; - } - m_sigma[k] = -m_sigma[k]; - } - - for (j = k+1; j < n; ++j) - { - if ((k < nct) && (m_sigma[k] != 0.0)) - { - // Apply the transformation. - Scalar t = matA.col(k).end(m-k).eigen2_dot(matA.col(j).end(m-k)); // FIXME dot product or cwise prod + .sum() ?? - t = -t/matA(k,k); - matA.col(j).end(m-k) += t * matA.col(k).end(m-k); - } - - // Place the k-th row of A into e for the - // subsequent calculation of the row transformation. - e[j] = matA(k,j); - } - - // Place the transformation in U for subsequent back multiplication. - if (wantu & (k < nct)) - m_matU.col(k).end(m-k) = matA.col(k).end(m-k); - - if (k < nrt) - { - // Compute the k-th row transformation and place the - // k-th super-diagonal in e[k]. - e[k] = e.end(n-k-1).norm(); - if (e[k] != 0.0) - { - if (e[k+1] < 0.0) - e[k] = -e[k]; - e.end(n-k-1) /= e[k]; - e[k+1] += 1.0; - } - e[k] = -e[k]; - if ((k+1 < m) & (e[k] != 0.0)) - { - // Apply the transformation. - work.end(m-k-1) = matA.corner(BottomRight,m-k-1,n-k-1) * e.end(n-k-1); - for (j = k+1; j < n; ++j) - matA.col(j).end(m-k-1) += (-e[j]/e[k+1]) * work.end(m-k-1); - } - - // Place the transformation in V for subsequent back multiplication. - if (wantv) - m_matV.col(k).end(n-k-1) = e.end(n-k-1); - } - } - - - // Set up the final bidiagonal matrix or order p. - int p = (std::min)(n,m+1); - if (nct < n) - m_sigma[nct] = matA(nct,nct); - if (m < p) - m_sigma[p-1] = 0.0; - if (nrt+1 < p) - e[nrt] = matA(nrt,p-1); - e[p-1] = 0.0; - - // If required, generate U. - if (wantu) - { - for (j = nct; j < nu; ++j) - { - m_matU.col(j).setZero(); - m_matU(j,j) = 1.0; - } - for (k = nct-1; k >= 0; k--) - { - if (m_sigma[k] != 0.0) - { - for (j = k+1; j < nu; ++j) - { - Scalar t = m_matU.col(k).end(m-k).eigen2_dot(m_matU.col(j).end(m-k)); // FIXME is it really a dot product we want ? - t = -t/m_matU(k,k); - m_matU.col(j).end(m-k) += t * m_matU.col(k).end(m-k); - } - m_matU.col(k).end(m-k) = - m_matU.col(k).end(m-k); - m_matU(k,k) = Scalar(1) + m_matU(k,k); - if (k-1>0) - m_matU.col(k).start(k-1).setZero(); - } - else - { - m_matU.col(k).setZero(); - m_matU(k,k) = 1.0; - } - } - } - - // If required, generate V. - if (wantv) - { - for (k = n-1; k >= 0; k--) - { - if ((k < nrt) & (e[k] != 0.0)) - { - for (j = k+1; j < nu; ++j) - { - Scalar t = m_matV.col(k).end(n-k-1).eigen2_dot(m_matV.col(j).end(n-k-1)); // FIXME is it really a dot product we want ? - t = -t/m_matV(k+1,k); - m_matV.col(j).end(n-k-1) += t * m_matV.col(k).end(n-k-1); - } - } - m_matV.col(k).setZero(); - m_matV(k,k) = 1.0; - } - } - - // Main iteration loop for the singular values. - int pp = p-1; - int iter = 0; - Scalar eps = ei_pow(Scalar(2),ei_is_same_type::ret ? Scalar(-23) : Scalar(-52)); - while (p > 0) - { - int k=0; - int kase=0; - - // Here is where a test for too many iterations would go. - - // This section of the program inspects for - // negligible elements in the s and e arrays. On - // completion the variables kase and k are set as follows. - - // kase = 1 if s(p) and e[k-1] are negligible and k

= -1; --k) - { - if (k == -1) - break; - if (ei_abs(e[k]) <= eps*(ei_abs(m_sigma[k]) + ei_abs(m_sigma[k+1]))) - { - e[k] = 0.0; - break; - } - } - if (k == p-2) - { - kase = 4; - } - else - { - int ks; - for (ks = p-1; ks >= k; --ks) - { - if (ks == k) - break; - Scalar t = (ks != p ? ei_abs(e[ks]) : Scalar(0)) + (ks != k+1 ? ei_abs(e[ks-1]) : Scalar(0)); - if (ei_abs(m_sigma[ks]) <= eps*t) - { - m_sigma[ks] = 0.0; - break; - } - } - if (ks == k) - { - kase = 3; - } - else if (ks == p-1) - { - kase = 1; - } - else - { - kase = 2; - k = ks; - } - } - ++k; - - // Perform the task indicated by kase. - switch (kase) - { - - // Deflate negligible s(p). - case 1: - { - Scalar f(e[p-2]); - e[p-2] = 0.0; - for (j = p-2; j >= k; --j) - { - Scalar t(numext::hypot(m_sigma[j],f)); - Scalar cs(m_sigma[j]/t); - Scalar sn(f/t); - m_sigma[j] = t; - if (j != k) - { - f = -sn*e[j-1]; - e[j-1] = cs*e[j-1]; - } - if (wantv) - { - for (i = 0; i < n; ++i) - { - t = cs*m_matV(i,j) + sn*m_matV(i,p-1); - m_matV(i,p-1) = -sn*m_matV(i,j) + cs*m_matV(i,p-1); - m_matV(i,j) = t; - } - } - } - } - break; - - // Split at negligible s(k). - case 2: - { - Scalar f(e[k-1]); - e[k-1] = 0.0; - for (j = k; j < p; ++j) - { - Scalar t(numext::hypot(m_sigma[j],f)); - Scalar cs( m_sigma[j]/t); - Scalar sn(f/t); - m_sigma[j] = t; - f = -sn*e[j]; - e[j] = cs*e[j]; - if (wantu) - { - for (i = 0; i < m; ++i) - { - t = cs*m_matU(i,j) + sn*m_matU(i,k-1); - m_matU(i,k-1) = -sn*m_matU(i,j) + cs*m_matU(i,k-1); - m_matU(i,j) = t; - } - } - } - } - break; - - // Perform one qr step. - case 3: - { - // Calculate the shift. - Scalar scale = (std::max)((std::max)((std::max)((std::max)( - ei_abs(m_sigma[p-1]),ei_abs(m_sigma[p-2])),ei_abs(e[p-2])), - ei_abs(m_sigma[k])),ei_abs(e[k])); - Scalar sp = m_sigma[p-1]/scale; - Scalar spm1 = m_sigma[p-2]/scale; - Scalar epm1 = e[p-2]/scale; - Scalar sk = m_sigma[k]/scale; - Scalar ek = e[k]/scale; - Scalar b = ((spm1 + sp)*(spm1 - sp) + epm1*epm1)/Scalar(2); - Scalar c = (sp*epm1)*(sp*epm1); - Scalar shift(0); - if ((b != 0.0) || (c != 0.0)) - { - shift = ei_sqrt(b*b + c); - if (b < 0.0) - shift = -shift; - shift = c/(b + shift); - } - Scalar f = (sk + sp)*(sk - sp) + shift; - Scalar g = sk*ek; - - // Chase zeros. - - for (j = k; j < p-1; ++j) - { - Scalar t = numext::hypot(f,g); - Scalar cs = f/t; - Scalar sn = g/t; - if (j != k) - e[j-1] = t; - f = cs*m_sigma[j] + sn*e[j]; - e[j] = cs*e[j] - sn*m_sigma[j]; - g = sn*m_sigma[j+1]; - m_sigma[j+1] = cs*m_sigma[j+1]; - if (wantv) - { - for (i = 0; i < n; ++i) - { - t = cs*m_matV(i,j) + sn*m_matV(i,j+1); - m_matV(i,j+1) = -sn*m_matV(i,j) + cs*m_matV(i,j+1); - m_matV(i,j) = t; - } - } - t = numext::hypot(f,g); - cs = f/t; - sn = g/t; - m_sigma[j] = t; - f = cs*e[j] + sn*m_sigma[j+1]; - m_sigma[j+1] = -sn*e[j] + cs*m_sigma[j+1]; - g = sn*e[j+1]; - e[j+1] = cs*e[j+1]; - if (wantu && (j < m-1)) - { - for (i = 0; i < m; ++i) - { - t = cs*m_matU(i,j) + sn*m_matU(i,j+1); - m_matU(i,j+1) = -sn*m_matU(i,j) + cs*m_matU(i,j+1); - m_matU(i,j) = t; - } - } - } - e[p-2] = f; - iter = iter + 1; - } - break; - - // Convergence. - case 4: - { - // Make the singular values positive. - if (m_sigma[k] <= 0.0) - { - m_sigma[k] = m_sigma[k] < Scalar(0) ? -m_sigma[k] : Scalar(0); - if (wantv) - m_matV.col(k).start(pp+1) = -m_matV.col(k).start(pp+1); - } - - // Order the singular values. - while (k < pp) - { - if (m_sigma[k] >= m_sigma[k+1]) - break; - Scalar t = m_sigma[k]; - m_sigma[k] = m_sigma[k+1]; - m_sigma[k+1] = t; - if (wantv && (k < n-1)) - m_matV.col(k).swap(m_matV.col(k+1)); - if (wantu && (k < m-1)) - m_matU.col(k).swap(m_matU.col(k+1)); - ++k; - } - iter = 0; - p--; - } - break; - } // end big switch - } // end iterations -} - -template -SVD& SVD::sort() -{ - int mu = m_matU.rows(); - int mv = m_matV.rows(); - int n = m_matU.cols(); - - for (int i=0; i p) - { - k = j; - p = m_sigma.coeff(j); - } - } - if (k != i) - { - m_sigma.coeffRef(k) = m_sigma.coeff(i); // i.e. - m_sigma.coeffRef(i) = p; // swaps the i-th and the k-th elements - - int j = mu; - for(int s=0; j!=0; ++s, --j) - std::swap(m_matU.coeffRef(s,i), m_matU.coeffRef(s,k)); - - j = mv; - for (int s=0; j!=0; ++s, --j) - std::swap(m_matV.coeffRef(s,i), m_matV.coeffRef(s,k)); - } - } - return *this; -} - -/** \returns the solution of \f$ A x = b \f$ using the current SVD decomposition of A. - * The parts of the solution corresponding to zero singular values are ignored. - * - * \sa MatrixBase::svd(), LU::solve(), LLT::solve() - */ -template -template -bool SVD::solve(const MatrixBase &b, ResultType* result) const -{ - ei_assert(b.rows() == m_matU.rows()); - - Scalar maxVal = m_sigma.cwise().abs().maxCoeff(); - for (int j=0; j aux = m_matU.transpose() * b.col(j); - - for (int i = 0; i col(j) = m_matV * aux; - } - return true; -} - -/** Computes the polar decomposition of the matrix, as a product unitary x positive. - * - * If either pointer is zero, the corresponding computation is skipped. - * - * Only for square matrices. - * - * \sa computePositiveUnitary(), computeRotationScaling() - */ -template -template -void SVD::computeUnitaryPositive(UnitaryType *unitary, - PositiveType *positive) const -{ - ei_assert(m_matU.cols() == m_matV.cols() && "Polar decomposition is only for square matrices"); - if(unitary) *unitary = m_matU * m_matV.adjoint(); - if(positive) *positive = m_matV * m_sigma.asDiagonal() * m_matV.adjoint(); -} - -/** Computes the polar decomposition of the matrix, as a product positive x unitary. - * - * If either pointer is zero, the corresponding computation is skipped. - * - * Only for square matrices. - * - * \sa computeUnitaryPositive(), computeRotationScaling() - */ -template -template -void SVD::computePositiveUnitary(UnitaryType *positive, - PositiveType *unitary) const -{ - ei_assert(m_matU.rows() == m_matV.rows() && "Polar decomposition is only for square matrices"); - if(unitary) *unitary = m_matU * m_matV.adjoint(); - if(positive) *positive = m_matU * m_sigma.asDiagonal() * m_matU.adjoint(); -} - -/** decomposes the matrix as a product rotation x scaling, the scaling being - * not necessarily positive. - * - * If either pointer is zero, the corresponding computation is skipped. - * - * This method requires the Geometry module. - * - * \sa computeScalingRotation(), computeUnitaryPositive() - */ -template -template -void SVD::computeRotationScaling(RotationType *rotation, ScalingType *scaling) const -{ - ei_assert(m_matU.rows() == m_matV.rows() && "Polar decomposition is only for square matrices"); - Scalar x = (m_matU * m_matV.adjoint()).determinant(); // so x has absolute value 1 - Matrix sv(m_sigma); - sv.coeffRef(0) *= x; - if(scaling) scaling->lazyAssign(m_matV * sv.asDiagonal() * m_matV.adjoint()); - if(rotation) - { - MatrixType m(m_matU); - m.col(0) /= x; - rotation->lazyAssign(m * m_matV.adjoint()); - } -} - -/** decomposes the matrix as a product scaling x rotation, the scaling being - * not necessarily positive. - * - * If either pointer is zero, the corresponding computation is skipped. - * - * This method requires the Geometry module. - * - * \sa computeRotationScaling(), computeUnitaryPositive() - */ -template -template -void SVD::computeScalingRotation(ScalingType *scaling, RotationType *rotation) const -{ - ei_assert(m_matU.rows() == m_matV.rows() && "Polar decomposition is only for square matrices"); - Scalar x = (m_matU * m_matV.adjoint()).determinant(); // so x has absolute value 1 - Matrix sv(m_sigma); - sv.coeffRef(0) *= x; - if(scaling) scaling->lazyAssign(m_matU * sv.asDiagonal() * m_matU.adjoint()); - if(rotation) - { - MatrixType m(m_matU); - m.col(0) /= x; - rotation->lazyAssign(m * m_matV.adjoint()); - } -} - - -/** \svd_module - * \returns the SVD decomposition of \c *this - */ -template -inline SVD::PlainObject> -MatrixBase::svd() const -{ - return SVD(derived()); -} - -} // end namespace Eigen - -#endif // EIGEN2_SVD_H diff --git a/Eigen/src/Eigen2Support/TriangularSolver.h b/Eigen/src/Eigen2Support/TriangularSolver.h deleted file mode 100644 index ebbeb3b49..000000000 --- a/Eigen/src/Eigen2Support/TriangularSolver.h +++ /dev/null @@ -1,42 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2010 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_TRIANGULAR_SOLVER2_H -#define EIGEN_TRIANGULAR_SOLVER2_H - -namespace Eigen { - -const unsigned int UnitDiagBit = UnitDiag; -const unsigned int SelfAdjointBit = SelfAdjoint; -const unsigned int UpperTriangularBit = Upper; -const unsigned int LowerTriangularBit = Lower; - -const unsigned int UpperTriangular = Upper; -const unsigned int LowerTriangular = Lower; -const unsigned int UnitUpperTriangular = UnitUpper; -const unsigned int UnitLowerTriangular = UnitLower; - -template -template -typename ExpressionType::PlainObject -Flagged::solveTriangular(const MatrixBase& other) const -{ - return m_matrix.template triangularView().solve(other.derived()); -} - -template -template -void Flagged::solveTriangularInPlace(const MatrixBase& other) const -{ - m_matrix.template triangularView().solveInPlace(other.derived()); -} - -} // end namespace Eigen - -#endif // EIGEN_TRIANGULAR_SOLVER2_H diff --git a/Eigen/src/Eigen2Support/VectorBlock.h b/Eigen/src/Eigen2Support/VectorBlock.h deleted file mode 100644 index 71a8080a9..000000000 --- a/Eigen/src/Eigen2Support/VectorBlock.h +++ /dev/null @@ -1,94 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008-2009 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN2_VECTORBLOCK_H -#define EIGEN2_VECTORBLOCK_H - -namespace Eigen { - -/** \deprecated use DenseMase::head(Index) */ -template -inline VectorBlock -MatrixBase::start(Index size) -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), 0, size); -} - -/** \deprecated use DenseMase::head(Index) */ -template -inline const VectorBlock -MatrixBase::start(Index size) const -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), 0, size); -} - -/** \deprecated use DenseMase::tail(Index) */ -template -inline VectorBlock -MatrixBase::end(Index size) -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), this->size() - size, size); -} - -/** \deprecated use DenseMase::tail(Index) */ -template -inline const VectorBlock -MatrixBase::end(Index size) const -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), this->size() - size, size); -} - -/** \deprecated use DenseMase::head() */ -template -template -inline VectorBlock -MatrixBase::start() -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), 0); -} - -/** \deprecated use DenseMase::head() */ -template -template -inline const VectorBlock -MatrixBase::start() const -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), 0); -} - -/** \deprecated use DenseMase::tail() */ -template -template -inline VectorBlock -MatrixBase::end() -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), size() - Size); -} - -/** \deprecated use DenseMase::tail() */ -template -template -inline const VectorBlock -MatrixBase::end() const -{ - EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) - return VectorBlock(derived(), size() - Size); -} - -} // end namespace Eigen - -#endif // EIGEN2_VECTORBLOCK_H diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index b8146d04d..d10eba201 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -346,40 +346,6 @@ template class SelfAdjointEigenSolver */ static const int m_maxIterations = 30; - #ifdef EIGEN2_SUPPORT - EIGEN_DEVICE_FUNC - SelfAdjointEigenSolver(const MatrixType& matrix, bool computeEigenvectors) - : m_eivec(matrix.rows(), matrix.cols()), - m_eivalues(matrix.cols()), - m_subdiag(matrix.rows() > 1 ? matrix.rows() - 1 : 1), - m_isInitialized(false) - { - compute(matrix, computeEigenvectors); - } - - EIGEN_DEVICE_FUNC - SelfAdjointEigenSolver(const MatrixType& matA, const MatrixType& matB, bool computeEigenvectors = true) - : m_eivec(matA.cols(), matA.cols()), - m_eivalues(matA.cols()), - m_subdiag(matA.cols() > 1 ? matA.cols() - 1 : 1), - m_isInitialized(false) - { - static_cast*>(this)->compute(matA, matB, computeEigenvectors ? ComputeEigenvectors : EigenvaluesOnly); - } - - EIGEN_DEVICE_FUNC - void compute(const MatrixType& matrix, bool computeEigenvectors) - { - compute(matrix, computeEigenvectors ? ComputeEigenvectors : EigenvaluesOnly); - } - - EIGEN_DEVICE_FUNC - void compute(const MatrixType& matA, const MatrixType& matB, bool computeEigenvectors = true) - { - compute(matA, matB, computeEigenvectors ? ComputeEigenvectors : EigenvaluesOnly); - } - #endif // EIGEN2_SUPPORT - protected: MatrixType m_eivec; RealVectorType m_eivalues; diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index 1d389ecac..91383db31 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -481,7 +481,6 @@ MatrixBase::partialPivLu() const } #endif -#if EIGEN2_SUPPORT_STAGE > STAGE20_RESOLVE_API_CONFLICTS /** \lu_module * * Synonym of partialPivLu(). @@ -499,8 +498,6 @@ MatrixBase::lu() const } #endif -#endif - } // end namespace Eigen #endif // EIGEN_PARTIALLU_H diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index bbcf7fb1c..c64d74da0 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -369,17 +369,6 @@ template class SparseMatrixBase : public EigenBase template Derived& operator*=(const SparseMatrixBase& other); - #ifdef EIGEN2_SUPPORT - // deprecated - template - typename internal::plain_matrix_type_column_major::type - solveTriangular(const MatrixBase& other) const; - - // deprecated - template - void solveTriangularInPlace(MatrixBase& other) const; - #endif // EIGEN2_SUPPORT - template inline const SparseTriangularView triangularView() const; diff --git a/Eigen/src/SparseCore/TriangularSolver.h b/Eigen/src/SparseCore/TriangularSolver.h index cb8ad82b4..f958ab300 100644 --- a/Eigen/src/SparseCore/TriangularSolver.h +++ b/Eigen/src/SparseCore/TriangularSolver.h @@ -305,30 +305,6 @@ void SparseTriangularView::solveInPlace(SparseMatrixBase -template -void SparseMatrixBase::solveTriangularInPlace(MatrixBase& other) const -{ - this->template triangular().solveInPlace(other); -} - -/** \deprecated */ -template -template -typename internal::plain_matrix_type_column_major::type -SparseMatrixBase::solveTriangular(const MatrixBase& other) const -{ - typename internal::plain_matrix_type_column_major::type res(other); - derived().solveTriangularInPlace(res); - return res; -} -#endif // EIGEN2_SUPPORT - } // end namespace Eigen #endif // EIGEN_SPARSETRIANGULARSOLVER_H diff --git a/doc/A05_PortingFrom2To3.dox b/doc/A05_PortingFrom2To3.dox index d885b4f6d..47011aec0 100644 --- a/doc/A05_PortingFrom2To3.dox +++ b/doc/A05_PortingFrom2To3.dox @@ -9,11 +9,8 @@ and gives tips to help porting your application from Eigen2 to Eigen3. \section CompatibilitySupport Eigen2 compatibility support -In order to ease the switch from Eigen2 to Eigen3, Eigen3 features \subpage Eigen2SupportModes "Eigen2 support modes". - -The quick way to enable this is to define the \c EIGEN2_SUPPORT preprocessor token \b before including any Eigen header (typically it should be set in your project options). - -A more powerful, \em staged migration path is also provided, which may be useful to migrate larger projects from Eigen2 to Eigen3. This is explained in the \ref Eigen2SupportModes "Eigen 2 support modes" page. +Up to version 3.2 %Eigen provides Eigen2 support modes. These are removed now, because they were barely used anymore and became hard to maintain after internal re-designs. +You can still use them by first porting your code to Eigen 3.2. \section Using The USING_PART_OF_NAMESPACE_EIGEN macro diff --git a/doc/A10_Eigen2SupportModes.dox b/doc/A10_Eigen2SupportModes.dox deleted file mode 100644 index abfdb4683..000000000 --- a/doc/A10_Eigen2SupportModes.dox +++ /dev/null @@ -1,93 +0,0 @@ -namespace Eigen { - -/** \page Eigen2SupportModes Eigen 2 support modes - -This page documents the Eigen2 support modes, a powerful tool to help migrating your project from Eigen 2 to Eigen 3. -Don't miss our page on \ref Eigen2ToEigen3 "API changes" between Eigen 2 and Eigen 3. - -\eigenAutoToc - -\section EIGEN2_SUPPORT_Macro The quick way: define EIGEN2_SUPPORT - -By defining EIGEN2_SUPPORT before including any Eigen 3 header, you get back a large part of the Eigen 2 API, while keeping the Eigen 3 API and ABI unchanged. - -This defaults to the \ref Stage30 "stage 30" described below. - -The rest of this page describes an optional, more powerful \em staged migration path. - -\section StagedMigrationPathOverview Overview of the staged migration path - -The primary reason why EIGEN2_SUPPORT alone may not be enough to migrate a large project from Eigen 2 to Eigen 3 is that some of the Eigen 2 API is inherently incompatible with the Eigen 3 API. This happens when the same identifier is used in Eigen 2 and in Eigen 3 with different meanings. To help migrate projects that rely on such API, we provide a staged migration path allowing to perform the migration \em incrementally. - -It goes as follows: -\li Step 0: start with a project using Eigen 2. -\li Step 1: build your project against Eigen 3 with \ref Stage10 "Eigen 2 support stage 10". This mode enables maximum compatibility with the Eigen 2 API, with just a few exceptions. -\li Step 2: build your project against Eigen 3 with \ref Stage20 "Eigen 2 support stage 20". This mode forces you to add eigen2_ prefixes to the Eigen2 identifiers that conflict with Eigen 3 API. -\li Step 3: build your project against Eigen 3 with \ref Stage30 "Eigen 2 support stage 30". This mode enables the full Eigen 3 API. -\li Step 4: build your project against Eigen 3 with \ref Stage40 "Eigen 2 support stage 40". This mode enables the full Eigen 3 strictness on matters, such as const-correctness, where Eigen 2 was looser. -\li Step 5: build your project against Eigen 3 without any Eigen 2 support mode. - -\section Stage10 Stage 10: define EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API - -Enable this mode by defining the EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API preprocessor macro before including any Eigen 3 header. - -This mode maximizes support for the Eigen 2 API. As a result, it does not offer the full Eigen 3 API. Also, it doesn't offer quite 100% of the Eigen 2 API. - -The part of the Eigen 3 API that is not present in this mode, is Eigen 3's Geometry module. Indeed, this mode completely replaces it by a copy of Eigen 2's Geometry module. - -The parts of the API that are still not 100% Eigen 2 compatible in this mode are: -\li Dot products over complex numbers. Eigen 2's dot product was linear in the first variable. Eigen 3's dot product is linear in the second variable. In other words, the Eigen 2 code \code x.dot(y) \endcode is equivalent to the Eigen 3 code \code y.dot(x) \endcode In yet other words, dot products are complex-conjugated in Eigen 3 compared to Eigen 2. The switch to the new convention was commanded by common usage, especially with the notation \f$ x^Ty \f$ for dot products of column-vectors. -\li The Sparse module. -\li Certain fine details of linear algebraic decompositions. For example, LDLT decomposition is now pivoting in Eigen 3 whereas it wasn't in Eigen 2, so code that was relying on its underlying matrix structure will break. -\li Usage of Eigen types in STL containers, \ref Eigen2ToEigen3 "as explained on this page". - -\section Stage20 Stage 20: define EIGEN2_SUPPORT_STAGE20_RESOLVE_API_CONFLICTS - -Enable this mode by defining the EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API preprocessor macro before including any Eigen 3 header. - -This mode removes the Eigen 2 API that is directly conflicting with Eigen 3 API. Instead, these bits of Eigen 2 API remain available with eigen2_ prefixes. The main examples of such API are: -\li the whole Geometry module. For example, replace \c Quaternion by \c eigen2_Quaternion, replace \c Transform3f by \c eigen2_Transform3f, etc. -\li the lu() method to obtain a LU decomposition. Replace by eigen2_lu(). - -There is also one more eigen2_-prefixed identifier that you should know about, even though its use is not checked at compile time by this mode: the dot() method. As was discussed above, over complex numbers, its meaning is different between Eigen 2 and Eigen 3. You can use eigen2_dot() to get the Eigen 2 behavior. - -\section Stage30 Stage 30: define EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API - -Enable this mode by defining the EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API preprocessor macro before including any Eigen 3 header. Also, this mode is what you get by default when you just define EIGEN2_SUPPORT. - -This mode gives you the full unaltered Eigen 3 API, while still keeping as much support as possible for the Eigen 2 API. - -The eigen2_-prefixed identifiers are still available, but at this stage you should now replace them by Eigen 3 identifiers. Have a look at our page on \ref Eigen2ToEigen3 "API changes" between Eigen 2 and Eigen 3. - -\section Stage40 Stage 40: define EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS - -Enable this mode by defining the EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS preprocessor macro before including any Eigen 3 header. - -This mode tightens the last bits of strictness, especially const-correctness, that had to be loosened to support what Eigen 2 allowed. For example, this code compiled in Eigen 2: -\code -const float array[4]; -x = Map(array); -\endcode -That allowed to circumvent constness. This is no longer allowed in Eigen 3. If you have to map const data in Eigen 3, map it as a const-qualified type. However, rather than explictly constructing Map objects, we strongly encourage you to use the static Map methods instead, as they take care of all of this for you: -\code -const float array[4]; -x = Vector4f::Map(array); -\endcode -This lets Eigen do the right thing for you and works equally well in Eigen 2 and in Eigen 3. - -\section FinallyDropAllEigen2Support Finally drop all Eigen 2 support - -Stage 40 is the first where it's "comfortable" to stay for a little longer period, since it preserves 100% Eigen 3 compatibility. However, we still encourage you to complete your migration as quickly as possible. While we do run the Eigen 2 test suite against Eigen 3's stage 10 support mode, we can't guarantee the same level of support and quality assurance for Eigen 2 support as we do for Eigen 3 itself, especially not in the long term. \ref Eigen2ToEigen3 "This page" describes a large part of the changes that you may need to perform. - -\section ABICompatibility What about ABI compatibility? - -It goes as follows: -\li Stage 10 already is ABI compatible with Eigen 3 for the basic (Matrix, Array, SparseMatrix...) types. However, since this stage uses a copy of Eigen 2's Geometry module instead of Eigen 3's own Geometry module, the ABI in the Geometry module is not Eigen 3 compatible. -\li Stage 20 removes the Eigen 3-incompatible Eigen 2 Geometry module (it remains available with eigen2_ prefix). So at this stage, all the identifiers that exist in Eigen 3 have the Eigen 3 ABI (and API). -\li Stage 30 introduces the remaining Eigen 3 identifiers. So at this stage, you have the full Eigen 3 ABI. -\li Stage 40 is no different than Stage 30 in these matters. - - -*/ - -} diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in index 509bd357a..2ceee5e20 100644 --- a/doc/Doxyfile.in +++ b/doc/Doxyfile.in @@ -1584,7 +1584,6 @@ PREDEFINED = EIGEN_EMPTY_STRUCT \ EIGEN_QT_SUPPORT \ EIGEN_STRONG_INLINE=inline \ EIGEN_DEVICE_FUNC= \ - "EIGEN2_SUPPORT_STAGE=99" \ "EIGEN_MAKE_CWISE_BINARY_OP(METHOD,FUNCTOR)=template const CwiseBinaryOp, const Derived, const OtherDerived> METHOD(const EIGEN_CURRENT_STORAGE_BASE_CLASS &other) const;" \ "EIGEN_CWISE_PRODUCT_RETURN_TYPE(LHS,RHS)=CwiseBinaryOp, const LHS, const RHS>" @@ -1601,8 +1600,7 @@ EXPAND_AS_DEFINED = EIGEN_MAKE_TYPEDEFS \ EIGEN_CWISE_BINOP_RETURN_TYPE \ EIGEN_CURRENT_STORAGE_BASE_CLASS \ EIGEN_MATHFUNC_IMPL \ - _EIGEN_GENERIC_PUBLIC_INTERFACE \ - EIGEN2_SUPPORT + _EIGEN_GENERIC_PUBLIC_INTERFACE # If the SKIP_FUNCTION_MACROS tag is set to YES (the default) then # doxygen's preprocessor will remove all references to function-like macros diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox index 6e40f919f..e67991fb7 100644 --- a/doc/PreprocessorDirectives.dox +++ b/doc/PreprocessorDirectives.dox @@ -5,7 +5,7 @@ namespace Eigen { You can control some aspects of %Eigen by defining the preprocessor tokens using \c \#define. These macros should be defined before any %Eigen headers are included. Often they are best set in the project options. -This page lists the preprocesor tokens recognised by %Eigen. +This page lists the preprocessor tokens recognized by %Eigen. \eigenAutoToc @@ -18,10 +18,9 @@ one option, and other parts (or libraries that you use) are compiled with anothe fail to link or exhibit subtle bugs. Nevertheless, these options can be useful for people who know what they are doing. - - \b EIGEN2_SUPPORT - if defined, enables the Eigen2 compatibility mode. This is meant to ease the transition - of Eigen2 to Eigen3 (see \ref Eigen2ToEigen3). Not defined by default. - - \b EIGEN2_SUPPORT_STAGEnn_xxx (for various values of nn and xxx) - staged migration path from Eigen2 to - Eigen3; see \ref Eigen2SupportModes. + - \b EIGEN2_SUPPORT and \b EIGEN2_SUPPORT_STAGEnn_xxx are disabled starting from the 3.3 release. + Defining one of these will raise a compile-error. If you need to compile Eigen2 code, + check this site. - \b EIGEN_DEFAULT_DENSE_INDEX_TYPE - the type for column and row indices in matrices, vectors and array (DenseBase::Index). Set to \c std::ptrdiff_t by default. - \b EIGEN_DEFAULT_IO_FORMAT - the IOFormat to use when printing a matrix if no %IOFormat is specified. @@ -55,7 +54,7 @@ run time. However, these assertions do cost time and can thus be turned off. \section TopicPreprocessorDirectivesPerformance Alignment, vectorization and performance tweaking - - \b EIGEN_MALLOC_ALREADY_ALIGNED - Can be set to 0 or 1 to tell whether default system malloc already + - \b EIGEN_MALLOC_ALREADY_ALIGNED - Can be set to 0 or 1 to tell whether default system \c malloc already returns aligned buffers. In not defined, then this information is automatically deduced from the compiler and system preprocessor tokens. - \b EIGEN_DONT_ALIGN - disables alignment completely. %Eigen will not try to align its objects and does not diff --git a/doc/examples/MatrixBase_cwise_const.cpp b/doc/examples/MatrixBase_cwise_const.cpp deleted file mode 100644 index 23700e0b6..000000000 --- a/doc/examples/MatrixBase_cwise_const.cpp +++ /dev/null @@ -1,18 +0,0 @@ -#define EIGEN2_SUPPORT -#include -#include - -using namespace Eigen; -using namespace std; - -int main() -{ - Matrix3i m = Matrix3i::Random(); - cout << "Here is the matrix m:" << endl << m << endl; - Matrix3i n = Matrix3i::Random(); - cout << "And here is the matrix n:" << endl << n << endl; - cout << "The coefficient-wise product of m and n is:" << endl; - cout << m.cwise() * n << endl; - cout << "Taking the cube of the coefficients of m yields:" << endl; - cout << m.cwise().pow(3) << endl; -} diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index c2d827051..62c9c78a6 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -157,7 +157,6 @@ ei_add_test(vectorization_logic) ei_add_test(basicstuff) ei_add_test(linearstructure) ei_add_test(integer_types) -ei_add_test(cwiseop) ei_add_test(unalignedcount) ei_add_test(exceptions) ei_add_test(redux) @@ -258,8 +257,6 @@ if(QT4_FOUND) ei_add_test(qtvector "" "${QT_QTCORE_LIBRARY}") endif(QT4_FOUND) -ei_add_test(eigen2support) - if(UMFPACK_FOUND) ei_add_test(umfpack_support "" "${UMFPACK_ALL_LIBS}") endif() @@ -303,7 +300,7 @@ ei_add_property(EIGEN_TESTING_SUMMARY "Sparse lib flags: ${SPARSE_LIBS}\n") option(EIGEN_TEST_EIGEN2 "Run whole Eigen2 test suite against EIGEN2_SUPPORT" OFF) if(EIGEN_TEST_EIGEN2) - add_subdirectory(eigen2) + message(WARNING "The Eigen2 test suite has been removed") endif() diff --git a/test/cwiseop.cpp b/test/cwiseop.cpp deleted file mode 100644 index 247fa2a09..000000000 --- a/test/cwiseop.cpp +++ /dev/null @@ -1,182 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#define EIGEN2_SUPPORT -#define EIGEN_NO_STATIC_ASSERT -#include "main.h" -#include - -#ifdef min -#undef min -#endif - -#ifdef max -#undef max -#endif - -using namespace std; - -template struct AddIfNull { - const Scalar operator() (const Scalar a, const Scalar b) const {return a<=1e-3 ? b : a;} - enum { Cost = NumTraits::AddCost }; -}; - -template -typename Eigen::internal::enable_if::IsInteger,typename MatrixType::Scalar>::type -cwiseops_real_only(MatrixType& m1, MatrixType& m2, MatrixType& m3, MatrixType& mones) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - - VERIFY_IS_APPROX(m1.cwise() / m2, m1.cwise() * (m2.cwise().inverse())); - m3 = m1.cwise().abs().cwise().sqrt(); - VERIFY_IS_APPROX(m3.cwise().square(), m1.cwise().abs()); - VERIFY_IS_APPROX(m1.cwise().square().cwise().sqrt(), m1.cwise().abs()); - VERIFY_IS_APPROX(m1.cwise().abs().cwise().log().cwise().exp() , m1.cwise().abs()); - - VERIFY_IS_APPROX(m1.cwise().pow(2), m1.cwise().square()); - m3 = (m1.cwise().abs().cwise()<=RealScalar(0.01)).select(mones,m1); - VERIFY_IS_APPROX(m3.cwise().pow(-1), m3.cwise().inverse()); - m3 = m1.cwise().abs(); - VERIFY_IS_APPROX(m3.cwise().pow(RealScalar(0.5)), m3.cwise().sqrt()); - -// VERIFY_IS_APPROX(m1.cwise().tan(), m1.cwise().sin().cwise() / m1.cwise().cos()); - VERIFY_IS_APPROX(mones, m1.cwise().sin().cwise().square() + m1.cwise().cos().cwise().square()); - m3 = m1; - m3.cwise() /= m2; - VERIFY_IS_APPROX(m3, m1.cwise() / m2); - - return Scalar(0); -} - -template -typename Eigen::internal::enable_if::IsInteger,typename MatrixType::Scalar>::type -cwiseops_real_only(MatrixType& , MatrixType& , MatrixType& , MatrixType& ) -{ - return 0; -} - -template void cwiseops(const MatrixType& m) -{ - typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - typedef Matrix VectorType; - - Index rows = m.rows(); - Index cols = m.cols(); - - MatrixType m1 = MatrixType::Random(rows, cols), - m1bis = m1, - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - m4(rows, cols), - mzero = MatrixType::Zero(rows, cols), - mones = MatrixType::Ones(rows, cols), - identity = Matrix - ::Identity(rows, rows); - VectorType vzero = VectorType::Zero(rows), - vones = VectorType::Ones(rows), - v3(rows); - - Index r = internal::random(0, rows-1), - c = internal::random(0, cols-1); - - Scalar s1 = internal::random(); - - // test Zero, Ones, Constant, and the set* variants - m3 = MatrixType::Constant(rows, cols, s1); - for (int j=0; j >(mones); - - VERIFY_IS_APPROX(m1.cwise().pow(2), m1.cwise().abs2()); - VERIFY_IS_APPROX(m1.cwise().pow(2), m1.cwise().square()); - VERIFY_IS_APPROX(m1.cwise().pow(3), m1.cwise().cube()); - - VERIFY_IS_APPROX(m1 + mones, m1.cwise()+Scalar(1)); - VERIFY_IS_APPROX(m1 - mones, m1.cwise()-Scalar(1)); - m3 = m1; m3.cwise() += 1; - VERIFY_IS_APPROX(m1 + mones, m3); - m3 = m1; m3.cwise() -= 1; - VERIFY_IS_APPROX(m1 - mones, m3); - - VERIFY_IS_APPROX(m2, m2.cwise() * mones); - VERIFY_IS_APPROX(m1.cwise() * m2, m2.cwise() * m1); - m3 = m1; - m3.cwise() *= m2; - VERIFY_IS_APPROX(m3, m1.cwise() * m2); - - VERIFY_IS_APPROX(mones, m2.cwise()/m2); - - // check min - VERIFY_IS_APPROX( m1.cwise().min(m2), m2.cwise().min(m1) ); - VERIFY_IS_APPROX( m1.cwise().min(m1+mones), m1 ); - VERIFY_IS_APPROX( m1.cwise().min(m1-mones), m1-mones ); - - // check max - VERIFY_IS_APPROX( m1.cwise().max(m2), m2.cwise().max(m1) ); - VERIFY_IS_APPROX( m1.cwise().max(m1-mones), m1 ); - VERIFY_IS_APPROX( m1.cwise().max(m1+mones), m1+mones ); - - VERIFY( (m1.cwise() == m1).all() ); - VERIFY( (m1.cwise() != m2).any() ); - VERIFY(!(m1.cwise() == (m1+mones)).any() ); - if (rows*cols>1) - { - m3 = m1; - m3(r,c) += 1; - VERIFY( (m1.cwise() == m3).any() ); - VERIFY( !(m1.cwise() == m3).all() ); - } - VERIFY( (m1.cwise().min(m2).cwise() <= m2).all() ); - VERIFY( (m1.cwise().max(m2).cwise() >= m2).all() ); - VERIFY( (m1.cwise().min(m2).cwise() < (m1+mones)).all() ); - VERIFY( (m1.cwise().max(m2).cwise() > (m1-mones)).all() ); - - VERIFY( (m1.cwise()(), Scalar(1)))).all() ); - VERIFY( !(m1.cwise()(), Scalar(1)))).all() ); - VERIFY( !(m1.cwise()>m1bis.unaryExpr(bind2nd(plus(), Scalar(1)))).any() ); - - cwiseops_real_only(m1, m2, m3, mones); -} - -void test_cwiseop() -{ - for(int i = 0; i < g_repeat ; i++) { - CALL_SUBTEST_1( cwiseops(Matrix()) ); - CALL_SUBTEST_2( cwiseops(Matrix4d()) ); - CALL_SUBTEST_3( cwiseops(MatrixXf(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); - CALL_SUBTEST_4( cwiseops(MatrixXf(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); - CALL_SUBTEST_5( cwiseops(MatrixXi(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); - CALL_SUBTEST_6( cwiseops(MatrixXd(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); - } -} diff --git a/test/eigen2/CMakeLists.txt b/test/eigen2/CMakeLists.txt deleted file mode 100644 index 41a02f4ad..000000000 --- a/test/eigen2/CMakeLists.txt +++ /dev/null @@ -1,65 +0,0 @@ -add_custom_target(eigen2_buildtests) -add_custom_target(eigen2_check COMMAND "ctest -R eigen2") -add_dependencies(eigen2_check eigen2_buildtests) -add_dependencies(buildtests eigen2_buildtests) - -add_definitions("-DEIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API") - -# Disable unused warnings for this module -# As EIGEN2 support is deprecated, it is not really worth fixing them -ei_add_cxx_compiler_flag("-Wno-unused-local-typedefs") -ei_add_cxx_compiler_flag("-Wno-unused-but-set-variable") - -ei_add_test(eigen2_meta) -ei_add_test(eigen2_sizeof) -ei_add_test(eigen2_dynalloc) -ei_add_test(eigen2_nomalloc) -#ei_add_test(eigen2_first_aligned) -ei_add_test(eigen2_mixingtypes) -#ei_add_test(eigen2_packetmath) -ei_add_test(eigen2_unalignedassert) -#ei_add_test(eigen2_vectorization_logic) -ei_add_test(eigen2_basicstuff) -ei_add_test(eigen2_linearstructure) -ei_add_test(eigen2_cwiseop) -ei_add_test(eigen2_sum) -ei_add_test(eigen2_product_small) -ei_add_test(eigen2_product_large ${EI_OFLAG}) -ei_add_test(eigen2_adjoint) -ei_add_test(eigen2_submatrices) -ei_add_test(eigen2_miscmatrices) -ei_add_test(eigen2_commainitializer) -ei_add_test(eigen2_smallvectors) -ei_add_test(eigen2_map) -ei_add_test(eigen2_array) -ei_add_test(eigen2_triangular) -ei_add_test(eigen2_cholesky " " "${GSL_LIBRARIES}") -ei_add_test(eigen2_lu ${EI_OFLAG}) -ei_add_test(eigen2_determinant ${EI_OFLAG}) -ei_add_test(eigen2_inverse) -ei_add_test(eigen2_qr) -ei_add_test(eigen2_eigensolver " " "${GSL_LIBRARIES}") -ei_add_test(eigen2_svd) -ei_add_test(eigen2_geometry) -ei_add_test(eigen2_geometry_with_eigen2_prefix) -ei_add_test(eigen2_hyperplane) -ei_add_test(eigen2_parametrizedline) -ei_add_test(eigen2_alignedbox) -ei_add_test(eigen2_regression) -ei_add_test(eigen2_stdvector) -ei_add_test(eigen2_newstdvector) -if(QT4_FOUND) - ei_add_test(eigen2_qtvector " " "${QT_QTCORE_LIBRARY}") -endif(QT4_FOUND) -# no support for eigen2 sparse module -# if(NOT EIGEN_DEFAULT_TO_ROW_MAJOR) -# ei_add_test(eigen2_sparse_vector) -# ei_add_test(eigen2_sparse_basic) -# ei_add_test(eigen2_sparse_solvers " " "${SPARSE_LIBS}") -# ei_add_test(eigen2_sparse_product) -# endif() -ei_add_test(eigen2_swap) -ei_add_test(eigen2_visitor) -ei_add_test(eigen2_bug_132) - -ei_add_test(eigen2_prec_inverse_4x4 ${EI_OFLAG}) diff --git a/test/eigen2/eigen2_adjoint.cpp b/test/eigen2/eigen2_adjoint.cpp deleted file mode 100644 index 8ec9c9202..000000000 --- a/test/eigen2/eigen2_adjoint.cpp +++ /dev/null @@ -1,101 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void adjoint(const MatrixType& m) -{ - /* this test covers the following files: - Transpose.h Conjugate.h Dot.h - */ - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - typedef Matrix SquareMatrixType; - int rows = m.rows(); - int cols = m.cols(); - - RealScalar largerEps = test_precision(); - if (ei_is_same_type::ret) - largerEps = RealScalar(1e-3f); - - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - mzero = MatrixType::Zero(rows, cols), - identity = SquareMatrixType::Identity(rows, rows), - square = SquareMatrixType::Random(rows, rows); - VectorType v1 = VectorType::Random(rows), - v2 = VectorType::Random(rows), - v3 = VectorType::Random(rows), - vzero = VectorType::Zero(rows); - - Scalar s1 = ei_random(), - s2 = ei_random(); - - // check basic compatibility of adjoint, transpose, conjugate - VERIFY_IS_APPROX(m1.transpose().conjugate().adjoint(), m1); - VERIFY_IS_APPROX(m1.adjoint().conjugate().transpose(), m1); - - // check multiplicative behavior - VERIFY_IS_APPROX((m1.adjoint() * m2).adjoint(), m2.adjoint() * m1); - VERIFY_IS_APPROX((s1 * m1).adjoint(), ei_conj(s1) * m1.adjoint()); - - // check basic properties of dot, norm, norm2 - typedef typename NumTraits::Real RealScalar; - VERIFY(ei_isApprox((s1 * v1 + s2 * v2).eigen2_dot(v3), s1 * v1.eigen2_dot(v3) + s2 * v2.eigen2_dot(v3), largerEps)); - VERIFY(ei_isApprox(v3.eigen2_dot(s1 * v1 + s2 * v2), ei_conj(s1)*v3.eigen2_dot(v1)+ei_conj(s2)*v3.eigen2_dot(v2), largerEps)); - VERIFY_IS_APPROX(ei_conj(v1.eigen2_dot(v2)), v2.eigen2_dot(v1)); - VERIFY_IS_APPROX(ei_real(v1.eigen2_dot(v1)), v1.squaredNorm()); - if(NumTraits::HasFloatingPoint) - VERIFY_IS_APPROX(v1.squaredNorm(), v1.norm() * v1.norm()); - VERIFY_IS_MUCH_SMALLER_THAN(ei_abs(vzero.eigen2_dot(v1)), static_cast(1)); - if(NumTraits::HasFloatingPoint) - VERIFY_IS_MUCH_SMALLER_THAN(vzero.norm(), static_cast(1)); - - // check compatibility of dot and adjoint - VERIFY(ei_isApprox(v1.eigen2_dot(square * v2), (square.adjoint() * v1).eigen2_dot(v2), largerEps)); - - // like in testBasicStuff, test operator() to check const-qualification - int r = ei_random(0, rows-1), - c = ei_random(0, cols-1); - VERIFY_IS_APPROX(m1.conjugate()(r,c), ei_conj(m1(r,c))); - VERIFY_IS_APPROX(m1.adjoint()(c,r), ei_conj(m1(r,c))); - - if(NumTraits::HasFloatingPoint) - { - // check that Random().normalized() works: tricky as the random xpr must be evaluated by - // normalized() in order to produce a consistent result. - VERIFY_IS_APPROX(VectorType::Random(rows).normalized().norm(), RealScalar(1)); - } - - // check inplace transpose - m3 = m1; - m3.transposeInPlace(); - VERIFY_IS_APPROX(m3,m1.transpose()); - m3.transposeInPlace(); - VERIFY_IS_APPROX(m3,m1); - -} - -void test_eigen2_adjoint() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( adjoint(Matrix()) ); - CALL_SUBTEST_2( adjoint(Matrix3d()) ); - CALL_SUBTEST_3( adjoint(Matrix4f()) ); - CALL_SUBTEST_4( adjoint(MatrixXcf(4, 4)) ); - CALL_SUBTEST_5( adjoint(MatrixXi(8, 12)) ); - CALL_SUBTEST_6( adjoint(MatrixXf(21, 21)) ); - } - // test a large matrix only once - CALL_SUBTEST_7( adjoint(Matrix()) ); -} - diff --git a/test/eigen2/eigen2_alignedbox.cpp b/test/eigen2/eigen2_alignedbox.cpp deleted file mode 100644 index 35043b958..000000000 --- a/test/eigen2/eigen2_alignedbox.cpp +++ /dev/null @@ -1,60 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include -#include -#include - -template void alignedbox(const BoxType& _box) -{ - /* this test covers the following files: - AlignedBox.h - */ - - const int dim = _box.dim(); - typedef typename BoxType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - VectorType p0 = VectorType::Random(dim); - VectorType p1 = VectorType::Random(dim); - RealScalar s1 = ei_random(0,1); - - BoxType b0(dim); - BoxType b1(VectorType::Random(dim),VectorType::Random(dim)); - BoxType b2; - - b0.extend(p0); - b0.extend(p1); - VERIFY(b0.contains(p0*s1+(Scalar(1)-s1)*p1)); - VERIFY(!b0.contains(p0 + (1+s1)*(p1-p0))); - - (b2 = b0).extend(b1); - VERIFY(b2.contains(b0)); - VERIFY(b2.contains(b1)); - VERIFY_IS_APPROX(b2.clamp(b0), b0); - - // casting - const int Dim = BoxType::AmbientDimAtCompileTime; - typedef typename GetDifferentType::type OtherScalar; - AlignedBox hp1f = b0.template cast(); - VERIFY_IS_APPROX(hp1f.template cast(),b0); - AlignedBox hp1d = b0.template cast(); - VERIFY_IS_APPROX(hp1d.template cast(),b0); -} - -void test_eigen2_alignedbox() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( alignedbox(AlignedBox()) ); - CALL_SUBTEST_2( alignedbox(AlignedBox()) ); - CALL_SUBTEST_3( alignedbox(AlignedBox()) ); - } -} diff --git a/test/eigen2/eigen2_array.cpp b/test/eigen2/eigen2_array.cpp deleted file mode 100644 index c1ff40ce7..000000000 --- a/test/eigen2/eigen2_array.cpp +++ /dev/null @@ -1,142 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -template void array(const MatrixType& m) -{ - /* this test covers the following files: - Array.cpp - */ - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - int rows = m.rows(); - int cols = m.cols(); - - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols); - - Scalar s1 = ei_random(), - s2 = ei_random(); - - // scalar addition - VERIFY_IS_APPROX(m1.cwise() + s1, s1 + m1.cwise()); - VERIFY_IS_APPROX(m1.cwise() + s1, MatrixType::Constant(rows,cols,s1) + m1); - VERIFY_IS_APPROX((m1*Scalar(2)).cwise() - s2, (m1+m1) - MatrixType::Constant(rows,cols,s2) ); - m3 = m1; - m3.cwise() += s2; - VERIFY_IS_APPROX(m3, m1.cwise() + s2); - m3 = m1; - m3.cwise() -= s1; - VERIFY_IS_APPROX(m3, m1.cwise() - s1); - - // reductions - VERIFY_IS_APPROX(m1.colwise().sum().sum(), m1.sum()); - VERIFY_IS_APPROX(m1.rowwise().sum().sum(), m1.sum()); - if (!ei_isApprox(m1.sum(), (m1+m2).sum())) - VERIFY_IS_NOT_APPROX(((m1+m2).rowwise().sum()).sum(), m1.sum()); - VERIFY_IS_APPROX(m1.colwise().sum(), m1.colwise().redux(internal::scalar_sum_op())); -} - -template void comparisons(const MatrixType& m) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - int rows = m.rows(); - int cols = m.cols(); - - int r = ei_random(0, rows-1), - c = ei_random(0, cols-1); - - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols); - - VERIFY(((m1.cwise() + Scalar(1)).cwise() > m1).all()); - VERIFY(((m1.cwise() - Scalar(1)).cwise() < m1).all()); - if (rows*cols>1) - { - m3 = m1; - m3(r,c) += 1; - VERIFY(! (m1.cwise() < m3).all() ); - VERIFY(! (m1.cwise() > m3).all() ); - } - - // comparisons to scalar - VERIFY( (m1.cwise() != (m1(r,c)+1) ).any() ); - VERIFY( (m1.cwise() > (m1(r,c)-1) ).any() ); - VERIFY( (m1.cwise() < (m1(r,c)+1) ).any() ); - VERIFY( (m1.cwise() == m1(r,c) ).any() ); - - // test Select - VERIFY_IS_APPROX( (m1.cwise()m2).select(m1,m2), m1.cwise().max(m2) ); - Scalar mid = (m1.cwise().abs().minCoeff() + m1.cwise().abs().maxCoeff())/Scalar(2); - for (int j=0; j=MatrixType::Constant(rows,cols,mid)) - .select(m1,0), m3); - // even shorter version: - VERIFY_IS_APPROX( (m1.cwise().abs().cwise()RealScalar(0.1)).count() == rows*cols); - VERIFY_IS_APPROX(((m1.cwise().abs().cwise()+1).cwise()>RealScalar(0.1)).colwise().count().template cast(), RowVectorXi::Constant(cols,rows)); - VERIFY_IS_APPROX(((m1.cwise().abs().cwise()+1).cwise()>RealScalar(0.1)).rowwise().count().template cast(), VectorXi::Constant(rows, cols)); -} - -template void lpNorm(const VectorType& v) -{ - VectorType u = VectorType::Random(v.size()); - - VERIFY_IS_APPROX(u.template lpNorm(), u.cwise().abs().maxCoeff()); - VERIFY_IS_APPROX(u.template lpNorm<1>(), u.cwise().abs().sum()); - VERIFY_IS_APPROX(u.template lpNorm<2>(), ei_sqrt(u.cwise().abs().cwise().square().sum())); - VERIFY_IS_APPROX(ei_pow(u.template lpNorm<5>(), typename VectorType::RealScalar(5)), u.cwise().abs().cwise().pow(5).sum()); -} - -void test_eigen2_array() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( array(Matrix()) ); - CALL_SUBTEST_2( array(Matrix2f()) ); - CALL_SUBTEST_3( array(Matrix4d()) ); - CALL_SUBTEST_4( array(MatrixXcf(3, 3)) ); - CALL_SUBTEST_5( array(MatrixXf(8, 12)) ); - CALL_SUBTEST_6( array(MatrixXi(8, 12)) ); - } - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( comparisons(Matrix()) ); - CALL_SUBTEST_2( comparisons(Matrix2f()) ); - CALL_SUBTEST_3( comparisons(Matrix4d()) ); - CALL_SUBTEST_5( comparisons(MatrixXf(8, 12)) ); - CALL_SUBTEST_6( comparisons(MatrixXi(8, 12)) ); - } - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( lpNorm(Matrix()) ); - CALL_SUBTEST_2( lpNorm(Vector2f()) ); - CALL_SUBTEST_3( lpNorm(Vector3d()) ); - CALL_SUBTEST_4( lpNorm(Vector4f()) ); - CALL_SUBTEST_5( lpNorm(VectorXf(16)) ); - CALL_SUBTEST_7( lpNorm(VectorXcd(10)) ); - } -} diff --git a/test/eigen2/eigen2_basicstuff.cpp b/test/eigen2/eigen2_basicstuff.cpp deleted file mode 100644 index 4fa16d5ae..000000000 --- a/test/eigen2/eigen2_basicstuff.cpp +++ /dev/null @@ -1,108 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void basicStuff(const MatrixType& m) -{ - typedef typename MatrixType::Scalar Scalar; - typedef Matrix VectorType; - - int rows = m.rows(); - int cols = m.cols(); - - // this test relies a lot on Random.h, and there's not much more that we can do - // to test it, hence I consider that we will have tested Random.h - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - mzero = MatrixType::Zero(rows, cols), - identity = Matrix - ::Identity(rows, rows), - square = Matrix::Random(rows, rows); - VectorType v1 = VectorType::Random(rows), - v2 = VectorType::Random(rows), - vzero = VectorType::Zero(rows); - - Scalar x = ei_random(); - - int r = ei_random(0, rows-1), - c = ei_random(0, cols-1); - - m1.coeffRef(r,c) = x; - VERIFY_IS_APPROX(x, m1.coeff(r,c)); - m1(r,c) = x; - VERIFY_IS_APPROX(x, m1(r,c)); - v1.coeffRef(r) = x; - VERIFY_IS_APPROX(x, v1.coeff(r)); - v1(r) = x; - VERIFY_IS_APPROX(x, v1(r)); - v1[r] = x; - VERIFY_IS_APPROX(x, v1[r]); - - VERIFY_IS_APPROX( v1, v1); - VERIFY_IS_NOT_APPROX( v1, 2*v1); - VERIFY_IS_MUCH_SMALLER_THAN( vzero, v1); - if(NumTraits::HasFloatingPoint) - VERIFY_IS_MUCH_SMALLER_THAN( vzero, v1.norm()); - VERIFY_IS_NOT_MUCH_SMALLER_THAN(v1, v1); - VERIFY_IS_APPROX( vzero, v1-v1); - VERIFY_IS_APPROX( m1, m1); - VERIFY_IS_NOT_APPROX( m1, 2*m1); - VERIFY_IS_MUCH_SMALLER_THAN( mzero, m1); - VERIFY_IS_NOT_MUCH_SMALLER_THAN(m1, m1); - VERIFY_IS_APPROX( mzero, m1-m1); - - // always test operator() on each read-only expression class, - // in order to check const-qualifiers. - // indeed, if an expression class (here Zero) is meant to be read-only, - // hence has no _write() method, the corresponding MatrixBase method (here zero()) - // should return a const-qualified object so that it is the const-qualified - // operator() that gets called, which in turn calls _read(). - VERIFY_IS_MUCH_SMALLER_THAN(MatrixType::Zero(rows,cols)(r,c), static_cast(1)); - - // now test copying a row-vector into a (column-)vector and conversely. - square.col(r) = square.row(r).eval(); - Matrix rv(rows); - Matrix cv(rows); - rv = square.row(r); - cv = square.col(r); - VERIFY_IS_APPROX(rv, cv.transpose()); - - if(cols!=1 && rows!=1 && MatrixType::SizeAtCompileTime!=Dynamic) - { - VERIFY_RAISES_ASSERT(m1 = (m2.block(0,0, rows-1, cols-1))); - } - - VERIFY_IS_APPROX(m3 = m1,m1); - MatrixType m4; - VERIFY_IS_APPROX(m4 = m1,m1); - - // test swap - m3 = m1; - m1.swap(m2); - VERIFY_IS_APPROX(m3, m2); - if(rows*cols>=3) - { - VERIFY_IS_NOT_APPROX(m3, m1); - } -} - -void test_eigen2_basicstuff() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( basicStuff(Matrix()) ); - CALL_SUBTEST_2( basicStuff(Matrix4d()) ); - CALL_SUBTEST_3( basicStuff(MatrixXcf(3, 3)) ); - CALL_SUBTEST_4( basicStuff(MatrixXi(8, 12)) ); - CALL_SUBTEST_5( basicStuff(MatrixXcd(20, 20)) ); - CALL_SUBTEST_6( basicStuff(Matrix()) ); - CALL_SUBTEST_7( basicStuff(Matrix(10,10)) ); - } -} diff --git a/test/eigen2/eigen2_bug_132.cpp b/test/eigen2/eigen2_bug_132.cpp deleted file mode 100644 index 7fe3610ce..000000000 --- a/test/eigen2/eigen2_bug_132.cpp +++ /dev/null @@ -1,26 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2010 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -void test_eigen2_bug_132() { - int size = 100; - MatrixXd A(size, size); - VectorXd b(size), c(size); - { - VectorXd y = A.transpose() * (b-c); // bug 132: infinite recursion in coeffRef - VectorXd z = (b-c).transpose() * A; // bug 132: infinite recursion in coeffRef - } - - // the following ones weren't failing, but let's include them for completeness: - { - VectorXd y = A * (b-c); - VectorXd z = (b-c).transpose() * A.transpose(); - } -} diff --git a/test/eigen2/eigen2_cholesky.cpp b/test/eigen2/eigen2_cholesky.cpp deleted file mode 100644 index 9c4b6f561..000000000 --- a/test/eigen2/eigen2_cholesky.cpp +++ /dev/null @@ -1,113 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#define EIGEN_NO_ASSERTION_CHECKING -#include "main.h" -#include -#include - -#ifdef HAS_GSL -#include "gsl_helper.h" -#endif - -template void cholesky(const MatrixType& m) -{ - /* this test covers the following files: - LLT.h LDLT.h - */ - int rows = m.rows(); - int cols = m.cols(); - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix SquareMatrixType; - typedef Matrix VectorType; - - MatrixType a0 = MatrixType::Random(rows,cols); - VectorType vecB = VectorType::Random(rows), vecX(rows); - MatrixType matB = MatrixType::Random(rows,cols), matX(rows,cols); - SquareMatrixType symm = a0 * a0.adjoint(); - // let's make sure the matrix is not singular or near singular - MatrixType a1 = MatrixType::Random(rows,cols); - symm += a1 * a1.adjoint(); - - #ifdef HAS_GSL - if (ei_is_same_type::ret) - { - typedef GslTraits Gsl; - typename Gsl::Matrix gMatA=0, gSymm=0; - typename Gsl::Vector gVecB=0, gVecX=0; - convert(symm, gSymm); - convert(symm, gMatA); - convert(vecB, gVecB); - convert(vecB, gVecX); - Gsl::cholesky(gMatA); - Gsl::cholesky_solve(gMatA, gVecB, gVecX); - VectorType vecX(rows), _vecX, _vecB; - convert(gVecX, _vecX); - symm.llt().solve(vecB, &vecX); - Gsl::prod(gSymm, gVecX, gVecB); - convert(gVecB, _vecB); - // test gsl itself ! - VERIFY_IS_APPROX(vecB, _vecB); - VERIFY_IS_APPROX(vecX, _vecX); - - Gsl::free(gMatA); - Gsl::free(gSymm); - Gsl::free(gVecB); - Gsl::free(gVecX); - } - #endif - - { - LDLT ldlt(symm); - VERIFY(ldlt.isPositiveDefinite()); - // in eigen3, LDLT is pivoting - //VERIFY_IS_APPROX(symm, ldlt.matrixL() * ldlt.vectorD().asDiagonal() * ldlt.matrixL().adjoint()); - ldlt.solve(vecB, &vecX); - VERIFY_IS_APPROX(symm * vecX, vecB); - ldlt.solve(matB, &matX); - VERIFY_IS_APPROX(symm * matX, matB); - } - - { - LLT chol(symm); - VERIFY(chol.isPositiveDefinite()); - VERIFY_IS_APPROX(symm, chol.matrixL() * chol.matrixL().adjoint()); - chol.solve(vecB, &vecX); - VERIFY_IS_APPROX(symm * vecX, vecB); - chol.solve(matB, &matX); - VERIFY_IS_APPROX(symm * matX, matB); - } - -#if 0 // cholesky is not rank-revealing anyway - // test isPositiveDefinite on non definite matrix - if (rows>4) - { - SquareMatrixType symm = a0.block(0,0,rows,cols-4) * a0.block(0,0,rows,cols-4).adjoint(); - LLT chol(symm); - VERIFY(!chol.isPositiveDefinite()); - LDLT cholnosqrt(symm); - VERIFY(!cholnosqrt.isPositiveDefinite()); - } -#endif -} - -void test_eigen2_cholesky() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( cholesky(Matrix()) ); - CALL_SUBTEST_2( cholesky(Matrix2d()) ); - CALL_SUBTEST_3( cholesky(Matrix3f()) ); - CALL_SUBTEST_4( cholesky(Matrix4d()) ); - CALL_SUBTEST_5( cholesky(MatrixXcd(7,7)) ); - CALL_SUBTEST_6( cholesky(MatrixXf(17,17)) ); - CALL_SUBTEST_7( cholesky(MatrixXd(33,33)) ); - } -} diff --git a/test/eigen2/eigen2_commainitializer.cpp b/test/eigen2/eigen2_commainitializer.cpp deleted file mode 100644 index e0f901e0b..000000000 --- a/test/eigen2/eigen2_commainitializer.cpp +++ /dev/null @@ -1,46 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -void test_eigen2_commainitializer() -{ - Matrix3d m3; - Matrix4d m4; - - VERIFY_RAISES_ASSERT( (m3 << 1, 2, 3, 4, 5, 6, 7, 8) ); - - #ifndef _MSC_VER - VERIFY_RAISES_ASSERT( (m3 << 1, 2, 3, 4, 5, 6, 7, 8, 9, 10) ); - #endif - - double data[] = {1, 2, 3, 4, 5, 6, 7, 8, 9}; - Matrix3d ref = Map >(data); - - m3 = Matrix3d::Random(); - m3 << 1, 2, 3, 4, 5, 6, 7, 8, 9; - VERIFY_IS_APPROX(m3, ref ); - - Vector3d vec[3]; - vec[0] << 1, 4, 7; - vec[1] << 2, 5, 8; - vec[2] << 3, 6, 9; - m3 = Matrix3d::Random(); - m3 << vec[0], vec[1], vec[2]; - VERIFY_IS_APPROX(m3, ref); - - vec[0] << 1, 2, 3; - vec[1] << 4, 5, 6; - vec[2] << 7, 8, 9; - m3 = Matrix3d::Random(); - m3 << vec[0].transpose(), - 4, 5, 6, - vec[2].transpose(); - VERIFY_IS_APPROX(m3, ref); -} diff --git a/test/eigen2/eigen2_cwiseop.cpp b/test/eigen2/eigen2_cwiseop.cpp deleted file mode 100644 index 4391d19b4..000000000 --- a/test/eigen2/eigen2_cwiseop.cpp +++ /dev/null @@ -1,158 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include -#include - -using namespace std; - -template struct AddIfNull { - const Scalar operator() (const Scalar a, const Scalar b) const {return a<=1e-3 ? b : a;} - enum { Cost = NumTraits::AddCost }; -}; - -template void cwiseops(const MatrixType& m) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - int rows = m.rows(); - int cols = m.cols(); - - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - m4(rows, cols), - mzero = MatrixType::Zero(rows, cols), - mones = MatrixType::Ones(rows, cols), - identity = Matrix - ::Identity(rows, rows), - square = Matrix::Random(rows, rows); - VectorType v1 = VectorType::Random(rows), - v2 = VectorType::Random(rows), - vzero = VectorType::Zero(rows), - vones = VectorType::Ones(rows), - v3(rows); - - int r = ei_random(0, rows-1), - c = ei_random(0, cols-1); - - Scalar s1 = ei_random(); - - // test Zero, Ones, Constant, and the set* variants - m3 = MatrixType::Constant(rows, cols, s1); - for (int j=0; j >(mones); - - VERIFY_IS_APPROX(m1.cwise().pow(2), m1.cwise().abs2()); - VERIFY_IS_APPROX(m1.cwise().pow(2), m1.cwise().square()); - VERIFY_IS_APPROX(m1.cwise().pow(3), m1.cwise().cube()); - - VERIFY_IS_APPROX(m1 + mones, m1.cwise()+Scalar(1)); - VERIFY_IS_APPROX(m1 - mones, m1.cwise()-Scalar(1)); - m3 = m1; m3.cwise() += 1; - VERIFY_IS_APPROX(m1 + mones, m3); - m3 = m1; m3.cwise() -= 1; - VERIFY_IS_APPROX(m1 - mones, m3); - - VERIFY_IS_APPROX(m2, m2.cwise() * mones); - VERIFY_IS_APPROX(m1.cwise() * m2, m2.cwise() * m1); - m3 = m1; - m3.cwise() *= m2; - VERIFY_IS_APPROX(m3, m1.cwise() * m2); - - VERIFY_IS_APPROX(mones, m2.cwise()/m2); - if(NumTraits::HasFloatingPoint) - { - VERIFY_IS_APPROX(m1.cwise() / m2, m1.cwise() * (m2.cwise().inverse())); - m3 = m1.cwise().abs().cwise().sqrt(); - VERIFY_IS_APPROX(m3.cwise().square(), m1.cwise().abs()); - VERIFY_IS_APPROX(m1.cwise().square().cwise().sqrt(), m1.cwise().abs()); - VERIFY_IS_APPROX(m1.cwise().abs().cwise().log().cwise().exp() , m1.cwise().abs()); - - VERIFY_IS_APPROX(m1.cwise().pow(2), m1.cwise().square()); - m3 = (m1.cwise().abs().cwise()<=RealScalar(0.01)).select(mones,m1); - VERIFY_IS_APPROX(m3.cwise().pow(-1), m3.cwise().inverse()); - m3 = m1.cwise().abs(); - VERIFY_IS_APPROX(m3.cwise().pow(RealScalar(0.5)), m3.cwise().sqrt()); - -// VERIFY_IS_APPROX(m1.cwise().tan(), m1.cwise().sin().cwise() / m1.cwise().cos()); - VERIFY_IS_APPROX(mones, m1.cwise().sin().cwise().square() + m1.cwise().cos().cwise().square()); - m3 = m1; - m3.cwise() /= m2; - VERIFY_IS_APPROX(m3, m1.cwise() / m2); - } - - // check min - VERIFY_IS_APPROX( m1.cwise().min(m2), m2.cwise().min(m1) ); - VERIFY_IS_APPROX( m1.cwise().min(m1+mones), m1 ); - VERIFY_IS_APPROX( m1.cwise().min(m1-mones), m1-mones ); - - // check max - VERIFY_IS_APPROX( m1.cwise().max(m2), m2.cwise().max(m1) ); - VERIFY_IS_APPROX( m1.cwise().max(m1-mones), m1 ); - VERIFY_IS_APPROX( m1.cwise().max(m1+mones), m1+mones ); - - VERIFY( (m1.cwise() == m1).all() ); - VERIFY( (m1.cwise() != m2).any() ); - VERIFY(!(m1.cwise() == (m1+mones)).any() ); - if (rows*cols>1) - { - m3 = m1; - m3(r,c) += 1; - VERIFY( (m1.cwise() == m3).any() ); - VERIFY( !(m1.cwise() == m3).all() ); - } - VERIFY( (m1.cwise().min(m2).cwise() <= m2).all() ); - VERIFY( (m1.cwise().max(m2).cwise() >= m2).all() ); - VERIFY( (m1.cwise().min(m2).cwise() < (m1+mones)).all() ); - VERIFY( (m1.cwise().max(m2).cwise() > (m1-mones)).all() ); - - VERIFY( (m1.cwise()(), Scalar(1)))).all() ); - VERIFY( !(m1.cwise()(), Scalar(1)))).all() ); - VERIFY( !(m1.cwise()>m1.unaryExpr(bind2nd(plus(), Scalar(1)))).any() ); -} - -void test_eigen2_cwiseop() -{ - for(int i = 0; i < g_repeat ; i++) { - CALL_SUBTEST_1( cwiseops(Matrix()) ); - CALL_SUBTEST_2( cwiseops(Matrix4d()) ); - CALL_SUBTEST_3( cwiseops(MatrixXf(3, 3)) ); - CALL_SUBTEST_3( cwiseops(MatrixXf(22, 22)) ); - CALL_SUBTEST_4( cwiseops(MatrixXi(8, 12)) ); - CALL_SUBTEST_5( cwiseops(MatrixXd(20, 20)) ); - } -} diff --git a/test/eigen2/eigen2_determinant.cpp b/test/eigen2/eigen2_determinant.cpp deleted file mode 100644 index c7b4ad053..000000000 --- a/test/eigen2/eigen2_determinant.cpp +++ /dev/null @@ -1,61 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Benoit Jacob -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -template void determinant(const MatrixType& m) -{ - /* this test covers the following files: - Determinant.h - */ - int size = m.rows(); - - MatrixType m1(size, size), m2(size, size); - m1.setRandom(); - m2.setRandom(); - typedef typename MatrixType::Scalar Scalar; - Scalar x = ei_random(); - VERIFY_IS_APPROX(MatrixType::Identity(size, size).determinant(), Scalar(1)); - VERIFY_IS_APPROX((m1*m2).determinant(), m1.determinant() * m2.determinant()); - if(size==1) return; - int i = ei_random(0, size-1); - int j; - do { - j = ei_random(0, size-1); - } while(j==i); - m2 = m1; - m2.row(i).swap(m2.row(j)); - VERIFY_IS_APPROX(m2.determinant(), -m1.determinant()); - m2 = m1; - m2.col(i).swap(m2.col(j)); - VERIFY_IS_APPROX(m2.determinant(), -m1.determinant()); - VERIFY_IS_APPROX(m2.determinant(), m2.transpose().determinant()); - VERIFY_IS_APPROX(ei_conj(m2.determinant()), m2.adjoint().determinant()); - m2 = m1; - m2.row(i) += x*m2.row(j); - VERIFY_IS_APPROX(m2.determinant(), m1.determinant()); - m2 = m1; - m2.row(i) *= x; - VERIFY_IS_APPROX(m2.determinant(), m1.determinant() * x); -} - -void test_eigen2_determinant() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( determinant(Matrix()) ); - CALL_SUBTEST_2( determinant(Matrix()) ); - CALL_SUBTEST_3( determinant(Matrix()) ); - CALL_SUBTEST_4( determinant(Matrix()) ); - CALL_SUBTEST_5( determinant(Matrix, 10, 10>()) ); - CALL_SUBTEST_6( determinant(MatrixXd(20, 20)) ); - } - CALL_SUBTEST_6( determinant(MatrixXd(200, 200)) ); -} diff --git a/test/eigen2/eigen2_dynalloc.cpp b/test/eigen2/eigen2_dynalloc.cpp deleted file mode 100644 index 1891a9e33..000000000 --- a/test/eigen2/eigen2_dynalloc.cpp +++ /dev/null @@ -1,131 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -#if EIGEN_ARCH_WANTS_ALIGNMENT -#define ALIGNMENT 16 -#else -#define ALIGNMENT 1 -#endif - -void check_handmade_aligned_malloc() -{ - for(int i = 1; i < 1000; i++) - { - char *p = (char*)ei_handmade_aligned_malloc(i); - VERIFY(std::size_t(p)%ALIGNMENT==0); - // if the buffer is wrongly allocated this will give a bad write --> check with valgrind - for(int j = 0; j < i; j++) p[j]=0; - ei_handmade_aligned_free(p); - } -} - -void check_aligned_malloc() -{ - for(int i = 1; i < 1000; i++) - { - char *p = (char*)ei_aligned_malloc(i); - VERIFY(std::size_t(p)%ALIGNMENT==0); - // if the buffer is wrongly allocated this will give a bad write --> check with valgrind - for(int j = 0; j < i; j++) p[j]=0; - ei_aligned_free(p); - } -} - -void check_aligned_new() -{ - for(int i = 1; i < 1000; i++) - { - float *p = ei_aligned_new(i); - VERIFY(std::size_t(p)%ALIGNMENT==0); - // if the buffer is wrongly allocated this will give a bad write --> check with valgrind - for(int j = 0; j < i; j++) p[j]=0; - ei_aligned_delete(p,i); - } -} - -void check_aligned_stack_alloc() -{ - for(int i = 1; i < 1000; i++) - { - ei_declare_aligned_stack_constructed_variable(float, p, i, 0); - VERIFY(std::size_t(p)%ALIGNMENT==0); - // if the buffer is wrongly allocated this will give a bad write --> check with valgrind - for(int j = 0; j < i; j++) p[j]=0; - } -} - - -// test compilation with both a struct and a class... -struct MyStruct -{ - EIGEN_MAKE_ALIGNED_OPERATOR_NEW - char dummychar; - Vector4f avec; -}; - -class MyClassA -{ - public: - EIGEN_MAKE_ALIGNED_OPERATOR_NEW - char dummychar; - Vector4f avec; -}; - -template void check_dynaligned() -{ - T* obj = new T; - VERIFY(std::size_t(obj)%ALIGNMENT==0); - delete obj; -} - -void test_eigen2_dynalloc() -{ - // low level dynamic memory allocation - CALL_SUBTEST(check_handmade_aligned_malloc()); - CALL_SUBTEST(check_aligned_malloc()); - CALL_SUBTEST(check_aligned_new()); - CALL_SUBTEST(check_aligned_stack_alloc()); - - for (int i=0; i() ); - CALL_SUBTEST( check_dynaligned() ); - CALL_SUBTEST( check_dynaligned() ); - CALL_SUBTEST( check_dynaligned() ); - CALL_SUBTEST( check_dynaligned() ); - } - - // check static allocation, who knows ? - { - MyStruct foo0; VERIFY(std::size_t(foo0.avec.data())%ALIGNMENT==0); - MyClassA fooA; VERIFY(std::size_t(fooA.avec.data())%ALIGNMENT==0); - } - - // dynamic allocation, single object - for (int i=0; iavec.data())%ALIGNMENT==0); - MyClassA *fooA = new MyClassA(); VERIFY(std::size_t(fooA->avec.data())%ALIGNMENT==0); - delete foo0; - delete fooA; - } - - // dynamic allocation, array - const int N = 10; - for (int i=0; iavec.data())%ALIGNMENT==0); - MyClassA *fooA = new MyClassA[N]; VERIFY(std::size_t(fooA->avec.data())%ALIGNMENT==0); - delete[] foo0; - delete[] fooA; - } - -} diff --git a/test/eigen2/eigen2_eigensolver.cpp b/test/eigen2/eigen2_eigensolver.cpp deleted file mode 100644 index 48b4ace43..000000000 --- a/test/eigen2/eigen2_eigensolver.cpp +++ /dev/null @@ -1,146 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -#ifdef HAS_GSL -#include "gsl_helper.h" -#endif - -template void selfadjointeigensolver(const MatrixType& m) -{ - /* this test covers the following files: - EigenSolver.h, SelfAdjointEigenSolver.h (and indirectly: Tridiagonalization.h) - */ - int rows = m.rows(); - int cols = m.cols(); - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - typedef Matrix RealVectorType; - typedef typename std::complex::Real> Complex; - - RealScalar largerEps = 10*test_precision(); - - MatrixType a = MatrixType::Random(rows,cols); - MatrixType a1 = MatrixType::Random(rows,cols); - MatrixType symmA = a.adjoint() * a + a1.adjoint() * a1; - - MatrixType b = MatrixType::Random(rows,cols); - MatrixType b1 = MatrixType::Random(rows,cols); - MatrixType symmB = b.adjoint() * b + b1.adjoint() * b1; - - SelfAdjointEigenSolver eiSymm(symmA); - // generalized eigen pb - SelfAdjointEigenSolver eiSymmGen(symmA, symmB); - - #ifdef HAS_GSL - if (ei_is_same_type::ret) - { - typedef GslTraits Gsl; - typename Gsl::Matrix gEvec=0, gSymmA=0, gSymmB=0; - typename GslTraits::Vector gEval=0; - RealVectorType _eval; - MatrixType _evec; - convert(symmA, gSymmA); - convert(symmB, gSymmB); - convert(symmA, gEvec); - gEval = GslTraits::createVector(rows); - - Gsl::eigen_symm(gSymmA, gEval, gEvec); - convert(gEval, _eval); - convert(gEvec, _evec); - - // test gsl itself ! - VERIFY((symmA * _evec).isApprox(_evec * _eval.asDiagonal(), largerEps)); - - // compare with eigen - VERIFY_IS_APPROX(_eval, eiSymm.eigenvalues()); - VERIFY_IS_APPROX(_evec.cwise().abs(), eiSymm.eigenvectors().cwise().abs()); - - // generalized pb - Gsl::eigen_symm_gen(gSymmA, gSymmB, gEval, gEvec); - convert(gEval, _eval); - convert(gEvec, _evec); - // test GSL itself: - VERIFY((symmA * _evec).isApprox(symmB * (_evec * _eval.asDiagonal()), largerEps)); - - // compare with eigen - MatrixType normalized_eivec = eiSymmGen.eigenvectors()*eiSymmGen.eigenvectors().colwise().norm().asDiagonal().inverse(); - VERIFY_IS_APPROX(_eval, eiSymmGen.eigenvalues()); - VERIFY_IS_APPROX(_evec.cwiseAbs(), normalized_eivec.cwiseAbs()); - - Gsl::free(gSymmA); - Gsl::free(gSymmB); - GslTraits::free(gEval); - Gsl::free(gEvec); - } - #endif - - VERIFY((symmA * eiSymm.eigenvectors()).isApprox( - eiSymm.eigenvectors() * eiSymm.eigenvalues().asDiagonal(), largerEps)); - - // generalized eigen problem Ax = lBx - VERIFY((symmA * eiSymmGen.eigenvectors()).isApprox( - symmB * (eiSymmGen.eigenvectors() * eiSymmGen.eigenvalues().asDiagonal()), largerEps)); - - MatrixType sqrtSymmA = eiSymm.operatorSqrt(); - VERIFY_IS_APPROX(symmA, sqrtSymmA*sqrtSymmA); - VERIFY_IS_APPROX(sqrtSymmA, symmA*eiSymm.operatorInverseSqrt()); -} - -template void eigensolver(const MatrixType& m) -{ - /* this test covers the following files: - EigenSolver.h - */ - int rows = m.rows(); - int cols = m.cols(); - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - typedef Matrix RealVectorType; - typedef typename std::complex::Real> Complex; - - // RealScalar largerEps = 10*test_precision(); - - MatrixType a = MatrixType::Random(rows,cols); - MatrixType a1 = MatrixType::Random(rows,cols); - MatrixType symmA = a.adjoint() * a + a1.adjoint() * a1; - - EigenSolver ei0(symmA); - VERIFY_IS_APPROX(symmA * ei0.pseudoEigenvectors(), ei0.pseudoEigenvectors() * ei0.pseudoEigenvalueMatrix()); - VERIFY_IS_APPROX((symmA.template cast()) * (ei0.pseudoEigenvectors().template cast()), - (ei0.pseudoEigenvectors().template cast()) * (ei0.eigenvalues().asDiagonal())); - - EigenSolver ei1(a); - VERIFY_IS_APPROX(a * ei1.pseudoEigenvectors(), ei1.pseudoEigenvectors() * ei1.pseudoEigenvalueMatrix()); - VERIFY_IS_APPROX(a.template cast() * ei1.eigenvectors(), - ei1.eigenvectors() * ei1.eigenvalues().asDiagonal()); - -} - -void test_eigen2_eigensolver() -{ - for(int i = 0; i < g_repeat; i++) { - // very important to test a 3x3 matrix since we provide a special path for it - CALL_SUBTEST_1( selfadjointeigensolver(Matrix3f()) ); - CALL_SUBTEST_2( selfadjointeigensolver(Matrix4d()) ); - CALL_SUBTEST_3( selfadjointeigensolver(MatrixXf(7,7)) ); - CALL_SUBTEST_4( selfadjointeigensolver(MatrixXcd(5,5)) ); - CALL_SUBTEST_5( selfadjointeigensolver(MatrixXd(19,19)) ); - - CALL_SUBTEST_6( eigensolver(Matrix4f()) ); - CALL_SUBTEST_5( eigensolver(MatrixXd(17,17)) ); - } -} - diff --git a/test/eigen2/eigen2_first_aligned.cpp b/test/eigen2/eigen2_first_aligned.cpp deleted file mode 100644 index 51bb3cad1..000000000 --- a/test/eigen2/eigen2_first_aligned.cpp +++ /dev/null @@ -1,49 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template -void test_eigen2_first_aligned_helper(Scalar *array, int size) -{ - const int packet_size = sizeof(Scalar) * ei_packet_traits::size; - VERIFY(((std::size_t(array) + sizeof(Scalar) * ei_alignmentOffset(array, size)) % packet_size) == 0); -} - -template -void test_eigen2_none_aligned_helper(Scalar *array, int size) -{ - VERIFY(ei_packet_traits::size == 1 || ei_alignmentOffset(array, size) == size); -} - -struct some_non_vectorizable_type { float x; }; - -void test_eigen2_first_aligned() -{ - EIGEN_ALIGN_128 float array_float[100]; - test_first_aligned_helper(array_float, 50); - test_first_aligned_helper(array_float+1, 50); - test_first_aligned_helper(array_float+2, 50); - test_first_aligned_helper(array_float+3, 50); - test_first_aligned_helper(array_float+4, 50); - test_first_aligned_helper(array_float+5, 50); - - EIGEN_ALIGN_128 double array_double[100]; - test_first_aligned_helper(array_double, 50); - test_first_aligned_helper(array_double+1, 50); - test_first_aligned_helper(array_double+2, 50); - - double *array_double_plus_4_bytes = (double*)(std::size_t(array_double)+4); - test_none_aligned_helper(array_double_plus_4_bytes, 50); - test_none_aligned_helper(array_double_plus_4_bytes+1, 50); - - some_non_vectorizable_type array_nonvec[100]; - test_first_aligned_helper(array_nonvec, 100); - test_none_aligned_helper(array_nonvec, 100); -} diff --git a/test/eigen2/eigen2_geometry.cpp b/test/eigen2/eigen2_geometry.cpp deleted file mode 100644 index 70b4ab5f1..000000000 --- a/test/eigen2/eigen2_geometry.cpp +++ /dev/null @@ -1,431 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include -#include -#include - -template void geometry(void) -{ - /* this test covers the following files: - Cross.h Quaternion.h, Transform.cpp - */ - - typedef Matrix Matrix2; - typedef Matrix Matrix3; - typedef Matrix Matrix4; - typedef Matrix Vector2; - typedef Matrix Vector3; - typedef Matrix Vector4; - typedef Quaternion Quaternionx; - typedef AngleAxis AngleAxisx; - typedef Transform Transform2; - typedef Transform Transform3; - typedef Scaling Scaling2; - typedef Scaling Scaling3; - typedef Translation Translation2; - typedef Translation Translation3; - - Scalar largeEps = test_precision(); - if (ei_is_same_type::ret) - largeEps = 1e-2f; - - Vector3 v0 = Vector3::Random(), - v1 = Vector3::Random(), - v2 = Vector3::Random(); - Vector2 u0 = Vector2::Random(); - Matrix3 matrot1; - - Scalar a = ei_random(-Scalar(M_PI), Scalar(M_PI)); - - // cross product - VERIFY_IS_MUCH_SMALLER_THAN(v1.cross(v2).eigen2_dot(v1), Scalar(1)); - Matrix3 m; - m << v0.normalized(), - (v0.cross(v1)).normalized(), - (v0.cross(v1).cross(v0)).normalized(); - VERIFY(m.isUnitary()); - - // Quaternion: Identity(), setIdentity(); - Quaternionx q1, q2; - q2.setIdentity(); - VERIFY_IS_APPROX(Quaternionx(Quaternionx::Identity()).coeffs(), q2.coeffs()); - q1.coeffs().setRandom(); - VERIFY_IS_APPROX(q1.coeffs(), (q1*q2).coeffs()); - - // unitOrthogonal - VERIFY_IS_MUCH_SMALLER_THAN(u0.unitOrthogonal().eigen2_dot(u0), Scalar(1)); - VERIFY_IS_MUCH_SMALLER_THAN(v0.unitOrthogonal().eigen2_dot(v0), Scalar(1)); - VERIFY_IS_APPROX(u0.unitOrthogonal().norm(), Scalar(1)); - VERIFY_IS_APPROX(v0.unitOrthogonal().norm(), Scalar(1)); - - - VERIFY_IS_APPROX(v0, AngleAxisx(a, v0.normalized()) * v0); - VERIFY_IS_APPROX(-v0, AngleAxisx(Scalar(M_PI), v0.unitOrthogonal()) * v0); - VERIFY_IS_APPROX(ei_cos(a)*v0.squaredNorm(), v0.eigen2_dot(AngleAxisx(a, v0.unitOrthogonal()) * v0)); - m = AngleAxisx(a, v0.normalized()).toRotationMatrix().adjoint(); - VERIFY_IS_APPROX(Matrix3::Identity(), m * AngleAxisx(a, v0.normalized())); - VERIFY_IS_APPROX(Matrix3::Identity(), AngleAxisx(a, v0.normalized()) * m); - - q1 = AngleAxisx(a, v0.normalized()); - q2 = AngleAxisx(a, v1.normalized()); - - // angular distance - Scalar refangle = ei_abs(AngleAxisx(q1.inverse()*q2).angle()); - if (refangle>Scalar(M_PI)) - refangle = Scalar(2)*Scalar(M_PI) - refangle; - - if((q1.coeffs()-q2.coeffs()).norm() > 10*largeEps) - { - VERIFY(ei_isApprox(q1.angularDistance(q2), refangle, largeEps)); - } - - // rotation matrix conversion - VERIFY_IS_APPROX(q1 * v2, q1.toRotationMatrix() * v2); - VERIFY_IS_APPROX(q1 * q2 * v2, - q1.toRotationMatrix() * q2.toRotationMatrix() * v2); - - VERIFY( (q2*q1).isApprox(q1*q2, largeEps) || !(q2 * q1 * v2).isApprox( - q1.toRotationMatrix() * q2.toRotationMatrix() * v2)); - - q2 = q1.toRotationMatrix(); - VERIFY_IS_APPROX(q1*v1,q2*v1); - - matrot1 = AngleAxisx(Scalar(0.1), Vector3::UnitX()) - * AngleAxisx(Scalar(0.2), Vector3::UnitY()) - * AngleAxisx(Scalar(0.3), Vector3::UnitZ()); - VERIFY_IS_APPROX(matrot1 * v1, - AngleAxisx(Scalar(0.1), Vector3(1,0,0)).toRotationMatrix() - * (AngleAxisx(Scalar(0.2), Vector3(0,1,0)).toRotationMatrix() - * (AngleAxisx(Scalar(0.3), Vector3(0,0,1)).toRotationMatrix() * v1))); - - // angle-axis conversion - AngleAxisx aa = q1; - VERIFY_IS_APPROX(q1 * v1, Quaternionx(aa) * v1); - VERIFY_IS_NOT_APPROX(q1 * v1, Quaternionx(AngleAxisx(aa.angle()*2,aa.axis())) * v1); - - // from two vector creation - VERIFY_IS_APPROX(v2.normalized(),(q2.setFromTwoVectors(v1,v2)*v1).normalized()); - VERIFY_IS_APPROX(v2.normalized(),(q2.setFromTwoVectors(v1,v2)*v1).normalized()); - - // inverse and conjugate - VERIFY_IS_APPROX(q1 * (q1.inverse() * v1), v1); - VERIFY_IS_APPROX(q1 * (q1.conjugate() * v1), v1); - - // AngleAxis - VERIFY_IS_APPROX(AngleAxisx(a,v1.normalized()).toRotationMatrix(), - Quaternionx(AngleAxisx(a,v1.normalized())).toRotationMatrix()); - - AngleAxisx aa1; - m = q1.toRotationMatrix(); - aa1 = m; - VERIFY_IS_APPROX(AngleAxisx(m).toRotationMatrix(), - Quaternionx(m).toRotationMatrix()); - - // Transform - // TODO complete the tests ! - a = 0; - while (ei_abs(a)(-Scalar(0.4)*Scalar(M_PI), Scalar(0.4)*Scalar(M_PI)); - q1 = AngleAxisx(a, v0.normalized()); - Transform3 t0, t1, t2; - // first test setIdentity() and Identity() - t0.setIdentity(); - VERIFY_IS_APPROX(t0.matrix(), Transform3::MatrixType::Identity()); - t0.matrix().setZero(); - t0 = Transform3::Identity(); - VERIFY_IS_APPROX(t0.matrix(), Transform3::MatrixType::Identity()); - - t0.linear() = q1.toRotationMatrix(); - t1.setIdentity(); - t1.linear() = q1.toRotationMatrix(); - - v0 << 50, 2, 1;//= ei_random_matrix().cwiseProduct(Vector3(10,2,0.5)); - t0.scale(v0); - t1.prescale(v0); - - VERIFY_IS_APPROX( (t0 * Vector3(1,0,0)).norm(), v0.x()); - //VERIFY(!ei_isApprox((t1 * Vector3(1,0,0)).norm(), v0.x())); - - t0.setIdentity(); - t1.setIdentity(); - v1 << 1, 2, 3; - t0.linear() = q1.toRotationMatrix(); - t0.pretranslate(v0); - t0.scale(v1); - t1.linear() = q1.conjugate().toRotationMatrix(); - t1.prescale(v1.cwise().inverse()); - t1.translate(-v0); - - VERIFY((t0.matrix() * t1.matrix()).isIdentity(test_precision())); - - t1.fromPositionOrientationScale(v0, q1, v1); - VERIFY_IS_APPROX(t1.matrix(), t0.matrix()); - VERIFY_IS_APPROX(t1*v1, t0*v1); - - t0.setIdentity(); t0.scale(v0).rotate(q1.toRotationMatrix()); - t1.setIdentity(); t1.scale(v0).rotate(q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - t0.setIdentity(); t0.scale(v0).rotate(AngleAxisx(q1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - VERIFY_IS_APPROX(t0.scale(a).matrix(), t1.scale(Vector3::Constant(a)).matrix()); - VERIFY_IS_APPROX(t0.prescale(a).matrix(), t1.prescale(Vector3::Constant(a)).matrix()); - - // More transform constructors, operator=, operator*= - - Matrix3 mat3 = Matrix3::Random(); - Matrix4 mat4; - mat4 << mat3 , Vector3::Zero() , Vector4::Zero().transpose(); - Transform3 tmat3(mat3), tmat4(mat4); - tmat4.matrix()(3,3) = Scalar(1); - VERIFY_IS_APPROX(tmat3.matrix(), tmat4.matrix()); - - Scalar a3 = ei_random(-Scalar(M_PI), Scalar(M_PI)); - Vector3 v3 = Vector3::Random().normalized(); - AngleAxisx aa3(a3, v3); - Transform3 t3(aa3); - Transform3 t4; - t4 = aa3; - VERIFY_IS_APPROX(t3.matrix(), t4.matrix()); - t4.rotate(AngleAxisx(-a3,v3)); - VERIFY_IS_APPROX(t4.matrix(), Matrix4::Identity()); - t4 *= aa3; - VERIFY_IS_APPROX(t3.matrix(), t4.matrix()); - - v3 = Vector3::Random(); - Translation3 tv3(v3); - Transform3 t5(tv3); - t4 = tv3; - VERIFY_IS_APPROX(t5.matrix(), t4.matrix()); - t4.translate(-v3); - VERIFY_IS_APPROX(t4.matrix(), Matrix4::Identity()); - t4 *= tv3; - VERIFY_IS_APPROX(t5.matrix(), t4.matrix()); - - Scaling3 sv3(v3); - Transform3 t6(sv3); - t4 = sv3; - VERIFY_IS_APPROX(t6.matrix(), t4.matrix()); - t4.scale(v3.cwise().inverse()); - VERIFY_IS_APPROX(t4.matrix(), Matrix4::Identity()); - t4 *= sv3; - VERIFY_IS_APPROX(t6.matrix(), t4.matrix()); - - // matrix * transform - VERIFY_IS_APPROX(Transform3(t3.matrix()*t4).matrix(), Transform3(t3*t4).matrix()); - - // chained Transform product - VERIFY_IS_APPROX(((t3*t4)*t5).matrix(), (t3*(t4*t5)).matrix()); - - // check that Transform product doesn't have aliasing problems - t5 = t4; - t5 = t5*t5; - VERIFY_IS_APPROX(t5, t4*t4); - - // 2D transformation - Transform2 t20, t21; - Vector2 v20 = Vector2::Random(); - Vector2 v21 = Vector2::Random(); - for (int k=0; k<2; ++k) - if (ei_abs(v21[k])(a).toRotationMatrix(); - VERIFY_IS_APPROX(t20.fromPositionOrientationScale(v20,a,v21).matrix(), - t21.pretranslate(v20).scale(v21).matrix()); - - t21.setIdentity(); - t21.linear() = Rotation2D(-a).toRotationMatrix(); - VERIFY( (t20.fromPositionOrientationScale(v20,a,v21) - * (t21.prescale(v21.cwise().inverse()).translate(-v20))).matrix().isIdentity(test_precision()) ); - - // Transform - new API - // 3D - t0.setIdentity(); - t0.rotate(q1).scale(v0).translate(v0); - // mat * scaling and mat * translation - t1 = (Matrix3(q1) * Scaling3(v0)) * Translation3(v0); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // mat * transformation and scaling * translation - t1 = Matrix3(q1) * (Scaling3(v0) * Translation3(v0)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - t0.setIdentity(); - t0.prerotate(q1).prescale(v0).pretranslate(v0); - // translation * scaling and transformation * mat - t1 = (Translation3(v0) * Scaling3(v0)) * Matrix3(q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // scaling * mat and translation * mat - t1 = Translation3(v0) * (Scaling3(v0) * Matrix3(q1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - t0.setIdentity(); - t0.scale(v0).translate(v0).rotate(q1); - // translation * mat and scaling * transformation - t1 = Scaling3(v0) * (Translation3(v0) * Matrix3(q1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // transformation * scaling - t0.scale(v0); - t1 = t1 * Scaling3(v0); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // transformation * translation - t0.translate(v0); - t1 = t1 * Translation3(v0); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // translation * transformation - t0.pretranslate(v0); - t1 = Translation3(v0) * t1; - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // transform * quaternion - t0.rotate(q1); - t1 = t1 * q1; - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // translation * quaternion - t0.translate(v1).rotate(q1); - t1 = t1 * (Translation3(v1) * q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // scaling * quaternion - t0.scale(v1).rotate(q1); - t1 = t1 * (Scaling3(v1) * q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // quaternion * transform - t0.prerotate(q1); - t1 = q1 * t1; - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // quaternion * translation - t0.rotate(q1).translate(v1); - t1 = t1 * (q1 * Translation3(v1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // quaternion * scaling - t0.rotate(q1).scale(v1); - t1 = t1 * (q1 * Scaling3(v1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // translation * vector - t0.setIdentity(); - t0.translate(v0); - VERIFY_IS_APPROX(t0 * v1, Translation3(v0) * v1); - - // scaling * vector - t0.setIdentity(); - t0.scale(v0); - VERIFY_IS_APPROX(t0 * v1, Scaling3(v0) * v1); - - // test transform inversion - t0.setIdentity(); - t0.translate(v0); - t0.linear().setRandom(); - VERIFY_IS_APPROX(t0.inverse(Affine), t0.matrix().inverse()); - t0.setIdentity(); - t0.translate(v0).rotate(q1); - VERIFY_IS_APPROX(t0.inverse(Isometry), t0.matrix().inverse()); - - // test extract rotation and scaling - t0.setIdentity(); - t0.translate(v0).rotate(q1).scale(v1); - VERIFY_IS_APPROX(t0.rotation() * v1, Matrix3(q1) * v1); - - Matrix3 mat_rotation, mat_scaling; - t0.setIdentity(); - t0.translate(v0).rotate(q1).scale(v1); - t0.computeRotationScaling(&mat_rotation, &mat_scaling); - VERIFY_IS_APPROX(t0.linear(), mat_rotation * mat_scaling); - VERIFY_IS_APPROX(mat_rotation*mat_rotation.adjoint(), Matrix3::Identity()); - VERIFY_IS_APPROX(mat_rotation.determinant(), Scalar(1)); - t0.computeScalingRotation(&mat_scaling, &mat_rotation); - VERIFY_IS_APPROX(t0.linear(), mat_scaling * mat_rotation); - VERIFY_IS_APPROX(mat_rotation*mat_rotation.adjoint(), Matrix3::Identity()); - VERIFY_IS_APPROX(mat_rotation.determinant(), Scalar(1)); - - // test casting - Transform t1f = t1.template cast(); - VERIFY_IS_APPROX(t1f.template cast(),t1); - Transform t1d = t1.template cast(); - VERIFY_IS_APPROX(t1d.template cast(),t1); - - Translation3 tr1(v0); - Translation tr1f = tr1.template cast(); - VERIFY_IS_APPROX(tr1f.template cast(),tr1); - Translation tr1d = tr1.template cast(); - VERIFY_IS_APPROX(tr1d.template cast(),tr1); - - Scaling3 sc1(v0); - Scaling sc1f = sc1.template cast(); - VERIFY_IS_APPROX(sc1f.template cast(),sc1); - Scaling sc1d = sc1.template cast(); - VERIFY_IS_APPROX(sc1d.template cast(),sc1); - - Quaternion q1f = q1.template cast(); - VERIFY_IS_APPROX(q1f.template cast(),q1); - Quaternion q1d = q1.template cast(); - VERIFY_IS_APPROX(q1d.template cast(),q1); - - AngleAxis aa1f = aa1.template cast(); - VERIFY_IS_APPROX(aa1f.template cast(),aa1); - AngleAxis aa1d = aa1.template cast(); - VERIFY_IS_APPROX(aa1d.template cast(),aa1); - - Rotation2D r2d1(ei_random()); - Rotation2D r2d1f = r2d1.template cast(); - VERIFY_IS_APPROX(r2d1f.template cast(),r2d1); - Rotation2D r2d1d = r2d1.template cast(); - VERIFY_IS_APPROX(r2d1d.template cast(),r2d1); - - m = q1; -// m.col(1) = Vector3(0,ei_random(),ei_random()).normalized(); -// m.col(0) = Vector3(-1,0,0).normalized(); -// m.col(2) = m.col(0).cross(m.col(1)); - #define VERIFY_EULER(I,J,K, X,Y,Z) { \ - Vector3 ea = m.eulerAngles(I,J,K); \ - Matrix3 m1 = Matrix3(AngleAxisx(ea[0], Vector3::Unit##X()) * AngleAxisx(ea[1], Vector3::Unit##Y()) * AngleAxisx(ea[2], Vector3::Unit##Z())); \ - VERIFY_IS_APPROX(m, Matrix3(AngleAxisx(ea[0], Vector3::Unit##X()) * AngleAxisx(ea[1], Vector3::Unit##Y()) * AngleAxisx(ea[2], Vector3::Unit##Z()))); \ - } - VERIFY_EULER(0,1,2, X,Y,Z); - VERIFY_EULER(0,1,0, X,Y,X); - VERIFY_EULER(0,2,1, X,Z,Y); - VERIFY_EULER(0,2,0, X,Z,X); - - VERIFY_EULER(1,2,0, Y,Z,X); - VERIFY_EULER(1,2,1, Y,Z,Y); - VERIFY_EULER(1,0,2, Y,X,Z); - VERIFY_EULER(1,0,1, Y,X,Y); - - VERIFY_EULER(2,0,1, Z,X,Y); - VERIFY_EULER(2,0,2, Z,X,Z); - VERIFY_EULER(2,1,0, Z,Y,X); - VERIFY_EULER(2,1,2, Z,Y,Z); - - // colwise/rowwise cross product - mat3.setRandom(); - Vector3 vec3 = Vector3::Random(); - Matrix3 mcross; - int i = ei_random(0,2); - mcross = mat3.colwise().cross(vec3); - VERIFY_IS_APPROX(mcross.col(i), mat3.col(i).cross(vec3)); - mcross = mat3.rowwise().cross(vec3); - VERIFY_IS_APPROX(mcross.row(i), mat3.row(i).cross(vec3)); - - -} - -void test_eigen2_geometry() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( geometry() ); - CALL_SUBTEST_2( geometry() ); - } -} diff --git a/test/eigen2/eigen2_geometry_with_eigen2_prefix.cpp b/test/eigen2/eigen2_geometry_with_eigen2_prefix.cpp deleted file mode 100644 index f83b57249..000000000 --- a/test/eigen2/eigen2_geometry_with_eigen2_prefix.cpp +++ /dev/null @@ -1,434 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#define EIGEN2_SUPPORT_STAGE15_RESOLVE_API_CONFLICTS_WARN - -#include "main.h" -#include -#include -#include - -template void geometry(void) -{ - /* this test covers the following files: - Cross.h Quaternion.h, Transform.cpp - */ - - typedef Matrix Matrix2; - typedef Matrix Matrix3; - typedef Matrix Matrix4; - typedef Matrix Vector2; - typedef Matrix Vector3; - typedef Matrix Vector4; - typedef eigen2_Quaternion Quaternionx; - typedef eigen2_AngleAxis AngleAxisx; - typedef eigen2_Transform Transform2; - typedef eigen2_Transform Transform3; - typedef eigen2_Scaling Scaling2; - typedef eigen2_Scaling Scaling3; - typedef eigen2_Translation Translation2; - typedef eigen2_Translation Translation3; - - Scalar largeEps = test_precision(); - if (ei_is_same_type::ret) - largeEps = 1e-2f; - - Vector3 v0 = Vector3::Random(), - v1 = Vector3::Random(), - v2 = Vector3::Random(); - Vector2 u0 = Vector2::Random(); - Matrix3 matrot1; - - Scalar a = ei_random(-Scalar(M_PI), Scalar(M_PI)); - - // cross product - VERIFY_IS_MUCH_SMALLER_THAN(v1.cross(v2).eigen2_dot(v1), Scalar(1)); - Matrix3 m; - m << v0.normalized(), - (v0.cross(v1)).normalized(), - (v0.cross(v1).cross(v0)).normalized(); - VERIFY(m.isUnitary()); - - // Quaternion: Identity(), setIdentity(); - Quaternionx q1, q2; - q2.setIdentity(); - VERIFY_IS_APPROX(Quaternionx(Quaternionx::Identity()).coeffs(), q2.coeffs()); - q1.coeffs().setRandom(); - VERIFY_IS_APPROX(q1.coeffs(), (q1*q2).coeffs()); - - // unitOrthogonal - VERIFY_IS_MUCH_SMALLER_THAN(u0.unitOrthogonal().eigen2_dot(u0), Scalar(1)); - VERIFY_IS_MUCH_SMALLER_THAN(v0.unitOrthogonal().eigen2_dot(v0), Scalar(1)); - VERIFY_IS_APPROX(u0.unitOrthogonal().norm(), Scalar(1)); - VERIFY_IS_APPROX(v0.unitOrthogonal().norm(), Scalar(1)); - - - VERIFY_IS_APPROX(v0, AngleAxisx(a, v0.normalized()) * v0); - VERIFY_IS_APPROX(-v0, AngleAxisx(Scalar(M_PI), v0.unitOrthogonal()) * v0); - VERIFY_IS_APPROX(ei_cos(a)*v0.squaredNorm(), v0.eigen2_dot(AngleAxisx(a, v0.unitOrthogonal()) * v0)); - m = AngleAxisx(a, v0.normalized()).toRotationMatrix().adjoint(); - VERIFY_IS_APPROX(Matrix3::Identity(), m * AngleAxisx(a, v0.normalized())); - VERIFY_IS_APPROX(Matrix3::Identity(), AngleAxisx(a, v0.normalized()) * m); - - q1 = AngleAxisx(a, v0.normalized()); - q2 = AngleAxisx(a, v1.normalized()); - - // angular distance - Scalar refangle = ei_abs(AngleAxisx(q1.inverse()*q2).angle()); - if (refangle>Scalar(M_PI)) - refangle = Scalar(2)*Scalar(M_PI) - refangle; - - if((q1.coeffs()-q2.coeffs()).norm() > 10*largeEps) - { - VERIFY(ei_isApprox(q1.angularDistance(q2), refangle, largeEps)); - } - - // rotation matrix conversion - VERIFY_IS_APPROX(q1 * v2, q1.toRotationMatrix() * v2); - VERIFY_IS_APPROX(q1 * q2 * v2, - q1.toRotationMatrix() * q2.toRotationMatrix() * v2); - - VERIFY( (q2*q1).isApprox(q1*q2, largeEps) || !(q2 * q1 * v2).isApprox( - q1.toRotationMatrix() * q2.toRotationMatrix() * v2)); - - q2 = q1.toRotationMatrix(); - VERIFY_IS_APPROX(q1*v1,q2*v1); - - matrot1 = AngleAxisx(Scalar(0.1), Vector3::UnitX()) - * AngleAxisx(Scalar(0.2), Vector3::UnitY()) - * AngleAxisx(Scalar(0.3), Vector3::UnitZ()); - VERIFY_IS_APPROX(matrot1 * v1, - AngleAxisx(Scalar(0.1), Vector3(1,0,0)).toRotationMatrix() - * (AngleAxisx(Scalar(0.2), Vector3(0,1,0)).toRotationMatrix() - * (AngleAxisx(Scalar(0.3), Vector3(0,0,1)).toRotationMatrix() * v1))); - - // angle-axis conversion - AngleAxisx aa = q1; - VERIFY_IS_APPROX(q1 * v1, Quaternionx(aa) * v1); - VERIFY_IS_NOT_APPROX(q1 * v1, Quaternionx(AngleAxisx(aa.angle()*2,aa.axis())) * v1); - - // from two vector creation - VERIFY_IS_APPROX(v2.normalized(),(q2.setFromTwoVectors(v1,v2)*v1).normalized()); - VERIFY_IS_APPROX(v2.normalized(),(q2.setFromTwoVectors(v1,v2)*v1).normalized()); - - // inverse and conjugate - VERIFY_IS_APPROX(q1 * (q1.inverse() * v1), v1); - VERIFY_IS_APPROX(q1 * (q1.conjugate() * v1), v1); - - // AngleAxis - VERIFY_IS_APPROX(AngleAxisx(a,v1.normalized()).toRotationMatrix(), - Quaternionx(AngleAxisx(a,v1.normalized())).toRotationMatrix()); - - AngleAxisx aa1; - m = q1.toRotationMatrix(); - aa1 = m; - VERIFY_IS_APPROX(AngleAxisx(m).toRotationMatrix(), - Quaternionx(m).toRotationMatrix()); - - // Transform - // TODO complete the tests ! - a = 0; - while (ei_abs(a)(-Scalar(0.4)*Scalar(M_PI), Scalar(0.4)*Scalar(M_PI)); - q1 = AngleAxisx(a, v0.normalized()); - Transform3 t0, t1, t2; - // first test setIdentity() and Identity() - t0.setIdentity(); - VERIFY_IS_APPROX(t0.matrix(), Transform3::MatrixType::Identity()); - t0.matrix().setZero(); - t0 = Transform3::Identity(); - VERIFY_IS_APPROX(t0.matrix(), Transform3::MatrixType::Identity()); - - t0.linear() = q1.toRotationMatrix(); - t1.setIdentity(); - t1.linear() = q1.toRotationMatrix(); - - v0 << 50, 2, 1;//= ei_random_matrix().cwiseProduct(Vector3(10,2,0.5)); - t0.scale(v0); - t1.prescale(v0); - - VERIFY_IS_APPROX( (t0 * Vector3(1,0,0)).norm(), v0.x()); - //VERIFY(!ei_isApprox((t1 * Vector3(1,0,0)).norm(), v0.x())); - - t0.setIdentity(); - t1.setIdentity(); - v1 << 1, 2, 3; - t0.linear() = q1.toRotationMatrix(); - t0.pretranslate(v0); - t0.scale(v1); - t1.linear() = q1.conjugate().toRotationMatrix(); - t1.prescale(v1.cwise().inverse()); - t1.translate(-v0); - - VERIFY((t0.matrix() * t1.matrix()).isIdentity(test_precision())); - - t1.fromPositionOrientationScale(v0, q1, v1); - VERIFY_IS_APPROX(t1.matrix(), t0.matrix()); - VERIFY_IS_APPROX(t1*v1, t0*v1); - - t0.setIdentity(); t0.scale(v0).rotate(q1.toRotationMatrix()); - t1.setIdentity(); t1.scale(v0).rotate(q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - t0.setIdentity(); t0.scale(v0).rotate(AngleAxisx(q1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - VERIFY_IS_APPROX(t0.scale(a).matrix(), t1.scale(Vector3::Constant(a)).matrix()); - VERIFY_IS_APPROX(t0.prescale(a).matrix(), t1.prescale(Vector3::Constant(a)).matrix()); - - // More transform constructors, operator=, operator*= - - Matrix3 mat3 = Matrix3::Random(); - Matrix4 mat4; - mat4 << mat3 , Vector3::Zero() , Vector4::Zero().transpose(); - Transform3 tmat3(mat3), tmat4(mat4); - tmat4.matrix()(3,3) = Scalar(1); - VERIFY_IS_APPROX(tmat3.matrix(), tmat4.matrix()); - - Scalar a3 = ei_random(-Scalar(M_PI), Scalar(M_PI)); - Vector3 v3 = Vector3::Random().normalized(); - AngleAxisx aa3(a3, v3); - Transform3 t3(aa3); - Transform3 t4; - t4 = aa3; - VERIFY_IS_APPROX(t3.matrix(), t4.matrix()); - t4.rotate(AngleAxisx(-a3,v3)); - VERIFY_IS_APPROX(t4.matrix(), Matrix4::Identity()); - t4 *= aa3; - VERIFY_IS_APPROX(t3.matrix(), t4.matrix()); - - v3 = Vector3::Random(); - Translation3 tv3(v3); - Transform3 t5(tv3); - t4 = tv3; - VERIFY_IS_APPROX(t5.matrix(), t4.matrix()); - t4.translate(-v3); - VERIFY_IS_APPROX(t4.matrix(), Matrix4::Identity()); - t4 *= tv3; - VERIFY_IS_APPROX(t5.matrix(), t4.matrix()); - - Scaling3 sv3(v3); - Transform3 t6(sv3); - t4 = sv3; - VERIFY_IS_APPROX(t6.matrix(), t4.matrix()); - t4.scale(v3.cwise().inverse()); - VERIFY_IS_APPROX(t4.matrix(), Matrix4::Identity()); - t4 *= sv3; - VERIFY_IS_APPROX(t6.matrix(), t4.matrix()); - - // matrix * transform - VERIFY_IS_APPROX(Transform3(t3.matrix()*t4).matrix(), Transform3(t3*t4).matrix()); - - // chained Transform product - VERIFY_IS_APPROX(((t3*t4)*t5).matrix(), (t3*(t4*t5)).matrix()); - - // check that Transform product doesn't have aliasing problems - t5 = t4; - t5 = t5*t5; - VERIFY_IS_APPROX(t5, t4*t4); - - // 2D transformation - Transform2 t20, t21; - Vector2 v20 = Vector2::Random(); - Vector2 v21 = Vector2::Random(); - for (int k=0; k<2; ++k) - if (ei_abs(v21[k])(a).toRotationMatrix(); - VERIFY_IS_APPROX(t20.fromPositionOrientationScale(v20,a,v21).matrix(), - t21.pretranslate(v20).scale(v21).matrix()); - - t21.setIdentity(); - t21.linear() = Rotation2D(-a).toRotationMatrix(); - VERIFY( (t20.fromPositionOrientationScale(v20,a,v21) - * (t21.prescale(v21.cwise().inverse()).translate(-v20))).matrix().isIdentity(test_precision()) ); - - // Transform - new API - // 3D - t0.setIdentity(); - t0.rotate(q1).scale(v0).translate(v0); - // mat * scaling and mat * translation - t1 = (Matrix3(q1) * Scaling3(v0)) * Translation3(v0); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // mat * transformation and scaling * translation - t1 = Matrix3(q1) * (Scaling3(v0) * Translation3(v0)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - t0.setIdentity(); - t0.prerotate(q1).prescale(v0).pretranslate(v0); - // translation * scaling and transformation * mat - t1 = (Translation3(v0) * Scaling3(v0)) * Matrix3(q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // scaling * mat and translation * mat - t1 = Translation3(v0) * (Scaling3(v0) * Matrix3(q1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - t0.setIdentity(); - t0.scale(v0).translate(v0).rotate(q1); - // translation * mat and scaling * transformation - t1 = Scaling3(v0) * (Translation3(v0) * Matrix3(q1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // transformation * scaling - t0.scale(v0); - t1 = t1 * Scaling3(v0); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // transformation * translation - t0.translate(v0); - t1 = t1 * Translation3(v0); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - // translation * transformation - t0.pretranslate(v0); - t1 = Translation3(v0) * t1; - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // transform * quaternion - t0.rotate(q1); - t1 = t1 * q1; - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // translation * quaternion - t0.translate(v1).rotate(q1); - t1 = t1 * (Translation3(v1) * q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // scaling * quaternion - t0.scale(v1).rotate(q1); - t1 = t1 * (Scaling3(v1) * q1); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // quaternion * transform - t0.prerotate(q1); - t1 = q1 * t1; - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // quaternion * translation - t0.rotate(q1).translate(v1); - t1 = t1 * (q1 * Translation3(v1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // quaternion * scaling - t0.rotate(q1).scale(v1); - t1 = t1 * (q1 * Scaling3(v1)); - VERIFY_IS_APPROX(t0.matrix(), t1.matrix()); - - // translation * vector - t0.setIdentity(); - t0.translate(v0); - VERIFY_IS_APPROX(t0 * v1, Translation3(v0) * v1); - - // scaling * vector - t0.setIdentity(); - t0.scale(v0); - VERIFY_IS_APPROX(t0 * v1, Scaling3(v0) * v1); - - // test transform inversion - t0.setIdentity(); - t0.translate(v0); - t0.linear().setRandom(); - VERIFY_IS_APPROX(t0.inverse(Affine), t0.matrix().inverse()); - t0.setIdentity(); - t0.translate(v0).rotate(q1); - VERIFY_IS_APPROX(t0.inverse(Isometry), t0.matrix().inverse()); - - // test extract rotation and scaling - t0.setIdentity(); - t0.translate(v0).rotate(q1).scale(v1); - VERIFY_IS_APPROX(t0.rotation() * v1, Matrix3(q1) * v1); - - Matrix3 mat_rotation, mat_scaling; - t0.setIdentity(); - t0.translate(v0).rotate(q1).scale(v1); - t0.computeRotationScaling(&mat_rotation, &mat_scaling); - VERIFY_IS_APPROX(t0.linear(), mat_rotation * mat_scaling); - VERIFY_IS_APPROX(mat_rotation*mat_rotation.adjoint(), Matrix3::Identity()); - VERIFY_IS_APPROX(mat_rotation.determinant(), Scalar(1)); - t0.computeScalingRotation(&mat_scaling, &mat_rotation); - VERIFY_IS_APPROX(t0.linear(), mat_scaling * mat_rotation); - VERIFY_IS_APPROX(mat_rotation*mat_rotation.adjoint(), Matrix3::Identity()); - VERIFY_IS_APPROX(mat_rotation.determinant(), Scalar(1)); - - // test casting - eigen2_Transform t1f = t1.template cast(); - VERIFY_IS_APPROX(t1f.template cast(),t1); - eigen2_Transform t1d = t1.template cast(); - VERIFY_IS_APPROX(t1d.template cast(),t1); - - Translation3 tr1(v0); - eigen2_Translation tr1f = tr1.template cast(); - VERIFY_IS_APPROX(tr1f.template cast(),tr1); - eigen2_Translation tr1d = tr1.template cast(); - VERIFY_IS_APPROX(tr1d.template cast(),tr1); - - Scaling3 sc1(v0); - eigen2_Scaling sc1f = sc1.template cast(); - VERIFY_IS_APPROX(sc1f.template cast(),sc1); - eigen2_Scaling sc1d = sc1.template cast(); - VERIFY_IS_APPROX(sc1d.template cast(),sc1); - - eigen2_Quaternion q1f = q1.template cast(); - VERIFY_IS_APPROX(q1f.template cast(),q1); - eigen2_Quaternion q1d = q1.template cast(); - VERIFY_IS_APPROX(q1d.template cast(),q1); - - eigen2_AngleAxis aa1f = aa1.template cast(); - VERIFY_IS_APPROX(aa1f.template cast(),aa1); - eigen2_AngleAxis aa1d = aa1.template cast(); - VERIFY_IS_APPROX(aa1d.template cast(),aa1); - - eigen2_Rotation2D r2d1(ei_random()); - eigen2_Rotation2D r2d1f = r2d1.template cast(); - VERIFY_IS_APPROX(r2d1f.template cast(),r2d1); - eigen2_Rotation2D r2d1d = r2d1.template cast(); - VERIFY_IS_APPROX(r2d1d.template cast(),r2d1); - - m = q1; -// m.col(1) = Vector3(0,ei_random(),ei_random()).normalized(); -// m.col(0) = Vector3(-1,0,0).normalized(); -// m.col(2) = m.col(0).cross(m.col(1)); - #define VERIFY_EULER(I,J,K, X,Y,Z) { \ - Vector3 ea = m.eulerAngles(I,J,K); \ - Matrix3 m1 = Matrix3(AngleAxisx(ea[0], Vector3::Unit##X()) * AngleAxisx(ea[1], Vector3::Unit##Y()) * AngleAxisx(ea[2], Vector3::Unit##Z())); \ - VERIFY_IS_APPROX(m, Matrix3(AngleAxisx(ea[0], Vector3::Unit##X()) * AngleAxisx(ea[1], Vector3::Unit##Y()) * AngleAxisx(ea[2], Vector3::Unit##Z()))); \ - } - VERIFY_EULER(0,1,2, X,Y,Z); - VERIFY_EULER(0,1,0, X,Y,X); - VERIFY_EULER(0,2,1, X,Z,Y); - VERIFY_EULER(0,2,0, X,Z,X); - - VERIFY_EULER(1,2,0, Y,Z,X); - VERIFY_EULER(1,2,1, Y,Z,Y); - VERIFY_EULER(1,0,2, Y,X,Z); - VERIFY_EULER(1,0,1, Y,X,Y); - - VERIFY_EULER(2,0,1, Z,X,Y); - VERIFY_EULER(2,0,2, Z,X,Z); - VERIFY_EULER(2,1,0, Z,Y,X); - VERIFY_EULER(2,1,2, Z,Y,Z); - - // colwise/rowwise cross product - mat3.setRandom(); - Vector3 vec3 = Vector3::Random(); - Matrix3 mcross; - int i = ei_random(0,2); - mcross = mat3.colwise().cross(vec3); - VERIFY_IS_APPROX(mcross.col(i), mat3.col(i).cross(vec3)); - mcross = mat3.rowwise().cross(vec3); - VERIFY_IS_APPROX(mcross.row(i), mat3.row(i).cross(vec3)); - - -} - -void test_eigen2_geometry_with_eigen2_prefix() -{ - std::cout << "eigen2 support: " << EIGEN2_SUPPORT_STAGE << std::endl; - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( geometry() ); - CALL_SUBTEST_2( geometry() ); - } -} diff --git a/test/eigen2/eigen2_hyperplane.cpp b/test/eigen2/eigen2_hyperplane.cpp deleted file mode 100644 index f3f85e14d..000000000 --- a/test/eigen2/eigen2_hyperplane.cpp +++ /dev/null @@ -1,126 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include -#include -#include - -template void hyperplane(const HyperplaneType& _plane) -{ - /* this test covers the following files: - Hyperplane.h - */ - - const int dim = _plane.dim(); - typedef typename HyperplaneType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - typedef Matrix MatrixType; - - VectorType p0 = VectorType::Random(dim); - VectorType p1 = VectorType::Random(dim); - - VectorType n0 = VectorType::Random(dim).normalized(); - VectorType n1 = VectorType::Random(dim).normalized(); - - HyperplaneType pl0(n0, p0); - HyperplaneType pl1(n1, p1); - HyperplaneType pl2 = pl1; - - Scalar s0 = ei_random(); - Scalar s1 = ei_random(); - - VERIFY_IS_APPROX( n1.eigen2_dot(n1), Scalar(1) ); - - VERIFY_IS_MUCH_SMALLER_THAN( pl0.absDistance(p0), Scalar(1) ); - VERIFY_IS_APPROX( pl1.signedDistance(p1 + n1 * s0), s0 ); - VERIFY_IS_MUCH_SMALLER_THAN( pl1.signedDistance(pl1.projection(p0)), Scalar(1) ); - VERIFY_IS_MUCH_SMALLER_THAN( pl1.absDistance(p1 + pl1.normal().unitOrthogonal() * s1), Scalar(1) ); - - // transform - if (!NumTraits::IsComplex) - { - MatrixType rot = MatrixType::Random(dim,dim).qr().matrixQ(); - Scaling scaling(VectorType::Random()); - Translation translation(VectorType::Random()); - - pl2 = pl1; - VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot).absDistance(rot * p1), Scalar(1) ); - pl2 = pl1; - VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot,Isometry).absDistance(rot * p1), Scalar(1) ); - pl2 = pl1; - VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot*scaling).absDistance((rot*scaling) * p1), Scalar(1) ); - pl2 = pl1; - VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot*scaling*translation) - .absDistance((rot*scaling*translation) * p1), Scalar(1) ); - pl2 = pl1; - VERIFY_IS_MUCH_SMALLER_THAN( pl2.transform(rot*translation,Isometry) - .absDistance((rot*translation) * p1), Scalar(1) ); - } - - // casting - const int Dim = HyperplaneType::AmbientDimAtCompileTime; - typedef typename GetDifferentType::type OtherScalar; - Hyperplane hp1f = pl1.template cast(); - VERIFY_IS_APPROX(hp1f.template cast(),pl1); - Hyperplane hp1d = pl1.template cast(); - VERIFY_IS_APPROX(hp1d.template cast(),pl1); -} - -template void lines() -{ - typedef Hyperplane HLine; - typedef ParametrizedLine PLine; - typedef Matrix Vector; - typedef Matrix CoeffsType; - - for(int i = 0; i < 10; i++) - { - Vector center = Vector::Random(); - Vector u = Vector::Random(); - Vector v = Vector::Random(); - Scalar a = ei_random(); - while (ei_abs(a-1) < 1e-4) a = ei_random(); - while (u.norm() < 1e-4) u = Vector::Random(); - while (v.norm() < 1e-4) v = Vector::Random(); - - HLine line_u = HLine::Through(center + u, center + a*u); - HLine line_v = HLine::Through(center + v, center + a*v); - - // the line equations should be normalized so that a^2+b^2=1 - VERIFY_IS_APPROX(line_u.normal().norm(), Scalar(1)); - VERIFY_IS_APPROX(line_v.normal().norm(), Scalar(1)); - - Vector result = line_u.intersection(line_v); - - // the lines should intersect at the point we called "center" - VERIFY_IS_APPROX(result, center); - - // check conversions between two types of lines - PLine pl(line_u); // gcc 3.3 will commit suicide if we don't name this variable - CoeffsType converted_coeffs(HLine(pl).coeffs()); - converted_coeffs *= line_u.coeffs()(0)/converted_coeffs(0); - VERIFY(line_u.coeffs().isApprox(converted_coeffs)); - } -} - -void test_eigen2_hyperplane() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( hyperplane(Hyperplane()) ); - CALL_SUBTEST_2( hyperplane(Hyperplane()) ); - CALL_SUBTEST_3( hyperplane(Hyperplane()) ); - CALL_SUBTEST_4( hyperplane(Hyperplane,5>()) ); - CALL_SUBTEST_5( lines() ); - CALL_SUBTEST_6( lines() ); - } -} diff --git a/test/eigen2/eigen2_inverse.cpp b/test/eigen2/eigen2_inverse.cpp deleted file mode 100644 index 5de1dfefa..000000000 --- a/test/eigen2/eigen2_inverse.cpp +++ /dev/null @@ -1,63 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -template void inverse(const MatrixType& m) -{ - /* this test covers the following files: - Inverse.h - */ - int rows = m.rows(); - int cols = m.cols(); - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - MatrixType m1 = MatrixType::Random(rows, cols), - m2(rows, cols), - mzero = MatrixType::Zero(rows, cols), - identity = MatrixType::Identity(rows, rows); - - while(ei_abs(m1.determinant()) < RealScalar(0.1) && rows <= 8) - { - m1 = MatrixType::Random(rows, cols); - } - - m2 = m1.inverse(); - VERIFY_IS_APPROX(m1, m2.inverse() ); - - m1.computeInverse(&m2); - VERIFY_IS_APPROX(m1, m2.inverse() ); - - VERIFY_IS_APPROX((Scalar(2)*m2).inverse(), m2.inverse()*Scalar(0.5)); - - VERIFY_IS_APPROX(identity, m1.inverse() * m1 ); - VERIFY_IS_APPROX(identity, m1 * m1.inverse() ); - - VERIFY_IS_APPROX(m1, m1.inverse().inverse() ); - - // since for the general case we implement separately row-major and col-major, test that - VERIFY_IS_APPROX(m1.transpose().inverse(), m1.inverse().transpose()); -} - -void test_eigen2_inverse() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( inverse(Matrix()) ); - CALL_SUBTEST_2( inverse(Matrix2d()) ); - CALL_SUBTEST_3( inverse(Matrix3f()) ); - CALL_SUBTEST_4( inverse(Matrix4f()) ); - CALL_SUBTEST_5( inverse(MatrixXf(8,8)) ); - CALL_SUBTEST_6( inverse(MatrixXcd(7,7)) ); - } -} diff --git a/test/eigen2/eigen2_linearstructure.cpp b/test/eigen2/eigen2_linearstructure.cpp deleted file mode 100644 index 22f02c396..000000000 --- a/test/eigen2/eigen2_linearstructure.cpp +++ /dev/null @@ -1,84 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void linearStructure(const MatrixType& m) -{ - /* this test covers the following files: - Sum.h Difference.h Opposite.h ScalarMultiple.h - */ - - typedef typename MatrixType::Scalar Scalar; - typedef Matrix VectorType; - - int rows = m.rows(); - int cols = m.cols(); - - // this test relies a lot on Random.h, and there's not much more that we can do - // to test it, hence I consider that we will have tested Random.h - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - mzero = MatrixType::Zero(rows, cols); - - Scalar s1 = ei_random(); - while (ei_abs(s1)<1e-3) s1 = ei_random(); - - int r = ei_random(0, rows-1), - c = ei_random(0, cols-1); - - VERIFY_IS_APPROX(-(-m1), m1); - VERIFY_IS_APPROX(m1+m1, 2*m1); - VERIFY_IS_APPROX(m1+m2-m1, m2); - VERIFY_IS_APPROX(-m2+m1+m2, m1); - VERIFY_IS_APPROX(m1*s1, s1*m1); - VERIFY_IS_APPROX((m1+m2)*s1, s1*m1+s1*m2); - VERIFY_IS_APPROX((-m1+m2)*s1, -s1*m1+s1*m2); - m3 = m2; m3 += m1; - VERIFY_IS_APPROX(m3, m1+m2); - m3 = m2; m3 -= m1; - VERIFY_IS_APPROX(m3, m2-m1); - m3 = m2; m3 *= s1; - VERIFY_IS_APPROX(m3, s1*m2); - if(NumTraits::HasFloatingPoint) - { - m3 = m2; m3 /= s1; - VERIFY_IS_APPROX(m3, m2/s1); - } - - // again, test operator() to check const-qualification - VERIFY_IS_APPROX((-m1)(r,c), -(m1(r,c))); - VERIFY_IS_APPROX((m1-m2)(r,c), (m1(r,c))-(m2(r,c))); - VERIFY_IS_APPROX((m1+m2)(r,c), (m1(r,c))+(m2(r,c))); - VERIFY_IS_APPROX((s1*m1)(r,c), s1*(m1(r,c))); - VERIFY_IS_APPROX((m1*s1)(r,c), (m1(r,c))*s1); - if(NumTraits::HasFloatingPoint) - VERIFY_IS_APPROX((m1/s1)(r,c), (m1(r,c))/s1); - - // use .block to disable vectorization and compare to the vectorized version - VERIFY_IS_APPROX(m1+m1.block(0,0,rows,cols), m1+m1); - VERIFY_IS_APPROX(m1.cwise() * m1.block(0,0,rows,cols), m1.cwise() * m1); - VERIFY_IS_APPROX(m1 - m1.block(0,0,rows,cols), m1 - m1); - VERIFY_IS_APPROX(m1.block(0,0,rows,cols) * s1, m1 * s1); -} - -void test_eigen2_linearstructure() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( linearStructure(Matrix()) ); - CALL_SUBTEST_2( linearStructure(Matrix2f()) ); - CALL_SUBTEST_3( linearStructure(Vector3d()) ); - CALL_SUBTEST_4( linearStructure(Matrix4d()) ); - CALL_SUBTEST_5( linearStructure(MatrixXcf(3, 3)) ); - CALL_SUBTEST_6( linearStructure(MatrixXf(8, 12)) ); - CALL_SUBTEST_7( linearStructure(MatrixXi(8, 12)) ); - CALL_SUBTEST_8( linearStructure(MatrixXcd(20, 20)) ); - } -} diff --git a/test/eigen2/eigen2_lu.cpp b/test/eigen2/eigen2_lu.cpp deleted file mode 100644 index e993b1c7c..000000000 --- a/test/eigen2/eigen2_lu.cpp +++ /dev/null @@ -1,122 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -template -void doSomeRankPreservingOperations(Eigen::MatrixBase& m) -{ - typedef typename Derived::RealScalar RealScalar; - for(int a = 0; a < 3*(m.rows()+m.cols()); a++) - { - RealScalar d = Eigen::ei_random(-1,1); - int i = Eigen::ei_random(0,m.rows()-1); // i is a random row number - int j; - do { - j = Eigen::ei_random(0,m.rows()-1); - } while (i==j); // j is another one (must be different) - m.row(i) += d * m.row(j); - - i = Eigen::ei_random(0,m.cols()-1); // i is a random column number - do { - j = Eigen::ei_random(0,m.cols()-1); - } while (i==j); // j is another one (must be different) - m.col(i) += d * m.col(j); - } -} - -template void lu_non_invertible() -{ - /* this test covers the following files: - LU.h - */ - // NOTE there seems to be a problem with too small sizes -- could easily lie in the doSomeRankPreservingOperations function - int rows = ei_random(20,200), cols = ei_random(20,200), cols2 = ei_random(20,200); - int rank = ei_random(1, std::min(rows, cols)-1); - - MatrixType m1(rows, cols), m2(cols, cols2), m3(rows, cols2), k(1,1); - m1 = MatrixType::Random(rows,cols); - if(rows <= cols) - for(int i = rank; i < rows; i++) m1.row(i).setZero(); - else - for(int i = rank; i < cols; i++) m1.col(i).setZero(); - doSomeRankPreservingOperations(m1); - - LU lu(m1); - typename LU::KernelResultType m1kernel = lu.kernel(); - typename LU::ImageResultType m1image = lu.image(); - - VERIFY(rank == lu.rank()); - VERIFY(cols - lu.rank() == lu.dimensionOfKernel()); - VERIFY(!lu.isInjective()); - VERIFY(!lu.isInvertible()); - VERIFY(lu.isSurjective() == (lu.rank() == rows)); - VERIFY((m1 * m1kernel).isMuchSmallerThan(m1)); - VERIFY(m1image.lu().rank() == rank); - MatrixType sidebyside(m1.rows(), m1.cols() + m1image.cols()); - sidebyside << m1, m1image; - VERIFY(sidebyside.lu().rank() == rank); - m2 = MatrixType::Random(cols,cols2); - m3 = m1*m2; - m2 = MatrixType::Random(cols,cols2); - lu.solve(m3, &m2); - VERIFY_IS_APPROX(m3, m1*m2); - /* solve now always returns true - m3 = MatrixType::Random(rows,cols2); - VERIFY(!lu.solve(m3, &m2)); - */ -} - -template void lu_invertible() -{ - /* this test covers the following files: - LU.h - */ - typedef typename NumTraits::Real RealScalar; - int size = ei_random(10,200); - - MatrixType m1(size, size), m2(size, size), m3(size, size); - m1 = MatrixType::Random(size,size); - - if (ei_is_same_type::ret) - { - // let's build a matrix more stable to inverse - MatrixType a = MatrixType::Random(size,size*2); - m1 += a * a.adjoint(); - } - - LU lu(m1); - VERIFY(0 == lu.dimensionOfKernel()); - VERIFY(size == lu.rank()); - VERIFY(lu.isInjective()); - VERIFY(lu.isSurjective()); - VERIFY(lu.isInvertible()); - VERIFY(lu.image().lu().isInvertible()); - m3 = MatrixType::Random(size,size); - lu.solve(m3, &m2); - VERIFY_IS_APPROX(m3, m1*m2); - VERIFY_IS_APPROX(m2, lu.inverse()*m3); - m3 = MatrixType::Random(size,size); - VERIFY(lu.solve(m3, &m2)); -} - -void test_eigen2_lu() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( lu_non_invertible() ); - CALL_SUBTEST_2( lu_non_invertible() ); - CALL_SUBTEST_3( lu_non_invertible() ); - CALL_SUBTEST_4( lu_non_invertible() ); - CALL_SUBTEST_1( lu_invertible() ); - CALL_SUBTEST_2( lu_invertible() ); - CALL_SUBTEST_3( lu_invertible() ); - CALL_SUBTEST_4( lu_invertible() ); - } -} diff --git a/test/eigen2/eigen2_map.cpp b/test/eigen2/eigen2_map.cpp deleted file mode 100644 index 4a1c4e11a..000000000 --- a/test/eigen2/eigen2_map.cpp +++ /dev/null @@ -1,114 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2007-2010 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void map_class_vector(const VectorType& m) -{ - typedef typename VectorType::Scalar Scalar; - - int size = m.size(); - - // test Map.h - Scalar* array1 = ei_aligned_new(size); - Scalar* array2 = ei_aligned_new(size); - Scalar* array3 = new Scalar[size+1]; - Scalar* array3unaligned = std::size_t(array3)%16 == 0 ? array3+1 : array3; - - Map(array1, size) = VectorType::Random(size); - Map(array2, size) = Map(array1, size); - Map(array3unaligned, size) = Map((const Scalar*)array1, size); // test non-const-correctness support in eigen2 - VectorType ma1 = Map(array1, size); - VectorType ma2 = Map(array2, size); - VectorType ma3 = Map(array3unaligned, size); - VERIFY_IS_APPROX(ma1, ma2); - VERIFY_IS_APPROX(ma1, ma3); - - ei_aligned_delete(array1, size); - ei_aligned_delete(array2, size); - delete[] array3; -} - -template void map_class_matrix(const MatrixType& m) -{ - typedef typename MatrixType::Scalar Scalar; - - int rows = m.rows(), cols = m.cols(), size = rows*cols; - - // test Map.h - Scalar* array1 = ei_aligned_new(size); - for(int i = 0; i < size; i++) array1[i] = Scalar(1); - Scalar* array2 = ei_aligned_new(size); - for(int i = 0; i < size; i++) array2[i] = Scalar(1); - Scalar* array3 = new Scalar[size+1]; - for(int i = 0; i < size+1; i++) array3[i] = Scalar(1); - Scalar* array3unaligned = std::size_t(array3)%16 == 0 ? array3+1 : array3; - Map(array1, rows, cols) = MatrixType::Ones(rows,cols); - Map(array2, rows, cols) = Map((const Scalar*)array1, rows, cols); // test non-const-correctness support in eigen2 - Map(array3unaligned, rows, cols) = Map(array1, rows, cols); - MatrixType ma1 = Map(array1, rows, cols); - MatrixType ma2 = Map(array2, rows, cols); - VERIFY_IS_APPROX(ma1, ma2); - MatrixType ma3 = Map(array3unaligned, rows, cols); - VERIFY_IS_APPROX(ma1, ma3); - - ei_aligned_delete(array1, size); - ei_aligned_delete(array2, size); - delete[] array3; -} - -template void map_static_methods(const VectorType& m) -{ - typedef typename VectorType::Scalar Scalar; - - int size = m.size(); - - // test Map.h - Scalar* array1 = ei_aligned_new(size); - Scalar* array2 = ei_aligned_new(size); - Scalar* array3 = new Scalar[size+1]; - Scalar* array3unaligned = std::size_t(array3)%16 == 0 ? array3+1 : array3; - - VectorType::MapAligned(array1, size) = VectorType::Random(size); - VectorType::Map(array2, size) = VectorType::Map(array1, size); - VectorType::Map(array3unaligned, size) = VectorType::Map(array1, size); - VectorType ma1 = VectorType::Map(array1, size); - VectorType ma2 = VectorType::MapAligned(array2, size); - VectorType ma3 = VectorType::Map(array3unaligned, size); - VERIFY_IS_APPROX(ma1, ma2); - VERIFY_IS_APPROX(ma1, ma3); - - ei_aligned_delete(array1, size); - ei_aligned_delete(array2, size); - delete[] array3; -} - - -void test_eigen2_map() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( map_class_vector(Matrix()) ); - CALL_SUBTEST_2( map_class_vector(Vector4d()) ); - CALL_SUBTEST_3( map_class_vector(RowVector4f()) ); - CALL_SUBTEST_4( map_class_vector(VectorXcf(8)) ); - CALL_SUBTEST_5( map_class_vector(VectorXi(12)) ); - - CALL_SUBTEST_1( map_class_matrix(Matrix()) ); - CALL_SUBTEST_2( map_class_matrix(Matrix4d()) ); - CALL_SUBTEST_6( map_class_matrix(Matrix()) ); - CALL_SUBTEST_4( map_class_matrix(MatrixXcf(ei_random(1,10),ei_random(1,10))) ); - CALL_SUBTEST_5( map_class_matrix(MatrixXi(ei_random(1,10),ei_random(1,10))) ); - - CALL_SUBTEST_1( map_static_methods(Matrix()) ); - CALL_SUBTEST_2( map_static_methods(Vector3f()) ); - CALL_SUBTEST_7( map_static_methods(RowVector3d()) ); - CALL_SUBTEST_4( map_static_methods(VectorXcd(8)) ); - CALL_SUBTEST_5( map_static_methods(VectorXf(12)) ); - } -} diff --git a/test/eigen2/eigen2_meta.cpp b/test/eigen2/eigen2_meta.cpp deleted file mode 100644 index 1d01bd84d..000000000 --- a/test/eigen2/eigen2_meta.cpp +++ /dev/null @@ -1,60 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -void test_eigen2_meta() -{ - typedef float & FloatRef; - typedef const float & ConstFloatRef; - - VERIFY((ei_meta_if<(3<4),ei_meta_true, ei_meta_false>::ret::ret)); - VERIFY(( ei_is_same_type::ret)); - VERIFY((!ei_is_same_type::ret)); - VERIFY((!ei_is_same_type::ret)); - VERIFY((!ei_is_same_type::ret)); - - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - VERIFY(( ei_is_same_type::type >::ret)); - - VERIFY(ei_meta_sqrt<1>::ret == 1); - #define VERIFY_META_SQRT(X) VERIFY(ei_meta_sqrt::ret == int(ei_sqrt(double(X)))) - VERIFY_META_SQRT(2); - VERIFY_META_SQRT(3); - VERIFY_META_SQRT(4); - VERIFY_META_SQRT(5); - VERIFY_META_SQRT(6); - VERIFY_META_SQRT(8); - VERIFY_META_SQRT(9); - VERIFY_META_SQRT(15); - VERIFY_META_SQRT(16); - VERIFY_META_SQRT(17); - VERIFY_META_SQRT(255); - VERIFY_META_SQRT(256); - VERIFY_META_SQRT(257); - VERIFY_META_SQRT(1023); - VERIFY_META_SQRT(1024); - VERIFY_META_SQRT(1025); -} diff --git a/test/eigen2/eigen2_miscmatrices.cpp b/test/eigen2/eigen2_miscmatrices.cpp deleted file mode 100644 index 8bbb20cc8..000000000 --- a/test/eigen2/eigen2_miscmatrices.cpp +++ /dev/null @@ -1,48 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void miscMatrices(const MatrixType& m) -{ - /* this test covers the following files: - DiagonalMatrix.h Ones.h - */ - - typedef typename MatrixType::Scalar Scalar; - typedef Matrix VectorType; - typedef Matrix RowVectorType; - int rows = m.rows(); - int cols = m.cols(); - - int r = ei_random(0, rows-1), r2 = ei_random(0, rows-1), c = ei_random(0, cols-1); - VERIFY_IS_APPROX(MatrixType::Ones(rows,cols)(r,c), static_cast(1)); - MatrixType m1 = MatrixType::Ones(rows,cols); - VERIFY_IS_APPROX(m1(r,c), static_cast(1)); - VectorType v1 = VectorType::Random(rows); - v1[0]; - Matrix - square = v1.asDiagonal(); - if(r==r2) VERIFY_IS_APPROX(square(r,r2), v1[r]); - else VERIFY_IS_MUCH_SMALLER_THAN(square(r,r2), static_cast(1)); - square = MatrixType::Zero(rows, rows); - square.diagonal() = VectorType::Ones(rows); - VERIFY_IS_APPROX(square, MatrixType::Identity(rows, rows)); -} - -void test_eigen2_miscmatrices() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( miscMatrices(Matrix()) ); - CALL_SUBTEST_2( miscMatrices(Matrix4d()) ); - CALL_SUBTEST_3( miscMatrices(MatrixXcf(3, 3)) ); - CALL_SUBTEST_4( miscMatrices(MatrixXi(8, 12)) ); - CALL_SUBTEST_5( miscMatrices(MatrixXcd(20, 20)) ); - } -} diff --git a/test/eigen2/eigen2_mixingtypes.cpp b/test/eigen2/eigen2_mixingtypes.cpp deleted file mode 100644 index fb5ac5dda..000000000 --- a/test/eigen2/eigen2_mixingtypes.cpp +++ /dev/null @@ -1,77 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_NO_STATIC_ASSERT -#define EIGEN_NO_STATIC_ASSERT // turn static asserts into runtime asserts in order to check them -#endif - -#ifndef EIGEN_DONT_VECTORIZE -#define EIGEN_DONT_VECTORIZE // SSE intrinsics aren't designed to allow mixing types -#endif - -#include "main.h" - - -template void mixingtypes(int size = SizeAtCompileType) -{ - typedef Matrix Mat_f; - typedef Matrix Mat_d; - typedef Matrix, SizeAtCompileType, SizeAtCompileType> Mat_cf; - typedef Matrix, SizeAtCompileType, SizeAtCompileType> Mat_cd; - typedef Matrix Vec_f; - typedef Matrix Vec_d; - typedef Matrix, SizeAtCompileType, 1> Vec_cf; - typedef Matrix, SizeAtCompileType, 1> Vec_cd; - - Mat_f mf(size,size); - Mat_d md(size,size); - Mat_cf mcf(size,size); - Mat_cd mcd(size,size); - Vec_f vf(size,1); - Vec_d vd(size,1); - Vec_cf vcf(size,1); - Vec_cd vcd(size,1); - - mf+mf; - VERIFY_RAISES_ASSERT(mf+md); - VERIFY_RAISES_ASSERT(mf+mcf); - VERIFY_RAISES_ASSERT(vf=vd); - VERIFY_RAISES_ASSERT(vf+=vd); - VERIFY_RAISES_ASSERT(mcd=md); - - mf*mf; - md*mcd; - mcd*md; - mf*vcf; - mcf*vf; - mcf *= mf; - vcd = md*vcd; - vcf = mcf*vf; -#if 0 - // these are know generating hard build errors in eigen3 - VERIFY_RAISES_ASSERT(mf*md); - VERIFY_RAISES_ASSERT(mcf*mcd); - VERIFY_RAISES_ASSERT(mcf*vcd); - VERIFY_RAISES_ASSERT(vcf = mf*vf); - - vf.eigen2_dot(vf); - VERIFY_RAISES_ASSERT(vd.eigen2_dot(vf)); - VERIFY_RAISES_ASSERT(vcf.eigen2_dot(vf)); // yeah eventually we should allow this but i'm too lazy to make that change now in Dot.h - // especially as that might be rewritten as cwise product .sum() which would make that automatic. -#endif -} - -void test_eigen2_mixingtypes() -{ - // check that our operator new is indeed called: - CALL_SUBTEST_1(mixingtypes<3>()); - CALL_SUBTEST_2(mixingtypes<4>()); - CALL_SUBTEST_3(mixingtypes(20)); -} diff --git a/test/eigen2/eigen2_newstdvector.cpp b/test/eigen2/eigen2_newstdvector.cpp deleted file mode 100644 index 5f9009901..000000000 --- a/test/eigen2/eigen2_newstdvector.cpp +++ /dev/null @@ -1,149 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#define EIGEN_USE_NEW_STDVECTOR -#include "main.h" -#include -#include - -template -void check_stdvector_matrix(const MatrixType& m) -{ - int rows = m.rows(); - int cols = m.cols(); - MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols); - std::vector > v(10, MatrixType(rows,cols)), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.resize(22,y); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((std::size_t)&(v[22]) == (std::size_t)&(v[21]) + sizeof(MatrixType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - MatrixType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; i -void check_stdvector_transform(const TransformType&) -{ - typedef typename TransformType::MatrixType MatrixType; - TransformType x(MatrixType::Random()), y(MatrixType::Random()); - std::vector > v(10), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.resize(22,y); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((std::size_t)&(v[22]) == (std::size_t)&(v[21]) + sizeof(TransformType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - TransformType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; i -void check_stdvector_quaternion(const QuaternionType&) -{ - typedef typename QuaternionType::Coefficients Coefficients; - QuaternionType x(Coefficients::Random()), y(Coefficients::Random()); - std::vector > v(10), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.resize(22,y); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((std::size_t)&(v[22]) == (std::size_t)&(v[21]) + sizeof(QuaternionType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - QuaternionType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; i -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// this hack is needed to make this file compiles with -pedantic (gcc) -#ifdef __GNUC__ -#define throw(X) -#endif -// discard stack allocation as that too bypasses malloc -#define EIGEN_STACK_ALLOCATION_LIMIT 0 -// any heap allocation will raise an assert -#define EIGEN_NO_MALLOC - -#include "main.h" - -template void nomalloc(const MatrixType& m) -{ - /* this test check no dynamic memory allocation are issued with fixed-size matrices - */ - - typedef typename MatrixType::Scalar Scalar; - typedef Matrix VectorType; - - int rows = m.rows(); - int cols = m.cols(); - - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - mzero = MatrixType::Zero(rows, cols), - identity = Matrix - ::Identity(rows, rows), - square = Matrix - ::Random(rows, rows); - VectorType v1 = VectorType::Random(rows), - v2 = VectorType::Random(rows), - vzero = VectorType::Zero(rows); - - Scalar s1 = ei_random(); - - int r = ei_random(0, rows-1), - c = ei_random(0, cols-1); - - VERIFY_IS_APPROX((m1+m2)*s1, s1*m1+s1*m2); - VERIFY_IS_APPROX((m1+m2)(r,c), (m1(r,c))+(m2(r,c))); - VERIFY_IS_APPROX(m1.cwise() * m1.block(0,0,rows,cols), m1.cwise() * m1); - VERIFY_IS_APPROX((m1*m1.transpose())*m2, m1*(m1.transpose()*m2)); -} - -void test_eigen2_nomalloc() -{ - // check that our operator new is indeed called: - VERIFY_RAISES_ASSERT(MatrixXd dummy = MatrixXd::Random(3,3)); - CALL_SUBTEST_1( nomalloc(Matrix()) ); - CALL_SUBTEST_2( nomalloc(Matrix4d()) ); - CALL_SUBTEST_3( nomalloc(Matrix()) ); -} diff --git a/test/eigen2/eigen2_packetmath.cpp b/test/eigen2/eigen2_packetmath.cpp deleted file mode 100644 index b1f325fe7..000000000 --- a/test/eigen2/eigen2_packetmath.cpp +++ /dev/null @@ -1,132 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -// using namespace Eigen; - -template bool areApprox(const Scalar* a, const Scalar* b, int size) -{ - for (int i=0; i const complex& min(const complex& a, const complex& b) -{ return a.real() < b.real() ? a : b; } - -template<> const complex& max(const complex& a, const complex& b) -{ return a.real() < b.real() ? b : a; } - -} - -template void packetmath() -{ - typedef typename ei_packet_traits::type Packet; - const int PacketSize = ei_packet_traits::size; - - const int size = PacketSize*4; - EIGEN_ALIGN_128 Scalar data1[ei_packet_traits::size*4]; - EIGEN_ALIGN_128 Scalar data2[ei_packet_traits::size*4]; - EIGEN_ALIGN_128 Packet packets[PacketSize*2]; - EIGEN_ALIGN_128 Scalar ref[ei_packet_traits::size*4]; - for (int i=0; i(); - data2[i] = ei_random(); - } - - ei_pstore(data2, ei_pload(data1)); - VERIFY(areApprox(data1, data2, PacketSize) && "aligned load/store"); - - for (int offset=0; offset(packets[0], packets[1]); - else if (offset==1) ei_palign<1>(packets[0], packets[1]); - else if (offset==2) ei_palign<2>(packets[0], packets[1]); - else if (offset==3) ei_palign<3>(packets[0], packets[1]); - ei_pstore(data2, packets[0]); - - for (int i=0; i Vector; - VERIFY(areApprox(ref, data2, PacketSize) && "ei_palign"); - } - - CHECK_CWISE(REF_ADD, ei_padd); - CHECK_CWISE(REF_SUB, ei_psub); - CHECK_CWISE(REF_MUL, ei_pmul); - #ifndef EIGEN_VECTORIZE_ALTIVEC - if (!ei_is_same_type::ret) - CHECK_CWISE(REF_DIV, ei_pdiv); - #endif - CHECK_CWISE(std::min, ei_pmin); - CHECK_CWISE(std::max, ei_pmax); - - for (int i=0; i() ); - CALL_SUBTEST_2( packetmath() ); - CALL_SUBTEST_3( packetmath() ); - CALL_SUBTEST_4( packetmath >() ); - } -} diff --git a/test/eigen2/eigen2_parametrizedline.cpp b/test/eigen2/eigen2_parametrizedline.cpp deleted file mode 100644 index 814728870..000000000 --- a/test/eigen2/eigen2_parametrizedline.cpp +++ /dev/null @@ -1,62 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include -#include -#include - -template void parametrizedline(const LineType& _line) -{ - /* this test covers the following files: - ParametrizedLine.h - */ - - const int dim = _line.dim(); - typedef typename LineType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - typedef Matrix MatrixType; - - VectorType p0 = VectorType::Random(dim); - VectorType p1 = VectorType::Random(dim); - - VectorType d0 = VectorType::Random(dim).normalized(); - - LineType l0(p0, d0); - - Scalar s0 = ei_random(); - Scalar s1 = ei_abs(ei_random()); - - VERIFY_IS_MUCH_SMALLER_THAN( l0.distance(p0), RealScalar(1) ); - VERIFY_IS_MUCH_SMALLER_THAN( l0.distance(p0+s0*d0), RealScalar(1) ); - VERIFY_IS_APPROX( (l0.projection(p1)-p1).norm(), l0.distance(p1) ); - VERIFY_IS_MUCH_SMALLER_THAN( l0.distance(l0.projection(p1)), RealScalar(1) ); - VERIFY_IS_APPROX( Scalar(l0.distance((p0+s0*d0) + d0.unitOrthogonal() * s1)), s1 ); - - // casting - const int Dim = LineType::AmbientDimAtCompileTime; - typedef typename GetDifferentType::type OtherScalar; - ParametrizedLine hp1f = l0.template cast(); - VERIFY_IS_APPROX(hp1f.template cast(),l0); - ParametrizedLine hp1d = l0.template cast(); - VERIFY_IS_APPROX(hp1d.template cast(),l0); -} - -void test_eigen2_parametrizedline() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( parametrizedline(ParametrizedLine()) ); - CALL_SUBTEST_2( parametrizedline(ParametrizedLine()) ); - CALL_SUBTEST_3( parametrizedline(ParametrizedLine()) ); - CALL_SUBTEST_4( parametrizedline(ParametrizedLine,5>()) ); - } -} diff --git a/test/eigen2/eigen2_prec_inverse_4x4.cpp b/test/eigen2/eigen2_prec_inverse_4x4.cpp deleted file mode 100644 index 8bfa55694..000000000 --- a/test/eigen2/eigen2_prec_inverse_4x4.cpp +++ /dev/null @@ -1,84 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include -#include - -template std::string type_name() { return "other"; } -template<> std::string type_name() { return "float"; } -template<> std::string type_name() { return "double"; } -template<> std::string type_name() { return "int"; } -template<> std::string type_name >() { return "complex"; } -template<> std::string type_name >() { return "complex"; } -template<> std::string type_name >() { return "complex"; } - -#define EIGEN_DEBUG_VAR(x) std::cerr << #x << " = " << x << std::endl; - -template inline typename NumTraits::Real epsilon() -{ - return std::numeric_limits::Real>::epsilon(); -} - -template void inverse_permutation_4x4() -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::RealScalar RealScalar; - Vector4i indices(0,1,2,3); - for(int i = 0; i < 24; ++i) - { - MatrixType m = MatrixType::Zero(); - m(indices(0),0) = 1; - m(indices(1),1) = 1; - m(indices(2),2) = 1; - m(indices(3),3) = 1; - MatrixType inv = m.inverse(); - double error = double( (m*inv-MatrixType::Identity()).norm() / epsilon() ); - VERIFY(error == 0.0); - std::next_permutation(indices.data(),indices.data()+4); - } -} - -template void inverse_general_4x4(int repeat) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::RealScalar RealScalar; - double error_sum = 0., error_max = 0.; - for(int i = 0; i < repeat; ++i) - { - MatrixType m; - RealScalar absdet; - do { - m = MatrixType::Random(); - absdet = ei_abs(m.determinant()); - } while(absdet < 10 * epsilon()); - MatrixType inv = m.inverse(); - double error = double( (m*inv-MatrixType::Identity()).norm() * absdet / epsilon() ); - error_sum += error; - error_max = std::max(error_max, error); - } - std::cerr << "inverse_general_4x4, Scalar = " << type_name() << std::endl; - double error_avg = error_sum / repeat; - EIGEN_DEBUG_VAR(error_avg); - EIGEN_DEBUG_VAR(error_max); - VERIFY(error_avg < (NumTraits::IsComplex ? 8.0 : 1.25)); - VERIFY(error_max < (NumTraits::IsComplex ? 64.0 : 20.0)); -} - -void test_eigen2_prec_inverse_4x4() -{ - CALL_SUBTEST_1((inverse_permutation_4x4())); - CALL_SUBTEST_1(( inverse_general_4x4(200000 * g_repeat) )); - - CALL_SUBTEST_2((inverse_permutation_4x4 >())); - CALL_SUBTEST_2(( inverse_general_4x4 >(200000 * g_repeat) )); - - CALL_SUBTEST_3((inverse_permutation_4x4())); - CALL_SUBTEST_3((inverse_general_4x4(50000 * g_repeat))); -} diff --git a/test/eigen2/eigen2_product_large.cpp b/test/eigen2/eigen2_product_large.cpp deleted file mode 100644 index 5149ef748..000000000 --- a/test/eigen2/eigen2_product_large.cpp +++ /dev/null @@ -1,45 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "product.h" - -void test_eigen2_product_large() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( product(MatrixXf(ei_random(1,320), ei_random(1,320))) ); - CALL_SUBTEST_2( product(MatrixXd(ei_random(1,320), ei_random(1,320))) ); - CALL_SUBTEST_3( product(MatrixXi(ei_random(1,320), ei_random(1,320))) ); - CALL_SUBTEST_4( product(MatrixXcf(ei_random(1,50), ei_random(1,50))) ); - CALL_SUBTEST_5( product(Matrix(ei_random(1,320), ei_random(1,320))) ); - } - -#ifdef EIGEN_TEST_PART_6 - { - // test a specific issue in DiagonalProduct - int N = 1000000; - VectorXf v = VectorXf::Ones(N); - MatrixXf m = MatrixXf::Ones(N,3); - m = (v+v).asDiagonal() * m; - VERIFY_IS_APPROX(m, MatrixXf::Constant(N,3,2)); - } - - { - // test deferred resizing in Matrix::operator= - MatrixXf a = MatrixXf::Random(10,4), b = MatrixXf::Random(4,10), c = a; - VERIFY_IS_APPROX((a = a * b), (c * b).eval()); - } - - { - MatrixXf mat1(10,10); mat1.setRandom(); - MatrixXf mat2(32,10); mat2.setRandom(); - MatrixXf result = mat1.row(2)*mat2.transpose(); - VERIFY_IS_APPROX(result, (mat1.row(2)*mat2.transpose()).eval()); - } -#endif -} diff --git a/test/eigen2/eigen2_product_small.cpp b/test/eigen2/eigen2_product_small.cpp deleted file mode 100644 index 4cd8c102f..000000000 --- a/test/eigen2/eigen2_product_small.cpp +++ /dev/null @@ -1,22 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#define EIGEN_NO_STATIC_ASSERT -#include "product.h" - -void test_eigen2_product_small() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( product(Matrix()) ); - CALL_SUBTEST_2( product(Matrix()) ); - CALL_SUBTEST_3( product(Matrix3d()) ); - CALL_SUBTEST_4( product(Matrix4d()) ); - CALL_SUBTEST_5( product(Matrix4f()) ); - } -} diff --git a/test/eigen2/eigen2_qr.cpp b/test/eigen2/eigen2_qr.cpp deleted file mode 100644 index 76977e4c1..000000000 --- a/test/eigen2/eigen2_qr.cpp +++ /dev/null @@ -1,69 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -template void qr(const MatrixType& m) -{ - /* this test covers the following files: - QR.h - */ - int rows = m.rows(); - int cols = m.cols(); - - typedef typename MatrixType::Scalar Scalar; - typedef Matrix SquareMatrixType; - typedef Matrix VectorType; - - MatrixType a = MatrixType::Random(rows,cols); - QR qrOfA(a); - VERIFY_IS_APPROX(a, qrOfA.matrixQ() * qrOfA.matrixR()); - VERIFY_IS_NOT_APPROX(a+MatrixType::Identity(rows, cols), qrOfA.matrixQ() * qrOfA.matrixR()); - - #if 0 // eigenvalues module not yet ready - SquareMatrixType b = a.adjoint() * a; - - // check tridiagonalization - Tridiagonalization tridiag(b); - VERIFY_IS_APPROX(b, tridiag.matrixQ() * tridiag.matrixT() * tridiag.matrixQ().adjoint()); - - // check hessenberg decomposition - HessenbergDecomposition hess(b); - VERIFY_IS_APPROX(b, hess.matrixQ() * hess.matrixH() * hess.matrixQ().adjoint()); - VERIFY_IS_APPROX(tridiag.matrixT(), hess.matrixH()); - b = SquareMatrixType::Random(cols,cols); - hess.compute(b); - VERIFY_IS_APPROX(b, hess.matrixQ() * hess.matrixH() * hess.matrixQ().adjoint()); - #endif -} - -void test_eigen2_qr() -{ - for(int i = 0; i < 1; i++) { - CALL_SUBTEST_1( qr(Matrix2f()) ); - CALL_SUBTEST_2( qr(Matrix4d()) ); - CALL_SUBTEST_3( qr(MatrixXf(12,8)) ); - CALL_SUBTEST_4( qr(MatrixXcd(5,5)) ); - CALL_SUBTEST_4( qr(MatrixXcd(7,3)) ); - } - -#ifdef EIGEN_TEST_PART_5 - // small isFullRank test - { - Matrix3d mat; - mat << 1, 45, 1, 2, 2, 2, 1, 2, 3; - VERIFY(mat.qr().isFullRank()); - mat << 1, 1, 1, 2, 2, 2, 1, 2, 3; - //always returns true in eigen2support - //VERIFY(!mat.qr().isFullRank()); - } - -#endif -} diff --git a/test/eigen2/eigen2_qtvector.cpp b/test/eigen2/eigen2_qtvector.cpp deleted file mode 100644 index 6cfb58a26..000000000 --- a/test/eigen2/eigen2_qtvector.cpp +++ /dev/null @@ -1,158 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#define EIGEN_WORK_AROUND_QT_BUG_CALLING_WRONG_OPERATOR_NEW_FIXED_IN_QT_4_5 - -#include "main.h" - -#include -#include - -#include - -template -void check_qtvector_matrix(const MatrixType& m) -{ - int rows = m.rows(); - int cols = m.cols(); - MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols); - QVector v(10, MatrixType(rows,cols)), w(20, y); - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], y); - } - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.fill(y,22); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((size_t)&(v[22]) == (size_t)&(v[21]) + sizeof(MatrixType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - MatrixType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(int i=23; i -void check_qtvector_transform(const TransformType&) -{ - typedef typename TransformType::MatrixType MatrixType; - TransformType x(MatrixType::Random()), y(MatrixType::Random()); - QVector v(10), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.fill(y,22); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((size_t)&(v[22]) == (size_t)&(v[21]) + sizeof(TransformType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - TransformType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; int(i) -void check_qtvector_quaternion(const QuaternionType&) -{ - typedef typename QuaternionType::Coefficients Coefficients; - QuaternionType x(Coefficients::Random()), y(Coefficients::Random()); - QVector v(10), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.fill(y,22); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((size_t)&(v[22]) == (size_t)&(v[21]) + sizeof(QuaternionType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - QuaternionType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; int(i) -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -template -void makeNoisyCohyperplanarPoints(int numPoints, - VectorType **points, - HyperplaneType *hyperplane, - typename VectorType::Scalar noiseAmplitude) -{ - typedef typename VectorType::Scalar Scalar; - const int size = points[0]->size(); - // pick a random hyperplane, store the coefficients of its equation - hyperplane->coeffs().resize(size + 1); - for(int j = 0; j < size + 1; j++) - { - do { - hyperplane->coeffs().coeffRef(j) = ei_random(); - } while(ei_abs(hyperplane->coeffs().coeff(j)) < 0.5); - } - - // now pick numPoints random points on this hyperplane - for(int i = 0; i < numPoints; i++) - { - VectorType& cur_point = *(points[i]); - do - { - cur_point = VectorType::Random(size)/*.normalized()*/; - // project cur_point onto the hyperplane - Scalar x = - (hyperplane->coeffs().start(size).cwise()*cur_point).sum(); - cur_point *= hyperplane->coeffs().coeff(size) / x; - } while( cur_point.norm() < 0.5 - || cur_point.norm() > 2.0 ); - } - - // add some noise to these points - for(int i = 0; i < numPoints; i++ ) - *(points[i]) += noiseAmplitude * VectorType::Random(size); -} - -template -void check_linearRegression(int numPoints, - VectorType **points, - const VectorType& original, - typename VectorType::Scalar tolerance) -{ - int size = points[0]->size(); - assert(size==2); - VectorType result(size); - linearRegression(numPoints, points, &result, 1); - typename VectorType::Scalar error = (result - original).norm() / original.norm(); - VERIFY(ei_abs(error) < ei_abs(tolerance)); -} - -template -void check_fitHyperplane(int numPoints, - VectorType **points, - const HyperplaneType& original, - typename VectorType::Scalar tolerance) -{ - int size = points[0]->size(); - HyperplaneType result(size); - fitHyperplane(numPoints, points, &result); - result.coeffs() *= original.coeffs().coeff(size)/result.coeffs().coeff(size); - typename VectorType::Scalar error = (result.coeffs() - original.coeffs()).norm() / original.coeffs().norm(); - std::cout << ei_abs(error) << " xxx " << ei_abs(tolerance) << std::endl; - VERIFY(ei_abs(error) < ei_abs(tolerance)); -} - -void test_eigen2_regression() -{ - for(int i = 0; i < g_repeat; i++) - { -#ifdef EIGEN_TEST_PART_1 - { - Vector2f points2f [1000]; - Vector2f *points2f_ptrs [1000]; - for(int i = 0; i < 1000; i++) points2f_ptrs[i] = &(points2f[i]); - Vector2f coeffs2f; - Hyperplane coeffs3f; - makeNoisyCohyperplanarPoints(1000, points2f_ptrs, &coeffs3f, 0.01f); - coeffs2f[0] = -coeffs3f.coeffs()[0]/coeffs3f.coeffs()[1]; - coeffs2f[1] = -coeffs3f.coeffs()[2]/coeffs3f.coeffs()[1]; - CALL_SUBTEST(check_linearRegression(10, points2f_ptrs, coeffs2f, 0.05f)); - CALL_SUBTEST(check_linearRegression(100, points2f_ptrs, coeffs2f, 0.01f)); - CALL_SUBTEST(check_linearRegression(1000, points2f_ptrs, coeffs2f, 0.002f)); - } -#endif -#ifdef EIGEN_TEST_PART_2 - { - Vector2f points2f [1000]; - Vector2f *points2f_ptrs [1000]; - for(int i = 0; i < 1000; i++) points2f_ptrs[i] = &(points2f[i]); - Hyperplane coeffs3f; - makeNoisyCohyperplanarPoints(1000, points2f_ptrs, &coeffs3f, 0.01f); - CALL_SUBTEST(check_fitHyperplane(10, points2f_ptrs, coeffs3f, 0.05f)); - CALL_SUBTEST(check_fitHyperplane(100, points2f_ptrs, coeffs3f, 0.01f)); - CALL_SUBTEST(check_fitHyperplane(1000, points2f_ptrs, coeffs3f, 0.002f)); - } -#endif -#ifdef EIGEN_TEST_PART_3 - { - Vector4d points4d [1000]; - Vector4d *points4d_ptrs [1000]; - for(int i = 0; i < 1000; i++) points4d_ptrs[i] = &(points4d[i]); - Hyperplane coeffs5d; - makeNoisyCohyperplanarPoints(1000, points4d_ptrs, &coeffs5d, 0.01); - CALL_SUBTEST(check_fitHyperplane(10, points4d_ptrs, coeffs5d, 0.05)); - CALL_SUBTEST(check_fitHyperplane(100, points4d_ptrs, coeffs5d, 0.01)); - CALL_SUBTEST(check_fitHyperplane(1000, points4d_ptrs, coeffs5d, 0.002)); - } -#endif -#ifdef EIGEN_TEST_PART_4 - { - VectorXcd *points11cd_ptrs[1000]; - for(int i = 0; i < 1000; i++) points11cd_ptrs[i] = new VectorXcd(11); - Hyperplane,Dynamic> *coeffs12cd = new Hyperplane,Dynamic>(11); - makeNoisyCohyperplanarPoints(1000, points11cd_ptrs, coeffs12cd, 0.01); - CALL_SUBTEST(check_fitHyperplane(100, points11cd_ptrs, *coeffs12cd, 0.025)); - CALL_SUBTEST(check_fitHyperplane(1000, points11cd_ptrs, *coeffs12cd, 0.006)); - delete coeffs12cd; - for(int i = 0; i < 1000; i++) delete points11cd_ptrs[i]; - } -#endif - } -} diff --git a/test/eigen2/eigen2_sizeof.cpp b/test/eigen2/eigen2_sizeof.cpp deleted file mode 100644 index ec1af5a06..000000000 --- a/test/eigen2/eigen2_sizeof.cpp +++ /dev/null @@ -1,31 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void verifySizeOf(const MatrixType&) -{ - typedef typename MatrixType::Scalar Scalar; - if (MatrixType::RowsAtCompileTime!=Dynamic && MatrixType::ColsAtCompileTime!=Dynamic) - VERIFY(sizeof(MatrixType)==sizeof(Scalar)*MatrixType::SizeAtCompileTime); - else - VERIFY(sizeof(MatrixType)==sizeof(Scalar*) + 2 * sizeof(typename MatrixType::Index)); -} - -void test_eigen2_sizeof() -{ - CALL_SUBTEST( verifySizeOf(Matrix()) ); - CALL_SUBTEST( verifySizeOf(Matrix4d()) ); - CALL_SUBTEST( verifySizeOf(Matrix()) ); - CALL_SUBTEST( verifySizeOf(Matrix()) ); - CALL_SUBTEST( verifySizeOf(MatrixXcf(3, 3)) ); - CALL_SUBTEST( verifySizeOf(MatrixXi(8, 12)) ); - CALL_SUBTEST( verifySizeOf(MatrixXcd(20, 20)) ); - CALL_SUBTEST( verifySizeOf(Matrix()) ); -} diff --git a/test/eigen2/eigen2_smallvectors.cpp b/test/eigen2/eigen2_smallvectors.cpp deleted file mode 100644 index 03962b17d..000000000 --- a/test/eigen2/eigen2_smallvectors.cpp +++ /dev/null @@ -1,42 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void smallVectors() -{ - typedef Matrix V2; - typedef Matrix V3; - typedef Matrix V4; - Scalar x1 = ei_random(), - x2 = ei_random(), - x3 = ei_random(), - x4 = ei_random(); - V2 v2(x1, x2); - V3 v3(x1, x2, x3); - V4 v4(x1, x2, x3, x4); - VERIFY_IS_APPROX(x1, v2.x()); - VERIFY_IS_APPROX(x1, v3.x()); - VERIFY_IS_APPROX(x1, v4.x()); - VERIFY_IS_APPROX(x2, v2.y()); - VERIFY_IS_APPROX(x2, v3.y()); - VERIFY_IS_APPROX(x2, v4.y()); - VERIFY_IS_APPROX(x3, v3.z()); - VERIFY_IS_APPROX(x3, v4.z()); - VERIFY_IS_APPROX(x4, v4.w()); -} - -void test_eigen2_smallvectors() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST( smallVectors() ); - CALL_SUBTEST( smallVectors() ); - CALL_SUBTEST( smallVectors() ); - } -} diff --git a/test/eigen2/eigen2_sparse_basic.cpp b/test/eigen2/eigen2_sparse_basic.cpp deleted file mode 100644 index 049077670..000000000 --- a/test/eigen2/eigen2_sparse_basic.cpp +++ /dev/null @@ -1,317 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Daniel Gomez Ferro -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "sparse.h" - -template -bool test_random_setter(SparseMatrix& sm, const DenseType& ref, const std::vector& nonzeroCoords) -{ - typedef SparseMatrix SparseType; - { - sm.setZero(); - SetterType w(sm); - std::vector remaining = nonzeroCoords; - while(!remaining.empty()) - { - int i = ei_random(0,remaining.size()-1); - w(remaining[i].x(),remaining[i].y()) = ref.coeff(remaining[i].x(),remaining[i].y()); - remaining[i] = remaining.back(); - remaining.pop_back(); - } - } - return sm.isApprox(ref); -} - -template -bool test_random_setter(DynamicSparseMatrix& sm, const DenseType& ref, const std::vector& nonzeroCoords) -{ - sm.setZero(); - std::vector remaining = nonzeroCoords; - while(!remaining.empty()) - { - int i = ei_random(0,remaining.size()-1); - sm.coeffRef(remaining[i].x(),remaining[i].y()) = ref.coeff(remaining[i].x(),remaining[i].y()); - remaining[i] = remaining.back(); - remaining.pop_back(); - } - return sm.isApprox(ref); -} - -template void sparse_basic(const SparseMatrixType& ref) -{ - const int rows = ref.rows(); - const int cols = ref.cols(); - typedef typename SparseMatrixType::Scalar Scalar; - enum { Flags = SparseMatrixType::Flags }; - - double density = std::max(8./(rows*cols), 0.01); - typedef Matrix DenseMatrix; - typedef Matrix DenseVector; - Scalar eps = 1e-6; - - SparseMatrixType m(rows, cols); - DenseMatrix refMat = DenseMatrix::Zero(rows, cols); - DenseVector vec1 = DenseVector::Random(rows); - Scalar s1 = ei_random(); - - std::vector zeroCoords; - std::vector nonzeroCoords; - initSparse(density, refMat, m, 0, &zeroCoords, &nonzeroCoords); - - if (zeroCoords.size()==0 || nonzeroCoords.size()==0) - return; - - // test coeff and coeffRef - for (int i=0; i<(int)zeroCoords.size(); ++i) - { - VERIFY_IS_MUCH_SMALLER_THAN( m.coeff(zeroCoords[i].x(),zeroCoords[i].y()), eps ); - if(ei_is_same_type >::ret) - VERIFY_RAISES_ASSERT( m.coeffRef(zeroCoords[0].x(),zeroCoords[0].y()) = 5 ); - } - VERIFY_IS_APPROX(m, refMat); - - m.coeffRef(nonzeroCoords[0].x(), nonzeroCoords[0].y()) = Scalar(5); - refMat.coeffRef(nonzeroCoords[0].x(), nonzeroCoords[0].y()) = Scalar(5); - - VERIFY_IS_APPROX(m, refMat); - /* - // test InnerIterators and Block expressions - for (int t=0; t<10; ++t) - { - int j = ei_random(0,cols-1); - int i = ei_random(0,rows-1); - int w = ei_random(1,cols-j-1); - int h = ei_random(1,rows-i-1); - -// VERIFY_IS_APPROX(m.block(i,j,h,w), refMat.block(i,j,h,w)); - for(int c=0; c w(m); -// for (int i=0; icoeffRef(nonzeroCoords[i].x(),nonzeroCoords[i].y()) = refMat.coeff(nonzeroCoords[i].x(),nonzeroCoords[i].y()); -// } -// } -// VERIFY_IS_APPROX(m, refMat); - - // random setter -// { -// m.setZero(); -// VERIFY_IS_NOT_APPROX(m, refMat); -// SparseSetter w(m); -// std::vector remaining = nonzeroCoords; -// while(!remaining.empty()) -// { -// int i = ei_random(0,remaining.size()-1); -// w->coeffRef(remaining[i].x(),remaining[i].y()) = refMat.coeff(remaining[i].x(),remaining[i].y()); -// remaining[i] = remaining.back(); -// remaining.pop_back(); -// } -// } -// VERIFY_IS_APPROX(m, refMat); - - VERIFY(( test_random_setter >(m,refMat,nonzeroCoords) )); - #ifdef EIGEN_UNORDERED_MAP_SUPPORT - VERIFY(( test_random_setter >(m,refMat,nonzeroCoords) )); - #endif - #ifdef _DENSE_HASH_MAP_H_ - VERIFY(( test_random_setter >(m,refMat,nonzeroCoords) )); - #endif - #ifdef _SPARSE_HASH_MAP_H_ - VERIFY(( test_random_setter >(m,refMat,nonzeroCoords) )); - #endif - - // test fillrand - { - DenseMatrix m1(rows,cols); - m1.setZero(); - SparseMatrixType m2(rows,cols); - m2.startFill(); - for (int j=0; j(0,rows-1); - if (m1.coeff(i,j)==Scalar(0)) - m2.fillrand(i,j) = m1(i,j) = ei_random(); - } - } - m2.endFill(); - VERIFY_IS_APPROX(m2,m1); - } - - // test RandomSetter - /*{ - SparseMatrixType m1(rows,cols), m2(rows,cols); - DenseMatrix refM1 = DenseMatrix::Zero(rows, rows); - initSparse(density, refM1, m1); - { - Eigen::RandomSetter setter(m2); - for (int j=0; j(density, refM1, m1); - initSparse(density, refM2, m2); - initSparse(density, refM3, m3); - initSparse(density, refM4, m4); - - VERIFY_IS_APPROX(m1+m2, refM1+refM2); - VERIFY_IS_APPROX(m1+m2+m3, refM1+refM2+refM3); - VERIFY_IS_APPROX(m3.cwise()*(m1+m2), refM3.cwise()*(refM1+refM2)); - VERIFY_IS_APPROX(m1*s1-m2, refM1*s1-refM2); - - VERIFY_IS_APPROX(m1*=s1, refM1*=s1); - VERIFY_IS_APPROX(m1/=s1, refM1/=s1); - - VERIFY_IS_APPROX(m1+=m2, refM1+=refM2); - VERIFY_IS_APPROX(m1-=m2, refM1-=refM2); - - VERIFY_IS_APPROX(m1.col(0).eigen2_dot(refM2.row(0)), refM1.col(0).eigen2_dot(refM2.row(0))); - - refM4.setRandom(); - // sparse cwise* dense - VERIFY_IS_APPROX(m3.cwise()*refM4, refM3.cwise()*refM4); -// VERIFY_IS_APPROX(m3.cwise()/refM4, refM3.cwise()/refM4); - } - - // test innerVector() - { - DenseMatrix refMat2 = DenseMatrix::Zero(rows, rows); - SparseMatrixType m2(rows, rows); - initSparse(density, refMat2, m2); - int j0 = ei_random(0,rows-1); - int j1 = ei_random(0,rows-1); - VERIFY_IS_APPROX(m2.innerVector(j0), refMat2.col(j0)); - VERIFY_IS_APPROX(m2.innerVector(j0)+m2.innerVector(j1), refMat2.col(j0)+refMat2.col(j1)); - //m2.innerVector(j0) = 2*m2.innerVector(j1); - //refMat2.col(j0) = 2*refMat2.col(j1); - //VERIFY_IS_APPROX(m2, refMat2); - } - - // test innerVectors() - { - DenseMatrix refMat2 = DenseMatrix::Zero(rows, rows); - SparseMatrixType m2(rows, rows); - initSparse(density, refMat2, m2); - int j0 = ei_random(0,rows-2); - int j1 = ei_random(0,rows-2); - int n0 = ei_random(1,rows-std::max(j0,j1)); - VERIFY_IS_APPROX(m2.innerVectors(j0,n0), refMat2.block(0,j0,rows,n0)); - VERIFY_IS_APPROX(m2.innerVectors(j0,n0)+m2.innerVectors(j1,n0), - refMat2.block(0,j0,rows,n0)+refMat2.block(0,j1,rows,n0)); - //m2.innerVectors(j0,n0) = m2.innerVectors(j0,n0) + m2.innerVectors(j1,n0); - //refMat2.block(0,j0,rows,n0) = refMat2.block(0,j0,rows,n0) + refMat2.block(0,j1,rows,n0); - } - - // test transpose - { - DenseMatrix refMat2 = DenseMatrix::Zero(rows, rows); - SparseMatrixType m2(rows, rows); - initSparse(density, refMat2, m2); - VERIFY_IS_APPROX(m2.transpose().eval(), refMat2.transpose().eval()); - VERIFY_IS_APPROX(m2.transpose(), refMat2.transpose()); - } - - // test prune - { - SparseMatrixType m2(rows, rows); - DenseMatrix refM2(rows, rows); - refM2.setZero(); - int countFalseNonZero = 0; - int countTrueNonZero = 0; - m2.startFill(); - for (int j=0; j(0,1); - if (x<0.1) - { - // do nothing - } - else if (x<0.5) - { - countFalseNonZero++; - m2.fill(i,j) = Scalar(0); - } - else - { - countTrueNonZero++; - m2.fill(i,j) = refM2(i,j) = Scalar(1); - } - } - m2.endFill(); - VERIFY(countFalseNonZero+countTrueNonZero == m2.nonZeros()); - VERIFY_IS_APPROX(m2, refM2); - m2.prune(1); - VERIFY(countTrueNonZero==m2.nonZeros()); - VERIFY_IS_APPROX(m2, refM2); - } -} - -void test_eigen2_sparse_basic() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( sparse_basic(SparseMatrix(8, 8)) ); - CALL_SUBTEST_2( sparse_basic(SparseMatrix >(16, 16)) ); - CALL_SUBTEST_1( sparse_basic(SparseMatrix(33, 33)) ); - - CALL_SUBTEST_3( sparse_basic(DynamicSparseMatrix(8, 8)) ); - } -} diff --git a/test/eigen2/eigen2_sparse_product.cpp b/test/eigen2/eigen2_sparse_product.cpp deleted file mode 100644 index d28e76dff..000000000 --- a/test/eigen2/eigen2_sparse_product.cpp +++ /dev/null @@ -1,115 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Daniel Gomez Ferro -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "sparse.h" - -template void sparse_product(const SparseMatrixType& ref) -{ - const int rows = ref.rows(); - const int cols = ref.cols(); - typedef typename SparseMatrixType::Scalar Scalar; - enum { Flags = SparseMatrixType::Flags }; - - double density = std::max(8./(rows*cols), 0.01); - typedef Matrix DenseMatrix; - typedef Matrix DenseVector; - - // test matrix-matrix product - { - DenseMatrix refMat2 = DenseMatrix::Zero(rows, rows); - DenseMatrix refMat3 = DenseMatrix::Zero(rows, rows); - DenseMatrix refMat4 = DenseMatrix::Zero(rows, rows); - DenseMatrix dm4 = DenseMatrix::Zero(rows, rows); - SparseMatrixType m2(rows, rows); - SparseMatrixType m3(rows, rows); - SparseMatrixType m4(rows, rows); - initSparse(density, refMat2, m2); - initSparse(density, refMat3, m3); - initSparse(density, refMat4, m4); - VERIFY_IS_APPROX(m4=m2*m3, refMat4=refMat2*refMat3); - VERIFY_IS_APPROX(m4=m2.transpose()*m3, refMat4=refMat2.transpose()*refMat3); - VERIFY_IS_APPROX(m4=m2.transpose()*m3.transpose(), refMat4=refMat2.transpose()*refMat3.transpose()); - VERIFY_IS_APPROX(m4=m2*m3.transpose(), refMat4=refMat2*refMat3.transpose()); - - // sparse * dense - VERIFY_IS_APPROX(dm4=m2*refMat3, refMat4=refMat2*refMat3); - VERIFY_IS_APPROX(dm4=m2*refMat3.transpose(), refMat4=refMat2*refMat3.transpose()); - VERIFY_IS_APPROX(dm4=m2.transpose()*refMat3, refMat4=refMat2.transpose()*refMat3); - VERIFY_IS_APPROX(dm4=m2.transpose()*refMat3.transpose(), refMat4=refMat2.transpose()*refMat3.transpose()); - - // dense * sparse - VERIFY_IS_APPROX(dm4=refMat2*m3, refMat4=refMat2*refMat3); - VERIFY_IS_APPROX(dm4=refMat2*m3.transpose(), refMat4=refMat2*refMat3.transpose()); - VERIFY_IS_APPROX(dm4=refMat2.transpose()*m3, refMat4=refMat2.transpose()*refMat3); - VERIFY_IS_APPROX(dm4=refMat2.transpose()*m3.transpose(), refMat4=refMat2.transpose()*refMat3.transpose()); - - VERIFY_IS_APPROX(m3=m3*m3, refMat3=refMat3*refMat3); - } - - // test matrix - diagonal product - if(false) // it compiles, but the precision is terrible. probably doesn't matter in this branch.... - { - DenseMatrix refM2 = DenseMatrix::Zero(rows, rows); - DenseMatrix refM3 = DenseMatrix::Zero(rows, rows); - DiagonalMatrix d1(DenseVector::Random(rows)); - SparseMatrixType m2(rows, rows); - SparseMatrixType m3(rows, rows); - initSparse(density, refM2, m2); - initSparse(density, refM3, m3); - VERIFY_IS_APPROX(m3=m2*d1, refM3=refM2*d1); - VERIFY_IS_APPROX(m3=m2.transpose()*d1, refM3=refM2.transpose()*d1); - VERIFY_IS_APPROX(m3=d1*m2, refM3=d1*refM2); - VERIFY_IS_APPROX(m3=d1*m2.transpose(), refM3=d1 * refM2.transpose()); - } - - // test self adjoint products - { - DenseMatrix b = DenseMatrix::Random(rows, rows); - DenseMatrix x = DenseMatrix::Random(rows, rows); - DenseMatrix refX = DenseMatrix::Random(rows, rows); - DenseMatrix refUp = DenseMatrix::Zero(rows, rows); - DenseMatrix refLo = DenseMatrix::Zero(rows, rows); - DenseMatrix refS = DenseMatrix::Zero(rows, rows); - SparseMatrixType mUp(rows, rows); - SparseMatrixType mLo(rows, rows); - SparseMatrixType mS(rows, rows); - do { - initSparse(density, refUp, mUp, ForceRealDiag|/*ForceNonZeroDiag|*/MakeUpperTriangular); - } while (refUp.isZero()); - refLo = refUp.transpose().conjugate(); - mLo = mUp.transpose().conjugate(); - refS = refUp + refLo; - refS.diagonal() *= 0.5; - mS = mUp + mLo; - for (int k=0; k()*b, refX=refS*b); - VERIFY_IS_APPROX(x=mLo.template marked()*b, refX=refS*b); - VERIFY_IS_APPROX(x=mS.template marked()*b, refX=refS*b); - } - -} - -void test_eigen2_sparse_product() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( sparse_product(SparseMatrix(8, 8)) ); - CALL_SUBTEST_2( sparse_product(SparseMatrix >(16, 16)) ); - CALL_SUBTEST_1( sparse_product(SparseMatrix(33, 33)) ); - - CALL_SUBTEST_3( sparse_product(DynamicSparseMatrix(8, 8)) ); - } -} diff --git a/test/eigen2/eigen2_sparse_solvers.cpp b/test/eigen2/eigen2_sparse_solvers.cpp deleted file mode 100644 index 3aef27ab4..000000000 --- a/test/eigen2/eigen2_sparse_solvers.cpp +++ /dev/null @@ -1,200 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Daniel Gomez Ferro -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "sparse.h" - -template void -initSPD(double density, - Matrix& refMat, - SparseMatrix& sparseMat) -{ - Matrix aux(refMat.rows(),refMat.cols()); - initSparse(density,refMat,sparseMat); - refMat = refMat * refMat.adjoint(); - for (int k=0; k<2; ++k) - { - initSparse(density,aux,sparseMat,ForceNonZeroDiag); - refMat += aux * aux.adjoint(); - } - sparseMat.startFill(); - for (int j=0 ; j void sparse_solvers(int rows, int cols) -{ - double density = std::max(8./(rows*cols), 0.01); - typedef Matrix DenseMatrix; - typedef Matrix DenseVector; - // Scalar eps = 1e-6; - - DenseVector vec1 = DenseVector::Random(rows); - - std::vector zeroCoords; - std::vector nonzeroCoords; - - // test triangular solver - { - DenseVector vec2 = vec1, vec3 = vec1; - SparseMatrix m2(rows, cols); - DenseMatrix refMat2 = DenseMatrix::Zero(rows, cols); - - // lower - initSparse(density, refMat2, m2, ForceNonZeroDiag|MakeLowerTriangular, &zeroCoords, &nonzeroCoords); - VERIFY_IS_APPROX(refMat2.template marked().solveTriangular(vec2), - m2.template marked().solveTriangular(vec3)); - - // lower - transpose - initSparse(density, refMat2, m2, ForceNonZeroDiag|MakeLowerTriangular, &zeroCoords, &nonzeroCoords); - VERIFY_IS_APPROX(refMat2.template marked().transpose().solveTriangular(vec2), - m2.template marked().transpose().solveTriangular(vec3)); - - // upper - initSparse(density, refMat2, m2, ForceNonZeroDiag|MakeUpperTriangular, &zeroCoords, &nonzeroCoords); - VERIFY_IS_APPROX(refMat2.template marked().solveTriangular(vec2), - m2.template marked().solveTriangular(vec3)); - - // upper - transpose - initSparse(density, refMat2, m2, ForceNonZeroDiag|MakeUpperTriangular, &zeroCoords, &nonzeroCoords); - VERIFY_IS_APPROX(refMat2.template marked().transpose().solveTriangular(vec2), - m2.template marked().transpose().solveTriangular(vec3)); - } - - // test LLT - { - // TODO fix the issue with complex (see SparseLLT::solveInPlace) - SparseMatrix m2(rows, cols); - DenseMatrix refMat2(rows, cols); - - DenseVector b = DenseVector::Random(cols); - DenseVector refX(cols), x(cols); - - initSPD(density, refMat2, m2); - - refMat2.llt().solve(b, &refX); - typedef SparseMatrix SparseSelfAdjointMatrix; - if (!NumTraits::IsComplex) - { - x = b; - SparseLLT (m2).solveInPlace(x); - VERIFY(refX.isApprox(x,test_precision()) && "LLT: default"); - } - #ifdef EIGEN_CHOLMOD_SUPPORT - x = b; - SparseLLT(m2).solveInPlace(x); - VERIFY(refX.isApprox(x,test_precision()) && "LLT: cholmod"); - #endif - if (!NumTraits::IsComplex) - { - #ifdef EIGEN_TAUCS_SUPPORT - x = b; - SparseLLT(m2,IncompleteFactorization).solveInPlace(x); - VERIFY(refX.isApprox(x,test_precision()) && "LLT: taucs (IncompleteFactorization)"); - x = b; - SparseLLT(m2,SupernodalMultifrontal).solveInPlace(x); - VERIFY(refX.isApprox(x,test_precision()) && "LLT: taucs (SupernodalMultifrontal)"); - x = b; - SparseLLT(m2,SupernodalLeftLooking).solveInPlace(x); - VERIFY(refX.isApprox(x,test_precision()) && "LLT: taucs (SupernodalLeftLooking)"); - #endif - } - } - - // test LDLT - if (!NumTraits::IsComplex) - { - // TODO fix the issue with complex (see SparseLDLT::solveInPlace) - SparseMatrix m2(rows, cols); - DenseMatrix refMat2(rows, cols); - - DenseVector b = DenseVector::Random(cols); - DenseVector refX(cols), x(cols); - - //initSPD(density, refMat2, m2); - initSparse(density, refMat2, m2, ForceNonZeroDiag|MakeUpperTriangular, 0, 0); - refMat2 += refMat2.adjoint(); - refMat2.diagonal() *= 0.5; - - refMat2.ldlt().solve(b, &refX); - typedef SparseMatrix SparseSelfAdjointMatrix; - x = b; - SparseLDLT ldlt(m2); - if (ldlt.succeeded()) - ldlt.solveInPlace(x); - VERIFY(refX.isApprox(x,test_precision()) && "LDLT: default"); - } - - // test LU - { - static int count = 0; - SparseMatrix m2(rows, cols); - DenseMatrix refMat2(rows, cols); - - DenseVector b = DenseVector::Random(cols); - DenseVector refX(cols), x(cols); - - initSparse(density, refMat2, m2, ForceNonZeroDiag, &zeroCoords, &nonzeroCoords); - - LU refLu(refMat2); - refLu.solve(b, &refX); - #if defined(EIGEN_SUPERLU_SUPPORT) || defined(EIGEN_UMFPACK_SUPPORT) - Scalar refDet = refLu.determinant(); - #endif - x.setZero(); - // // SparseLU > (m2).solve(b,&x); - // // VERIFY(refX.isApprox(x,test_precision()) && "LU: default"); - #ifdef EIGEN_SUPERLU_SUPPORT - { - x.setZero(); - SparseLU,SuperLU> slu(m2); - if (slu.succeeded()) - { - if (slu.solve(b,&x)) { - VERIFY(refX.isApprox(x,test_precision()) && "LU: SuperLU"); - } - // std::cerr << refDet << " == " << slu.determinant() << "\n"; - if (count==0) { - VERIFY_IS_APPROX(refDet,slu.determinant()); // FIXME det is not very stable for complex - } - } - } - #endif - #ifdef EIGEN_UMFPACK_SUPPORT - { - // check solve - x.setZero(); - SparseLU,UmfPack> slu(m2); - if (slu.succeeded()) { - if (slu.solve(b,&x)) { - if (count==0) { - VERIFY(refX.isApprox(x,test_precision()) && "LU: umfpack"); // FIXME solve is not very stable for complex - } - } - VERIFY_IS_APPROX(refDet,slu.determinant()); - // TODO check the extracted data - //std::cerr << slu.matrixL() << "\n"; - } - } - #endif - count++; - } - -} - -void test_eigen2_sparse_solvers() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( sparse_solvers(8, 8) ); - CALL_SUBTEST_2( sparse_solvers >(16, 16) ); - CALL_SUBTEST_1( sparse_solvers(101, 101) ); - } -} diff --git a/test/eigen2/eigen2_sparse_vector.cpp b/test/eigen2/eigen2_sparse_vector.cpp deleted file mode 100644 index e6d2d77a1..000000000 --- a/test/eigen2/eigen2_sparse_vector.cpp +++ /dev/null @@ -1,84 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Daniel Gomez Ferro -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "sparse.h" - -template void sparse_vector(int rows, int cols) -{ - double densityMat = std::max(8./(rows*cols), 0.01); - double densityVec = std::max(8./float(rows), 0.1); - typedef Matrix DenseMatrix; - typedef Matrix DenseVector; - typedef SparseVector SparseVectorType; - typedef SparseMatrix SparseMatrixType; - Scalar eps = 1e-6; - - SparseMatrixType m1(rows,cols); - SparseVectorType v1(rows), v2(rows), v3(rows); - DenseMatrix refM1 = DenseMatrix::Zero(rows, cols); - DenseVector refV1 = DenseVector::Random(rows), - refV2 = DenseVector::Random(rows), - refV3 = DenseVector::Random(rows); - - std::vector zerocoords, nonzerocoords; - initSparse(densityVec, refV1, v1, &zerocoords, &nonzerocoords); - initSparse(densityMat, refM1, m1); - - initSparse(densityVec, refV2, v2); - initSparse(densityVec, refV3, v3); - - Scalar s1 = ei_random(); - - // test coeff and coeffRef - for (unsigned int i=0; i(8, 8) ); - CALL_SUBTEST_2( sparse_vector >(16, 16) ); - CALL_SUBTEST_1( sparse_vector(299, 535) ); - } -} - diff --git a/test/eigen2/eigen2_stdvector.cpp b/test/eigen2/eigen2_stdvector.cpp deleted file mode 100644 index 6ab05c20a..000000000 --- a/test/eigen2/eigen2_stdvector.cpp +++ /dev/null @@ -1,148 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include -#include "main.h" -#include - -template -void check_stdvector_matrix(const MatrixType& m) -{ - int rows = m.rows(); - int cols = m.cols(); - MatrixType x = MatrixType::Random(rows,cols), y = MatrixType::Random(rows,cols); - std::vector > v(10, MatrixType(rows,cols)), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.resize(22,y); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((std::size_t)&(v[22]) == (std::size_t)&(v[21]) + sizeof(MatrixType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - MatrixType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; i -void check_stdvector_transform(const TransformType&) -{ - typedef typename TransformType::MatrixType MatrixType; - TransformType x(MatrixType::Random()), y(MatrixType::Random()); - std::vector > v(10), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.resize(22,y); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((std::size_t)&(v[22]) == (std::size_t)&(v[21]) + sizeof(TransformType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - TransformType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; i -void check_stdvector_quaternion(const QuaternionType&) -{ - typedef typename QuaternionType::Coefficients Coefficients; - QuaternionType x(Coefficients::Random()), y(Coefficients::Random()); - std::vector > v(10), w(20, y); - v[5] = x; - w[6] = v[5]; - VERIFY_IS_APPROX(w[6], v[5]); - v = w; - for(int i = 0; i < 20; i++) - { - VERIFY_IS_APPROX(w[i], v[i]); - } - - v.resize(21); - v[20] = x; - VERIFY_IS_APPROX(v[20], x); - v.resize(22,y); - VERIFY_IS_APPROX(v[21], y); - v.push_back(x); - VERIFY_IS_APPROX(v[22], x); - VERIFY((std::size_t)&(v[22]) == (std::size_t)&(v[21]) + sizeof(QuaternionType)); - - // do a lot of push_back such that the vector gets internally resized - // (with memory reallocation) - QuaternionType* ref = &w[0]; - for(int i=0; i<30 || ((ref==&w[0]) && i<300); ++i) - v.push_back(w[i%w.size()]); - for(unsigned int i=23; i -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -// check minor separately in order to avoid the possible creation of a zero-sized -// array. Comes from a compilation error with gcc-3.4 or gcc-4 with -ansi -pedantic. -// Another solution would be to declare the array like this: T m_data[Size==0?1:Size]; in ei_matrix_storage -// but this is probably not bad to raise such an error at compile time... -template struct CheckMinor -{ - typedef Matrix MatrixType; - CheckMinor(MatrixType& m1, int r1, int c1) - { - int rows = m1.rows(); - int cols = m1.cols(); - - Matrix mi = m1.minor(0,0).eval(); - VERIFY_IS_APPROX(mi, m1.block(1,1,rows-1,cols-1)); - mi = m1.minor(r1,c1); - VERIFY_IS_APPROX(mi.transpose(), m1.transpose().minor(c1,r1)); - //check operator(), both constant and non-constant, on minor() - m1.minor(r1,c1)(0,0) = m1.minor(0,0)(0,0); - } -}; - -template struct CheckMinor -{ - typedef Matrix MatrixType; - CheckMinor(MatrixType&, int, int) {} -}; - -template void submatrices(const MatrixType& m) -{ - /* this test covers the following files: - Row.h Column.h Block.h Minor.h DiagonalCoeffs.h - */ - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::RealScalar RealScalar; - typedef Matrix VectorType; - typedef Matrix RowVectorType; - int rows = m.rows(); - int cols = m.cols(); - - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - mzero = MatrixType::Zero(rows, cols), - ones = MatrixType::Ones(rows, cols), - identity = Matrix - ::Identity(rows, rows), - square = Matrix - ::Random(rows, rows); - VectorType v1 = VectorType::Random(rows), - v2 = VectorType::Random(rows), - v3 = VectorType::Random(rows), - vzero = VectorType::Zero(rows); - - Scalar s1 = ei_random(); - - int r1 = ei_random(0,rows-1); - int r2 = ei_random(r1,rows-1); - int c1 = ei_random(0,cols-1); - int c2 = ei_random(c1,cols-1); - - //check row() and col() - VERIFY_IS_APPROX(m1.col(c1).transpose(), m1.transpose().row(c1)); - VERIFY_IS_APPROX(square.row(r1).eigen2_dot(m1.col(c1)), (square.lazy() * m1.conjugate())(r1,c1)); - //check operator(), both constant and non-constant, on row() and col() - m1.row(r1) += s1 * m1.row(r2); - m1.col(c1) += s1 * m1.col(c2); - - //check block() - Matrix b1(1,1); b1(0,0) = m1(r1,c1); - RowVectorType br1(m1.block(r1,0,1,cols)); - VectorType bc1(m1.block(0,c1,rows,1)); - VERIFY_IS_APPROX(b1, m1.block(r1,c1,1,1)); - VERIFY_IS_APPROX(m1.row(r1), br1); - VERIFY_IS_APPROX(m1.col(c1), bc1); - //check operator(), both constant and non-constant, on block() - m1.block(r1,c1,r2-r1+1,c2-c1+1) = s1 * m2.block(0, 0, r2-r1+1,c2-c1+1); - m1.block(r1,c1,r2-r1+1,c2-c1+1)(r2-r1,c2-c1) = m2.block(0, 0, r2-r1+1,c2-c1+1)(0,0); - - //check minor() - CheckMinor checkminor(m1,r1,c1); - - //check diagonal() - VERIFY_IS_APPROX(m1.diagonal(), m1.transpose().diagonal()); - m2.diagonal() = 2 * m1.diagonal(); - m2.diagonal()[0] *= 3; - VERIFY_IS_APPROX(m2.diagonal()[0], static_cast(6) * m1.diagonal()[0]); - - enum { - BlockRows = EIGEN_SIZE_MIN_PREFER_FIXED(MatrixType::RowsAtCompileTime,2), - BlockCols = EIGEN_SIZE_MIN_PREFER_FIXED(MatrixType::ColsAtCompileTime,5) - }; - if (rows>=5 && cols>=8) - { - // test fixed block() as lvalue - m1.template block(1,1) *= s1; - // test operator() on fixed block() both as constant and non-constant - m1.template block(1,1)(0, 3) = m1.template block<2,5>(1,1)(1,2); - // check that fixed block() and block() agree - Matrix b = m1.template block(3,3); - VERIFY_IS_APPROX(b, m1.block(3,3,BlockRows,BlockCols)); - } - - if (rows>2) - { - // test sub vectors - VERIFY_IS_APPROX(v1.template start<2>(), v1.block(0,0,2,1)); - VERIFY_IS_APPROX(v1.template start<2>(), v1.start(2)); - VERIFY_IS_APPROX(v1.template start<2>(), v1.segment(0,2)); - VERIFY_IS_APPROX(v1.template start<2>(), v1.template segment<2>(0)); - int i = rows-2; - VERIFY_IS_APPROX(v1.template end<2>(), v1.block(i,0,2,1)); - VERIFY_IS_APPROX(v1.template end<2>(), v1.end(2)); - VERIFY_IS_APPROX(v1.template end<2>(), v1.segment(i,2)); - VERIFY_IS_APPROX(v1.template end<2>(), v1.template segment<2>(i)); - i = ei_random(0,rows-2); - VERIFY_IS_APPROX(v1.segment(i,2), v1.template segment<2>(i)); - } - - // stress some basic stuffs with block matrices - VERIFY(ei_real(ones.col(c1).sum()) == RealScalar(rows)); - VERIFY(ei_real(ones.row(r1).sum()) == RealScalar(cols)); - - VERIFY(ei_real(ones.col(c1).eigen2_dot(ones.col(c2))) == RealScalar(rows)); - VERIFY(ei_real(ones.row(r1).eigen2_dot(ones.row(r2))) == RealScalar(cols)); -} - -void test_eigen2_submatrices() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( submatrices(Matrix()) ); - CALL_SUBTEST_2( submatrices(Matrix4d()) ); - CALL_SUBTEST_3( submatrices(MatrixXcf(3, 3)) ); - CALL_SUBTEST_4( submatrices(MatrixXi(8, 12)) ); - CALL_SUBTEST_5( submatrices(MatrixXcd(20, 20)) ); - CALL_SUBTEST_6( submatrices(MatrixXf(20, 20)) ); - } -} diff --git a/test/eigen2/eigen2_sum.cpp b/test/eigen2/eigen2_sum.cpp deleted file mode 100644 index b47057caa..000000000 --- a/test/eigen2/eigen2_sum.cpp +++ /dev/null @@ -1,71 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void matrixSum(const MatrixType& m) -{ - typedef typename MatrixType::Scalar Scalar; - - int rows = m.rows(); - int cols = m.cols(); - - MatrixType m1 = MatrixType::Random(rows, cols); - - VERIFY_IS_MUCH_SMALLER_THAN(MatrixType::Zero(rows, cols).sum(), Scalar(1)); - VERIFY_IS_APPROX(MatrixType::Ones(rows, cols).sum(), Scalar(float(rows*cols))); // the float() here to shut up excessive MSVC warning about int->complex conversion being lossy - Scalar x = Scalar(0); - for(int i = 0; i < rows; i++) for(int j = 0; j < cols; j++) x += m1(i,j); - VERIFY_IS_APPROX(m1.sum(), x); -} - -template void vectorSum(const VectorType& w) -{ - typedef typename VectorType::Scalar Scalar; - int size = w.size(); - - VectorType v = VectorType::Random(size); - for(int i = 1; i < size; i++) - { - Scalar s = Scalar(0); - for(int j = 0; j < i; j++) s += v[j]; - VERIFY_IS_APPROX(s, v.start(i).sum()); - } - - for(int i = 0; i < size-1; i++) - { - Scalar s = Scalar(0); - for(int j = i; j < size; j++) s += v[j]; - VERIFY_IS_APPROX(s, v.end(size-i).sum()); - } - - for(int i = 0; i < size/2; i++) - { - Scalar s = Scalar(0); - for(int j = i; j < size-i; j++) s += v[j]; - VERIFY_IS_APPROX(s, v.segment(i, size-2*i).sum()); - } -} - -void test_eigen2_sum() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( matrixSum(Matrix()) ); - CALL_SUBTEST_2( matrixSum(Matrix2f()) ); - CALL_SUBTEST_3( matrixSum(Matrix4d()) ); - CALL_SUBTEST_4( matrixSum(MatrixXcf(3, 3)) ); - CALL_SUBTEST_5( matrixSum(MatrixXf(8, 12)) ); - CALL_SUBTEST_6( matrixSum(MatrixXi(8, 12)) ); - } - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_5( vectorSum(VectorXf(5)) ); - CALL_SUBTEST_7( vectorSum(VectorXd(10)) ); - CALL_SUBTEST_5( vectorSum(VectorXf(33)) ); - } -} diff --git a/test/eigen2/eigen2_svd.cpp b/test/eigen2/eigen2_svd.cpp deleted file mode 100644 index d4689a56f..000000000 --- a/test/eigen2/eigen2_svd.cpp +++ /dev/null @@ -1,87 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include - -template void svd(const MatrixType& m) -{ - /* this test covers the following files: - SVD.h - */ - int rows = m.rows(); - int cols = m.cols(); - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - MatrixType a = MatrixType::Random(rows,cols); - Matrix b = - Matrix::Random(rows,1); - Matrix x(cols,1), x2(cols,1); - - RealScalar largerEps = test_precision(); - if (ei_is_same_type::ret) - largerEps = 1e-3f; - - { - SVD svd(a); - MatrixType sigma = MatrixType::Zero(rows,cols); - MatrixType matU = MatrixType::Zero(rows,rows); - sigma.block(0,0,cols,cols) = svd.singularValues().asDiagonal(); - matU.block(0,0,rows,cols) = svd.matrixU(); - VERIFY_IS_APPROX(a, matU * sigma * svd.matrixV().transpose()); - } - - - if (rows==cols) - { - if (ei_is_same_type::ret) - { - MatrixType a1 = MatrixType::Random(rows,cols); - a += a * a.adjoint() + a1 * a1.adjoint(); - } - SVD svd(a); - svd.solve(b, &x); - VERIFY_IS_APPROX(a * x,b); - } - - - if(rows==cols) - { - SVD svd(a); - MatrixType unitary, positive; - svd.computeUnitaryPositive(&unitary, &positive); - VERIFY_IS_APPROX(unitary * unitary.adjoint(), MatrixType::Identity(unitary.rows(),unitary.rows())); - VERIFY_IS_APPROX(positive, positive.adjoint()); - for(int i = 0; i < rows; i++) VERIFY(positive.diagonal()[i] >= 0); // cheap necessary (not sufficient) condition for positivity - VERIFY_IS_APPROX(unitary*positive, a); - - svd.computePositiveUnitary(&positive, &unitary); - VERIFY_IS_APPROX(unitary * unitary.adjoint(), MatrixType::Identity(unitary.rows(),unitary.rows())); - VERIFY_IS_APPROX(positive, positive.adjoint()); - for(int i = 0; i < rows; i++) VERIFY(positive.diagonal()[i] >= 0); // cheap necessary (not sufficient) condition for positivity - VERIFY_IS_APPROX(positive*unitary, a); - } -} - -void test_eigen2_svd() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( svd(Matrix3f()) ); - CALL_SUBTEST_2( svd(Matrix4d()) ); - CALL_SUBTEST_3( svd(MatrixXf(7,7)) ); - CALL_SUBTEST_4( svd(MatrixXd(14,7)) ); - // complex are not implemented yet -// CALL_SUBTEST( svd(MatrixXcd(6,6)) ); -// CALL_SUBTEST( svd(MatrixXcf(3,3)) ); - SVD s; - MatrixXf m = MatrixXf::Random(10,1); - s.compute(m); - } -} diff --git a/test/eigen2/eigen2_swap.cpp b/test/eigen2/eigen2_swap.cpp deleted file mode 100644 index f3a8846d9..000000000 --- a/test/eigen2/eigen2_swap.cpp +++ /dev/null @@ -1,83 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#define EIGEN_NO_STATIC_ASSERT -#include "main.h" - -template -struct other_matrix_type -{ - typedef int type; -}; - -template -struct other_matrix_type > -{ - typedef Matrix<_Scalar, _Rows, _Cols, _Options^RowMajor, _MaxRows, _MaxCols> type; -}; - -template void swap(const MatrixType& m) -{ - typedef typename other_matrix_type::type OtherMatrixType; - typedef typename MatrixType::Scalar Scalar; - - ei_assert((!ei_is_same_type::ret)); - int rows = m.rows(); - int cols = m.cols(); - - // construct 3 matrix guaranteed to be distinct - MatrixType m1 = MatrixType::Random(rows,cols); - MatrixType m2 = MatrixType::Random(rows,cols) + Scalar(100) * MatrixType::Identity(rows,cols); - OtherMatrixType m3 = OtherMatrixType::Random(rows,cols) + Scalar(200) * OtherMatrixType::Identity(rows,cols); - - MatrixType m1_copy = m1; - MatrixType m2_copy = m2; - OtherMatrixType m3_copy = m3; - - // test swapping 2 matrices of same type - m1.swap(m2); - VERIFY_IS_APPROX(m1,m2_copy); - VERIFY_IS_APPROX(m2,m1_copy); - m1 = m1_copy; - m2 = m2_copy; - - // test swapping 2 matrices of different types - m1.swap(m3); - VERIFY_IS_APPROX(m1,m3_copy); - VERIFY_IS_APPROX(m3,m1_copy); - m1 = m1_copy; - m3 = m3_copy; - - // test swapping matrix with expression - m1.swap(m2.block(0,0,rows,cols)); - VERIFY_IS_APPROX(m1,m2_copy); - VERIFY_IS_APPROX(m2,m1_copy); - m1 = m1_copy; - m2 = m2_copy; - - // test swapping two expressions of different types - m1.transpose().swap(m3.transpose()); - VERIFY_IS_APPROX(m1,m3_copy); - VERIFY_IS_APPROX(m3,m1_copy); - m1 = m1_copy; - m3 = m3_copy; - - // test assertion on mismatching size -- matrix case - VERIFY_RAISES_ASSERT(m1.swap(m1.row(0))); - // test assertion on mismatching size -- xpr case - VERIFY_RAISES_ASSERT(m1.row(0).swap(m1)); -} - -void test_eigen2_swap() -{ - CALL_SUBTEST_1( swap(Matrix3f()) ); // fixed size, no vectorization - CALL_SUBTEST_1( swap(Matrix4d()) ); // fixed size, possible vectorization - CALL_SUBTEST_1( swap(MatrixXd(3,3)) ); // dyn size, no vectorization - CALL_SUBTEST_1( swap(MatrixXf(30,30)) ); // dyn size, possible vectorization -} diff --git a/test/eigen2/eigen2_triangular.cpp b/test/eigen2/eigen2_triangular.cpp deleted file mode 100644 index 3748c7d71..000000000 --- a/test/eigen2/eigen2_triangular.cpp +++ /dev/null @@ -1,158 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void triangular(const MatrixType& m) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef Matrix VectorType; - - RealScalar largerEps = 10*test_precision(); - - int rows = m.rows(); - int cols = m.cols(); - - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - m4(rows, cols), - r1(rows, cols), - r2(rows, cols), - mzero = MatrixType::Zero(rows, cols), - mones = MatrixType::Ones(rows, cols), - identity = Matrix - ::Identity(rows, rows), - square = Matrix - ::Random(rows, rows); - VectorType v1 = VectorType::Random(rows), - v2 = VectorType::Random(rows), - vzero = VectorType::Zero(rows); - - MatrixType m1up = m1.template part(); - MatrixType m2up = m2.template part(); - - if (rows*cols>1) - { - VERIFY(m1up.isUpperTriangular()); - VERIFY(m2up.transpose().isLowerTriangular()); - VERIFY(!m2.isLowerTriangular()); - } - -// VERIFY_IS_APPROX(m1up.transpose() * m2, m1.upper().transpose().lower() * m2); - - // test overloaded operator+= - r1.setZero(); - r2.setZero(); - r1.template part() += m1; - r2 += m1up; - VERIFY_IS_APPROX(r1,r2); - - // test overloaded operator= - m1.setZero(); - m1.template part() = (m2.transpose() * m2).lazy(); - m3 = m2.transpose() * m2; - VERIFY_IS_APPROX(m3.template part().transpose(), m1); - - // test overloaded operator= - m1.setZero(); - m1.template part() = (m2.transpose() * m2).lazy(); - VERIFY_IS_APPROX(m3.template part(), m1); - - VERIFY_IS_APPROX(m3.template part(), m3.diagonal().asDiagonal()); - - m1 = MatrixType::Random(rows, cols); - for (int i=0; i(); - - Transpose trm4(m4); - // test back and forward subsitution - m3 = m1.template part(); - VERIFY(m3.template marked().solveTriangular(m3).cwise().abs().isIdentity(test_precision())); - VERIFY(m3.transpose().template marked() - .solveTriangular(m3.transpose()).cwise().abs().isIdentity(test_precision())); - // check M * inv(L) using in place API - m4 = m3; - m3.transpose().template marked().solveTriangularInPlace(trm4); - VERIFY(m4.cwise().abs().isIdentity(test_precision())); - - m3 = m1.template part(); - VERIFY(m3.template marked().solveTriangular(m3).cwise().abs().isIdentity(test_precision())); - VERIFY(m3.transpose().template marked() - .solveTriangular(m3.transpose()).cwise().abs().isIdentity(test_precision())); - // check M * inv(U) using in place API - m4 = m3; - m3.transpose().template marked().solveTriangularInPlace(trm4); - VERIFY(m4.cwise().abs().isIdentity(test_precision())); - - m3 = m1.template part(); - VERIFY(m2.isApprox(m3 * (m3.template marked().solveTriangular(m2)), largerEps)); - m3 = m1.template part(); - VERIFY(m2.isApprox(m3 * (m3.template marked().solveTriangular(m2)), largerEps)); - - VERIFY((m1.template part() * m2.template part()).isUpperTriangular()); - - // test swap - m1.setOnes(); - m2.setZero(); - m2.template part().swap(m1); - m3.setZero(); - m3.template part().setOnes(); - VERIFY_IS_APPROX(m2,m3); - -} - -void selfadjoint() -{ - Matrix2i m; - m << 1, 2, - 3, 4; - - Matrix2i m1 = Matrix2i::Zero(); - m1.part() = m; - Matrix2i ref1; - ref1 << 1, 2, - 2, 4; - VERIFY(m1 == ref1); - - Matrix2i m2 = Matrix2i::Zero(); - m2.part() = m.part(); - Matrix2i ref2; - ref2 << 1, 2, - 2, 4; - VERIFY(m2 == ref2); - - Matrix2i m3 = Matrix2i::Zero(); - m3.part() = m.part(); - Matrix2i ref3; - ref3 << 1, 0, - 0, 4; - VERIFY(m3 == ref3); - - // example inspired from bug 159 - int array[] = {1, 2, 3, 4}; - Matrix2i::Map(array).part() = Matrix2i::Random().part(); - - std::cout << "hello\n" << array << std::endl; -} - -void test_eigen2_triangular() -{ - CALL_SUBTEST_8( selfadjoint() ); - for(int i = 0; i < g_repeat ; i++) { - CALL_SUBTEST_1( triangular(Matrix()) ); - CALL_SUBTEST_2( triangular(Matrix()) ); - CALL_SUBTEST_3( triangular(Matrix3d()) ); - CALL_SUBTEST_4( triangular(MatrixXcf(4, 4)) ); - CALL_SUBTEST_5( triangular(Matrix,8, 8>()) ); - CALL_SUBTEST_6( triangular(MatrixXd(17,17)) ); - CALL_SUBTEST_7( triangular(Matrix(5, 5)) ); - } -} diff --git a/test/eigen2/eigen2_unalignedassert.cpp b/test/eigen2/eigen2_unalignedassert.cpp deleted file mode 100644 index d10b6f538..000000000 --- a/test/eigen2/eigen2_unalignedassert.cpp +++ /dev/null @@ -1,116 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -struct Good1 -{ - MatrixXd m; // good: m will allocate its own array, taking care of alignment. - Good1() : m(20,20) {} -}; - -struct Good2 -{ - Matrix3d m; // good: m's size isn't a multiple of 16 bytes, so m doesn't have to be aligned -}; - -struct Good3 -{ - Vector2f m; // good: same reason -}; - -struct Bad4 -{ - Vector2d m; // bad: sizeof(m)%16==0 so alignment is required -}; - -struct Bad5 -{ - Matrix m; // bad: same reason -}; - -struct Bad6 -{ - Matrix m; // bad: same reason -}; - -struct Good7 -{ - EIGEN_MAKE_ALIGNED_OPERATOR_NEW - Vector2d m; - float f; // make the struct have sizeof%16!=0 to make it a little more tricky when we allow an array of 2 such objects -}; - -struct Good8 -{ - EIGEN_MAKE_ALIGNED_OPERATOR_NEW - float f; // try the f at first -- the EIGEN_ALIGN_128 attribute of m should make that still work - Matrix4f m; -}; - -struct Good9 -{ - Matrix m; // good: no alignment requested - float f; -}; - -template struct Depends -{ - EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(Align) - Vector2d m; - float f; -}; - -template -void check_unalignedassert_good() -{ - T *x, *y; - x = new T; - delete x; - y = new T[2]; - delete[] y; -} - -#if EIGEN_ARCH_WANTS_ALIGNMENT -template -void check_unalignedassert_bad() -{ - float buf[sizeof(T)+16]; - float *unaligned = buf; - while((reinterpret_cast(unaligned)&0xf)==0) ++unaligned; // make sure unaligned is really unaligned - T *x = ::new(static_cast(unaligned)) T; - x->~T(); -} -#endif - -void unalignedassert() -{ - check_unalignedassert_good(); - check_unalignedassert_good(); - check_unalignedassert_good(); -#if EIGEN_ARCH_WANTS_ALIGNMENT - VERIFY_RAISES_ASSERT(check_unalignedassert_bad()); - VERIFY_RAISES_ASSERT(check_unalignedassert_bad()); - VERIFY_RAISES_ASSERT(check_unalignedassert_bad()); -#endif - - check_unalignedassert_good(); - check_unalignedassert_good(); - check_unalignedassert_good(); - check_unalignedassert_good >(); - -#if EIGEN_ARCH_WANTS_ALIGNMENT - VERIFY_RAISES_ASSERT(check_unalignedassert_bad >()); -#endif -} - -void test_eigen2_unalignedassert() -{ - CALL_SUBTEST(unalignedassert()); -} diff --git a/test/eigen2/eigen2_visitor.cpp b/test/eigen2/eigen2_visitor.cpp deleted file mode 100644 index 4781991de..000000000 --- a/test/eigen2/eigen2_visitor.cpp +++ /dev/null @@ -1,116 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" - -template void matrixVisitor(const MatrixType& p) -{ - typedef typename MatrixType::Scalar Scalar; - - int rows = p.rows(); - int cols = p.cols(); - - // construct a random matrix where all coefficients are different - MatrixType m; - m = MatrixType::Random(rows, cols); - for(int i = 0; i < m.size(); i++) - for(int i2 = 0; i2 < i; i2++) - while(m(i) == m(i2)) // yes, == - m(i) = ei_random(); - - Scalar minc = Scalar(1000), maxc = Scalar(-1000); - int minrow=0,mincol=0,maxrow=0,maxcol=0; - for(int j = 0; j < cols; j++) - for(int i = 0; i < rows; i++) - { - if(m(i,j) < minc) - { - minc = m(i,j); - minrow = i; - mincol = j; - } - if(m(i,j) > maxc) - { - maxc = m(i,j); - maxrow = i; - maxcol = j; - } - } - int eigen_minrow, eigen_mincol, eigen_maxrow, eigen_maxcol; - Scalar eigen_minc, eigen_maxc; - eigen_minc = m.minCoeff(&eigen_minrow,&eigen_mincol); - eigen_maxc = m.maxCoeff(&eigen_maxrow,&eigen_maxcol); - VERIFY(minrow == eigen_minrow); - VERIFY(maxrow == eigen_maxrow); - VERIFY(mincol == eigen_mincol); - VERIFY(maxcol == eigen_maxcol); - VERIFY_IS_APPROX(minc, eigen_minc); - VERIFY_IS_APPROX(maxc, eigen_maxc); - VERIFY_IS_APPROX(minc, m.minCoeff()); - VERIFY_IS_APPROX(maxc, m.maxCoeff()); -} - -template void vectorVisitor(const VectorType& w) -{ - typedef typename VectorType::Scalar Scalar; - - int size = w.size(); - - // construct a random vector where all coefficients are different - VectorType v; - v = VectorType::Random(size); - for(int i = 0; i < size; i++) - for(int i2 = 0; i2 < i; i2++) - while(v(i) == v(i2)) // yes, == - v(i) = ei_random(); - - Scalar minc = Scalar(1000), maxc = Scalar(-1000); - int minidx=0,maxidx=0; - for(int i = 0; i < size; i++) - { - if(v(i) < minc) - { - minc = v(i); - minidx = i; - } - if(v(i) > maxc) - { - maxc = v(i); - maxidx = i; - } - } - int eigen_minidx, eigen_maxidx; - Scalar eigen_minc, eigen_maxc; - eigen_minc = v.minCoeff(&eigen_minidx); - eigen_maxc = v.maxCoeff(&eigen_maxidx); - VERIFY(minidx == eigen_minidx); - VERIFY(maxidx == eigen_maxidx); - VERIFY_IS_APPROX(minc, eigen_minc); - VERIFY_IS_APPROX(maxc, eigen_maxc); - VERIFY_IS_APPROX(minc, v.minCoeff()); - VERIFY_IS_APPROX(maxc, v.maxCoeff()); -} - -void test_eigen2_visitor() -{ - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_1( matrixVisitor(Matrix()) ); - CALL_SUBTEST_2( matrixVisitor(Matrix2f()) ); - CALL_SUBTEST_3( matrixVisitor(Matrix4d()) ); - CALL_SUBTEST_4( matrixVisitor(MatrixXd(8, 12)) ); - CALL_SUBTEST_5( matrixVisitor(Matrix(20, 20)) ); - CALL_SUBTEST_6( matrixVisitor(MatrixXi(8, 12)) ); - } - for(int i = 0; i < g_repeat; i++) { - CALL_SUBTEST_7( vectorVisitor(Vector4f()) ); - CALL_SUBTEST_4( vectorVisitor(VectorXd(10)) ); - CALL_SUBTEST_4( vectorVisitor(RowVectorXd(10)) ); - CALL_SUBTEST_8( vectorVisitor(VectorXf(33)) ); - } -} diff --git a/test/eigen2/gsl_helper.h b/test/eigen2/gsl_helper.h deleted file mode 100644 index d1d854533..000000000 --- a/test/eigen2/gsl_helper.h +++ /dev/null @@ -1,175 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_GSL_HELPER -#define EIGEN_GSL_HELPER - -#include - -#include -#include -#include -#include -#include -#include - -namespace Eigen { - -template::IsComplex> struct GslTraits -{ - typedef gsl_matrix* Matrix; - typedef gsl_vector* Vector; - static Matrix createMatrix(int rows, int cols) { return gsl_matrix_alloc(rows,cols); } - static Vector createVector(int size) { return gsl_vector_alloc(size); } - static void free(Matrix& m) { gsl_matrix_free(m); m=0; } - static void free(Vector& m) { gsl_vector_free(m); m=0; } - static void prod(const Matrix& m, const Vector& v, Vector& x) { gsl_blas_dgemv(CblasNoTrans,1,m,v,0,x); } - static void cholesky(Matrix& m) { gsl_linalg_cholesky_decomp(m); } - static void cholesky_solve(const Matrix& m, const Vector& b, Vector& x) { gsl_linalg_cholesky_solve(m,b,x); } - static void eigen_symm(const Matrix& m, Vector& eval, Matrix& evec) - { - gsl_eigen_symmv_workspace * w = gsl_eigen_symmv_alloc(m->size1); - Matrix a = createMatrix(m->size1, m->size2); - gsl_matrix_memcpy(a, m); - gsl_eigen_symmv(a,eval,evec,w); - gsl_eigen_symmv_sort(eval, evec, GSL_EIGEN_SORT_VAL_ASC); - gsl_eigen_symmv_free(w); - free(a); - } - static void eigen_symm_gen(const Matrix& m, const Matrix& _b, Vector& eval, Matrix& evec) - { - gsl_eigen_gensymmv_workspace * w = gsl_eigen_gensymmv_alloc(m->size1); - Matrix a = createMatrix(m->size1, m->size2); - Matrix b = createMatrix(_b->size1, _b->size2); - gsl_matrix_memcpy(a, m); - gsl_matrix_memcpy(b, _b); - gsl_eigen_gensymmv(a,b,eval,evec,w); - gsl_eigen_symmv_sort(eval, evec, GSL_EIGEN_SORT_VAL_ASC); - gsl_eigen_gensymmv_free(w); - free(a); - } -}; - -template struct GslTraits -{ - typedef gsl_matrix_complex* Matrix; - typedef gsl_vector_complex* Vector; - static Matrix createMatrix(int rows, int cols) { return gsl_matrix_complex_alloc(rows,cols); } - static Vector createVector(int size) { return gsl_vector_complex_alloc(size); } - static void free(Matrix& m) { gsl_matrix_complex_free(m); m=0; } - static void free(Vector& m) { gsl_vector_complex_free(m); m=0; } - static void cholesky(Matrix& m) { gsl_linalg_complex_cholesky_decomp(m); } - static void cholesky_solve(const Matrix& m, const Vector& b, Vector& x) { gsl_linalg_complex_cholesky_solve(m,b,x); } - static void prod(const Matrix& m, const Vector& v, Vector& x) - { gsl_blas_zgemv(CblasNoTrans,gsl_complex_rect(1,0),m,v,gsl_complex_rect(0,0),x); } - static void eigen_symm(const Matrix& m, gsl_vector* &eval, Matrix& evec) - { - gsl_eigen_hermv_workspace * w = gsl_eigen_hermv_alloc(m->size1); - Matrix a = createMatrix(m->size1, m->size2); - gsl_matrix_complex_memcpy(a, m); - gsl_eigen_hermv(a,eval,evec,w); - gsl_eigen_hermv_sort(eval, evec, GSL_EIGEN_SORT_VAL_ASC); - gsl_eigen_hermv_free(w); - free(a); - } - static void eigen_symm_gen(const Matrix& m, const Matrix& _b, gsl_vector* &eval, Matrix& evec) - { - gsl_eigen_genhermv_workspace * w = gsl_eigen_genhermv_alloc(m->size1); - Matrix a = createMatrix(m->size1, m->size2); - Matrix b = createMatrix(_b->size1, _b->size2); - gsl_matrix_complex_memcpy(a, m); - gsl_matrix_complex_memcpy(b, _b); - gsl_eigen_genhermv(a,b,eval,evec,w); - gsl_eigen_hermv_sort(eval, evec, GSL_EIGEN_SORT_VAL_ASC); - gsl_eigen_genhermv_free(w); - free(a); - } -}; - -template -void convert(const MatrixType& m, gsl_matrix* &res) -{ -// if (res) -// gsl_matrix_free(res); - res = gsl_matrix_alloc(m.rows(), m.cols()); - for (int i=0 ; i -void convert(const gsl_matrix* m, MatrixType& res) -{ - res.resize(int(m->size1), int(m->size2)); - for (int i=0 ; i -void convert(const VectorType& m, gsl_vector* &res) -{ - if (res) gsl_vector_free(res); - res = gsl_vector_alloc(m.size()); - for (int i=0 ; i -void convert(const gsl_vector* m, VectorType& res) -{ - res.resize (m->size); - for (int i=0 ; i -void convert(const MatrixType& m, gsl_matrix_complex* &res) -{ - res = gsl_matrix_complex_alloc(m.rows(), m.cols()); - for (int i=0 ; i -void convert(const gsl_matrix_complex* m, MatrixType& res) -{ - res.resize(int(m->size1), int(m->size2)); - for (int i=0 ; i -void convert(const VectorType& m, gsl_vector_complex* &res) -{ - res = gsl_vector_complex_alloc(m.size()); - for (int i=0 ; i -void convert(const gsl_vector_complex* m, VectorType& res) -{ - res.resize(m->size); - for (int i=0 ; i -// Copyright (C) 2008 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include -#include -#include -#include -#include - -#ifndef EIGEN_TEST_FUNC -#error EIGEN_TEST_FUNC must be defined -#endif - -#define DEFAULT_REPEAT 10 - -namespace Eigen -{ - static std::vector g_test_stack; - static int g_repeat; -} - -#define EI_PP_MAKE_STRING2(S) #S -#define EI_PP_MAKE_STRING(S) EI_PP_MAKE_STRING2(S) - -#define EI_PP_CAT2(a,b) a ## b -#define EI_PP_CAT(a,b) EI_PP_CAT2(a,b) - -#ifndef EIGEN_NO_ASSERTION_CHECKING - - namespace Eigen - { - static const bool should_raise_an_assert = false; - - // Used to avoid to raise two exceptions at a time in which - // case the exception is not properly caught. - // This may happen when a second exceptions is raise in a destructor. - static bool no_more_assert = false; - - struct eigen_assert_exception - { - eigen_assert_exception(void) {} - ~eigen_assert_exception() { Eigen::no_more_assert = false; } - }; - } - - // If EIGEN_DEBUG_ASSERTS is defined and if no assertion is raised while - // one should have been, then the list of excecuted assertions is printed out. - // - // EIGEN_DEBUG_ASSERTS is not enabled by default as it - // significantly increases the compilation time - // and might even introduce side effects that would hide - // some memory errors. - #ifdef EIGEN_DEBUG_ASSERTS - - namespace Eigen - { - static bool ei_push_assert = false; - static std::vector eigen_assert_list; - } - - #define eigen_assert(a) \ - if( (!(a)) && (!no_more_assert) ) \ - { \ - Eigen::no_more_assert = true; \ - throw Eigen::eigen_assert_exception(); \ - } \ - else if (Eigen::ei_push_assert) \ - { \ - eigen_assert_list.push_back(std::string(EI_PP_MAKE_STRING(__FILE__)" ("EI_PP_MAKE_STRING(__LINE__)") : "#a) ); \ - } - - #define VERIFY_RAISES_ASSERT(a) \ - { \ - Eigen::no_more_assert = false; \ - try { \ - Eigen::eigen_assert_list.clear(); \ - Eigen::ei_push_assert = true; \ - a; \ - Eigen::ei_push_assert = false; \ - std::cerr << "One of the following asserts should have been raised:\n"; \ - for (uint ai=0 ; ai - - -#define VERIFY(a) do { if (!(a)) { \ - std::cerr << "Test " << g_test_stack.back() << " failed in "EI_PP_MAKE_STRING(__FILE__) << " (" << EI_PP_MAKE_STRING(__LINE__) << ")" \ - << std::endl << " " << EI_PP_MAKE_STRING(a) << std::endl << std::endl; \ - abort(); \ - } } while (0) - -#define VERIFY_IS_APPROX(a, b) VERIFY(test_ei_isApprox(a, b)) -#define VERIFY_IS_NOT_APPROX(a, b) VERIFY(!test_ei_isApprox(a, b)) -#define VERIFY_IS_MUCH_SMALLER_THAN(a, b) VERIFY(test_ei_isMuchSmallerThan(a, b)) -#define VERIFY_IS_NOT_MUCH_SMALLER_THAN(a, b) VERIFY(!test_ei_isMuchSmallerThan(a, b)) -#define VERIFY_IS_APPROX_OR_LESS_THAN(a, b) VERIFY(test_ei_isApproxOrLessThan(a, b)) -#define VERIFY_IS_NOT_APPROX_OR_LESS_THAN(a, b) VERIFY(!test_ei_isApproxOrLessThan(a, b)) - -#define CALL_SUBTEST(FUNC) do { \ - g_test_stack.push_back(EI_PP_MAKE_STRING(FUNC)); \ - FUNC; \ - g_test_stack.pop_back(); \ - } while (0) - -namespace Eigen { - -template inline typename NumTraits::Real test_precision(); -template<> inline int test_precision() { return 0; } -template<> inline float test_precision() { return 1e-3f; } -template<> inline double test_precision() { return 1e-6; } -template<> inline float test_precision >() { return test_precision(); } -template<> inline double test_precision >() { return test_precision(); } -template<> inline long double test_precision() { return 1e-6; } - -inline bool test_ei_isApprox(const int& a, const int& b) -{ return ei_isApprox(a, b, test_precision()); } -inline bool test_ei_isMuchSmallerThan(const int& a, const int& b) -{ return ei_isMuchSmallerThan(a, b, test_precision()); } -inline bool test_ei_isApproxOrLessThan(const int& a, const int& b) -{ return ei_isApproxOrLessThan(a, b, test_precision()); } - -inline bool test_ei_isApprox(const float& a, const float& b) -{ return ei_isApprox(a, b, test_precision()); } -inline bool test_ei_isMuchSmallerThan(const float& a, const float& b) -{ return ei_isMuchSmallerThan(a, b, test_precision()); } -inline bool test_ei_isApproxOrLessThan(const float& a, const float& b) -{ return ei_isApproxOrLessThan(a, b, test_precision()); } - -inline bool test_ei_isApprox(const double& a, const double& b) -{ return ei_isApprox(a, b, test_precision()); } -inline bool test_ei_isMuchSmallerThan(const double& a, const double& b) -{ return ei_isMuchSmallerThan(a, b, test_precision()); } -inline bool test_ei_isApproxOrLessThan(const double& a, const double& b) -{ return ei_isApproxOrLessThan(a, b, test_precision()); } - -inline bool test_ei_isApprox(const std::complex& a, const std::complex& b) -{ return ei_isApprox(a, b, test_precision >()); } -inline bool test_ei_isMuchSmallerThan(const std::complex& a, const std::complex& b) -{ return ei_isMuchSmallerThan(a, b, test_precision >()); } - -inline bool test_ei_isApprox(const std::complex& a, const std::complex& b) -{ return ei_isApprox(a, b, test_precision >()); } -inline bool test_ei_isMuchSmallerThan(const std::complex& a, const std::complex& b) -{ return ei_isMuchSmallerThan(a, b, test_precision >()); } - -inline bool test_ei_isApprox(const long double& a, const long double& b) -{ return ei_isApprox(a, b, test_precision()); } -inline bool test_ei_isMuchSmallerThan(const long double& a, const long double& b) -{ return ei_isMuchSmallerThan(a, b, test_precision()); } -inline bool test_ei_isApproxOrLessThan(const long double& a, const long double& b) -{ return ei_isApproxOrLessThan(a, b, test_precision()); } - -template -inline bool test_ei_isApprox(const Type1& a, const Type2& b) -{ - return a.isApprox(b, test_precision()); -} - -template -inline bool test_ei_isMuchSmallerThan(const MatrixBase& m1, - const MatrixBase& m2) -{ - return m1.isMuchSmallerThan(m2, test_precision::Scalar>()); -} - -template -inline bool test_ei_isMuchSmallerThan(const MatrixBase& m, - const typename NumTraits::Scalar>::Real& s) -{ - return m.isMuchSmallerThan(s, test_precision::Scalar>()); -} - -} // end namespace Eigen - -template struct GetDifferentType; - -template<> struct GetDifferentType { typedef double type; }; -template<> struct GetDifferentType { typedef float type; }; -template struct GetDifferentType > -{ typedef std::complex::type> type; }; - -// forward declaration of the main test function -void EI_PP_CAT(test_,EIGEN_TEST_FUNC)(); - -using namespace Eigen; - -#ifdef EIGEN_TEST_PART_1 -#define CALL_SUBTEST_1(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_1(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_2 -#define CALL_SUBTEST_2(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_2(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_3 -#define CALL_SUBTEST_3(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_3(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_4 -#define CALL_SUBTEST_4(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_4(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_5 -#define CALL_SUBTEST_5(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_5(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_6 -#define CALL_SUBTEST_6(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_6(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_7 -#define CALL_SUBTEST_7(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_7(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_8 -#define CALL_SUBTEST_8(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_8(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_9 -#define CALL_SUBTEST_9(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_9(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_10 -#define CALL_SUBTEST_10(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_10(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_11 -#define CALL_SUBTEST_11(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_11(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_12 -#define CALL_SUBTEST_12(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_12(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_13 -#define CALL_SUBTEST_13(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_13(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_14 -#define CALL_SUBTEST_14(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_14(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_15 -#define CALL_SUBTEST_15(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_15(FUNC) -#endif - -#ifdef EIGEN_TEST_PART_16 -#define CALL_SUBTEST_16(FUNC) CALL_SUBTEST(FUNC) -#else -#define CALL_SUBTEST_16(FUNC) -#endif - - - -int main(int argc, char *argv[]) -{ - bool has_set_repeat = false; - bool has_set_seed = false; - bool need_help = false; - unsigned int seed = 0; - int repeat = DEFAULT_REPEAT; - - for(int i = 1; i < argc; i++) - { - if(argv[i][0] == 'r') - { - if(has_set_repeat) - { - std::cout << "Argument " << argv[i] << " conflicting with a former argument" << std::endl; - return 1; - } - repeat = std::atoi(argv[i]+1); - has_set_repeat = true; - if(repeat <= 0) - { - std::cout << "Invalid \'repeat\' value " << argv[i]+1 << std::endl; - return 1; - } - } - else if(argv[i][0] == 's') - { - if(has_set_seed) - { - std::cout << "Argument " << argv[i] << " conflicting with a former argument" << std::endl; - return 1; - } - seed = int(std::strtoul(argv[i]+1, 0, 10)); - has_set_seed = true; - bool ok = seed!=0; - if(!ok) - { - std::cout << "Invalid \'seed\' value " << argv[i]+1 << std::endl; - return 1; - } - } - else - { - need_help = true; - } - } - - if(need_help) - { - std::cout << "This test application takes the following optional arguments:" << std::endl; - std::cout << " rN Repeat each test N times (default: " << DEFAULT_REPEAT << ")" << std::endl; - std::cout << " sN Use N as seed for random numbers (default: based on current time)" << std::endl; - return 1; - } - - if(!has_set_seed) seed = (unsigned int) std::time(NULL); - if(!has_set_repeat) repeat = DEFAULT_REPEAT; - - std::cout << "Initializing random number generator with seed " << seed << std::endl; - std::srand(seed); - std::cout << "Repeating each test " << repeat << " times" << std::endl; - - Eigen::g_repeat = repeat; - Eigen::g_test_stack.push_back(EI_PP_MAKE_STRING(EIGEN_TEST_FUNC)); - - EI_PP_CAT(test_,EIGEN_TEST_FUNC)(); - return 0; -} - - - diff --git a/test/eigen2/product.h b/test/eigen2/product.h deleted file mode 100644 index 2c9604d9a..000000000 --- a/test/eigen2/product.h +++ /dev/null @@ -1,132 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2006-2008 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "main.h" -#include -#include - -template -bool areNotApprox(const MatrixBase& m1, const MatrixBase& m2, typename Derived1::RealScalar epsilon = precision()) -{ - return !((m1-m2).cwise().abs2().maxCoeff() < epsilon * epsilon - * std::max(m1.cwise().abs2().maxCoeff(), m2.cwise().abs2().maxCoeff())); -} - -template void product(const MatrixType& m) -{ - /* this test covers the following files: - Identity.h Product.h - */ - - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::FloatingPoint FloatingPoint; - typedef Matrix RowVectorType; - typedef Matrix ColVectorType; - typedef Matrix RowSquareMatrixType; - typedef Matrix ColSquareMatrixType; - typedef Matrix OtherMajorMatrixType; - - int rows = m.rows(); - int cols = m.cols(); - - // this test relies a lot on Random.h, and there's not much more that we can do - // to test it, hence I consider that we will have tested Random.h - MatrixType m1 = MatrixType::Random(rows, cols), - m2 = MatrixType::Random(rows, cols), - m3(rows, cols), - mzero = MatrixType::Zero(rows, cols); - RowSquareMatrixType - identity = RowSquareMatrixType::Identity(rows, rows), - square = RowSquareMatrixType::Random(rows, rows), - res = RowSquareMatrixType::Random(rows, rows); - ColSquareMatrixType - square2 = ColSquareMatrixType::Random(cols, cols), - res2 = ColSquareMatrixType::Random(cols, cols); - RowVectorType v1 = RowVectorType::Random(rows), - v2 = RowVectorType::Random(rows), - vzero = RowVectorType::Zero(rows); - ColVectorType vc2 = ColVectorType::Random(cols), vcres(cols); - OtherMajorMatrixType tm1 = m1; - - Scalar s1 = ei_random(); - - int r = ei_random(0, rows-1), - c = ei_random(0, cols-1); - - // begin testing Product.h: only associativity for now - // (we use Transpose.h but this doesn't count as a test for it) - - VERIFY_IS_APPROX((m1*m1.transpose())*m2, m1*(m1.transpose()*m2)); - m3 = m1; - m3 *= m1.transpose() * m2; - VERIFY_IS_APPROX(m3, m1 * (m1.transpose()*m2)); - VERIFY_IS_APPROX(m3, m1.lazy() * (m1.transpose()*m2)); - - // continue testing Product.h: distributivity - VERIFY_IS_APPROX(square*(m1 + m2), square*m1+square*m2); - VERIFY_IS_APPROX(square*(m1 - m2), square*m1-square*m2); - - // continue testing Product.h: compatibility with ScalarMultiple.h - VERIFY_IS_APPROX(s1*(square*m1), (s1*square)*m1); - VERIFY_IS_APPROX(s1*(square*m1), square*(m1*s1)); - - // again, test operator() to check const-qualification - s1 += (square.lazy() * m1)(r,c); - - // test Product.h together with Identity.h - VERIFY_IS_APPROX(v1, identity*v1); - VERIFY_IS_APPROX(v1.transpose(), v1.transpose() * identity); - // again, test operator() to check const-qualification - VERIFY_IS_APPROX(MatrixType::Identity(rows, cols)(r,c), static_cast(r==c)); - - if (rows!=cols) - VERIFY_RAISES_ASSERT(m3 = m1*m1); - - // test the previous tests were not screwed up because operator* returns 0 - // (we use the more accurate default epsilon) - if (NumTraits::HasFloatingPoint && std::min(rows,cols)>1) - { - VERIFY(areNotApprox(m1.transpose()*m2,m2.transpose()*m1)); - } - - // test optimized operator+= path - res = square; - res += (m1 * m2.transpose()).lazy(); - VERIFY_IS_APPROX(res, square + m1 * m2.transpose()); - if (NumTraits::HasFloatingPoint && std::min(rows,cols)>1) - { - VERIFY(areNotApprox(res,square + m2 * m1.transpose())); - } - vcres = vc2; - vcres += (m1.transpose() * v1).lazy(); - VERIFY_IS_APPROX(vcres, vc2 + m1.transpose() * v1); - tm1 = m1; - VERIFY_IS_APPROX(tm1.transpose() * v1, m1.transpose() * v1); - VERIFY_IS_APPROX(v1.transpose() * tm1, v1.transpose() * m1); - - // test submatrix and matrix/vector product - for (int i=0; i::HasFloatingPoint && std::min(rows,cols)>1) - { - VERIFY(areNotApprox(res2,square2 + m2.transpose() * m1)); - } -} - diff --git a/test/eigen2/sparse.h b/test/eigen2/sparse.h deleted file mode 100644 index e12f89990..000000000 --- a/test/eigen2/sparse.h +++ /dev/null @@ -1,154 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. Eigen itself is part of the KDE project. -// -// Copyright (C) 2008 Daniel Gomez Ferro -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_TESTSPARSE_H - -#include "main.h" - -#if EIGEN_GNUC_AT_LEAST(4,0) && !defined __ICC -#include -#define EIGEN_UNORDERED_MAP_SUPPORT -namespace std { - using std::tr1::unordered_map; -} -#endif - -#ifdef EIGEN_GOOGLEHASH_SUPPORT - #include -#endif - -#include -#include -#include - -enum { - ForceNonZeroDiag = 1, - MakeLowerTriangular = 2, - MakeUpperTriangular = 4, - ForceRealDiag = 8 -}; - -/* Initializes both a sparse and dense matrix with same random values, - * and a ratio of \a density non zero entries. - * \param flags is a union of ForceNonZeroDiag, MakeLowerTriangular and MakeUpperTriangular - * allowing to control the shape of the matrix. - * \param zeroCoords and nonzeroCoords allows to get the coordinate lists of the non zero, - * and zero coefficients respectively. - */ -template void -initSparse(double density, - Matrix& refMat, - SparseMatrix& sparseMat, - int flags = 0, - std::vector* zeroCoords = 0, - std::vector* nonzeroCoords = 0) -{ - sparseMat.startFill(int(refMat.rows()*refMat.cols()*density)); - for(int j=0; j(0,1) < density) ? ei_random() : Scalar(0); - if ((flags&ForceNonZeroDiag) && (i==j)) - { - v = ei_random()*Scalar(3.); - v = v*v + Scalar(5.); - } - if ((flags & MakeLowerTriangular) && j>i) - v = Scalar(0); - else if ((flags & MakeUpperTriangular) && jpush_back(Vector2i(i,j)); - } - else if (zeroCoords) - { - zeroCoords->push_back(Vector2i(i,j)); - } - refMat(i,j) = v; - } - } - sparseMat.endFill(); -} - -template void -initSparse(double density, - Matrix& refMat, - DynamicSparseMatrix& sparseMat, - int flags = 0, - std::vector* zeroCoords = 0, - std::vector* nonzeroCoords = 0) -{ - sparseMat.startFill(int(refMat.rows()*refMat.cols()*density)); - for(int j=0; j(0,1) < density) ? ei_random() : Scalar(0); - if ((flags&ForceNonZeroDiag) && (i==j)) - { - v = ei_random()*Scalar(3.); - v = v*v + Scalar(5.); - } - if ((flags & MakeLowerTriangular) && j>i) - v = Scalar(0); - else if ((flags & MakeUpperTriangular) && jpush_back(Vector2i(i,j)); - } - else if (zeroCoords) - { - zeroCoords->push_back(Vector2i(i,j)); - } - refMat(i,j) = v; - } - } - sparseMat.endFill(); -} - -template void -initSparse(double density, - Matrix& refVec, - SparseVector& sparseVec, - std::vector* zeroCoords = 0, - std::vector* nonzeroCoords = 0) -{ - sparseVec.reserve(int(refVec.size()*density)); - sparseVec.setZero(); - for(int i=0; i(0,1) < density) ? ei_random() : Scalar(0); - if (v!=Scalar(0)) - { - sparseVec.fill(i) = v; - if (nonzeroCoords) - nonzeroCoords->push_back(i); - } - else if (zeroCoords) - zeroCoords->push_back(i); - refVec[i] = v; - } -} - -#endif // EIGEN_TESTSPARSE_H diff --git a/unsupported/Eigen/SVD b/unsupported/Eigen/SVD index 7cc059280..900a8aa60 100644 --- a/unsupported/Eigen/SVD +++ b/unsupported/Eigen/SVD @@ -29,10 +29,6 @@ #include "../../Eigen/src/SVD/JacobiSVD_MKL.h" #endif -#ifdef EIGEN2_SUPPORT -#include "../../Eigen/src/Eigen2Support/SVD.h" -#endif - #include "../../Eigen/src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_SVD_MODULE_H From 6f846ef9c6ef3e838d1361a30dc11f65ddff967d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 17:51:53 +0200 Subject: [PATCH 0573/1103] Split StorageKind promotion into two helpers: one for products, and one for coefficient-wise operations. --- Eigen/src/Core/CwiseBinaryOp.h | 15 +++--- Eigen/src/Core/PermutationMatrix.h | 16 ------ Eigen/src/Core/Product.h | 15 +++--- Eigen/src/Core/ProductBase.h | 5 +- Eigen/src/Core/products/CoeffBasedProduct.h | 5 +- Eigen/src/Core/util/Constants.h | 4 -- Eigen/src/Core/util/XprHelper.h | 58 +++++++++++++++++++-- Eigen/src/SparseCore/SparseCwiseBinaryOp.h | 6 --- 8 files changed, 77 insertions(+), 47 deletions(-) diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index 355f3bdc8..b4662907e 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -56,8 +56,9 @@ struct traits > typename Rhs::Scalar ) >::type Scalar; - typedef typename promote_storage_type::StorageKind, - typename traits::StorageKind>::ret StorageKind; + typedef typename cwise_promote_storage_type::StorageKind, + typename traits::StorageKind, + BinaryOp>::ret StorageKind; typedef typename promote_index_type::Index, typename traits::Index>::type Index; typedef typename Lhs::Nested LhsNested; @@ -98,8 +99,9 @@ template class CwiseBinaryOp : internal::no_assignment_operator, public CwiseBinaryOpImpl< BinaryOp, LhsType, RhsType, - typename internal::promote_storage_type::StorageKind, - typename internal::traits::StorageKind>::ret> + typename internal::cwise_promote_storage_type::StorageKind, + typename internal::traits::StorageKind, + BinaryOp>::ret> { public: @@ -108,8 +110,9 @@ class CwiseBinaryOp : internal::no_assignment_operator, typedef typename CwiseBinaryOpImpl< BinaryOp, LhsType, RhsType, - typename internal::promote_storage_type::StorageKind, - typename internal::traits::StorageKind>::ret>::Base Base; + typename internal::cwise_promote_storage_type::StorageKind, + typename internal::traits::StorageKind, + BinaryOp>::ret>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseBinaryOp) typedef typename internal::nested::type LhsNested; diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 4399e1d64..658473397 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -465,22 +465,6 @@ class Map, struct PermutationStorage {}; -#ifdef EIGEN_TEST_EVALUATORS - -// storage type of product of permutation wrapper with dense - -namespace internal { - -template<> struct promote_storage_type -{ typedef Dense ret; }; - -template<> struct promote_storage_type -{ typedef Dense ret; }; - -} // end namespace internal - -#endif // EIGEN_TEST_EVALUATORS - template class TranspositionsWrapper; namespace internal { template diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index e381ac46a..cb4d4c924 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -62,8 +62,9 @@ struct traits > typedef MatrixXpr XprKind; typedef typename product_result_scalar::Scalar Scalar; - typedef typename promote_storage_type::StorageKind, - typename traits::StorageKind>::ret StorageKind; + typedef typename product_promote_storage_type::StorageKind, + typename traits::StorageKind, + internal::product_type::ret>::ret StorageKind; typedef typename promote_index_type::Index, typename traits::Index>::type Index; @@ -94,8 +95,9 @@ struct traits > template class Product : public ProductImpl<_Lhs,_Rhs,Option, - typename internal::promote_storage_type::StorageKind, - typename internal::traits<_Rhs>::StorageKind>::ret> + typename internal::product_promote_storage_type::StorageKind, + typename internal::traits<_Rhs>::StorageKind, + internal::product_type<_Lhs,_Rhs>::ret>::ret> { public: @@ -104,8 +106,9 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, typedef typename ProductImpl< Lhs, Rhs, Option, - typename internal::promote_storage_type::ret>::Base Base; + typename internal::product_promote_storage_type::ret>::ret>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Product) typedef typename internal::nested::type LhsNested; diff --git a/Eigen/src/Core/ProductBase.h b/Eigen/src/Core/ProductBase.h index 3b2246fd8..4a54f5f81 100644 --- a/Eigen/src/Core/ProductBase.h +++ b/Eigen/src/Core/ProductBase.h @@ -27,8 +27,9 @@ struct traits > typedef typename remove_all<_Lhs>::type Lhs; typedef typename remove_all<_Rhs>::type Rhs; typedef typename scalar_product_traits::ReturnType Scalar; - typedef typename promote_storage_type::StorageKind, - typename traits::StorageKind>::ret StorageKind; + typedef typename product_promote_storage_type::StorageKind, + typename traits::StorageKind, + 0>::ret StorageKind; typedef typename promote_index_type::Index, typename traits::Index>::type Index; enum { diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h index 85e8ac2c7..76806fd62 100644 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ b/Eigen/src/Core/products/CoeffBasedProduct.h @@ -42,8 +42,9 @@ struct traits > typedef typename remove_all::type _LhsNested; typedef typename remove_all::type _RhsNested; typedef typename scalar_product_traits::ReturnType Scalar; - typedef typename promote_storage_type::StorageKind, - typename traits<_RhsNested>::StorageKind>::ret StorageKind; + typedef typename product_promote_storage_type::StorageKind, + typename traits<_RhsNested>::StorageKind, + 0>::ret StorageKind; typedef typename promote_index_type::Index, typename traits<_RhsNested>::Index>::type Index; diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 131f5d793..4b55b3a9a 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -449,8 +449,6 @@ struct MatrixXpr {}; /** The type used to identify an array expression */ struct ArrayXpr {}; - -#ifdef EIGEN_ENABLE_EVALUATORS // An evaluator must define its shape. By default, it can be one of the following: struct DenseShape { static std::string debugName() { return "DenseShape"; } }; struct DiagonalShape { static std::string debugName() { return "DiagonalShape"; } }; @@ -459,8 +457,6 @@ struct TriangularShape { static std::string debugName() { return "TriangularSha struct SelfAdjointShape { static std::string debugName() { return "SelfAdjointShape"; } }; struct PermutationShape { static std::string debugName() { return "PermutationShape"; } }; struct SparseShape { static std::string debugName() { return "SparseShape"; } }; -#endif - } // end namespace Eigen diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 19adbefd7..70f2c566f 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -231,6 +231,10 @@ template struct plain_matrix_type { typedef typename plain_matrix_type_dense::XprKind>::type type; }; +template struct plain_matrix_type +{ + typedef typename T::PlainObject type; +}; template struct plain_matrix_type_dense { @@ -273,6 +277,11 @@ template struct eval // > type; }; +template struct eval +{ + typedef typename plain_matrix_type::type type; +}; + // for matrices, no need to evaluate, just use a const reference to avoid a useless copy template struct eval, Dense> @@ -515,12 +524,51 @@ template struct cast_return_type const XprType&,CastType>::type type; }; -template struct promote_storage_type; +/** \internal Specify the "storage kind" of applying a coefficient-wise + * binary operations between two expressions of kinds A and B respectively. + * The template parameter Functor permits to specialize the resulting storage kind wrt to + * the functor. + * The default rules are as follows: + * \code + * A op A -> A + * A op dense -> dense + * dense op B -> dense + * A * dense -> A + * dense * B -> B + * \endcode + */ +template struct cwise_promote_storage_type; -template struct promote_storage_type -{ - typedef A ret; -}; +template struct cwise_promote_storage_type { typedef A ret; }; +template struct cwise_promote_storage_type { typedef Dense ret; }; +template struct cwise_promote_storage_type > { typedef Dense ret; }; +template struct cwise_promote_storage_type { typedef Dense ret; }; +template struct cwise_promote_storage_type { typedef Dense ret; }; +template struct cwise_promote_storage_type > { typedef A ret; }; +template struct cwise_promote_storage_type > { typedef B ret; }; + +/** \internal Specify the "storage kind" of multiplying an expression of kind A with kind B. + * The template parameter ProductTag permits to specialize the resulting storage kind wrt to + * some compile-time properties of the product: GemmProduct, GemvProduct, OuterProduct, InnerProduct. + * The default rules are as follows: + * \code + * K * K -> K + * dense * K -> dense + * K * dense -> dense + * diag * K -> K + * K * diag -> K + * \endcode + */ +template struct product_promote_storage_type; + +template struct product_promote_storage_type { typedef A ret;}; +template struct product_promote_storage_type { typedef Dense ret;}; +template struct product_promote_storage_type { typedef Dense ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; +template struct product_promote_storage_type { typedef A ret; }; +template struct product_promote_storage_type { typedef B ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; /** \internal gives the plain matrix or array type to store a row/column/diagonal of a matrix type. * \param Scalar optional parameter allowing to pass a different scalar type than the one of the MatrixType. diff --git a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h index dcb095dcf..68ea6cb11 100644 --- a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h @@ -31,12 +31,6 @@ namespace Eigen { namespace internal { -template<> struct promote_storage_type -{ typedef Sparse ret; }; - -template<> struct promote_storage_type -{ typedef Sparse ret; }; - template::StorageKind, typename _RhsStorageMode = typename traits::StorageKind> From 1e6f53e070ffb3d386bea3cda5e37569c0f11b37 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 17:52:58 +0200 Subject: [PATCH 0574/1103] Use DiagonalShape as the storage kind of DiagonalBase<>. --- Eigen/src/Core/DiagonalMatrix.h | 23 ++++------------------- 1 file changed, 4 insertions(+), 19 deletions(-) diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index 7a0b736f1..e26d24eb2 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -156,7 +156,7 @@ struct traits > : traits > { typedef Matrix<_Scalar,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1> DiagonalVectorType; - typedef Dense StorageKind; + typedef DiagonalShape StorageKind; typedef DenseIndex Index; enum { Flags = LvalueBit | NoPreferredStorageOrderBit @@ -280,7 +280,7 @@ struct traits > typedef _DiagonalVectorType DiagonalVectorType; typedef typename DiagonalVectorType::Scalar Scalar; typedef typename DiagonalVectorType::Index Index; - typedef typename DiagonalVectorType::StorageKind StorageKind; + typedef DiagonalShape StorageKind; typedef typename traits::XprKind XprKind; enum { RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, @@ -364,23 +364,8 @@ bool MatrixBase::isDiagonal(const RealScalar& prec) const #ifdef EIGEN_ENABLE_EVALUATORS namespace internal { - -// TODO currently a diagonal expression has the form DiagonalMatrix<> or DiagonalWrapper -// in the future diagonal-ness should be defined by the expression traits -template -struct evaluator_traits > -{ - typedef typename storage_kind_to_evaluator_kind::Kind Kind; - typedef DiagonalShape Shape; - static const int AssumeAliasing = 0; -}; -template -struct evaluator_traits > -{ - typedef typename storage_kind_to_evaluator_kind::Kind Kind; - typedef DiagonalShape Shape; - static const int AssumeAliasing = 0; -}; + +template<> struct storage_kind_to_shape { typedef DiagonalShape Shape; }; struct Diagonal2Dense {}; From 7390af91b63e4f250bddd5971eab44bae3a89f32 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 17:53:18 +0200 Subject: [PATCH 0575/1103] Implement evaluators for sparse*dense products --- Eigen/SparseCore | 2 +- Eigen/src/SparseCore/SparseDenseProduct.h | 334 +++++++++++++++------- Eigen/src/SparseCore/SparseMatrixBase.h | 34 ++- 3 files changed, 252 insertions(+), 118 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index f74df3038..7cbfb47f2 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -52,10 +52,10 @@ struct Sparse {}; #include "src/SparseCore/ConservativeSparseSparseProduct.h" #include "src/SparseCore/SparseSparseProductWithPruning.h" #include "src/SparseCore/SparseProduct.h" +#include "src/SparseCore/SparseDenseProduct.h" #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" -#include "src/SparseCore/SparseDenseProduct.h" #include "src/SparseCore/SparseTriangularView.h" #include "src/SparseCore/SparseSelfAdjointView.h" #include "src/SparseCore/TriangularSolver.h" diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 610833f3b..2a23365c6 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2010 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -12,6 +12,152 @@ namespace Eigen { +namespace internal { + +template +struct sparse_time_dense_product_impl; + +template +struct sparse_time_dense_product_impl +{ + typedef typename internal::remove_all::type Lhs; + typedef typename internal::remove_all::type Rhs; + typedef typename internal::remove_all::type Res; + typedef typename Lhs::Index Index; +#ifndef EIGEN_TEST_EVALUATORS + typedef typename Lhs::InnerIterator LhsInnerIterator; +#else + typedef typename evaluator::InnerIterator LhsInnerIterator; +#endif + static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) + { +#ifndef EIGEN_TEST_EVALUATORS + const Lhs &lhsEval(lhs); +#else + typename evaluator::type lhsEval(lhs); +#endif + for(Index c=0; c +struct scalar_product_traits > +{ + enum { + Defined = 1 + }; + typedef typename CwiseUnaryOp, T2>::PlainObject ReturnType; +}; +template +struct sparse_time_dense_product_impl +{ + typedef typename internal::remove_all::type Lhs; + typedef typename internal::remove_all::type Rhs; + typedef typename internal::remove_all::type Res; + typedef typename Lhs::Index Index; +#ifndef EIGEN_TEST_EVALUATORS + typedef typename Lhs::InnerIterator LhsInnerIterator; +#else + typedef typename evaluator::InnerIterator LhsInnerIterator; +#endif + static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) + { +#ifndef EIGEN_TEST_EVALUATORS + const Lhs &lhsEval(lhs); +#else + typename evaluator::type lhsEval(lhs); +#endif + for(Index c=0; c::ReturnType rhs_j(alpha * rhs.coeff(j,c)); + for(LhsInnerIterator it(lhsEval,j); it ;++it) + res.coeffRef(it.index(),c) += it.value() * rhs_j; + } + } + } +}; + +template +struct sparse_time_dense_product_impl +{ + typedef typename internal::remove_all::type Lhs; + typedef typename internal::remove_all::type Rhs; + typedef typename internal::remove_all::type Res; + typedef typename Lhs::Index Index; +#ifndef EIGEN_TEST_EVALUATORS + typedef typename Lhs::InnerIterator LhsInnerIterator; +#else + typedef typename evaluator::InnerIterator LhsInnerIterator; +#endif + static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) + { +#ifndef EIGEN_TEST_EVALUATORS + const Lhs &lhsEval(lhs); +#else + typename evaluator::type lhsEval(lhs); +#endif + for(Index j=0; j +struct sparse_time_dense_product_impl +{ + typedef typename internal::remove_all::type Lhs; + typedef typename internal::remove_all::type Rhs; + typedef typename internal::remove_all::type Res; + typedef typename Lhs::Index Index; +#ifndef EIGEN_TEST_EVALUATORS + typedef typename Lhs::InnerIterator LhsInnerIterator; +#else + typedef typename evaluator::InnerIterator LhsInnerIterator; +#endif + static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) + { +#ifndef EIGEN_TEST_EVALUATORS + const Lhs &lhsEval(lhs); +#else + typename evaluator::type lhsEval(lhs); +#endif + for(Index j=0; j +inline void sparse_time_dense_product(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) +{ + sparse_time_dense_product_impl::run(lhs, rhs, res, alpha); +} + +} // end namespace internal + +#ifndef EIGEN_TEST_EVALUATORS template struct SparseDenseProductReturnType { typedef SparseTimeDenseProduct Type; @@ -138,111 +284,6 @@ struct traits > typedef MatrixXpr XprKind; }; -template -struct sparse_time_dense_product_impl; - -template -struct sparse_time_dense_product_impl -{ - typedef typename internal::remove_all::type Lhs; - typedef typename internal::remove_all::type Rhs; - typedef typename internal::remove_all::type Res; - typedef typename Lhs::Index Index; - typedef typename Lhs::InnerIterator LhsInnerIterator; - static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) - { - for(Index c=0; c -struct scalar_product_traits > -{ - enum { - Defined = 1 - }; - typedef typename CwiseUnaryOp, T2>::PlainObject ReturnType; -}; -template -struct sparse_time_dense_product_impl -{ - typedef typename internal::remove_all::type Lhs; - typedef typename internal::remove_all::type Rhs; - typedef typename internal::remove_all::type Res; - typedef typename Lhs::InnerIterator LhsInnerIterator; - typedef typename Lhs::Index Index; - static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) - { - for(Index c=0; c::ReturnType rhs_j(alpha * rhs.coeff(j,c)); - for(LhsInnerIterator it(lhs,j); it ;++it) - res.coeffRef(it.index(),c) += it.value() * rhs_j; - } - } - } -}; - -template -struct sparse_time_dense_product_impl -{ - typedef typename internal::remove_all::type Lhs; - typedef typename internal::remove_all::type Rhs; - typedef typename internal::remove_all::type Res; - typedef typename Lhs::InnerIterator LhsInnerIterator; - typedef typename Lhs::Index Index; - static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) - { - for(Index j=0; j -struct sparse_time_dense_product_impl -{ - typedef typename internal::remove_all::type Lhs; - typedef typename internal::remove_all::type Rhs; - typedef typename internal::remove_all::type Res; - typedef typename Lhs::InnerIterator LhsInnerIterator; - typedef typename Lhs::Index Index; - static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) - { - for(Index j=0; j -inline void sparse_time_dense_product(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) -{ - sparse_time_dense_product_impl::run(lhs, rhs, res, alpha); -} - } // end namespace internal template @@ -305,6 +346,87 @@ SparseMatrixBase::operator*(const MatrixBase &other) cons { return typename SparseDenseProductReturnType::Type(derived(), other.derived()); } +#endif // EIGEN_TEST_EVALUATORS + +#ifdef EIGEN_TEST_EVALUATORS + +namespace internal { + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + typedef typename nested_eval::type LhsNested; + typedef typename nested_eval::type RhsNested; + LhsNested lhsNested(lhs); + RhsNested rhsNested(rhs); + + dst.setZero(); + internal::sparse_time_dense_product(lhsNested, rhsNested, dst, typename Dest::Scalar(1)); + } +}; + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + typedef typename nested_eval::type LhsNested; + typedef typename nested_eval::type RhsNested; + LhsNested lhsNested(lhs); + RhsNested rhsNested(rhs); + + dst.setZero(); + // transpoe everything + Transpose dstT(dst); + internal::sparse_time_dense_product(rhsNested.transpose(), lhsNested.transpose(), dstT, typename Dest::Scalar(1)); + } +}; + +template +struct product_evaluator, ProductTag, SparseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +template +struct product_evaluator, ProductTag, DenseShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +} // end namespace internal + +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index 3a81916fb..3bc5af86d 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -282,6 +282,17 @@ template class SparseMatrixBase : public EigenBase const SparseDiagonalProduct operator*(const DiagonalBase &lhs, const SparseMatrixBase& rhs) { return SparseDiagonalProduct(lhs.derived(), rhs.derived()); } + + /** dense * sparse (return a dense object unless it is an outer product) */ + template friend + const typename DenseSparseProductReturnType::Type + operator*(const MatrixBase& lhs, const Derived& rhs) + { return typename DenseSparseProductReturnType::Type(lhs.derived(),rhs); } + + /** sparse * dense (returns a dense object unless it is an outer product) */ + template + const typename SparseDenseProductReturnType::Type + operator*(const MatrixBase &other) const; #else // EIGEN_TEST_EVALUATORS // sparse * diagonal template @@ -299,18 +310,19 @@ template class SparseMatrixBase : public EigenBase template const Product operator*(const SparseMatrixBase &other) const; -#endif // EIGEN_TEST_EVALUATORS - - /** dense * sparse (return a dense object unless it is an outer product) */ - template friend - const typename DenseSparseProductReturnType::Type - operator*(const MatrixBase& lhs, const Derived& rhs) - { return typename DenseSparseProductReturnType::Type(lhs.derived(),rhs); } - - /** sparse * dense (returns a dense object unless it is an outer product) */ + + // sparse * dense template - const typename SparseDenseProductReturnType::Type - operator*(const MatrixBase &other) const; + const Product + operator*(const MatrixBase &other) const + { return Product(derived(), other.derived()); } + + // dense * sparse + template friend + const Product + operator*(const MatrixBase &lhs, const SparseMatrixBase& rhs) + { return Product(lhs.derived(), rhs.derived()); } +#endif // EIGEN_TEST_EVALUATORS /** \returns an expression of P H P^-1 where H is the matrix represented by \c *this */ SparseSymmetricPermutationProduct twistedBy(const PermutationMatrix& perm) const From 8f4cdbbc8f9b2214b906412701722a150cba3460 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 18:04:30 +0200 Subject: [PATCH 0576/1103] Fix typo in dense * diagonal evaluator. --- Eigen/src/Core/ProductEvaluators.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 7298c51b1..9436e200b 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -858,7 +858,7 @@ struct product_evaluator, ProductTag, DenseShape, typedef Product XprType; typedef typename XprType::PlainObject PlainObject; - enum { StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; + enum { StorageOrder = int(Lhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal()) From bf334b8ae51725754f525c2ffcfbd83ffc55ff2e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 1 Jul 2014 22:29:04 +0200 Subject: [PATCH 0577/1103] Fix regeression in bicgstab: the threshold used to detect the need for a restart was much too large. --- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index 7a46b51fa..dc524c225 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -61,6 +61,7 @@ bool bicgstab(const MatrixType& mat, const Rhs& rhs, Dest& x, VectorType s(n), t(n); RealScalar tol2 = tol*tol; + RealScalar eps2 = NumTraits::epsilon()*NumTraits::epsilon(); int i = 0; int restarts = 0; @@ -69,7 +70,7 @@ bool bicgstab(const MatrixType& mat, const Rhs& rhs, Dest& x, Scalar rho_old = rho; rho = r0.dot(r); - if (internal::isMuchSmallerThan(rho,r0_sqnorm)) + if (abs(rho) < eps2*r0_sqnorm) { // The new residual vector became too orthogonal to the arbitrarily choosen direction r0 // Let's restart with a new r0: From 0a8e4712d1f3cfe8830783f59c6c9987ea34701c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 2 Jul 2014 16:13:05 +0200 Subject: [PATCH 0578/1103] Do not attempt to include on Windows CE --- Eigen/Core | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/Core b/Eigen/Core index a135afc6a..16e388fa4 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -205,7 +205,7 @@ #endif // required for __cpuid, needs to be included after cmath -#if defined(_MSC_VER) && (defined(_M_IX86)||defined(_M_X64)) +#if defined(_MSC_VER) && (defined(_M_IX86)||defined(_M_X64)) && (!defined(_WIN32_WCE)) #include #endif From 998455a5707e1312d3058672d6d337a20158d86e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 2 Jul 2014 23:04:46 +0200 Subject: [PATCH 0579/1103] LDLT is not rank-revealing, so we should not attempt to use the biggest diagonal elements as thresholds. --- Eigen/src/Cholesky/LDLT.h | 27 +++++++++++++-------------- test/cholesky.cpp | 20 ++++++++++++++++++++ 2 files changed, 33 insertions(+), 14 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 1d84438e2..67a99a0af 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -258,23 +258,13 @@ template<> struct ldlt_inplace return true; } - RealScalar cutoff(0), biggest_in_corner; - for (Index k = 0; k < size; ++k) { // Find largest diagonal element Index index_of_biggest_in_corner; - biggest_in_corner = mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner); + mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner); index_of_biggest_in_corner += k; - if(k == 0) - { - // The biggest overall is the point of reference to which further diagonals - // are compared; if any diagonal is negligible compared - // to the largest overall, the algorithm bails. - cutoff = abs(NumTraits::epsilon() * biggest_in_corner); - } - transpositions.coeffRef(k) = index_of_biggest_in_corner; if(k != index_of_biggest_in_corner) { @@ -311,7 +301,11 @@ template<> struct ldlt_inplace A21.noalias() -= A20 * temp.head(k); } - if((rs>0) && (abs(mat.coeffRef(k,k)) > cutoff)) + // In some previous versions of Eigen (e.g., 3.2.1), the scaling was omitted if the pivot + // was smaller than the cutoff value. However, soince LDLT is not rank-revealing + // we should only make sure we do not introduce INF or NaN values. + // LAPACK also uses 0 as the cutoff value. + if((rs>0) && (abs(mat.coeffRef(k,k)) > RealScalar(0))) A21 /= mat.coeffRef(k,k); RealScalar realAkk = numext::real(mat.coeffRef(k,k)); @@ -495,8 +489,13 @@ struct solve_retval, Rhs> typedef typename LDLTType::Scalar Scalar; typedef typename LDLTType::RealScalar RealScalar; const Diagonal vectorD = dec().vectorD(); - RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() * NumTraits::epsilon(), - RealScalar(1) / NumTraits::highest()); // motivated by LAPACK's xGELSS + // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon + // as motivated by LAPACK's xGELSS: + // RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() *NumTraits::epsilon(),RealScalar(1) / NumTraits::highest()); + // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest + // diagonal element is not well justified and to numerical issues in some cases. + // Moreover, Lapack's xSYTRS routines use 0 for the tolerance. + RealScalar tolerance = RealScalar(1) / NumTraits::highest(); for (Index i = 0; i < vectorD.size(); ++i) { if(abs(vectorD(i)) > tolerance) dst.row(i) /= vectorD(i); diff --git a/test/cholesky.cpp b/test/cholesky.cpp index 569318f83..9aa7da7bb 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -68,6 +68,7 @@ template void cholesky(const MatrixType& m) Index cols = m.cols(); typedef typename MatrixType::Scalar Scalar; + typedef typename NumTraits::Real RealScalar; typedef Matrix SquareMatrixType; typedef Matrix VectorType; @@ -207,6 +208,25 @@ template void cholesky(const MatrixType& m) vecX = ldltlo.solve(vecB); VERIFY_IS_APPROX(A * vecX, vecB); } + + // check matrices with wide spectrum + if(rows>=3) + { + RealScalar s = (std::min)(16,std::numeric_limits::max_exponent10/8); + Matrix a = Matrix::Random(rows,rows); + Matrix d = Matrix::Random(rows); + for(int k=0; k(-s,s)); + SquareMatrixType A = a * d.asDiagonal() * a.adjoint(); + // Make sure a solution exists: + vecX.setRandom(); + vecB = A * vecX; + vecX.setZero(); + ldltlo.compute(A); + VERIFY_IS_APPROX(A, ldltlo.reconstructedMatrix()); + vecX = ldltlo.solve(vecB); + VERIFY_IS_APPROX(A * vecX, vecB); + } } // update/downdate From 3a9f9faada5de3a13611d7543014c5ef503eff58 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 4 Jul 2014 12:48:24 +0200 Subject: [PATCH 0580/1103] Fix unused typedef warning --- Eigen/src/Cholesky/LDLT.h | 1 - 1 file changed, 1 deletion(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 67a99a0af..ef81030f8 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -486,7 +486,6 @@ struct solve_retval, Rhs> using std::abs; EIGEN_USING_STD_MATH(max); typedef typename LDLTType::MatrixType MatrixType; - typedef typename LDLTType::Scalar Scalar; typedef typename LDLTType::RealScalar RealScalar; const Diagonal vectorD = dec().vectorD(); // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon From f36538049669a7efee57f2b1e3c60bf8bf3976bb Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 4 Jul 2014 12:52:55 +0200 Subject: [PATCH 0581/1103] Fix regression introduced by 3117036b80075390dbc46f60aa0d595e5a44661b : Matrix(int) did not compile if Scalar is not constructible from int. Now this falls back to the (Index size) constructor. --- Eigen/src/Core/PlainObjectBase.h | 4 ++-- Eigen/src/Core/util/Meta.h | 19 +++++++++++++++++++ 2 files changed, 21 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index b5b25ef70..ae5b342ce 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -692,7 +692,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if::type* = 0) + EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if::value,T>::type* = 0) { EIGEN_STATIC_ASSERT(bool(NumTraits::IsInteger), FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) @@ -700,7 +700,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type } template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if::type* = 0) + EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if::value,T>::type* = 0) { EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) m_storage.data()[0] = val0; diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h index e4e4d4a87..795197f59 100644 --- a/Eigen/src/Core/util/Meta.h +++ b/Eigen/src/Core/util/Meta.h @@ -80,6 +80,25 @@ template struct add_const_on_value_type { typedef T const template struct add_const_on_value_type { typedef T const* const type; }; template struct add_const_on_value_type { typedef T const* const type; }; + +template +struct is_convertible +{ +private: + struct yes {int a[1];}; + struct no {int a[2];}; + + template + static yes test (const T&) {} + + template static no test (...) {} + +public: + static From ms_from; + enum { value = sizeof(test(ms_from))==sizeof(yes) }; +}; + + /** \internal Allows to enable/disable an overload * according to a compile time condition. */ From 8ee38d2db69aa7245d4aa1c5cb6a177a866da20f Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 5 Jul 2014 21:48:48 +0800 Subject: [PATCH 0582/1103] Fix dox for namespaces --- Eigen/src/Jacobi/Jacobi.h | 2 +- unsupported/Eigen/src/IterativeSolvers/Scaling.h | 6 ++++-- 2 files changed, 5 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h index 956f72d57..da9fb53d0 100644 --- a/Eigen/src/Jacobi/Jacobi.h +++ b/Eigen/src/Jacobi/Jacobi.h @@ -255,13 +255,13 @@ void JacobiRotation::makeGivens(const Scalar& p, const Scalar& q, Scalar * Implementation of MatrixBase methods ****************************************************************************************/ +namespace internal { /** \jacobi_module * Applies the clock wise 2D rotation \a j to the set of 2D vectors of cordinates \a x and \a y: * \f$ \left ( \begin{array}{cc} x \\ y \end{array} \right ) = J \left ( \begin{array}{cc} x \\ y \end{array} \right ) \f$ * * \sa MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight() */ -namespace internal { template void apply_rotation_in_the_plane(VectorX& _x, VectorY& _y, const JacobiRotation& j); } diff --git a/unsupported/Eigen/src/IterativeSolvers/Scaling.h b/unsupported/Eigen/src/IterativeSolvers/Scaling.h index 4fd439202..9189fddf1 100644 --- a/unsupported/Eigen/src/IterativeSolvers/Scaling.h +++ b/unsupported/Eigen/src/IterativeSolvers/Scaling.h @@ -9,6 +9,10 @@ #ifndef EIGEN_ITERSCALING_H #define EIGEN_ITERSCALING_H + +namespace Eigen { +using std::abs; + /** * \ingroup IterativeSolvers_Module * \brief iterative scaling algorithm to equilibrate rows and column norms in matrices @@ -41,8 +45,6 @@ * * \sa \ref IncompleteLUT */ -namespace Eigen { -using std::abs; template class IterScaling { From 1a817d3b7090c0cecd648003e0b24d434967d36e Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 5 Jul 2014 21:50:05 +0800 Subject: [PATCH 0583/1103] Document internal namespace --- Eigen/Core | 3 +++ unsupported/Eigen/doxygen.hpp | 19 +++++++++++++++++++ 2 files changed, 22 insertions(+) create mode 100644 unsupported/Eigen/doxygen.hpp diff --git a/Eigen/Core b/Eigen/Core index 16e388fa4..3f1efb538 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -220,6 +220,9 @@ /** \brief Namespace containing all symbols from the %Eigen library. */ namespace Eigen { +/** \brief Namespace containing low-level routines from the %Eigen library. */ +namespace internal {} + inline static const char *SimdInstructionSetsInUse(void) { #if defined(EIGEN_VECTORIZE_AVX) return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2"; diff --git a/unsupported/Eigen/doxygen.hpp b/unsupported/Eigen/doxygen.hpp new file mode 100644 index 000000000..031488527 --- /dev/null +++ b/unsupported/Eigen/doxygen.hpp @@ -0,0 +1,19 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Chen-Pang He +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +// As its name suggests, this file is intended to be read by Doxygen. +// There is nothing defined so we need no #include guard. + +//! \brief Namespace containing all symbols from the %Eigen library. +namespace Eigen { + +//! \brief Namespace containing low-level routines from the %Eigen library. +namespace internal {} + +} From 7a915f68467df10fb321b4bfa6c139669b00daaa Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 5 Jul 2014 22:41:58 +0800 Subject: [PATCH 0584/1103] Move Doxygen-only stuff to *.dox --- Eigen/Core | 3 --- doc/Manual.dox | 3 +++ unsupported/Eigen/doxygen.hpp | 19 ------------------- unsupported/doc/Overview.dox | 3 +++ 4 files changed, 6 insertions(+), 22 deletions(-) delete mode 100644 unsupported/Eigen/doxygen.hpp diff --git a/Eigen/Core b/Eigen/Core index 3f1efb538..16e388fa4 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -220,9 +220,6 @@ /** \brief Namespace containing all symbols from the %Eigen library. */ namespace Eigen { -/** \brief Namespace containing low-level routines from the %Eigen library. */ -namespace internal {} - inline static const char *SimdInstructionSetsInUse(void) { #if defined(EIGEN_VECTORIZE_AVX) return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2"; diff --git a/doc/Manual.dox b/doc/Manual.dox index 0cbe792d5..d6d4d4689 100644 --- a/doc/Manual.dox +++ b/doc/Manual.dox @@ -159,4 +159,7 @@ namespace Eigen { \ingroup Geometry_Reference */ /** \addtogroup Splines_Module \ingroup Geometry_Reference */ + +/** \brief Namespace containing low-level routines from the %Eigen library. */ +namespace internal {} } diff --git a/unsupported/Eigen/doxygen.hpp b/unsupported/Eigen/doxygen.hpp deleted file mode 100644 index 031488527..000000000 --- a/unsupported/Eigen/doxygen.hpp +++ /dev/null @@ -1,19 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2014 Chen-Pang He -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// As its name suggests, this file is intended to be read by Doxygen. -// There is nothing defined so we need no #include guard. - -//! \brief Namespace containing all symbols from the %Eigen library. -namespace Eigen { - -//! \brief Namespace containing low-level routines from the %Eigen library. -namespace internal {} - -} diff --git a/unsupported/doc/Overview.dox b/unsupported/doc/Overview.dox index d048377df..93b2a8927 100644 --- a/unsupported/doc/Overview.dox +++ b/unsupported/doc/Overview.dox @@ -1,3 +1,4 @@ +/// \brief Namespace containing all symbols from the %Eigen library. namespace Eigen { /** \mainpage Eigen's unsupported modules @@ -22,4 +23,6 @@ subject to be included in Eigen in the future. */ +/// \brief Namespace containing low-level routines from the %Eigen library. +namespace internal {} } From 4860da2de12cd940d506db0f20e39ef2cb7266ba Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sat, 5 Jul 2014 23:01:27 +0800 Subject: [PATCH 0585/1103] Percent "Eigen" in dox to prevent linking if not referring to the Eigen namespace --- unsupported/doc/Overview.dox | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/unsupported/doc/Overview.dox b/unsupported/doc/Overview.dox index 93b2a8927..ef361b35b 100644 --- a/unsupported/doc/Overview.dox +++ b/unsupported/doc/Overview.dox @@ -1,15 +1,15 @@ /// \brief Namespace containing all symbols from the %Eigen library. namespace Eigen { -/** \mainpage Eigen's unsupported modules +/** \mainpage %Eigen's unsupported modules -This is the API documentation for Eigen's unsupported modules. +This is the API documentation for %Eigen's unsupported modules. These modules are contributions from various users. They are provided "as is", without any support. Click on the \e Modules tab at the top of this page to get a list of all unsupported modules. -Don't miss the official Eigen documentation. +Don't miss the official Eigen documentation. */ @@ -19,7 +19,7 @@ Don't miss the official Eigen documentation. The unsupported modules are contributions from various users. They are provided "as is", without any support. Nevertheless, some of them are -subject to be included in Eigen in the future. +subject to be included in %Eigen in the future. */ From ac1bb3e5b3349bf744c3905fec55295e6bb024ab Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Jul 2014 14:22:33 +0200 Subject: [PATCH 0586/1103] bug #770: fix out of bounds access --- unsupported/Eigen/MPRealSupport | 3 +++ 1 file changed, 3 insertions(+) diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport index 35d77e5bd..8e22b13c7 100644 --- a/unsupported/Eigen/MPRealSupport +++ b/unsupported/Eigen/MPRealSupport @@ -157,6 +157,9 @@ int main() void operator()(mpreal* res, Index resStride, const mpreal* blockA, const mpreal* blockB, Index rows, Index depth, Index cols, mpreal alpha, Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0) { + if(rows==0 || cols==0 || depth==0) + return; + mpreal acc1(0,mpfr_get_prec(blockA[0].mpfr_srcptr())), tmp (0,mpfr_get_prec(blockA[0].mpfr_srcptr())); From 1cb71a8782c5aa8d75562ebf2c36d1581e7799ef Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 18 Jul 2014 14:34:58 +0200 Subject: [PATCH 0587/1103] bug #138: Make building of internal documentation configurable via cmake flag --- doc/CMakeLists.txt | 8 ++++++++ doc/Doxyfile.in | 10 +++++----- 2 files changed, 13 insertions(+), 5 deletions(-) diff --git a/doc/CMakeLists.txt b/doc/CMakeLists.txt index 8a493031c..1b8aaf9aa 100644 --- a/doc/CMakeLists.txt +++ b/doc/CMakeLists.txt @@ -10,12 +10,20 @@ if(CMAKE_COMPILER_IS_GNUCXX) endif(CMAKE_SYSTEM_NAME MATCHES Linux) endif(CMAKE_COMPILER_IS_GNUCXX) +option(EIGEN_INTERNAL_DOCUMENTATION "Build internal documentation" OFF) + + # Set some Doxygen flags set(EIGEN_DOXY_PROJECT_NAME "Eigen") set(EIGEN_DOXY_OUTPUT_DIRECTORY_SUFFIX "") set(EIGEN_DOXY_INPUT "\"${Eigen_SOURCE_DIR}/Eigen\" \"${Eigen_SOURCE_DIR}/doc\"") set(EIGEN_DOXY_HTML_COLORSTYLE_HUE "220") set(EIGEN_DOXY_TAGFILES "") +if(EIGEN_INTERNAL_DOCUMENTATION) + set(EIGEN_DOXY_INTERNAL "YES") +else(EIGEN_INTERNAL_DOCUMENTATION) + set(EIGEN_DOXY_INTERNAL "NO") +endif(EIGEN_INTERNAL_DOCUMENTATION) configure_file( ${CMAKE_CURRENT_SOURCE_DIR}/Doxyfile.in diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in index 2ceee5e20..5d82add72 100644 --- a/doc/Doxyfile.in +++ b/doc/Doxyfile.in @@ -460,7 +460,7 @@ HIDE_IN_BODY_DOCS = NO # to NO (the default) then the documentation will be excluded. # Set it to YES to include the internal documentation. -INTERNAL_DOCS = NO +INTERNAL_DOCS = ${EIGEN_DOXY_INTERNAL} # If the CASE_SENSE_NAMES tag is set to NO then Doxygen will only generate # file names in lower-case letters. If set to YES upper-case letters are also @@ -480,7 +480,7 @@ HIDE_SCOPE_NAMES = NO # will put a list of the files that are included by a file in the documentation # of that file. -SHOW_INCLUDE_FILES = NO +SHOW_INCLUDE_FILES = ${EIGEN_DOXY_INTERNAL} # If the FORCE_LOCAL_INCLUDES tag is set to YES then Doxygen # will list include files with double quotes in the documentation @@ -546,7 +546,7 @@ STRICT_PROTO_MATCHING = NO # disable (NO) the todo list. This list is created by putting \todo # commands in the documentation. -GENERATE_TODOLIST = NO +GENERATE_TODOLIST = ${EIGEN_DOXY_INTERNAL} # The GENERATE_TESTLIST tag can be used to enable (YES) or # disable (NO) the test list. This list is created by putting \test @@ -558,13 +558,13 @@ GENERATE_TESTLIST = NO # disable (NO) the bug list. This list is created by putting \bug # commands in the documentation. -GENERATE_BUGLIST = NO +GENERATE_BUGLIST = ${EIGEN_DOXY_INTERNAL} # The GENERATE_DEPRECATEDLIST tag can be used to enable (YES) or # disable (NO) the deprecated list. This list is created by putting # \deprecated commands in the documentation. -GENERATE_DEPRECATEDLIST= NO +GENERATE_DEPRECATEDLIST= ${EIGEN_DOXY_INTERNAL} # The ENABLED_SECTIONS tag can be used to enable conditional # documentation sections, marked by \if sectionname ... \endif. From 68eafc10b1165eb83461187e549a54a618495916 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 18 Jul 2014 15:42:12 +0200 Subject: [PATCH 0588/1103] Add note to EIGEN_DONT_PARALLELIZE into preprocessor documentation page (requested in IRC) --- doc/PreprocessorDirectives.dox | 2 ++ 1 file changed, 2 insertions(+) diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox index e67991fb7..96d273823 100644 --- a/doc/PreprocessorDirectives.dox +++ b/doc/PreprocessorDirectives.dox @@ -61,6 +61,8 @@ run time. However, these assertions do cost time and can thus be turned off. expect that any objects passed to it are aligned. This will turn off vectorization. Not defined by default. - \b EIGEN_DONT_ALIGN_STATICALLY - disables alignment of arrays on the stack. Not defined by default, unless \c EIGEN_DONT_ALIGN is defined. + - \b EIGEN_DONT_PARALLELIZE - if defined, this disables multi-threading. This is only relevant if you enabled OpenMP. + See \ref TopicMultiThreading for details. - \b EIGEN_DONT_VECTORIZE - disables explicit vectorization when defined. Not defined by default, unless alignment is disabled by %Eigen's platform test or the user defining \c EIGEN_DONT_ALIGN. - \b EIGEN_FAST_MATH - enables some optimizations which might affect the accuracy of the result. This currently From ef4a86d6b8eb7334ffa3a5a55f8432d723f8a502 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 18 Jul 2014 16:39:58 +0200 Subject: [PATCH 0589/1103] Fix trivial warnings in MPRealSupport --- unsupported/Eigen/MPRealSupport | 6 +++--- unsupported/test/mpreal/mpreal.h | 4 ++-- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport index 8e22b13c7..632de3854 100644 --- a/unsupported/Eigen/MPRealSupport +++ b/unsupported/Eigen/MPRealSupport @@ -88,9 +88,9 @@ int main() inline static Real epsilon (const Real& x) { return mpfr::machine_epsilon(x); } inline static Real dummy_precision() - { - unsigned int weak_prec = ((mpfr::mpreal::get_default_prec()-1) * 90) / 100; - return mpfr::machine_epsilon(weak_prec); + { + mpfr_prec_t weak_prec = ((mpfr::mpreal::get_default_prec()-1) * 90) / 100; + return mpfr::machine_epsilon(weak_prec); } }; diff --git a/unsupported/test/mpreal/mpreal.h b/unsupported/test/mpreal/mpreal.h index 4c29d2ac2..dddda7dd9 100644 --- a/unsupported/test/mpreal/mpreal.h +++ b/unsupported/test/mpreal/mpreal.h @@ -1698,7 +1698,7 @@ inline bool isregular(const mpreal& op){ return (mpfr_regular_p(op.mpfr_srcpt ////////////////////////////////////////////////////////////////////////// // Type Converters -inline bool mpreal::toBool (mp_rnd_t mode) const { return mpfr_zero_p (mpfr_srcptr()) == 0; } +inline bool mpreal::toBool (mp_rnd_t /*mode*/) const { return mpfr_zero_p (mpfr_srcptr()) == 0; } inline long mpreal::toLong (mp_rnd_t mode) const { return mpfr_get_si (mpfr_srcptr(), mode); } inline unsigned long mpreal::toULong (mp_rnd_t mode) const { return mpfr_get_ui (mpfr_srcptr(), mode); } inline float mpreal::toFloat (mp_rnd_t mode) const { return mpfr_get_flt(mpfr_srcptr(), mode); } @@ -3070,4 +3070,4 @@ namespace std } -#endif /* __MPREAL_H__ */ \ No newline at end of file +#endif /* __MPREAL_H__ */ From 8e19027130b1e31cfab4e09b67c155c07cf85ff6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 20 Jul 2014 14:03:22 +0200 Subject: [PATCH 0590/1103] bug #826: fix 64 to 32 bits conversion warning when calling Matrix(long) --- Eigen/src/Core/PlainObjectBase.h | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index ae5b342ce..e4e3e8903 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -705,6 +705,17 @@ class PlainObjectBase : public internal::dense_xpr_base::type EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) m_storage.data()[0] = val0; } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Index& val0, + typename internal::enable_if< (!internal::is_same::value) + && Base::SizeAtCompileTime==1 + && internal::is_convertible::value,T>::type* = 0) + { + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) + m_storage.data()[0] = Scalar(val0); + } template EIGEN_DEVICE_FUNC From d4cc1bdc7f27b856c4aefc939254ec851c831bf0 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 20 Jul 2014 14:22:58 +0200 Subject: [PATCH 0591/1103] Make the ordering method of SimplicialL[D]LT user configurable. --- Eigen/src/SparseCholesky/SimplicialCholesky.h | 42 ++++++++++--------- test/simplicial_cholesky.cpp | 41 ++++++++++-------- 2 files changed, 46 insertions(+), 37 deletions(-) diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index f41d7e010..e1f96ba5a 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -37,6 +37,7 @@ class SimplicialCholeskyBase : internal::noncopyable { public: typedef typename internal::traits::MatrixType MatrixType; + typedef typename internal::traits::OrderingType OrderingType; enum { UpLo = internal::traits::UpLo }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; @@ -240,15 +241,16 @@ class SimplicialCholeskyBase : internal::noncopyable RealScalar m_shiftScale; }; -template class SimplicialLLT; -template class SimplicialLDLT; -template class SimplicialCholesky; +template > class SimplicialLLT; +template > class SimplicialLDLT; +template > class SimplicialCholesky; namespace internal { -template struct traits > +template struct traits > { typedef _MatrixType MatrixType; + typedef _Ordering OrderingType; enum { UpLo = _UpLo }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::Index Index; @@ -259,9 +261,10 @@ template struct traits struct traits > +template struct traits > { typedef _MatrixType MatrixType; + typedef _Ordering OrderingType; enum { UpLo = _UpLo }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::Index Index; @@ -272,9 +275,10 @@ template struct traits struct traits > +template struct traits > { typedef _MatrixType MatrixType; + typedef _Ordering OrderingType; enum { UpLo = _UpLo }; }; @@ -294,11 +298,12 @@ template struct traits * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower * or Upper. Default is Lower. + * \tparam _Ordering The ordering method to use, either AMDOrdering<> or NaturalOrdering<>. Default is AMDOrdering<> * - * \sa class SimplicialLDLT + * \sa class SimplicialLDLT, class AMDOrdering, class NaturalOrdering */ -template - class SimplicialLLT : public SimplicialCholeskyBase > +template + class SimplicialLLT : public SimplicialCholeskyBase > { public: typedef _MatrixType MatrixType; @@ -382,11 +387,12 @@ public: * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<> * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower * or Upper. Default is Lower. + * \tparam _Ordering The ordering method to use, either AMDOrdering<> or NaturalOrdering<>. Default is AMDOrdering<> * - * \sa class SimplicialLLT + * \sa class SimplicialLLT, class AMDOrdering, class NaturalOrdering */ -template - class SimplicialLDLT : public SimplicialCholeskyBase > +template + class SimplicialLDLT : public SimplicialCholeskyBase > { public: typedef _MatrixType MatrixType; @@ -467,8 +473,8 @@ public: * * \sa class SimplicialLDLT, class SimplicialLLT */ -template - class SimplicialCholesky : public SimplicialCholeskyBase > +template + class SimplicialCholesky : public SimplicialCholeskyBase > { public: typedef _MatrixType MatrixType; @@ -612,15 +618,13 @@ void SimplicialCholeskyBase::ordering(const MatrixType& a, CholMatrixTy { eigen_assert(a.rows()==a.cols()); const Index size = a.rows(); - // TODO allows to configure the permutation // Note that amd compute the inverse permutation { CholMatrixType C; C = a.template selfadjointView(); - // remove diagonal entries: - // seems not to be needed - // C.prune(keep_diag()); - internal::minimum_degree_ordering(C, m_Pinv); + + OrderingType ordering; + ordering(C,m_Pinv); } if(m_Pinv.size()>0) diff --git a/test/simplicial_cholesky.cpp b/test/simplicial_cholesky.cpp index e93a52e9c..786468421 100644 --- a/test/simplicial_cholesky.cpp +++ b/test/simplicial_cholesky.cpp @@ -11,26 +11,31 @@ template void test_simplicial_cholesky_T() { - SimplicialCholesky, Lower> chol_colmajor_lower; - SimplicialCholesky, Upper> chol_colmajor_upper; - SimplicialLLT, Lower> llt_colmajor_lower; - SimplicialLDLT, Upper> llt_colmajor_upper; - SimplicialLDLT, Lower> ldlt_colmajor_lower; - SimplicialLDLT, Upper> ldlt_colmajor_upper; + SimplicialCholesky, Lower> chol_colmajor_lower_amd; + SimplicialCholesky, Upper> chol_colmajor_upper_amd; + SimplicialLLT, Lower> llt_colmajor_lower_amd; + SimplicialLLT, Upper> llt_colmajor_upper_amd; + SimplicialLDLT, Lower> ldlt_colmajor_lower_amd; + SimplicialLDLT, Upper> ldlt_colmajor_upper_amd; + SimplicialLDLT, Lower, NaturalOrdering > ldlt_colmajor_lower_nat; + SimplicialLDLT, Upper, NaturalOrdering > ldlt_colmajor_upper_nat; - check_sparse_spd_solving(chol_colmajor_lower); - check_sparse_spd_solving(chol_colmajor_upper); - check_sparse_spd_solving(llt_colmajor_lower); - check_sparse_spd_solving(llt_colmajor_upper); - check_sparse_spd_solving(ldlt_colmajor_lower); - check_sparse_spd_solving(ldlt_colmajor_upper); + check_sparse_spd_solving(chol_colmajor_lower_amd); + check_sparse_spd_solving(chol_colmajor_upper_amd); + check_sparse_spd_solving(llt_colmajor_lower_amd); + check_sparse_spd_solving(llt_colmajor_upper_amd); + check_sparse_spd_solving(ldlt_colmajor_lower_amd); + check_sparse_spd_solving(ldlt_colmajor_upper_amd); - check_sparse_spd_determinant(chol_colmajor_lower); - check_sparse_spd_determinant(chol_colmajor_upper); - check_sparse_spd_determinant(llt_colmajor_lower); - check_sparse_spd_determinant(llt_colmajor_upper); - check_sparse_spd_determinant(ldlt_colmajor_lower); - check_sparse_spd_determinant(ldlt_colmajor_upper); + check_sparse_spd_determinant(chol_colmajor_lower_amd); + check_sparse_spd_determinant(chol_colmajor_upper_amd); + check_sparse_spd_determinant(llt_colmajor_lower_amd); + check_sparse_spd_determinant(llt_colmajor_upper_amd); + check_sparse_spd_determinant(ldlt_colmajor_lower_amd); + check_sparse_spd_determinant(ldlt_colmajor_upper_amd); + + check_sparse_spd_solving(ldlt_colmajor_lower_nat); + check_sparse_spd_solving(ldlt_colmajor_upper_nat); } void test_simplicial_cholesky() From 339f14b8d1b73db0366afc3a497d566cecf72e1b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 21 Jul 2014 13:43:48 +0200 Subject: [PATCH 0592/1103] bug #826: document caveats in 1x1 and 2x1 constructors. --- Eigen/src/Core/Matrix.h | 22 ++++++++++++++++++---- doc/PreprocessorDirectives.dox | 13 +++++++++++++ 2 files changed, 31 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index cd2ff91e7..8c95ee3ca 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -257,9 +257,15 @@ class Matrix /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors * - * Note that this is only useful for dynamic-size vectors. For fixed-size vectors, - * it is redundant to pass the dimension here, so it makes more sense to use the default - * constructor Matrix() instead. + * This is useful for dynamic-size vectors. For fixed-size vectors, + * it is redundant to pass these parameters, so one should use the default constructor + * Matrix() instead. + * + * \warning This constructor is disabled for fixed-size \c 1x1 matrices. For instance, + * calling Matrix(1) will call the initialization constructor: Matrix(const Scalar&). + * For fixed-size \c 1x1 matrices it is thefore recommended to use the default + * constructor Matrix() instead, especilly when using one of the non standard + * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives). */ EIGEN_STRONG_INLINE explicit Matrix(Index dim); /** \brief Constructs an initialized 1x1 matrix with the given coefficient */ @@ -268,9 +274,17 @@ class Matrix * * This is useful for dynamic-size matrices. For fixed-size matrices, * it is redundant to pass these parameters, so one should use the default constructor - * Matrix() instead. */ + * Matrix() instead. + * + * \warning This constructor is disabled for fixed-size \c 1x2 and \c 2x1 vectors. For instance, + * calling Matrix2f(2,1) will call the initialization constructor: Matrix(const Scalar& x, const Scalar& y). + * For fixed-size \c 1x2 or \c 2x1 vectors it is thefore recommended to use the default + * constructor Matrix() instead, especilly when using one of the non standard + * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives). + */ EIGEN_DEVICE_FUNC Matrix(Index rows, Index cols); + /** \brief Constructs an initialized 2D vector with given coefficients */ Matrix(const Scalar& x, const Scalar& y); #endif diff --git a/doc/PreprocessorDirectives.dox b/doc/PreprocessorDirectives.dox index 96d273823..4be2167ef 100644 --- a/doc/PreprocessorDirectives.dox +++ b/doc/PreprocessorDirectives.dox @@ -27,10 +27,23 @@ are doing. Defaults to the %IOFormat constructed by the default constructor IOFormat::IOFormat(). - \b EIGEN_INITIALIZE_MATRICES_BY_ZERO - if defined, all entries of newly constructed matrices and arrays are initialized to zero, as are new entries in matrices and arrays after resizing. Not defined by default. + \warning The unary (resp. binary) constructor of \c 1x1 (resp. \c 2x1 or \c 1x2) fixed size matrices is + always interpreted as an initialization constructor where the argument(s) are the coefficient values + and not the sizes. For instance, \code Vector2d v(2,1); \endcode will create a vector with coeficients [2,1], + and \b not a \c 2x1 vector initialized with zeros (i.e., [0,0]). If such cases might occur, then it is + recommended to use the default constructor with a explicit call to resize: + \code + Matrix v; + v.resize(size); + Matrix m; + m.resize(rows,cols); + \endcode - \b EIGEN_INITIALIZE_MATRICES_BY_NAN - if defined, all entries of newly constructed matrices and arrays are initialized to NaN, as are new entries in matrices and arrays after resizing. This option is especially useful for debugging purpose, though a memory tool like valgrind is preferable. Not defined by default. + \warning See the documentation of \c EIGEN_INITIALIZE_MATRICES_BY_ZERO for a discussion on a limitations + of these macros when applied to \c 1x1, \c 1x2, and \c 2x1 fixed-size matrices. - \b EIGEN_NO_AUTOMATIC_RESIZING - if defined, the matrices (or arrays) on both sides of an assignment a = b have to be of the same size; otherwise, %Eigen automatically resizes \c a so that it is of the correct size. Not defined by default. From 58687aa5e638d365d5e41c1e6c66cbfc44fce85f Mon Sep 17 00:00:00 2001 From: Moritz Klammler Date: Sun, 6 Jul 2014 06:58:13 +0200 Subject: [PATCH 0593/1103] Avoid memory leak when constructor of user-defined type throws exception. The added check `ctorleak.cpp` demonstrates how the leak can be reproduced. The test appears to pass but it is leaking the storage of the (not created) matrix. I don't know how to make this test fail in the existing test suite but you can run it through Valgrind (or another debugger) to verify the leak. $ ./check.sh ctorleak && valgrind --leak-check=full ./test/ctorleak This patch fixes this leak by adding some try-catch-delete-rethrow blocks to `Eigen/src/Core/util/Memory.h`. --- Eigen/src/Core/util/Memory.h | 83 ++++++++++++++++++++++++++++++------ test/CMakeLists.txt | 2 + test/ctorleak.cpp | 43 +++++++++++++++++++ 3 files changed, 114 insertions(+), 14 deletions(-) create mode 100644 test/ctorleak.cpp diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index dd9285714..4b3dcde3a 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -338,15 +338,6 @@ template<> inline void* conditional_aligned_realloc(void* ptr, size_t new *** Construction/destruction of array elements *** *****************************************************************************/ -/** \internal Constructs the elements of an array. - * The \a size parameter tells on how many objects to call the constructor of T. - */ -template inline T* construct_elements_of_array(T *ptr, size_t size) -{ - for (size_t i=0; i < size; ++i) ::new (ptr + i) T; - return ptr; -} - /** \internal Destructs the elements of an array. * The \a size parameters tells on how many objects to call the destructor of T. */ @@ -357,6 +348,24 @@ template inline void destruct_elements_of_array(T *ptr, size_t size) while(size) ptr[--size].~T(); } +/** \internal Constructs the elements of an array. + * The \a size parameter tells on how many objects to call the constructor of T. + */ +template inline T* construct_elements_of_array(T *ptr, size_t size) +{ + size_t i; + try + { + for (i = 0; i < size; ++i) ::new (ptr + i) T; + return ptr; + } + catch (...) + { + destruct_elements_of_array(ptr, i); + throw; + } +} + /***************************************************************************** *** Implementation of aligned new/delete-like functions *** *****************************************************************************/ @@ -376,14 +385,30 @@ template inline T* aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(aligned_malloc(sizeof(T)*size)); - return construct_elements_of_array(result, size); + try + { + return construct_elements_of_array(result, size); + } + catch (...) + { + aligned_free(result); + throw; + } } template inline T* conditional_aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); - return construct_elements_of_array(result, size); + try + { + return construct_elements_of_array(result, size); + } + catch (...) + { + conditional_aligned_free(result); + throw; + } } /** \internal Deletes objects constructed with aligned_new @@ -412,7 +437,17 @@ template inline T* conditional_aligned_realloc_new(T* pt destruct_elements_of_array(pts+new_size, old_size-new_size); T *result = reinterpret_cast(conditional_aligned_realloc(reinterpret_cast(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(new_size > old_size) - construct_elements_of_array(result+old_size, new_size-old_size); + { + try + { + construct_elements_of_array(result+old_size, new_size-old_size); + } + catch (...) + { + conditional_aligned_free(result); + throw; + } + } return result; } @@ -422,7 +457,17 @@ template inline T* conditional_aligned_new_auto(size_t s check_size_for_overflow(size); T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); if(NumTraits::RequireInitialization) - construct_elements_of_array(result, size); + { + try + { + construct_elements_of_array(result, size); + } + catch (...) + { + conditional_aligned_free(result); + throw; + } + } return result; } @@ -434,7 +479,17 @@ template inline T* conditional_aligned_realloc_new_auto( destruct_elements_of_array(pts+new_size, old_size-new_size); T *result = reinterpret_cast(conditional_aligned_realloc(reinterpret_cast(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(NumTraits::RequireInitialization && (new_size > old_size)) - construct_elements_of_array(result+old_size, new_size-old_size); + { + try + { + construct_elements_of_array(result+old_size, new_size-old_size); + } + catch (...) + { + conditional_aligned_free(result); + throw; + } + } return result; } diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 4521d07e4..fed5c0e06 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -251,6 +251,8 @@ ei_add_test(bicgstab) ei_add_test(sparselu) ei_add_test(sparseqr) +ei_add_test(ctorleak) + # ei_add_test(denseLM) if(QT4_FOUND) diff --git a/test/ctorleak.cpp b/test/ctorleak.cpp new file mode 100644 index 000000000..72ab94b66 --- /dev/null +++ b/test/ctorleak.cpp @@ -0,0 +1,43 @@ +#include "main.h" + +#include // std::exception + +struct Foo +{ + int dummy; + Foo() { if (!internal::random(0, 10)) throw Foo::Fail(); } + class Fail : public std::exception {}; +}; + +namespace Eigen +{ + template<> + struct NumTraits + { + typedef double Real; + typedef double NonInteger; + typedef double Nested; + enum + { + IsComplex = 0, + IsInteger = 1, + ReadCost = -1, + AddCost = -1, + MulCost = -1, + IsSigned = 1, + RequireInitialization = 1 + }; + static inline Real epsilon() { return 1.0; } + static inline Real dummy_epsilon() { return 0.0; } + }; +} + +void test_ctorleak() +{ + try + { + Matrix m(14, 92); + eigen_assert(false); // not reached + } + catch (const Foo::Fail&) { /* ignore */ } +} From 85777fc131ca8e22305ac4da41e4f18a870c2797 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 6 Jul 2014 13:45:54 +0800 Subject: [PATCH 0594/1103] Mark internal namespace as \internal --- doc/Manual.dox | 2 +- unsupported/doc/Overview.dox | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/doc/Manual.dox b/doc/Manual.dox index d6d4d4689..43af857a5 100644 --- a/doc/Manual.dox +++ b/doc/Manual.dox @@ -160,6 +160,6 @@ namespace Eigen { /** \addtogroup Splines_Module \ingroup Geometry_Reference */ -/** \brief Namespace containing low-level routines from the %Eigen library. */ +/** \internal \brief Namespace containing low-level routines from the %Eigen library. */ namespace internal {} } diff --git a/unsupported/doc/Overview.dox b/unsupported/doc/Overview.dox index ef361b35b..45464a545 100644 --- a/unsupported/doc/Overview.dox +++ b/unsupported/doc/Overview.dox @@ -23,6 +23,6 @@ subject to be included in %Eigen in the future. */ -/// \brief Namespace containing low-level routines from the %Eigen library. +/// \internal \brief Namespace containing low-level routines from the %Eigen library. namespace internal {} } From 2bf58316ee120e52a193d69104c57781f94ff6a5 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Sun, 6 Jul 2014 13:49:43 +0800 Subject: [PATCH 0595/1103] Fix dox at internal::tridiagonal_qr_step --- Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index d10eba201..fc8ecaa6f 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -355,6 +355,7 @@ template class SelfAdjointEigenSolver bool m_eigenvectorsOk; }; +namespace internal { /** \internal * * \eigenvalues_module \ingroup Eigenvalues_Module @@ -371,7 +372,6 @@ template class SelfAdjointEigenSolver * Implemented from Golub's "Matrix Computations", algorithm 8.3.2: * "implicit symmetric QR step with Wilkinson shift" */ -namespace internal { template EIGEN_DEVICE_FUNC static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n); From b9ee880f0786f12701b32ac5b49ff58141d207f6 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Mon, 7 Jul 2014 21:28:00 +0800 Subject: [PATCH 0596/1103] chmod -x Eigen/src/Core/GenericPacketMath.h --- Eigen/src/Core/GenericPacketMath.h | 0 1 file changed, 0 insertions(+), 0 deletions(-) mode change 100755 => 100644 Eigen/src/Core/GenericPacketMath.h diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h old mode 100755 new mode 100644 From e4b6979334312eefeb63eb7941b48cd95ce7d343 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Mon, 7 Jul 2014 21:32:33 +0800 Subject: [PATCH 0597/1103] Find OpenBLAS more aggressively. This made a difference on Fedora 20 --- bench/btl/cmake/FindOPENBLAS.cmake | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/bench/btl/cmake/FindOPENBLAS.cmake b/bench/btl/cmake/FindOPENBLAS.cmake index c76fc251c..2a0919436 100644 --- a/bench/btl/cmake/FindOPENBLAS.cmake +++ b/bench/btl/cmake/FindOPENBLAS.cmake @@ -3,7 +3,7 @@ if (OPENBLAS_LIBRARIES) set(OPENBLAS_FIND_QUIETLY TRUE) endif (OPENBLAS_LIBRARIES) -find_file(OPENBLAS_LIBRARIES libopenblas.so PATHS /usr/lib $ENV{OPENBLASDIR} ${LIB_INSTALL_DIR}) +find_file(OPENBLAS_LIBRARIES NAMES libopenblas.so libopenblas.so.0 PATHS /usr/lib /usr/lib64 $ENV{OPENBLASDIR} ${LIB_INSTALL_DIR}) find_library(OPENBLAS_LIBRARIES openblas PATHS $ENV{OPENBLASDIR} ${LIB_INSTALL_DIR}) if(OPENBLAS_LIBRARIES AND CMAKE_COMPILER_IS_GNUCXX) From 1eefa5a84192da236df601f067eb90e5bd5a9379 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Mon, 7 Jul 2014 22:55:28 +0800 Subject: [PATCH 0598/1103] Find benchmark opponents also in /usr/lib64 --- bench/btl/cmake/FindACML.cmake | 2 ++ bench/btl/cmake/FindATLAS.cmake | 13 ++++--------- bench/btl/cmake/FindCBLAS.cmake | 1 + 3 files changed, 7 insertions(+), 9 deletions(-) diff --git a/bench/btl/cmake/FindACML.cmake b/bench/btl/cmake/FindACML.cmake index f45ae1b0d..4989fa2f4 100644 --- a/bench/btl/cmake/FindACML.cmake +++ b/bench/btl/cmake/FindACML.cmake @@ -17,6 +17,7 @@ find_file(ACML_LIBRARIES libacml_mp.so PATHS /usr/lib + /usr/lib64 $ENV{ACMLDIR}/lib ${LIB_INSTALL_DIR} ) @@ -35,6 +36,7 @@ if(NOT ACML_LIBRARIES) libacml.so libacml_mv.so PATHS /usr/lib + /usr/lib64 $ENV{ACMLDIR}/lib ${LIB_INSTALL_DIR} ) diff --git a/bench/btl/cmake/FindATLAS.cmake b/bench/btl/cmake/FindATLAS.cmake index 14b1dee09..4136a989d 100644 --- a/bench/btl/cmake/FindATLAS.cmake +++ b/bench/btl/cmake/FindATLAS.cmake @@ -3,18 +3,13 @@ if (ATLAS_LIBRARIES) set(ATLAS_FIND_QUIETLY TRUE) endif (ATLAS_LIBRARIES) -find_file(ATLAS_LIB libatlas.so.3 PATHS /usr/lib $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) +find_file(ATLAS_LIB libatlas.so.3 PATHS /usr/lib /usr/lib/atlas /usr/lib64 /usr/lib64/atlas $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) find_library(ATLAS_LIB satlas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) -find_file(ATLAS_LAPACK liblapack_atlas.so.3 PATHS /usr/lib $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) -find_library(ATLAS_LAPACK lapack_atlas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) +find_file(ATLAS_LAPACK NAMES liblapack_atlas.so.3 liblapack.so.3 PATHS /usr/lib /usr/lib/atlas /usr/lib64 /usr/lib64/atlas $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) +find_library(ATLAS_LAPACK NAMES lapack_atlas lapack PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) -if(NOT ATLAS_LAPACK) - find_file(ATLAS_LAPACK liblapack.so.3 PATHS /usr/lib/atlas $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) - find_library(ATLAS_LAPACK lapack PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) -endif(NOT ATLAS_LAPACK) - -find_file(ATLAS_F77BLAS libf77blas.so.3 PATHS /usr/lib $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) +find_file(ATLAS_F77BLAS libf77blas.so.3 PATHS /usr/lib /usr/lib/atlas /usr/lib64 /usr/lib64/atlas $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) find_library(ATLAS_F77BLAS f77blas PATHS $ENV{ATLASDIR} ${LIB_INSTALL_DIR}) if(ATLAS_LIB AND ATLAS_CBLAS AND ATLAS_LAPACK AND ATLAS_F77BLAS) diff --git a/bench/btl/cmake/FindCBLAS.cmake b/bench/btl/cmake/FindCBLAS.cmake index 554f0291b..ce0f2f2b2 100644 --- a/bench/btl/cmake/FindCBLAS.cmake +++ b/bench/btl/cmake/FindCBLAS.cmake @@ -23,6 +23,7 @@ find_file(CBLAS_LIBRARIES libcblas.so.3 PATHS /usr/lib + /usr/lib64 $ENV{CBLASDIR}/lib ${LIB_INSTALL_DIR} ) From 47981c5925caa8316205ea84b17616dd69073678 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 7 Jul 2014 14:07:57 -0700 Subject: [PATCH 0599/1103] Added support for tensor slicing --- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 343 +++++++++++++++++- 1 file changed, 327 insertions(+), 16 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 764bba4e6..55954a3a7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -20,10 +20,9 @@ namespace Eigen { * */ namespace internal { -template -struct traits > : public traits +template +struct traits > : public traits { - // Type promotion to handle the case where the types of the lhs and the rhs are different. typedef typename XprType::Scalar Scalar; typedef typename internal::packet_traits::type Packet; typedef typename traits::StorageKind StorageKind; @@ -32,24 +31,24 @@ struct traits > : public traits::type _Nested; }; -template -struct eval, Eigen::Dense> +template +struct eval, Eigen::Dense> { - typedef const TensorReshapingOp& type; + typedef const TensorReshapingOp& type; }; -template -struct nested, 1, typename eval >::type> +template +struct nested, 1, typename eval >::type> { - typedef TensorReshapingOp type; + typedef TensorReshapingOp type; }; } // end namespace internal -template -class TensorReshapingOp : public TensorBase > +template +class TensorReshapingOp : public TensorBase, WriteAccessors> { public: typedef typename Eigen::internal::traits::Scalar Scalar; @@ -71,16 +70,27 @@ class TensorReshapingOp : public TensorBase::type& expression() const { return m_xpr; } + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorReshapingOp& operator = (const OtherDerived& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + protected: typename XprType::Nested m_xpr; const NewDimensions m_dims; }; -template -struct TensorEvaluator, Device> +// Eval as rvalue +template +struct TensorEvaluator, Device> { - typedef TensorReshapingOp XprType; + typedef TensorReshapingOp XprType; typedef NewDimensions Dimensions; enum { @@ -88,7 +98,7 @@ struct TensorEvaluator, Device> PacketAccess = TensorEvaluator::PacketAccess, }; - TensorEvaluator(const XprType& op, const Device& device) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_impl(op.expression(), device), m_dimensions(op.dimensions()) { } @@ -96,7 +106,7 @@ struct TensorEvaluator, Device> typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; - const Dimensions& dimensions() const { return m_dimensions; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { m_impl.evalSubExprsIfNeeded(); @@ -116,12 +126,313 @@ struct TensorEvaluator, Device> return m_impl.template packet(index); } + protected: + NewDimensions m_dimensions; + TensorEvaluator m_impl; +}; + + +// Eval as lvalue +// TODO(bsteiner): share the code with the evaluator for rvalue reshapes. +template +struct TensorEvaluator, Device> +{ + typedef TensorReshapingOp XprType; + typedef NewDimensions Dimensions; + + enum { + IsAligned = TensorEvaluator::IsAligned, + PacketAccess = TensorEvaluator::PacketAccess, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_dimensions(op.dimensions()) + { } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_impl.evalSubExprsIfNeeded(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + return m_impl.coeff(index); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) + { + return m_impl.coeffRef(index); + } + template EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketReturnType& x) + { + m_impl.template writePacket(index, x); + } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return m_impl.template packet(index); + } + private: NewDimensions m_dimensions; TensorEvaluator m_impl; }; +/** \class TensorSlicing + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor slicing class. + * + * + */ +namespace internal { +template +struct traits > : public traits +{ + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorSlicingOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorSlicingOp type; +}; + +} // end namespace internal + + + +template +class TensorSlicingOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorSlicingOp(const XprType& expr, const StartIndices& indices, const Sizes& sizes) + : m_xpr(expr), m_indices(indices), m_sizes(sizes) {} + + EIGEN_DEVICE_FUNC + const StartIndices& startIndices() const { return m_indices; } + EIGEN_DEVICE_FUNC + const Sizes& sizes() const { return m_sizes; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorSlicingOp& operator = (const OtherDerived& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + + protected: + typename XprType::Nested m_xpr; + const StartIndices m_indices; + const Sizes m_sizes; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorSlicingOp XprType; + static const int NumDims = internal::array_size::value; + + enum { + IsAligned = TensorEvaluator::IsAligned, + PacketAccess = /*TensorEvaluator::PacketAccess*/false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_dimensions(op.sizes()), m_offsets(op.startIndices()) + { + for (int i = 0; i < internal::array_size::value; ++i) { + eigen_assert(m_impl.dimensions()[i] >= op.sizes()[i] + op.startIndices()[i]); + } + + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + } else { + m_inputStrides[0] = 1; + } + } + + const Sizes& output_dims = op.sizes(); + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1]; + } else { + m_outputStrides[0] = 1; + } + } + } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef Sizes Dimensions; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_impl.evalSubExprsIfNeeded(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i >= 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + return m_impl.coeff(inputIndex); + } + + /* template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return m_impl.template packet(index); + }*/ + + private: + Dimensions m_dimensions; + array m_outputStrides; + array m_inputStrides; + const StartIndices m_offsets; + TensorEvaluator m_impl; +}; + + +// Eval as lvalue +// TODO(bsteiner): share the code with the evaluator for rvalue slices. +template +struct TensorEvaluator, Device> +{ + typedef TensorSlicingOp XprType; + static const int NumDims = internal::array_size::value; + + enum { + IsAligned = TensorEvaluator::IsAligned, + PacketAccess = /*TensorEvaluator::PacketAccess*/false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_dimensions(op.sizes()), m_offsets(op.startIndices()) + { + for (int i = 0; i < internal::array_size::value; ++i) { + eigen_assert(m_impl.dimensions()[i] >= op.sizes()[i] + op.startIndices()[i]); + } + + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + } else { + m_inputStrides[0] = 1; + } + } + + const Sizes& output_dims = op.sizes(); + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1]; + } else { + m_outputStrides[0] = 1; + } + } + } + + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef Sizes Dimensions; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { + m_impl.evalSubExprsIfNeeded(); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i >= 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + return m_impl.coeff(inputIndex); + } + + /* template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return m_impl.template packet(index); + }*/ + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) + { + Index inputIndex = 0; + for (int i = NumDims - 1; i >= 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + return m_impl.coeffRef(inputIndex); + } + + private: + Dimensions m_dimensions; + array m_outputStrides; + array m_inputStrides; + const StartIndices m_offsets; + TensorEvaluator m_impl; +}; + + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H From bc072c5cba4cb6e9e7a6fd5f1e8f0e1231203223 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 7 Jul 2014 14:08:45 -0700 Subject: [PATCH 0600/1103] Added support for tensor slicing --- .../Eigen/CXX11/src/Tensor/TensorBase.h | 22 ++++++++++++++++--- 1 file changed, 19 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 6b53d2a3d..527d47c57 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -204,11 +204,16 @@ class TensorBase return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); } - // Morphing operators (slicing tbd). + // Morphing operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - const TensorReshapingOp + const TensorReshapingOp reshape(const NewDimensions& newDimensions) const { - return TensorReshapingOp(derived(), newDimensions); + return TensorReshapingOp(derived(), newDimensions); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorSlicingOp + slice(const StartIndices& startIndices, const Sizes& sizes) const { + return TensorSlicingOp(derived(), startIndices, sizes); } // Force the evaluation of the expression. @@ -257,6 +262,17 @@ class TensorBase : public TensorBase, const Derived, const OtherDerived>(derived(), other.derived()); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorReshapingOp + reshape(const NewDimensions& newDimensions) { + return TensorReshapingOp(derived(), newDimensions); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorSlicingOp + slice(const StartIndices& startIndices, const Sizes& sizes) const { + return TensorSlicingOp(derived(), startIndices, sizes); + } + // Select the device on which to evaluate the expression. template TensorDevice device(const DeviceType& device) { From 7d53633e05986c61ce90e7fc36862d529c0cc036 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 7 Jul 2014 14:10:36 -0700 Subject: [PATCH 0601/1103] Added support for tensor slicing --- unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h | 1 + 1 file changed, 1 insertion(+) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index c0dffbd0c..5d6e7776a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -25,6 +25,7 @@ template class TensorReductionOp; template class TensorContractionOp; template class TensorConvolutionOp; template class TensorReshapingOp; +template class TensorSlicingOp; template class TensorAssignOp; template class TensorEvalToOp; From 7fa83e7374b38e9900e77cf2b29b54919f65d1ef Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 09:56:09 +0200 Subject: [PATCH 0602/1103] Fix LDLT with semi-definite complex matrices: owing to round-off errors, the diagonal was not real. Also exploit the fact that the diagonal is real in the rest of LDLT From 0dfb73d46ada6a9749a24a946186c8a8c2472dc5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 10:04:27 +0200 Subject: [PATCH 0603/1103] Fix LDLT with semi-definite complex matrices: owing to round-off errors, the diagonal was not real. Also exploit the fact that the diagonal is real in the rest of LDLT --- Eigen/src/Cholesky/LDLT.h | 12 ++++++------ test/cholesky.cpp | 2 +- 2 files changed, 7 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index ef81030f8..6881e1ca8 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -295,7 +295,7 @@ template<> struct ldlt_inplace if(k>0) { - temp.head(k) = mat.diagonal().head(k).asDiagonal() * A10.adjoint(); + temp.head(k) = mat.diagonal().real().head(k).asDiagonal() * A10.adjoint(); mat.coeffRef(k,k) -= (A10 * temp.head(k)).value(); if(rs>0) A21.noalias() -= A20 * temp.head(k); @@ -305,10 +305,10 @@ template<> struct ldlt_inplace // was smaller than the cutoff value. However, soince LDLT is not rank-revealing // we should only make sure we do not introduce INF or NaN values. // LAPACK also uses 0 as the cutoff value. - if((rs>0) && (abs(mat.coeffRef(k,k)) > RealScalar(0))) - A21 /= mat.coeffRef(k,k); - RealScalar realAkk = numext::real(mat.coeffRef(k,k)); + if((rs>0) && (abs(realAkk) > RealScalar(0))) + A21 /= realAkk; + if (sign == PositiveSemiDef) { if (realAkk < 0) sign = Indefinite; } else if (sign == NegativeSemiDef) { @@ -487,7 +487,7 @@ struct solve_retval, Rhs> EIGEN_USING_STD_MATH(max); typedef typename LDLTType::MatrixType MatrixType; typedef typename LDLTType::RealScalar RealScalar; - const Diagonal vectorD = dec().vectorD(); + const typename Diagonal::RealReturnType vectorD(dec().vectorD()); // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon // as motivated by LAPACK's xGELSS: // RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() *NumTraits::epsilon(),RealScalar(1) / NumTraits::highest()); @@ -552,7 +552,7 @@ MatrixType LDLT::reconstructedMatrix() const // L^* P res = matrixU() * res; // D(L^*P) - res = vectorD().asDiagonal() * res; + res = vectorD().real().asDiagonal() * res; // L(DL^*P) res = matrixL() * res; // P^T (LDL^*P) diff --git a/test/cholesky.cpp b/test/cholesky.cpp index 9aa7da7bb..1d147bd7a 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -209,7 +209,7 @@ template void cholesky(const MatrixType& m) VERIFY_IS_APPROX(A * vecX, vecB); } - // check matrices with wide spectrum + // check matrices with a wide spectrum if(rows>=3) { RealScalar s = (std::min)(16,std::numeric_limits::max_exponent10/8); From 904509fbb664b3165531e6e4d7234d0765425ad9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 10:28:09 +0200 Subject: [PATCH 0604/1103] Move using std::abs from Eigen's namespace to function scope. --- unsupported/Eigen/src/IterativeSolvers/Scaling.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/IterativeSolvers/Scaling.h b/unsupported/Eigen/src/IterativeSolvers/Scaling.h index 9189fddf1..d113e6e90 100644 --- a/unsupported/Eigen/src/IterativeSolvers/Scaling.h +++ b/unsupported/Eigen/src/IterativeSolvers/Scaling.h @@ -11,7 +11,6 @@ #define EIGEN_ITERSCALING_H namespace Eigen { -using std::abs; /** * \ingroup IterativeSolvers_Module @@ -73,6 +72,7 @@ class IterScaling */ void compute (const MatrixType& mat) { + using std::abs; int m = mat.rows(); int n = mat.cols(); eigen_assert((m>0 && m == n) && "Please give a non - empty matrix"); From e25e6748521818fb13270f7351df8538eb54a056 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 8 Jul 2014 13:57:26 +0200 Subject: [PATCH 0605/1103] bug #837: Always re-align the result of EIGEN_ALLOCA. --- Eigen/src/Core/util/Memory.h | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 9718ef6a8..dd9285714 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -607,12 +607,9 @@ template class aligned_stack_memory_handler * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token. */ #ifdef EIGEN_ALLOCA - // The native alloca() that comes with llvm aligns buffer on 16 bytes even when AVX is enabled. -#if defined(__arm__) || defined(_WIN32) || EIGEN_ALIGN_BYTES > 16 - #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast((reinterpret_cast(EIGEN_ALLOCA(SIZE+EIGEN_ALIGN_BYTES)) & ~(size_t(EIGEN_ALIGN_BYTES-1))) + EIGEN_ALIGN_BYTES) - #else - #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA - #endif + // We always manually re-align the result of EIGEN_ALLOCA. + // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment. + #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast((reinterpret_cast(EIGEN_ALLOCA(SIZE+EIGEN_ALIGN_BYTES-1)) + EIGEN_ALIGN_BYTES-1) & ~(size_t(EIGEN_ALIGN_BYTES-1))) #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \ Eigen::internal::check_size_for_overflow(SIZE); \ From b47ef1431f2e4d1218253df179b593cc5c269aa7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 16:47:11 +0200 Subject: [PATCH 0606/1103] Fix many long to int implicit conversions --- Eigen/src/Core/GeneralProduct.h | 4 +- Eigen/src/Core/GenericPacketMath.h | 4 +- Eigen/src/Core/Visitor.h | 2 +- Eigen/src/Core/arch/AVX/Complex.h | 4 +- Eigen/src/Core/arch/AVX/PacketMath.h | 8 ++-- Eigen/src/Core/arch/AltiVec/Complex.h | 4 +- Eigen/src/Core/arch/AltiVec/PacketMath.h | 8 ++-- Eigen/src/Core/arch/NEON/Complex.h | 4 +- Eigen/src/Core/arch/NEON/PacketMath.h | 8 ++-- Eigen/src/Core/arch/SSE/Complex.h | 4 +- Eigen/src/Core/arch/SSE/PacketMath.h | 12 +++--- .../Core/products/SelfadjointMatrixVector.h | 4 +- .../Core/products/TriangularMatrixVector.h | 2 +- Eigen/src/SparseCore/SparseBlock.h | 40 ++++++++++--------- Eigen/src/SparseCore/TriangularSolver.h | 30 ++++++++------ test/cholesky.cpp | 6 +-- test/mapstaticmethods.cpp | 6 ++- test/product_notemporary.cpp | 5 +-- test/ref.cpp | 5 +-- test/sparse.h | 4 +- test/sparse_basic.cpp | 8 ++-- 21 files changed, 89 insertions(+), 83 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 229d12c3f..624b8b6e8 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -445,7 +445,7 @@ template<> struct gemv_selector if(!evalToDest) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = dest.size(); + Index size = dest.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif if(!alphaIsCompatible) @@ -510,7 +510,7 @@ template<> struct gemv_selector if(!DirectlyUseRhs) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = actualRhs.size(); + Index size = actualRhs.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif Map(actualRhsPtr, actualRhs.size()) = actualRhs; diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 4523d2263..a1fcb82fb 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -233,10 +233,10 @@ template EIGEN_DEVICE_FUNC inline void pstore( template EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from) { (*to) = from; } - template EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, int /*stride*/) + template EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, DenseIndex /*stride*/) { return ploadu(from); } - template EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, int /*stride*/) + template EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, DenseIndex /*stride*/) { pstore(to, from); } /** \internal tries to do cache prefetching of \a addr */ diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h index 64867b7a2..6f4b9ec35 100644 --- a/Eigen/src/Core/Visitor.h +++ b/Eigen/src/Core/Visitor.h @@ -194,7 +194,7 @@ DenseBase::minCoeff(IndexType* index) const EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived) internal::min_coeff_visitor minVisitor; this->visit(minVisitor); - *index = (RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row; + *index = IndexType((RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row); return minVisitor.res; } diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index 9ced85132..b3cf7bb26 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -92,7 +92,7 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } -template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packet4cf>(const std::complex* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packet4cf>(const std::complex* from, DenseIndex stride) { return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]), std::imag(from[2*stride]), std::real(from[2*stride]), @@ -100,7 +100,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packe std::imag(from[0*stride]), std::real(from[0*stride]))); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet4cf>(std::complex* to, const Packet4cf& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet4cf>(std::complex* to, const Packet4cf& from, DenseIndex stride) { __m128 low = _mm256_extractf128_ps(from.v, 0); to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)), diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 8b8307d75..66b97bd69 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -224,17 +224,17 @@ template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet8i& // NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available // NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4); -template<> EIGEN_DEVICE_FUNC inline Packet8f pgather(const float* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet8f pgather(const float* from, DenseIndex stride) { return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride], from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline Packet4d pgather(const double* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4d pgather(const double* from, DenseIndex stride) { return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet8f& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet8f& from, DenseIndex stride) { __m128 low = _mm256_extractf128_ps(from, 0); to[stride*0] = _mm_cvtss_f32(low); @@ -248,7 +248,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, co to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)); to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet4d& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet4d& from, DenseIndex stride) { __m128d low = _mm256_extractf128_pd(from, 0); to[stride*0] = _mm_cvtsd_f64(low); diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 5409ddedd..73fa62643 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -68,14 +68,14 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, DenseIndex stride) { std::complex EIGEN_ALIGN16 af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return Packet2cf(vec_ld(0, (const float*)af)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, DenseIndex stride) { std::complex EIGEN_ALIGN16 af[2]; vec_st(from.v, 0, (float*)af); diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 0e9adf450..aadfc17be 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -190,7 +190,7 @@ pbroadcast4(const int *a, a3 = vec_splat(a3, 3); } -template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, DenseIndex stride) { float EIGEN_ALIGN16 af[4]; af[0] = from[0*stride]; @@ -199,7 +199,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const floa af[3] = from[3*stride]; return vec_ld(0, af); } -template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, DenseIndex stride) { int EIGEN_ALIGN16 ai[4]; ai[0] = from[0*stride]; @@ -208,7 +208,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* f ai[3] = from[3*stride]; return vec_ld(0, ai); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, DenseIndex stride) { float EIGEN_ALIGN16 af[4]; vec_st(from, 0, af); @@ -217,7 +217,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, co to[2*stride] = af[2]; to[3*stride] = af[3]; } -template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, DenseIndex stride) { int EIGEN_ALIGN16 ai[4]; vec_st(from, 0, ai); diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 259f2e7b8..42e7733d7 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -111,7 +111,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } -template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, DenseIndex stride) { Packet4f res; res = vsetq_lane_f32(std::real(from[0*stride]), res, 0); @@ -121,7 +121,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packe return Packet2cf(res); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, DenseIndex stride) { to[stride*0] = std::complex(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1)); to[stride*1] = std::complex(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3)); diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index e5eb06f36..380b76ae9 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -222,7 +222,7 @@ template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& f template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); } -template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, DenseIndex stride) { Packet4f res; res = vsetq_lane_f32(from[0*stride], res, 0); @@ -231,7 +231,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const floa res = vsetq_lane_f32(from[3*stride], res, 3); return res; } -template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, DenseIndex stride) { Packet4i res; res = vsetq_lane_s32(from[0*stride], res, 0); @@ -241,14 +241,14 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* f return res; } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, DenseIndex stride) { to[stride*0] = vgetq_lane_f32(from, 0); to[stride*1] = vgetq_lane_f32(from, 1); to[stride*2] = vgetq_lane_f32(from, 2); to[stride*3] = vgetq_lane_f32(from, 3); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, DenseIndex stride) { to[stride*0] = vgetq_lane_s32(from, 0); to[stride*1] = vgetq_lane_s32(from, 1); diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 758183c18..414c4cb6a 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -114,13 +114,13 @@ template<> EIGEN_STRONG_INLINE void pstore >(std::complex EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); } -template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, DenseIndex stride) { return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]), std::imag(from[0*stride]), std::real(from[0*stride]))); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, DenseIndex stride) { to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)), _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1))); diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 6912f3bc3..380afe77c 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -383,32 +383,32 @@ template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), Packet2d(_mm_castps_pd(from))); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), Packet2d(_mm_castsi128_pd(from))); } -template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, DenseIndex stride) { return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, DenseIndex stride) { return _mm_set_pd(from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, DenseIndex stride) { return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, DenseIndex stride) { to[stride*0] = _mm_cvtss_f32(from); to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1)); to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2)); to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, DenseIndex stride) { to[stride*0] = _mm_cvtsd_f64(from); to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, DenseIndex stride) { to[stride*0] = _mm_cvtsi128_si32(from); to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1)); diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index fdc81205a..26e787949 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -218,7 +218,7 @@ struct SelfadjointProductMatrix if(!EvalToDest) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = dest.size(); + Index size = dest.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif MappedDest(actualDestPtr, dest.size()) = dest; @@ -227,7 +227,7 @@ struct SelfadjointProductMatrix if(!UseRhs) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = rhs.size(); + Index size = rhs.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif Map(actualRhsPtr, rhs.size()) = rhs; diff --git a/Eigen/src/Core/products/TriangularMatrixVector.h b/Eigen/src/Core/products/TriangularMatrixVector.h index 6117d5a82..817768481 100644 --- a/Eigen/src/Core/products/TriangularMatrixVector.h +++ b/Eigen/src/Core/products/TriangularMatrixVector.h @@ -322,7 +322,7 @@ template<> struct trmv_selector if(!DirectlyUseRhs) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = actualRhs.size(); + Index size = actualRhs.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif Map(actualRhsPtr, actualRhs.size()) = actualRhs; diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 5b95cc33f..cbcd7a299 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -50,11 +50,11 @@ public: Index m_outer; }; - inline BlockImpl(const XprType& xpr, int i) + inline BlockImpl(const XprType& xpr, Index i) : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) {} - inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) + inline BlockImpl(const XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) : m_matrix(xpr), m_outerStart(IsRowMajor ? startRow : startCol), m_outerSize(IsRowMajor ? blockRows : blockCols) {} @@ -65,7 +65,7 @@ public: { Index nnz = 0; Index end = m_outerStart + m_outerSize.value(); - for(int j=m_outerStart; j,BlockRows,BlockCols,true : public internal::sparse_matrix_block_impl,BlockRows,BlockCols> { public: + typedef _Index Index; typedef SparseMatrix<_Scalar, _Options, _Index> SparseMatrixType; typedef internal::sparse_matrix_block_impl Base; - inline BlockImpl(SparseMatrixType& xpr, int i) + inline BlockImpl(SparseMatrixType& xpr, Index i) : Base(xpr, i) {} - inline BlockImpl(SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) + inline BlockImpl(SparseMatrixType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) : Base(xpr, startRow, startCol, blockRows, blockCols) {} @@ -282,13 +283,14 @@ class BlockImpl,BlockRows,BlockCol : public internal::sparse_matrix_block_impl,BlockRows,BlockCols> { public: + typedef _Index Index; typedef const SparseMatrix<_Scalar, _Options, _Index> SparseMatrixType; typedef internal::sparse_matrix_block_impl Base; - inline BlockImpl(SparseMatrixType& xpr, int i) + inline BlockImpl(SparseMatrixType& xpr, Index i) : Base(xpr, i) {} - inline BlockImpl(SparseMatrixType& xpr, int startRow, int startCol, int blockRows, int blockCols) + inline BlockImpl(SparseMatrixType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) : Base(xpr, startRow, startCol, blockRows, blockCols) {} @@ -362,7 +364,7 @@ public: /** Column or Row constructor */ - inline BlockImpl(const XprType& xpr, int i) + inline BlockImpl(const XprType& xpr, Index i) : m_matrix(xpr), m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0), m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0), @@ -372,32 +374,32 @@ public: /** Dynamic-size constructor */ - inline BlockImpl(const XprType& xpr, int startRow, int startCol, int blockRows, int blockCols) + inline BlockImpl(const XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols) : m_matrix(xpr), m_startRow(startRow), m_startCol(startCol), m_blockRows(blockRows), m_blockCols(blockCols) {} - inline int rows() const { return m_blockRows.value(); } - inline int cols() const { return m_blockCols.value(); } + inline Index rows() const { return m_blockRows.value(); } + inline Index cols() const { return m_blockCols.value(); } - inline Scalar& coeffRef(int row, int col) + inline Scalar& coeffRef(Index row, Index col) { return m_matrix.const_cast_derived() .coeffRef(row + m_startRow.value(), col + m_startCol.value()); } - inline const Scalar coeff(int row, int col) const + inline const Scalar coeff(Index row, Index col) const { return m_matrix.coeff(row + m_startRow.value(), col + m_startCol.value()); } - inline Scalar& coeffRef(int index) + inline Scalar& coeffRef(Index index) { return m_matrix.const_cast_derived() .coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); } - inline const Scalar coeff(int index) const + inline const Scalar coeff(Index index) const { return m_matrix .coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index), diff --git a/Eigen/src/SparseCore/TriangularSolver.h b/Eigen/src/SparseCore/TriangularSolver.h index f958ab300..dd55522a7 100644 --- a/Eigen/src/SparseCore/TriangularSolver.h +++ b/Eigen/src/SparseCore/TriangularSolver.h @@ -28,15 +28,16 @@ template struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; static void run(const Lhs& lhs, Rhs& other) { - for(int col=0 ; col struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; static void run(const Lhs& lhs, Rhs& other) { - for(int col=0 ; col=0 ; --i) + for(Index i=lhs.rows()-1 ; i>=0 ; --i) { Scalar tmp = other.coeff(i,col); Scalar l_ii = 0; @@ -100,11 +102,12 @@ template struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; static void run(const Lhs& lhs, Rhs& other) { - for(int col=0 ; col struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; static void run(const Lhs& lhs, Rhs& other) { - for(int col=0 ; col=0; --i) + for(Index i=lhs.cols()-1; i>=0; --i) { Scalar& tmp = other.coeffRef(i,col); if (tmp!=Scalar(0)) // optimization when other is actually sparse @@ -209,7 +213,7 @@ struct sparse_solve_triangular_sparse_selector { typedef typename Rhs::Scalar Scalar; typedef typename promote_index_type::Index, - typename traits::Index>::type Index; + typename traits::Index>::type Index; static void run(const Lhs& lhs, Rhs& other) { const bool IsLower = (UpLo==Lower); @@ -219,7 +223,7 @@ struct sparse_solve_triangular_sparse_selector Rhs res(other.rows(), other.cols()); res.reserve(other.nonZeros()); - for(int col=0 ; col tempVector.coeffRef(rhsIt.index()) = rhsIt.value(); } - for(int i=IsLower?0:lhs.cols()-1; + for(Index i=IsLower?0:lhs.cols()-1; IsLower?i=0; i+=IsLower?1:-1) { @@ -267,7 +271,7 @@ struct sparse_solve_triangular_sparse_selector } - int count = 0; + Index count = 0; // FIXME compute a reference value to filter zeros for (typename AmbiVector::Iterator it(tempVector/*,1e-12*/); it; ++it) { diff --git a/test/cholesky.cpp b/test/cholesky.cpp index 1d147bd7a..a883192ab 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -181,7 +181,7 @@ template void cholesky(const MatrixType& m) if(rows>=3) { SquareMatrixType A = symm; - int c = internal::random(0,rows-2); + Index c = internal::random(0,rows-2); A.bottomRightCorner(c,c).setZero(); // Make sure a solution exists: vecX.setRandom(); @@ -196,7 +196,7 @@ template void cholesky(const MatrixType& m) // check non-full rank matrices if(rows>=3) { - int r = internal::random(1,rows-1); + Index r = internal::random(1,rows-1); Matrix a = Matrix::Random(rows,r); SquareMatrixType A = a * a.adjoint(); // Make sure a solution exists: @@ -215,7 +215,7 @@ template void cholesky(const MatrixType& m) RealScalar s = (std::min)(16,std::numeric_limits::max_exponent10/8); Matrix a = Matrix::Random(rows,rows); Matrix d = Matrix::Random(rows); - for(int k=0; k(-s,s)); SquareMatrixType A = a * d.asDiagonal() * a.adjoint(); // Make sure a solution exists: diff --git a/test/mapstaticmethods.cpp b/test/mapstaticmethods.cpp index 5b512bde4..06272d106 100644 --- a/test/mapstaticmethods.cpp +++ b/test/mapstaticmethods.cpp @@ -69,7 +69,8 @@ struct mapstaticmethods_impl { static void run(const PlainObjectType& m) { - int rows = m.rows(), cols = m.cols(); + typedef typename PlainObjectType::Index Index; + Index rows = m.rows(), cols = m.cols(); int i = internal::random(2,5), j = internal::random(2,5); @@ -115,7 +116,8 @@ struct mapstaticmethods_impl { static void run(const PlainObjectType& v) { - int size = v.size(); + typedef typename PlainObjectType::Index Index; + Index size = v.size(); int i = internal::random(2,5); diff --git a/test/product_notemporary.cpp b/test/product_notemporary.cpp index 258d238e2..3a9df618b 100644 --- a/test/product_notemporary.cpp +++ b/test/product_notemporary.cpp @@ -7,13 +7,12 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. -static int nb_temporaries; +static long int nb_temporaries; -inline void on_temporary_creation(int size) { +inline void on_temporary_creation(long int size) { // here's a great place to set a breakpoint when debugging failures in this test! if(size!=0) nb_temporaries++; } - #define EIGEN_DENSE_STORAGE_CTOR_PLUGIN { on_temporary_creation(size); } diff --git a/test/ref.cpp b/test/ref.cpp index 19e81549c..d91e3b54c 100644 --- a/test/ref.cpp +++ b/test/ref.cpp @@ -12,13 +12,12 @@ #undef EIGEN_DEFAULT_TO_ROW_MAJOR #endif -static int nb_temporaries; +static long int nb_temporaries; -inline void on_temporary_creation(int) { +inline void on_temporary_creation(long int) { // here's a great place to set a breakpoint when debugging failures in this test! nb_temporaries++; } - #define EIGEN_DENSE_STORAGE_CTOR_PLUGIN { on_temporary_creation(size); } diff --git a/test/sparse.h b/test/sparse.h index e19a76316..81ab9e873 100644 --- a/test/sparse.h +++ b/test/sparse.h @@ -71,7 +71,7 @@ initSparse(double density, //sparseMat.startVec(j); for(Index i=0; i(0,1) < density) ? internal::random() : Scalar(0); @@ -163,7 +163,7 @@ initSparse(double density, { sparseVec.reserve(int(refVec.size()*density)); sparseVec.setZero(); - for(Index i=0; i(0,1) < density) ? internal::random() : Scalar(0); if (v!=Scalar(0)) diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index 6c620f037..4c9b9111e 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -147,7 +147,7 @@ template void sparse_basic(const SparseMatrixType& re DenseMatrix m1(rows,cols); m1.setZero(); SparseMatrixType m2(rows,cols); - VectorXi r(VectorXi::Constant(m2.outerSize(), ((mode%2)==0) ? m2.innerSize() : std::max(1,m2.innerSize()/8))); + VectorXi r(VectorXi::Constant(m2.outerSize(), ((mode%2)==0) ? int(m2.innerSize()) : std::max(1,int(m2.innerSize())/8))); m2.reserve(r); for (int k=0; k void sparse_basic(const SparseMatrixType& re VERIFY_IS_APPROX(m2.innerVector(j0)+m2.innerVector(j1), refMat2.col(j0)+refMat2.col(j1)); SparseMatrixType m3(rows,rows); - m3.reserve(VectorXi::Constant(rows,rows/2)); + m3.reserve(VectorXi::Constant(rows,int(rows/2))); for(Index j=0; j void sparse_basic(const SparseMatrixType& re { typedef Triplet TripletType; std::vector triplets; - int ntriplets = rows*cols; + Index ntriplets = rows*cols; triplets.reserve(ntriplets); DenseMatrix refMat(rows,cols); refMat.setZero(); - for(int i=0;i(0,rows-1); Index c = internal::random(0,cols-1); From 5c4733f6e41dc09a1d700c780c8a3492906e127a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 18:38:34 +0200 Subject: [PATCH 0607/1103] Fix bug #809: unused variable warning --- Eigen/src/OrderingMethods/Ordering.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/OrderingMethods/Ordering.h b/Eigen/src/OrderingMethods/Ordering.h index 4e0609784..f3c31f9cb 100644 --- a/Eigen/src/OrderingMethods/Ordering.h +++ b/Eigen/src/OrderingMethods/Ordering.h @@ -136,12 +136,12 @@ class COLAMDOrdering Index stats [COLAMD_STATS]; internal::colamd_set_defaults(knobs); - Index info; IndexVector p(n+1), A(Alen); for(Index i=0; i <= n; i++) p(i) = mat.outerIndexPtr()[i]; for(Index i=0; i < nnz; i++) A(i) = mat.innerIndexPtr()[i]; // Call Colamd routine to compute the ordering - info = internal::colamd(m, n, Alen, A.data(), p.data(), knobs, stats); + Index info = internal::colamd(m, n, Alen, A.data(), p.data(), knobs, stats); + EIGEN_UNUSED_VARIABLE(info); eigen_assert( info && "COLAMD failed " ); perm.resize(n); From 5214466b7a38ec7684ea685d5d9c56ae2cae678e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 19:01:49 +0200 Subject: [PATCH 0608/1103] Fix implicit long to int conversions in blas interface --- blas/level1_impl.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/blas/level1_impl.h b/blas/level1_impl.h index 98e4c0963..e623bd178 100644 --- a/blas/level1_impl.h +++ b/blas/level1_impl.h @@ -59,7 +59,7 @@ int EIGEN_CAT(EIGEN_CAT(i,SCALAR_SUFFIX),amax_)(int *n, RealScalar *px, int *inc DenseIndex ret; if(*incx==1) make_vector(x,*n).cwiseAbs().maxCoeff(&ret); else make_vector(x,*n,std::abs(*incx)).cwiseAbs().maxCoeff(&ret); - return ret+1; + return int(ret)+1; } int EIGEN_CAT(EIGEN_CAT(i,SCALAR_SUFFIX),amin_)(int *n, RealScalar *px, int *incx) @@ -70,7 +70,7 @@ int EIGEN_CAT(EIGEN_CAT(i,SCALAR_SUFFIX),amin_)(int *n, RealScalar *px, int *inc DenseIndex ret; if(*incx==1) make_vector(x,*n).cwiseAbs().minCoeff(&ret); else make_vector(x,*n,std::abs(*incx)).cwiseAbs().minCoeff(&ret); - return ret+1; + return int(ret)+1; } int EIGEN_BLAS_FUNC(rotg)(RealScalar *pa, RealScalar *pb, RealScalar *pc, RealScalar *ps) From e3557e8dd292e71e033677926cc3ab749d2696b3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 18:58:41 +0200 Subject: [PATCH 0609/1103] bug #808: use double instead of float for the increasing size ratio in CompressedStorage::resize (grafted from 0e0ae400843cd9a459e5306d4d6560dbea759a42 ) --- Eigen/src/SparseCore/CompressedStorage.h | 4 ++-- Eigen/src/SparseCore/SparseMatrix.h | 10 +++++----- 2 files changed, 7 insertions(+), 7 deletions(-) diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index 10d516a66..a667cb56e 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -83,10 +83,10 @@ class CompressedStorage reallocate(m_size); } - void resize(size_t size, float reserveSizeFactor = 0) + void resize(size_t size, double reserveSizeFactor = 0) { if (m_allocatedSize::Scalar& Sparse size_t p = m_outerIndex[outer+1]; ++m_outerIndex[outer+1]; - float reallocRatio = 1; + double reallocRatio = 1; if (m_data.allocatedSize()<=m_data.size()) { // if there is no preallocated memory, let's reserve a minimum of 32 elements @@ -1190,13 +1190,13 @@ EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& Sparse { // we need to reallocate the data, to reduce multiple reallocations // we use a smart resize algorithm based on the current filling ratio - // in addition, we use float to avoid integers overflows - float nnzEstimate = float(m_outerIndex[outer])*float(m_outerSize)/float(outer+1); - reallocRatio = (nnzEstimate-float(m_data.size()))/float(m_data.size()); + // in addition, we use double to avoid integers overflows + double nnzEstimate = double(m_outerIndex[outer])*double(m_outerSize)/double(outer+1); + reallocRatio = (nnzEstimate-double(m_data.size()))/double(m_data.size()); // furthermore we bound the realloc ratio to: // 1) reduce multiple minor realloc when the matrix is almost filled // 2) avoid to allocate too much memory when the matrix is almost empty - reallocRatio = (std::min)((std::max)(reallocRatio,1.5f),8.f); + reallocRatio = (std::min)((std::max)(reallocRatio,1.5),8.); } } m_data.resize(m_data.size()+1,reallocRatio); From 77d57cd6813d4326543c68f7fd64901402e05d17 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 8 Jul 2014 19:07:58 +0200 Subject: [PATCH 0610/1103] bug #808: fix implicit conversions from int/longint to float/double --- Eigen/src/Core/functors/NullaryFunctors.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/functors/NullaryFunctors.h b/Eigen/src/Core/functors/NullaryFunctors.h index 950acd93b..be03fbf52 100644 --- a/Eigen/src/Core/functors/NullaryFunctors.h +++ b/Eigen/src/Core/functors/NullaryFunctors.h @@ -92,7 +92,7 @@ struct linspaced_op_impl template EIGEN_STRONG_INLINE const Packet packetOp(Index i) const - { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1(i),m_interPacket))); } + { return internal::padd(m_lowPacket, pmul(m_stepPacket, padd(pset1(Scalar(i)),m_interPacket))); } const Scalar m_low; const Scalar m_step; @@ -112,7 +112,7 @@ template struct functor_traits< linspaced_o template struct linspaced_op { typedef typename packet_traits::type Packet; - linspaced_op(const Scalar& low, const Scalar& high, DenseIndex num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/(num_steps-1))) {} + linspaced_op(const Scalar& low, const Scalar& high, DenseIndex num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1))) {} template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); } From 1967e7f2f37b0b2c17cb4ead26f43915d342c663 Mon Sep 17 00:00:00 2001 From: Chen-Pang He Date: Wed, 9 Jul 2014 03:32:32 +0800 Subject: [PATCH 0611/1103] Fix bug #839 --- Eigen/src/Core/TriangularMatrix.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index cdd341965..72792d21b 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -323,22 +323,22 @@ template class TriangularView /** Efficient triangular matrix times vector/matrix product */ template EIGEN_DEVICE_FUNC - TriangularProduct + TriangularProduct operator*(const MatrixBase& rhs) const { return TriangularProduct - + (m_matrix, rhs.derived()); } /** Efficient vector/matrix times triangular matrix product */ template friend EIGEN_DEVICE_FUNC - TriangularProduct + TriangularProduct operator*(const MatrixBase& lhs, const TriangularView& rhs) { return TriangularProduct - + (lhs.derived(),rhs.m_matrix); } From c285fda7f40ca161e6c8e66481d9a68e50613c48 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 8 Jul 2014 16:30:48 -0700 Subject: [PATCH 0612/1103] Extended the functionality of the TensorDeviceType classes --- .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 59 ++++++++++++++++++- 1 file changed, 56 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h index 142edda14..b9c8c19fe 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -21,6 +21,12 @@ struct DefaultDevice { EIGEN_STRONG_INLINE void deallocate(void* buffer) const { internal::aligned_free(buffer); } + EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const { + ::memcpy(dst, src, n); + } + EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { + ::memset(buffer, c, n); + } }; @@ -28,7 +34,7 @@ struct DefaultDevice { // We should really use a thread pool here but first we need to find a portable thread pool library. #ifdef EIGEN_USE_THREADS struct ThreadPoolDevice { - ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : /*pool_(pool), */num_threads_(num_cores) { } + ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : /*pool_(pool), */num_threads_(num_cores) { } size_t numThreads() const { return num_threads_; } EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { @@ -37,6 +43,12 @@ struct ThreadPoolDevice { EIGEN_STRONG_INLINE void deallocate(void* buffer) const { internal::aligned_free(buffer); } + EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const { + ::memcpy(dst, src, n); + } + EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { + ::memset(buffer, c, n); + } private: // todo: NUMA, ... @@ -47,20 +59,61 @@ struct ThreadPoolDevice { // GPU offloading #ifdef EIGEN_USE_GPU +static int m_numMultiProcessors = 0; +static int m_maxThreadsPerBlock = 0; +static int m_maxThreadsPerMultiProcessor = 0; + +static inline int getNumCudaMultiProcessors() { + if (m_numMultiProcessors == 0) { + cudaDeviceProp deviceProp; + cudaGetDeviceProperties(&deviceProp, 0); + m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock; + m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor; + m_numMultiProcessors = deviceProp.multiProcessorCount; + } + return m_numMultiProcessors; +} +static inline int maxCudaThreadsPerBlock() { + if (m_maxThreadsPerBlock == 0) { + cudaDeviceProp deviceProp; + cudaGetDeviceProperties(&deviceProp, 0); + m_numMultiProcessors = deviceProp.multiProcessorCount; + m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor; + m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock; + } + return m_maxThreadsPerBlock; +} +static inline int maxCudaThreadsPerMultiProcessor() { + if (m_maxThreadsPerBlock == 0) { + cudaDeviceProp deviceProp; + cudaGetDeviceProperties(&deviceProp, 0); + m_numMultiProcessors = deviceProp.multiProcessorCount; + m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock; + m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor; + } + return m_maxThreadsPerMultiProcessor; +} + struct GpuDevice { // The cudastream is not owned: the caller is responsible for its initialization and eventual destruction. GpuDevice(const cudaStream_t* stream) : stream_(stream) { eigen_assert(stream); } EIGEN_STRONG_INLINE const cudaStream_t& stream() const { return *stream_; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { + /*EIGEN_DEVICE_FUNC*/ EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { void* result; cudaMalloc(&result, num_bytes); return result; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const { + /*EIGEN_DEVICE_FUNC */EIGEN_STRONG_INLINE void deallocate(void* buffer) const { cudaFree(buffer); } + EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const { + cudaMemcpyAsync(dst, src, n, cudaMemcpyDeviceToDevice, *stream_); + } + EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { + cudaMemsetAsync(buffer, c, n, *stream_); + } private: // TODO: multigpu. From cc1bacea5b6b532728a001f8cfcf762e5385dcef Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 8 Jul 2014 16:39:28 -0700 Subject: [PATCH 0613/1103] Improved the efficiency of the tensor evaluation code on thread pools and gpus. --- .../Eigen/CXX11/src/Tensor/TensorExecutor.h | 50 +++++++------------ 1 file changed, 19 insertions(+), 31 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index 3e41f3290..f50f839fc 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -77,17 +77,17 @@ struct TensorExecutor #ifdef EIGEN_USE_THREADS template struct EvalRange { - static void run(Evaluator& evaluator, const Index first, const Index last) { + static void run(Evaluator* evaluator, const Index first, const Index last) { eigen_assert(last > first); for (Index i = first; i < last; ++i) { - evaluator.evalScalar(i); + evaluator->evalScalar(i); } } }; template struct EvalRange { - static void run(Evaluator& evaluator, const Index first, const Index last,) { + static void run(Evaluator* evaluator, const Index first, const Index last) { eigen_assert(last > first); Index i = first; @@ -96,12 +96,12 @@ struct EvalRange { eigen_assert(first % PacketSize == 0); Index lastPacket = last - (last % PacketSize); for (; i < lastPacket; i += PacketSize) { - evaluator.evalPacket(i); + evaluator->evalPacket(i); } } for (; i < last; ++i) { - evaluator.evalScalar(i); + evaluator->evalScalar(i); } } }; @@ -112,24 +112,23 @@ struct TensorExecutor typedef typename Expression::Index Index; static inline void run(const Expression& expr, const ThreadPoolDevice& device) { - TensorEvaluator evaluator(expr, device); + typedef TensorEvaluator Evaluator; + Evaluator evaluator(expr, device); evaluator.evalSubExprsIfNeeded(); const Index size = evaluator.dimensions().TotalSize(); - static const int PacketSize = Vectorizable ? unpacket_traits::PacketReturnType>::size : 1; + static const int PacketSize = Vectorizable ? unpacket_traits::size : 1; int blocksz = std::ceil(static_cast(size)/device.numThreads()) + PacketSize - 1; const Index blocksize = std::max(PacketSize, (blocksz - (blocksz % PacketSize))); const Index numblocks = size / blocksize; - TensorEvaluator single_threaded_eval(expr, DefaultDevice()); - Index i = 0; vector > results; results.reserve(numblocks); for (int i = 0; i < numblocks; ++i) { - results.push_back(std::async(std::launch::async, &EvalRange, Index>::run, single_threaded_eval, i*blocksize, (i+1)*blocksize)); + results.push_back(std::async(std::launch::async, &EvalRange::run, &evaluator, i*blocksize, (i+1)*blocksize)); } for (int i = 0; i < numblocks; ++i) { @@ -137,7 +136,7 @@ struct TensorExecutor } if (numblocks * blocksize < size) { - EvalRange, Index>::run(single_threaded_eval, numblocks * blocksize, size, nullptr); + EvalRange::run(&evaluator, numblocks * blocksize, size); } evaluator.cleanup(); @@ -149,15 +148,11 @@ struct TensorExecutor // GPU: the evaluation of the expression is offloaded to a GPU. #if defined(EIGEN_USE_GPU) && defined(__CUDACC__) template -__global__ void EigenMetaKernelNoCheck(Evaluator eval) { - const int index = blockIdx.x * blockDim.x + threadIdx.x; - eval.evalScalar(index); -} -template -__global__ void EigenMetaKernelPeel(Evaluator eval, int peel_start_offset, int size) { - const int index = peel_start_offset + blockIdx.x * blockDim.x + threadIdx.x; - if (index < size) { - eval.evalScalar(index); +__global__ void EigenMetaKernel(Evaluator eval, unsigned int size) { + const int first_index = blockIdx.x * blockDim.x + threadIdx.x; + const int step_size = blockDim.x * gridDim.x; + for (int i = first_index; i < size; i += step_size) { + eval.evalScalar(i); } } @@ -169,19 +164,12 @@ struct TensorExecutor { TensorEvaluator evaluator(expr, device); evaluator.evalSubExprsIfNeeded(); + const int num_blocks = getNumCudaMultiProcessors() * maxCudaThreadsPerMultiProcessor() / maxCudaThreadsPerBlock(); + const int block_size = maxCudaThreadsPerBlock(); const Index size = evaluator.dimensions().TotalSize(); - const int block_size = std::min(size, 32*32); - const int num_blocks = size / block_size; - EigenMetaKernelNoCheck > <<>>(evaluator); - - const int remaining_items = size % block_size; - if (remaining_items > 0) { - const int peel_start_offset = num_blocks * block_size; - const int peel_block_size = std::min(size, 32); - const int peel_num_blocks = (remaining_items + peel_block_size - 1) / peel_block_size; - EigenMetaKernelPeel > <<>>(evaluator, peel_start_offset, size); - } + EigenMetaKernel > <<>>(evaluator, size); + eigen_assert(cudaGetLastError() == cudaSuccess); evaluator.cleanup(); } }; From ea0906dfd877b3be91b5b0a28d2040ec360b1d3a Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 8 Jul 2014 16:43:28 -0700 Subject: [PATCH 0614/1103] Improved evaluation of tensor expressions when used as rvalues --- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 46 ++++++++++++++++++- 1 file changed, 45 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index 5c8b079da..ac9829ce9 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -23,6 +23,7 @@ namespace Eigen { * leading to lvalues (slicing, reshaping, etc...) */ +// Generic evaluator template struct TensorEvaluator { @@ -38,7 +39,7 @@ struct TensorEvaluator PacketAccess = Derived::PacketAccess, }; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(Derived& m, const Device&) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device&) : m_data(const_cast(m.data())), m_dims(m.dimensions()) { } @@ -75,6 +76,49 @@ struct TensorEvaluator }; +// Default evaluator for rvalues +template +struct TensorEvaluator +{ + typedef typename Derived::Index Index; + typedef typename Derived::Scalar Scalar; + typedef typename Derived::Packet Packet; + typedef typename Derived::Scalar CoeffReturnType; + typedef typename Derived::Packet PacketReturnType; + typedef typename Derived::Dimensions Dimensions; + + enum { + IsAligned = Derived::IsAligned, + PacketAccess = Derived::PacketAccess, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device&) + : m_data(m.data()), m_dims(m.dimensions()) + { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dims; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { + eigen_assert(m_data); + return m_data[index]; + } + + template EIGEN_STRONG_INLINE + PacketReturnType packet(Index index) const + { + return internal::ploadt(m_data + index); + } + + protected: + const Scalar* m_data; + Dimensions m_dims; +}; + + + // -------------------- CwiseNullaryOp -------------------- From 54fbbe7b4ec085260b6ad79a9eb912416482516d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 9 Jul 2014 13:06:06 +0200 Subject: [PATCH 0615/1103] Add unit test for bug #839. --- test/triangular.cpp | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/test/triangular.cpp b/test/triangular.cpp index 54320390b..936c2aef3 100644 --- a/test/triangular.cpp +++ b/test/triangular.cpp @@ -113,6 +113,13 @@ template void triangular_square(const MatrixType& m) m3.setZero(); m3.template triangularView().setOnes(); VERIFY_IS_APPROX(m2,m3); + + m1.setRandom(); + m3 = m1.template triangularView(); + Matrix m5(cols, internal::random(1,20)); m5.setRandom(); + Matrix m6(internal::random(1,20), rows); m6.setRandom(); + VERIFY_IS_APPROX(m1.template triangularView() * m5, m3*m5); + VERIFY_IS_APPROX(m6*m1.template triangularView(), m6*m3); } From 62f948c56a782689947e4bbdcaab2b5cb0247ac5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 9 Jul 2014 16:01:24 +0200 Subject: [PATCH 0616/1103] Generalize unit testing of pscatter --- test/packetmath.cpp | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-) diff --git a/test/packetmath.cpp b/test/packetmath.cpp index a51d31dbd..e5dc473c2 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -408,14 +408,17 @@ template void packetmath_scatter_gather() { for (int i=0; i()/RealScalar(PacketSize); } - EIGEN_ALIGN_DEFAULT Scalar buffer[PacketSize*11]; - memset(buffer, 0, 11*sizeof(Packet)); + + int stride = internal::random(1,20); + + EIGEN_ALIGN_DEFAULT Scalar buffer[PacketSize*20]; + memset(buffer, 0, 20*sizeof(Packet)); Packet packet = internal::pload(data1); - internal::pscatter(buffer, packet, 11); + internal::pscatter(buffer, packet, stride); - for (int i = 0; i < PacketSize*11; ++i) { - if ((i%11) == 0) { - VERIFY(isApproxAbs(buffer[i], data1[i/11], refvalue) && "pscatter"); + for (int i = 0; i < PacketSize*20; ++i) { + if ((i%stride) == 0 && i Date: Wed, 9 Jul 2014 16:54:15 +0200 Subject: [PATCH 0617/1103] Determine version of Metis library. Apparently, at least version 5.x is needed for Eigen/MetisSupport. Marked some internal variables as advanced --- cmake/FindCholmod.cmake | 2 +- cmake/FindFFTW.cmake | 2 +- cmake/FindMetis.cmake | 38 ++++++++++++++++++++++++++++++++++++-- test/CMakeLists.txt | 3 ++- 4 files changed, 40 insertions(+), 5 deletions(-) diff --git a/cmake/FindCholmod.cmake b/cmake/FindCholmod.cmake index 7b3046d45..23239c300 100644 --- a/cmake/FindCholmod.cmake +++ b/cmake/FindCholmod.cmake @@ -86,4 +86,4 @@ include(FindPackageHandleStandardArgs) find_package_handle_standard_args(CHOLMOD DEFAULT_MSG CHOLMOD_INCLUDES CHOLMOD_LIBRARIES) -mark_as_advanced(CHOLMOD_INCLUDES CHOLMOD_LIBRARIES AMD_LIBRARY COLAMD_LIBRARY SUITESPARSE_LIBRARY) +mark_as_advanced(CHOLMOD_INCLUDES CHOLMOD_LIBRARIES AMD_LIBRARY COLAMD_LIBRARY SUITESPARSE_LIBRARY CAMD_LIBRARY CCOLAMD_LIBRARY CHOLMOD_METIS_LIBRARY) diff --git a/cmake/FindFFTW.cmake b/cmake/FindFFTW.cmake index a9e9925e7..6c4dc9ab4 100644 --- a/cmake/FindFFTW.cmake +++ b/cmake/FindFFTW.cmake @@ -115,5 +115,5 @@ include(FindPackageHandleStandardArgs) find_package_handle_standard_args(FFTW DEFAULT_MSG FFTW_INCLUDES FFTW_LIBRARIES) -mark_as_advanced(FFTW_INCLUDES FFTW_LIBRARIES) +mark_as_advanced(FFTW_INCLUDES FFTW_LIBRARIES FFTW_LIB FFTWF_LIB FFTWL_LIB) diff --git a/cmake/FindMetis.cmake b/cmake/FindMetis.cmake index 627c3e9ae..e0040d320 100644 --- a/cmake/FindMetis.cmake +++ b/cmake/FindMetis.cmake @@ -10,16 +10,50 @@ find_path(METIS_INCLUDES PATHS $ENV{METISDIR} ${INCLUDE_INSTALL_DIR} - PATH_SUFFIXES + PATH_SUFFIXES + . metis include ) +macro(_metis_check_version) + file(READ "${METIS_INCLUDES}/metis.h" _metis_version_header) + + string(REGEX MATCH "define[ \t]+METIS_VER_MAJOR[ \t]+([0-9]+)" _metis_major_version_match "${_metis_version_header}") + set(METIS_MAJOR_VERSION "${CMAKE_MATCH_1}") + string(REGEX MATCH "define[ \t]+METIS_VER_MINOR[ \t]+([0-9]+)" _metis_minor_version_match "${_metis_version_header}") + set(METIS_MINOR_VERSION "${CMAKE_MATCH_1}") + string(REGEX MATCH "define[ \t]+METIS_VER_SUBMINOR[ \t]+([0-9]+)" _metis_subminor_version_match "${_metis_version_header}") + set(METIS_SUBMINOR_VERSION "${CMAKE_MATCH_1}") + if(NOT METIS_MAJOR_VERSION) + message(WARNING "Could not determine Metis version. Assuming version 4.0.0") + set(METIS_VERSION 4.0.0) + else() + set(METIS_VERSION ${METIS_MAJOR_VERSION}.${METIS_MINOR_VERSION}.${METIS_SUBMINOR_VERSION}) + endif() + if(${METIS_VERSION} VERSION_LESS ${Metis_FIND_VERSION}) + set(METIS_VERSION_OK FALSE) + else() + set(METIS_VERSION_OK TRUE) + endif() + + if(NOT METIS_VERSION_OK) + message(STATUS "Metis version ${METIS_VERSION} found in ${METIS_INCLUDES}, " + "but at least version ${Metis_FIND_VERSION} is required") + endif(NOT METIS_VERSION_OK) +endmacro(_metis_check_version) + + if(METIS_INCLUDES AND Metis_FIND_VERSION) + _metis_check_version() + else() + set(METIS_VERSION_OK TRUE) + endif() + find_library(METIS_LIBRARIES metis PATHS $ENV{METISDIR} ${LIB_INSTALL_DIR} PATH_SUFFIXES lib) include(FindPackageHandleStandardArgs) find_package_handle_standard_args(METIS DEFAULT_MSG - METIS_INCLUDES METIS_LIBRARIES) + METIS_INCLUDES METIS_LIBRARIES METIS_VERSION_OK) mark_as_advanced(METIS_INCLUDES METIS_LIBRARIES) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 62c9c78a6..4521d07e4 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -82,7 +82,7 @@ endif() find_package(Pastix) find_package(Scotch) -find_package(Metis) +find_package(Metis 5.0 REQUIRED) if(PASTIX_FOUND) add_definitions("-DEIGEN_PASTIX_SUPPORT") include_directories(${PASTIX_INCLUDES}) @@ -299,6 +299,7 @@ ei_add_property(EIGEN_TESTING_SUMMARY "CXX_FLAGS: ${CMAKE_CXX_FLAGS}\n") ei_add_property(EIGEN_TESTING_SUMMARY "Sparse lib flags: ${SPARSE_LIBS}\n") option(EIGEN_TEST_EIGEN2 "Run whole Eigen2 test suite against EIGEN2_SUPPORT" OFF) +mark_as_advanced(EIGEN_TEST_EIGEN2) if(EIGEN_TEST_EIGEN2) message(WARNING "The Eigen2 test suite has been removed") endif() From 23bb592a2d0de937c68fc1c819642bbcfa2c1e4f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 9 Jul 2014 17:21:16 +0200 Subject: [PATCH 0618/1103] Fix unit test when using 80bits FPU --- test/block.cpp | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/test/block.cpp b/test/block.cpp index 9ed5d7bc5..269acd28e 100644 --- a/test/block.cpp +++ b/test/block.cpp @@ -141,11 +141,11 @@ template void block(const MatrixType& m) VERIFY_IS_EQUAL( (m1.transpose().block(c1,r1,c2-c1+1,r2-r1+1).col(0)) , (m1.row(r1).segment(c1,c2-c1+1)).transpose() ); // expressions without direct access - VERIFY_IS_EQUAL( ((m1+m2).block(r1,c1,rows-r1,cols-c1).block(r2-r1,c2-c1,rows-r2,cols-c2)) , ((m1+m2).block(r2,c2,rows-r2,cols-c2)) ); - VERIFY_IS_EQUAL( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).row(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)) ); - VERIFY_IS_EQUAL( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).col(0)) , ((m1+m2).col(c1).segment(r1,r2-r1+1)) ); - VERIFY_IS_EQUAL( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).transpose().col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() ); - VERIFY_IS_EQUAL( ((m1+m2).transpose().block(c1,r1,c2-c1+1,r2-r1+1).col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() ); + VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,rows-r1,cols-c1).block(r2-r1,c2-c1,rows-r2,cols-c2)) , ((m1+m2).block(r2,c2,rows-r2,cols-c2)) ); + VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).row(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)) ); + VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).col(0)) , ((m1+m2).col(c1).segment(r1,r2-r1+1)) ); + VERIFY_IS_APPROX( ((m1+m2).block(r1,c1,r2-r1+1,c2-c1+1).transpose().col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() ); + VERIFY_IS_APPROX( ((m1+m2).transpose().block(c1,r1,c2-c1+1,r2-r1+1).col(0)) , ((m1+m2).row(r1).segment(c1,c2-c1+1)).transpose() ); // evaluation into plain matrices from expressions with direct access (stress MapBase) DynamicMatrixType dm; From 25b2f6624d092ed99d0c4936de0c83c9ea4a024d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 9 Jul 2014 12:48:34 -0700 Subject: [PATCH 0619/1103] Improved the speed of slicing operations. --- unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 55954a3a7..f6f67afa7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -320,11 +320,12 @@ struct TensorEvaluator, Devi EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { Index inputIndex = 0; - for (int i = NumDims - 1; i >= 0; --i) { + for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; index -= idx * m_outputStrides[i]; } + inputIndex += (index + m_offsets[0]); return m_impl.coeff(inputIndex); } @@ -399,11 +400,12 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { Index inputIndex = 0; - for (int i = NumDims - 1; i >= 0; --i) { + for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; index -= idx * m_outputStrides[i]; } + inputIndex += (index + m_offsets[0]); return m_impl.coeff(inputIndex); } @@ -416,11 +418,12 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) { Index inputIndex = 0; - for (int i = NumDims - 1; i >= 0; --i) { + for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; index -= idx * m_outputStrides[i]; } + inputIndex += (index + m_offsets[0]); return m_impl.coeffRef(inputIndex); } From e955725ff1434396cc41beda6f5989bef0940704 Mon Sep 17 00:00:00 2001 From: Kolja Brix Date: Thu, 10 Jul 2014 08:20:55 +0200 Subject: [PATCH 0620/1103] Fix GMRES: Initialize essential Householder vector with correct dimension. Add check if initial guess is already a sufficient approximation. --- unsupported/Eigen/src/IterativeSolvers/GMRES.h | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index 073367506..c8c84069e 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2011 Gael Guennebaud -// Copyright (C) 2012 Kolja Brix +// Copyright (C) 2012, 2014 Kolja Brix // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -72,16 +72,20 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition VectorType p0 = rhs - mat*x; VectorType r0 = precond.solve(p0); -// RealScalar r0_sqnorm = r0.squaredNorm(); + + // is initial guess already good enough? + if(abs(r0.norm()) < tol) { + return true; + } VectorType w = VectorType::Zero(restart + 1); - FMatrixType H = FMatrixType::Zero(m, restart + 1); + FMatrixType H = FMatrixType::Zero(m, restart + 1); // Hessenberg matrix VectorType tau = VectorType::Zero(restart + 1); std::vector < JacobiRotation < Scalar > > G(restart); // generate first Householder vector - VectorType e; + VectorType e(m-1); RealScalar beta; r0.makeHouseholder(e, tau.coeffRef(0), beta); w(0)=(Scalar) beta; From f27f55bee3efc2cafd01cb07d3faadf7eb490f66 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 10 Jul 2014 14:59:18 +0200 Subject: [PATCH 0621/1103] Make MatrixBase::makeHouseholder resize its output vector if it is zero --- Eigen/src/Householder/Householder.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Householder/Householder.h b/Eigen/src/Householder/Householder.h index 32112af9b..9d33e1e62 100644 --- a/Eigen/src/Householder/Householder.h +++ b/Eigen/src/Householder/Householder.h @@ -80,7 +80,7 @@ void MatrixBase::makeHouseholder( { tau = RealScalar(0); beta = numext::real(c0); - essential.setZero(); + essential.setZero(size()-1); } else { From cf7cf7b4907859381e823d99588de94d5e31bd72 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 10 Jul 2014 16:12:13 +0200 Subject: [PATCH 0622/1103] Backed out of changeset 6089:f27f55bee3efc2cafd01cb07d3faadf7eb490f66 Unfortunately this breaks things at other places --- Eigen/src/Householder/Householder.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Householder/Householder.h b/Eigen/src/Householder/Householder.h index 9d33e1e62..32112af9b 100644 --- a/Eigen/src/Householder/Householder.h +++ b/Eigen/src/Householder/Householder.h @@ -80,7 +80,7 @@ void MatrixBase::makeHouseholder( { tau = RealScalar(0); beta = numext::real(c0); - essential.setZero(size()-1); + essential.setZero(); } else { From d1460d92782ce0f7b90a8a52d44d50b44b167f98 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 10 Jul 2014 16:23:20 +0200 Subject: [PATCH 0623/1103] stride must be DenseIndex not int --- Eigen/src/Core/arch/AVX/Complex.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index b3cf7bb26..aa5aa1e34 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -310,13 +310,13 @@ template<> EIGEN_STRONG_INLINE Packet2cd ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } -template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather, Packet2cd>(const std::complex* from, int stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather, Packet2cd>(const std::complex* from, DenseIndex stride) { return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]), std::imag(from[0*stride]), std::real(from[0*stride]))); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cd>(std::complex* to, const Packet2cd& from, int stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cd>(std::complex* to, const Packet2cd& from, DenseIndex stride) { __m128d low = _mm256_extractf128_pd(from.v, 0); to[stride*0] = std::complex(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1))); From b1169ce40cf615ec2cd8eb116ae905b48a849a3a Mon Sep 17 00:00:00 2001 From: Jeff Date: Thu, 10 Jul 2014 12:03:42 -0600 Subject: [PATCH 0624/1103] Fixed index that would cause crash with two point, two derivative interpolation. Added static_cast. --- unsupported/Eigen/src/Splines/SplineFitting.h | 2 +- unsupported/test/splines.cpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/src/Splines/SplineFitting.h b/unsupported/Eigen/src/Splines/SplineFitting.h index 85de38b25..d3c245fa9 100644 --- a/unsupported/Eigen/src/Splines/SplineFitting.h +++ b/unsupported/Eigen/src/Splines/SplineFitting.h @@ -128,7 +128,7 @@ namespace Eigen newKnotIndex = -1; - ParameterVectorType temporaryParameters(numParameters); + ParameterVectorType temporaryParameters(numParameters + 1); KnotVectorType derivativeKnots(numInternalDerivatives); for (unsigned int i = 0; i < numAverageKnots - 1; ++i) { diff --git a/unsupported/test/splines.cpp b/unsupported/test/splines.cpp index 0b43eefb8..97665af96 100644 --- a/unsupported/test/splines.cpp +++ b/unsupported/test/splines.cpp @@ -252,7 +252,7 @@ void check_global_interpolation_with_derivatives2d() VectorXd derivativeIndices(numPoints); for (Eigen::DenseIndex i = 0; i < numPoints; ++i) - derivativeIndices(i) = i; + derivativeIndices(i) = static_cast(i); const Spline2d spline = SplineFitting::InterpolateWithDerivatives( points, derivatives, derivativeIndices, degree); From ffd3654f6738bab79db010e02cd67660ecca62c1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 10 Jul 2014 11:09:46 -0700 Subject: [PATCH 0625/1103] Vectorized the evaluation of expressions involving tensor slices. --- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 108 ++++++++++++++++-- 1 file changed, 98 insertions(+), 10 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index f6f67afa7..3b42c8514 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -273,8 +273,10 @@ struct TensorEvaluator, Devi static const int NumDims = internal::array_size::value; enum { - IsAligned = TensorEvaluator::IsAligned, - PacketAccess = /*TensorEvaluator::PacketAccess*/false, + // Alignment can't be guaranteed at compile time since it depends on the + // slice offsets and sizes. + IsAligned = /*TensorEvaluator::IsAligned*/false, + PacketAccess = TensorEvaluator::PacketAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -329,11 +331,40 @@ struct TensorEvaluator, Devi return m_impl.coeff(inputIndex); } - /* template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - return m_impl.template packet(index); - }*/ + static const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); + + Index inputIndices[] = {0, 0}; + Index indices[] = {index, index + packetSize - 1}; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / m_outputStrides[i]; + const Index idx1 = indices[1] / m_outputStrides[i]; + inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; + inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += (indices[0] + m_offsets[0]); + inputIndices[1] += (indices[1] + m_offsets[0]); + if (inputIndices[1] - inputIndices[0] == packetSize - 1) { + PacketReturnType rslt = m_impl.template packet(inputIndices[0]); + return rslt; + } + else { + CoeffReturnType values[packetSize]; + values[0] = m_impl.coeff(inputIndices[0]); + values[packetSize-1] = m_impl.coeff(inputIndices[1]); + for (int i = 1; i < packetSize-1; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + } private: Dimensions m_dimensions; @@ -353,8 +384,8 @@ struct TensorEvaluator, Device> static const int NumDims = internal::array_size::value; enum { - IsAligned = TensorEvaluator::IsAligned, - PacketAccess = /*TensorEvaluator::PacketAccess*/false, + IsAligned = /*TensorEvaluator::IsAligned*/false, + PacketAccess = TensorEvaluator::PacketAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -409,11 +440,38 @@ struct TensorEvaluator, Device> return m_impl.coeff(inputIndex); } - /* template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - return m_impl.template packet(index); - }*/ + static const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + Index inputIndices[] = {0, 0}; + Index indices[] = {index, index + packetSize - 1}; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / m_outputStrides[i]; + const Index idx1 = indices[1] / m_outputStrides[i]; + inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; + inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += (indices[0] + m_offsets[0]); + inputIndices[1] += (indices[1] + m_offsets[0]); + if (inputIndices[1] - inputIndices[0] == packetSize - 1) { + PacketReturnType rslt = m_impl.template packet(inputIndices[0]); + return rslt; + } + else { + CoeffReturnType values[packetSize]; + values[0] = m_impl.coeff(inputIndices[0]); + values[packetSize-1] = m_impl.coeff(inputIndices[1]); + for (int i = 1; i < packetSize-1; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) { @@ -427,6 +485,36 @@ struct TensorEvaluator, Device> return m_impl.coeffRef(inputIndex); } + template EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketReturnType& x) + { + static const int packetSize = internal::unpacket_traits::size; + Index inputIndices[] = {0, 0}; + Index indices[] = {index, index + packetSize - 1}; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / m_outputStrides[i]; + const Index idx1 = indices[1] / m_outputStrides[i]; + inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; + inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += (indices[0] + m_offsets[0]); + inputIndices[1] += (indices[1] + m_offsets[0]); + if (inputIndices[1] - inputIndices[0] == packetSize - 1) { + m_impl.template writePacket(inputIndices[0], x); + } + else { + CoeffReturnType values[packetSize]; + internal::pstore(values, x); + m_impl.coeffRef(inputIndices[0]) = values[0]; + m_impl.coeffRef(inputIndices[1]) = values[packetSize-1]; + for (int i = 1; i < packetSize-1; ++i) { + coeffRef(index+i) = values[i]; + } + } + } + private: Dimensions m_dimensions; array m_outputStrides; From 9b7a6f0122f6817a3c12bc75803d4270cd9db507 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 10 Jul 2014 11:27:27 -0700 Subject: [PATCH 0626/1103] Added tests for tensor slicing --- unsupported/test/cxx11_tensor_morphing.cpp | 132 ++++++++++++++++++++- 1 file changed, 130 insertions(+), 2 deletions(-) diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp index 21af9e0b5..fbfdaadb7 100644 --- a/unsupported/test/cxx11_tensor_morphing.cpp +++ b/unsupported/test/cxx11_tensor_morphing.cpp @@ -52,8 +52,7 @@ static void test_reshape_in_expr() { TensorMap> tensor2(m2.data(), 3,5,7,11,13); Tensor::Dimensions newDims1{{2,3*5*7*11}}; Tensor::Dimensions newDims2{{3*5*7*11,13}}; - typedef Tensor::DimensionPair DimPair; - array contract_along{{DimPair(1, 0)}}; + array::DimensionPair, 1> contract_along{{1, 0}}; Tensor tensor3(2,13); tensor3 = tensor1.reshape(newDims1).contract(tensor2.reshape(newDims2), contract_along); @@ -65,8 +64,137 @@ static void test_reshape_in_expr() { } } + +static void test_reshape_as_lvalue() +{ + Tensor tensor(2,3,7); + tensor.setRandom(); + + Tensor tensor2d(6,7); + Tensor::Dimensions dim{{2,3,7}}; + tensor2d.reshape(dim) = tensor; + + Tensor tensor5d(2,3,1,7,1); + tensor5d.reshape(dim).device(Eigen::DefaultDevice()) = tensor; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(tensor2d(i+2*j,k), tensor(i,j,k)); + VERIFY_IS_EQUAL(tensor5d(i,j,0,k,0), tensor(i,j,k)); + } + } + } +} + + +static void test_simple_slice() +{ + Tensor tensor(2,3,5,7,11); + tensor.setRandom(); + + Tensor slice1(1,1,1,1,1); + Eigen::DSizes indices(Eigen::array(1,2,3,4,5)); + Eigen::DSizes sizes(Eigen::array(1,1,1,1,1)); + slice1 = tensor.slice(indices, sizes); + VERIFY_IS_EQUAL(slice1(0,0,0,0,0), tensor(1,2,3,4,5)); + + Tensor slice2(1,1,2,2,3); + Eigen::DSizes indices2(Eigen::array(1,1,3,4,5)); + Eigen::DSizes sizes2(Eigen::array(1,1,2,2,3)); + slice2 = tensor.slice(indices2, sizes2); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 2; ++j) { + for (int k = 0; k < 3; ++k) { + VERIFY_IS_EQUAL(slice2(0,0,i,j,k), tensor(1,1,3+i,4+j,5+k)); + } + } + } +} + + +static void test_slice_in_expr() { + MatrixXf m1(7,7); + MatrixXf m2(3,3); + m1.setRandom(); + m2.setRandom(); + + MatrixXf m3 = m1.block(1, 2, 3, 3) * m2.block(0, 2, 3, 1); + + TensorMap> tensor1(m1.data(), 7, 7); + TensorMap> tensor2(m2.data(), 3, 3); + Tensor tensor3(3,1); + array::DimensionPair, 1> contract_along{{1, 0}}; + + Eigen::DSizes indices1(Eigen::array(1,2)); + Eigen::DSizes sizes1(Eigen::array(3,3)); + Eigen::DSizes indices2(Eigen::array(0,2)); + Eigen::DSizes sizes2(Eigen::array(3,1)); + tensor3 = tensor1.slice(indices1, sizes1).contract(tensor2.slice(indices2, sizes2), contract_along); + + Map res(tensor3.data(), 3, 1); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 1; ++j) { + VERIFY_IS_APPROX(res(i,j), m3(i,j)); + } + } +} + + +static void test_slice_as_lvalue() +{ + Tensor tensor1(2,2,7); + tensor1.setRandom(); + Tensor tensor2(2,2,7); + tensor2.setRandom(); + Tensor tensor3(4,3,5); + tensor3.setRandom(); + Tensor tensor4(4,3,2); + tensor4.setRandom(); + + Tensor result(4,5,7); + Eigen::DSizes sizes12(Eigen::array(2,2,7)); + Eigen::DSizes first_slice(Eigen::array(0,0,0)); + result.slice(first_slice, sizes12) = tensor1; + Eigen::DSizes second_slice(Eigen::array(2,0,0)); + result.slice(second_slice, sizes12).device(Eigen::DefaultDevice()) = tensor2; + + Eigen::DSizes sizes3(Eigen::array(4,3,5)); + Eigen::DSizes third_slice(Eigen::array(0,2,0)); + result.slice(third_slice, sizes3) = tensor3; + + Eigen::DSizes sizes4(Eigen::array(4,3,2)); + Eigen::DSizes fourth_slice(Eigen::array(0,2,5)); + result.slice(fourth_slice, sizes4) = tensor4; + + for (int j = 0; j < 2; ++j) { + for (int k = 0; k < 7; ++k) { + for (int i = 0; i < 2; ++i) { + VERIFY_IS_EQUAL(result(i,j,k), tensor1(i,j,k)); + VERIFY_IS_EQUAL(result(i+2,j,k), tensor2(i,j,k)); + } + } + } + for (int i = 0; i < 4; ++i) { + for (int j = 2; j < 5; ++j) { + for (int k = 0; k < 5; ++k) { + VERIFY_IS_EQUAL(result(i,j,k), tensor3(i,j-2,k)); + } + for (int k = 5; k < 7; ++k) { + VERIFY_IS_EQUAL(result(i,j,k), tensor4(i,j-2,k-5)); + } + } + } +} + + void test_cxx11_tensor_morphing() { CALL_SUBTEST(test_simple_reshape()); CALL_SUBTEST(test_reshape_in_expr()); + CALL_SUBTEST(test_reshape_as_lvalue()); + + CALL_SUBTEST(test_simple_slice()); + CALL_SUBTEST(test_slice_in_expr()); + CALL_SUBTEST(test_slice_as_lvalue()); } From 40bb98e76acbe6e077903e15896c100ee6cced39 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 10 Jul 2014 11:29:51 -0700 Subject: [PATCH 0627/1103] Added primitives to compare tensor dimensions --- .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 54 +++++++++++++++++++ 1 file changed, 54 insertions(+) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 3e5687915..3b169a06f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -210,6 +210,60 @@ struct DSizes : array { }; +namespace internal { + +template struct array_size > { + static const size_t value = NumDims; +}; +template struct array_size > { + static const size_t value = NumDims; +}; +#ifndef EIGEN_EMULATE_CXX11_META_H +template struct array_size > { +static const size_t value = Sizes::count; +}; +template struct array_size > { +static const size_t value = Sizes::count; +}; +#else +template struct array_size > { + static const size_t value = Sizes::count; +}; +template struct array_size > { + static const size_t value = Sizes::count; +}; +template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_get(const Sizes& a) { + return get::Base>::value; +}; + +#endif + + +template +struct sizes_match_up_to_dim { + static inline bool run(Dims1& dims1, Dims2& dims2) { + return (array_get(dims1) == array_get(dims2)) & + sizes_match_up_to_dim::run(dims1, dims2); + } +}; +template +struct sizes_match_up_to_dim { + static inline bool run(Dims1& dims1, Dims2& dims2) { + return (array_get<0>(dims1) == array_get<0>(dims2)); + } +}; + +template +bool dimensions_match(Dims1& dims1, Dims2& dims2) { + if (array_size::value != array_size::value) { + return false; + } + return sizes_match_up_to_dim::value-1>::run(dims1, dims2); +} + +} // end namespace internal + + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H From df604e4f499edcb8ddc17692685cf736067a95f2 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 11 Jul 2014 16:24:49 +0200 Subject: [PATCH 0628/1103] Fix inner iterator on an outer-vector --- Eigen/src/SparseCore/SparseBlock.h | 1 + 1 file changed, 1 insertion(+) diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index cbcd7a299..491cc72b0 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -523,6 +523,7 @@ namespace internal { while(m_outerPos Date: Fri, 11 Jul 2014 16:25:36 +0200 Subject: [PATCH 0629/1103] Fix bug #838: fix dense * sparse and sparse * dense outer products --- Eigen/src/SparseCore/SparseDenseProduct.h | 33 +++++++++++---- test/sparse_product.cpp | 51 +++++++++++------------ 2 files changed, 50 insertions(+), 34 deletions(-) diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 610833f3b..a3772c255 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -19,7 +19,10 @@ template struct SparseDenseProductRet template struct SparseDenseProductReturnType { - typedef SparseDenseOuterProduct Type; + typedef typename internal::conditional< + Lhs::IsRowMajor, + SparseDenseOuterProduct, + SparseDenseOuterProduct >::type Type; }; template struct DenseSparseProductReturnType @@ -29,7 +32,10 @@ template struct DenseSparseProductRet template struct DenseSparseProductReturnType { - typedef SparseDenseOuterProduct Type; + typedef typename internal::conditional< + Rhs::IsRowMajor, + SparseDenseOuterProduct, + SparseDenseOuterProduct >::type Type; }; namespace internal { @@ -114,17 +120,30 @@ class SparseDenseOuterProduct::InnerIterator : public _LhsNes typedef typename SparseDenseOuterProduct::Index Index; public: EIGEN_STRONG_INLINE InnerIterator(const SparseDenseOuterProduct& prod, Index outer) - : Base(prod.lhs(), 0), m_outer(outer), m_factor(prod.rhs().coeff(outer)) - { - } + : Base(prod.lhs(), 0), m_outer(outer), m_factor(get(prod.rhs(), outer, typename internal::traits::StorageKind() )) + { } inline Index outer() const { return m_outer; } - inline Index row() const { return Transpose ? Base::row() : m_outer; } - inline Index col() const { return Transpose ? m_outer : Base::row(); } + inline Index row() const { return Transpose ? m_outer : Base::index(); } + inline Index col() const { return Transpose ? Base::index() : m_outer; } inline Scalar value() const { return Base::value() * m_factor; } protected: + static Scalar get(const _RhsNested &rhs, Index outer, Dense = Dense()) + { + return rhs.coeff(outer); + } + + static Scalar get(const _RhsNested &rhs, Index outer, Sparse = Sparse()) + { + typename Traits::_RhsNested::InnerIterator it(rhs, outer); + if (it && it.index()==0) + return it.value(); + + return Scalar(0); + } + Index m_outer; Scalar m_factor; }; diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index a2ea9d5b7..2c8b97e5b 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -9,32 +9,6 @@ #include "sparse.h" -template struct test_outer; - -template struct test_outer { - static void run(SparseMatrixType& m2, SparseMatrixType& m4, DenseMatrix& refMat2, DenseMatrix& refMat4) { - typedef typename SparseMatrixType::Index Index; - Index c = internal::random(0,m2.cols()-1); - Index c1 = internal::random(0,m2.cols()-1); - VERIFY_IS_APPROX(m4=m2.col(c)*refMat2.col(c1).transpose(), refMat4=refMat2.col(c)*refMat2.col(c1).transpose()); - VERIFY_IS_APPROX(m4=refMat2.col(c1)*m2.col(c).transpose(), refMat4=refMat2.col(c1)*refMat2.col(c).transpose()); - } -}; - -template struct test_outer { - static void run(SparseMatrixType& m2, SparseMatrixType& m4, DenseMatrix& refMat2, DenseMatrix& refMat4) { - typedef typename SparseMatrixType::Index Index; - Index r = internal::random(0,m2.rows()-1); - Index c1 = internal::random(0,m2.cols()-1); - VERIFY_IS_APPROX(m4=m2.row(r).transpose()*refMat2.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat2.col(c1).transpose()); - VERIFY_IS_APPROX(m4=refMat2.col(c1)*m2.row(r), refMat4=refMat2.col(c1)*refMat2.row(r)); - } -}; - -// (m2,m4,refMat2,refMat4,dv1); -// VERIFY_IS_APPROX(m4=m2.innerVector(c)*dv1.transpose(), refMat4=refMat2.colVector(c)*dv1.transpose()); -// VERIFY_IS_APPROX(m4=dv1*mcm.col(c).transpose(), refMat4=dv1*refMat2.col(c).transpose()); - template void sparse_product() { typedef typename SparseMatrixType::Index Index; @@ -119,7 +93,30 @@ template void sparse_product() VERIFY_IS_APPROX(dm4=refMat2t.transpose()*m3t.transpose(), refMat4=refMat2t.transpose()*refMat3t.transpose()); // sparse * dense and dense * sparse outer product - test_outer::run(m2,m4,refMat2,refMat4); + { + Index c = internal::random(0,depth-1); + Index r = internal::random(0,rows-1); + Index c1 = internal::random(0,cols-1); + Index r1 = internal::random(0,depth-1); + + VERIFY_IS_APPROX( m4=m2.col(c)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); + VERIFY_IS_APPROX(dm4=m2.col(c)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); + + VERIFY_IS_APPROX(m4=refMat3.col(c1)*m2.col(c).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); + VERIFY_IS_APPROX(dm4=refMat3.col(c1)*m2.col(c).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); + + VERIFY_IS_APPROX( m4=refMat3.row(r1).transpose()*m2.col(c).transpose(), refMat4=refMat3.row(r1).transpose()*refMat2.col(c).transpose()); + VERIFY_IS_APPROX(dm4=refMat3.row(r1).transpose()*m2.col(c).transpose(), refMat4=refMat3.row(r1).transpose()*refMat2.col(c).transpose()); + + VERIFY_IS_APPROX( m4=m2.row(r).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); + VERIFY_IS_APPROX(dm4=m2.row(r).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); + + VERIFY_IS_APPROX( m4=refMat3.col(c1)*m2.row(r), refMat4=refMat3.col(c1)*refMat2.row(r)); + VERIFY_IS_APPROX(dm4=refMat3.col(c1)*m2.row(r), refMat4=refMat3.col(c1)*refMat2.row(r)); + + VERIFY_IS_APPROX( m4=refMat3.row(r1).transpose()*m2.row(r), refMat4=refMat3.row(r1).transpose()*refMat2.row(r)); + VERIFY_IS_APPROX(dm4=refMat3.row(r1).transpose()*m2.row(r), refMat4=refMat3.row(r1).transpose()*refMat2.row(r)); + } VERIFY_IS_APPROX(m6=m6*m6, refMat6=refMat6*refMat6); From a20e2462bf760fa87d7228d42fac88c188da5a6d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 11 Jul 2014 17:15:26 +0200 Subject: [PATCH 0630/1103] Fix bug #838: detect outer products from either the lhs or rhs --- Eigen/src/SparseCore/SparseUtil.h | 7 +++++-- test/sparse_product.cpp | 6 +++++- 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 05023858b..02c19d18f 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -84,8 +84,11 @@ template class DenseTimeSparseProduct; template class SparseDenseOuterProduct; template struct SparseSparseProductReturnType; -template::ColsAtCompileTime> struct DenseSparseProductReturnType; -template::ColsAtCompileTime> struct SparseDenseProductReturnType; +template::ColsAtCompileTime,internal::traits::RowsAtCompileTime)> struct DenseSparseProductReturnType; + +template::ColsAtCompileTime,internal::traits::RowsAtCompileTime)> struct SparseDenseProductReturnType; template class SparseSymmetricPermutationProduct; namespace internal { diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index 2c8b97e5b..7a76acf57 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -100,18 +100,22 @@ template void sparse_product() Index r1 = internal::random(0,depth-1); VERIFY_IS_APPROX( m4=m2.col(c)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); + VERIFY_IS_APPROX( m4=m2.middleCols(c,1)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); VERIFY_IS_APPROX(dm4=m2.col(c)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); - + VERIFY_IS_APPROX(m4=refMat3.col(c1)*m2.col(c).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); + VERIFY_IS_APPROX(m4=refMat3.col(c1)*m2.middleCols(c,1).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); VERIFY_IS_APPROX(dm4=refMat3.col(c1)*m2.col(c).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); VERIFY_IS_APPROX( m4=refMat3.row(r1).transpose()*m2.col(c).transpose(), refMat4=refMat3.row(r1).transpose()*refMat2.col(c).transpose()); VERIFY_IS_APPROX(dm4=refMat3.row(r1).transpose()*m2.col(c).transpose(), refMat4=refMat3.row(r1).transpose()*refMat2.col(c).transpose()); VERIFY_IS_APPROX( m4=m2.row(r).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); + VERIFY_IS_APPROX( m4=m2.middleRows(r,1).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); VERIFY_IS_APPROX(dm4=m2.row(r).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); VERIFY_IS_APPROX( m4=refMat3.col(c1)*m2.row(r), refMat4=refMat3.col(c1)*refMat2.row(r)); + VERIFY_IS_APPROX( m4=refMat3.col(c1)*m2.middleRows(r,1), refMat4=refMat3.col(c1)*refMat2.row(r)); VERIFY_IS_APPROX(dm4=refMat3.col(c1)*m2.row(r), refMat4=refMat3.col(c1)*refMat2.row(r)); VERIFY_IS_APPROX( m4=refMat3.row(r1).transpose()*m2.row(r), refMat4=refMat3.row(r1).transpose()*refMat2.row(r)); From 4f440b81237a2624e4bf927b279b4a399469b28d Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 14 Jul 2014 14:36:20 +0200 Subject: [PATCH 0631/1103] Test vectorization logic for int --- test/vectorization_logic.cpp | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp index 09b46660b..b069f0771 100644 --- a/test/vectorization_logic.cpp +++ b/test/vectorization_logic.cpp @@ -77,8 +77,9 @@ bool test_redux(const Xpr&, int traversal, int unrolling) template::Vectorizable> struct vectorization_logic { + typedef internal::packet_traits PacketTraits; enum { - PacketSize = internal::packet_traits::size + PacketSize = PacketTraits::size }; static void run() { @@ -151,7 +152,7 @@ template::Vectori LinearTraversal,CompleteUnrolling)); VERIFY(test_assign(Matrix3(),Matrix3().cwiseQuotient(Matrix3()), - LinearVectorizedTraversal,CompleteUnrolling)); + PacketTraits::HasDiv ? LinearVectorizedTraversal : LinearTraversal,CompleteUnrolling)); VERIFY(test_assign(Matrix(),Matrix()+Matrix(), LinearTraversal,NoUnrolling)); @@ -209,6 +210,7 @@ void test_vectorization_logic() #ifdef EIGEN_VECTORIZE + CALL_SUBTEST( vectorization_logic::run() ); CALL_SUBTEST( vectorization_logic::run() ); CALL_SUBTEST( vectorization_logic::run() ); CALL_SUBTEST( vectorization_logic >::run() ); From 2bdb3b1afdbc07d54fec43edff92138f82492941 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 15 Jul 2014 11:00:16 +0200 Subject: [PATCH 0632/1103] Extend dense*sparse product unit tests --- test/sparse_product.cpp | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index 7a76acf57..6d32132bc 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -77,17 +77,27 @@ template void sparse_product() m4 = m2; refMat4 = refMat2; VERIFY_IS_APPROX(m4=m4*m3, refMat4=refMat4*refMat3); - // sparse * dense + // sparse * dense matrix VERIFY_IS_APPROX(dm4=m2*refMat3, refMat4=refMat2*refMat3); VERIFY_IS_APPROX(dm4=m2*refMat3t.transpose(), refMat4=refMat2*refMat3t.transpose()); VERIFY_IS_APPROX(dm4=m2t.transpose()*refMat3, refMat4=refMat2t.transpose()*refMat3); VERIFY_IS_APPROX(dm4=m2t.transpose()*refMat3t.transpose(), refMat4=refMat2t.transpose()*refMat3t.transpose()); + VERIFY_IS_APPROX(dm4=m2*refMat3, refMat4=refMat2*refMat3); + VERIFY_IS_APPROX(dm4=dm4+m2*refMat3, refMat4=refMat4+refMat2*refMat3); VERIFY_IS_APPROX(dm4=m2*(refMat3+refMat3), refMat4=refMat2*(refMat3+refMat3)); VERIFY_IS_APPROX(dm4=m2t.transpose()*(refMat3+refMat5)*0.5, refMat4=refMat2t.transpose()*(refMat3+refMat5)*0.5); + + // sparse * dense vector + VERIFY_IS_APPROX(dm4.col(0)=m2*refMat3.col(0), refMat4.col(0)=refMat2*refMat3.col(0)); + VERIFY_IS_APPROX(dm4.col(0)=m2*refMat3t.transpose().col(0), refMat4.col(0)=refMat2*refMat3t.transpose().col(0)); + VERIFY_IS_APPROX(dm4.col(0)=m2t.transpose()*refMat3.col(0), refMat4.col(0)=refMat2t.transpose()*refMat3.col(0)); + VERIFY_IS_APPROX(dm4.col(0)=m2t.transpose()*refMat3t.transpose().col(0), refMat4.col(0)=refMat2t.transpose()*refMat3t.transpose().col(0)); // dense * sparse VERIFY_IS_APPROX(dm4=refMat2*m3, refMat4=refMat2*refMat3); + VERIFY_IS_APPROX(dm4=dm4+refMat2*m3, refMat4=refMat4+refMat2*refMat3); + VERIFY_IS_APPROX(dm4+=refMat2*m3, refMat4+=refMat2*refMat3); VERIFY_IS_APPROX(dm4=refMat2*m3t.transpose(), refMat4=refMat2*refMat3t.transpose()); VERIFY_IS_APPROX(dm4=refMat2t.transpose()*m3, refMat4=refMat2t.transpose()*refMat3); VERIFY_IS_APPROX(dm4=refMat2t.transpose()*m3t.transpose(), refMat4=refMat2t.transpose()*refMat3t.transpose()); From a0d1aac6c5f8e3419788cdc1c63537ae8c5760fa Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 15 Jul 2014 11:15:36 +0200 Subject: [PATCH 0633/1103] Extend unit test of dense triangular solvers --- test/product_trsolve.cpp | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-) diff --git a/test/product_trsolve.cpp b/test/product_trsolve.cpp index 69892b3a8..4b97fa9d6 100644 --- a/test/product_trsolve.cpp +++ b/test/product_trsolve.cpp @@ -84,10 +84,18 @@ void test_product_trsolve() CALL_SUBTEST_4((trsolve,Dynamic,Dynamic>(internal::random(1,EIGEN_TEST_MAX_SIZE/2),internal::random(1,EIGEN_TEST_MAX_SIZE/2)))); // vectors - CALL_SUBTEST_1((trsolve(internal::random(1,EIGEN_TEST_MAX_SIZE)))); - CALL_SUBTEST_5((trsolve,Dynamic,1>(internal::random(1,EIGEN_TEST_MAX_SIZE)))); - CALL_SUBTEST_6((trsolve())); - CALL_SUBTEST_7((trsolve())); - CALL_SUBTEST_8((trsolve,4,1>())); + CALL_SUBTEST_5((trsolve(internal::random(1,EIGEN_TEST_MAX_SIZE)))); + CALL_SUBTEST_6((trsolve(internal::random(1,EIGEN_TEST_MAX_SIZE)))); + CALL_SUBTEST_7((trsolve,Dynamic,1>(internal::random(1,EIGEN_TEST_MAX_SIZE)))); + CALL_SUBTEST_8((trsolve,Dynamic,1>(internal::random(1,EIGEN_TEST_MAX_SIZE)))); + + // meta-unrollers + CALL_SUBTEST_9((trsolve())); + CALL_SUBTEST_10((trsolve())); + CALL_SUBTEST_11((trsolve,4,1>())); + CALL_SUBTEST_12((trsolve())); + CALL_SUBTEST_13((trsolve())); + CALL_SUBTEST_14((trsolve())); + } } From 0a945687b73cfe139c65ac2b5b001ee3dcc3c5d0 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Tue, 15 Jul 2014 11:02:51 +0000 Subject: [PATCH 0634/1103] Added HasDiv=1 to Altivec PacketMath.h, now vectorization_logic test passes. Added comments to the constants, indicative of the actual values --- Eigen/src/Core/arch/AltiVec/PacketMath.h | 24 +++++++++++++----------- 1 file changed, 13 insertions(+), 11 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index aadfc17be..b43e8ace3 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Konstantinos Margaritis +// Copyright (C) 2008-2014 Konstantinos Margaritis // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -62,17 +62,17 @@ typedef __vector unsigned char Packet16uc; // Define global static constants: static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; -static Packet16uc p16uc_REVERSE = {12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; -static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); -static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; +static Packet16uc p16uc_REVERSE = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; +static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); //{ 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15} +static Packet16uc p16uc_DUPLICATE = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; -static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); -static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); -static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); -static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); -static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); -static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); -static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); +static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16} +static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1} +static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0} +static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000} template<> struct packet_traits : default_packet_traits { @@ -82,8 +82,10 @@ template<> struct packet_traits : default_packet_traits Vectorizable = 1, AlignedOnScalar = 1, size=4, + HasHalfPacket=0, // FIXME check the Has* + HasDiv = 1, HasSin = 0, HasCos = 0, HasLog = 0, From 338d2ec42b99a17f4fa1759f27bf8735b4aea974 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 16 Jul 2014 13:17:06 +0200 Subject: [PATCH 0635/1103] bug #826: fix is_convertible for MSVC and add minimalistic unit test for is_convertible --- Eigen/src/Core/util/Meta.h | 23 ++++++++++++++++------- test/meta.cpp | 24 ++++++++++++++++++++++++ 2 files changed, 40 insertions(+), 7 deletions(-) diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h index 795197f59..b99b8849e 100644 --- a/Eigen/src/Core/util/Meta.h +++ b/Eigen/src/Core/util/Meta.h @@ -82,22 +82,31 @@ template struct add_const_on_value_type { typedef T template -struct is_convertible +struct is_convertible_impl { private: + struct any_conversion + { + template any_conversion(const volatile T&); + template any_conversion(T&); + }; struct yes {int a[1];}; struct no {int a[2];}; - - template - static yes test (const T&) {} - - template static no test (...) {} + + static yes test(const To&, int); + static no test(any_conversion, ...); public: static From ms_from; - enum { value = sizeof(test(ms_from))==sizeof(yes) }; + enum { value = sizeof(test(ms_from, 0))==sizeof(yes) }; }; +template +struct is_convertible +{ + enum { value = is_convertible_impl::type, + typename remove_all::type>::value }; +}; /** \internal Allows to enable/disable an overload * according to a compile time condition. diff --git a/test/meta.cpp b/test/meta.cpp index 3302c5887..b8dea68e8 100644 --- a/test/meta.cpp +++ b/test/meta.cpp @@ -9,6 +9,12 @@ #include "main.h" +template +bool check_is_convertible(const From&, const To&) +{ + return internal::is_convertible::value; +} + void test_meta() { VERIFY((internal::conditional<(3<4),internal::true_type, internal::false_type>::type::value)); @@ -52,6 +58,24 @@ void test_meta() VERIFY(( internal::is_same::type >::value)); VERIFY(( internal::is_same::type >::value)); + VERIFY(( internal::is_convertible::value )); + VERIFY(( internal::is_convertible::value )); + VERIFY(( internal::is_convertible::value )); + VERIFY((!internal::is_convertible,double>::value )); + VERIFY(( internal::is_convertible::value )); +// VERIFY((!internal::is_convertible::value )); //does not work because the conversion is prevented by a static assertion + VERIFY((!internal::is_convertible::value )); + VERIFY((!internal::is_convertible::value )); + { + float f; + MatrixXf A, B; + VectorXf a, b; + VERIFY(( check_is_convertible(a.dot(b), f) )); + VERIFY(( check_is_convertible(a.transpose()*b, f) )); + VERIFY((!check_is_convertible(A*B, f) )); + VERIFY(( check_is_convertible(A*B, A) )); + } + VERIFY(internal::meta_sqrt<1>::ret == 1); #define VERIFY_META_SQRT(X) VERIFY(internal::meta_sqrt::ret == int(std::sqrt(double(X)))) VERIFY_META_SQRT(2); From cd0b433540cf8722be689d96a915e184c9cbb22b Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 16 Jul 2014 15:41:10 +0200 Subject: [PATCH 0636/1103] Regression test for bug #714. Note that the bug only occurs on some compilers and is not fixed yet --- test/product_large.cpp | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/test/product_large.cpp b/test/product_large.cpp index 03d7bd8ed..11531aa1d 100644 --- a/test/product_large.cpp +++ b/test/product_large.cpp @@ -61,4 +61,13 @@ void test_product_large() VERIFY_IS_APPROX(r2, (mat1.row(2)*mat2).eval()); } #endif + + // Regression test for bug 714: +#ifdef EIGEN_HAS_OPENMP + std::cout << "Testing omp_set_dynamic(1)\n"; + omp_set_dynamic(1); + for(int i = 0; i < g_repeat; i++) { + CALL_SUBTEST_6( product(Matrix(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); + } +#endif } From a53f2b0e4339df24c9075fe2804917803bca0b2e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 16 Jul 2014 17:00:54 +0200 Subject: [PATCH 0637/1103] bug #838: add unit test for fill-in in sparse outer product and fix abusive fill-in. --- Eigen/src/SparseCore/SparseDenseProduct.h | 14 ++++---- test/sparse_product.cpp | 43 ++++++++++++++--------- 2 files changed, 35 insertions(+), 22 deletions(-) diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index a3772c255..295b2e81f 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -120,31 +120,33 @@ class SparseDenseOuterProduct::InnerIterator : public _LhsNes typedef typename SparseDenseOuterProduct::Index Index; public: EIGEN_STRONG_INLINE InnerIterator(const SparseDenseOuterProduct& prod, Index outer) - : Base(prod.lhs(), 0), m_outer(outer), m_factor(get(prod.rhs(), outer, typename internal::traits::StorageKind() )) - { } + : Base(prod.lhs(), 0), m_outer(outer), m_empty(false), m_factor(get(prod.rhs(), outer, typename internal::traits::StorageKind() )) + {} inline Index outer() const { return m_outer; } inline Index row() const { return Transpose ? m_outer : Base::index(); } inline Index col() const { return Transpose ? Base::index() : m_outer; } inline Scalar value() const { return Base::value() * m_factor; } + inline operator bool() const { return Base::operator bool() && !m_empty; } protected: - static Scalar get(const _RhsNested &rhs, Index outer, Dense = Dense()) + Scalar get(const _RhsNested &rhs, Index outer, Dense = Dense()) const { return rhs.coeff(outer); } - static Scalar get(const _RhsNested &rhs, Index outer, Sparse = Sparse()) + Scalar get(const _RhsNested &rhs, Index outer, Sparse = Sparse()) { typename Traits::_RhsNested::InnerIterator it(rhs, outer); - if (it && it.index()==0) + if (it && it.index()==0 && it.value()!=0) return it.value(); - + m_empty = true; return Scalar(0); } Index m_outer; + bool m_empty; Scalar m_factor; }; diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index 7a76acf57..0f52164c8 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -98,28 +98,39 @@ template void sparse_product() Index r = internal::random(0,rows-1); Index c1 = internal::random(0,cols-1); Index r1 = internal::random(0,depth-1); + DenseMatrix dm5 = DenseMatrix::Random(depth, cols); - VERIFY_IS_APPROX( m4=m2.col(c)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); - VERIFY_IS_APPROX( m4=m2.middleCols(c,1)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); - VERIFY_IS_APPROX(dm4=m2.col(c)*refMat3.col(c1).transpose(), refMat4=refMat2.col(c)*refMat3.col(c1).transpose()); + VERIFY_IS_APPROX( m4=m2.col(c)*dm5.col(c1).transpose(), refMat4=refMat2.col(c)*dm5.col(c1).transpose()); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX( m4=m2.middleCols(c,1)*dm5.col(c1).transpose(), refMat4=refMat2.col(c)*dm5.col(c1).transpose()); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX(dm4=m2.col(c)*dm5.col(c1).transpose(), refMat4=refMat2.col(c)*dm5.col(c1).transpose()); - VERIFY_IS_APPROX(m4=refMat3.col(c1)*m2.col(c).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); - VERIFY_IS_APPROX(m4=refMat3.col(c1)*m2.middleCols(c,1).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); - VERIFY_IS_APPROX(dm4=refMat3.col(c1)*m2.col(c).transpose(), refMat4=refMat3.col(c1)*refMat2.col(c).transpose()); + VERIFY_IS_APPROX(m4=dm5.col(c1)*m2.col(c).transpose(), refMat4=dm5.col(c1)*refMat2.col(c).transpose()); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX(m4=dm5.col(c1)*m2.middleCols(c,1).transpose(), refMat4=dm5.col(c1)*refMat2.col(c).transpose()); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX(dm4=dm5.col(c1)*m2.col(c).transpose(), refMat4=dm5.col(c1)*refMat2.col(c).transpose()); - VERIFY_IS_APPROX( m4=refMat3.row(r1).transpose()*m2.col(c).transpose(), refMat4=refMat3.row(r1).transpose()*refMat2.col(c).transpose()); - VERIFY_IS_APPROX(dm4=refMat3.row(r1).transpose()*m2.col(c).transpose(), refMat4=refMat3.row(r1).transpose()*refMat2.col(c).transpose()); + VERIFY_IS_APPROX( m4=dm5.row(r1).transpose()*m2.col(c).transpose(), refMat4=dm5.row(r1).transpose()*refMat2.col(c).transpose()); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX(dm4=dm5.row(r1).transpose()*m2.col(c).transpose(), refMat4=dm5.row(r1).transpose()*refMat2.col(c).transpose()); - VERIFY_IS_APPROX( m4=m2.row(r).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); - VERIFY_IS_APPROX( m4=m2.middleRows(r,1).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); - VERIFY_IS_APPROX(dm4=m2.row(r).transpose()*refMat3.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*refMat3.col(c1).transpose()); + VERIFY_IS_APPROX( m4=m2.row(r).transpose()*dm5.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*dm5.col(c1).transpose()); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX( m4=m2.middleRows(r,1).transpose()*dm5.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*dm5.col(c1).transpose()); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX(dm4=m2.row(r).transpose()*dm5.col(c1).transpose(), refMat4=refMat2.row(r).transpose()*dm5.col(c1).transpose()); - VERIFY_IS_APPROX( m4=refMat3.col(c1)*m2.row(r), refMat4=refMat3.col(c1)*refMat2.row(r)); - VERIFY_IS_APPROX( m4=refMat3.col(c1)*m2.middleRows(r,1), refMat4=refMat3.col(c1)*refMat2.row(r)); - VERIFY_IS_APPROX(dm4=refMat3.col(c1)*m2.row(r), refMat4=refMat3.col(c1)*refMat2.row(r)); + VERIFY_IS_APPROX( m4=dm5.col(c1)*m2.row(r), refMat4=dm5.col(c1)*refMat2.row(r)); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX( m4=dm5.col(c1)*m2.middleRows(r,1), refMat4=dm5.col(c1)*refMat2.row(r)); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX(dm4=dm5.col(c1)*m2.row(r), refMat4=dm5.col(c1)*refMat2.row(r)); - VERIFY_IS_APPROX( m4=refMat3.row(r1).transpose()*m2.row(r), refMat4=refMat3.row(r1).transpose()*refMat2.row(r)); - VERIFY_IS_APPROX(dm4=refMat3.row(r1).transpose()*m2.row(r), refMat4=refMat3.row(r1).transpose()*refMat2.row(r)); + VERIFY_IS_APPROX( m4=dm5.row(r1).transpose()*m2.row(r), refMat4=dm5.row(r1).transpose()*refMat2.row(r)); + VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); + VERIFY_IS_APPROX(dm4=dm5.row(r1).transpose()*m2.row(r), refMat4=dm5.row(r1).transpose()*refMat2.row(r)); } VERIFY_IS_APPROX(m6=m6*m6, refMat6=refMat6*refMat6); From 424c3ad26654c1cdaa526d73fda435ccfb0b5c79 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Jul 2014 09:41:33 +0200 Subject: [PATCH 0638/1103] bug #842: fix specialized product for mpreal --- unsupported/Eigen/MPRealSupport | 20 +++----------------- unsupported/test/mpreal_support.cpp | 4 ++-- 2 files changed, 5 insertions(+), 19 deletions(-) diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport index d4b03647d..6a65a601b 100644 --- a/unsupported/Eigen/MPRealSupport +++ b/unsupported/Eigen/MPRealSupport @@ -139,9 +139,8 @@ int main() public: typedef mpfr::mpreal ResScalar; enum { - nr = 2, // must be 2 for proper packing... + nr = 1, mr = 1, - WorkSpaceFactor = nr, LhsProgress = 1, RhsProgress = 1 }; @@ -154,9 +153,9 @@ int main() EIGEN_DONT_INLINE void operator()(mpreal* res, Index resStride, const mpreal* blockA, const mpreal* blockB, Index rows, Index depth, Index cols, mpreal alpha, - Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0, mpreal* /*unpackedB*/ = 0) + Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0) { - mpreal acc1, acc2, tmp; + mpreal acc1, tmp; if(strideA==-1) strideA = depth; if(strideB==-1) strideB = depth; @@ -165,33 +164,20 @@ int main() { Index actual_nr = (std::min)(nr,cols-j); mpreal *C1 = res + j*resStride; - mpreal *C2 = res + (j+1)*resStride; for(Index i=0; i(blockB) + j*strideB + offsetB*actual_nr; mpreal *A = const_cast(blockA) + i*strideA + offsetA; acc1 = 0; - acc2 = 0; for(Index k=0; k #include "main.h" #include #include @@ -32,12 +33,11 @@ void test_mpreal_support() VERIFY_IS_APPROX(A.array().abs2().sqrt(), A.array().abs()); VERIFY_IS_APPROX(A.array().sin(), sin(A.array())); VERIFY_IS_APPROX(A.array().cos(), cos(A.array())); - // Cholesky X = S.selfadjointView().llt().solve(B); VERIFY_IS_APPROX((S.selfadjointView()*X).eval(),B); - + // partial LU X = A.lu().solve(B); VERIFY_IS_APPROX((A*X).eval(),B); From 14c8793a70ca3cddae234216f7e927c8bcd9d998 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 17 Jul 2014 11:14:14 +0200 Subject: [PATCH 0639/1103] Remove unnecessary include --- unsupported/test/mpreal_support.cpp | 1 - 1 file changed, 1 deletion(-) diff --git a/unsupported/test/mpreal_support.cpp b/unsupported/test/mpreal_support.cpp index 13237886e..1aa9e786a 100644 --- a/unsupported/test/mpreal_support.cpp +++ b/unsupported/test/mpreal_support.cpp @@ -1,4 +1,3 @@ -#include #include "main.h" #include #include From 40b74411e4ba49990bef69175a866f53587f8c59 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Jul 2014 11:59:51 +0200 Subject: [PATCH 0640/1103] bug #842: update mpreal copy (fix compilation with clang) --- unsupported/test/mpreal/mpreal.h | 1030 ++++++++++++++++-------------- 1 file changed, 563 insertions(+), 467 deletions(-) diff --git a/unsupported/test/mpreal/mpreal.h b/unsupported/test/mpreal/mpreal.h index 70d37ab71..4c29d2ac2 100644 --- a/unsupported/test/mpreal/mpreal.h +++ b/unsupported/test/mpreal/mpreal.h @@ -1,59 +1,47 @@ /* - Multi-precision floating point number class for C++. + MPFR C++: Multi-precision floating point number class for C++. Based on MPFR library: http://mpfr.org Project homepage: http://www.holoborodko.com/pavel/mpfr Contact e-mail: pavel@holoborodko.com - Copyright (c) 2008-2013 Pavel Holoborodko + Copyright (c) 2008-2014 Pavel Holoborodko Contributors: Dmitriy Gubanov, Konstantin Holoborodko, Brian Gladman, Helmut Jarausch, Fokko Beekhof, Ulrich Mutze, Heinz van Saanen, Pere Constans, Peter van Hoof, Gael Guennebaud, Tsai Chia Cheng, - Alexei Zubanov, Jauhien Piatlicki, Victor Berger, John Westwood. + Alexei Zubanov, Jauhien Piatlicki, Victor Berger, John Westwood, + Petr Aleksandrov, Orion Poplawski, Charles Karney. - **************************************************************************** - This library is free software; you can redistribute it and/or - modify it under the terms of the GNU Lesser General Public - License as published by the Free Software Foundation; either - version 2.1 of the License, or (at your option) any later version. + Licensing: + (A) MPFR C++ is under GNU General Public License ("GPL"). + + (B) Non-free licenses may also be purchased from the author, for users who + do not want their programs protected by the GPL. - This library is distributed in the hope that it will be useful, + The non-free licenses are for users that wish to use MPFR C++ in + their products but are unwilling to release their software + under the GPL (which would require them to release source code + and allow free redistribution). + + Such users can purchase an unlimited-use license from the author. + Contact us for more details. + + GNU General Public License ("GPL") copyright permissions statement: + ************************************************************************** + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Lesser General Public License for more details. + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. - You should have received a copy of the GNU Lesser General Public - License along with this library; if not, write to the Free Software - Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA - - **************************************************************************** - Redistribution and use in source and binary forms, with or without - modification, are permitted provided that the following conditions - are met: - - 1. Redistributions of source code must retain the above copyright - notice, this list of conditions and the following disclaimer. - - 2. Redistributions in binary form must reproduce the above copyright - notice, this list of conditions and the following disclaimer in the - documentation and/or other materials provided with the distribution. - - 3. The name of the author may not be used to endorse or promote products - derived from this software without specific prior written permission. - - THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND - ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE - IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE - ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE - FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL - DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS - OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) - HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT - LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY - OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF - SUCH DAMAGE. + You should have received a copy of the GNU General Public License + along with this program. If not, see . */ #ifndef __MPREAL_H__ @@ -71,6 +59,15 @@ // Options #define MPREAL_HAVE_INT64_SUPPORT // Enable int64_t support if possible. Available only for MSVC 2010 & GCC. #define MPREAL_HAVE_MSVC_DEBUGVIEW // Enable Debugger Visualizer for "Debug" builds in MSVC. +#define MPREAL_HAVE_DYNAMIC_STD_NUMERIC_LIMITS // Enable extended std::numeric_limits specialization. + // Meaning that "digits", "round_style" and similar members are defined as functions, not constants. + // See std::numeric_limits at the end of the file for more information. + +// Library version +#define MPREAL_VERSION_MAJOR 3 +#define MPREAL_VERSION_MINOR 5 +#define MPREAL_VERSION_PATCHLEVEL 9 +#define MPREAL_VERSION_STRING "3.5.9" // Detect compiler using signatures from http://predef.sourceforge.net/ #if defined(__GNUC__) && defined(__INTEL_COMPILER) @@ -83,17 +80,29 @@ #define IsInf(x) std::isinf(x) // GNU C/C++ (and/or other compilers), just hope for C99 conformance #endif -// Detect support for r-value references (move semantic). Borrowed from Eigen. // A Clang feature extension to determine compiler features. -// We use it to determine 'cxx_rvalue_references' #ifndef __has_feature -# define __has_feature(x) 0 + #define __has_feature(x) 0 #endif +// Detect support for r-value references (move semantic). Borrowed from Eigen. #if (__has_feature(cxx_rvalue_references) || \ - defined(__GXX_EXPERIMENTAL_CXX0X__) || \ - (defined(_MSC_VER) && _MSC_VER >= 1600)) -#define MPREAL_HAVE_MOVE_SUPPORT + defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L || \ + (defined(_MSC_VER) && _MSC_VER >= 1600)) + + #define MPREAL_HAVE_MOVE_SUPPORT + + // Use fields in mpfr_t structure to check if it was initialized / set dummy initialization + #define mpfr_is_initialized(x) (0 != (x)->_mpfr_d) + #define mpfr_set_uninitialized(x) ((x)->_mpfr_d = 0 ) +#endif + +// Detect support for explicit converters. +#if (__has_feature(cxx_explicit_conversions) || \ + defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L || \ + (defined(_MSC_VER) && _MSC_VER >= 1800)) + + #define MPREAL_HAVE_EXPLICIT_CONVERTERS #endif // Detect available 64-bit capabilities @@ -112,7 +121,7 @@ #endif #elif defined (__GNUC__) && defined(__linux__) - #if defined(__amd64__) || defined(__amd64) || defined(__x86_64__) || defined(__x86_64) || defined(__ia64) || defined(__itanium__) || defined(_M_IA64) + #if defined(__amd64__) || defined(__amd64) || defined(__x86_64__) || defined(__x86_64) || defined(__ia64) || defined(__itanium__) || defined(_M_IA64) || defined (__PPC64__) #undef MPREAL_HAVE_INT64_SUPPORT // Remove all shaman dances for x64 builds since #undef MPFR_USE_INTMAX_T // GCC already supports x64 as of "long int" is 64-bit integer, nothing left to do #else @@ -175,7 +184,7 @@ public: mpreal(const int u, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t mode = mpreal::get_default_rnd()); // Construct mpreal from mpfr_t structure. - // shared = true allows to avoid deep copy, so that mpreal and 'u' shared same data & pointers. + // shared = true allows to avoid deep copy, so that mpreal and 'u' share the same data & pointers. mpreal(const mpfr_t u, bool shared = false); #if defined (MPREAL_HAVE_INT64_SUPPORT) @@ -315,14 +324,34 @@ public: friend bool operator == (const mpreal& a, const double b); // Type Conversion operators + bool toBool (mp_rnd_t mode = GMP_RNDZ) const; long toLong (mp_rnd_t mode = GMP_RNDZ) const; unsigned long toULong (mp_rnd_t mode = GMP_RNDZ) const; + float toFloat (mp_rnd_t mode = GMP_RNDN) const; double toDouble (mp_rnd_t mode = GMP_RNDN) const; long double toLDouble (mp_rnd_t mode = GMP_RNDN) const; +#if defined (MPREAL_HAVE_EXPLICIT_CONVERTERS) + explicit operator bool () const { return toBool(); } + explicit operator int () const { return toLong(); } + explicit operator long () const { return toLong(); } + explicit operator long long () const { return toLong(); } + explicit operator unsigned () const { return toULong(); } + explicit operator unsigned long () const { return toULong(); } + explicit operator unsigned long long () const { return toULong(); } + explicit operator float () const { return toFloat(); } + explicit operator double () const { return toDouble(); } + explicit operator long double () const { return toLDouble(); } +#endif + #if defined (MPREAL_HAVE_INT64_SUPPORT) int64_t toInt64 (mp_rnd_t mode = GMP_RNDZ) const; uint64_t toUInt64 (mp_rnd_t mode = GMP_RNDZ) const; + + #if defined (MPREAL_HAVE_EXPLICIT_CONVERTERS) + explicit operator int64_t () const { return toInt64(); } + explicit operator uint64_t () const { return toUInt64(); } + #endif #endif // Get raw pointers so that mpreal can be directly used in raw mpfr_* functions @@ -331,122 +360,124 @@ public: ::mpfr_srcptr mpfr_srcptr() const; // Convert mpreal to string with n significant digits in base b - // n = 0 -> convert with the maximum available digits - std::string toString(int n = 0, int b = 10, mp_rnd_t mode = mpreal::get_default_rnd()) const; + // n = -1 -> convert with the maximum available digits + std::string toString(int n = -1, int b = 10, mp_rnd_t mode = mpreal::get_default_rnd()) const; #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - std::string toString(const std::string& format) const; + std::string toString(const std::string& format) const; #endif - // Math Functions - friend const mpreal sqr (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal sqrt(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal sqrt(const unsigned long int v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal cbrt(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal root(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal pow (const mpreal& a, const mpreal& b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal pow (const mpreal& a, const mpz_t b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal pow (const mpreal& a, const unsigned long int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal pow (const mpreal& a, const long int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal pow (const unsigned long int a, const mpreal& b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal pow (const unsigned long int a, const unsigned long int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal fabs(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + std::ostream& output(std::ostream& os) const; - friend const mpreal abs(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal dim(const mpreal& a, const mpreal& b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend inline const mpreal mul_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend inline const mpreal mul_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend inline const mpreal div_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend inline const mpreal div_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + // Math Functions + friend const mpreal sqr (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal sqrt(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal sqrt(const unsigned long int v, mp_rnd_t rnd_mode); + friend const mpreal cbrt(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal root(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode); + friend const mpreal pow (const mpreal& a, const mpreal& b, mp_rnd_t rnd_mode); + friend const mpreal pow (const mpreal& a, const mpz_t b, mp_rnd_t rnd_mode); + friend const mpreal pow (const mpreal& a, const unsigned long int b, mp_rnd_t rnd_mode); + friend const mpreal pow (const mpreal& a, const long int b, mp_rnd_t rnd_mode); + friend const mpreal pow (const unsigned long int a, const mpreal& b, mp_rnd_t rnd_mode); + friend const mpreal pow (const unsigned long int a, const unsigned long int b, mp_rnd_t rnd_mode); + friend const mpreal fabs(const mpreal& v, mp_rnd_t rnd_mode); + + friend const mpreal abs(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal dim(const mpreal& a, const mpreal& b, mp_rnd_t rnd_mode); + friend inline const mpreal mul_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode); + friend inline const mpreal mul_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode); + friend inline const mpreal div_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode); + friend inline const mpreal div_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode); friend int cmpabs(const mpreal& a,const mpreal& b); - friend const mpreal log (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal log2 (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal log10(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal exp (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal exp2 (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal exp10(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal log1p(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal expm1(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal log (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal log2 (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal log10(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal exp (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal exp2 (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal exp10(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal log1p(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal expm1(const mpreal& v, mp_rnd_t rnd_mode); - friend const mpreal cos(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal sin(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal tan(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal sec(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal csc(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal cot(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend int sin_cos(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal cos(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal sin(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal tan(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal sec(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal csc(const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal cot(const mpreal& v, mp_rnd_t rnd_mode); + friend int sin_cos(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode); - friend const mpreal acos (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal asin (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal atan (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal atan2 (const mpreal& y, const mpreal& x, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal acot (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal asec (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal acsc (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal acos (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal asin (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal atan (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal atan2 (const mpreal& y, const mpreal& x, mp_rnd_t rnd_mode); + friend const mpreal acot (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal asec (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal acsc (const mpreal& v, mp_rnd_t rnd_mode); - friend const mpreal cosh (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal sinh (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal tanh (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal sech (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal csch (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal coth (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal acosh (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal asinh (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal atanh (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal acoth (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal asech (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal acsch (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal cosh (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal sinh (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal tanh (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal sech (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal csch (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal coth (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal acosh (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal asinh (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal atanh (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal acoth (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal asech (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal acsch (const mpreal& v, mp_rnd_t rnd_mode); - friend const mpreal hypot (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal hypot (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); - friend const mpreal fac_ui (unsigned long int v, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal eint (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal fac_ui (unsigned long int v, mp_prec_t prec, mp_rnd_t rnd_mode); + friend const mpreal eint (const mpreal& v, mp_rnd_t rnd_mode); - friend const mpreal gamma (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal lngamma (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal lgamma (const mpreal& v, int *signp = 0, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal zeta (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal erf (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal erfc (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal besselj0 (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal besselj1 (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal besseljn (long n, const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal bessely0 (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal bessely1 (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal besselyn (long n, const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal fma (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal fms (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal agm (const mpreal& v1, const mpreal& v2, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal sum (const mpreal tab[], unsigned long int n, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal gamma (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal lngamma (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal lgamma (const mpreal& v, int *signp, mp_rnd_t rnd_mode); + friend const mpreal zeta (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal erf (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal erfc (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal besselj0 (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal besselj1 (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal besseljn (long n, const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal bessely0 (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal bessely1 (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal besselyn (long n, const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal fma (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode); + friend const mpreal fms (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode); + friend const mpreal agm (const mpreal& v1, const mpreal& v2, mp_rnd_t rnd_mode); + friend const mpreal sum (const mpreal tab[], unsigned long int n, mp_rnd_t rnd_mode); friend int sgn(const mpreal& v); // returns -1 or +1 // MPFR 2.4.0 Specifics #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - friend int sinh_cosh (mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal li2 (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal fmod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal rec_sqrt (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend int sinh_cosh (mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal li2 (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal fmod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); + friend const mpreal rec_sqrt (const mpreal& v, mp_rnd_t rnd_mode); // MATLAB's semantic equivalents - friend const mpreal rem (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); // Remainder after division - friend const mpreal mod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); // Modulus after division + friend const mpreal rem (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); // Remainder after division + friend const mpreal mod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); // Modulus after division #endif // MPFR 3.0.0 Specifics #if (MPFR_VERSION >= MPFR_VERSION_NUM(3,0,0)) - friend const mpreal digamma (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal ai (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal urandom (gmp_randstate_t& state, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); // use gmp_randinit_default() to init state, gmp_randclear() to clear - friend const mpreal grandom (gmp_randstate_t& state, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); // use gmp_randinit_default() to init state, gmp_randclear() to clear - friend const mpreal grandom (unsigned int seed = 0); + friend const mpreal digamma (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal ai (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal urandom (gmp_randstate_t& state, mp_rnd_t rnd_mode); // use gmp_randinit_default() to init state, gmp_randclear() to clear + friend const mpreal grandom (gmp_randstate_t& state, mp_rnd_t rnd_mode); // use gmp_randinit_default() to init state, gmp_randclear() to clear + friend const mpreal grandom (unsigned int seed); #endif // Uniformly distributed random number generation in [0,1] using // Mersenne-Twister algorithm by default. // Use parameter to setup seed, e.g.: random((unsigned)time(NULL)) // Check urandom() for more precise control. - friend const mpreal random(unsigned int seed = 0); + friend const mpreal random(unsigned int seed); // Exponent and mantissa manipulation friend const mpreal frexp(const mpreal& v, mp_exp_t* exp); @@ -458,31 +489,31 @@ public: // Constants // don't forget to call mpfr_free_cache() for every thread where you are using const-functions - friend const mpreal const_log2 (mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal const_pi (mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal const_euler (mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal const_catalan (mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal const_log2 (mp_prec_t prec, mp_rnd_t rnd_mode); + friend const mpreal const_pi (mp_prec_t prec, mp_rnd_t rnd_mode); + friend const mpreal const_euler (mp_prec_t prec, mp_rnd_t rnd_mode); + friend const mpreal const_catalan (mp_prec_t prec, mp_rnd_t rnd_mode); // returns +inf iff sign>=0 otherwise -inf - friend const mpreal const_infinity(int sign = 1, mp_prec_t prec = mpreal::get_default_prec(), mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal const_infinity(int sign, mp_prec_t prec); // Output/ Input friend std::ostream& operator<<(std::ostream& os, const mpreal& v); friend std::istream& operator>>(std::istream& is, mpreal& v); // Integer Related Functions - friend const mpreal rint (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal rint (const mpreal& v, mp_rnd_t rnd_mode); friend const mpreal ceil (const mpreal& v); friend const mpreal floor(const mpreal& v); friend const mpreal round(const mpreal& v); friend const mpreal trunc(const mpreal& v); - friend const mpreal rint_ceil (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal rint_floor (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal rint_round (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal rint_trunc (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal frac (const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal remainder ( const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal remquo (long* q, const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal rint_ceil (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal rint_floor (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal rint_round (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal rint_trunc (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal frac (const mpreal& v, mp_rnd_t rnd_mode); + friend const mpreal remainder ( const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); + friend const mpreal remquo (long* q, const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); // Miscellaneous Functions friend const mpreal nexttoward (const mpreal& x, const mpreal& y); @@ -549,8 +580,8 @@ public: // Efficient swapping of two mpreal values - needed for std algorithms friend void swap(mpreal& x, mpreal& y); - friend const mpreal fmax(const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); - friend const mpreal fmin(const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + friend const mpreal fmax(const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); + friend const mpreal fmin(const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode); private: // Human friendly Debug Preview in Visual Studio. @@ -579,16 +610,16 @@ public: // Default constructor: creates mp number and initializes it to 0. inline mpreal::mpreal() { - mpfr_init2(mp,mpreal::get_default_prec()); - mpfr_set_ui(mp,0,mpreal::get_default_rnd()); + mpfr_init2 (mpfr_ptr(), mpreal::get_default_prec()); + mpfr_set_ui(mpfr_ptr(), 0, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const mpreal& u) { - mpfr_init2(mp,mpfr_get_prec(u.mp)); - mpfr_set(mp,u.mp,mpreal::get_default_rnd()); + mpfr_init2(mpfr_ptr(),mpfr_get_prec(u.mpfr_srcptr())); + mpfr_set (mpfr_ptr(),u.mpfr_srcptr(),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; } @@ -596,8 +627,7 @@ inline mpreal::mpreal(const mpreal& u) #ifdef MPREAL_HAVE_MOVE_SUPPORT inline mpreal::mpreal(mpreal&& other) { - mpfr_ptr()->_mpfr_d = 0; // make sure "other" holds no pinter to actual data - + mpfr_set_uninitialized(mpfr_ptr()); // make sure "other" holds no pinter to actual data mpfr_swap(mpfr_ptr(), other.mpfr_ptr()); MPREAL_MSVC_DEBUGVIEW_CODE; @@ -616,12 +646,12 @@ inline mpreal::mpreal(const mpfr_t u, bool shared) { if(shared) { - std::memcpy(mp, u, sizeof(mpfr_t)); + std::memcpy(mpfr_ptr(), u, sizeof(mpfr_t)); } else { - mpfr_init2(mp, mpfr_get_prec(u)); - mpfr_set (mp, u, mpreal::get_default_rnd()); + mpfr_init2(mpfr_ptr(), mpfr_get_prec(u)); + mpfr_set (mpfr_ptr(), u, mpreal::get_default_rnd()); } MPREAL_MSVC_DEBUGVIEW_CODE; @@ -629,40 +659,40 @@ inline mpreal::mpreal(const mpfr_t u, bool shared) inline mpreal::mpreal(const mpf_t u) { - mpfr_init2(mp,(mp_prec_t) mpf_get_prec(u)); // (gmp: mp_bitcnt_t) unsigned long -> long (mpfr: mp_prec_t) - mpfr_set_f(mp,u,mpreal::get_default_rnd()); + mpfr_init2(mpfr_ptr(),(mp_prec_t) mpf_get_prec(u)); // (gmp: mp_bitcnt_t) unsigned long -> long (mpfr: mp_prec_t) + mpfr_set_f(mpfr_ptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const mpz_t u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_z(mp,u,mode); + mpfr_init2(mpfr_ptr(), prec); + mpfr_set_z(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const mpq_t u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_q(mp,u,mode); + mpfr_init2(mpfr_ptr(), prec); + mpfr_set_q(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const double u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp, prec); + mpfr_init2(mpfr_ptr(), prec); #if (MPREAL_DOUBLE_BITS_OVERFLOW > -1) - if(fits_in_bits(u, MPREAL_DOUBLE_BITS_OVERFLOW)) - { - mpfr_set_d(mp, u, mode); - }else - throw conversion_overflow(); + if(fits_in_bits(u, MPREAL_DOUBLE_BITS_OVERFLOW)) + { + mpfr_set_d(mpfr_ptr(), u, mode); + }else + throw conversion_overflow(); #else - mpfr_set_d(mp, u, mode); + mpfr_set_d(mpfr_ptr(), u, mode); #endif MPREAL_MSVC_DEBUGVIEW_CODE; @@ -670,40 +700,40 @@ inline mpreal::mpreal(const double u, mp_prec_t prec, mp_rnd_t mode) inline mpreal::mpreal(const long double u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_ld(mp,u,mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_ld(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const unsigned long int u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_ui(mp,u,mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_ui(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const unsigned int u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_ui(mp,u,mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_ui(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const long int u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_si(mp,u,mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_si(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const int u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_si(mp,u,mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_si(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } @@ -711,16 +741,16 @@ inline mpreal::mpreal(const int u, mp_prec_t prec, mp_rnd_t mode) #if defined (MPREAL_HAVE_INT64_SUPPORT) inline mpreal::mpreal(const uint64_t u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_uj(mp, u, mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_uj(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const int64_t u, mp_prec_t prec, mp_rnd_t mode) { - mpfr_init2(mp,prec); - mpfr_set_sj(mp, u, mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_sj(mpfr_ptr(), u, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } @@ -728,24 +758,26 @@ inline mpreal::mpreal(const int64_t u, mp_prec_t prec, mp_rnd_t mode) inline mpreal::mpreal(const char* s, mp_prec_t prec, int base, mp_rnd_t mode) { - mpfr_init2(mp, prec); - mpfr_set_str(mp, s, base, mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_str(mpfr_ptr(), s, base, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mpreal::mpreal(const std::string& s, mp_prec_t prec, int base, mp_rnd_t mode) { - mpfr_init2(mp, prec); - mpfr_set_str(mp, s.c_str(), base, mode); + mpfr_init2 (mpfr_ptr(), prec); + mpfr_set_str(mpfr_ptr(), s.c_str(), base, mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline void mpreal::clear(::mpfr_ptr x) { - // Use undocumented field in mpfr_t structure to check if it was initialized - if(0 != x->_mpfr_d) mpfr_clear(x); +#ifdef MPREAL_HAVE_MOVE_SUPPORT + if(mpfr_is_initialized(x)) +#endif + mpfr_clear(x); } inline mpreal::~mpreal() @@ -821,6 +853,9 @@ const mpreal sqrt(const int v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); const mpreal sqrt(const long double v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); const mpreal sqrt(const double v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); +// abs +inline const mpreal abs(const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()); + ////////////////////////////////////////////////////////////////////////// // pow const mpreal pow(const mpreal& a, const unsigned int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); @@ -873,6 +908,11 @@ const mpreal pow(const double a, const unsigned int b, mp_rnd_t rnd_mode = mprea const mpreal pow(const double a, const long int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); const mpreal pow(const double a, const int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); +inline const mpreal mul_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); +inline const mpreal mul_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); +inline const mpreal div_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); +inline const mpreal div_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode = mpreal::get_default_rnd()); + ////////////////////////////////////////////////////////////////////////// // Estimate machine epsilon for the given precision // Returns smallest eps such that 1.0 + eps != 1.0 @@ -920,15 +960,15 @@ inline mpreal& mpreal::operator=(const mpreal& v) { if (this != &v) { - mp_prec_t tp = mpfr_get_prec(mp); - mp_prec_t vp = mpfr_get_prec(v.mp); + mp_prec_t tp = mpfr_get_prec( mpfr_srcptr()); + mp_prec_t vp = mpfr_get_prec(v.mpfr_srcptr()); - if(tp != vp){ - clear(mpfr_ptr()); - mpfr_init2(mpfr_ptr(), vp); - } + if(tp != vp){ + clear(mpfr_ptr()); + mpfr_init2(mpfr_ptr(), vp); + } - mpfr_set(mp, v.mp, mpreal::get_default_rnd()); + mpfr_set(mpfr_ptr(), v.mpfr_srcptr(), mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; } @@ -937,7 +977,7 @@ inline mpreal& mpreal::operator=(const mpreal& v) inline mpreal& mpreal::operator=(const mpf_t v) { - mpfr_set_f(mp, v, mpreal::get_default_rnd()); + mpfr_set_f(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -945,7 +985,7 @@ inline mpreal& mpreal::operator=(const mpf_t v) inline mpreal& mpreal::operator=(const mpz_t v) { - mpfr_set_z(mp, v, mpreal::get_default_rnd()); + mpfr_set_z(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -953,7 +993,7 @@ inline mpreal& mpreal::operator=(const mpz_t v) inline mpreal& mpreal::operator=(const mpq_t v) { - mpfr_set_q(mp, v, mpreal::get_default_rnd()); + mpfr_set_q(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -961,7 +1001,7 @@ inline mpreal& mpreal::operator=(const mpq_t v) inline mpreal& mpreal::operator=(const long double v) { - mpfr_set_ld(mp, v, mpreal::get_default_rnd()); + mpfr_set_ld(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -970,22 +1010,22 @@ inline mpreal& mpreal::operator=(const long double v) inline mpreal& mpreal::operator=(const double v) { #if (MPREAL_DOUBLE_BITS_OVERFLOW > -1) - if(fits_in_bits(v, MPREAL_DOUBLE_BITS_OVERFLOW)) - { - mpfr_set_d(mp,v,mpreal::get_default_rnd()); - }else - throw conversion_overflow(); + if(fits_in_bits(v, MPREAL_DOUBLE_BITS_OVERFLOW)) + { + mpfr_set_d(mpfr_ptr(),v,mpreal::get_default_rnd()); + }else + throw conversion_overflow(); #else - mpfr_set_d(mp,v,mpreal::get_default_rnd()); + mpfr_set_d(mpfr_ptr(),v,mpreal::get_default_rnd()); #endif - MPREAL_MSVC_DEBUGVIEW_CODE; + MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator=(const unsigned long int v) { - mpfr_set_ui(mp, v, mpreal::get_default_rnd()); + mpfr_set_ui(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -993,7 +1033,7 @@ inline mpreal& mpreal::operator=(const unsigned long int v) inline mpreal& mpreal::operator=(const unsigned int v) { - mpfr_set_ui(mp, v, mpreal::get_default_rnd()); + mpfr_set_ui(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -1001,7 +1041,7 @@ inline mpreal& mpreal::operator=(const unsigned int v) inline mpreal& mpreal::operator=(const long int v) { - mpfr_set_si(mp, v, mpreal::get_default_rnd()); + mpfr_set_si(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -1009,7 +1049,7 @@ inline mpreal& mpreal::operator=(const long int v) inline mpreal& mpreal::operator=(const int v) { - mpfr_set_si(mp, v, mpreal::get_default_rnd()); + mpfr_set_si(mpfr_ptr(), v, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; @@ -1026,11 +1066,11 @@ inline mpreal& mpreal::operator=(const char* s) mpfr_t t; - mpfr_init2(t, mpfr_get_prec(mp)); + mpfr_init2(t, mpfr_get_prec(mpfr_srcptr())); if(0 == mpfr_set_str(t, s, 10, mpreal::get_default_rnd())) { - mpfr_set(mp, t, mpreal::get_default_rnd()); + mpfr_set(mpfr_ptr(), t, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; } @@ -1049,11 +1089,11 @@ inline mpreal& mpreal::operator=(const std::string& s) mpfr_t t; - mpfr_init2(t, mpfr_get_prec(mp)); + mpfr_init2(t, mpfr_get_prec(mpfr_srcptr())); if(0 == mpfr_set_str(t, s.c_str(), 10, mpreal::get_default_rnd())) { - mpfr_set(mp, t, mpreal::get_default_rnd()); + mpfr_set(mpfr_ptr(), t, mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; } @@ -1066,7 +1106,7 @@ inline mpreal& mpreal::operator=(const std::string& s) // + Addition inline mpreal& mpreal::operator+=(const mpreal& v) { - mpfr_add(mp,mp,v.mp,mpreal::get_default_rnd()); + mpfr_add(mpfr_ptr(), mpfr_srcptr(), v.mpfr_srcptr(), mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1080,14 +1120,14 @@ inline mpreal& mpreal::operator+=(const mpf_t u) inline mpreal& mpreal::operator+=(const mpz_t u) { - mpfr_add_z(mp,mp,u,mpreal::get_default_rnd()); + mpfr_add_z(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator+=(const mpq_t u) { - mpfr_add_q(mp,mp,u,mpreal::get_default_rnd()); + mpfr_add_q(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1102,7 +1142,7 @@ inline mpreal& mpreal::operator+= (const long double u) inline mpreal& mpreal::operator+= (const double u) { #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - mpfr_add_d(mp,mp,u,mpreal::get_default_rnd()); + mpfr_add_d(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); #else *this += mpreal(u); #endif @@ -1113,28 +1153,28 @@ inline mpreal& mpreal::operator+= (const double u) inline mpreal& mpreal::operator+=(const unsigned long int u) { - mpfr_add_ui(mp,mp,u,mpreal::get_default_rnd()); + mpfr_add_ui(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator+=(const unsigned int u) { - mpfr_add_ui(mp,mp,u,mpreal::get_default_rnd()); + mpfr_add_ui(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator+=(const long int u) { - mpfr_add_si(mp,mp,u,mpreal::get_default_rnd()); + mpfr_add_si(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator+=(const int u) { - mpfr_add_si(mp,mp,u,mpreal::get_default_rnd()); + mpfr_add_si(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1154,9 +1194,9 @@ inline const mpreal mpreal::operator+()const { return mpreal(*this); } inline const mpreal operator+(const mpreal& a, const mpreal& b) { - mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); - mpfr_add(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); - return c; + mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpfr_add(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); + return c; } inline mpreal& mpreal::operator++() @@ -1187,21 +1227,21 @@ inline const mpreal mpreal::operator-- (int) // - Subtraction inline mpreal& mpreal::operator-=(const mpreal& v) { - mpfr_sub(mp,mp,v.mp,mpreal::get_default_rnd()); + mpfr_sub(mpfr_ptr(),mpfr_srcptr(),v.mpfr_srcptr(),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator-=(const mpz_t v) { - mpfr_sub_z(mp,mp,v,mpreal::get_default_rnd()); + mpfr_sub_z(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator-=(const mpq_t v) { - mpfr_sub_q(mp,mp,v,mpreal::get_default_rnd()); + mpfr_sub_q(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1216,7 +1256,7 @@ inline mpreal& mpreal::operator-=(const long double v) inline mpreal& mpreal::operator-=(const double v) { #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - mpfr_sub_d(mp,mp,v,mpreal::get_default_rnd()); + mpfr_sub_d(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); #else *this -= mpreal(v); #endif @@ -1227,28 +1267,28 @@ inline mpreal& mpreal::operator-=(const double v) inline mpreal& mpreal::operator-=(const unsigned long int v) { - mpfr_sub_ui(mp,mp,v,mpreal::get_default_rnd()); + mpfr_sub_ui(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator-=(const unsigned int v) { - mpfr_sub_ui(mp,mp,v,mpreal::get_default_rnd()); + mpfr_sub_ui(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator-=(const long int v) { - mpfr_sub_si(mp,mp,v,mpreal::get_default_rnd()); + mpfr_sub_si(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator-=(const int v) { - mpfr_sub_si(mp,mp,v,mpreal::get_default_rnd()); + mpfr_sub_si(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1256,15 +1296,15 @@ inline mpreal& mpreal::operator-=(const int v) inline const mpreal mpreal::operator-()const { mpreal u(*this); - mpfr_neg(u.mp,u.mp,mpreal::get_default_rnd()); + mpfr_neg(u.mpfr_ptr(),u.mpfr_srcptr(),mpreal::get_default_rnd()); return u; } inline const mpreal operator-(const mpreal& a, const mpreal& b) { - mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); - mpfr_sub(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); - return c; + mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpfr_sub(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); + return c; } inline const mpreal operator-(const double b, const mpreal& a) @@ -1312,21 +1352,21 @@ inline const mpreal operator-(const int b, const mpreal& a) // * Multiplication inline mpreal& mpreal::operator*= (const mpreal& v) { - mpfr_mul(mp,mp,v.mp,mpreal::get_default_rnd()); + mpfr_mul(mpfr_ptr(),mpfr_srcptr(),v.mpfr_srcptr(),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator*=(const mpz_t v) { - mpfr_mul_z(mp,mp,v,mpreal::get_default_rnd()); + mpfr_mul_z(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator*=(const mpq_t v) { - mpfr_mul_q(mp,mp,v,mpreal::get_default_rnd()); + mpfr_mul_q(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1341,7 +1381,7 @@ inline mpreal& mpreal::operator*=(const long double v) inline mpreal& mpreal::operator*=(const double v) { #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - mpfr_mul_d(mp,mp,v,mpreal::get_default_rnd()); + mpfr_mul_d(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); #else *this *= mpreal(v); #endif @@ -1351,58 +1391,58 @@ inline mpreal& mpreal::operator*=(const double v) inline mpreal& mpreal::operator*=(const unsigned long int v) { - mpfr_mul_ui(mp,mp,v,mpreal::get_default_rnd()); + mpfr_mul_ui(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator*=(const unsigned int v) { - mpfr_mul_ui(mp,mp,v,mpreal::get_default_rnd()); + mpfr_mul_ui(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator*=(const long int v) { - mpfr_mul_si(mp,mp,v,mpreal::get_default_rnd()); + mpfr_mul_si(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator*=(const int v) { - mpfr_mul_si(mp,mp,v,mpreal::get_default_rnd()); + mpfr_mul_si(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline const mpreal operator*(const mpreal& a, const mpreal& b) { - mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); - mpfr_mul(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); - return c; + mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); + mpfr_mul(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); + return c; } ////////////////////////////////////////////////////////////////////////// // / Division inline mpreal& mpreal::operator/=(const mpreal& v) { - mpfr_div(mp,mp,v.mp,mpreal::get_default_rnd()); + mpfr_div(mpfr_ptr(),mpfr_srcptr(),v.mpfr_srcptr(),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator/=(const mpz_t v) { - mpfr_div_z(mp,mp,v,mpreal::get_default_rnd()); + mpfr_div_z(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator/=(const mpq_t v) { - mpfr_div_q(mp,mp,v,mpreal::get_default_rnd()); + mpfr_div_q(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1417,7 +1457,7 @@ inline mpreal& mpreal::operator/=(const long double v) inline mpreal& mpreal::operator/=(const double v) { #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - mpfr_div_d(mp,mp,v,mpreal::get_default_rnd()); + mpfr_div_d(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); #else *this /= mpreal(v); #endif @@ -1427,72 +1467,72 @@ inline mpreal& mpreal::operator/=(const double v) inline mpreal& mpreal::operator/=(const unsigned long int v) { - mpfr_div_ui(mp,mp,v,mpreal::get_default_rnd()); + mpfr_div_ui(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator/=(const unsigned int v) { - mpfr_div_ui(mp,mp,v,mpreal::get_default_rnd()); + mpfr_div_ui(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator/=(const long int v) { - mpfr_div_si(mp,mp,v,mpreal::get_default_rnd()); + mpfr_div_si(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator/=(const int v) { - mpfr_div_si(mp,mp,v,mpreal::get_default_rnd()); + mpfr_div_si(mpfr_ptr(),mpfr_srcptr(),v,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline const mpreal operator/(const mpreal& a, const mpreal& b) { - mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_ptr()), mpfr_get_prec(b.mpfr_ptr()))); - mpfr_div(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); - return c; + mpreal c(0, (std::max)(mpfr_get_prec(a.mpfr_srcptr()), mpfr_get_prec(b.mpfr_srcptr()))); + mpfr_div(c.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), mpreal::get_default_rnd()); + return c; } inline const mpreal operator/(const unsigned long int b, const mpreal& a) { - mpreal x(0, mpfr_get_prec(a.mpfr_ptr())); + mpreal x(0, mpfr_get_prec(a.mpfr_srcptr())); mpfr_ui_div(x.mpfr_ptr(), b, a.mpfr_srcptr(), mpreal::get_default_rnd()); return x; } inline const mpreal operator/(const unsigned int b, const mpreal& a) { - mpreal x(0, mpfr_get_prec(a.mpfr_ptr())); + mpreal x(0, mpfr_get_prec(a.mpfr_srcptr())); mpfr_ui_div(x.mpfr_ptr(), b, a.mpfr_srcptr(), mpreal::get_default_rnd()); return x; } inline const mpreal operator/(const long int b, const mpreal& a) { - mpreal x(0, mpfr_get_prec(a.mpfr_ptr())); - mpfr_si_div(x.mpfr_ptr(), b, a.mpfr_srcptr(),mpreal::get_default_rnd()); + mpreal x(0, mpfr_get_prec(a.mpfr_srcptr())); + mpfr_si_div(x.mpfr_ptr(), b, a.mpfr_srcptr(), mpreal::get_default_rnd()); return x; } inline const mpreal operator/(const int b, const mpreal& a) { - mpreal x(0, mpfr_get_prec(a.mpfr_ptr())); - mpfr_si_div(x.mpfr_ptr(), b, a.mpfr_srcptr(),mpreal::get_default_rnd()); + mpreal x(0, mpfr_get_prec(a.mpfr_srcptr())); + mpfr_si_div(x.mpfr_ptr(), b, a.mpfr_srcptr(), mpreal::get_default_rnd()); return x; } inline const mpreal operator/(const double b, const mpreal& a) { #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - mpreal x(0, mpfr_get_prec(a.mpfr_ptr())); - mpfr_d_div(x.mpfr_ptr(), b, a.mpfr_srcptr(),mpreal::get_default_rnd()); + mpreal x(0, mpfr_get_prec(a.mpfr_srcptr())); + mpfr_d_div(x.mpfr_ptr(), b, a.mpfr_srcptr(), mpreal::get_default_rnd()); return x; #else mpreal x(0, mpfr_get_prec(a.mpfr_ptr())); @@ -1505,56 +1545,56 @@ inline const mpreal operator/(const double b, const mpreal& a) // Shifts operators - Multiplication/Division by power of 2 inline mpreal& mpreal::operator<<=(const unsigned long int u) { - mpfr_mul_2ui(mp,mp,u,mpreal::get_default_rnd()); + mpfr_mul_2ui(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator<<=(const unsigned int u) { - mpfr_mul_2ui(mp,mp,static_cast(u),mpreal::get_default_rnd()); + mpfr_mul_2ui(mpfr_ptr(),mpfr_srcptr(),static_cast(u),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator<<=(const long int u) { - mpfr_mul_2si(mp,mp,u,mpreal::get_default_rnd()); + mpfr_mul_2si(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator<<=(const int u) { - mpfr_mul_2si(mp,mp,static_cast(u),mpreal::get_default_rnd()); + mpfr_mul_2si(mpfr_ptr(),mpfr_srcptr(),static_cast(u),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator>>=(const unsigned long int u) { - mpfr_div_2ui(mp,mp,u,mpreal::get_default_rnd()); + mpfr_div_2ui(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator>>=(const unsigned int u) { - mpfr_div_2ui(mp,mp,static_cast(u),mpreal::get_default_rnd()); + mpfr_div_2ui(mpfr_ptr(),mpfr_srcptr(),static_cast(u),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator>>=(const long int u) { - mpfr_div_2si(mp,mp,u,mpreal::get_default_rnd()); + mpfr_div_2si(mpfr_ptr(),mpfr_srcptr(),u,mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::operator>>=(const int u) { - mpfr_div_2si(mp,mp,static_cast(u),mpreal::get_default_rnd()); + mpfr_div_2si(mpfr_ptr(),mpfr_srcptr(),static_cast(u),mpreal::get_default_rnd()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1603,7 +1643,7 @@ inline const mpreal operator>>(const mpreal& v, const int k) inline const mpreal mul_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode) { mpreal x(v); - mpfr_mul_2ui(x.mp,v.mp,k,rnd_mode); + mpfr_mul_2ui(x.mpfr_ptr(),v.mpfr_srcptr(),k,rnd_mode); return x; } @@ -1611,61 +1651,63 @@ inline const mpreal mul_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_m inline const mpreal mul_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode) { mpreal x(v); - mpfr_mul_2si(x.mp,v.mp,k,rnd_mode); + mpfr_mul_2si(x.mpfr_ptr(),v.mpfr_srcptr(),k,rnd_mode); return x; } inline const mpreal div_2ui(const mpreal& v, unsigned long int k, mp_rnd_t rnd_mode) { mpreal x(v); - mpfr_div_2ui(x.mp,v.mp,k,rnd_mode); + mpfr_div_2ui(x.mpfr_ptr(),v.mpfr_srcptr(),k,rnd_mode); return x; } inline const mpreal div_2si(const mpreal& v, long int k, mp_rnd_t rnd_mode) { mpreal x(v); - mpfr_div_2si(x.mp,v.mp,k,rnd_mode); + mpfr_div_2si(x.mpfr_ptr(),v.mpfr_srcptr(),k,rnd_mode); return x; } ////////////////////////////////////////////////////////////////////////// //Boolean operators -inline bool operator > (const mpreal& a, const mpreal& b){ return (mpfr_greater_p(a.mp,b.mp) !=0); } -inline bool operator >= (const mpreal& a, const mpreal& b){ return (mpfr_greaterequal_p(a.mp,b.mp) !=0); } -inline bool operator < (const mpreal& a, const mpreal& b){ return (mpfr_less_p(a.mp,b.mp) !=0); } -inline bool operator <= (const mpreal& a, const mpreal& b){ return (mpfr_lessequal_p(a.mp,b.mp) !=0); } -inline bool operator == (const mpreal& a, const mpreal& b){ return (mpfr_equal_p(a.mp,b.mp) !=0); } -inline bool operator != (const mpreal& a, const mpreal& b){ return (mpfr_lessgreater_p(a.mp,b.mp) !=0); } +inline bool operator > (const mpreal& a, const mpreal& b){ return (mpfr_greater_p (a.mpfr_srcptr(),b.mpfr_srcptr()) !=0 ); } +inline bool operator >= (const mpreal& a, const mpreal& b){ return (mpfr_greaterequal_p (a.mpfr_srcptr(),b.mpfr_srcptr()) !=0 ); } +inline bool operator < (const mpreal& a, const mpreal& b){ return (mpfr_less_p (a.mpfr_srcptr(),b.mpfr_srcptr()) !=0 ); } +inline bool operator <= (const mpreal& a, const mpreal& b){ return (mpfr_lessequal_p (a.mpfr_srcptr(),b.mpfr_srcptr()) !=0 ); } +inline bool operator == (const mpreal& a, const mpreal& b){ return (mpfr_equal_p (a.mpfr_srcptr(),b.mpfr_srcptr()) !=0 ); } +inline bool operator != (const mpreal& a, const mpreal& b){ return (mpfr_lessgreater_p (a.mpfr_srcptr(),b.mpfr_srcptr()) !=0 ); } -inline bool operator == (const mpreal& a, const unsigned long int b ){ return (mpfr_cmp_ui(a.mp,b) == 0); } -inline bool operator == (const mpreal& a, const unsigned int b ){ return (mpfr_cmp_ui(a.mp,b) == 0); } -inline bool operator == (const mpreal& a, const long int b ){ return (mpfr_cmp_si(a.mp,b) == 0); } -inline bool operator == (const mpreal& a, const int b ){ return (mpfr_cmp_si(a.mp,b) == 0); } -inline bool operator == (const mpreal& a, const long double b ){ return (mpfr_cmp_ld(a.mp,b) == 0); } -inline bool operator == (const mpreal& a, const double b ){ return (mpfr_cmp_d(a.mp,b) == 0); } +inline bool operator == (const mpreal& a, const unsigned long int b ){ return (mpfr_cmp_ui(a.mpfr_srcptr(),b) == 0 ); } +inline bool operator == (const mpreal& a, const unsigned int b ){ return (mpfr_cmp_ui(a.mpfr_srcptr(),b) == 0 ); } +inline bool operator == (const mpreal& a, const long int b ){ return (mpfr_cmp_si(a.mpfr_srcptr(),b) == 0 ); } +inline bool operator == (const mpreal& a, const int b ){ return (mpfr_cmp_si(a.mpfr_srcptr(),b) == 0 ); } +inline bool operator == (const mpreal& a, const long double b ){ return (mpfr_cmp_ld(a.mpfr_srcptr(),b) == 0 ); } +inline bool operator == (const mpreal& a, const double b ){ return (mpfr_cmp_d (a.mpfr_srcptr(),b) == 0 ); } -inline bool isnan (const mpreal& v){ return (mpfr_nan_p(v.mp) != 0); } -inline bool isinf (const mpreal& v){ return (mpfr_inf_p(v.mp) != 0); } -inline bool isfinite (const mpreal& v){ return (mpfr_number_p(v.mp) != 0); } -inline bool iszero (const mpreal& v){ return (mpfr_zero_p(v.mp) != 0); } -inline bool isint (const mpreal& v){ return (mpfr_integer_p(v.mp) != 0); } +inline bool isnan (const mpreal& op){ return (mpfr_nan_p (op.mpfr_srcptr()) != 0 ); } +inline bool isinf (const mpreal& op){ return (mpfr_inf_p (op.mpfr_srcptr()) != 0 ); } +inline bool isfinite (const mpreal& op){ return (mpfr_number_p (op.mpfr_srcptr()) != 0 ); } +inline bool iszero (const mpreal& op){ return (mpfr_zero_p (op.mpfr_srcptr()) != 0 ); } +inline bool isint (const mpreal& op){ return (mpfr_integer_p(op.mpfr_srcptr()) != 0 ); } #if (MPFR_VERSION >= MPFR_VERSION_NUM(3,0,0)) -inline bool isregular(const mpreal& v){ return (mpfr_regular_p(v.mp));} +inline bool isregular(const mpreal& op){ return (mpfr_regular_p(op.mpfr_srcptr()));} #endif ////////////////////////////////////////////////////////////////////////// // Type Converters -inline long mpreal::toLong (mp_rnd_t mode) const { return mpfr_get_si(mp, mode); } -inline unsigned long mpreal::toULong (mp_rnd_t mode) const { return mpfr_get_ui(mp, mode); } -inline double mpreal::toDouble (mp_rnd_t mode) const { return mpfr_get_d (mp, mode); } -inline long double mpreal::toLDouble(mp_rnd_t mode) const { return mpfr_get_ld(mp, mode); } +inline bool mpreal::toBool (mp_rnd_t mode) const { return mpfr_zero_p (mpfr_srcptr()) == 0; } +inline long mpreal::toLong (mp_rnd_t mode) const { return mpfr_get_si (mpfr_srcptr(), mode); } +inline unsigned long mpreal::toULong (mp_rnd_t mode) const { return mpfr_get_ui (mpfr_srcptr(), mode); } +inline float mpreal::toFloat (mp_rnd_t mode) const { return mpfr_get_flt(mpfr_srcptr(), mode); } +inline double mpreal::toDouble (mp_rnd_t mode) const { return mpfr_get_d (mpfr_srcptr(), mode); } +inline long double mpreal::toLDouble(mp_rnd_t mode) const { return mpfr_get_ld (mpfr_srcptr(), mode); } #if defined (MPREAL_HAVE_INT64_SUPPORT) -inline int64_t mpreal::toInt64 (mp_rnd_t mode) const{ return mpfr_get_sj(mp, mode); } -inline uint64_t mpreal::toUInt64(mp_rnd_t mode) const{ return mpfr_get_uj(mp, mode); } +inline int64_t mpreal::toInt64 (mp_rnd_t mode) const{ return mpfr_get_sj(mpfr_srcptr(), mode); } +inline uint64_t mpreal::toUInt64(mp_rnd_t mode) const{ return mpfr_get_uj(mpfr_srcptr(), mode); } #endif inline ::mpfr_ptr mpreal::mpfr_ptr() { return mp; } @@ -1689,7 +1731,7 @@ inline std::string mpreal::toString(const std::string& format) const if( !format.empty() ) { - if(!(mpfr_asprintf(&s,format.c_str(),mp) < 0)) + if(!(mpfr_asprintf(&s, format.c_str(), mpfr_srcptr()) < 0)) { out = std::string(s); @@ -1704,20 +1746,19 @@ inline std::string mpreal::toString(const std::string& format) const inline std::string mpreal::toString(int n, int b, mp_rnd_t mode) const { + // TODO: Add extended format specification (f, e, rounding mode) as it done in output operator (void)b; (void)mode; #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) - // Use MPFR native function for output - char format[128]; - int digits; + std::ostringstream format; - digits = n > 0 ? n : bits2digits(mpfr_get_prec(mp)); + int digits = (n >= 0) ? n : bits2digits(mpfr_get_prec(mpfr_srcptr())); + + format << "%." << digits << "RNg"; - sprintf(format,"%%.%dRNg",digits); // Default format - - return toString(std::string(format)); + return toString(format.str()); #else @@ -1735,8 +1776,8 @@ inline std::string mpreal::toString(int n, int b, mp_rnd_t mode) const if(mpfr_zero_p(mp)) return "0"; if(mpfr_nan_p(mp)) return "NaN"; - s = mpfr_get_str(NULL,&exp,b,0,mp,mode); - ns = mpfr_get_str(NULL,&exp,b,n,mp,mode); + s = mpfr_get_str(NULL, &exp, b, 0, mp, mode); + ns = mpfr_get_str(NULL, &exp, b, (std::max)(0,n), mp, mode); if(s!=NULL && ns!=NULL) { @@ -1821,17 +1862,43 @@ inline std::string mpreal::toString(int n, int b, mp_rnd_t mode) const ////////////////////////////////////////////////////////////////////////// // I/O +inline std::ostream& mpreal::output(std::ostream& os) const +{ + std::ostringstream format; + const std::ios::fmtflags flags = os.flags(); + + format << ((flags & std::ios::showpos) ? "%+" : "%"); + if (os.precision() >= 0) + format << '.' << os.precision() << "R*" + << ((flags & std::ios::floatfield) == std::ios::fixed ? 'f' : + (flags & std::ios::floatfield) == std::ios::scientific ? 'e' : + 'g'); + else + format << "R*e"; + + char *s = NULL; + if(!(mpfr_asprintf(&s, format.str().c_str(), + mpfr::mpreal::get_default_rnd(), + mpfr_srcptr()) + < 0)) + { + os << std::string(s); + mpfr_free_str(s); + } + return os; +} + inline std::ostream& operator<<(std::ostream& os, const mpreal& v) { - return os<(os.precision())); + return v.output(os); } inline std::istream& operator>>(std::istream &is, mpreal& v) { - // ToDo, use cout::hexfloat and other flags to setup base + // TODO: use cout::hexfloat and other flags to setup base std::string tmp; is >> tmp; - mpfr_set_str(v.mp, tmp.c_str(), 10, mpreal::get_default_rnd()); + mpfr_set_str(v.mpfr_ptr(), tmp.c_str(), 10, mpreal::get_default_rnd()); return is; } @@ -1844,53 +1911,53 @@ inline mp_prec_t digits2bits(int d) { const double LOG2_10 = 3.3219280948873624; - return (mp_prec_t) std::ceil( d * LOG2_10 ); + return mp_prec_t(std::ceil( d * LOG2_10 )); } inline int bits2digits(mp_prec_t b) { const double LOG10_2 = 0.30102999566398119; - return (int) std::floor( b * LOG10_2 ); + return int(std::floor( b * LOG10_2 )); } ////////////////////////////////////////////////////////////////////////// // Set/Get number properties -inline int sgn(const mpreal& v) +inline int sgn(const mpreal& op) { - int r = mpfr_signbit(v.mp); - return (r>0?-1:1); + int r = mpfr_signbit(op.mpfr_srcptr()); + return (r > 0? -1 : 1); } inline mpreal& mpreal::setSign(int sign, mp_rnd_t RoundingMode) { - mpfr_setsign(mp,mp,(sign < 0 ? 1 : 0),RoundingMode); + mpfr_setsign(mpfr_ptr(), mpfr_srcptr(), (sign < 0 ? 1 : 0), RoundingMode); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline int mpreal::getPrecision() const { - return mpfr_get_prec(mp); + return int(mpfr_get_prec(mpfr_srcptr())); } inline mpreal& mpreal::setPrecision(int Precision, mp_rnd_t RoundingMode) { - mpfr_prec_round(mp, Precision, RoundingMode); + mpfr_prec_round(mpfr_ptr(), Precision, RoundingMode); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::setInf(int sign) { - mpfr_set_inf(mp,sign); + mpfr_set_inf(mpfr_ptr(), sign); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } inline mpreal& mpreal::setNan() { - mpfr_set_nan(mp); + mpfr_set_nan(mpfr_ptr()); MPREAL_MSVC_DEBUGVIEW_CODE; return *this; } @@ -1899,9 +1966,9 @@ inline mpreal& mpreal::setZero(int sign) { #if (MPFR_VERSION >= MPFR_VERSION_NUM(3,0,0)) - mpfr_set_zero(mp, sign); + mpfr_set_zero(mpfr_ptr(), sign); #else - mpfr_set_si(mp, 0, (mpfr_get_default_rounding_mode)()); + mpfr_set_si(mpfr_ptr(), 0, (mpfr_get_default_rounding_mode)()); setSign(sign); #endif @@ -1911,23 +1978,23 @@ inline mpreal& mpreal::setZero(int sign) inline mp_prec_t mpreal::get_prec() const { - return mpfr_get_prec(mp); + return mpfr_get_prec(mpfr_srcptr()); } inline void mpreal::set_prec(mp_prec_t prec, mp_rnd_t rnd_mode) { - mpfr_prec_round(mp,prec,rnd_mode); + mpfr_prec_round(mpfr_ptr(),prec,rnd_mode); MPREAL_MSVC_DEBUGVIEW_CODE; } inline mp_exp_t mpreal::get_exp () { - return mpfr_get_exp(mp); + return mpfr_get_exp(mpfr_srcptr()); } inline int mpreal::set_exp (mp_exp_t e) { - int x = mpfr_set_exp(mp, e); + int x = mpfr_set_exp(mpfr_ptr(), e); MPREAL_MSVC_DEBUGVIEW_CODE; return x; } @@ -1945,7 +2012,7 @@ inline const mpreal ldexp(const mpreal& v, mp_exp_t exp) mpreal x(v); // rounding is not important since we just increasing the exponent - mpfr_mul_2si(x.mp,x.mp,exp,mpreal::get_default_rnd()); + mpfr_mul_2si(x.mpfr_ptr(), x.mpfr_srcptr(), exp, mpreal::get_default_rnd()); return x; } @@ -1960,9 +2027,9 @@ inline mpreal machine_epsilon(const mpreal& x) /* the smallest eps such that x + eps != x */ if( x < 0) { - return nextabove(-x)+x; + return nextabove(-x) + x; }else{ - return nextabove(x)-x; + return nextabove( x) - x; } } @@ -1997,22 +2064,22 @@ inline bool isEqualFuzzy(const mpreal& a, const mpreal& b) inline const mpreal modf(const mpreal& v, mpreal& n) { - mpreal frac(v); + mpreal f(v); // rounding is not important since we are using the same number - mpfr_frac(frac.mp,frac.mp,mpreal::get_default_rnd()); - mpfr_trunc(n.mp,v.mp); - return frac; + mpfr_frac (f.mpfr_ptr(),f.mpfr_srcptr(),mpreal::get_default_rnd()); + mpfr_trunc(n.mpfr_ptr(),v.mpfr_srcptr()); + return f; } inline int mpreal::check_range (int t, mp_rnd_t rnd_mode) { - return mpfr_check_range(mp,t,rnd_mode); + return mpfr_check_range(mpfr_ptr(),t,rnd_mode); } inline int mpreal::subnormalize (int t,mp_rnd_t rnd_mode) { - int r = mpfr_subnormalize(mp,t,rnd_mode); + int r = mpfr_subnormalize(mpfr_ptr(),t,rnd_mode); MPREAL_MSVC_DEBUGVIEW_CODE; return r; } @@ -2065,8 +2132,11 @@ inline mp_exp_t mpreal::get_emax_max (void) mpfr_##f(y.mpfr_ptr(), x.mpfr_srcptr(), r); \ return y; -inline const mpreal sqr (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(sqr ); } -inline const mpreal sqrt (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(sqrt); } +inline const mpreal sqr (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) +{ MPREAL_UNARY_MATH_FUNCTION_BODY(sqr ); } + +inline const mpreal sqrt (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) +{ MPREAL_UNARY_MATH_FUNCTION_BODY(sqrt); } inline const mpreal sqrt(const unsigned long int x, mp_rnd_t r) { @@ -2092,14 +2162,14 @@ inline const mpreal sqrt(const int v, mp_rnd_t rnd_mode) else return mpreal().setNan(); // NaN } -inline const mpreal root(const mpreal& x, unsigned long int k, mp_rnd_t r) +inline const mpreal root(const mpreal& x, unsigned long int k, mp_rnd_t r = mpreal::get_default_rnd()) { mpreal y(0, mpfr_get_prec(x.mpfr_srcptr())); mpfr_root(y.mpfr_ptr(), x.mpfr_srcptr(), k, r); return y; } -inline const mpreal dim(const mpreal& a, const mpreal& b, mp_rnd_t r) +inline const mpreal dim(const mpreal& a, const mpreal& b, mp_rnd_t r = mpreal::get_default_rnd()) { mpreal y(0, mpfr_get_prec(a.mpfr_srcptr())); mpfr_dim(y.mpfr_ptr(), a.mpfr_srcptr(), b.mpfr_srcptr(), r); @@ -2111,7 +2181,7 @@ inline int cmpabs(const mpreal& a,const mpreal& b) return mpfr_cmpabs(a.mpfr_ptr(), b.mpfr_srcptr()); } -inline int sin_cos(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode) +inline int sin_cos(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { return mpfr_sin_cos(s.mpfr_ptr(), c.mpfr_ptr(), v.mpfr_srcptr(), rnd_mode); } @@ -2119,84 +2189,85 @@ inline int sin_cos(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode) inline const mpreal sqrt (const long double v, mp_rnd_t rnd_mode) { return sqrt(mpreal(v),rnd_mode); } inline const mpreal sqrt (const double v, mp_rnd_t rnd_mode) { return sqrt(mpreal(v),rnd_mode); } -inline const mpreal cbrt (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(cbrt ); } -inline const mpreal fabs (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(abs ); } -inline const mpreal abs (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(abs ); } -inline const mpreal log (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(log ); } -inline const mpreal log2 (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(log2 ); } -inline const mpreal log10 (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(log10); } -inline const mpreal exp (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(exp ); } -inline const mpreal exp2 (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(exp2 ); } -inline const mpreal exp10 (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(exp10); } -inline const mpreal cos (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(cos ); } -inline const mpreal sin (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(sin ); } -inline const mpreal tan (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(tan ); } -inline const mpreal sec (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(sec ); } -inline const mpreal csc (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(csc ); } -inline const mpreal cot (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(cot ); } -inline const mpreal acos (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(acos ); } -inline const mpreal asin (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(asin ); } -inline const mpreal atan (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(atan ); } +inline const mpreal cbrt (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(cbrt ); } +inline const mpreal fabs (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(abs ); } +inline const mpreal abs (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(abs ); } +inline const mpreal log (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(log ); } +inline const mpreal log2 (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(log2 ); } +inline const mpreal log10 (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(log10); } +inline const mpreal exp (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(exp ); } +inline const mpreal exp2 (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(exp2 ); } +inline const mpreal exp10 (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(exp10); } +inline const mpreal cos (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(cos ); } +inline const mpreal sin (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(sin ); } +inline const mpreal tan (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(tan ); } +inline const mpreal sec (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(sec ); } +inline const mpreal csc (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(csc ); } +inline const mpreal cot (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(cot ); } +inline const mpreal acos (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(acos ); } +inline const mpreal asin (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(asin ); } +inline const mpreal atan (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(atan ); } -inline const mpreal acot (const mpreal& v, mp_rnd_t r) { return atan (1/v, r); } -inline const mpreal asec (const mpreal& v, mp_rnd_t r) { return acos (1/v, r); } -inline const mpreal acsc (const mpreal& v, mp_rnd_t r) { return asin (1/v, r); } -inline const mpreal acoth (const mpreal& v, mp_rnd_t r) { return atanh(1/v, r); } -inline const mpreal asech (const mpreal& v, mp_rnd_t r) { return acosh(1/v, r); } -inline const mpreal acsch (const mpreal& v, mp_rnd_t r) { return asinh(1/v, r); } +inline const mpreal acot (const mpreal& v, mp_rnd_t r = mpreal::get_default_rnd()) { return atan (1/v, r); } +inline const mpreal asec (const mpreal& v, mp_rnd_t r = mpreal::get_default_rnd()) { return acos (1/v, r); } +inline const mpreal acsc (const mpreal& v, mp_rnd_t r = mpreal::get_default_rnd()) { return asin (1/v, r); } +inline const mpreal acoth (const mpreal& v, mp_rnd_t r = mpreal::get_default_rnd()) { return atanh(1/v, r); } +inline const mpreal asech (const mpreal& v, mp_rnd_t r = mpreal::get_default_rnd()) { return acosh(1/v, r); } +inline const mpreal acsch (const mpreal& v, mp_rnd_t r = mpreal::get_default_rnd()) { return asinh(1/v, r); } -inline const mpreal cosh (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(cosh ); } -inline const mpreal sinh (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(sinh ); } -inline const mpreal tanh (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(tanh ); } -inline const mpreal sech (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(sech ); } -inline const mpreal csch (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(csch ); } -inline const mpreal coth (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(coth ); } -inline const mpreal acosh (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(acosh); } -inline const mpreal asinh (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(asinh); } -inline const mpreal atanh (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(atanh); } +inline const mpreal cosh (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(cosh ); } +inline const mpreal sinh (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(sinh ); } +inline const mpreal tanh (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(tanh ); } +inline const mpreal sech (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(sech ); } +inline const mpreal csch (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(csch ); } +inline const mpreal coth (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(coth ); } +inline const mpreal acosh (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(acosh); } +inline const mpreal asinh (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(asinh); } +inline const mpreal atanh (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(atanh); } -inline const mpreal log1p (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(log1p ); } -inline const mpreal expm1 (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(expm1 ); } -inline const mpreal eint (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(eint ); } -inline const mpreal gamma (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(gamma ); } -inline const mpreal lngamma (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(lngamma); } -inline const mpreal zeta (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(zeta ); } -inline const mpreal erf (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(erf ); } -inline const mpreal erfc (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(erfc ); } -inline const mpreal besselj0(const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(j0 ); } -inline const mpreal besselj1(const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(j1 ); } -inline const mpreal bessely0(const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(y0 ); } -inline const mpreal bessely1(const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(y1 ); } +inline const mpreal log1p (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(log1p ); } +inline const mpreal expm1 (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(expm1 ); } +inline const mpreal eint (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(eint ); } +inline const mpreal gamma (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(gamma ); } +inline const mpreal lngamma (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(lngamma); } +inline const mpreal zeta (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(zeta ); } +inline const mpreal erf (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(erf ); } +inline const mpreal erfc (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(erfc ); } +inline const mpreal besselj0(const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(j0 ); } +inline const mpreal besselj1(const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(j1 ); } +inline const mpreal bessely0(const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(y0 ); } +inline const mpreal bessely1(const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(y1 ); } -inline const mpreal atan2 (const mpreal& y, const mpreal& x, mp_rnd_t rnd_mode) +inline const mpreal atan2 (const mpreal& y, const mpreal& x, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); mpfr_atan2(a.mpfr_ptr(), y.mpfr_srcptr(), x.mpfr_srcptr(), rnd_mode); return a; } -inline const mpreal hypot (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) +inline const mpreal hypot (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); mpfr_hypot(a.mpfr_ptr(), x.mpfr_srcptr(), y.mpfr_srcptr(), rnd_mode); return a; } -inline const mpreal remainder (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) +inline const mpreal remainder (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); mpfr_remainder(a.mpfr_ptr(), x.mpfr_srcptr(), y.mpfr_srcptr(), rnd_mode); return a; } -inline const mpreal remquo (long* q, const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) +inline const mpreal remquo (long* q, const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a(0,(std::max)(y.getPrecision(), x.getPrecision())); mpfr_remquo(a.mpfr_ptr(),q, x.mpfr_srcptr(), y.mpfr_srcptr(), rnd_mode); return a; } -inline const mpreal fac_ui (unsigned long int v, mp_prec_t prec, mp_rnd_t rnd_mode) +inline const mpreal fac_ui (unsigned long int v, mp_prec_t prec = mpreal::get_default_prec(), + mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(0, prec); mpfr_fac_ui(x.mpfr_ptr(),v,rnd_mode); @@ -2204,33 +2275,33 @@ inline const mpreal fac_ui (unsigned long int v, mp_prec_t prec, mp_rnd_t rnd_mo } -inline const mpreal lgamma (const mpreal& v, int *signp, mp_rnd_t rnd_mode) +inline const mpreal lgamma (const mpreal& v, int *signp = 0, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(v); int tsignp; - if(signp) mpfr_lgamma(x.mp, signp,v.mp,rnd_mode); - else mpfr_lgamma(x.mp,&tsignp,v.mp,rnd_mode); + if(signp) mpfr_lgamma(x.mpfr_ptr(), signp,v.mpfr_srcptr(),rnd_mode); + else mpfr_lgamma(x.mpfr_ptr(),&tsignp,v.mpfr_srcptr(),rnd_mode); return x; } -inline const mpreal besseljn (long n, const mpreal& x, mp_rnd_t r) +inline const mpreal besseljn (long n, const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { - mpreal y(0, mpfr_get_prec(x.mpfr_srcptr())); + mpreal y(0, x.getPrecision()); mpfr_jn(y.mpfr_ptr(), n, x.mpfr_srcptr(), r); return y; } -inline const mpreal besselyn (long n, const mpreal& x, mp_rnd_t r) +inline const mpreal besselyn (long n, const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { - mpreal y(0, mpfr_get_prec(x.mpfr_srcptr())); + mpreal y(0, x.getPrecision()); mpfr_yn(y.mpfr_ptr(), n, x.mpfr_srcptr(), r); return y; } -inline const mpreal fma (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode) +inline const mpreal fma (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a; mp_prec_t p1, p2, p3; @@ -2245,7 +2316,7 @@ inline const mpreal fma (const mpreal& v1, const mpreal& v2, const mpreal& v3, m return a; } -inline const mpreal fms (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode) +inline const mpreal fms (const mpreal& v1, const mpreal& v2, const mpreal& v3, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a; mp_prec_t p1, p2, p3; @@ -2260,7 +2331,7 @@ inline const mpreal fms (const mpreal& v1, const mpreal& v2, const mpreal& v3, m return a; } -inline const mpreal agm (const mpreal& v1, const mpreal& v2, mp_rnd_t rnd_mode) +inline const mpreal agm (const mpreal& v1, const mpreal& v2, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a; mp_prec_t p1, p2; @@ -2275,7 +2346,7 @@ inline const mpreal agm (const mpreal& v1, const mpreal& v2, mp_rnd_t rnd_mode) return a; } -inline const mpreal sum (const mpreal tab[], unsigned long int n, mp_rnd_t rnd_mode) +inline const mpreal sum (const mpreal tab[], unsigned long int n, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x; mpfr_ptr* t; @@ -2292,23 +2363,23 @@ inline const mpreal sum (const mpreal tab[], unsigned long int n, mp_rnd_t rnd_m // MPFR 2.4.0 Specifics #if (MPFR_VERSION >= MPFR_VERSION_NUM(2,4,0)) -inline int sinh_cosh(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode) +inline int sinh_cosh(mpreal& s, mpreal& c, const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { return mpfr_sinh_cosh(s.mp,c.mp,v.mp,rnd_mode); } -inline const mpreal li2 (const mpreal& x, mp_rnd_t r) +inline const mpreal li2 (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(li2); } -inline const mpreal rem (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) +inline const mpreal rem (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { /* R = rem(X,Y) if Y != 0, returns X - n * Y where n = trunc(X/Y). */ return fmod(x, y, rnd_mode); } -inline const mpreal mod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) +inline const mpreal mod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { (void)rnd_mode; @@ -2333,7 +2404,7 @@ inline const mpreal mod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) return m; } -inline const mpreal fmod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) +inline const mpreal fmod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal a; mp_prec_t yp, xp; @@ -2348,7 +2419,7 @@ inline const mpreal fmod (const mpreal& x, const mpreal& y, mp_rnd_t rnd_mode) return a; } -inline const mpreal rec_sqrt(const mpreal& v, mp_rnd_t rnd_mode) +inline const mpreal rec_sqrt(const mpreal& v, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(v); mpfr_rec_sqrt(x.mp,v.mp,rnd_mode); @@ -2359,41 +2430,41 @@ inline const mpreal rec_sqrt(const mpreal& v, mp_rnd_t rnd_mode) ////////////////////////////////////////////////////////////////////////// // MPFR 3.0.0 Specifics #if (MPFR_VERSION >= MPFR_VERSION_NUM(3,0,0)) -inline const mpreal digamma (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(digamma); } -inline const mpreal ai (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(ai); } +inline const mpreal digamma (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(digamma); } +inline const mpreal ai (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(ai); } #endif // MPFR 3.0.0 Specifics ////////////////////////////////////////////////////////////////////////// // Constants -inline const mpreal const_log2 (mp_prec_t p, mp_rnd_t r) +inline const mpreal const_log2 (mp_prec_t p = mpreal::get_default_prec(), mp_rnd_t r = mpreal::get_default_rnd()) { mpreal x(0, p); mpfr_const_log2(x.mpfr_ptr(), r); return x; } -inline const mpreal const_pi (mp_prec_t p, mp_rnd_t r) +inline const mpreal const_pi (mp_prec_t p = mpreal::get_default_prec(), mp_rnd_t r = mpreal::get_default_rnd()) { mpreal x(0, p); mpfr_const_pi(x.mpfr_ptr(), r); return x; } -inline const mpreal const_euler (mp_prec_t p, mp_rnd_t r) +inline const mpreal const_euler (mp_prec_t p = mpreal::get_default_prec(), mp_rnd_t r = mpreal::get_default_rnd()) { mpreal x(0, p); mpfr_const_euler(x.mpfr_ptr(), r); return x; } -inline const mpreal const_catalan (mp_prec_t p, mp_rnd_t r) +inline const mpreal const_catalan (mp_prec_t p = mpreal::get_default_prec(), mp_rnd_t r = mpreal::get_default_rnd()) { mpreal x(0, p); mpfr_const_catalan(x.mpfr_ptr(), r); return x; } -inline const mpreal const_infinity (int sign, mp_prec_t p, mp_rnd_t /*r*/) +inline const mpreal const_infinity (int sign = 1, mp_prec_t p = mpreal::get_default_prec()) { mpreal x(0, p); mpfr_set_inf(x.mpfr_ptr(), sign); @@ -2430,12 +2501,12 @@ inline const mpreal trunc(const mpreal& v) return x; } -inline const mpreal rint (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint ); } -inline const mpreal rint_ceil (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_ceil ); } -inline const mpreal rint_floor (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_floor); } -inline const mpreal rint_round (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_round); } -inline const mpreal rint_trunc (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_trunc); } -inline const mpreal frac (const mpreal& x, mp_rnd_t r) { MPREAL_UNARY_MATH_FUNCTION_BODY(frac ); } +inline const mpreal rint (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint ); } +inline const mpreal rint_ceil (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_ceil ); } +inline const mpreal rint_floor (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_floor); } +inline const mpreal rint_round (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_round); } +inline const mpreal rint_trunc (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(rint_trunc); } +inline const mpreal frac (const mpreal& x, mp_rnd_t r = mpreal::get_default_rnd()) { MPREAL_UNARY_MATH_FUNCTION_BODY(frac ); } ////////////////////////////////////////////////////////////////////////// // Miscellaneous Functions @@ -2443,14 +2514,14 @@ inline void swap (mpreal& a, mpreal& b) { mpfr_swap(a.mp,b inline const mpreal (max)(const mpreal& x, const mpreal& y){ return (x>y?x:y); } inline const mpreal (min)(const mpreal& x, const mpreal& y){ return (x= MPFR_VERSION_NUM(3,0,0)) +#if (MPFR_VERSION >= MPFR_VERSION_NUM(3,1,0)) // use gmp_randinit_default() to init state, gmp_randclear() to clear -inline const mpreal urandom (gmp_randstate_t& state, mp_rnd_t rnd_mode) +inline const mpreal urandom (gmp_randstate_t& state, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x; mpfr_urandom(x.mp,state,rnd_mode); return x; } -inline const mpreal grandom (gmp_randstate_t& state, mp_rnd_t rnd_mode) +inline const mpreal grandom (gmp_randstate_t& state, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x; mpfr_grandom(x.mp, NULL, state, rnd_mode); @@ -2516,7 +2587,7 @@ inline const mpreal random2 (mp_size_t size, mp_exp_t exp) // a = random(seed); <- initialization & first random number generation // a = random(); <- next random numbers generation // seed != 0 -inline const mpreal random(unsigned int seed) +inline const mpreal random(unsigned int seed = 0) { #if (MPFR_VERSION >= MPFR_VERSION_NUM(3,0,0)) @@ -2541,7 +2612,7 @@ inline const mpreal random(unsigned int seed) } #if (MPFR_VERSION >= MPFR_VERSION_NUM(3,0,0)) -inline const mpreal grandom(unsigned int seed) +inline const mpreal grandom(unsigned int seed = 0) { static gmp_randstate_t state; static bool isFirstTime = true; @@ -2578,21 +2649,21 @@ inline bool mpreal::fits_in_bits(double x, int n) return IsInf(x) || (std::modf ( std::ldexp ( std::frexp ( x, &i ), n ), &t ) == 0.0); } -inline const mpreal pow(const mpreal& a, const mpreal& b, mp_rnd_t rnd_mode) +inline const mpreal pow(const mpreal& a, const mpreal& b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(a); mpfr_pow(x.mp,x.mp,b.mp,rnd_mode); return x; } -inline const mpreal pow(const mpreal& a, const mpz_t b, mp_rnd_t rnd_mode) +inline const mpreal pow(const mpreal& a, const mpz_t b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(a); mpfr_pow_z(x.mp,x.mp,b,rnd_mode); return x; } -inline const mpreal pow(const mpreal& a, const unsigned long int b, mp_rnd_t rnd_mode) +inline const mpreal pow(const mpreal& a, const unsigned long int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(a); mpfr_pow_ui(x.mp,x.mp,b,rnd_mode); @@ -2604,7 +2675,7 @@ inline const mpreal pow(const mpreal& a, const unsigned int b, mp_rnd_t rnd_mode return pow(a,static_cast(b),rnd_mode); } -inline const mpreal pow(const mpreal& a, const long int b, mp_rnd_t rnd_mode) +inline const mpreal pow(const mpreal& a, const long int b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(a); mpfr_pow_si(x.mp,x.mp,b,rnd_mode); @@ -2626,7 +2697,7 @@ inline const mpreal pow(const mpreal& a, const double b, mp_rnd_t rnd_mode) return pow(a,mpreal(b),rnd_mode); } -inline const mpreal pow(const unsigned long int a, const mpreal& b, mp_rnd_t rnd_mode) +inline const mpreal pow(const unsigned long int a, const mpreal& b, mp_rnd_t rnd_mode = mpreal::get_default_rnd()) { mpreal x(a); mpfr_ui_pow(x.mp,a,b.mp,rnd_mode); @@ -2641,13 +2712,13 @@ inline const mpreal pow(const unsigned int a, const mpreal& b, mp_rnd_t rnd_mode inline const mpreal pow(const long int a, const mpreal& b, mp_rnd_t rnd_mode) { if (a>=0) return pow(static_cast(a),b,rnd_mode); - else return pow(mpreal(a),b,rnd_mode); + else return pow(mpreal(a),b,rnd_mode); } inline const mpreal pow(const int a, const mpreal& b, mp_rnd_t rnd_mode) { if (a>=0) return pow(static_cast(a),b,rnd_mode); - else return pow(mpreal(a),b,rnd_mode); + else return pow(mpreal(a),b,rnd_mode); } inline const mpreal pow(const long double a, const mpreal& b, mp_rnd_t rnd_mode) @@ -2676,13 +2747,13 @@ inline const mpreal pow(const unsigned long int a, const unsigned int b, mp_rnd_ inline const mpreal pow(const unsigned long int a, const long int b, mp_rnd_t rnd_mode) { if(b>0) return pow(a,static_cast(b),rnd_mode); //mpfr_ui_pow_ui - else return pow(a,mpreal(b),rnd_mode); //mpfr_ui_pow + else return pow(a,mpreal(b),rnd_mode); //mpfr_ui_pow } inline const mpreal pow(const unsigned long int a, const int b, mp_rnd_t rnd_mode) { if(b>0) return pow(a,static_cast(b),rnd_mode); //mpfr_ui_pow_ui - else return pow(a,mpreal(b),rnd_mode); //mpfr_ui_pow + else return pow(a,mpreal(b),rnd_mode); //mpfr_ui_pow } inline const mpreal pow(const unsigned long int a, const long double b, mp_rnd_t rnd_mode) @@ -2879,7 +2950,7 @@ inline const mpreal pow(const double a, const int b, mp_rnd_t rnd_mode) // Non-throwing swap C++ idiom: http://en.wikibooks.org/wiki/More_C%2B%2B_Idioms/Non-throwing_swap namespace std { - // only allowed to extend namespace std with specializations + // we are allowed to extend namespace std with specializations only template <> inline void swap(mpfr::mpreal& x, mpfr::mpreal& y) { @@ -2907,20 +2978,6 @@ namespace std static const bool tinyness_before = true; static const float_denorm_style has_denorm = denorm_absent; - - inline static float_round_style round_style() - { - mp_rnd_t r = mpfr::mpreal::get_default_rnd(); - - switch (r) - { - case GMP_RNDN: return round_to_nearest; - case GMP_RNDZ: return round_toward_zero; - case GMP_RNDU: return round_toward_infinity; - case GMP_RNDD: return round_toward_neg_infinity; - default: return round_indeterminate; - } - } inline static mpfr::mpreal (min) (mp_prec_t precision = mpfr::mpreal::get_default_prec()) { return mpfr::minval(precision); } inline static mpfr::mpreal (max) (mp_prec_t precision = mpfr::mpreal::get_default_prec()) { return mpfr::maxval(precision); } @@ -2928,7 +2985,7 @@ namespace std // Returns smallest eps such that 1 + eps != 1 (classic machine epsilon) inline static mpfr::mpreal epsilon(mp_prec_t precision = mpfr::mpreal::get_default_prec()) { return mpfr::machine_epsilon(precision); } - + // Returns smallest eps such that x + eps != x (relative machine epsilon) inline static mpfr::mpreal epsilon(const mpfr::mpreal& x) { return mpfr::machine_epsilon(x); } @@ -2951,10 +3008,30 @@ namespace std MPREAL_PERMISSIVE_EXPR static const int min_exponent10 = (int) (MPFR_EMIN_DEFAULT * 0.3010299956639811); MPREAL_PERMISSIVE_EXPR static const int max_exponent10 = (int) (MPFR_EMAX_DEFAULT * 0.3010299956639811); - // Should be constant according to standard, but 'digits' depends on precision in MPFR +#ifdef MPREAL_HAVE_DYNAMIC_STD_NUMERIC_LIMITS - inline static int digits() { return mpfr::mpreal::get_default_prec(); } - inline static int digits(const mpfr::mpreal& x) { return x.getPrecision(); } + // Following members should be constant according to standard, but they can be variable in MPFR + // So we define them as functions here. + // + // This is preferable way for std::numeric_limits specialization. + // But it is incompatible with standard std::numeric_limits and might not work with other libraries, e.g. boost. + // See below for compatible implementation. + inline static float_round_style round_style() + { + mp_rnd_t r = mpfr::mpreal::get_default_rnd(); + + switch (r) + { + case GMP_RNDN: return round_to_nearest; + case GMP_RNDZ: return round_toward_zero; + case GMP_RNDU: return round_toward_infinity; + case GMP_RNDD: return round_toward_neg_infinity; + default: return round_indeterminate; + } + } + + inline static int digits() { return int(mpfr::mpreal::get_default_prec()); } + inline static int digits(const mpfr::mpreal& x) { return x.getPrecision(); } inline static int digits10(mp_prec_t precision = mpfr::mpreal::get_default_prec()) { @@ -2970,8 +3047,27 @@ namespace std { return digits10(precision); } +#else + // Digits and round_style are NOT constants when it comes to mpreal. + // If possible, please use functions digits() and round_style() defined above. + // + // These (default) values are preserved for compatibility with existing libraries, e.g. boost. + // Change them accordingly to your application. + // + // For example, if you use 256 bits of precision uniformly in your program, then: + // digits = 256 + // digits10 = 77 + // max_digits10 = 78 + // + // Approximate formula for decimal digits is: digits10 = floor(log10(2) * digits). See bits2digits() for more details. + + static const std::float_round_style round_style = round_to_nearest; + static const int digits = 53; + static const int digits10 = 15; + static const int max_digits10 = 16; +#endif }; } -#endif /* __MPREAL_H__ */ +#endif /* __MPREAL_H__ */ \ No newline at end of file From 84ad8ce7e30e26eb235906cd569d1f2b7f3798c5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Jul 2014 12:00:56 +0200 Subject: [PATCH 0641/1103] Fix bug #770: workaround thread safety in mpreal --- unsupported/Eigen/MPRealSupport | 28 ++++++++++++++-------------- 1 file changed, 14 insertions(+), 14 deletions(-) diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport index 6a65a601b..de4f11934 100644 --- a/unsupported/Eigen/MPRealSupport +++ b/unsupported/Eigen/MPRealSupport @@ -146,8 +146,8 @@ int main() }; }; - template - struct gebp_kernel + template + struct gebp_kernel { typedef mpfr::mpreal mpreal; @@ -155,34 +155,34 @@ int main() void operator()(mpreal* res, Index resStride, const mpreal* blockA, const mpreal* blockB, Index rows, Index depth, Index cols, mpreal alpha, Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0) { - mpreal acc1, tmp; + mpreal acc1(0,mpfr_get_prec(blockA[0].mpfr_srcptr())), + tmp (0,mpfr_get_prec(blockA[0].mpfr_srcptr())); if(strideA==-1) strideA = depth; if(strideB==-1) strideB = depth; - for(Index j=0; j)(nr,cols-j); - mpreal *C1 = res + j*resStride; - for(Index i=0; i(blockB) + j*strideB + offsetB*actual_nr; - mpreal *A = const_cast(blockA) + i*strideA + offsetA; + mpreal *C1 = res + j*resStride; + + const mpreal *A = blockA + i*strideA + offsetA; + const mpreal *B = blockB + j*strideB + offsetB; + acc1 = 0; for(Index k=0; k Date: Thu, 17 Jul 2014 12:06:20 +0200 Subject: [PATCH 0642/1103] bug #842: warn user about MPFR++ being under the GPL --- unsupported/Eigen/MPRealSupport | 2 ++ 1 file changed, 2 insertions(+) diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport index de4f11934..35d77e5bd 100644 --- a/unsupported/Eigen/MPRealSupport +++ b/unsupported/Eigen/MPRealSupport @@ -27,6 +27,8 @@ namespace Eigen { * via the MPFR C++ * library which itself is built upon MPFR/GMP. * + * \warning MPFR C++ is licensed under the GPL. + * * You can find a copy of MPFR C++ that is known to be compatible in the unsupported/test/mpreal folder. * * Here is an example: From 77af4cc3c9fac237d8fcf32379137b14c203033f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 17 Jul 2014 13:34:26 +0200 Subject: [PATCH 0643/1103] bug #397: add a warning for 64 to 32 bit integer conversion and fix many of these warning by splitting the index type used for storage and as size/coefficient indexes in PermutationMatrix and Transpositions. --- CMakeLists.txt | 7 +- Eigen/src/Cholesky/LDLT.h | 8 +- Eigen/src/Core/PermutationMatrix.h | 60 ++++++++------- Eigen/src/Core/Transpositions.h | 77 +++++++++++--------- Eigen/src/LU/PartialPivLU.h | 6 +- Eigen/src/SparseCore/SparseCwiseBinaryOp.h | 11 ++- Eigen/src/SparseCore/SparseDenseProduct.h | 6 +- Eigen/src/SparseCore/SparseDiagonalProduct.h | 14 ++-- Eigen/src/SparseCore/SparseMatrix.h | 6 +- Eigen/src/SparseCore/SparseMatrixBase.h | 18 ++--- Eigen/src/SparseCore/SparseSelfAdjointView.h | 2 +- 11 files changed, 120 insertions(+), 95 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index a719e47fd..470095680 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -108,7 +108,8 @@ endif() set(EIGEN_TEST_MAX_SIZE "320" CACHE STRING "Maximal matrix/vector size, default is 320") macro(ei_add_cxx_compiler_flag FLAG) - string(REGEX REPLACE "-" "" SFLAG ${FLAG}) + string(REGEX REPLACE "-" "" SFLAG1 ${FLAG}) + string(REGEX REPLACE "\\+" "p" SFLAG ${SFLAG1}) check_cxx_compiler_flag(${FLAG} COMPILER_SUPPORT_${SFLAG}) if(COMPILER_SUPPORT_${SFLAG}) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${FLAG}") @@ -142,6 +143,9 @@ if(NOT MSVC) ei_add_cxx_compiler_flag("-Wpointer-arith") ei_add_cxx_compiler_flag("-Wwrite-strings") ei_add_cxx_compiler_flag("-Wformat-security") + ei_add_cxx_compiler_flag("-Wshorten-64-to-32") + ei_add_cxx_compiler_flag("-Wenum-conversion") + ei_add_cxx_compiler_flag("-Wc++11-extensions") ei_add_cxx_compiler_flag("-Wno-psabi") ei_add_cxx_compiler_flag("-Wno-variadic-macros") @@ -153,6 +157,7 @@ if(NOT MSVC) ei_add_cxx_compiler_flag("-wd981") # disable ICC's "operands are evaluated in unspecified order" remark ei_add_cxx_compiler_flag("-wd2304") # disbale ICC's "warning #2304: non-explicit constructor with single argument may cause implicit type conversion" produced by -Wnon-virtual-dtor + # The -ansi flag must be added last, otherwise it is also used as a linker flag by check_cxx_compiler_flag making it fails # Moreover we should not set both -strict-ansi and -ansi check_cxx_compiler_flag("-strict-ansi" COMPILER_SUPPORT_STRICTANSI) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 6881e1ca8..aa9784e54 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -246,6 +246,7 @@ template<> struct ldlt_inplace typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; typedef typename MatrixType::Index Index; + typedef typename TranspositionType::StorageIndexType IndexType; eigen_assert(mat.rows()==mat.cols()); const Index size = mat.rows(); @@ -265,7 +266,7 @@ template<> struct ldlt_inplace mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner); index_of_biggest_in_corner += k; - transpositions.coeffRef(k) = index_of_biggest_in_corner; + transpositions.coeffRef(k) = IndexType(index_of_biggest_in_corner); if(k != index_of_biggest_in_corner) { // apply the transposition while taking care to consider only @@ -274,7 +275,7 @@ template<> struct ldlt_inplace mat.row(k).head(k).swap(mat.row(index_of_biggest_in_corner).head(k)); mat.col(k).tail(s).swap(mat.col(index_of_biggest_in_corner).tail(s)); std::swap(mat.coeffRef(k,k),mat.coeffRef(index_of_biggest_in_corner,index_of_biggest_in_corner)); - for(int i=k+1;i template LDLT& LDLT::rankUpdate(const MatrixBase& w, const typename NumTraits::Real& sigma) { + typedef typename TranspositionType::StorageIndexType IndexType; const Index size = w.rows(); if (m_isInitialized) { @@ -453,7 +455,7 @@ LDLT& LDLT::rankUpdate(const MatrixBase=0 ? internal::PositiveSemiDef : internal::NegativeSemiDef; m_isInitialized = true; diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 1297b8413..009873c78 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -66,11 +66,11 @@ class PermutationBase : public EigenBase MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime, MaxColsAtCompileTime = Traits::MaxColsAtCompileTime }; - typedef typename Traits::Scalar Scalar; + typedef typename Traits::StorageIndexType StorageIndexType; typedef typename Traits::Index Index; - typedef Matrix + typedef Matrix DenseMatrixType; - typedef PermutationMatrix + typedef PermutationMatrix PlainPermutationType; using Base::derived; #endif @@ -147,7 +147,7 @@ class PermutationBase : public EigenBase /** Sets *this to be the identity permutation matrix */ void setIdentity() { - for(Index i = 0; i < size(); ++i) + for(StorageIndexType i = 0; i < size(); ++i) indices().coeffRef(i) = i; } @@ -173,8 +173,8 @@ class PermutationBase : public EigenBase eigen_assert(i>=0 && j>=0 && i * * \param SizeAtCompileTime the number of rows/cols, or Dynamic * \param MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it. - * \param IndexType the interger type of the indices + * \param StorageIndexType the interger type of the indices * * This class represents a permutation matrix, internally stored as a vector of integers. * @@ -270,17 +270,18 @@ class PermutationBase : public EigenBase */ namespace internal { -template -struct traits > - : traits > +template +struct traits > + : traits > { - typedef IndexType Index; - typedef Matrix IndicesType; + typedef Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; + typedef typename IndicesType::Index Index; + typedef _StorageIndexType StorageIndexType; }; } -template -class PermutationMatrix : public PermutationBase > +template +class PermutationMatrix : public PermutationBase > { typedef PermutationBase Base; typedef internal::traits Traits; @@ -288,6 +289,8 @@ class PermutationMatrix : public PermutationBase -struct traits,_PacketAccess> > - : traits > +template +struct traits,_PacketAccess> > + : traits > { - typedef IndexType Index; - typedef Map, _PacketAccess> IndicesType; + typedef Map, _PacketAccess> IndicesType; + typedef typename IndicesType::Index Index; + typedef _StorageIndexType StorageIndexType; }; } -template -class Map,_PacketAccess> - : public PermutationBase,_PacketAccess> > +template +class Map,_PacketAccess> + : public PermutationBase,_PacketAccess> > { typedef PermutationBase Base; typedef internal::traits Traits; @@ -403,14 +407,15 @@ class Map, #ifndef EIGEN_PARSED_BY_DOXYGEN typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar Index; + typedef typename IndicesType::Scalar StorageIndexType; + typedef typename IndicesType::Index Index; #endif - inline Map(const Index* indicesPtr) + inline Map(const StorageIndexType* indicesPtr) : m_indices(indicesPtr) {} - inline Map(const Index* indicesPtr, Index size) + inline Map(const StorageIndexType* indicesPtr, Index size) : m_indices(indicesPtr,size) {} @@ -466,7 +471,8 @@ struct traits > { typedef PermutationStorage StorageKind; typedef typename _IndicesType::Scalar Scalar; - typedef typename _IndicesType::Scalar Index; + typedef typename _IndicesType::Scalar StorageIndexType; + typedef typename _IndicesType::Index Index; typedef _IndicesType IndicesType; enum { RowsAtCompileTime = _IndicesType::SizeAtCompileTime, diff --git a/Eigen/src/Core/Transpositions.h b/Eigen/src/Core/Transpositions.h index e4ba0756f..92261118f 100644 --- a/Eigen/src/Core/Transpositions.h +++ b/Eigen/src/Core/Transpositions.h @@ -53,7 +53,8 @@ class TranspositionsBase public: typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar Index; + typedef typename IndicesType::Scalar StorageIndexType; + typedef typename IndicesType::Index Index; Derived& derived() { return *static_cast(this); } const Derived& derived() const { return *static_cast(this); } @@ -81,17 +82,17 @@ class TranspositionsBase inline Index size() const { return indices().size(); } /** Direct access to the underlying index vector */ - inline const Index& coeff(Index i) const { return indices().coeff(i); } + inline const StorageIndexType& coeff(Index i) const { return indices().coeff(i); } /** Direct access to the underlying index vector */ - inline Index& coeffRef(Index i) { return indices().coeffRef(i); } + inline StorageIndexType& coeffRef(Index i) { return indices().coeffRef(i); } /** Direct access to the underlying index vector */ - inline const Index& operator()(Index i) const { return indices()(i); } + inline const StorageIndexType& operator()(Index i) const { return indices()(i); } /** Direct access to the underlying index vector */ - inline Index& operator()(Index i) { return indices()(i); } + inline StorageIndexType& operator()(Index i) { return indices()(i); } /** Direct access to the underlying index vector */ - inline const Index& operator[](Index i) const { return indices()(i); } + inline const StorageIndexType& operator[](Index i) const { return indices()(i); } /** Direct access to the underlying index vector */ - inline Index& operator[](Index i) { return indices()(i); } + inline StorageIndexType& operator[](Index i) { return indices()(i); } /** const version of indices(). */ const IndicesType& indices() const { return derived().indices(); } @@ -99,7 +100,7 @@ class TranspositionsBase IndicesType& indices() { return derived().indices(); } /** Resizes to given size. */ - inline void resize(int newSize) + inline void resize(Index newSize) { indices().resize(newSize); } @@ -107,7 +108,7 @@ class TranspositionsBase /** Sets \c *this to represents an identity transformation */ void setIdentity() { - for(int i = 0; i < indices().size(); ++i) + for(StorageIndexType i = 0; i < indices().size(); ++i) coeffRef(i) = i; } @@ -144,23 +145,26 @@ class TranspositionsBase }; namespace internal { -template -struct traits > +template +struct traits > { - typedef IndexType Index; - typedef Matrix IndicesType; + typedef Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; + typedef typename IndicesType::Index Index; + typedef _StorageIndexType StorageIndexType; }; } -template -class Transpositions : public TranspositionsBase > +template +class Transpositions : public TranspositionsBase > { typedef internal::traits Traits; public: typedef TranspositionsBase Base; typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar Index; + typedef typename IndicesType::Scalar StorageIndexType; + typedef typename IndicesType::Index Index; + inline Transpositions() {} @@ -215,30 +219,32 @@ class Transpositions : public TranspositionsBase -struct traits,_PacketAccess> > +template +struct traits,_PacketAccess> > { - typedef IndexType Index; - typedef Map, _PacketAccess> IndicesType; + typedef Map, _PacketAccess> IndicesType; + typedef typename IndicesType::Index Index; + typedef _StorageIndexType StorageIndexType; }; } -template -class Map,PacketAccess> - : public TranspositionsBase,PacketAccess> > +template +class Map,PacketAccess> + : public TranspositionsBase,PacketAccess> > { typedef internal::traits Traits; public: typedef TranspositionsBase Base; typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar Index; + typedef typename IndicesType::Scalar StorageIndexType; + typedef typename IndicesType::Index Index; - inline Map(const Index* indicesPtr) + inline Map(const StorageIndexType* indicesPtr) : m_indices(indicesPtr) {} - inline Map(const Index* indicesPtr, Index size) + inline Map(const StorageIndexType* indicesPtr, Index size) : m_indices(indicesPtr,size) {} @@ -275,7 +281,8 @@ namespace internal { template struct traits > { - typedef typename _IndicesType::Scalar Index; + typedef typename _IndicesType::Scalar StorageIndexType; + typedef typename _IndicesType::Index Index; typedef _IndicesType IndicesType; }; } @@ -289,7 +296,8 @@ class TranspositionsWrapper typedef TranspositionsBase Base; typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar Index; + typedef typename IndicesType::Scalar StorageIndexType; + typedef typename IndicesType::Index Index; inline TranspositionsWrapper(IndicesType& a_indices) : m_indices(a_indices) @@ -363,24 +371,25 @@ struct transposition_matrix_product_retval { typedef typename remove_all::type MatrixTypeNestedCleaned; typedef typename TranspositionType::Index Index; + typedef typename TranspositionType::StorageIndexType StorageIndexType; transposition_matrix_product_retval(const TranspositionType& tr, const MatrixType& matrix) : m_transpositions(tr), m_matrix(matrix) {} - inline int rows() const { return m_matrix.rows(); } - inline int cols() const { return m_matrix.cols(); } + inline Index rows() const { return m_matrix.rows(); } + inline Index cols() const { return m_matrix.cols(); } template inline void evalTo(Dest& dst) const { - const int size = m_transpositions.size(); - Index j = 0; + const Index size = m_transpositions.size(); + StorageIndexType j = 0; if(!(is_same::value && extract_data(dst) == extract_data(m_matrix))) dst = m_matrix; - for(int k=(Transposed?size-1:0) ; Transposed?k>=0:k=0:k -void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, typename TranspositionType::Index& nb_transpositions) +void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, typename TranspositionType::StorageIndexType& nb_transpositions) { eigen_assert(lu.cols() == row_transpositions.size()); eigen_assert((&row_transpositions.coeffRef(1)-&row_transpositions.coeffRef(0)) == 1); partial_lu_impl - + ::blocked_lu(lu.rows(), lu.cols(), &lu.coeffRef(0,0), lu.outerStride(), &row_transpositions.coeffRef(0), nb_transpositions); } @@ -396,7 +396,7 @@ PartialPivLU& PartialPivLU::compute(const MatrixType& ma m_rowsTranspositions.resize(size); - typename TranspositionType::Index nb_transpositions; + typename TranspositionType::StorageIndexType nb_transpositions; internal::partial_lu_inplace(m_lu, m_rowsTranspositions, nb_transpositions); m_det_p = (nb_transpositions%2) ? -1 : 1; diff --git a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h index ec86ca933..60fdd214a 100644 --- a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h @@ -69,7 +69,6 @@ class CwiseBinaryOpImpl::InnerIterator : public internal::sparse_cwise_binary_op_inner_iterator_selector::InnerIterator> { public: - typedef typename Lhs::Index Index; typedef internal::sparse_cwise_binary_op_inner_iterator_selector< BinaryOp,Lhs,Rhs, InnerIterator> Base; @@ -95,11 +94,11 @@ class sparse_cwise_binary_op_inner_iterator_selector CwiseBinaryXpr; typedef typename traits::Scalar Scalar; + typedef typename traits::Index Index; typedef typename traits::_LhsNested _LhsNested; typedef typename traits::_RhsNested _RhsNested; typedef typename _LhsNested::InnerIterator LhsIterator; typedef typename _RhsNested::InnerIterator RhsIterator; - typedef typename Lhs::Index Index; public: @@ -161,11 +160,11 @@ class sparse_cwise_binary_op_inner_iterator_selector, Lhs, typedef scalar_product_op BinaryFunc; typedef CwiseBinaryOp CwiseBinaryXpr; typedef typename CwiseBinaryXpr::Scalar Scalar; + typedef typename CwiseBinaryXpr::Index Index; typedef typename traits::_LhsNested _LhsNested; typedef typename _LhsNested::InnerIterator LhsIterator; typedef typename traits::_RhsNested _RhsNested; typedef typename _RhsNested::InnerIterator RhsIterator; - typedef typename Lhs::Index Index; public: EIGEN_STRONG_INLINE sparse_cwise_binary_op_inner_iterator_selector(const CwiseBinaryXpr& xpr, Index outer) @@ -215,15 +214,15 @@ class sparse_cwise_binary_op_inner_iterator_selector, Lhs, typedef scalar_product_op BinaryFunc; typedef CwiseBinaryOp CwiseBinaryXpr; typedef typename CwiseBinaryXpr::Scalar Scalar; + typedef typename CwiseBinaryXpr::Index Index; typedef typename traits::_LhsNested _LhsNested; typedef typename traits::RhsNested RhsNested; typedef typename _LhsNested::InnerIterator LhsIterator; - typedef typename Lhs::Index Index; enum { IsRowMajor = (int(Lhs::Flags)&RowMajorBit)==RowMajorBit }; public: EIGEN_STRONG_INLINE sparse_cwise_binary_op_inner_iterator_selector(const CwiseBinaryXpr& xpr, Index outer) - : m_rhs(xpr.rhs()), m_lhsIter(xpr.lhs(),outer), m_functor(xpr.functor()), m_outer(outer) + : m_rhs(xpr.rhs()), m_lhsIter(xpr.lhs(),typename _LhsNested::Index(outer)), m_functor(xpr.functor()), m_outer(outer) {} EIGEN_STRONG_INLINE Derived& operator++() @@ -256,9 +255,9 @@ class sparse_cwise_binary_op_inner_iterator_selector, Lhs, typedef scalar_product_op BinaryFunc; typedef CwiseBinaryOp CwiseBinaryXpr; typedef typename CwiseBinaryXpr::Scalar Scalar; + typedef typename CwiseBinaryXpr::Index Index; typedef typename traits::_RhsNested _RhsNested; typedef typename _RhsNested::InnerIterator RhsIterator; - typedef typename Lhs::Index Index; enum { IsRowMajor = (int(Rhs::Flags)&RowMajorBit)==RowMajorBit }; public: diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 295b2e81f..4a7813296 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -102,8 +102,8 @@ class SparseDenseOuterProduct EIGEN_STATIC_ASSERT(Tr,YOU_MADE_A_PROGRAMMING_MISTAKE); } - EIGEN_STRONG_INLINE Index rows() const { return Tr ? m_rhs.rows() : m_lhs.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return Tr ? m_lhs.cols() : m_rhs.cols(); } + EIGEN_STRONG_INLINE Index rows() const { return Tr ? Index(m_rhs.rows()) : m_lhs.rows(); } + EIGEN_STRONG_INLINE Index cols() const { return Tr ? m_lhs.cols() : Index(m_rhs.cols()); } EIGEN_STRONG_INLINE const _LhsNested& lhs() const { return m_lhs; } EIGEN_STRONG_INLINE const _RhsNested& rhs() const { return m_rhs; } @@ -139,7 +139,7 @@ class SparseDenseOuterProduct::InnerIterator : public _LhsNes Scalar get(const _RhsNested &rhs, Index outer, Sparse = Sparse()) { typename Traits::_RhsNested::InnerIterator it(rhs, outer); - if (it && it.index()==0 && it.value()!=0) + if (it && it.index()==0 && it.value()!=Scalar(0)) return it.value(); m_empty = true; return Scalar(0); diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index 1bb590e64..c056b4914 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -32,8 +32,10 @@ struct traits > typedef typename remove_all::type _Lhs; typedef typename remove_all::type _Rhs; typedef typename _Lhs::Scalar Scalar; - typedef typename promote_index_type::Index, - typename traits::Index>::type Index; + // propagate the index type of the sparse matrix + typedef typename conditional< is_diagonal<_Lhs>::ret, + typename traits::Index, + typename traits::Index>::type Index; typedef Sparse StorageKind; typedef MatrixXpr XprKind; enum { @@ -90,8 +92,8 @@ class SparseDiagonalProduct eigen_assert(lhs.cols() == rhs.rows() && "invalid sparse matrix * diagonal matrix product"); } - EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); } + EIGEN_STRONG_INLINE Index rows() const { return Index(m_lhs.rows()); } + EIGEN_STRONG_INLINE Index cols() const { return Index(m_rhs.cols()); } EIGEN_STRONG_INLINE const _LhsNested& lhs() const { return m_lhs; } EIGEN_STRONG_INLINE const _RhsNested& rhs() const { return m_rhs; } @@ -109,7 +111,7 @@ class sparse_diagonal_product_inner_iterator_selector : public CwiseUnaryOp,const Rhs>::InnerIterator { typedef typename CwiseUnaryOp,const Rhs>::InnerIterator Base; - typedef typename Lhs::Index Index; + typedef typename Rhs::Index Index; public: inline sparse_diagonal_product_inner_iterator_selector( const SparseDiagonalProductType& expr, Index outer) @@ -129,7 +131,7 @@ class sparse_diagonal_product_inner_iterator_selector scalar_product_op, const typename Rhs::ConstInnerVectorReturnType, const typename Lhs::DiagonalVectorType>::InnerIterator Base; - typedef typename Lhs::Index Index; + typedef typename Rhs::Index Index; Index m_outer; public: inline sparse_diagonal_product_inner_iterator_selector( diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index ed935c40d..2ed2f3ebd 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -800,7 +800,9 @@ protected: template void initAssignment(const Other& other) { - resize(other.rows(), other.cols()); + eigen_assert( other.rows() == typename Other::Index(Index(other.rows())) + && other.cols() == typename Other::Index(Index(other.cols())) ); + resize(Index(other.rows()), Index(other.cols())); if(m_innerNonZeros) { std::free(m_innerNonZeros); @@ -940,7 +942,7 @@ void set_from_triplets(const InputIterator& begin, const InputIterator& end, Spa enum { IsRowMajor = SparseMatrixType::IsRowMajor }; typedef typename SparseMatrixType::Scalar Scalar; typedef typename SparseMatrixType::Index Index; - SparseMatrix trMat(mat.rows(),mat.cols()); + SparseMatrix trMat(mat.rows(),mat.cols()); if(begin!=end) { diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index c64d74da0..1050cf3f1 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -202,20 +202,20 @@ template class SparseMatrixBase : public EigenBase inline Derived& assign(const OtherDerived& other) { const bool transpose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); - const Index outerSize = (int(OtherDerived::Flags) & RowMajorBit) ? other.rows() : other.cols(); + const Index outerSize = (int(OtherDerived::Flags) & RowMajorBit) ? Index(other.rows()) : Index(other.cols()); if ((!transpose) && other.isRValue()) { // eval without temporary - derived().resize(other.rows(), other.cols()); + derived().resize(Index(other.rows()), Index(other.cols())); derived().setZero(); derived().reserve((std::max)(this->rows(),this->cols())*2); for (Index j=0; j class SparseMatrixBase : public EigenBase enum { Flip = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit) }; - const Index outerSize = other.outerSize(); + const Index outerSize = Index(other.outerSize()); //typedef typename internal::conditional, Derived>::type TempType; // thanks to shallow copies, we always eval to a tempary - Derived temp(other.rows(), other.cols()); + Derived temp(Index(other.rows()), Index(other.cols())); temp.reserve((std::max)(this->rows(),this->cols())*2); for (Index j=0; j class SparseMatrixBase : public EigenBase { dst.setZero(); for (Index j=0; j Date: Thu, 17 Jul 2014 17:09:15 +0200 Subject: [PATCH 0644/1103] bug #843: fix jacobisvd for complexes and extend respective unit test to chack with random tricky matrices --- Eigen/src/SVD/JacobiSVD.h | 12 +++--- test/jacobisvd.cpp | 88 +++++++++++++++++++++++++++++---------- 2 files changed, 74 insertions(+), 26 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 439eb5d29..412daa746 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -375,17 +375,19 @@ struct svd_precondition_2x2_block_to_be_real Scalar z; JacobiRotation rot; RealScalar n = sqrt(numext::abs2(work_matrix.coeff(p,p)) + numext::abs2(work_matrix.coeff(q,p))); + if(n==0) { z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q); work_matrix.row(p) *= z; if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z); if(work_matrix.coeff(q,q)!=Scalar(0)) + { z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q); - else - z = Scalar(0); - work_matrix.row(q) *= z; - if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z); + work_matrix.row(q) *= z; + if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z); + } + // otherwise the second row is already zero, so we have nothing to do. } else { @@ -835,7 +837,7 @@ JacobiSVD::compute(const MatrixType& matrix, unsig if(m_computeThinV) m_matrixV.setIdentity(m_cols, m_diagSize); } - // Scaling factor to reducover/under-flows + // Scaling factor to reduce over/under-flows RealScalar scale = m_workMatrix.cwiseAbs().maxCoeff(); if(scale==RealScalar(0)) scale = RealScalar(1); m_workMatrix /= scale; diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index d441a6eca..36721b496 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -67,6 +67,7 @@ template void jacobisvd_solve(const MatrixType& m, unsigned int computationOptions) { typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; typedef typename MatrixType::Index Index; Index rows = m.rows(); Index cols = m.cols(); @@ -81,9 +82,37 @@ void jacobisvd_solve(const MatrixType& m, unsigned int computationOptions) RhsType rhs = RhsType::Random(rows, internal::random(1, cols)); JacobiSVD svd(m, computationOptions); + + if(internal::is_same::value) svd.setThreshold(1e-8); + else if(internal::is_same::value) svd.setThreshold(1e-4); + SolutionType x = svd.solve(rhs); + + RealScalar residual = (m*x-rhs).norm(); + // Check that there is no significantly better solution in the neighborhood of x + if(!test_isMuchSmallerThan(residual,rhs.norm())) + { + // If the residual is very small, then we have an exact solution, so we are already good. + for(int k=0;k::epsilon(); + RealScalar residual_y = (m*y-rhs).norm(); + VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); + + y.row(k) = x.row(k).array() - 2*NumTraits::epsilon(); + residual_y = (m*y-rhs).norm(); + VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); + } + } + // evaluate normal equation which works also for least-squares solutions - VERIFY_IS_APPROX(m.adjoint()*m*x,m.adjoint()*rhs); + if(internal::is_same::value) + { + // This test is not stable with single precision. + // This is probably because squaring m signicantly affects the precision. + VERIFY_IS_APPROX(m.adjoint()*m*x,m.adjoint()*rhs); + } // check minimal norm solutions { @@ -139,10 +168,9 @@ void jacobisvd_test_all_computation_options(const MatrixType& m) if (QRPreconditioner == NoQRPreconditioner && m.rows() != m.cols()) return; JacobiSVD fullSvd(m, ComputeFullU|ComputeFullV); - - jacobisvd_check_full(m, fullSvd); - jacobisvd_solve(m, ComputeFullU | ComputeFullV); - + CALL_SUBTEST(( jacobisvd_check_full(m, fullSvd) )); + CALL_SUBTEST(( jacobisvd_solve(m, ComputeFullU | ComputeFullV) )); + #if defined __INTEL_COMPILER // remark #111: statement is unreachable #pragma warning disable 111 @@ -150,20 +178,20 @@ void jacobisvd_test_all_computation_options(const MatrixType& m) if(QRPreconditioner == FullPivHouseholderQRPreconditioner) return; - jacobisvd_compare_to_full(m, ComputeFullU, fullSvd); - jacobisvd_compare_to_full(m, ComputeFullV, fullSvd); - jacobisvd_compare_to_full(m, 0, fullSvd); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeFullU, fullSvd) )); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeFullV, fullSvd) )); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, 0, fullSvd) )); if (MatrixType::ColsAtCompileTime == Dynamic) { // thin U/V are only available with dynamic number of columns - jacobisvd_compare_to_full(m, ComputeFullU|ComputeThinV, fullSvd); - jacobisvd_compare_to_full(m, ComputeThinV, fullSvd); - jacobisvd_compare_to_full(m, ComputeThinU|ComputeFullV, fullSvd); - jacobisvd_compare_to_full(m, ComputeThinU , fullSvd); - jacobisvd_compare_to_full(m, ComputeThinU|ComputeThinV, fullSvd); - jacobisvd_solve(m, ComputeFullU | ComputeThinV); - jacobisvd_solve(m, ComputeThinU | ComputeFullV); - jacobisvd_solve(m, ComputeThinU | ComputeThinV); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeFullU|ComputeThinV, fullSvd) )); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinV, fullSvd) )); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinU|ComputeFullV, fullSvd) )); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinU , fullSvd) )); + CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinU|ComputeThinV, fullSvd) )); + CALL_SUBTEST(( jacobisvd_solve(m, ComputeFullU | ComputeThinV) )); + CALL_SUBTEST(( jacobisvd_solve(m, ComputeThinU | ComputeFullV) )); + CALL_SUBTEST(( jacobisvd_solve(m, ComputeThinU | ComputeThinV) )); // test reconstruction typedef typename MatrixType::Index Index; @@ -176,12 +204,29 @@ void jacobisvd_test_all_computation_options(const MatrixType& m) template void jacobisvd(const MatrixType& a = MatrixType(), bool pickrandom = true) { - MatrixType m = pickrandom ? MatrixType::Random(a.rows(), a.cols()) : a; + MatrixType m = a; + if(pickrandom) + { + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + typedef typename MatrixType::Index Index; + Index diagSize = (std::min)(a.rows(), a.cols()); + RealScalar s = std::numeric_limits::max_exponent10/4; + s = internal::random(1,s); + Matrix d = Matrix::Random(diagSize); + for(Index k=0; k(-s,s)); + m = Matrix::Random(a.rows(),diagSize) * d.asDiagonal() * Matrix::Random(diagSize,a.cols()); + // cancel some coeffs + Index n = internal::random(0,m.size()-1); + for(Index i=0; i(0,m.rows()-1), internal::random(0,m.cols()-1)) = Scalar(0); + } - jacobisvd_test_all_computation_options(m); - jacobisvd_test_all_computation_options(m); - jacobisvd_test_all_computation_options(m); - jacobisvd_test_all_computation_options(m); + CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); + CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); + CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); + CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); } template void jacobisvd_verify_assert(const MatrixType& m) @@ -384,6 +429,7 @@ void test_jacobisvd() TEST_SET_BUT_UNUSED_VARIABLE(r) TEST_SET_BUT_UNUSED_VARIABLE(c) + CALL_SUBTEST_10(( jacobisvd(MatrixXd(r,c)) )); CALL_SUBTEST_7(( jacobisvd(MatrixXf(r,c)) )); CALL_SUBTEST_8(( jacobisvd(MatrixXcd(r,c)) )); (void) r; From 36e6c9064fc68d5c47473f6d251da10e96ad42b3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 18 Jul 2014 14:19:18 +0200 Subject: [PATCH 0645/1103] bug #770: fix out of bounds access --- unsupported/Eigen/MPRealSupport | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport index 35d77e5bd..0584e470e 100644 --- a/unsupported/Eigen/MPRealSupport +++ b/unsupported/Eigen/MPRealSupport @@ -157,9 +157,12 @@ int main() void operator()(mpreal* res, Index resStride, const mpreal* blockA, const mpreal* blockB, Index rows, Index depth, Index cols, mpreal alpha, Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0) { + if(rows==0 || cols==0 || depth==0) + return; + mpreal acc1(0,mpfr_get_prec(blockA[0].mpfr_srcptr())), tmp (0,mpfr_get_prec(blockA[0].mpfr_srcptr())); - + if(strideA==-1) strideA = depth; if(strideB==-1) strideB = depth; From 529e6cb5529f626b7c5cf781bdcd5a5720f904c1 Mon Sep 17 00:00:00 2001 From: Moritz Klammler Date: Fri, 18 Jul 2014 23:19:56 +0200 Subject: [PATCH 0646/1103] Applied changes suggested by Christoph Hertzberg to c'tor leak fix. - Enclose exception handling in '#ifdef EIGEN_EXCEPTIONS'. - Use an object counter to demonstrate the bug more readily. --- Eigen/src/Core/util/Memory.h | 24 ++++++++++++++++++++++++ test/ctorleak.cpp | 28 +++++++++++++++++++++++++++- 2 files changed, 51 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 4b3dcde3a..9b3e84fb0 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -354,16 +354,20 @@ template inline void destruct_elements_of_array(T *ptr, size_t size) template inline T* construct_elements_of_array(T *ptr, size_t size) { size_t i; +#ifdef EIGEN_EXCEPTIONS try +#endif { for (i = 0; i < size; ++i) ::new (ptr + i) T; return ptr; } +#ifdef EIGEN_EXCEPTIONS catch (...) { destruct_elements_of_array(ptr, i); throw; } +#endif } /***************************************************************************** @@ -385,30 +389,38 @@ template inline T* aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(aligned_malloc(sizeof(T)*size)); +#ifdef EIGEN_EXCEPTIONS try +#endif { return construct_elements_of_array(result, size); } +#ifdef EIGEN_EXCEPTIONS catch (...) { aligned_free(result); throw; } +#endif } template inline T* conditional_aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); +#ifdef EIGEN_EXCEPTIONS try +#endif { return construct_elements_of_array(result, size); } +#ifdef EIGEN_EXCEPTIONS catch (...) { conditional_aligned_free(result); throw; } +#endif } /** \internal Deletes objects constructed with aligned_new @@ -438,15 +450,19 @@ template inline T* conditional_aligned_realloc_new(T* pt T *result = reinterpret_cast(conditional_aligned_realloc(reinterpret_cast(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(new_size > old_size) { +#ifdef EIGEN_EXCEPTIONS try +#endif { construct_elements_of_array(result+old_size, new_size-old_size); } +#ifdef EIGEN_EXCEPTIONS catch (...) { conditional_aligned_free(result); throw; } +#endif } return result; } @@ -458,15 +474,19 @@ template inline T* conditional_aligned_new_auto(size_t s T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); if(NumTraits::RequireInitialization) { +#ifdef EIGEN_EXCEPTIONS try +#endif { construct_elements_of_array(result, size); } +#ifdef EIGEN_EXCEPTIONS catch (...) { conditional_aligned_free(result); throw; } +#endif } return result; } @@ -480,15 +500,19 @@ template inline T* conditional_aligned_realloc_new_auto( T *result = reinterpret_cast(conditional_aligned_realloc(reinterpret_cast(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(NumTraits::RequireInitialization && (new_size > old_size)) { +#ifdef EIGEN_EXCEPTIONS try +#endif { construct_elements_of_array(result+old_size, new_size-old_size); } +#ifdef EIGEN_EXCEPTIONS catch (...) { conditional_aligned_free(result); throw; } +#endif } return result; } diff --git a/test/ctorleak.cpp b/test/ctorleak.cpp index 72ab94b66..f3f4411c8 100644 --- a/test/ctorleak.cpp +++ b/test/ctorleak.cpp @@ -4,11 +4,30 @@ struct Foo { + static unsigned object_count; + static unsigned object_limit; int dummy; - Foo() { if (!internal::random(0, 10)) throw Foo::Fail(); } + + Foo() + { +#ifdef EIGEN_EXCEPTIONS + // TODO: Is this the correct way to handle this? + if (Foo::object_count > Foo::object_limit) { throw Foo::Fail(); } +#endif + ++Foo::object_count; + } + + ~Foo() + { + --Foo::object_count; + } + class Fail : public std::exception {}; }; +unsigned Foo::object_count = 0; +unsigned Foo::object_limit = 0; + namespace Eigen { template<> @@ -34,10 +53,17 @@ namespace Eigen void test_ctorleak() { + Foo::object_count = 0; + Foo::object_limit = internal::random(0, 14 * 92 - 2); +#ifdef EIGEN_EXCEPTIONS try +#endif { Matrix m(14, 92); eigen_assert(false); // not reached } +#ifdef EIGEN_EXCEPTIONS catch (const Foo::Fail&) { /* ignore */ } +#endif + VERIFY_IS_EQUAL(static_cast(0), Foo::object_count); } From 3eba5e1101d8652483e1cf232a06dccf49a8a530 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 19 Jul 2014 14:55:56 +0200 Subject: [PATCH 0647/1103] Implement evaluator for sparse outer products --- Eigen/src/Core/CoreEvaluators.h | 2 +- Eigen/src/SparseCore/SparseAssign.h | 11 +- Eigen/src/SparseCore/SparseDenseProduct.h | 155 +++++++++++++++++++++- Eigen/src/SparseCore/SparseMatrix.h | 4 +- 4 files changed, 163 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index f9f229b09..b19a29e53 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -109,7 +109,7 @@ struct evaluator_base typedef evaluator type; typedef evaluator nestedType; - typedef typename ExpressionType::Index Index; + typedef typename traits::Index Index; // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices. typedef traits ExpressionTraits; }; diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index 066bc2de1..528d63e35 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -186,14 +186,14 @@ void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src) SrcEvaluatorType srcEvaluator(src); const bool transpose = (DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit); - const Index outerSize = (int(SrcEvaluatorType::Flags) & RowMajorBit) ? src.rows() : src.cols(); + const Index outerEvaluationSize = (SrcEvaluatorType::Flags&RowMajorBit) ? src.rows() : src.cols(); if ((!transpose) && src.isRValue()) { // eval without temporary dst.resize(src.rows(), src.cols()); dst.setZero(); dst.reserve((std::max)(src.rows(),src.cols())*2); - for (Index j=0; j dst.setZero(); typename internal::evaluator::type srcEval(src); typename internal::evaluator::type dstEval(dst); - for (Index j=0; j::Flags&RowMajorBit) ? src.rows() : src.cols(); + for (Index j=0; j::InnerIterator i(srcEval,j); i; ++i) dstEval.coeffRef(i.row(),i.col()) = i.value(); } diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 116edd62e..883e24acb 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -13,7 +13,10 @@ namespace Eigen { namespace internal { - + +template <> struct product_promote_storage_type { typedef Sparse ret; }; +template <> struct product_promote_storage_type { typedef Sparse ret; }; + template +// class sparse_dense_outer_product_iterator : public LhsIterator +// { +// typedef typename SparseDenseOuterProduct::Index Index; +// public: +// template +// EIGEN_STRONG_INLINE InnerIterator(const XprEval& prod, Index outer) +// : LhsIterator(prod.lhs(), 0), +// m_outer(outer), m_empty(false), m_factor(get(prod.rhs(), outer, typename internal::traits::StorageKind() )) +// {} +// +// inline Index outer() const { return m_outer; } +// inline Index row() const { return Transpose ? m_outer : Base::index(); } +// inline Index col() const { return Transpose ? Base::index() : m_outer; } +// +// inline Scalar value() const { return Base::value() * m_factor; } +// inline operator bool() const { return Base::operator bool() && !m_empty; } +// +// protected: +// Scalar get(const _RhsNested &rhs, Index outer, Dense = Dense()) const +// { +// return rhs.coeff(outer); +// } +// +// Scalar get(const _RhsNested &rhs, Index outer, Sparse = Sparse()) +// { +// typename Traits::_RhsNested::InnerIterator it(rhs, outer); +// if (it && it.index()==0 && it.value()!=Scalar(0)) +// return it.value(); +// m_empty = true; +// return Scalar(0); +// } +// +// Index m_outer; +// bool m_empty; +// Scalar m_factor; +// }; + +template +struct sparse_dense_outer_product_evaluator +{ +protected: + typedef typename conditional::type Lhs1; + typedef typename conditional::type Rhs; + typedef Product ProdXprType; + + // if the actual left-hand side is a dense vector, + // then build a sparse-view so that we can seamlessly iterator over it. + typedef typename conditional::StorageKind,Sparse>::value, + Lhs1, SparseView >::type Lhs; + typedef typename conditional::StorageKind,Sparse>::value, + Lhs1 const&, SparseView >::type LhsArg; + + typedef typename evaluator::type LhsEval; + typedef typename evaluator::type RhsEval; + typedef typename evaluator::InnerIterator LhsIterator; + typedef typename ProdXprType::Scalar Scalar; + typedef typename ProdXprType::Index Index; + +public: + enum { + Flags = Transpose ? RowMajorBit : 0, + CoeffReadCost = Dynamic + }; + + class InnerIterator : public LhsIterator + { + public: + InnerIterator(const sparse_dense_outer_product_evaluator &xprEval, Index outer) + : LhsIterator(xprEval.m_lhsXprImpl, 0), + m_outer(outer), + m_empty(false), + m_factor(get(xprEval.m_rhsXprImpl, outer, typename internal::traits::StorageKind() )) + {} + + EIGEN_STRONG_INLINE Index outer() const { return m_outer; } + EIGEN_STRONG_INLINE Index row() const { return Transpose ? m_outer : LhsIterator::index(); } + EIGEN_STRONG_INLINE Index col() const { return Transpose ? LhsIterator::index() : m_outer; } + + EIGEN_STRONG_INLINE Scalar value() const { return LhsIterator::value() * m_factor; } + EIGEN_STRONG_INLINE operator bool() const { return LhsIterator::operator bool() && (!m_empty); } + + + protected: + Scalar get(const RhsEval &rhs, Index outer, Dense = Dense()) const + { + return rhs.coeff(outer); + } + + Scalar get(const RhsEval &rhs, Index outer, Sparse = Sparse()) + { + typename RhsEval::InnerIterator it(rhs, outer); + if (it && it.index()==0 && it.value()!=Scalar(0)) + return it.value(); + m_empty = true; + return Scalar(0); + } + + Index m_outer; + bool m_empty; + Scalar m_factor; + }; + + sparse_dense_outer_product_evaluator(const Lhs &lhs, const Rhs &rhs) + : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs) + {} + + // transpose case + sparse_dense_outer_product_evaluator(const Rhs &rhs, const Lhs1 &lhs) + : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs) + {} + +protected: + const LhsArg m_lhs; + typename evaluator::nestedType m_lhsXprImpl; + typename evaluator::nestedType m_rhsXprImpl; +}; + +// sparse * dense outer product +template +struct product_evaluator, OuterProduct, SparseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : sparse_dense_outer_product_evaluator +{ + typedef sparse_dense_outer_product_evaluator Base; + + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + + product_evaluator(const XprType& xpr) + : Base(xpr.lhs(), xpr.rhs()) + {} + +}; + +template +struct product_evaluator, OuterProduct, DenseShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : sparse_dense_outer_product_evaluator +{ + typedef sparse_dense_outer_product_evaluator Base; + + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + + product_evaluator(const XprType& xpr) + : Base(xpr.lhs(), xpr.rhs()) + {} + +}; + } // end namespace internal #endif // EIGEN_TEST_EVALUATORS diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 9d0d6c3ab..36a024cc7 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -1132,7 +1132,7 @@ EIGEN_DONT_INLINE SparseMatrix& SparseMatrix::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - + const bool needToTranspose = (Flags & RowMajorBit) != (internal::evaluator::Flags & RowMajorBit); if (needToTranspose) { @@ -1140,7 +1140,7 @@ EIGEN_DONT_INLINE SparseMatrix& SparseMatrix::type OtherCopy; + typedef typename internal::nested_eval::type >::type OtherCopy; typedef typename internal::remove_all::type _OtherCopy; typedef internal::evaluator<_OtherCopy> OtherCopyEval; OtherCopy otherCopy(other.derived()); From 62f332fc041a33a151617d93e60503a8a9bbf043 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 19 Jul 2014 15:19:10 +0200 Subject: [PATCH 0648/1103] Make sure we evaluate into temporaries matching evaluator storage order requirements --- Eigen/src/Core/AssignEvaluator.h | 3 +-- Eigen/src/SparseCore/SparseMatrixBase.h | 3 +-- 2 files changed, 2 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index c45ac96f2..43755c2fd 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -695,9 +695,8 @@ void call_assignment(const Dst& dst, const Src& src) template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) { - #ifdef EIGEN_TEST_EVALUATORS - typename Src::PlainObject tmp(src); + typename plain_matrix_type::type tmp(src); #else typename Src::PlainObject tmp(src.rows(), src.cols()); call_assignment_no_alias(tmp, src, internal::assign_op()); diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index a6452fa4e..361a66067 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -102,10 +102,9 @@ template class SparseMatrixBase : public EigenBase Transpose >::type AdjointReturnType; - + // FIXME storage order do not match evaluator storage order typedef SparseMatrix PlainObject; - #ifndef EIGEN_PARSED_BY_DOXYGEN /** This is the "real scalar" type; if the \a Scalar type is already real numbers * (e.g. int, float or double) then \a RealScalar is just the same as \a Scalar. If From 50eef6dfc399b3276005498fc84de6f519708725 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 20 Jul 2014 15:16:34 +0200 Subject: [PATCH 0649/1103] Compilation fixes --- Eigen/src/Core/ProductEvaluators.h | 20 ++++++++++---------- Eigen/src/Core/util/ForwardDeclarations.h | 4 ++-- 2 files changed, 12 insertions(+), 12 deletions(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 9436e200b..dc91ddd4d 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -889,8 +889,8 @@ struct product_evaluator, ProductTag, DenseShape, * Products with permutation matrices ***************************************************************************/ -template -struct generic_product_impl +template +struct generic_product_impl { template static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) @@ -900,8 +900,8 @@ struct generic_product_impl } }; -template -struct generic_product_impl +template +struct generic_product_impl { template static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) @@ -911,8 +911,8 @@ struct generic_product_impl } }; -template -struct generic_product_impl, Rhs, PermutationShape, DenseShape, ProductType> +template +struct generic_product_impl, Rhs, PermutationShape, DenseShape, ProductTag> { template static void evalTo(Dest& dst, const Transpose& lhs, const Rhs& rhs) @@ -922,8 +922,8 @@ struct generic_product_impl, Rhs, PermutationShape, DenseShape, P } }; -template -struct generic_product_impl, DenseShape, PermutationShape, ProductType> +template +struct generic_product_impl, DenseShape, PermutationShape, ProductTag> { template static void evalTo(Dest& dst, const Lhs& lhs, const Transpose& rhs) @@ -935,7 +935,7 @@ struct generic_product_impl, DenseShape, PermutationShape, P // TODO: left/right and self-adj/symmetric/permutation look the same ... Too much boilerplate? template -struct product_evaluator, ProductTag, PermutationShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> +struct product_evaluator, ProductTag, PermutationShape, DenseShape, typename traits::Scalar, typename traits::Scalar> : public evaluator::PlainObject>::type { typedef Product XprType; @@ -954,7 +954,7 @@ protected: }; template -struct product_evaluator, ProductTag, DenseShape, PermutationShape, typename Lhs::Scalar, typename Rhs::Scalar> +struct product_evaluator, ProductTag, DenseShape, PermutationShape, typename traits::Scalar, typename traits::Scalar> : public evaluator::PlainObject>::type { typedef Product XprType; diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 274cfac00..99aa9b372 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -164,8 +164,8 @@ template< typename T, int ProductTag = internal::product_type::ret, typename LhsShape = typename evaluator_traits::Shape, typename RhsShape = typename evaluator_traits::Shape, - typename LhsScalar = typename T::Lhs::Scalar, - typename RhsScalar = typename T::Rhs::Scalar + typename LhsScalar = typename traits::Scalar, + typename RhsScalar = typename traits::Scalar > struct product_evaluator; } From 946b99dd5c537f7c93d477645c87fff079a58b6f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 21 Jul 2014 11:45:54 +0200 Subject: [PATCH 0650/1103] Extend qr unit test --- test/qr_fullpivoting.cpp | 6 +++++- test/vectorwiseop.cpp | 4 ++-- 2 files changed, 7 insertions(+), 3 deletions(-) diff --git a/test/qr_fullpivoting.cpp b/test/qr_fullpivoting.cpp index 511f2473f..601773404 100644 --- a/test/qr_fullpivoting.cpp +++ b/test/qr_fullpivoting.cpp @@ -40,7 +40,11 @@ template void qr() MatrixType c = qr.matrixQ() * r * qr.colsPermutation().inverse(); VERIFY_IS_APPROX(m1, c); - + + // stress the ReturnByValue mechanism + MatrixType tmp; + VERIFY_IS_APPROX(tmp.noalias() = qr.matrixQ() * r, (qr.matrixQ() * r).eval()); + MatrixType m2 = MatrixType::Random(cols,cols2); MatrixType m3 = m1*m2; m2 = MatrixType::Random(cols,cols2); diff --git a/test/vectorwiseop.cpp b/test/vectorwiseop.cpp index 6cd1acdda..1631d54c4 100644 --- a/test/vectorwiseop.cpp +++ b/test/vectorwiseop.cpp @@ -104,8 +104,8 @@ template void vectorwiseop_array(const ArrayType& m) m2 = m1; // yes, there might be an aliasing issue there but ".rowwise() /=" - // is suppposed to evaluate " m2.colwise().sum()" into to temporary to avoid - // evaluating the reducions multiple times + // is supposed to evaluate " m2.colwise().sum()" into a temporary to avoid + // evaluating the reduction multiple times if(ArrayType::RowsAtCompileTime>2 || ArrayType::RowsAtCompileTime==Dynamic) { m2.rowwise() /= m2.colwise().sum(); From 9b729f93a10a43a498f4c10f8d80c31a94ae7a0c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 21 Jul 2014 11:46:47 +0200 Subject: [PATCH 0651/1103] Resizing is done by call_assignment_noalias, so no need to perform it when dealing with aliasing. --- Eigen/src/Core/AssignEvaluator.h | 2 -- 1 file changed, 2 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 43755c2fd..1727a6a4a 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -702,8 +702,6 @@ void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable call_assignment_no_alias(tmp, src, internal::assign_op()); #endif - // resizing - dst.resize(tmp.rows(), tmp.cols()); call_assignment_no_alias(dst, tmp, func); } From 2a251ffab0f3abd1fcfe340a4bee61e352d83424 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Jul 2014 09:32:40 +0200 Subject: [PATCH 0652/1103] Implement evaluator for sparse-selfadjoint products --- Eigen/SparseCore | 3 +- Eigen/src/Core/GeneralProduct.h | 16 +- Eigen/src/Core/Product.h | 24 +- Eigen/src/Core/SelfAdjointView.h | 1 + Eigen/src/SparseCore/SparseAssign.h | 3 + Eigen/src/SparseCore/SparseDenseProduct.h | 39 -- Eigen/src/SparseCore/SparseMatrix.h | 4 + Eigen/src/SparseCore/SparseSelfAdjointView.h | 397 +++++++++++++++---- test/sparse_product.cpp | 10 +- 9 files changed, 352 insertions(+), 145 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 7cbfb47f2..69b413ec2 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -53,11 +53,12 @@ struct Sparse {}; #include "src/SparseCore/SparseSparseProductWithPruning.h" #include "src/SparseCore/SparseProduct.h" #include "src/SparseCore/SparseDenseProduct.h" +#include "src/SparseCore/SparseSelfAdjointView.h" #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" #include "src/SparseCore/SparseTriangularView.h" -#include "src/SparseCore/SparseSelfAdjointView.h" + #include "src/SparseCore/TriangularSolver.h" #endif diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 4c1922741..76271a87d 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -62,14 +62,14 @@ template struct product_type typedef typename remove_all::type _Lhs; typedef typename remove_all::type _Rhs; enum { - MaxRows = _Lhs::MaxRowsAtCompileTime, - Rows = _Lhs::RowsAtCompileTime, - MaxCols = _Rhs::MaxColsAtCompileTime, - Cols = _Rhs::ColsAtCompileTime, - MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::MaxColsAtCompileTime, - _Rhs::MaxRowsAtCompileTime), - Depth = EIGEN_SIZE_MIN_PREFER_FIXED(_Lhs::ColsAtCompileTime, - _Rhs::RowsAtCompileTime) + MaxRows = traits<_Lhs>::MaxRowsAtCompileTime, + Rows = traits<_Lhs>::RowsAtCompileTime, + MaxCols = traits<_Rhs>::MaxColsAtCompileTime, + Cols = traits<_Rhs>::ColsAtCompileTime, + MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::MaxColsAtCompileTime, + traits<_Rhs>::MaxRowsAtCompileTime), + Depth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::ColsAtCompileTime, + traits<_Rhs>::RowsAtCompileTime) }; // the splitting into different lines of code here, introducing the _select enums and the typedef below, diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index cb4d4c924..9e5e47d13 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -58,24 +58,26 @@ struct traits > { typedef typename remove_all::type LhsCleaned; typedef typename remove_all::type RhsCleaned; + typedef traits LhsTraits; + typedef traits RhsTraits; typedef MatrixXpr XprKind; typedef typename product_result_scalar::Scalar Scalar; - typedef typename product_promote_storage_type::StorageKind, - typename traits::StorageKind, + typedef typename product_promote_storage_type::ret>::ret StorageKind; - typedef typename promote_index_type::Index, - typename traits::Index>::type Index; + typedef typename promote_index_type::type Index; enum { - RowsAtCompileTime = LhsCleaned::RowsAtCompileTime, - ColsAtCompileTime = RhsCleaned::ColsAtCompileTime, - MaxRowsAtCompileTime = LhsCleaned::MaxRowsAtCompileTime, - MaxColsAtCompileTime = RhsCleaned::MaxColsAtCompileTime, + RowsAtCompileTime = LhsTraits::RowsAtCompileTime, + ColsAtCompileTime = RhsTraits::ColsAtCompileTime, + MaxRowsAtCompileTime = LhsTraits::MaxRowsAtCompileTime, + MaxColsAtCompileTime = RhsTraits::MaxColsAtCompileTime, // FIXME: only needed by GeneralMatrixMatrixTriangular - InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsCleaned::ColsAtCompileTime, RhsCleaned::RowsAtCompileTime), + InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsTraits::ColsAtCompileTime, RhsTraits::RowsAtCompileTime), #ifndef EIGEN_TEST_EVALUATORS // dummy, for evaluators unit test only @@ -84,8 +86,8 @@ struct traits > // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator. Flags = ( MaxRowsAtCompileTime==1 - || ((LhsCleaned::Flags&NoPreferredStorageOrderBit) && (RhsCleaned::Flags&RowMajorBit)) - || ((RhsCleaned::Flags&NoPreferredStorageOrderBit) && (LhsCleaned::Flags&RowMajorBit)) ) + || ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit)) + || ((RhsTraits::Flags&NoPreferredStorageOrderBit) && (LhsTraits::Flags&RowMajorBit)) ) ? RowMajorBit : (MaxColsAtCompileTime==1 ? 0 : NoPreferredStorageOrderBit) }; }; diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 9319ad87c..c8bdec5d4 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -328,6 +328,7 @@ struct triangular_assignment_selector // in the future selfadjoint-ness should be defined by the expression traits // such that Transpose > is valid. (currently TriangularBase::transpose() is overloaded to make it work) diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index 528d63e35..f3da8c6c4 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -127,6 +127,7 @@ template template Derived& SparseMatrixBase::operator=(const EigenBase &other) { + // TODO use the evaluator mechanism other.derived().evalTo(derived()); return derived(); } @@ -135,6 +136,7 @@ template template Derived& SparseMatrixBase::operator=(const ReturnByValue& other) { + // TODO use the evaluator mechanism other.evalTo(derived()); return derived(); } @@ -143,6 +145,7 @@ template template inline Derived& SparseMatrixBase::operator=(const SparseMatrixBase& other) { + // FIXME, by default sparse evaluation do not alias, so we should be able to bypass the generic call_assignment internal::call_assignment/*_no_alias*/(derived(), other.derived()); return derived(); } diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 883e24acb..8864b7308 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -448,45 +448,6 @@ protected: PlainObject m_result; }; - -// template -// class sparse_dense_outer_product_iterator : public LhsIterator -// { -// typedef typename SparseDenseOuterProduct::Index Index; -// public: -// template -// EIGEN_STRONG_INLINE InnerIterator(const XprEval& prod, Index outer) -// : LhsIterator(prod.lhs(), 0), -// m_outer(outer), m_empty(false), m_factor(get(prod.rhs(), outer, typename internal::traits::StorageKind() )) -// {} -// -// inline Index outer() const { return m_outer; } -// inline Index row() const { return Transpose ? m_outer : Base::index(); } -// inline Index col() const { return Transpose ? Base::index() : m_outer; } -// -// inline Scalar value() const { return Base::value() * m_factor; } -// inline operator bool() const { return Base::operator bool() && !m_empty; } -// -// protected: -// Scalar get(const _RhsNested &rhs, Index outer, Dense = Dense()) const -// { -// return rhs.coeff(outer); -// } -// -// Scalar get(const _RhsNested &rhs, Index outer, Sparse = Sparse()) -// { -// typename Traits::_RhsNested::InnerIterator it(rhs, outer); -// if (it && it.index()==0 && it.value()!=Scalar(0)) -// return it.value(); -// m_empty = true; -// return Scalar(0); -// } -// -// Index m_outer; -// bool m_empty; -// Scalar m_factor; -// }; - template struct sparse_dense_outer_product_evaluator { diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 36a024cc7..5f0c3d0a7 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -664,7 +664,11 @@ class SparseMatrix : m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0) { check_template_parameters(); +#ifndef EIGEN_TEST_EVALUATORS *this = other; +#else + Base::operator=(other); +#endif } /** Copy constructor (it performs a deep copy) */ diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 56c922929..8bd836883 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -12,13 +12,23 @@ namespace Eigen { +#ifndef EIGEN_TEST_EVALUATORS + +template +class SparseSelfAdjointTimeDenseProduct; + +template +class DenseTimeSparseSelfAdjointProduct; + +#endif // #ifndef EIGEN_TEST_EVALUATORS + /** \ingroup SparseCore_Module * \class SparseSelfAdjointView * * \brief Pseudo expression to manipulate a triangular sparse matrix as a selfadjoint matrix. * * \param MatrixType the type of the dense matrix storing the coefficients - * \param UpLo can be either \c #Lower or \c #Upper + * \param Mode can be either \c #Lower or \c #Upper * * This class is an expression of a sefladjoint matrix from a triangular part of a matrix * with given dense storage of the coefficients. It is the return type of MatrixBase::selfadjointView() @@ -26,37 +36,33 @@ namespace Eigen { * * \sa SparseMatrixBase::selfadjointView() */ -template -class SparseSelfAdjointTimeDenseProduct; - -template -class DenseTimeSparseSelfAdjointProduct; - namespace internal { -template -struct traits > : traits { +template +struct traits > : traits { }; -template +template void permute_symm_to_symm(const MatrixType& mat, SparseMatrix& _dest, const typename MatrixType::Index* perm = 0); -template +template void permute_symm_to_fullsymm(const MatrixType& mat, SparseMatrix& _dest, const typename MatrixType::Index* perm = 0); } -template class SparseSelfAdjointView - : public EigenBase > +template class SparseSelfAdjointView + : public EigenBase > { public: + + enum { Mode = _Mode }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::Index Index; typedef Matrix VectorI; typedef typename MatrixType::Nested MatrixTypeNested; typedef typename internal::remove_all::type _MatrixTypeNested; - + inline SparseSelfAdjointView(const MatrixType& matrix) : m_matrix(matrix) { eigen_assert(rows()==cols() && "SelfAdjointView is only for squared matrices"); @@ -74,40 +80,76 @@ template class SparseSelfAdjointView * Note that there is no algorithmic advantage of performing such a product compared to a general sparse-sparse matrix product. * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product. */ +#ifndef EIGEN_TEST_EVALUATORS template SparseSparseProduct operator*(const SparseMatrixBase& rhs) const { return SparseSparseProduct(*this, rhs.derived()); } +#else + template + Product + operator*(const SparseMatrixBase& rhs) const + { + return Product(*this, rhs.derived()); + } +#endif /** \returns an expression of the matrix product between a sparse matrix \a lhs and a sparse self-adjoint matrix \a rhs. * * Note that there is no algorithmic advantage of performing such a product compared to a general sparse-sparse matrix product. * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product. */ +#ifndef EIGEN_TEST_EVALUATORS template friend SparseSparseProduct operator*(const SparseMatrixBase& lhs, const SparseSelfAdjointView& rhs) { return SparseSparseProduct(lhs.derived(), rhs); } +#else // EIGEN_TEST_EVALUATORS + template friend + Product + operator*(const SparseMatrixBase& lhs, const SparseSelfAdjointView& rhs) + { + return Product(lhs.derived(), rhs); + } +#endif // EIGEN_TEST_EVALUATORS /** Efficient sparse self-adjoint matrix times dense vector/matrix product */ +#ifndef EIGEN_TEST_EVALUATORS template - SparseSelfAdjointTimeDenseProduct + SparseSelfAdjointTimeDenseProduct operator*(const MatrixBase& rhs) const { - return SparseSelfAdjointTimeDenseProduct(m_matrix, rhs.derived()); + return SparseSelfAdjointTimeDenseProduct(m_matrix, rhs.derived()); } +#else + template + Product + operator*(const MatrixBase& rhs) const + { + return Product(*this, rhs.derived()); + } +#endif /** Efficient dense vector/matrix times sparse self-adjoint matrix product */ +#ifndef EIGEN_TEST_EVALUATORS template friend - DenseTimeSparseSelfAdjointProduct + DenseTimeSparseSelfAdjointProduct operator*(const MatrixBase& lhs, const SparseSelfAdjointView& rhs) { - return DenseTimeSparseSelfAdjointProduct(lhs.derived(), rhs.m_matrix); + return DenseTimeSparseSelfAdjointProduct(lhs.derived(), rhs.m_matrix); } +#else + template friend + Product + operator*(const MatrixBase& lhs, const SparseSelfAdjointView& rhs) + { + return Product(lhs.derived(), rhs); + } +#endif /** Perform a symmetric rank K update of the selfadjoint matrix \c *this: * \f$ this = this + \alpha ( u u^* ) \f$ where \a u is a vector or matrix. @@ -123,30 +165,31 @@ template class SparseSelfAdjointView /** \internal triggered by sparse_matrix = SparseSelfadjointView; */ template void evalTo(SparseMatrix& _dest) const { - internal::permute_symm_to_fullsymm(m_matrix, _dest); + internal::permute_symm_to_fullsymm(m_matrix, _dest); } template void evalTo(DynamicSparseMatrix& _dest) const { // TODO directly evaluate into _dest; SparseMatrix tmp(_dest.rows(),_dest.cols()); - internal::permute_symm_to_fullsymm(m_matrix, tmp); + internal::permute_symm_to_fullsymm(m_matrix, tmp); _dest = tmp; } /** \returns an expression of P H P^-1 */ - SparseSymmetricPermutationProduct<_MatrixTypeNested,UpLo> twistedBy(const PermutationMatrix& perm) const +#ifndef EIGEN_TEST_EVALUATORS + SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode> twistedBy(const PermutationMatrix& perm) const { - return SparseSymmetricPermutationProduct<_MatrixTypeNested,UpLo>(m_matrix, perm); + return SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode>(m_matrix, perm); } - - template - SparseSelfAdjointView& operator=(const SparseSymmetricPermutationProduct& permutedMatrix) + + template + SparseSelfAdjointView& operator=(const SparseSymmetricPermutationProduct& permutedMatrix) { permutedMatrix.evalTo(*this); return *this; } - +#endif // EIGEN_TEST_EVALUATORS SparseSelfAdjointView& operator=(const SparseSelfAdjointView& src) { @@ -154,22 +197,18 @@ template class SparseSelfAdjointView return *this = src.twistedBy(pnull); } - template - SparseSelfAdjointView& operator=(const SparseSelfAdjointView& src) + template + SparseSelfAdjointView& operator=(const SparseSelfAdjointView& src) { PermutationMatrix pnull; return *this = src.twistedBy(pnull); } - - // const SparseLLT llt() const; - // const SparseLDLT ldlt() const; - protected: typename MatrixType::Nested m_matrix; - mutable VectorI m_countPerRow; - mutable VectorI m_countPerCol; + //mutable VectorI m_countPerRow; + //mutable VectorI m_countPerCol; }; /*************************************************************************** @@ -177,15 +216,15 @@ template class SparseSelfAdjointView ***************************************************************************/ template -template -const SparseSelfAdjointView SparseMatrixBase::selfadjointView() const +template +const SparseSelfAdjointView SparseMatrixBase::selfadjointView() const { return derived(); } template -template -SparseSelfAdjointView SparseMatrixBase::selfadjointView() +template +SparseSelfAdjointView SparseMatrixBase::selfadjointView() { return derived(); } @@ -194,16 +233,16 @@ SparseSelfAdjointView SparseMatrixBase::selfadjointView( * Implementation of SparseSelfAdjointView methods ***************************************************************************/ -template +template template -SparseSelfAdjointView& -SparseSelfAdjointView::rankUpdate(const SparseMatrixBase& u, const Scalar& alpha) +SparseSelfAdjointView& +SparseSelfAdjointView::rankUpdate(const SparseMatrixBase& u, const Scalar& alpha) { SparseMatrix tmp = u * u.adjoint(); if(alpha==Scalar(0)) - m_matrix.const_cast_derived() = tmp.template triangularView(); + m_matrix.const_cast_derived() = tmp.template triangularView(); else - m_matrix.const_cast_derived() += alpha * tmp.template triangularView(); + m_matrix.const_cast_derived() += alpha * tmp.template triangularView(); return *this; } @@ -212,18 +251,19 @@ SparseSelfAdjointView::rankUpdate(const SparseMatrixBase -struct traits > - : traits, Lhs, Rhs> > +template +struct traits > + : traits, Lhs, Rhs> > { typedef Dense StorageKind; }; } -template +template class SparseSelfAdjointTimeDenseProduct - : public ProductBase, Lhs, Rhs> + : public ProductBase, Lhs, Rhs> { public: EIGEN_PRODUCT_PUBLIC_INTERFACE(SparseSelfAdjointTimeDenseProduct) @@ -241,9 +281,9 @@ class SparseSelfAdjointTimeDenseProduct enum { LhsIsRowMajor = (_Lhs::Flags&RowMajorBit)==RowMajorBit, ProcessFirstHalf = - ((UpLo&(Upper|Lower))==(Upper|Lower)) - || ( (UpLo&Upper) && !LhsIsRowMajor) - || ( (UpLo&Lower) && LhsIsRowMajor), + ((Mode&(Upper|Lower))==(Upper|Lower)) + || ( (Mode&Upper) && !LhsIsRowMajor) + || ( (Mode&Lower) && LhsIsRowMajor), ProcessSecondHalf = !ProcessFirstHalf }; for (typename _Lhs::Index j=0; j -struct traits > - : traits, Lhs, Rhs> > +template +struct traits > + : traits, Lhs, Rhs> > {}; } -template +template class DenseTimeSparseSelfAdjointProduct - : public ProductBase, Lhs, Rhs> + : public ProductBase, Lhs, Rhs> { public: EIGEN_PRODUCT_PUBLIC_INTERFACE(DenseTimeSparseSelfAdjointProduct) @@ -301,16 +341,197 @@ class DenseTimeSparseSelfAdjointProduct DenseTimeSparseSelfAdjointProduct& operator=(const DenseTimeSparseSelfAdjointProduct&); }; +#else // EIGEN_TEST_EVALUATORS + +namespace internal { + +template +inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) +{ + EIGEN_ONLY_USED_FOR_DEBUG(alpha); + // TODO use alpha + eigen_assert(alpha==AlphaType(1) && "alpha != 1 is not implemented yet, sorry"); + + typedef typename evaluator::type LhsEval; + typedef typename evaluator::InnerIterator LhsIterator; + typedef typename SparseLhsType::Index Index; + typedef typename SparseLhsType::Scalar LhsScalar; + + enum { + LhsIsRowMajor = (LhsEval::Flags&RowMajorBit)==RowMajorBit, + ProcessFirstHalf = + ((Mode&(Upper|Lower))==(Upper|Lower)) + || ( (Mode&Upper) && !LhsIsRowMajor) + || ( (Mode&Lower) && LhsIsRowMajor), + ProcessSecondHalf = !ProcessFirstHalf + }; + + LhsEval lhsEval(lhs); + + for (Index j=0; j +// in the future selfadjoint-ness should be defined by the expression traits +// such that Transpose > is valid. (currently TriangularBase::transpose() is overloaded to make it work) +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef SparseSelfAdjointShape Shape; + + static const int AssumeAliasing = 0; +}; + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const LhsView& lhsView, const Rhs& rhs) + { + typedef typename LhsView::_MatrixTypeNested Lhs; + typedef typename nested_eval::type LhsNested; + typedef typename nested_eval::type RhsNested; + LhsNested lhsNested(lhsView.matrix()); + RhsNested rhsNested(rhs); + + dst.setZero(); + internal::sparse_selfadjoint_time_dense_product(lhsNested, rhsNested, dst, typename Dest::Scalar(1)); + } +}; + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const RhsView& rhsView) + { + typedef typename RhsView::_MatrixTypeNested Rhs; + typedef typename nested_eval::type LhsNested; + typedef typename nested_eval::type RhsNested; + LhsNested lhsNested(lhs); + RhsNested rhsNested(rhsView.matrix()); + + dst.setZero(); + // transpoe everything + Transpose dstT(dst); + internal::sparse_selfadjoint_time_dense_product(rhsNested.transpose(), lhsNested.transpose(), dstT, typename Dest::Scalar(1)); + } +}; + +template +struct product_evaluator, ProductTag, SparseSelfAdjointShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +template +struct product_evaluator, ProductTag, DenseShape, SparseSelfAdjointShape, typename Lhs::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +template +struct product_evaluator, ProductTag, SparseSelfAdjointShape, SparseShape, typename LhsView::Scalar, typename Rhs::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : /*m_lhs(xpr.lhs()),*/ m_result(xpr.rows(), xpr.cols()) + { + m_lhs = xpr.lhs(); + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, m_lhs, xpr.rhs()); + } + +protected: + typename Rhs::PlainObject m_lhs; + PlainObject m_result; +}; + +template +struct product_evaluator, ProductTag, SparseShape, SparseSelfAdjointShape, typename Lhs::Scalar, typename RhsView::Scalar> + : public evaluator::PlainObject>::type +{ + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_rhs(xpr.rhs()), m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), m_rhs); + } + +protected: + typename Lhs::PlainObject m_rhs; + PlainObject m_result; +}; + +} // namespace internal + +#endif // EIGEN_TEST_EVALUATORS + /*************************************************************************** * Implementation of symmetric copies and permutations ***************************************************************************/ namespace internal { - -template -struct traits > : traits { -}; -template +template void permute_symm_to_fullsymm(const MatrixType& mat, SparseMatrix& _dest, const typename MatrixType::Index* perm) { typedef typename MatrixType::Index Index; @@ -337,11 +558,11 @@ void permute_symm_to_fullsymm(const MatrixType& mat, SparseMatrixc) || ( UpLo==Upper && rc) || ( Mode==Upper && rc) || ( (UpLo&Upper)==Upper && rc) || ( (Mode&Upper)==Upper && r +template void permute_symm_to_symm(const MatrixType& mat, SparseMatrix& _dest, const typename MatrixType::Index* perm) { typedef typename MatrixType::Index Index; @@ -407,8 +628,8 @@ void permute_symm_to_symm(const MatrixType& mat, SparseMatrixj)) + if((int(SrcMode)==int(Lower) && ij)) continue; Index ip = perm ? perm[i] : i; - count[int(DstUpLo)==int(Lower) ? (std::min)(ip,jp) : (std::max)(ip,jp)]++; + count[int(DstMode)==int(Lower) ? (std::min)(ip,jp) : (std::max)(ip,jp)]++; } } dest.outerIndexPtr()[0] = 0; @@ -441,17 +662,17 @@ void permute_symm_to_symm(const MatrixType& mat, SparseMatrixj)) + if((int(SrcMode)==int(Lower) && ij)) continue; Index jp = perm ? perm[j] : j; Index ip = perm? perm[i] : i; - Index k = count[int(DstUpLo)==int(Lower) ? (std::min)(ip,jp) : (std::max)(ip,jp)]++; - dest.innerIndexPtr()[k] = int(DstUpLo)==int(Lower) ? (std::max)(ip,jp) : (std::min)(ip,jp); + Index k = count[int(DstMode)==int(Lower) ? (std::min)(ip,jp) : (std::max)(ip,jp)]++; + dest.innerIndexPtr()[k] = int(DstMode)==int(Lower) ? (std::max)(ip,jp) : (std::min)(ip,jp); if(!StorageOrderMatch) std::swap(ip,jp); - if( ((int(DstUpLo)==int(Lower) && ipjp))) + if( ((int(DstMode)==int(Lower) && ipjp))) dest.valuePtr()[k] = numext::conj(it.value()); else dest.valuePtr()[k] = it.value(); @@ -461,9 +682,19 @@ void permute_symm_to_symm(const MatrixType& mat, SparseMatrix +#ifndef EIGEN_TEST_EVALUATORS + +namespace internal { + +template +struct traits > : traits { +}; + +} + +template class SparseSymmetricPermutationProduct - : public EigenBase > + : public EigenBase > { public: typedef typename MatrixType::Scalar Scalar; @@ -485,15 +716,15 @@ class SparseSymmetricPermutationProduct template void evalTo(SparseMatrix& _dest) const { -// internal::permute_symm_to_fullsymm(m_matrix,_dest,m_perm.indices().data()); +// internal::permute_symm_to_fullsymm(m_matrix,_dest,m_perm.indices().data()); SparseMatrix tmp; - internal::permute_symm_to_fullsymm(m_matrix,tmp,m_perm.indices().data()); + internal::permute_symm_to_fullsymm(m_matrix,tmp,m_perm.indices().data()); _dest = tmp; } - template void evalTo(SparseSelfAdjointView& dest) const + template void evalTo(SparseSelfAdjointView& dest) const { - internal::permute_symm_to_symm(m_matrix,dest.matrix(),m_perm.indices().data()); + internal::permute_symm_to_symm(m_matrix,dest.matrix(),m_perm.indices().data()); } protected: @@ -502,6 +733,10 @@ class SparseSymmetricPermutationProduct }; +#else // EIGEN_TEST_EVALUATORS + +#endif // EIGEN_TEST_EVALUATORS + } // end namespace Eigen #endif // EIGEN_SPARSE_SELFADJOINTVIEW_H diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index 27bc548f8..fa9be5440 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -19,7 +19,7 @@ template void sparse_product() typedef typename SparseMatrixType::Scalar Scalar; enum { Flags = SparseMatrixType::Flags }; - double density = (std::max)(8./(rows*cols), 0.1); + double density = (std::max)(8./(rows*cols), 0.2); typedef Matrix DenseMatrix; typedef Matrix DenseVector; typedef Matrix RowDenseVector; @@ -109,7 +109,7 @@ template void sparse_product() Index c1 = internal::random(0,cols-1); Index r1 = internal::random(0,depth-1); DenseMatrix dm5 = DenseMatrix::Random(depth, cols); - + VERIFY_IS_APPROX( m4=m2.col(c)*dm5.col(c1).transpose(), refMat4=refMat2.col(c)*dm5.col(c1).transpose()); VERIFY_IS_EQUAL(m4.nonZeros(), (refMat4.array()!=0).count()); VERIFY_IS_APPROX( m4=m2.middleCols(c,1)*dm5.col(c1).transpose(), refMat4=refMat2.col(c)*dm5.col(c1).transpose()); @@ -153,11 +153,11 @@ template void sparse_product() RowSpVector rv0(depth), rv1; RowDenseVector drv0(depth), drv1(rv1); initSparse(2*density,drv0, rv0); - - VERIFY_IS_APPROX(cv1=rv0*m3, dcv1=drv0*refMat3); + + VERIFY_IS_APPROX(cv1=m3*cv0, dcv1=refMat3*dcv0); VERIFY_IS_APPROX(rv1=rv0*m3, drv1=drv0*refMat3); - VERIFY_IS_APPROX(cv1=m3*cv0, dcv1=refMat3*dcv0); VERIFY_IS_APPROX(cv1=m3t.adjoint()*cv0, dcv1=refMat3t.adjoint()*dcv0); + VERIFY_IS_APPROX(cv1=rv0*m3, dcv1=drv0*refMat3); VERIFY_IS_APPROX(rv1=m3*cv0, drv1=refMat3*dcv0); } From 6daa6a0d164ac3c225645e47f55238e9ba2a32cc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Jul 2014 11:35:56 +0200 Subject: [PATCH 0653/1103] Refactor TriangularView to handle both dense and sparse objects. Introduce a glu_shape helper to assemble sparse/dense shapes with triagular/seladjoint views. --- Eigen/SparseCore | 6 +- Eigen/src/Core/SolveTriangular.h | 10 +- Eigen/src/Core/TriangularMatrix.h | 467 ++++++++++-------- .../products/GeneralMatrixMatrixTriangular.h | 13 +- .../Core/products/TriangularMatrixMatrix.h | 2 + .../Core/products/TriangularMatrixVector.h | 4 +- Eigen/src/Core/util/Constants.h | 14 +- Eigen/src/Core/util/XprHelper.h | 3 + Eigen/src/SparseCholesky/SimplicialCholesky.h | 8 +- Eigen/src/SparseCore/MappedSparseMatrix.h | 28 ++ Eigen/src/SparseCore/SparseMatrixBase.h | 2 +- Eigen/src/SparseCore/SparseSelfAdjointView.h | 2 - Eigen/src/SparseCore/SparseTriangularView.h | 62 +-- Eigen/src/SparseCore/SparseUtil.h | 7 +- Eigen/src/SparseCore/TriangularSolver.h | 176 ++++++- 15 files changed, 520 insertions(+), 284 deletions(-) diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 69b413ec2..578469c1c 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -54,12 +54,12 @@ struct Sparse {}; #include "src/SparseCore/SparseProduct.h" #include "src/SparseCore/SparseDenseProduct.h" #include "src/SparseCore/SparseSelfAdjointView.h" +#include "src/SparseCore/SparseTriangularView.h" +#include "src/SparseCore/TriangularSolver.h" + #ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" -#include "src/SparseCore/SparseTriangularView.h" - -#include "src/SparseCore/TriangularSolver.h" #endif #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/src/Core/SolveTriangular.h b/Eigen/src/Core/SolveTriangular.h index ef17f288e..0f17e3a89 100644 --- a/Eigen/src/Core/SolveTriangular.h +++ b/Eigen/src/Core/SolveTriangular.h @@ -171,10 +171,10 @@ struct triangular_solver_selector { */ template template -void TriangularView::solveInPlace(const MatrixBase& _other) const +void TriangularViewImpl::solveInPlace(const MatrixBase& _other) const { OtherDerived& other = _other.const_cast_derived(); - eigen_assert( cols() == rows() && ((Side==OnTheLeft && cols() == other.rows()) || (Side==OnTheRight && cols() == other.cols())) ); + eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) ); eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower))); enum { copy = internal::traits::Flags & RowMajorBit && OtherDerived::IsVectorAtCompileTime }; @@ -183,7 +183,7 @@ void TriangularView::solveInPlace(const MatrixBase::type, - Side, Mode>::run(nestedExpression(), otherCopy); + Side, Mode>::run(derived().nestedExpression(), otherCopy); if (copy) other = otherCopy; @@ -213,9 +213,9 @@ void TriangularView::solveInPlace(const MatrixBase template const internal::triangular_solve_retval,Other> -TriangularView::solve(const MatrixBase& other) const +TriangularViewImpl::solve(const MatrixBase& other) const { - return internal::triangular_solve_retval(*this, other.derived()); + return internal::triangular_solve_retval(derived(), other.derived()); } namespace internal { diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 69bbaf6a9..25dab0e77 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -49,7 +49,7 @@ template class TriangularBase : public EigenBase typedef typename internal::traits::Scalar Scalar; typedef typename internal::traits::StorageKind StorageKind; typedef typename internal::traits::Index Index; - typedef typename internal::traits::DenseMatrixType DenseMatrixType; + typedef typename internal::traits::FullMatrixType DenseMatrixType; typedef DenseMatrixType DenseType; typedef Derived const& Nested; @@ -172,8 +172,8 @@ struct traits > : traits typedef typename nested::type MatrixTypeNested; typedef typename remove_reference::type MatrixTypeNestedNonRef; typedef typename remove_all::type MatrixTypeNestedCleaned; + typedef typename MatrixType::PlainObject FullMatrixType; typedef MatrixType ExpressionType; - typedef typename MatrixType::PlainObject DenseMatrixType; enum { Mode = _Mode, Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode @@ -185,22 +185,23 @@ struct traits > : traits }; } +#ifndef EIGEN_TEST_EVALUATORS template struct TriangularProduct; +#endif + +template class TriangularViewImpl; template class TriangularView - : public TriangularBase > + : public TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind > { public: - typedef TriangularBase Base; + typedef TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind > Base; typedef typename internal::traits::Scalar Scalar; - typedef _MatrixType MatrixType; - typedef typename internal::traits::DenseMatrixType DenseMatrixType; - typedef DenseMatrixType PlainObject; protected: typedef typename internal::traits::MatrixTypeNested MatrixTypeNested; @@ -210,8 +211,6 @@ template class TriangularView typedef typename internal::remove_all::type MatrixConjugateReturnType; public: - using Base::evalToLazy; - typedef typename internal::traits::StorageKind StorageKind; typedef typename internal::traits::Index Index; @@ -228,110 +227,19 @@ template class TriangularView EIGEN_DEVICE_FUNC inline TriangularView(const MatrixType& matrix) : m_matrix(matrix) {} + + using Base::operator=; EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } - EIGEN_DEVICE_FUNC - inline Index outerStride() const { return m_matrix.outerStride(); } - EIGEN_DEVICE_FUNC - inline Index innerStride() const { return m_matrix.innerStride(); } - -#ifdef EIGEN_TEST_EVALUATORS - - /** \sa MatrixBase::operator+=() */ - template - EIGEN_DEVICE_FUNC - TriangularView& operator+=(const DenseBase& other) { - internal::call_assignment_no_alias(*this, other.derived(), internal::add_assign_op()); - return *this; - } - /** \sa MatrixBase::operator-=() */ - template - EIGEN_DEVICE_FUNC - TriangularView& operator-=(const DenseBase& other) { - internal::call_assignment_no_alias(*this, other.derived(), internal::sub_assign_op()); - return *this; - } - -#else - /** \sa MatrixBase::operator+=() */ - template - EIGEN_DEVICE_FUNC - TriangularView& operator+=(const DenseBase& other) { return *this = m_matrix + other.derived(); } - /** \sa MatrixBase::operator-=() */ - template - EIGEN_DEVICE_FUNC - TriangularView& operator-=(const DenseBase& other) { return *this = m_matrix - other.derived(); } -#endif - /** \sa MatrixBase::operator*=() */ - EIGEN_DEVICE_FUNC - TriangularView& operator*=(const typename internal::traits::Scalar& other) { return *this = m_matrix * other; } - /** \sa MatrixBase::operator/=() */ - EIGEN_DEVICE_FUNC - TriangularView& operator/=(const typename internal::traits::Scalar& other) { return *this = m_matrix / other; } - - /** \sa MatrixBase::fill() */ - EIGEN_DEVICE_FUNC - void fill(const Scalar& value) { setConstant(value); } - /** \sa MatrixBase::setConstant() */ - EIGEN_DEVICE_FUNC - TriangularView& setConstant(const Scalar& value) - { return *this = MatrixType::Constant(rows(), cols(), value); } - /** \sa MatrixBase::setZero() */ - EIGEN_DEVICE_FUNC - TriangularView& setZero() { return setConstant(Scalar(0)); } - /** \sa MatrixBase::setOnes() */ - EIGEN_DEVICE_FUNC - TriangularView& setOnes() { return setConstant(Scalar(1)); } - - /** \sa MatrixBase::coeff() - * \warning the coordinates must fit into the referenced triangular part - */ - EIGEN_DEVICE_FUNC - inline Scalar coeff(Index row, Index col) const - { - Base::check_coordinates_internal(row, col); - return m_matrix.coeff(row, col); - } - - /** \sa MatrixBase::coeffRef() - * \warning the coordinates must fit into the referenced triangular part - */ - EIGEN_DEVICE_FUNC - inline Scalar& coeffRef(Index row, Index col) - { - Base::check_coordinates_internal(row, col); - return m_matrix.const_cast_derived().coeffRef(row, col); - } EIGEN_DEVICE_FUNC const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; } EIGEN_DEVICE_FUNC MatrixTypeNestedCleaned& nestedExpression() { return *const_cast(&m_matrix); } - /** Assigns a triangular matrix to a triangular part of a dense matrix */ - template - EIGEN_DEVICE_FUNC - TriangularView& operator=(const TriangularBase& other); - - template - EIGEN_DEVICE_FUNC - TriangularView& operator=(const MatrixBase& other); - - EIGEN_DEVICE_FUNC - TriangularView& operator=(const TriangularView& other) - { return *this = other.nestedExpression(); } - - template - EIGEN_DEVICE_FUNC - void lazyAssign(const TriangularBase& other); - - template - EIGEN_DEVICE_FUNC - void lazyAssign(const MatrixBase& other); - /** \sa MatrixBase::conjugate() */ EIGEN_DEVICE_FUNC inline TriangularView conjugate() @@ -360,78 +268,16 @@ template class TriangularView return m_matrix.transpose(); } -#ifdef EIGEN_TEST_EVALUATORS - - /** Efficient triangular matrix times vector/matrix product */ - template - EIGEN_DEVICE_FUNC - const Product - operator*(const MatrixBase& rhs) const - { - return Product(*this, rhs.derived()); - } - - /** Efficient vector/matrix times triangular matrix product */ - template friend - EIGEN_DEVICE_FUNC - const Product - operator*(const MatrixBase& lhs, const TriangularView& rhs) - { - return Product(lhs.derived(),rhs); - } - -#else // EIGEN_TEST_EVALUATORS - /** Efficient triangular matrix times vector/matrix product */ - template - EIGEN_DEVICE_FUNC - TriangularProduct - operator*(const MatrixBase& rhs) const - { - return TriangularProduct - - (m_matrix, rhs.derived()); - } - - /** Efficient vector/matrix times triangular matrix product */ - template friend - EIGEN_DEVICE_FUNC - TriangularProduct - operator*(const MatrixBase& lhs, const TriangularView& rhs) - { - return TriangularProduct - - (lhs.derived(),rhs.m_matrix); - } -#endif - - template - EIGEN_DEVICE_FUNC - inline const internal::triangular_solve_retval - solve(const MatrixBase& other) const; - - template - EIGEN_DEVICE_FUNC - void solveInPlace(const MatrixBase& other) const; - #ifdef EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC inline const Solve solve(const MatrixBase& other) const { return Solve(*this, other.derived()); } -#else // EIGEN_TEST_EVALUATORS - template - EIGEN_DEVICE_FUNC - inline const internal::triangular_solve_retval - solve(const MatrixBase& other) const - { return solve(other); } + + using Base::solve; #endif // EIGEN_TEST_EVALUATORS - template - EIGEN_DEVICE_FUNC - void solveInPlace(const MatrixBase& other) const - { return solveInPlace(other); } - EIGEN_DEVICE_FUNC const SelfAdjointView selfadjointView() const { @@ -445,30 +291,6 @@ template class TriangularView return SelfAdjointView(m_matrix); } - template - EIGEN_DEVICE_FUNC - void swap(TriangularBase const & other) - { - #ifdef EIGEN_TEST_EVALUATORS - call_assignment(*this, other.const_cast_derived(), internal::swap_assign_op()); - #else - TriangularView,Mode>(const_cast(m_matrix)).lazyAssign(other.const_cast_derived().nestedExpression()); - #endif - } - - // TODO: this overload is ambiguous and it should be deprecated (Gael) - template - EIGEN_DEVICE_FUNC - void swap(MatrixBase const & other) - { - #ifdef EIGEN_TEST_EVALUATORS - call_assignment(*this, other.const_cast_derived(), internal::swap_assign_op()); - #else - SwapWrapper swaper(const_cast(m_matrix)); - TriangularView,Mode>(swaper).lazyAssign(other.derived()); - #endif - } - EIGEN_DEVICE_FUNC Scalar determinant() const { @@ -479,13 +301,227 @@ template class TriangularView else return m_matrix.diagonal().prod(); } + + protected: + + MatrixTypeNested m_matrix; +}; + +template class TriangularViewImpl<_MatrixType,_Mode,Dense> + : public TriangularBase > +{ + public: + + typedef TriangularView<_MatrixType, _Mode> TriangularViewType; + typedef TriangularBase Base; + typedef typename internal::traits::Scalar Scalar; + + typedef _MatrixType MatrixType; + typedef typename MatrixType::PlainObject DenseMatrixType; + typedef DenseMatrixType PlainObject; + + public: + using Base::evalToLazy; + using Base::derived; + + typedef typename internal::traits::StorageKind StorageKind; + typedef typename internal::traits::Index Index; + + enum { + Mode = _Mode, + Flags = internal::traits::Flags + }; + + EIGEN_DEVICE_FUNC + inline Index outerStride() const { return derived().nestedExpression().outerStride(); } + EIGEN_DEVICE_FUNC + inline Index innerStride() const { return derived().nestedExpression().innerStride(); } + +#ifdef EIGEN_TEST_EVALUATORS + + /** \sa MatrixBase::operator+=() */ + template + EIGEN_DEVICE_FUNC + TriangularViewType& operator+=(const DenseBase& other) { + internal::call_assignment_no_alias(derived(), other.derived(), internal::add_assign_op()); + return derived(); + } + /** \sa MatrixBase::operator-=() */ + template + EIGEN_DEVICE_FUNC + TriangularViewType& operator-=(const DenseBase& other) { + internal::call_assignment_no_alias(derived(), other.derived(), internal::sub_assign_op()); + return derived(); + } + +#else + /** \sa MatrixBase::operator+=() */ + template + EIGEN_DEVICE_FUNC + TriangularViewType& operator+=(const DenseBase& other) { return *this = derived().nestedExpression() + other.derived(); } + /** \sa MatrixBase::operator-=() */ + template + EIGEN_DEVICE_FUNC + TriangularViewType& operator-=(const DenseBase& other) { return *this = derived().nestedExpression() - other.derived(); } +#endif + /** \sa MatrixBase::operator*=() */ + EIGEN_DEVICE_FUNC + TriangularViewType& operator*=(const typename internal::traits::Scalar& other) { return *this = derived().nestedExpression() * other; } + /** \sa MatrixBase::operator/=() */ + EIGEN_DEVICE_FUNC + TriangularViewType& operator/=(const typename internal::traits::Scalar& other) { return *this = derived().nestedExpression() / other; } + + /** \sa MatrixBase::fill() */ + EIGEN_DEVICE_FUNC + void fill(const Scalar& value) { setConstant(value); } + /** \sa MatrixBase::setConstant() */ + EIGEN_DEVICE_FUNC + TriangularViewType& setConstant(const Scalar& value) + { return *this = MatrixType::Constant(derived().rows(), derived().cols(), value); } + /** \sa MatrixBase::setZero() */ + EIGEN_DEVICE_FUNC + TriangularViewType& setZero() { return setConstant(Scalar(0)); } + /** \sa MatrixBase::setOnes() */ + EIGEN_DEVICE_FUNC + TriangularViewType& setOnes() { return setConstant(Scalar(1)); } + + /** \sa MatrixBase::coeff() + * \warning the coordinates must fit into the referenced triangular part + */ + EIGEN_DEVICE_FUNC + inline Scalar coeff(Index row, Index col) const + { + Base::check_coordinates_internal(row, col); + return derived().nestedExpression().coeff(row, col); + } + + /** \sa MatrixBase::coeffRef() + * \warning the coordinates must fit into the referenced triangular part + */ + EIGEN_DEVICE_FUNC + inline Scalar& coeffRef(Index row, Index col) + { + Base::check_coordinates_internal(row, col); + return derived().nestedExpression().const_cast_derived().coeffRef(row, col); + } + + /** Assigns a triangular matrix to a triangular part of a dense matrix */ + template + EIGEN_DEVICE_FUNC + TriangularViewType& operator=(const TriangularBase& other); + + template + EIGEN_DEVICE_FUNC + TriangularViewType& operator=(const MatrixBase& other); + + EIGEN_DEVICE_FUNC + TriangularViewType& operator=(const TriangularViewImpl& other) + { return *this = other.nestedExpression(); } + + template + EIGEN_DEVICE_FUNC + void lazyAssign(const TriangularBase& other); + + template + EIGEN_DEVICE_FUNC + void lazyAssign(const MatrixBase& other); + +#ifdef EIGEN_TEST_EVALUATORS + + /** Efficient triangular matrix times vector/matrix product */ + template + EIGEN_DEVICE_FUNC + const Product + operator*(const MatrixBase& rhs) const + { + return Product(derived(), rhs.derived()); + } + + /** Efficient vector/matrix times triangular matrix product */ + template friend + EIGEN_DEVICE_FUNC + const Product + operator*(const MatrixBase& lhs, const TriangularViewImpl& rhs) + { + return Product(lhs.derived(),rhs.derived()); + } + +#else // EIGEN_TEST_EVALUATORS + /** Efficient triangular matrix times vector/matrix product */ + template + EIGEN_DEVICE_FUNC + TriangularProduct + operator*(const MatrixBase& rhs) const + { + return TriangularProduct + + (derived().nestedExpression(), rhs.derived()); + } + + /** Efficient vector/matrix times triangular matrix product */ + template friend + EIGEN_DEVICE_FUNC + TriangularProduct + operator*(const MatrixBase& lhs, const TriangularVieImplw& rhs) + { + return TriangularProduct + + (lhs.derived(),rhs.nestedExpression()); + } +#endif + + template + EIGEN_DEVICE_FUNC + inline const internal::triangular_solve_retval + solve(const MatrixBase& other) const; + + template + EIGEN_DEVICE_FUNC + void solveInPlace(const MatrixBase& other) const; + +#ifndef EIGEN_TEST_EVALUATORS + template + EIGEN_DEVICE_FUNC + inline const internal::triangular_solve_retval + solve(const MatrixBase& other) const + { return solve(other); } +#endif // EIGEN_TEST_EVALUATORS + + template + EIGEN_DEVICE_FUNC + void solveInPlace(const MatrixBase& other) const + { return solveInPlace(other); } + + template + EIGEN_DEVICE_FUNC + void swap(TriangularBase const & other) + { + #ifdef EIGEN_TEST_EVALUATORS + call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op()); + #else + TriangularView,Mode>(const_cast(derived().nestedExpression())).lazyAssign(other.const_cast_derived().nestedExpression()); + #endif + } + + // TODO: this overload is ambiguous and it should be deprecated (Gael) + template + EIGEN_DEVICE_FUNC + void swap(MatrixBase const & other) + { + #ifdef EIGEN_TEST_EVALUATORS + call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op()); + #else + SwapWrapper swaper(const_cast(derived().nestedExpression())); + TriangularView,Mode>(swaper).lazyAssign(other.derived()); + #endif + } #ifndef EIGEN_TEST_EVALUATORS // TODO simplify the following: template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& operator=(const ProductBase& other) + EIGEN_STRONG_INLINE TriangularViewType& operator=(const ProductBase& other) { setZero(); return assignProduct(other,1); @@ -493,14 +529,14 @@ template class TriangularView template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& operator+=(const ProductBase& other) + EIGEN_STRONG_INLINE TriangularViewType& operator+=(const ProductBase& other) { return assignProduct(other,1); } template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& operator-=(const ProductBase& other) + EIGEN_STRONG_INLINE TriangularViewType& operator-=(const ProductBase& other) { return assignProduct(other,-1); } @@ -508,7 +544,7 @@ template class TriangularView template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& operator=(const ScaledProduct& other) + EIGEN_STRONG_INLINE TriangularViewType& operator=(const ScaledProduct& other) { setZero(); return assignProduct(other,other.alpha()); @@ -516,14 +552,14 @@ template class TriangularView template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& operator+=(const ScaledProduct& other) + EIGEN_STRONG_INLINE TriangularViewType& operator+=(const ScaledProduct& other) { return assignProduct(other,other.alpha()); } template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& operator-=(const ScaledProduct& other) + EIGEN_STRONG_INLINE TriangularViewType& operator-=(const ScaledProduct& other) { return assignProduct(other,-other.alpha()); } @@ -542,16 +578,15 @@ template class TriangularView template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& _assignProduct(const ProductType& prod, const Scalar& alpha); + EIGEN_STRONG_INLINE TriangularViewType& _assignProduct(const ProductType& prod, const Scalar& alpha); protected: #else protected: template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularView& assignProduct(const ProductBase& prod, const Scalar& alpha); + EIGEN_STRONG_INLINE TriangularViewType& assignProduct(const ProductBase& prod, const Scalar& alpha); #endif - MatrixTypeNested m_matrix; }; /*************************************************************************** @@ -736,18 +771,18 @@ struct triangular_assignment_selector template inline TriangularView& -TriangularView::operator=(const MatrixBase& other) +TriangularViewImpl::operator=(const MatrixBase& other) { - internal::call_assignment_no_alias(*this, other.derived(), internal::assign_op()); - return *this; + internal::call_assignment_no_alias(derived(), other.derived(), internal::assign_op()); + return derived(); } // FIXME should we keep that possibility template template -void TriangularView::lazyAssign(const MatrixBase& other) +void TriangularViewImpl::lazyAssign(const MatrixBase& other) { - internal::call_assignment(this->noalias(), other.template triangularView()); + internal::call_assignment(derived().noalias(), other.template triangularView()); } @@ -755,19 +790,19 @@ void TriangularView::lazyAssign(const MatrixBase template template inline TriangularView& -TriangularView::operator=(const TriangularBase& other) +TriangularViewImpl::operator=(const TriangularBase& other) { eigen_assert(Mode == int(OtherDerived::Mode)); - internal::call_assignment(*this, other.derived()); - return *this; + internal::call_assignment(derived(), other.derived()); + return derived(); } template template -void TriangularView::lazyAssign(const TriangularBase& other) +void TriangularViewImpl::lazyAssign(const TriangularBase& other) { eigen_assert(Mode == int(OtherDerived::Mode)); - internal::call_assignment(this->noalias(), other.derived()); + internal::call_assignment(derived().noalias(), other.derived()); } #else @@ -776,7 +811,7 @@ void TriangularView::lazyAssign(const TriangularBase template inline TriangularView& -TriangularView::operator=(const MatrixBase& other) +TriangularViewImpl::operator=(const MatrixBase& other) { if(OtherDerived::Flags & EvalBeforeAssigningBit) { @@ -786,13 +821,13 @@ TriangularView::operator=(const MatrixBase& othe } else lazyAssign(other.derived()); - return *this; + return derived(); } // FIXME should we keep that possibility template template -void TriangularView::lazyAssign(const MatrixBase& other) +void TriangularViewImp::lazyAssign(const MatrixBase& other) { enum { unroll = MatrixType::SizeAtCompileTime != Dynamic @@ -813,7 +848,7 @@ void TriangularView::lazyAssign(const MatrixBase template template inline TriangularView& -TriangularView::operator=(const TriangularBase& other) +TriangularViewImpl::operator=(const TriangularBase& other) { eigen_assert(Mode == int(OtherDerived::Mode)); if(internal::traits::Flags & EvalBeforeAssigningBit) @@ -824,12 +859,12 @@ TriangularView::operator=(const TriangularBase& } else lazyAssign(other.derived().nestedExpression()); - return *this; + return derived(); } template template -void TriangularView::lazyAssign(const TriangularBase& other) +void TriangularViewImpl::lazyAssign(const TriangularBase& other) { enum { unroll = MatrixType::SizeAtCompileTime != Dynamic @@ -1000,7 +1035,7 @@ template struct evaluator_traits > { typedef typename storage_kind_to_evaluator_kind::Kind Kind; - typedef TriangularShape Shape; + typedef typename glue_shapes::Shape, TriangularShape>::type Shape; // 1 if assignment A = B assumes aliasing when B is of type T and thus B needs to be evaluated into a // temporary; 0 if not. diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index b28f07bdc..a9d8352a9 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -264,22 +264,23 @@ struct general_product_to_triangular_selector #ifdef EIGEN_TEST_EVALUATORS template template -TriangularView& TriangularView::_assignProduct(const ProductType& prod, const Scalar& alpha) +TriangularView& TriangularViewImpl::_assignProduct(const ProductType& prod, const Scalar& alpha) { - eigen_assert(m_matrix.rows() == prod.rows() && m_matrix.cols() == prod.cols()); + eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols()); - general_product_to_triangular_selector::InnerSize==1>::run(m_matrix.const_cast_derived(), prod, alpha); + general_product_to_triangular_selector::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha); return *this; } #else template template -TriangularView& TriangularView::assignProduct(const ProductBase& prod, const Scalar& alpha) +TriangularView& TriangularViewImpl::assignProduct(const ProductBase& prod, const Scalar& alpha) { - eigen_assert(m_matrix.rows() == prod.rows() && m_matrix.cols() == prod.cols()); + eigen_assert(derived().rows() == prod.rows() && derived().cols() == prod.cols()); - general_product_to_triangular_selector::run(m_matrix.const_cast_derived(), prod.derived(), alpha); + general_product_to_triangular_selector + ::run(derived().nestedExpression().const_cast_derived(), prod.derived(), alpha); return *this; } diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index d93277cb8..fda6e2486 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -368,10 +368,12 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix struct traits > : traits, Lhs, Rhs> > {}; +#endif } // end namespace internal diff --git a/Eigen/src/Core/products/TriangularMatrixVector.h b/Eigen/src/Core/products/TriangularMatrixVector.h index 399de3092..19167c232 100644 --- a/Eigen/src/Core/products/TriangularMatrixVector.h +++ b/Eigen/src/Core/products/TriangularMatrixVector.h @@ -157,6 +157,8 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product struct traits > : traits, Lhs, Rhs> > @@ -166,7 +168,7 @@ template struct traits > : traits, Lhs, Rhs> > {}; - +#endif template struct trmv_selector; diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 4b55b3a9a..3ab8d0ed3 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -450,13 +450,13 @@ struct MatrixXpr {}; struct ArrayXpr {}; // An evaluator must define its shape. By default, it can be one of the following: -struct DenseShape { static std::string debugName() { return "DenseShape"; } }; -struct DiagonalShape { static std::string debugName() { return "DiagonalShape"; } }; -struct BandShape { static std::string debugName() { return "BandShape"; } }; -struct TriangularShape { static std::string debugName() { return "TriangularShape"; } }; -struct SelfAdjointShape { static std::string debugName() { return "SelfAdjointShape"; } }; -struct PermutationShape { static std::string debugName() { return "PermutationShape"; } }; -struct SparseShape { static std::string debugName() { return "SparseShape"; } }; +struct DenseShape { static std::string debugName() { return "DenseShape"; } }; +struct DiagonalShape { static std::string debugName() { return "DiagonalShape"; } }; +struct BandShape { static std::string debugName() { return "BandShape"; } }; +struct TriangularShape { static std::string debugName() { return "TriangularShape"; } }; +struct SelfAdjointShape { static std::string debugName() { return "SelfAdjointShape"; } }; +struct PermutationShape { static std::string debugName() { return "PermutationShape"; } }; +struct SparseShape { static std::string debugName() { return "SparseShape"; } }; } // end namespace Eigen diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 70f2c566f..7779fb3fa 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -638,6 +638,9 @@ template struct is_diagonal > template struct is_diagonal > { enum { ret = true }; }; +template struct glue_shapes; +template<> struct glue_shapes { typedef TriangularShape type; }; + } // end namespace internal // we require Lhs and Rhs to have the same scalar type. Currently there is no example of a binary functor diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index f41d7e010..21c7e0b39 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -253,8 +253,8 @@ template struct traits CholMatrixType; - typedef SparseTriangularView MatrixL; - typedef SparseTriangularView MatrixU; + typedef TriangularView MatrixL; + typedef TriangularView MatrixU; static inline MatrixL getL(const MatrixType& m) { return m; } static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); } }; @@ -266,8 +266,8 @@ template struct traits CholMatrixType; - typedef SparseTriangularView MatrixL; - typedef SparseTriangularView MatrixU; + typedef TriangularView MatrixL; + typedef TriangularView MatrixU; static inline MatrixL getL(const MatrixType& m) { return m; } static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); } }; diff --git a/Eigen/src/SparseCore/MappedSparseMatrix.h b/Eigen/src/SparseCore/MappedSparseMatrix.h index ab1a266a9..9205b906f 100644 --- a/Eigen/src/SparseCore/MappedSparseMatrix.h +++ b/Eigen/src/SparseCore/MappedSparseMatrix.h @@ -176,6 +176,34 @@ class MappedSparseMatrix::ReverseInnerIterator const Index m_end; }; +#ifdef EIGEN_ENABLE_EVALUATORS +namespace internal { + +template +struct evaluator > + : evaluator_base > +{ + typedef MappedSparseMatrix<_Scalar,_Options,_Index> MappedSparseMatrixType; + typedef typename MappedSparseMatrixType::InnerIterator InnerIterator; + typedef typename MappedSparseMatrixType::ReverseInnerIterator ReverseInnerIterator; + + enum { + CoeffReadCost = NumTraits<_Scalar>::ReadCost, + Flags = MappedSparseMatrixType::Flags + }; + + evaluator() : m_matrix(0) {} + evaluator(const MappedSparseMatrixType &mat) : m_matrix(&mat) {} + + operator MappedSparseMatrixType&() { return m_matrix->const_cast_derived(); } + operator const MappedSparseMatrixType&() const { return *m_matrix; } + + const MappedSparseMatrixType *m_matrix; +}; + +} +#endif + } // end namespace Eigen #endif // EIGEN_MAPPED_SPARSEMATRIX_H diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index 361a66067..4c5563755 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -333,7 +333,7 @@ template class SparseMatrixBase : public EigenBase Derived& operator*=(const SparseMatrixBase& other); template - inline const SparseTriangularView triangularView() const; + inline const TriangularView triangularView() const; template inline const SparseSelfAdjointView selfadjointView() const; template inline SparseSelfAdjointView selfadjointView(); diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 8bd836883..ff51fc435 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -392,8 +392,6 @@ inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, cons res.row(j) += i.value() * rhs.row(j); } } - -struct SparseSelfAdjointShape { static std::string debugName() { return "SparseSelfAdjointShape"; } }; // TODO currently a selfadjoint expression has the form SelfAdjointView<.,.> // in the future selfadjoint-ness should be defined by the expression traits diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index 333127b78..ad184208b 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -15,15 +15,15 @@ namespace Eigen { namespace internal { -template -struct traits > -: public traits -{}; +// template +// struct traits > +// : public traits +// {}; } // namespace internal -template class SparseTriangularView - : public SparseMatrixBase > +template class TriangularViewImpl + : public SparseMatrixBase > { enum { SkipFirst = ((Mode&Lower) && !(MatrixType::Flags&RowMajorBit)) || ((Mode&Upper) && (MatrixType::Flags&RowMajorBit)), @@ -31,45 +31,51 @@ template class SparseTriangularView SkipDiag = (Mode&ZeroDiag) ? 1 : 0, HasUnitDiag = (Mode&UnitDiag) ? 1 : 0 }; + + typedef TriangularView TriangularViewType; + +protected: + // dummy solve function to make TriangularView happy. + void solve() const; public: - EIGEN_SPARSE_PUBLIC_INTERFACE(SparseTriangularView) - + EIGEN_SPARSE_PUBLIC_INTERFACE(TriangularViewType) + class InnerIterator; class ReverseInnerIterator; - inline Index rows() const { return m_matrix.rows(); } - inline Index cols() const { return m_matrix.cols(); } - typedef typename MatrixType::Nested MatrixTypeNested; typedef typename internal::remove_reference::type MatrixTypeNestedNonRef; typedef typename internal::remove_all::type MatrixTypeNestedCleaned; - inline SparseTriangularView(const MatrixType& matrix) : m_matrix(matrix) {} - - /** \internal */ - inline const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; } - +#ifndef EIGEN_TEST_EVALUATORS template typename internal::plain_matrix_type_column_major::type solve(const MatrixBase& other) const; +#else // EIGEN_TEST_EVALUATORS + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const { + if(!(internal::is_same::value && internal::extract_data(dst) == internal::extract_data(rhs))) + dst = rhs; + this->template solveInPlace(dst); + } +#endif // EIGEN_TEST_EVALUATORS template void solveInPlace(MatrixBase& other) const; template void solveInPlace(SparseMatrixBase& other) const; - - protected: - MatrixTypeNested m_matrix; + }; -template -class SparseTriangularView::InnerIterator : public MatrixTypeNestedCleaned::InnerIterator +template +class TriangularViewImpl::InnerIterator : public MatrixTypeNestedCleaned::InnerIterator { typedef typename MatrixTypeNestedCleaned::InnerIterator Base; - typedef typename SparseTriangularView::Index Index; + typedef typename TriangularViewType::Index Index; public: - EIGEN_STRONG_INLINE InnerIterator(const SparseTriangularView& view, Index outer) + EIGEN_STRONG_INLINE InnerIterator(const TriangularViewImpl& view, Index outer) : Base(view.nestedExpression(), outer), m_returnOne(false) { if(SkipFirst) @@ -132,14 +138,14 @@ class SparseTriangularView::InnerIterator : public MatrixTypeNe bool m_returnOne; }; -template -class SparseTriangularView::ReverseInnerIterator : public MatrixTypeNestedCleaned::ReverseInnerIterator +template +class TriangularViewImpl::ReverseInnerIterator : public MatrixTypeNestedCleaned::ReverseInnerIterator { typedef typename MatrixTypeNestedCleaned::ReverseInnerIterator Base; - typedef typename SparseTriangularView::Index Index; + typedef typename TriangularViewImpl::Index Index; public: - EIGEN_STRONG_INLINE ReverseInnerIterator(const SparseTriangularView& view, Index outer) + EIGEN_STRONG_INLINE ReverseInnerIterator(const TriangularViewType& view, Index outer) : Base(view.nestedExpression(), outer) { eigen_assert((!HasUnitDiag) && "ReverseInnerIterator does not support yet triangular views with a unit diagonal"); @@ -168,7 +174,7 @@ class SparseTriangularView::ReverseInnerIterator : public Matri template template -inline const SparseTriangularView +inline const TriangularView SparseMatrixBase::triangularView() const { return derived(); diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 6e1db0ce8..686be6c92 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -94,7 +94,6 @@ template class Dynamic template class SparseVector; template class MappedSparseMatrix; -template class SparseTriangularView; template class SparseSelfAdjointView; template class SparseDiagonalProduct; template class SparseView; @@ -163,6 +162,12 @@ struct generic_xpr_base typedef SparseMatrixBase type; }; +struct SparseTriangularShape { static std::string debugName() { return "SparseTriangularShape"; } }; +struct SparseSelfAdjointShape { static std::string debugName() { return "SparseSelfAdjointShape"; } }; + +template<> struct glue_shapes { typedef SparseSelfAdjointShape type; }; +template<> struct glue_shapes { typedef SparseTriangularShape type; }; + } // end namespace internal /** \ingroup SparseCore_Module diff --git a/Eigen/src/SparseCore/TriangularSolver.h b/Eigen/src/SparseCore/TriangularSolver.h index dd55522a7..012a1bb75 100644 --- a/Eigen/src/SparseCore/TriangularSolver.h +++ b/Eigen/src/SparseCore/TriangularSolver.h @@ -23,6 +23,7 @@ template::Flags) & RowMajorBit> struct sparse_solve_triangular_selector; +#ifndef EIGEN_TEST_EVALUATORS // forward substitution, row-major template struct sparse_solve_triangular_selector @@ -163,13 +164,166 @@ struct sparse_solve_triangular_selector } }; +#else // EIGEN_TEST_EVALUATORS + +// forward substitution, row-major +template +struct sparse_solve_triangular_selector +{ + typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; + typedef typename evaluator::type LhsEval; + typedef typename evaluator::InnerIterator LhsIterator; + static void run(const Lhs& lhs, Rhs& other) + { + LhsEval lhsEval(lhs); + for(Index col=0 ; col +struct sparse_solve_triangular_selector +{ + typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; + typedef typename evaluator::type LhsEval; + typedef typename evaluator::InnerIterator LhsIterator; + static void run(const Lhs& lhs, Rhs& other) + { + LhsEval lhsEval(lhs); + for(Index col=0 ; col=0 ; --i) + { + Scalar tmp = other.coeff(i,col); + Scalar l_ii = 0; + LhsIterator it(lhsEval, i); + while(it && it.index() +struct sparse_solve_triangular_selector +{ + typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; + typedef typename evaluator::type LhsEval; + typedef typename evaluator::InnerIterator LhsIterator; + static void run(const Lhs& lhs, Rhs& other) + { + LhsEval lhsEval(lhs); + for(Index col=0 ; col +struct sparse_solve_triangular_selector +{ + typedef typename Rhs::Scalar Scalar; + typedef typename Lhs::Index Index; + typedef typename evaluator::type LhsEval; + typedef typename evaluator::InnerIterator LhsIterator; + static void run(const Lhs& lhs, Rhs& other) + { + LhsEval lhsEval(lhs); + for(Index col=0 ; col=0; --i) + { + Scalar& tmp = other.coeffRef(i,col); + if (tmp!=Scalar(0)) // optimization when other is actually sparse + { + if(!(Mode & UnitDiag)) + { + // TODO replace this by a binary search. make sure the binary search is safe for partially sorted elements + LhsIterator it(lhsEval, i); + while(it && it.index()!=i) + ++it; + eigen_assert(it && it.index()==i); + other.coeffRef(i,col) /= it.value(); + } + LhsIterator it(lhsEval, i); + for(; it && it.index() +template template -void SparseTriangularView::solveInPlace(MatrixBase& other) const +void TriangularViewImpl::solveInPlace(MatrixBase& other) const { - eigen_assert(m_matrix.cols() == m_matrix.rows() && m_matrix.cols() == other.rows()); + eigen_assert(derived().cols() == derived().rows() && derived().cols() == other.rows()); eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower))); enum { copy = internal::traits::Flags & RowMajorBit }; @@ -178,21 +332,23 @@ void SparseTriangularView::solveInPlace(MatrixBase::type, OtherDerived&>::type OtherCopy; OtherCopy otherCopy(other.derived()); - internal::sparse_solve_triangular_selector::type, Mode>::run(m_matrix, otherCopy); + internal::sparse_solve_triangular_selector::type, Mode>::run(derived().nestedExpression(), otherCopy); if (copy) other = otherCopy; } -template +#ifndef EIGEN_TEST_EVALUATORS +template template typename internal::plain_matrix_type_column_major::type -SparseTriangularView::solve(const MatrixBase& other) const +TriangularViewImpl::solve(const MatrixBase& other) const { typename internal::plain_matrix_type_column_major::type res(other); solveInPlace(res); return res; } +#endif // pure sparse path @@ -290,11 +446,11 @@ struct sparse_solve_triangular_sparse_selector } // end namespace internal -template +template template -void SparseTriangularView::solveInPlace(SparseMatrixBase& other) const +void TriangularViewImpl::solveInPlace(SparseMatrixBase& other) const { - eigen_assert(m_matrix.cols() == m_matrix.rows() && m_matrix.cols() == other.rows()); + eigen_assert(derived().cols() == derived().rows() && derived().cols() == other.rows()); eigen_assert( (!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower))); // enum { copy = internal::traits::Flags & RowMajorBit }; @@ -303,7 +459,7 @@ void SparseTriangularView::solveInPlace(SparseMatrixBase::type, OtherDerived&>::type OtherCopy; // OtherCopy otherCopy(other.derived()); - internal::sparse_solve_triangular_sparse_selector::run(m_matrix, other.derived()); + internal::sparse_solve_triangular_sparse_selector::run(derived().nestedExpression(), other.derived()); // if (copy) // other = otherCopy; From 4aac87251f16094c01e9c5c8bbf094cd471a2306 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Jul 2014 12:54:03 +0200 Subject: [PATCH 0654/1103] Re-enable a couple of unit tests with evaluators. --- Eigen/src/Core/SelfAdjointView.h | 2 +- Eigen/src/SparseCore/SparseMatrix.h | 12 ++++++++++-- test/CMakeLists.txt | 4 +++- test/main.h | 1 + test/sparse_basic.cpp | 9 +++++---- 5 files changed, 20 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index c8bdec5d4..546f61252 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -35,7 +35,7 @@ struct traits > : traits typedef typename nested::type MatrixTypeNested; typedef typename remove_all::type MatrixTypeNestedCleaned; typedef MatrixType ExpressionType; - typedef typename MatrixType::PlainObject DenseMatrixType; + typedef typename MatrixType::PlainObject FullMatrixType; enum { Mode = UpLo | SelfAdjoint, Flags = MatrixTypeNestedCleaned::Flags & (HereditaryBits) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 5f0c3d0a7..6080c272a 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -52,7 +52,9 @@ struct traits > MaxRowsAtCompileTime = Dynamic, MaxColsAtCompileTime = Dynamic, Flags = _Options | NestByRefBit | LvalueBit, +#ifndef EIGEN_TEST_EVALUATORS CoeffReadCost = NumTraits::ReadCost, +#endif SupportedAccessPatterns = InnerRandomAccessPattern }; }; @@ -74,8 +76,10 @@ struct traits, DiagIndex> ColsAtCompileTime = 1, MaxRowsAtCompileTime = Dynamic, MaxColsAtCompileTime = 1, - Flags = 0, - CoeffReadCost = _MatrixTypeNested::CoeffReadCost*10 + Flags = 0 +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = _MatrixTypeNested::CoeffReadCost*10 +#endif }; }; @@ -1343,6 +1347,8 @@ template struct evaluator > : evaluator_base > { + typedef _Scalar Scalar; + typedef _Index Index; typedef SparseMatrix<_Scalar,_Options,_Index> SparseMatrixType; typedef typename SparseMatrixType::InnerIterator InnerIterator; typedef typename SparseMatrixType::ReverseInnerIterator ReverseInnerIterator; @@ -1358,6 +1364,8 @@ struct evaluator > operator SparseMatrixType&() { return m_matrix->const_cast_derived(); } operator const SparseMatrixType&() const { return *m_matrix; } + Scalar coeff(Index row, Index col) const { return m_matrix->coeff(row,col); } + const SparseMatrixType *m_matrix; }; diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 154e62424..5b5b55c60 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -229,10 +229,12 @@ if(NOT EIGEN_TEST_EVALUATORS) ei_add_test(stdvector_overload) ei_add_test(stdlist) ei_add_test(stddeque) - ei_add_test(sparse_vector) +endif(NOT EIGEN_TEST_EVALUATORS) ei_add_test(sparse_basic) + ei_add_test(sparse_vector) ei_add_test(sparse_product) ei_add_test(sparse_solvers) +if(NOT EIGEN_TEST_EVALUATORS) ei_add_test(sparse_permutations) ei_add_test(simplicial_cholesky) ei_add_test(conjugate_gradient) diff --git a/test/main.h b/test/main.h index a3c157126..57996956d 100644 --- a/test/main.h +++ b/test/main.h @@ -72,6 +72,7 @@ namespace Eigen } #define TRACK std::cerr << __FILE__ << " " << __LINE__ << std::endl +// #define TRACK while() #define EI_PP_MAKE_STRING2(S) #S #define EI_PP_MAKE_STRING(S) EI_PP_MAKE_STRING2(S) diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index 4c9b9111e..c86534bad 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -201,9 +201,9 @@ template void sparse_basic(const SparseMatrixType& re VERIFY(m3.innerVector(j0).nonZeros() == m3.transpose().innerVector(j0).nonZeros()); - //m2.innerVector(j0) = 2*m2.innerVector(j1); - //refMat2.col(j0) = 2*refMat2.col(j1); - //VERIFY_IS_APPROX(m2, refMat2); +// m2.innerVector(j0) = 2*m2.innerVector(j1); +// refMat2.col(j0) = 2*refMat2.col(j1); +// VERIFY_IS_APPROX(m2, refMat2); } // test innerVectors() @@ -239,7 +239,7 @@ template void sparse_basic(const SparseMatrixType& re VERIFY_IS_APPROX(m2, refMat2); } - + // test basic computations { DenseMatrix refM1 = DenseMatrix::Zero(rows, rows); @@ -255,6 +255,7 @@ template void sparse_basic(const SparseMatrixType& re initSparse(density, refM3, m3); initSparse(density, refM4, m4); + VERIFY_IS_APPROX(m1*s1, refM1*s1); VERIFY_IS_APPROX(m1+m2, refM1+refM2); VERIFY_IS_APPROX(m1+m2+m3, refM1+refM2+refM3); VERIFY_IS_APPROX(m3.cwiseProduct(m1+m2), refM3.cwiseProduct(refM1+refM2)); From a8283e0ed2d3e42ba8e1c42f51cb08eacc301047 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 22 Jul 2014 13:16:44 +0200 Subject: [PATCH 0655/1103] Define EIGEN_TRY, EIGEN_CATCH, EIGEN_THROW as suggested by Moritz Klammer. Make it possible to run unit-tests with exceptions disabled via EIGEN_TEST_NO_EXCEPTIONS flag. Enhanced ctorleak unit-test --- CMakeLists.txt | 8 +- Eigen/Core | 16 ++-- Eigen/src/Core/util/Macros.h | 12 +++ Eigen/src/Core/util/Memory.h | 145 ++++++++++++++--------------------- test/CMakeLists.txt | 10 ++- test/ctorleak.cpp | 42 +++------- test/main.h | 16 +++- 7 files changed, 113 insertions(+), 136 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 470095680..96d6c8701 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -119,7 +119,7 @@ endmacro(ei_add_cxx_compiler_flag) if(NOT MSVC) # We assume that other compilers are partly compatible with GNUCC - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fexceptions") +# set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fexceptions") set(CMAKE_CXX_FLAGS_DEBUG "-g3") set(CMAKE_CXX_FLAGS_RELEASE "-g0 -O2") @@ -301,6 +301,12 @@ if(EIGEN_TEST_NO_EXPLICIT_ALIGNMENT) message(STATUS "Disabling alignment in tests/examples") endif() +option(EIGEN_TEST_NO_EXCEPTIONS "Disables C++ exceptions" OFF) +if(EIGEN_TEST_NO_EXCEPTIONS) + ei_add_cxx_compiler_flag("-fno-exceptions") + message(STATUS "Disabling exceptions in tests/examples") +endif() + option(EIGEN_TEST_C++0x "Enables all C++0x features." OFF) include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR}) diff --git a/Eigen/Core b/Eigen/Core index 16e388fa4..9a73fe37b 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -42,6 +42,14 @@ #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC; #endif +#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) + #define EIGEN_EXCEPTIONS +#endif + +#ifdef EIGEN_EXCEPTIONS + #include +#endif + // then include this file where all our macros are defined. It's really important to do it first because // it's where we do all the alignment settings (platform detection and honoring the user's will if he // defined e.g. EIGEN_DONT_ALIGN) so it needs to be done before we do anything with vectorization. @@ -209,14 +217,6 @@ #include #endif -#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) - #define EIGEN_EXCEPTIONS -#endif - -#ifdef EIGEN_EXCEPTIONS - #include -#endif - /** \brief Namespace containing all symbols from the %Eigen library. */ namespace Eigen { diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 71e42b125..5e9b0a112 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -452,4 +452,16 @@ namespace Eigen { const RHS \ > +#ifdef EIGEN_EXCEPTIONS +# define EIGEN_THROW_X(X) throw X +# define EIGEN_THROW throw +# define EIGEN_TRY try +# define EIGEN_CATCH(X) catch (X) +#else +# define EIGEN_THROW_X(X) std::abort() +# define EIGEN_THROW std::abort() +# define EIGEN_TRY if (true) +# define EIGEN_CATCH(X) else +#endif + #endif // EIGEN_MACROS_H diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 9b3e84fb0..38990566d 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -354,20 +354,16 @@ template inline void destruct_elements_of_array(T *ptr, size_t size) template inline T* construct_elements_of_array(T *ptr, size_t size) { size_t i; -#ifdef EIGEN_EXCEPTIONS - try -#endif - { + EIGEN_TRY + { for (i = 0; i < size; ++i) ::new (ptr + i) T; return ptr; - } -#ifdef EIGEN_EXCEPTIONS - catch (...) - { - destruct_elements_of_array(ptr, i); - throw; - } -#endif + } + EIGEN_CATCH(...) + { + destruct_elements_of_array(ptr, i); + EIGEN_THROW; + } } /***************************************************************************** @@ -389,38 +385,30 @@ template inline T* aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(aligned_malloc(sizeof(T)*size)); -#ifdef EIGEN_EXCEPTIONS - try -#endif - { - return construct_elements_of_array(result, size); - } -#ifdef EIGEN_EXCEPTIONS - catch (...) - { - aligned_free(result); - throw; - } -#endif + EIGEN_TRY + { + return construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + aligned_free(result); + EIGEN_THROW; + } } template inline T* conditional_aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); -#ifdef EIGEN_EXCEPTIONS - try -#endif - { - return construct_elements_of_array(result, size); - } -#ifdef EIGEN_EXCEPTIONS - catch (...) - { - conditional_aligned_free(result); - throw; - } -#endif + EIGEN_TRY + { + return construct_elements_of_array(result, size); + } + EIGEN_CATCH(...) + { + conditional_aligned_free(result); + EIGEN_THROW; + } } /** \internal Deletes objects constructed with aligned_new @@ -449,21 +437,17 @@ template inline T* conditional_aligned_realloc_new(T* pt destruct_elements_of_array(pts+new_size, old_size-new_size); T *result = reinterpret_cast(conditional_aligned_realloc(reinterpret_cast(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(new_size > old_size) + { + EIGEN_TRY { -#ifdef EIGEN_EXCEPTIONS - try -#endif - { - construct_elements_of_array(result+old_size, new_size-old_size); - } -#ifdef EIGEN_EXCEPTIONS - catch (...) - { - conditional_aligned_free(result); - throw; - } -#endif + construct_elements_of_array(result+old_size, new_size-old_size); } + EIGEN_CATCH(...) + { + conditional_aligned_free(result); + EIGEN_THROW; + } + } return result; } @@ -473,21 +457,17 @@ template inline T* conditional_aligned_new_auto(size_t s check_size_for_overflow(size); T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); if(NumTraits::RequireInitialization) + { + EIGEN_TRY { -#ifdef EIGEN_EXCEPTIONS - try -#endif - { - construct_elements_of_array(result, size); - } -#ifdef EIGEN_EXCEPTIONS - catch (...) - { - conditional_aligned_free(result); - throw; - } -#endif + construct_elements_of_array(result, size); } + EIGEN_CATCH(...) + { + conditional_aligned_free(result); + EIGEN_THROW; + } + } return result; } @@ -499,21 +479,17 @@ template inline T* conditional_aligned_realloc_new_auto( destruct_elements_of_array(pts+new_size, old_size-new_size); T *result = reinterpret_cast(conditional_aligned_realloc(reinterpret_cast(pts), sizeof(T)*new_size, sizeof(T)*old_size)); if(NumTraits::RequireInitialization && (new_size > old_size)) + { + EIGEN_TRY { -#ifdef EIGEN_EXCEPTIONS - try -#endif - { - construct_elements_of_array(result+old_size, new_size-old_size); - } -#ifdef EIGEN_EXCEPTIONS - catch (...) - { - conditional_aligned_free(result); - throw; - } -#endif + construct_elements_of_array(result+old_size, new_size-old_size); } + EIGEN_CATCH(...) + { + conditional_aligned_free(result); + EIGEN_THROW; + } + } return result; } @@ -713,20 +689,11 @@ template class aligned_stack_memory_handler *****************************************************************************/ #if EIGEN_ALIGN - #ifdef EIGEN_EXCEPTIONS - #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ + #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ void* operator new(size_t size, const std::nothrow_t&) throw() { \ - try { return Eigen::internal::conditional_aligned_malloc(size); } \ - catch (...) { return 0; } \ - return 0; \ + EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc(size); } \ + EIGEN_CATCH (...) { return 0; } \ } - #else - #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \ - void* operator new(size_t size, const std::nothrow_t&) throw() { \ - return Eigen::internal::conditional_aligned_malloc(size); \ - } - #endif - #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \ void *operator new(size_t size) { \ return Eigen::internal::conditional_aligned_malloc(size); \ diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index fed5c0e06..47aefddb8 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -158,7 +158,9 @@ ei_add_test(basicstuff) ei_add_test(linearstructure) ei_add_test(integer_types) ei_add_test(unalignedcount) -ei_add_test(exceptions) +if(NOT EIGEN_TEST_NO_EXCEPTIONS) + ei_add_test(exceptions) +endif() ei_add_test(redux) ei_add_test(visitor) ei_add_test(block) @@ -238,7 +240,9 @@ ei_add_test(nesting_ops "${CMAKE_CXX_FLAGS_DEBUG}") ei_add_test(zerosized) ei_add_test(dontalign) ei_add_test(evaluators) -ei_add_test(sizeoverflow) +if(NOT EIGEN_TEST_NO_EXCEPTIONS) + ei_add_test(sizeoverflow) +endif() ei_add_test(prec_inverse_4x4) ei_add_test(vectorwiseop) ei_add_test(special_numbers) @@ -251,8 +255,6 @@ ei_add_test(bicgstab) ei_add_test(sparselu) ei_add_test(sparseqr) -ei_add_test(ctorleak) - # ei_add_test(denseLM) if(QT4_FOUND) diff --git a/test/ctorleak.cpp b/test/ctorleak.cpp index f3f4411c8..145d91be4 100644 --- a/test/ctorleak.cpp +++ b/test/ctorleak.cpp @@ -28,42 +28,24 @@ struct Foo unsigned Foo::object_count = 0; unsigned Foo::object_limit = 0; -namespace Eigen -{ - template<> - struct NumTraits - { - typedef double Real; - typedef double NonInteger; - typedef double Nested; - enum - { - IsComplex = 0, - IsInteger = 1, - ReadCost = -1, - AddCost = -1, - MulCost = -1, - IsSigned = 1, - RequireInitialization = 1 - }; - static inline Real epsilon() { return 1.0; } - static inline Real dummy_epsilon() { return 0.0; } - }; -} void test_ctorleak() { + typedef DenseIndex Index; Foo::object_count = 0; - Foo::object_limit = internal::random(0, 14 * 92 - 2); + for(int i = 0; i < g_repeat; i++) { + Index rows = internal::random(2,EIGEN_TEST_MAX_SIZE), cols = internal::random(2,EIGEN_TEST_MAX_SIZE); + Foo::object_limit = internal::random(0, rows*cols - 2); #ifdef EIGEN_EXCEPTIONS - try -#endif + try { - Matrix m(14, 92); - eigen_assert(false); // not reached - } -#ifdef EIGEN_EXCEPTIONS - catch (const Foo::Fail&) { /* ignore */ } #endif + Matrix m(rows, cols); +#ifdef EIGEN_EXCEPTIONS + VERIFY(false); // not reached if exceptions are enabled + } + catch (const Foo::Fail&) { /* ignore */ } +#endif + } VERIFY_IS_EQUAL(static_cast(0), Foo::object_count); } diff --git a/test/main.h b/test/main.h index 3ccc2ae88..3295dcb71 100644 --- a/test/main.h +++ b/test/main.h @@ -117,13 +117,14 @@ namespace Eigen if(report_on_cerr_on_assert_failure) \ std::cerr << #a << " " __FILE__ << "(" << __LINE__ << ")\n"; \ Eigen::no_more_assert = true; \ - throw Eigen::eigen_assert_exception(); \ + EIGEN_THROW_X(Eigen::eigen_assert_exception()); \ } \ else if (Eigen::internal::push_assert) \ { \ eigen_assert_list.push_back(std::string(EI_PP_MAKE_STRING(__FILE__) " (" EI_PP_MAKE_STRING(__LINE__) ") : " #a) ); \ } + #ifdef EIGEN_EXCEPTIONS #define VERIFY_RAISES_ASSERT(a) \ { \ Eigen::no_more_assert = false; \ @@ -142,6 +143,7 @@ namespace Eigen Eigen::report_on_cerr_on_assert_failure = true; \ Eigen::internal::push_assert = false; \ } + #endif //EIGEN_EXCEPTIONS #elif !defined(__CUDACC__) // EIGEN_DEBUG_ASSERTS // see bug 89. The copy_bool here is working around a bug in gcc <= 4.3 @@ -152,9 +154,10 @@ namespace Eigen if(report_on_cerr_on_assert_failure) \ eigen_plain_assert(a); \ else \ - throw Eigen::eigen_assert_exception(); \ + EIGEN_THROW_X(Eigen::eigen_assert_exception()); \ } - #define VERIFY_RAISES_ASSERT(a) { \ + #ifdef EIGEN_EXCEPTIONS + #define VERIFY_RAISES_ASSERT(a) { \ Eigen::no_more_assert = false; \ Eigen::report_on_cerr_on_assert_failure = false; \ try { \ @@ -164,9 +167,14 @@ namespace Eigen catch (Eigen::eigen_assert_exception&) { VERIFY(true); } \ Eigen::report_on_cerr_on_assert_failure = true; \ } - + #endif //EIGEN_EXCEPTIONS #endif // EIGEN_DEBUG_ASSERTS +#ifndef VERIFY_RAISES_ASSERT + #define VERIFY_RAISES_ASSERT(a) \ + std::cout << "Can't VERIFY_RAISES_ASSERT( " #a " ) with exceptions disabled"; +#endif + #if !defined(__CUDACC__) #define EIGEN_USE_CUSTOM_ASSERT #endif From baa77ffe3878ec80f1b136856dbc31d21eac2a1e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Jul 2014 16:13:56 +0200 Subject: [PATCH 0656/1103] Fix max sizes at compile time of DiagonalWrapper --- Eigen/src/Core/DiagonalMatrix.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index 5f7e5fe33..801131b54 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -270,8 +270,8 @@ struct traits > enum { RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, - MaxRowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, - MaxColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime, + MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, + MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, Flags = (traits::Flags & LvalueBit) | NoPreferredStorageOrderBit #ifndef EIGEN_TEST_EVALUATORS , From 922694a2d1531b6c50fc26a96b7d998089da72bf Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Jul 2014 16:57:14 +0200 Subject: [PATCH 0657/1103] Extend fixed-size ctor unit test and fix conversion warning. --- Eigen/src/Core/PlainObjectBase.h | 10 ++++++++++ test/basicstuff.cpp | 28 +++++++++++++++++++++++----- 2 files changed, 33 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index e4e3e8903..b968b325e 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -681,6 +681,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) resize(nbRows,nbCols); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init2(const Scalar& val0, const Scalar& val1, typename internal::enable_if::type* = 0) @@ -689,6 +690,15 @@ class PlainObjectBase : public internal::dense_xpr_base::type m_storage.data()[0] = val0; m_storage.data()[1] = val1; } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init2(const Index& val0, const Index& val1, typename internal::enable_if<(!internal::is_same::value) && Base::SizeAtCompileTime==2,T1>::type* = 0) + { + EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2) + m_storage.data()[0] = Scalar(val0); + m_storage.data()[1] = Scalar(val1); + } template EIGEN_DEVICE_FUNC diff --git a/test/basicstuff.cpp b/test/basicstuff.cpp index 56370d591..70cefe19f 100644 --- a/test/basicstuff.cpp +++ b/test/basicstuff.cpp @@ -183,6 +183,7 @@ void fixedSizeMatrixConstruction() Scalar raw[4]; for(int k=0; k<4; ++k) raw[k] = internal::random(); + { Matrix m(raw); Array a(raw); @@ -200,18 +201,34 @@ void fixedSizeMatrixConstruction() VERIFY((a==Array(raw[0],raw[1],raw[2])).all()); } { - Matrix m(raw); - Array a(raw); + Matrix m(raw), m2( (DenseIndex(raw[0])), (DenseIndex(raw[1])) ); + Array a(raw), a2( (DenseIndex(raw[0])), (DenseIndex(raw[1])) ); for(int k=0; k<2; ++k) VERIFY(m(k) == raw[k]); for(int k=0; k<2; ++k) VERIFY(a(k) == raw[k]); VERIFY_IS_EQUAL(m,(Matrix(raw[0],raw[1]))); VERIFY((a==Array(raw[0],raw[1])).all()); + for(int k=0; k<2; ++k) VERIFY(m2(k) == DenseIndex(raw[k])); + for(int k=0; k<2; ++k) VERIFY(a2(k) == DenseIndex(raw[k])); } { - Matrix m(raw); - Array a(raw); + Matrix m(raw), m2( (DenseIndex(raw[0])), (DenseIndex(raw[1])) ); + Array a(raw), a2( (DenseIndex(raw[0])), (DenseIndex(raw[1])) ); + for(int k=0; k<2; ++k) VERIFY(m(k) == raw[k]); + for(int k=0; k<2; ++k) VERIFY(a(k) == raw[k]); + VERIFY_IS_EQUAL(m,(Matrix(raw[0],raw[1]))); + VERIFY((a==Array(raw[0],raw[1])).all()); + for(int k=0; k<2; ++k) VERIFY(m2(k) == DenseIndex(raw[k])); + for(int k=0; k<2; ++k) VERIFY(a2(k) == DenseIndex(raw[k])); + } + { + Matrix m(raw), m1(raw[0]), m2( (DenseIndex(raw[0])) ); + Array a(raw), a1(raw[0]), a2( (DenseIndex(raw[0])) ); VERIFY(m(0) == raw[0]); VERIFY(a(0) == raw[0]); + VERIFY(m1(0) == raw[0]); + VERIFY(a1(0) == raw[0]); + VERIFY(m2(0) == DenseIndex(raw[0])); + VERIFY(a2(0) == DenseIndex(raw[0])); VERIFY_IS_EQUAL(m,(Matrix(raw[0]))); VERIFY((a==Array(raw[0])).all()); } @@ -233,9 +250,10 @@ void test_basicstuff() } CALL_SUBTEST_1(fixedSizeMatrixConstruction()); - CALL_SUBTEST_1(fixedSizeMatrixConstruction()); + CALL_SUBTEST_1(fixedSizeMatrixConstruction()); CALL_SUBTEST_1(fixedSizeMatrixConstruction()); CALL_SUBTEST_1(fixedSizeMatrixConstruction()); + CALL_SUBTEST_1(fixedSizeMatrixConstruction()); CALL_SUBTEST_1(fixedSizeMatrixConstruction()); CALL_SUBTEST_2(casting()); From 7f15f27a9e239bd386541b9ff33ace35fb2a6994 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Jul 2014 17:01:34 +0200 Subject: [PATCH 0658/1103] Workaround ambiguous call of init1 with MSVC. --- Eigen/src/Core/PlainObjectBase.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index b968b325e..82d96eb92 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -721,7 +721,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type EIGEN_STRONG_INLINE void _init1(const Index& val0, typename internal::enable_if< (!internal::is_same::value) && Base::SizeAtCompileTime==1 - && internal::is_convertible::value,T>::type* = 0) + && internal::is_convertible::value,T*>::type* = 0) { EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) m_storage.data()[0] = Scalar(val0); From d1e9f39a9aac85eeec2e17f00aef24cb35537be3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 22 Jul 2014 18:28:19 +0200 Subject: [PATCH 0659/1103] Ambiguous call fixes for clang. --- Eigen/src/Core/PlainObjectBase.h | 11 ++++++++--- test/basicstuff.cpp | 10 ++++++++-- 2 files changed, 16 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 82d96eb92..f8292c793 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -693,7 +693,11 @@ class PlainObjectBase : public internal::dense_xpr_base::type template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void _init2(const Index& val0, const Index& val1, typename internal::enable_if<(!internal::is_same::value) && Base::SizeAtCompileTime==2,T1>::type* = 0) + EIGEN_STRONG_INLINE void _init2(const Index& val0, const Index& val1, + typename internal::enable_if< (!internal::is_same::value) + && (internal::is_same::value) + && (internal::is_same::value) + && Base::SizeAtCompileTime==2,T1>::type* = 0) { EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2) m_storage.data()[0] = Scalar(val0); @@ -719,8 +723,9 @@ class PlainObjectBase : public internal::dense_xpr_base::type template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init1(const Index& val0, - typename internal::enable_if< (!internal::is_same::value) - && Base::SizeAtCompileTime==1 + typename internal::enable_if< (!internal::is_same::value) + && (internal::is_same::value) + && Base::SizeAtCompileTime==1 && internal::is_convertible::value,T*>::type* = 0) { EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1) diff --git a/test/basicstuff.cpp b/test/basicstuff.cpp index 70cefe19f..3d7d74d4e 100644 --- a/test/basicstuff.cpp +++ b/test/basicstuff.cpp @@ -211,7 +211,10 @@ void fixedSizeMatrixConstruction() for(int k=0; k<2; ++k) VERIFY(a2(k) == DenseIndex(raw[k])); } { - Matrix m(raw), m2( (DenseIndex(raw[0])), (DenseIndex(raw[1])) ); + Matrix m(raw), + m2( (DenseIndex(raw[0])), (DenseIndex(raw[1])) ), + m3( (int(raw[0])), (int(raw[1])) ), + m4( (float(raw[0])), (float(raw[1])) ); Array a(raw), a2( (DenseIndex(raw[0])), (DenseIndex(raw[1])) ); for(int k=0; k<2; ++k) VERIFY(m(k) == raw[k]); for(int k=0; k<2; ++k) VERIFY(a(k) == raw[k]); @@ -219,9 +222,11 @@ void fixedSizeMatrixConstruction() VERIFY((a==Array(raw[0],raw[1])).all()); for(int k=0; k<2; ++k) VERIFY(m2(k) == DenseIndex(raw[k])); for(int k=0; k<2; ++k) VERIFY(a2(k) == DenseIndex(raw[k])); + for(int k=0; k<2; ++k) VERIFY(m3(k) == int(raw[k])); + for(int k=0; k<2; ++k) VERIFY(m4(k) == float(raw[k])); } { - Matrix m(raw), m1(raw[0]), m2( (DenseIndex(raw[0])) ); + Matrix m(raw), m1(raw[0]), m2( (DenseIndex(raw[0])) ), m3( (int(raw[0])) ); Array a(raw), a1(raw[0]), a2( (DenseIndex(raw[0])) ); VERIFY(m(0) == raw[0]); VERIFY(a(0) == raw[0]); @@ -229,6 +234,7 @@ void fixedSizeMatrixConstruction() VERIFY(a1(0) == raw[0]); VERIFY(m2(0) == DenseIndex(raw[0])); VERIFY(a2(0) == DenseIndex(raw[0])); + VERIFY(m3(0) == int(raw[0])); VERIFY_IS_EQUAL(m,(Matrix(raw[0]))); VERIFY((a==Array(raw[0])).all()); } From f7bb7ee3f36474163da7c7f6f88306d553238df2 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 22 Jul 2014 10:31:21 -0700 Subject: [PATCH 0660/1103] Fixed the assignment operator of the Tensor and TensorMap classes. --- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 11 +++++++++++ unsupported/Eigen/CXX11/src/Tensor/TensorMap.h | 9 ++++++++- 2 files changed, 19 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index 09601fc7d..547bb74d1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -229,6 +229,17 @@ class Tensor : public TensorBase > EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Tensor& operator=(const Tensor& other) + { + // FIXME: we need to resize the tensor to fix the dimensions of the other. + // Unfortunately this isn't possible yet when the rhs is an expression. + // resize(other.dimensions()); + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor& operator=(const OtherDerived& other) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index c97135b63..417717b90 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -241,9 +241,16 @@ template class TensorMap : public Tensor } #endif + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Self& operator=(const Self& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } template - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Self& operator=(const OtherDerived& other) { typedef TensorAssignOp Assign; From 1f371e78e659d6e5fd781aea93b6b9c7a0604aeb Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 22 Jul 2014 10:32:40 -0700 Subject: [PATCH 0661/1103] Added a few tests to validate the behavior of the assignment operator. --- unsupported/test/cxx11_tensor_assign.cpp | 43 ++++++++++++++++++++++++ unsupported/test/cxx11_tensor_simple.cpp | 2 +- 2 files changed, 44 insertions(+), 1 deletion(-) diff --git a/unsupported/test/cxx11_tensor_assign.cpp b/unsupported/test/cxx11_tensor_assign.cpp index c88872950..b024bed19 100644 --- a/unsupported/test/cxx11_tensor_assign.cpp +++ b/unsupported/test/cxx11_tensor_assign.cpp @@ -186,10 +186,53 @@ static void test_3d() } } +static void test_same_type() +{ + Tensor orig_tensor(5); + Tensor dest_tensor(5); + orig_tensor.setRandom(); + dest_tensor.setRandom(); + int* orig_data = orig_tensor.data(); + int* dest_data = dest_tensor.data(); + dest_tensor = orig_tensor; + VERIFY_IS_EQUAL(orig_tensor.data(), orig_data); + VERIFY_IS_EQUAL(dest_tensor.data(), dest_data); + for (int i = 0; i < 5; ++i) { + VERIFY_IS_EQUAL(dest_tensor(i), orig_tensor(i)); + } + + TensorFixedSize > orig_array; + TensorFixedSize > dest_array; + orig_array.setRandom(); + dest_array.setRandom(); + orig_data = orig_array.data(); + dest_data = dest_array.data(); + dest_array = orig_array; + VERIFY_IS_EQUAL(orig_array.data(), orig_data); + VERIFY_IS_EQUAL(dest_array.data(), dest_data); + for (int i = 0; i < 5; ++i) { + VERIFY_IS_EQUAL(dest_array(i), orig_array(i)); + } + + int orig[5] = {1, 2, 3, 4, 5}; + int dest[5] = {6, 7, 8, 9, 10}; + TensorMap > orig_map(orig, 5); + TensorMap > dest_map(dest, 5); + orig_data = orig_map.data(); + dest_data = dest_map.data(); + dest_map = orig_map; + VERIFY_IS_EQUAL(orig_map.data(), orig_data); + VERIFY_IS_EQUAL(dest_map.data(), dest_data); + for (int i = 0; i < 5; ++i) { + VERIFY_IS_EQUAL(dest[i], i+1); + } +} + void test_cxx11_tensor_assign() { CALL_SUBTEST(test_1d()); CALL_SUBTEST(test_2d()); CALL_SUBTEST(test_3d()); + CALL_SUBTEST(test_same_type()); } diff --git a/unsupported/test/cxx11_tensor_simple.cpp b/unsupported/test/cxx11_tensor_simple.cpp index 1f76033ea..1455f2a4c 100644 --- a/unsupported/test/cxx11_tensor_simple.cpp +++ b/unsupported/test/cxx11_tensor_simple.cpp @@ -244,7 +244,7 @@ static void test_simple_assign() epsilon(0,1,2) = epsilon(2,0,1) = epsilon(1,2,0) = 1; epsilon(2,1,0) = epsilon(0,2,1) = epsilon(1,0,2) = -1; - Tensor e2(2,3,1); + Tensor e2(3,3,3); e2.setZero(); VERIFY_IS_EQUAL((e2(1,2,0)), 0); From 2c625ec9ba64c6c09217190b494c9b08efb499e5 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Tue, 22 Jul 2014 20:46:03 +0000 Subject: [PATCH 0662/1103] Simplification of some Altivec constants, reuse existing constants and avoid loading from RAM esp in the case of p16uc_COMPLEX_TRANSPOSE* --- Eigen/src/Core/arch/AltiVec/Complex.h | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 73fa62643..13b874d0c 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -16,13 +16,14 @@ namespace internal { static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; static Packet16uc p16uc_COMPLEX_RE = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; -static Packet16uc p16uc_COMPLEX_IM = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +static Packet16uc p16uc_COMPLEX_IM = vec_sld(p16uc_DUPLICATE, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; static Packet16uc p16uc_COMPLEX_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);//{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; -static Packet16uc p16uc_PSET_HI = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 1));//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; -static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergeh((Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 2), (Packet4ui) vec_splat((Packet4ui)p16uc_FORWARD, 3));//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; -static Packet16uc p16uc_COMPLEX_TRANSPOSE_0 = { 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; -static Packet16uc p16uc_COMPLEX_TRANSPOSE_1 = { 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; +static Packet16uc p16uc_PSET_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; +static Packet16uc p16uc_COMPLEX_MASK16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);//{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; +static Packet16uc p16uc_COMPLEX_TRANSPOSE_0 = vec_add(p16uc_PSET_HI, p16uc_COMPLEX_MASK16);//{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; +static Packet16uc p16uc_COMPLEX_TRANSPOSE_1 = vec_add(p16uc_PSET_LO, p16uc_COMPLEX_MASK16);//{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; //---------- float ---------- struct Packet2cf From a0a87410d0cd76e930279f1a7c81c6da90148e59 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 24 Jul 2014 22:08:10 +0200 Subject: [PATCH 0663/1103] Fix bug #61: gemm was broken since we changed the blocking order --- blas/level3_impl.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/blas/level3_impl.h b/blas/level3_impl.h index 07dbc22ff..a05872666 100644 --- a/blas/level3_impl.h +++ b/blas/level3_impl.h @@ -56,7 +56,7 @@ int EIGEN_BLAS_FUNC(gemm)(char *opa, char *opb, int *m, int *n, int *k, RealScal else matrix(c, *m, *n, *ldc) *= beta; } - internal::gemm_blocking_space blocking(*m,*n,*k); + internal::gemm_blocking_space blocking(*m,*n,*k,true); int code = OP(*opa) | (OP(*opb) << 2); func[code](*m, *n, *k, a, *lda, b, *ldb, c, *ldc, alpha, blocking, 0); From 5f3d542b8a526d2fb8a179eb68f4a6d35f4f85a3 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Fri, 25 Jul 2014 13:34:03 +0100 Subject: [PATCH 0664/1103] Fix typo in MatrixExponential noticed by Markos. --- unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 05df5a102..160120d03 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -176,8 +176,8 @@ void matrix_exp_pade17(const MatrixType &A, MatrixType &U, MatrixType &V) const MatrixType A4 = A2 * A2; const MatrixType A6 = A4 * A2; const MatrixType A8 = A4 * A4; - V = b[17] * m_tmp1 + b[15] * A6 + b[13] * A4 + b[11] * A2; // used for temporary storage - matrixType tmp = A8 * V; + V = b[17] * A8 + b[15] * A6 + b[13] * A4 + b[11] * A2; // used for temporary storage + MatrixType tmp = A8 * V; tmp += b[9] * A8 + b[7] * A6 + b[5] * A4 + b[3] * A2 + b[1] * MatrixType::Identity(A.rows(), A.cols()); U.noalias() = A * tmp; From 2116e261fb27c795d153f171467cf7912ff3eec5 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 25 Jul 2014 09:47:59 -0700 Subject: [PATCH 0665/1103] Made sure that the data stored in fixed sized tensor is aligned. --- unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index c9d6517eb..0c4f8a3d6 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -38,7 +38,7 @@ template Date: Wed, 30 Jul 2014 09:32:35 +0200 Subject: [PATCH 0666/1103] add missing delete operator overloads --- Eigen/src/Core/util/Memory.h | 2 ++ test/dynalloc.cpp | 31 +++++++++++++++++++++++++++++++ 2 files changed, 33 insertions(+) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 38990566d..810ee786b 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -703,6 +703,8 @@ template class aligned_stack_memory_handler } \ void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free(ptr); } \ void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free(ptr); } \ + void operator delete(void * ptr, std::size_t /* sz */) throw() { Eigen::internal::conditional_aligned_free(ptr); } \ + void operator delete[](void * ptr, std::size_t /* sz */) throw() { Eigen::internal::conditional_aligned_free(ptr); } \ /* in-place new and delete. since (at least afaik) there is no actual */ \ /* memory allocated we can safely let the default implementation handle */ \ /* this particular case. */ \ diff --git a/test/dynalloc.cpp b/test/dynalloc.cpp index c98cc80f0..4f53ea6f0 100644 --- a/test/dynalloc.cpp +++ b/test/dynalloc.cpp @@ -93,6 +93,32 @@ template void check_dynaligned() } } +template void check_custom_new_delete() +{ + { + T* t = new T; + delete t; + } + + { + std::size_t N = internal::random(1,10); + T* t = new T[N]; + delete[] t; + } + +#ifdef EIGEN_ALIGN + { + T* t = static_cast((T::operator new)(sizeof(T))); + (T::operator delete)(t, sizeof(T)); + } + + { + T* t = static_cast((T::operator new)(sizeof(T))); + (T::operator delete)(t); + } +#endif +} + void test_dynalloc() { // low level dynamic memory allocation @@ -109,6 +135,11 @@ void test_dynalloc() CALL_SUBTEST(check_dynaligned() ); CALL_SUBTEST(check_dynaligned() ); CALL_SUBTEST(check_dynaligned() ); + + CALL_SUBTEST( check_custom_new_delete() ); + CALL_SUBTEST( check_custom_new_delete() ); + CALL_SUBTEST( check_custom_new_delete() ); + CALL_SUBTEST( check_custom_new_delete() ); } // check static allocation, who knows ? From 929e77192cb8c0839238ff95e7eb1b310c627176 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 30 Jul 2014 11:39:52 +0200 Subject: [PATCH 0667/1103] Various minor fixes --- Eigen/src/Core/ProductEvaluators.h | 4 +- Eigen/src/Core/TriangularMatrix.h | 32 ++--- .../products/GeneralMatrixMatrixTriangular.h | 4 +- Eigen/src/SparseCore/SparseAssign.h | 1 + Eigen/src/SparseCore/SparseSelfAdjointView.h | 2 +- Eigen/src/SparseCore/SparseTriangularView.h | 124 ++++++++++++++++-- Eigen/src/SparseCore/SparseUtil.h | 4 +- test/CMakeLists.txt | 1 - 8 files changed, 137 insertions(+), 35 deletions(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index dc91ddd4d..b04df00e7 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -602,7 +602,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - triangular_product_impl + triangular_product_impl ::run(dst, lhs.nestedExpression(), rhs, alpha); } }; @@ -636,7 +636,7 @@ struct generic_product_impl template static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha) { - triangular_product_impl::run(dst, lhs, rhs.nestedExpression(), alpha); + triangular_product_impl::run(dst, lhs, rhs.nestedExpression(), alpha); } }; diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 25dab0e77..0383ca9f5 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -176,7 +176,7 @@ struct traits > : traits typedef MatrixType ExpressionType; enum { Mode = _Mode, - Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) | Mode + Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) #ifndef EIGEN_TEST_EVALUATORS , CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost @@ -206,7 +206,6 @@ template class TriangularView protected: typedef typename internal::traits::MatrixTypeNested MatrixTypeNested; typedef typename internal::traits::MatrixTypeNestedNonRef MatrixTypeNestedNonRef; - typedef typename internal::traits::MatrixTypeNestedCleaned MatrixTypeNestedCleaned; typedef typename internal::remove_all::type MatrixConjugateReturnType; @@ -214,6 +213,7 @@ template class TriangularView typedef typename internal::traits::StorageKind StorageKind; typedef typename internal::traits::Index Index; + typedef typename internal::traits::MatrixTypeNestedCleaned NestedExpression; enum { Mode = _Mode, @@ -229,6 +229,8 @@ template class TriangularView {} using Base::operator=; + TriangularView& operator=(const TriangularView &other) + { return Base::operator=(other); } EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } @@ -236,9 +238,9 @@ template class TriangularView inline Index cols() const { return m_matrix.cols(); } EIGEN_DEVICE_FUNC - const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; } + const NestedExpression& nestedExpression() const { return m_matrix; } EIGEN_DEVICE_FUNC - MatrixTypeNestedCleaned& nestedExpression() { return *const_cast(&m_matrix); } + NestedExpression& nestedExpression() { return *const_cast(&m_matrix); } /** \sa MatrixBase::conjugate() */ EIGEN_DEVICE_FUNC @@ -416,7 +418,7 @@ template class TriangularViewImpl<_Mat EIGEN_DEVICE_FUNC TriangularViewType& operator=(const TriangularViewImpl& other) - { return *this = other.nestedExpression(); } + { return *this = other.derived().nestedExpression(); } template EIGEN_DEVICE_FUNC @@ -462,11 +464,11 @@ template class TriangularViewImpl<_Mat template friend EIGEN_DEVICE_FUNC TriangularProduct - operator*(const MatrixBase& lhs, const TriangularVieImplw& rhs) + operator*(const MatrixBase& lhs, const TriangularViewImpl& rhs) { return TriangularProduct - (lhs.derived(),rhs.nestedExpression()); + (lhs.derived(),rhs.derived().nestedExpression()); } #endif @@ -827,20 +829,20 @@ TriangularViewImpl::operator=(const MatrixBase template -void TriangularViewImp::lazyAssign(const MatrixBase& other) +void TriangularViewImpl::lazyAssign(const MatrixBase& other) { enum { unroll = MatrixType::SizeAtCompileTime != Dynamic && internal::traits::CoeffReadCost != Dynamic && MatrixType::SizeAtCompileTime*internal::traits::CoeffReadCost/2 <= EIGEN_UNROLLING_LIMIT }; - eigen_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols()); + eigen_assert(derived().rows() == other.rows() && derived().cols() == other.cols()); internal::triangular_assignment_selector ::run(m_matrix.const_cast_derived(), other.derived()); + >::run(derived().nestedExpression().const_cast_derived(), other.derived()); } @@ -872,13 +874,13 @@ void TriangularViewImpl::lazyAssign(const TriangularBas && MatrixType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT }; - eigen_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols()); + eigen_assert(derived().rows() == other.rows() && derived().cols() == other.cols()); internal::triangular_assignment_selector ::run(m_matrix.const_cast_derived(), other.derived().nestedExpression()); + >::run(derived().nestedExpression().const_cast_derived(), other.derived().nestedExpression()); } #endif // EIGEN_TEST_EVALUATORS @@ -1043,13 +1045,13 @@ struct evaluator_traits > }; template -struct evaluator > +struct unary_evaluator, IndexBased> : evaluator::type> { typedef TriangularView XprType; typedef evaluator::type> Base; - typedef evaluator type; - evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {} + typedef evaluator type; + unary_evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {} }; // Additional assignment kinds: diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index a9d8352a9..06c64714a 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -270,7 +270,7 @@ TriangularView& TriangularViewImpl::_ass general_product_to_triangular_selector::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha); - return *this; + return derived(); } #else template @@ -282,7 +282,7 @@ TriangularView& TriangularViewImpl::assi general_product_to_triangular_selector ::run(derived().nestedExpression().const_cast_derived(), prod.derived(), alpha); - return *this; + return derived(); } #endif } // end namespace Eigen diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index f3da8c6c4..c2091ee49 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -173,6 +173,7 @@ struct Sparse2Sparse {}; struct Sparse2Dense {}; template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; +template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; template<> struct AssignmentKind { typedef Sparse2Dense Kind; }; diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index ff51fc435..10cd755a4 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -238,7 +238,7 @@ template SparseSelfAdjointView& SparseSelfAdjointView::rankUpdate(const SparseMatrixBase& u, const Scalar& alpha) { - SparseMatrix tmp = u * u.adjoint(); + SparseMatrix tmp = u * u.adjoint(); if(alpha==Scalar(0)) m_matrix.const_cast_derived() = tmp.template triangularView(); else diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index ad184208b..7586a0a6e 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // Copyright (C) 2012 Désiré Nuentsa-Wakam // // This Source Code Form is subject to the terms of the Mozilla @@ -13,15 +13,6 @@ namespace Eigen { -namespace internal { - -// template -// struct traits > -// : public traits -// {}; - -} // namespace internal - template class TriangularViewImpl : public SparseMatrixBase > { @@ -76,7 +67,7 @@ class TriangularViewImpl::InnerIterator : public MatrixT public: EIGEN_STRONG_INLINE InnerIterator(const TriangularViewImpl& view, Index outer) - : Base(view.nestedExpression(), outer), m_returnOne(false) + : Base(view.derived().nestedExpression(), outer), m_returnOne(false) { if(SkipFirst) { @@ -146,7 +137,7 @@ class TriangularViewImpl::ReverseInnerIterator : public public: EIGEN_STRONG_INLINE ReverseInnerIterator(const TriangularViewType& view, Index outer) - : Base(view.nestedExpression(), outer) + : Base(view.derived().nestedExpression(), outer) { eigen_assert((!HasUnitDiag) && "ReverseInnerIterator does not support yet triangular views with a unit diagonal"); if(SkipLast) { @@ -172,6 +163,115 @@ class TriangularViewImpl::ReverseInnerIterator : public } }; +#ifdef EIGEN_TEST_EVALUATORS +namespace internal { + +template +struct unary_evaluator, IteratorBased> + : evaluator_base > +{ + typedef TriangularView XprType; + +protected: + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::Index Index; + typedef typename evaluator::InnerIterator EvalIterator; + + enum { SkipFirst = ((Mode&Lower) && !(ArgType::Flags&RowMajorBit)) + || ((Mode&Upper) && (ArgType::Flags&RowMajorBit)), + SkipLast = !SkipFirst, + SkipDiag = (Mode&ZeroDiag) ? 1 : 0, + HasUnitDiag = (Mode&UnitDiag) ? 1 : 0 + }; + +public: + + enum { + CoeffReadCost = evaluator::CoeffReadCost, + Flags = XprType::Flags + }; + + unary_evaluator(const XprType &xpr) : m_argImpl(xpr.nestedExpression()) {} + + class InnerIterator : public EvalIterator + { + typedef EvalIterator Base; + public: + + EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& xprEval, Index outer) + : Base(xprEval.m_argImpl,outer) + { + if(SkipFirst) + { + while((*this) && ((HasUnitDiag||SkipDiag) ? this->index()<=outer : this->index()=Base::outer())) + { + if((!SkipFirst) && Base::operator bool()) + Base::operator++(); + m_returnOne = true; + } + } + + EIGEN_STRONG_INLINE InnerIterator& operator++() + { + if(HasUnitDiag && m_returnOne) + m_returnOne = false; + else + { + Base::operator++(); + if(HasUnitDiag && (!SkipFirst) && ((!Base::operator bool()) || Base::index()>=Base::outer())) + { + if((!SkipFirst) && Base::operator bool()) + Base::operator++(); + m_returnOne = true; + } + } + return *this; + } + + EIGEN_STRONG_INLINE operator bool() const + { + if(HasUnitDiag && m_returnOne) + return true; + if(SkipFirst) return Base::operator bool(); + else + { + if (SkipDiag) return (Base::operator bool() && this->index() < this->outer()); + else return (Base::operator bool() && this->index() <= this->outer()); + } + } + +// inline Index row() const { return (ArgType::Flags&RowMajorBit ? Base::outer() : this->index()); } +// inline Index col() const { return (ArgType::Flags&RowMajorBit ? this->index() : Base::outer()); } + inline Index index() const + { + if(HasUnitDiag && m_returnOne) return Base::outer(); + else return Base::index(); + } + inline Scalar value() const + { + if(HasUnitDiag && m_returnOne) return Scalar(1); + else return Base::value(); + } + + protected: + bool m_returnOne; + private: + Scalar& valueRef(); + }; + +protected: + typename evaluator::type m_argImpl; +}; + +} // end namespace internal +#endif + template template inline const TriangularView diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 686be6c92..0183907a1 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -54,7 +54,7 @@ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) typedef typename Eigen::internal::traits::Index Index; \ enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ - Flags = Eigen::internal::traits::Flags, \ + Flags = Eigen::internal::traits::Flags, \ CoeffReadCost = Eigen::internal::traits::CoeffReadCost, \ SizeAtCompileTime = Base::SizeAtCompileTime, \ IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \ @@ -72,7 +72,7 @@ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) typedef typename Eigen::internal::traits::Index Index; \ enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ - Flags = Eigen::internal::traits::Flags, \ + Flags = Eigen::internal::traits::Flags, \ SizeAtCompileTime = Base::SizeAtCompileTime, \ IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \ using Base::derived; \ diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 5b5b55c60..a53ce22f5 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -216,7 +216,6 @@ ei_add_test(jacobi) ei_add_test(jacobisvd) ei_add_test(householder) if(NOT EIGEN_TEST_EVALUATORS) - ei_add_test(cwiseop) # Eigen2 related ei_add_test(geo_orthomethods) ei_add_test(geo_homogeneous) ei_add_test(geo_quaternion) From cd0ff253ec906f0b4604ec084df3d7e7cc6811d1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 30 Jul 2014 15:22:50 +0200 Subject: [PATCH 0668/1103] Make permutation compatible with sparse matrices --- CMakeLists.txt | 2 +- Eigen/SparseCore | 3 - Eigen/src/Core/PermutationMatrix.h | 9 +- Eigen/src/Core/util/Constants.h | 3 + Eigen/src/Core/util/XprHelper.h | 24 ++-- Eigen/src/SparseCore/SparseCwiseUnaryOp.h | 92 ++++++------- Eigen/src/SparseCore/SparseMatrix.h | 8 +- Eigen/src/SparseCore/SparsePermutation.h | 131 ++++++++++++++++++- Eigen/src/SparseCore/SparseSelfAdjointView.h | 12 +- test/CMakeLists.txt | 2 +- 10 files changed, 213 insertions(+), 73 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 470095680..cc3988196 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -143,7 +143,7 @@ if(NOT MSVC) ei_add_cxx_compiler_flag("-Wpointer-arith") ei_add_cxx_compiler_flag("-Wwrite-strings") ei_add_cxx_compiler_flag("-Wformat-security") - ei_add_cxx_compiler_flag("-Wshorten-64-to-32") +# ei_add_cxx_compiler_flag("-Wshorten-64-to-32") ei_add_cxx_compiler_flag("-Wenum-conversion") ei_add_cxx_compiler_flag("-Wc++11-extensions") diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 578469c1c..41cae58b0 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -56,11 +56,8 @@ struct Sparse {}; #include "src/SparseCore/SparseSelfAdjointView.h" #include "src/SparseCore/SparseTriangularView.h" #include "src/SparseCore/TriangularSolver.h" - -#ifndef EIGEN_TEST_EVALUATORS #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" -#endif #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 5d648c68c..84d64edb3 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -13,7 +13,8 @@ namespace Eigen { -template class PermutedImpl; +// TODO: this does not seems to be needed at all: +// template class PermutedImpl; /** \class PermutationBase * \ingroup Core_Module @@ -276,6 +277,7 @@ template > : traits > { + typedef PermutationStorage StorageKind; typedef Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; typedef typename IndicesType::Index Index; typedef _StorageIndexType StorageIndexType; @@ -397,6 +399,7 @@ template,_PacketAccess> > : traits > { + typedef PermutationStorage StorageKind; typedef Map, _PacketAccess> IndicesType; typedef typename IndicesType::Index Index; typedef _StorageIndexType StorageIndexType; @@ -468,8 +471,6 @@ class Map class TranspositionsWrapper; namespace internal { template @@ -665,6 +666,8 @@ struct traits > > } // end namespace internal +// TODO: the specificties should be handled by the evaluator, +// at the very least we should only specialize TransposeImpl template class Transpose > : public EigenBase > > diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 3ab8d0ed3..c0463583d 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -443,6 +443,9 @@ enum Action {GetAction, SetAction}; /** The type used to identify a dense storage. */ struct Dense {}; +/** The type used to identify a permutation storage. */ +struct PermutationStorage {}; + /** The type used to identify a matrix expression */ struct MatrixXpr {}; diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 7779fb3fa..5ed3b39b7 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -557,18 +557,26 @@ template struct cwise_promote_s * K * dense -> dense * diag * K -> K * K * diag -> K + * Perm * K -> K + * K * Perm -> K * \endcode */ template struct product_promote_storage_type; -template struct product_promote_storage_type { typedef A ret;}; -template struct product_promote_storage_type { typedef Dense ret;}; -template struct product_promote_storage_type { typedef Dense ret; }; -template struct product_promote_storage_type { typedef Dense ret; }; -template struct product_promote_storage_type { typedef A ret; }; -template struct product_promote_storage_type { typedef B ret; }; -template struct product_promote_storage_type { typedef Dense ret; }; -template struct product_promote_storage_type { typedef Dense ret; }; +template struct product_promote_storage_type { typedef A ret;}; +template struct product_promote_storage_type { typedef Dense ret;}; +template struct product_promote_storage_type { typedef Dense ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; + +template struct product_promote_storage_type { typedef A ret; }; +template struct product_promote_storage_type { typedef B ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; + +template struct product_promote_storage_type { typedef A ret; }; +template struct product_promote_storage_type { typedef B ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; +template struct product_promote_storage_type { typedef Dense ret; }; /** \internal gives the plain matrix or array type to store a row/column/diagonal of a matrix type. * \param Scalar optional parameter allowing to pass a different scalar type than the one of the MatrixType. diff --git a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h index 7a849c822..0e0f9f5f5 100644 --- a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h @@ -152,7 +152,7 @@ struct unary_evaluator, IteratorBased> typedef CwiseUnaryOp XprType; class InnerIterator; - class ReverseInnerIterator; +// class ReverseInnerIterator; enum { CoeffReadCost = evaluator::CoeffReadCost + functor_traits::Cost, @@ -163,7 +163,7 @@ struct unary_evaluator, IteratorBased> protected: typedef typename evaluator::InnerIterator EvalIterator; - typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; +// typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; const UnaryOp m_functor; typename evaluator::nestedType m_argImpl; @@ -192,28 +192,28 @@ class unary_evaluator, IteratorBased>::InnerIterat Scalar& valueRef(); }; -template -class unary_evaluator, IteratorBased>::ReverseInnerIterator - : public unary_evaluator, IteratorBased>::EvalReverseIterator -{ - typedef typename XprType::Scalar Scalar; - typedef typename unary_evaluator, IteratorBased>::EvalReverseIterator Base; - public: - - EIGEN_STRONG_INLINE ReverseInnerIterator(const XprType& unaryOp, typename XprType::Index outer) - : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) - {} - - EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() - { Base::operator--(); return *this; } - - EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } - - protected: - const UnaryOp m_functor; - private: - Scalar& valueRef(); -}; +// template +// class unary_evaluator, IteratorBased>::ReverseInnerIterator +// : public unary_evaluator, IteratorBased>::EvalReverseIterator +// { +// typedef typename XprType::Scalar Scalar; +// typedef typename unary_evaluator, IteratorBased>::EvalReverseIterator Base; +// public: +// +// EIGEN_STRONG_INLINE ReverseInnerIterator(const XprType& unaryOp, typename XprType::Index outer) +// : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) +// {} +// +// EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() +// { Base::operator--(); return *this; } +// +// EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } +// +// protected: +// const UnaryOp m_functor; +// private: +// Scalar& valueRef(); +// }; @@ -238,7 +238,7 @@ struct unary_evaluator, IteratorBased> protected: typedef typename evaluator::InnerIterator EvalIterator; - typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; +// typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; const ViewOp m_functor; typename evaluator::nestedType m_argImpl; @@ -266,27 +266,27 @@ class unary_evaluator, IteratorBased>::InnerItera const ViewOp m_functor; }; -template -class unary_evaluator, IteratorBased>::ReverseInnerIterator - : public unary_evaluator, IteratorBased>::EvalReverseIterator -{ - typedef typename XprType::Scalar Scalar; - typedef typename unary_evaluator, IteratorBased>::EvalReverseIterator Base; - public: - - EIGEN_STRONG_INLINE ReverseInnerIterator(const XprType& unaryOp, typename XprType::Index outer) - : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) - {} - - EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() - { Base::operator--(); return *this; } - - EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } - EIGEN_STRONG_INLINE Scalar& valueRef() { return m_functor(Base::valueRef()); } - - protected: - const ViewOp m_functor; -}; +// template +// class unary_evaluator, IteratorBased>::ReverseInnerIterator +// : public unary_evaluator, IteratorBased>::EvalReverseIterator +// { +// typedef typename XprType::Scalar Scalar; +// typedef typename unary_evaluator, IteratorBased>::EvalReverseIterator Base; +// public: +// +// EIGEN_STRONG_INLINE ReverseInnerIterator(const XprType& unaryOp, typename XprType::Index outer) +// : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) +// {} +// +// EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() +// { Base::operator--(); return *this; } +// +// EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); } +// EIGEN_STRONG_INLINE Scalar& valueRef() { return m_functor(Base::valueRef()); } +// +// protected: +// const ViewOp m_functor; +// }; } // end namespace internal diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 6080c272a..9e2dae1b3 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -736,8 +736,8 @@ class SparseMatrix return *this; } +#ifndef EIGEN_PARSED_BY_DOXYGEN #ifndef EIGEN_TEST_EVALUATORS - #ifndef EIGEN_PARSED_BY_DOXYGEN template inline SparseMatrix& operator=(const SparseSparseProduct& product) { return Base::operator=(product); } @@ -748,12 +748,12 @@ class SparseMatrix initAssignment(other); return Base::operator=(other.derived()); } - +#endif // EIGEN_TEST_EVALUATORS + template inline SparseMatrix& operator=(const EigenBase& other) { return Base::operator=(other.derived()); } - #endif -#endif // EIGEN_TEST_EVALUATORS +#endif // EIGEN_PARSED_BY_DOXYGEN template EIGEN_DONT_INLINE SparseMatrix& operator=(const SparseMatrixBase& other); diff --git a/Eigen/src/SparseCore/SparsePermutation.h b/Eigen/src/SparseCore/SparsePermutation.h index b85be93f6..ebfefab98 100644 --- a/Eigen/src/SparseCore/SparsePermutation.h +++ b/Eigen/src/SparseCore/SparsePermutation.h @@ -103,7 +103,7 @@ struct permut_sparsematrix_product_retval } - +#ifndef EIGEN_TEST_EVALUATORS /** \returns the matrix with the permutation applied to the columns */ @@ -143,6 +143,135 @@ operator*(const Transpose >& tperm, const SparseMat return internal::permut_sparsematrix_product_retval, SparseDerived, OnTheLeft, true>(tperm.nestedPermutation(), matrix.derived()); } +#else // EIGEN_TEST_EVALUATORS + +namespace internal { + +template struct product_promote_storage_type { typedef Sparse ret; }; +template struct product_promote_storage_type { typedef Sparse ret; }; + +// TODO, the following need cleaning, this is just a copy-past of the dense case + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + permut_sparsematrix_product_retval pmpr(lhs, rhs); + pmpr.evalTo(dst); + } +}; + +template +struct generic_product_impl +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs) + { + permut_sparsematrix_product_retval pmpr(rhs, lhs); + pmpr.evalTo(dst); + } +}; + +template +struct generic_product_impl, Rhs, PermutationShape, SparseShape, ProductTag> +{ + template + static void evalTo(Dest& dst, const Transpose& lhs, const Rhs& rhs) + { + permut_sparsematrix_product_retval pmpr(lhs.nestedPermutation(), rhs); + pmpr.evalTo(dst); + } +}; + +template +struct generic_product_impl, SparseShape, PermutationShape, ProductTag> +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Transpose& rhs) + { + permut_sparsematrix_product_retval pmpr(rhs.nestedPermutation(), lhs); + pmpr.evalTo(dst); + } +}; + +template +struct product_evaluator, ProductTag, PermutationShape, SparseShape, typename traits::Scalar, typename traits::Scalar> + : public evaluator >::ReturnType>::type +{ + typedef Product XprType; + typedef typename traits >::ReturnType PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +template +struct product_evaluator, ProductTag, SparseShape, PermutationShape, typename traits::Scalar, typename traits::Scalar> + : public evaluator >::ReturnType>::type +{ + typedef Product XprType; + typedef typename traits >::ReturnType PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +} // end namespace internal + +/** \returns the matrix with the permutation applied to the columns + */ +template +inline const Product +operator*(const SparseMatrixBase& matrix, const PermutationBase& perm) +{ return Product(matrix.derived(), perm.derived()); } + +/** \returns the matrix with the permutation applied to the rows + */ +template +inline const Product +operator*( const PermutationBase& perm, const SparseMatrixBase& matrix) +{ return Product(perm.derived(), matrix.derived()); } + + +// TODO, the following specializations should not be needed as Transpose should be a PermutationBase. +/** \returns the matrix with the inverse permutation applied to the columns. + */ +template +inline const Product > > +operator*(const SparseMatrixBase& matrix, const Transpose >& tperm) +{ + return Product > >(matrix.derived(), tperm); +} + +/** \returns the matrix with the inverse permutation applied to the rows. + */ +template +inline const Product >, SparseDerived> +operator*(const Transpose >& tperm, const SparseMatrixBase& matrix) +{ + return Product >, SparseDerived>(tperm, matrix.derived()); +} + +#endif // EIGEN_TEST_EVALUATORS + } // end namespace Eigen #endif // EIGEN_SPARSE_SELFADJOINTVIEW_H diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 10cd755a4..530ff27bf 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -20,7 +20,7 @@ class SparseSelfAdjointTimeDenseProduct; template class DenseTimeSparseSelfAdjointProduct; -#endif // #ifndef EIGEN_TEST_EVALUATORS +#endif // EIGEN_TEST_EVALUATORS /** \ingroup SparseCore_Module * \class SparseSelfAdjointView @@ -177,7 +177,7 @@ template class SparseSelfAdjointView } /** \returns an expression of P H P^-1 */ -#ifndef EIGEN_TEST_EVALUATORS +// #ifndef EIGEN_TEST_EVALUATORS SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode> twistedBy(const PermutationMatrix& perm) const { return SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode>(m_matrix, perm); @@ -189,7 +189,7 @@ template class SparseSelfAdjointView permutedMatrix.evalTo(*this); return *this; } -#endif // EIGEN_TEST_EVALUATORS +// #endif // EIGEN_TEST_EVALUATORS SparseSelfAdjointView& operator=(const SparseSelfAdjointView& src) { @@ -680,7 +680,7 @@ void permute_symm_to_symm(const MatrixType& mat, SparseMatrix Date: Thu, 31 Jul 2014 13:35:49 +0200 Subject: [PATCH 0669/1103] Call product_generic_impl by default, and remove lot of boilerplate code --- Eigen/src/Core/CoreEvaluators.h | 4 +- Eigen/src/Core/Product.h | 4 +- Eigen/src/Core/ProductEvaluators.h | 130 +------------- Eigen/src/Geometry/Homogeneous.h | 175 ++++++++++++++++++- Eigen/src/Geometry/Transform.h | 22 +++ Eigen/src/SparseCore/SparseDenseProduct.h | 42 +---- Eigen/src/SparseCore/SparseDiagonalProduct.h | 12 +- Eigen/src/SparseCore/SparsePermutation.h | 4 + Eigen/src/SparseCore/SparseProduct.h | 19 -- Eigen/src/SparseCore/SparseSelfAdjointView.h | 46 +---- test/main.h | 6 - 11 files changed, 223 insertions(+), 241 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index b19a29e53..66984e378 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -67,8 +67,8 @@ struct evaluator_traits_base // typedef evaluator nestedType; // by default, get evaluator kind and shape from storage - typedef typename storage_kind_to_evaluator_kind::Kind Kind; - typedef typename storage_kind_to_shape::Shape Shape; + typedef typename storage_kind_to_evaluator_kind::StorageKind>::Kind Kind; + typedef typename storage_kind_to_shape::StorageKind>::Shape Shape; // 1 if assignment A = B assumes aliasing when B is of type T and thus B needs to be evaluated into a // temporary; 0 if not. diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 9e5e47d13..0cf20f2e2 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -108,8 +108,8 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, typedef typename ProductImpl< Lhs, Rhs, Option, - typename internal::product_promote_storage_type::StorageKind, + typename internal::traits::StorageKind, internal::product_type::ret>::ret>::Base Base; EIGEN_GENERIC_PUBLIC_INTERFACE(Product) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index b04df00e7..8a63384a7 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -90,9 +90,10 @@ struct evaluator_traits > enum { AssumeAliasing = 1 }; }; -// The evaluator for default dense products creates a temporary and call generic_product_impl -template -struct product_evaluator, ProductTag, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> +// This is the default evaluator implementation for products: +// It creates a temporary and call generic_product_impl +template +struct product_evaluator, ProductTag, LhsShape, RhsShape, typename traits::Scalar, typename traits::Scalar> : public evaluator::PlainObject>::type { typedef Product XprType; @@ -118,7 +119,7 @@ struct product_evaluator, ProductTag, DenseSha // // generic_product_impl::evalTo(m_result, lhs, rhs); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); } protected: @@ -501,8 +502,8 @@ protected: }; template -struct product_evaluator, LazyCoeffBasedProductMode, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar > - : product_evaluator, CoeffBasedProductMode, DenseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar > +struct product_evaluator, LazyCoeffBasedProductMode, DenseShape, DenseShape, typename traits::Scalar, typename traits::Scalar > + : product_evaluator, CoeffBasedProductMode, DenseShape, DenseShape, typename traits::Scalar, typename traits::Scalar > { typedef Product XprType; typedef Product BaseProduct; @@ -607,26 +608,6 @@ struct generic_product_impl } }; -template -struct product_evaluator, ProductTag, TriangularShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - - template struct generic_product_impl : generic_product_impl_base > @@ -640,26 +621,6 @@ struct generic_product_impl } }; -template -struct product_evaluator, ProductTag, DenseShape, TriangularShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - - /*************************************************************************** * SelfAdjoint products @@ -681,26 +642,6 @@ struct generic_product_impl } }; -template -struct product_evaluator, ProductTag, SelfAdjointShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - - template struct generic_product_impl : generic_product_impl_base > @@ -714,24 +655,6 @@ struct generic_product_impl } }; -template -struct product_evaluator, ProductTag, DenseShape, SelfAdjointShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; /*************************************************************************** * Diagonal products @@ -933,45 +856,6 @@ struct generic_product_impl, DenseShape, PermutationShape, P } }; -// TODO: left/right and self-adj/symmetric/permutation look the same ... Too much boilerplate? -template -struct product_evaluator, ProductTag, PermutationShape, DenseShape, typename traits::Scalar, typename traits::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - -template -struct product_evaluator, ProductTag, DenseShape, PermutationShape, typename traits::Scalar, typename traits::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index 00e71d190..07bc22154 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -48,8 +48,10 @@ struct traits > TmpFlags = _MatrixTypeNested::Flags & HereditaryBits, Flags = ColsAtCompileTime==1 ? (TmpFlags & ~RowMajorBit) : RowsAtCompileTime==1 ? (TmpFlags | RowMajorBit) - : TmpFlags, - CoeffReadCost = _MatrixTypeNested::CoeffReadCost + : TmpFlags +#ifndef EIGEN_TEST_EVALUATORS + , CoeffReadCost = _MatrixTypeNested::CoeffReadCost +#endif // EIGEN_TEST_EVALUATORS }; }; @@ -63,6 +65,7 @@ template class Homogeneous { public: + typedef MatrixType NestedExpression; enum { Direction = _Direction }; typedef MatrixBase Base; @@ -74,7 +77,10 @@ template class Homogeneous inline Index rows() const { return m_matrix.rows() + (int(Direction)==Vertical ? 1 : 0); } inline Index cols() const { return m_matrix.cols() + (int(Direction)==Horizontal ? 1 : 0); } + + const NestedExpression& nestedExpression() const { return m_matrix; } +#ifndef EIGEN_TEST_EVALUATORS inline Scalar coeff(Index row, Index col) const { if( (int(Direction)==Vertical && row==m_matrix.rows()) @@ -106,6 +112,31 @@ template class Homogeneous eigen_assert(int(Direction)==Vertical); return internal::homogeneous_left_product_impl >(lhs,rhs.m_matrix); } +#else + template + inline const Product + operator* (const MatrixBase& rhs) const + { + eigen_assert(int(Direction)==Horizontal); + return Product(*this,rhs.derived()); + } + + template friend + inline const Product + operator* (const MatrixBase& lhs, const Homogeneous& rhs) + { + eigen_assert(int(Direction)==Vertical); + return Product(lhs.derived(),rhs); + } + + template friend + inline const Product, Homogeneous > + operator* (const Transform& lhs, const Homogeneous& rhs) + { + eigen_assert(int(Direction)==Vertical); + return Product, Homogeneous>(lhs,rhs); + } +#endif protected: typename MatrixType::Nested m_matrix; @@ -300,6 +331,146 @@ struct homogeneous_right_product_impl,Rhs> typename Rhs::Nested m_rhs; }; +#ifdef EIGEN_TEST_EVALUATORS +template +struct unary_evaluator, IndexBased> + : evaluator::PlainObject >::type +{ + typedef Homogeneous XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + unary_evaluator(const XprType& op) + : Base(), m_temp(op) + { + ::new (static_cast(this)) Base(m_temp); + } + +protected: + PlainObject m_temp; +}; + +// dense = homogeneous +template< typename DstXprType, typename ArgType, typename Scalar> +struct Assignment, internal::assign_op, Dense2Dense, Scalar> +{ + typedef Homogeneous SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dst.template topRows(src.nestedExpression().rows()) = src.nestedExpression(); +// dst.topRows(src.nestedExpression().rows()) = src.nestedExpression(); + dst.row(dst.rows()-1).setOnes(); + } +}; + +// dense = homogeneous +template< typename DstXprType, typename ArgType, typename Scalar> +struct Assignment, internal::assign_op, Dense2Dense, Scalar> +{ + typedef Homogeneous SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + dst.template leftCols(src.nestedExpression().cols()) = src.nestedExpression(); +// dst.leftCols(src.nestedExpression().cols()) = src.nestedExpression(); + dst.col(dst.cols()-1).setOnes(); + } +}; + +template +struct generic_product_impl, Rhs, DenseShape, DenseShape, ProductTag> +{ + template + static void evalTo(Dest& dst, const Homogeneous& lhs, const Rhs& rhs) + { + homogeneous_right_product_impl, Rhs>(lhs.nestedExpression(), rhs).evalTo(dst); + } +}; + +template +struct generic_product_impl, DenseShape, DenseShape, ProductTag> +{ + template + static void evalTo(Dest& dst, const Lhs& lhs, const Homogeneous& rhs) + { + homogeneous_left_product_impl, Lhs>(rhs.nestedExpression(), lhs).evalTo(dst); + } +}; + +template +struct generic_product_impl, Homogeneous, DenseShape, DenseShape, ProductTag> +{ + typedef Transform TransformType; + template + static void evalTo(Dest& dst, const TransformType& lhs, const Homogeneous& rhs) + { + homogeneous_left_product_impl, TransformType>(rhs.nestedExpression(), lhs).evalTo(dst); + } +}; + + +template +struct product_evaluator, Rhs, DefaultProduct>, ProductTag, DenseShape, DenseShape, typename traits::Scalar, typename traits::Scalar> + : public evaluator, Rhs, DefaultProduct>::PlainObject>::type +{ + typedef Homogeneous Lhs; + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +template +struct product_evaluator, DefaultProduct>, ProductTag, DenseShape, DenseShape, typename traits::Scalar, typename traits::Scalar> + : public evaluator, DefaultProduct>::PlainObject>::type +{ + typedef Homogeneous Rhs; + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +template +struct product_evaluator, Homogeneous, DefaultProduct>, ProductTag, DenseShape, DenseShape, Scalar, typename traits::Scalar> + : public evaluator, Homogeneous, DefaultProduct>::PlainObject>::type +{ + typedef Transform Lhs; + typedef Homogeneous Rhs; + typedef Product XprType; + typedef typename XprType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + product_evaluator(const XprType& xpr) + : m_result(xpr.rows(), xpr.cols()) + { + ::new (static_cast(this)) Base(m_result); + generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); + } + +protected: + PlainObject m_result; +}; + +#endif // EIGEN_TEST_EVALUATORS + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index cb93acf6b..54d05f9cf 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -62,6 +62,23 @@ struct transform_construct_from_matrix; template struct transform_take_affine_part; +#ifdef EIGEN_TEST_EVALUATORS +template +struct traits > +{ + typedef _Scalar Scalar; + typedef DenseIndex Index; + typedef Dense StorageKind; + enum { + RowsAtCompileTime = _Dim, + ColsAtCompileTime = _Dim, + MaxRowsAtCompileTime = _Dim, + MaxColsAtCompileTime = _Dim, + Flags = 0 + }; +}; +#endif + } // end namespace internal /** \geometry_module \ingroup Geometry_Module @@ -355,6 +372,11 @@ public: inline Transform& operator=(const QTransform& other); inline QTransform toQTransform(void) const; #endif + +#ifdef EIGEN_TEST_EVALUATORS + Index rows() const { return m_matrix.cols(); } + Index cols() const { return m_matrix.cols(); } +#endif /** shortcut for m_matrix(row,col); * \sa MatrixBase::operator(Index,Index) const */ diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 8864b7308..a715b8bde 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -410,44 +410,6 @@ struct generic_product_impl } }; -template -struct product_evaluator, ProductTag, SparseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - -template -struct product_evaluator, ProductTag, DenseShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - template struct sparse_dense_outer_product_evaluator { @@ -530,7 +492,7 @@ protected: // sparse * dense outer product template -struct product_evaluator, OuterProduct, SparseShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> +struct product_evaluator, OuterProduct, SparseShape, DenseShape, typename traits::Scalar, typename traits::Scalar> : sparse_dense_outer_product_evaluator { typedef sparse_dense_outer_product_evaluator Base; @@ -545,7 +507,7 @@ struct product_evaluator, OuterProduct, Sparse }; template -struct product_evaluator, OuterProduct, DenseShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> +struct product_evaluator, OuterProduct, DenseShape, SparseShape, typename traits::Scalar, typename traits::Scalar> : sparse_dense_outer_product_evaluator { typedef sparse_dense_outer_product_evaluator Base; diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index 4c51881e0..9f465a828 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -197,11 +197,11 @@ enum { template struct sparse_diagonal_product_evaluator; -template -struct product_evaluator, ProductTag, DiagonalShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> +template +struct product_evaluator, ProductTag, DiagonalShape, SparseShape, typename traits::Scalar, typename traits::Scalar> : public sparse_diagonal_product_evaluator { - typedef Product XprType; + typedef Product XprType; typedef evaluator type; typedef evaluator nestedType; enum { CoeffReadCost = Dynamic, Flags = Rhs::Flags&RowMajorBit }; // FIXME CoeffReadCost & Flags @@ -210,11 +210,11 @@ struct product_evaluator, ProductTag, DiagonalShape, product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) {} }; -template -struct product_evaluator, ProductTag, SparseShape, DiagonalShape, typename Lhs::Scalar, typename Rhs::Scalar> +template +struct product_evaluator, ProductTag, SparseShape, DiagonalShape, typename traits::Scalar, typename traits::Scalar> : public sparse_diagonal_product_evaluator, Lhs::Flags&RowMajorBit?SDP_AsCwiseProduct:SDP_AsScalarProduct> { - typedef Product XprType; + typedef Product XprType; typedef evaluator type; typedef evaluator nestedType; enum { CoeffReadCost = Dynamic, Flags = Lhs::Flags&RowMajorBit }; // FIXME CoeffReadCost & Flags diff --git a/Eigen/src/SparseCore/SparsePermutation.h b/Eigen/src/SparseCore/SparsePermutation.h index ebfefab98..a888ae9e1 100644 --- a/Eigen/src/SparseCore/SparsePermutation.h +++ b/Eigen/src/SparseCore/SparsePermutation.h @@ -196,6 +196,10 @@ struct generic_product_impl, SparseShape, PermutationShape, } }; +// TODO, the following two overloads are only needed to define the right temporary type through +// typename traits >::ReturnType +// while it should be correctly handled by traits >::PlainObject + template struct product_evaluator, ProductTag, PermutationShape, SparseShape, typename traits::Scalar, typename traits::Scalar> : public evaluator >::ReturnType>::type diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index 4e181d471..18f40b9d9 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -224,25 +224,6 @@ struct generic_product_impl } }; -template -struct product_evaluator, ProductTag, SparseShape, SparseShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - template struct evaluator > > : public evaluator::PlainObject>::type diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 530ff27bf..4235d6c4c 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -441,46 +441,11 @@ struct generic_product_impl -struct product_evaluator, ProductTag, SparseSelfAdjointShape, DenseShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - -template -struct product_evaluator, ProductTag, DenseShape, SparseSelfAdjointShape, typename Lhs::Scalar, typename Rhs::Scalar> - : public evaluator::PlainObject>::type -{ - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; +// NOTE: these two overloads are needed to evaluate the sparse sefladjoint view into a full sparse matrix +// TODO: maybe the copy could be handled by generic_product_impl so that these overloads would not be needed anymore template -struct product_evaluator, ProductTag, SparseSelfAdjointShape, SparseShape, typename LhsView::Scalar, typename Rhs::Scalar> +struct product_evaluator, ProductTag, SparseSelfAdjointShape, SparseShape, typename traits::Scalar, typename traits::Scalar> : public evaluator::PlainObject>::type { typedef Product XprType; @@ -488,9 +453,8 @@ struct product_evaluator, ProductTag, Spar typedef typename evaluator::type Base; product_evaluator(const XprType& xpr) - : /*m_lhs(xpr.lhs()),*/ m_result(xpr.rows(), xpr.cols()) + : m_lhs(xpr.lhs()), m_result(xpr.rows(), xpr.cols()) { - m_lhs = xpr.lhs(); ::new (static_cast(this)) Base(m_result); generic_product_impl::evalTo(m_result, m_lhs, xpr.rhs()); } @@ -501,7 +465,7 @@ protected: }; template -struct product_evaluator, ProductTag, SparseShape, SparseSelfAdjointShape, typename Lhs::Scalar, typename RhsView::Scalar> +struct product_evaluator, ProductTag, SparseShape, SparseSelfAdjointShape, typename traits::Scalar, typename traits::Scalar> : public evaluator::PlainObject>::type { typedef Product XprType; diff --git a/test/main.h b/test/main.h index 57996956d..e89b5a305 100644 --- a/test/main.h +++ b/test/main.h @@ -279,13 +279,7 @@ inline bool test_isApproxOrLessThan(const long double& a, const long double& b) template inline bool test_isApprox(const Type1& a, const Type2& b) { -#ifdef EIGEN_TEST_EVALUATORS - typename internal::eval::type a_eval(a); - typename internal::eval::type b_eval(b); - return a_eval.isApprox(b_eval, test_precision()); -#else return a.isApprox(b, test_precision()); -#endif } // The idea behind this function is to compare the two scalars a and b where From 5f5a8d97c02deeab6dbf977d40170197c75adc83 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 31 Jul 2014 13:43:19 +0200 Subject: [PATCH 0670/1103] Re-enable main unit tests which are now compiling and running fine with evaluators --- test/CMakeLists.txt | 44 +++++++++++++++++++++----------------------- 1 file changed, 21 insertions(+), 23 deletions(-) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index a9d0ed490..c7614e1af 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -215,32 +215,30 @@ ei_add_test(eigensolver_generalized_real) ei_add_test(jacobi) ei_add_test(jacobisvd) ei_add_test(householder) +ei_add_test(geo_orthomethods) +ei_add_test(geo_quaternion) +ei_add_test(geo_eulerangles) +ei_add_test(geo_parametrizedline) +ei_add_test(geo_alignedbox) if(NOT EIGEN_TEST_EVALUATORS) - ei_add_test(geo_orthomethods) - ei_add_test(geo_homogeneous) - ei_add_test(geo_quaternion) - ei_add_test(geo_transformations) - ei_add_test(geo_eulerangles) ei_add_test(geo_hyperplane) - ei_add_test(geo_parametrizedline) - ei_add_test(geo_alignedbox) - ei_add_test(stdvector) - ei_add_test(stdvector_overload) - ei_add_test(stdlist) - ei_add_test(stddeque) -endif(NOT EIGEN_TEST_EVALUATORS) - ei_add_test(sparse_basic) - ei_add_test(sparse_vector) - ei_add_test(sparse_product) - ei_add_test(sparse_solvers) - ei_add_test(sparse_permutations) - ei_add_test(simplicial_cholesky) - ei_add_test(conjugate_gradient) - ei_add_test(bicgstab) - ei_add_test(sparselu) - ei_add_test(sparseqr) -if(NOT EIGEN_TEST_EVALUATORS) + ei_add_test(geo_transformations) + ei_add_test(geo_homogeneous) endif(NOT EIGEN_TEST_EVALUATORS) +ei_add_test(stdvector) +ei_add_test(stdvector_overload) +ei_add_test(stdlist) +ei_add_test(stddeque) +ei_add_test(sparse_basic) +ei_add_test(sparse_vector) +ei_add_test(sparse_product) +ei_add_test(sparse_solvers) +ei_add_test(sparse_permutations) +ei_add_test(simplicial_cholesky) +ei_add_test(conjugate_gradient) +ei_add_test(bicgstab) +ei_add_test(sparselu) +ei_add_test(sparseqr) ei_add_test(umeyama) ei_add_test(nesting_ops "${CMAKE_CXX_FLAGS_DEBUG}") ei_add_test(zerosized) From 702a3c17db753b4ceaa234eac6591d0d1737015a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 31 Jul 2014 14:54:00 +0200 Subject: [PATCH 0671/1103] Make Transform exposes sizes: Dim+1 x Dim+1 for projective transform, and Dim x Dim+1 for all others --- Eigen/src/Geometry/Transform.h | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 54d05f9cf..a62548f12 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -70,10 +70,11 @@ struct traits > typedef DenseIndex Index; typedef Dense StorageKind; enum { - RowsAtCompileTime = _Dim, - ColsAtCompileTime = _Dim, - MaxRowsAtCompileTime = _Dim, - MaxColsAtCompileTime = _Dim, + Dim1 = _Dim==Dynamic ? _Dim : _Dim + 1, + RowsAtCompileTime = _Mode==Projective ? Dim1 : _Dim, + ColsAtCompileTime = Dim1, + MaxRowsAtCompileTime = RowsAtCompileTime, + MaxColsAtCompileTime = ColsAtCompileTime, Flags = 0 }; }; @@ -374,7 +375,7 @@ public: #endif #ifdef EIGEN_TEST_EVALUATORS - Index rows() const { return m_matrix.cols(); } + Index rows() const { return Mode==Projective ? m_matrix.cols() : m_matrix.cols()-1; } Index cols() const { return m_matrix.cols(); } #endif From db183ca7b31a366fa1a870927d68c3bf6de15c4d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 31 Jul 2014 14:54:54 +0200 Subject: [PATCH 0672/1103] Make minimal changes to make homogenous compatible with evaluators --- Eigen/src/Core/util/Constants.h | 1 + Eigen/src/Geometry/Homogeneous.h | 94 +++++++++----------------------- test/geo_homogeneous.cpp | 3 +- 3 files changed, 27 insertions(+), 71 deletions(-) diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index c0463583d..86817a989 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -454,6 +454,7 @@ struct ArrayXpr {}; // An evaluator must define its shape. By default, it can be one of the following: struct DenseShape { static std::string debugName() { return "DenseShape"; } }; +struct HomogeneousShape { static std::string debugName() { return "HomogeneousShape"; } }; struct DiagonalShape { static std::string debugName() { return "DiagonalShape"; } }; struct BandShape { static std::string debugName() { return "BandShape"; } }; struct TriangularShape { static std::string debugName() { return "TriangularShape"; } }; diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index 07bc22154..7a3b5de80 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -332,6 +332,18 @@ struct homogeneous_right_product_impl,Rhs> }; #ifdef EIGEN_TEST_EVALUATORS + +template +struct evaluator_traits > +{ + typedef typename storage_kind_to_evaluator_kind::Kind Kind; + typedef HomogeneousShape Shape; + static const int AssumeAliasing = 0; +}; + +template<> struct AssignmentKind { typedef Dense2Dense Kind; }; + + template struct unary_evaluator, IndexBased> : evaluator::PlainObject >::type @@ -355,10 +367,13 @@ template< typename DstXprType, typename ArgType, typename Scalar> struct Assignment, internal::assign_op, Dense2Dense, Scalar> { typedef Homogeneous SrcXprType; + // TODO clang generates garbage if this function is inlined. no valgrind error though. +#ifdef __clang__ + EIGEN_DONT_INLINE +#endif static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { dst.template topRows(src.nestedExpression().rows()) = src.nestedExpression(); -// dst.topRows(src.nestedExpression().rows()) = src.nestedExpression(); dst.row(dst.rows()-1).setOnes(); } }; @@ -368,16 +383,19 @@ template< typename DstXprType, typename ArgType, typename Scalar> struct Assignment, internal::assign_op, Dense2Dense, Scalar> { typedef Homogeneous SrcXprType; + // TODO clang generates garbage if this function is inlined. no valgrind error though. +#ifdef __clang__ + EIGEN_DONT_INLINE +#endif static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { dst.template leftCols(src.nestedExpression().cols()) = src.nestedExpression(); -// dst.leftCols(src.nestedExpression().cols()) = src.nestedExpression(); dst.col(dst.cols()-1).setOnes(); } }; template -struct generic_product_impl, Rhs, DenseShape, DenseShape, ProductTag> +struct generic_product_impl, Rhs, HomogeneousShape, DenseShape, ProductTag> { template static void evalTo(Dest& dst, const Homogeneous& lhs, const Rhs& rhs) @@ -387,88 +405,26 @@ struct generic_product_impl, Rhs, DenseShape, Den }; template -struct generic_product_impl, DenseShape, DenseShape, ProductTag> +struct generic_product_impl, DenseShape, HomogeneousShape, ProductTag> { template static void evalTo(Dest& dst, const Lhs& lhs, const Homogeneous& rhs) { - homogeneous_left_product_impl, Lhs>(rhs.nestedExpression(), lhs).evalTo(dst); + homogeneous_left_product_impl, Lhs>(lhs, rhs.nestedExpression()).evalTo(dst); } }; template -struct generic_product_impl, Homogeneous, DenseShape, DenseShape, ProductTag> +struct generic_product_impl, Homogeneous, DenseShape, HomogeneousShape, ProductTag> { typedef Transform TransformType; template static void evalTo(Dest& dst, const TransformType& lhs, const Homogeneous& rhs) { - homogeneous_left_product_impl, TransformType>(rhs.nestedExpression(), lhs).evalTo(dst); + homogeneous_left_product_impl, TransformType>(lhs, rhs.nestedExpression()).evalTo(dst); } }; - -template -struct product_evaluator, Rhs, DefaultProduct>, ProductTag, DenseShape, DenseShape, typename traits::Scalar, typename traits::Scalar> - : public evaluator, Rhs, DefaultProduct>::PlainObject>::type -{ - typedef Homogeneous Lhs; - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - -template -struct product_evaluator, DefaultProduct>, ProductTag, DenseShape, DenseShape, typename traits::Scalar, typename traits::Scalar> - : public evaluator, DefaultProduct>::PlainObject>::type -{ - typedef Homogeneous Rhs; - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - -template -struct product_evaluator, Homogeneous, DefaultProduct>, ProductTag, DenseShape, DenseShape, Scalar, typename traits::Scalar> - : public evaluator, Homogeneous, DefaultProduct>::PlainObject>::type -{ - typedef Transform Lhs; - typedef Homogeneous Rhs; - typedef Product XprType; - typedef typename XprType::PlainObject PlainObject; - typedef typename evaluator::type Base; - - product_evaluator(const XprType& xpr) - : m_result(xpr.rows(), xpr.cols()) - { - ::new (static_cast(this)) Base(m_result); - generic_product_impl::evalTo(m_result, xpr.lhs(), xpr.rhs()); - } - -protected: - PlainObject m_result; -}; - #endif // EIGEN_TEST_EVALUATORS } // end namespace internal diff --git a/test/geo_homogeneous.cpp b/test/geo_homogeneous.cpp index c91bde819..078894035 100644 --- a/test/geo_homogeneous.cpp +++ b/test/geo_homogeneous.cpp @@ -57,7 +57,6 @@ template void homogeneous(void) VERIFY_IS_APPROX((v0.transpose().rowwise().homogeneous().eval()) * t2, v0.transpose().rowwise().homogeneous() * t2); - m0.transpose().rowwise().homogeneous().eval(); VERIFY_IS_APPROX((m0.transpose().rowwise().homogeneous().eval()) * t2, m0.transpose().rowwise().homogeneous() * t2); @@ -82,7 +81,7 @@ template void homogeneous(void) VERIFY_IS_APPROX(aff * pts.colwise().homogeneous(), (aff * pts1).colwise().hnormalized()); VERIFY_IS_APPROX(caff * pts.colwise().homogeneous(), (caff * pts1).colwise().hnormalized()); VERIFY_IS_APPROX(proj * pts.colwise().homogeneous(), (proj * pts1)); - + VERIFY_IS_APPROX((aff * pts1).colwise().hnormalized(), aff * pts); VERIFY_IS_APPROX((caff * pts1).colwise().hnormalized(), caff * pts); From 26d2cdefd4dc291ba6613d5fd4127f6f1a6b5fa1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 31 Jul 2014 16:24:29 +0200 Subject: [PATCH 0673/1103] Fix 4x4 inverse via SSE for submatrices --- Eigen/src/LU/arch/Inverse_SSE.h | 11 ++++++++--- test/CMakeLists.txt | 8 +++----- test/inverse.cpp | 8 ++++++++ 3 files changed, 19 insertions(+), 8 deletions(-) diff --git a/Eigen/src/LU/arch/Inverse_SSE.h b/Eigen/src/LU/arch/Inverse_SSE.h index 60b7a2376..1f62ef14e 100644 --- a/Eigen/src/LU/arch/Inverse_SSE.h +++ b/Eigen/src/LU/arch/Inverse_SSE.h @@ -39,9 +39,11 @@ struct compute_inverse_size4 ResultAlignment = bool(ResultType::Flags&AlignedBit), StorageOrdersMatch = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit) }; + typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType; - static void run(const MatrixType& matrix, ResultType& result) + static void run(const MatrixType& mat, ResultType& result) { + ActualMatrixType matrix(mat); EIGEN_ALIGN16 const unsigned int _Sign_PNNP[4] = { 0x00000000, 0x80000000, 0x80000000, 0x00000000 }; // Load the full matrix into registers @@ -167,14 +169,17 @@ struct compute_inverse_size4 ResultAlignment = bool(ResultType::Flags&AlignedBit), StorageOrdersMatch = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit) }; - static void run(const MatrixType& matrix, ResultType& result) + typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType; + + static void run(const MatrixType& mat, ResultType& result) { + ActualMatrixType matrix(mat); const __m128d _Sign_NP = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0)); const __m128d _Sign_PN = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); // The inverse is calculated using "Divide and Conquer" technique. The // original matrix is divide into four 2x2 sub-matrices. Since each - // register of the matrix holds two element, the smaller matrices are + // register of the matrix holds two elements, the smaller matrices are // consisted of two registers. Hence we get a better locality of the // calculations. diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index c7614e1af..000d16c45 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -220,11 +220,9 @@ ei_add_test(geo_quaternion) ei_add_test(geo_eulerangles) ei_add_test(geo_parametrizedline) ei_add_test(geo_alignedbox) -if(NOT EIGEN_TEST_EVALUATORS) - ei_add_test(geo_hyperplane) - ei_add_test(geo_transformations) - ei_add_test(geo_homogeneous) -endif(NOT EIGEN_TEST_EVALUATORS) +ei_add_test(geo_hyperplane) +ei_add_test(geo_transformations) +ei_add_test(geo_homogeneous) ei_add_test(stdvector) ei_add_test(stdvector_overload) ei_add_test(stdlist) diff --git a/test/inverse.cpp b/test/inverse.cpp index 8187b088d..1195bcc76 100644 --- a/test/inverse.cpp +++ b/test/inverse.cpp @@ -68,6 +68,14 @@ template void inverse(const MatrixType& m) VERIFY_IS_MUCH_SMALLER_THAN(abs(det-m3.determinant()), RealScalar(1)); m3.computeInverseWithCheck(m4, invertible); VERIFY( rows==1 ? invertible : !invertible ); + + // check with submatrices + { + Matrix m3; + m3.setRandom(); + m2 = m3.template topLeftCorner().inverse(); + VERIFY_IS_APPROX( (m3.template topLeftCorner()), m2.inverse() ); + } #endif // check in-place inversion From d79516660c1f42ae0719ed353614d0977bd40153 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 31 Jul 2014 16:43:19 +0200 Subject: [PATCH 0674/1103] Make loadMarket use the sparse-matrix index type, thus enabling loading huge matrices. --- unsupported/Eigen/src/SparseExtra/MarketIO.h | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/src/SparseExtra/MarketIO.h b/unsupported/Eigen/src/SparseExtra/MarketIO.h index 1c40d3f7c..25ff4228d 100644 --- a/unsupported/Eigen/src/SparseExtra/MarketIO.h +++ b/unsupported/Eigen/src/SparseExtra/MarketIO.h @@ -133,6 +133,7 @@ template bool loadMarket(SparseMatrixType& mat, const std::string& filename) { typedef typename SparseMatrixType::Scalar Scalar; + typedef typename SparseMatrixType::Index Index; std::ifstream input(filename.c_str(),std::ios::in); if(!input) return false; @@ -142,11 +143,11 @@ bool loadMarket(SparseMatrixType& mat, const std::string& filename) bool readsizes = false; - typedef Triplet T; + typedef Triplet T; std::vector elements; - int M(-1), N(-1), NNZ(-1); - int count = 0; + Index M(-1), N(-1), NNZ(-1); + Index count = 0; while(input.getline(buffer, maxBuffersize)) { // skip comments @@ -169,7 +170,7 @@ bool loadMarket(SparseMatrixType& mat, const std::string& filename) } else { - int i(-1), j(-1); + Index i(-1), j(-1); Scalar value; if( internal::GetMarketLine(line, M, N, i, j, value) ) { From 647622281e5409e617854d35450afc0cd3a4dd49 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 31 Jul 2014 17:39:04 -0700 Subject: [PATCH 0675/1103] The tensor assignment code now resizes the destination tensor as needed. --- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index 547bb74d1..fdbe8df4c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -79,6 +79,7 @@ class Tensor : public TensorBase > }; static const int Options = Options_; + static const std::size_t NumIndices = NumIndices_; typedef DSizes Dimensions; @@ -232,11 +233,9 @@ class Tensor : public TensorBase > EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor& operator=(const Tensor& other) { - // FIXME: we need to resize the tensor to fix the dimensions of the other. - // Unfortunately this isn't possible yet when the rhs is an expression. - // resize(other.dimensions()); typedef TensorAssignOp Assign; Assign assign(*this, other); + resize(TensorEvaluator(assign, DefaultDevice()).dimensions()); internal::TensorExecutor::run(assign, DefaultDevice()); return *this; } @@ -244,11 +243,9 @@ class Tensor : public TensorBase > EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor& operator=(const OtherDerived& other) { - // FIXME: we need to resize the tensor to fix the dimensions of the other. - // Unfortunately this isn't possible yet when the rhs is an expression. - // resize(other.dimensions()); typedef TensorAssignOp Assign; Assign assign(*this, other); + resize(TensorEvaluator(assign, DefaultDevice()).dimensions()); internal::TensorExecutor::run(assign, DefaultDevice()); return *this; } From db76193bc7dba3c58503bc45f0c12f23ecd296e2 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Fri, 1 Aug 2014 14:41:49 +0200 Subject: [PATCH 0676/1103] Fix typo in PermutationMatrix (doc). --- Eigen/src/Core/PermutationMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 009873c78..8aa4c8bc5 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -262,7 +262,7 @@ class PermutationBase : public EigenBase * * \param SizeAtCompileTime the number of rows/cols, or Dynamic * \param MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it. - * \param StorageIndexType the interger type of the indices + * \param StorageIndexType the integer type of the indices * * This class represents a permutation matrix, internally stored as a vector of integers. * From fc13b37c552c6661c81f135cf94edc3de09d2000 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Aug 2014 14:47:33 +0200 Subject: [PATCH 0677/1103] Make cross product uses nested/nested_eval --- Eigen/src/Geometry/OrthoMethods.h | 30 ++++++++++++++++++++++++------ test/geo_orthomethods.cpp | 9 +++++++++ 2 files changed, 33 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Geometry/OrthoMethods.h b/Eigen/src/Geometry/OrthoMethods.h index 26be3ee5b..98e7a9bfc 100644 --- a/Eigen/src/Geometry/OrthoMethods.h +++ b/Eigen/src/Geometry/OrthoMethods.h @@ -30,8 +30,13 @@ MatrixBase::cross(const MatrixBase& other) const // Note that there is no need for an expression here since the compiler // optimize such a small temporary very well (even within a complex expression) +#ifndef EIGEN_TEST_EVALUATORS typename internal::nested::type lhs(derived()); typename internal::nested::type rhs(other.derived()); +#else + typename internal::nested_eval::type lhs(derived()); + typename internal::nested_eval::type rhs(other.derived()); +#endif return typename cross_product_return_type::type( numext::conj(lhs.coeff(1) * rhs.coeff(2) - lhs.coeff(2) * rhs.coeff(1)), numext::conj(lhs.coeff(2) * rhs.coeff(0) - lhs.coeff(0) * rhs.coeff(2)), @@ -76,8 +81,13 @@ MatrixBase::cross3(const MatrixBase& other) const EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Derived,4) EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,4) +#ifndef EIGEN_TEST_EVALUATORS typedef typename internal::nested::type DerivedNested; typedef typename internal::nested::type OtherDerivedNested; +#else + typedef typename internal::nested_eval::type DerivedNested; + typedef typename internal::nested_eval::type OtherDerivedNested; +#endif DerivedNested lhs(derived()); OtherDerivedNested rhs(other.derived()); @@ -103,21 +113,29 @@ VectorwiseOp::cross(const MatrixBase& ot EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,3) EIGEN_STATIC_ASSERT((internal::is_same::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + +#ifndef EIGEN_TEST_EVALUATORS + typename internal::nested::type mat(_expression()); + typename internal::nested::type vec(other.derived()); +#else + typename internal::nested_eval::type mat(_expression()); + typename internal::nested_eval::type vec(other.derived()); +#endif CrossReturnType res(_expression().rows(),_expression().cols()); if(Direction==Vertical) { eigen_assert(CrossReturnType::RowsAtCompileTime==3 && "the matrix must have exactly 3 rows"); - res.row(0) = (_expression().row(1) * other.coeff(2) - _expression().row(2) * other.coeff(1)).conjugate(); - res.row(1) = (_expression().row(2) * other.coeff(0) - _expression().row(0) * other.coeff(2)).conjugate(); - res.row(2) = (_expression().row(0) * other.coeff(1) - _expression().row(1) * other.coeff(0)).conjugate(); + res.row(0) = (mat.row(1) * vec.coeff(2) - mat.row(2) * vec.coeff(1)).conjugate(); + res.row(1) = (mat.row(2) * vec.coeff(0) - mat.row(0) * vec.coeff(2)).conjugate(); + res.row(2) = (mat.row(0) * vec.coeff(1) - mat.row(1) * vec.coeff(0)).conjugate(); } else { eigen_assert(CrossReturnType::ColsAtCompileTime==3 && "the matrix must have exactly 3 columns"); - res.col(0) = (_expression().col(1) * other.coeff(2) - _expression().col(2) * other.coeff(1)).conjugate(); - res.col(1) = (_expression().col(2) * other.coeff(0) - _expression().col(0) * other.coeff(2)).conjugate(); - res.col(2) = (_expression().col(0) * other.coeff(1) - _expression().col(1) * other.coeff(0)).conjugate(); + res.col(0) = (mat.col(1) * vec.coeff(2) - mat.col(2) * vec.coeff(1)).conjugate(); + res.col(1) = (mat.col(2) * vec.coeff(0) - mat.col(0) * vec.coeff(2)).conjugate(); + res.col(2) = (mat.col(0) * vec.coeff(1) - mat.col(1) * vec.coeff(0)).conjugate(); } return res; } diff --git a/test/geo_orthomethods.cpp b/test/geo_orthomethods.cpp index c836dae40..2b50c16e7 100644 --- a/test/geo_orthomethods.cpp +++ b/test/geo_orthomethods.cpp @@ -33,6 +33,7 @@ template void orthomethods_3() VERIFY_IS_MUCH_SMALLER_THAN(v1.dot(v1.cross(v2)), Scalar(1)); VERIFY_IS_MUCH_SMALLER_THAN(v1.cross(v2).dot(v2), Scalar(1)); VERIFY_IS_MUCH_SMALLER_THAN(v2.dot(v1.cross(v2)), Scalar(1)); + VERIFY_IS_MUCH_SMALLER_THAN(v1.cross(Vector3::Random()).dot(v1), Scalar(1)); Matrix3 mat3; mat3 << v0.normalized(), (v0.cross(v1)).normalized(), @@ -47,6 +48,13 @@ template void orthomethods_3() int i = internal::random(0,2); mcross = mat3.colwise().cross(vec3); VERIFY_IS_APPROX(mcross.col(i), mat3.col(i).cross(vec3)); + + VERIFY_IS_MUCH_SMALLER_THAN((mat3.transpose() * mat3.colwise().cross(vec3)).diagonal().cwiseAbs().sum(), Scalar(1)); + VERIFY_IS_MUCH_SMALLER_THAN((mat3.transpose() * mat3.colwise().cross(Vector3::Random())).diagonal().cwiseAbs().sum(), Scalar(1)); + + VERIFY_IS_MUCH_SMALLER_THAN((vec3.transpose() * mat3.colwise().cross(vec3)).cwiseAbs().sum(), Scalar(1)); + VERIFY_IS_MUCH_SMALLER_THAN((vec3.transpose() * Matrix3::Random().colwise().cross(vec3)).cwiseAbs().sum(), Scalar(1)); + mcross = mat3.rowwise().cross(vec3); VERIFY_IS_APPROX(mcross.row(i), mat3.row(i).cross(vec3)); @@ -57,6 +65,7 @@ template void orthomethods_3() v40.w() = v41.w() = v42.w() = 0; v42.template head<3>() = v40.template head<3>().cross(v41.template head<3>()); VERIFY_IS_APPROX(v40.cross3(v41), v42); + VERIFY_IS_MUCH_SMALLER_THAN(v40.cross3(Vector4::Random()).dot(v40), Scalar(1)); // check mixed product typedef Matrix RealVector3; From 2a3c3c49a19572b544b22820065c673cfac9bc04 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Aug 2014 14:48:22 +0200 Subject: [PATCH 0678/1103] Fix numerous nested versus nested_eval shortcomings --- Eigen/src/Core/Dot.h | 5 +++++ Eigen/src/Geometry/AlignedBox.h | 4 ++++ Eigen/src/LU/InverseImpl.h | 4 ++++ test/CMakeLists.txt | 4 ---- 4 files changed, 13 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h index d18b0099a..d6441c6a5 100644 --- a/Eigen/src/Core/Dot.h +++ b/Eigen/src/Core/Dot.h @@ -211,8 +211,13 @@ template bool MatrixBase::isOrthogonal (const MatrixBase& other, const RealScalar& prec) const { +#ifndef EIGEN_TEST_EVALUATORS typename internal::nested::type nested(derived()); typename internal::nested::type otherNested(other.derived()); +#else + typename internal::nested_eval::type nested(derived()); + typename internal::nested_eval::type otherNested(other.derived()); +#endif return numext::abs2(nested.dot(otherNested)) <= prec * prec * nested.squaredNorm() * otherNested.squaredNorm(); } diff --git a/Eigen/src/Geometry/AlignedBox.h b/Eigen/src/Geometry/AlignedBox.h index b6a2f0e24..1d1daaa61 100644 --- a/Eigen/src/Geometry/AlignedBox.h +++ b/Eigen/src/Geometry/AlignedBox.h @@ -71,7 +71,11 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim) template inline explicit AlignedBox(const MatrixBase& a_p) { +#ifndef EIGEN_TEST_EVALUATORS typename internal::nested::type p(a_p.derived()); +#else + typename internal::nested_eval::type p(a_p.derived()); +#endif m_min = p; m_max = p; } diff --git a/Eigen/src/LU/InverseImpl.h b/Eigen/src/LU/InverseImpl.h index 174dfbac5..e10fee48f 100644 --- a/Eigen/src/LU/InverseImpl.h +++ b/Eigen/src/LU/InverseImpl.h @@ -422,7 +422,11 @@ inline void MatrixBase::computeInverseAndDetWithCheck( // for larger sizes, evaluating has negligible cost and limits code size. typedef typename internal::conditional< RowsAtCompileTime == 2, +#ifndef EIGEN_TEST_EVALUATORS typename internal::remove_all::type>::type, +#else + typename internal::remove_all::type>::type, +#endif PlainObject >::type MatrixType; internal::compute_inverse_and_det_with_check::run diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 000d16c45..7555aaa38 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -251,8 +251,6 @@ ei_add_test(dense_storage) # # ei_add_test(denseLM) -if(NOT EIGEN_TEST_EVALUATORS) - if(QT4_FOUND) ei_add_test(qtvector "" "${QT_QTCORE_LIBRARY}") endif(QT4_FOUND) @@ -285,8 +283,6 @@ if(METIS_FOUND) ei_add_test(metis_support "" "${METIS_LIBRARIES}") endif() -endif(NOT EIGEN_TEST_EVALUATORS) - string(TOLOWER "${CMAKE_CXX_COMPILER}" cmake_cxx_compiler_tolower) if(cmake_cxx_compiler_tolower MATCHES "qcc") set(CXX_IS_QCC "ON") From 6f58a41097986d53d2ece3525550a7f7657d05ed Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Fri, 1 Aug 2014 15:35:45 +0200 Subject: [PATCH 0679/1103] Fix typos in Ref.h (doc). --- Eigen/src/Core/Ref.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index cd6d949c4..12c9c09f4 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -39,10 +39,10 @@ template& x); * \endcode * - * In the in-out case, the input argument must satisfies the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered. + * In the in-out case, the input argument must satisfy the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered. * By default, a Ref can reference any dense vector expression of float having a contiguous memory layout. - * Likewise, a Ref can reference any column major dense matrix expression of float whose column's elements are contiguously stored with - * the possibility to have a constant space inbetween each column, i.e.: the inner stride mmust be equal to 1, but the outer-stride (or leading dimension), + * Likewise, a Ref can reference any column-major dense matrix expression of float whose column's elements are contiguously stored with + * the possibility to have a constant space in-between each column, i.e. the inner stride must be equal to 1, but the outer stride (or leading dimension) * can be greater than the number of rows. * * In the const case, if the input expression does not match the above requirement, then it is evaluated into a temporary before being passed to the function. From c53f88297c8dc0a1622258df701bdc864ff9ee76 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Fri, 1 Aug 2014 15:43:47 +0200 Subject: [PATCH 0680/1103] Fix more typos in Ref.h (doc). --- Eigen/src/Core/Ref.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index 12c9c09f4..92614c6e2 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -19,17 +19,17 @@ template object can represent either a const expression or a l-value: * \code * // in-out argument: @@ -58,15 +58,15 @@ template > x); * foo3(A.row()); // OK * \endcode - * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involved more - * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overloads internally calling a + * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involve more + * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overload internally calling a * template function, e.g.: * \code * // in the .h: From c2ff44cbf3c58045a38120d28316a992c0cc0d57 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Aug 2014 16:23:30 +0200 Subject: [PATCH 0681/1103] Make assignment from general EigenBase object call evaluator, and support dense X= sparse --- Eigen/src/Core/EigenBase.h | 12 ++++++++++++ Eigen/src/SparseCore/SparseAssign.h | 25 +++++++++++++++++++++---- 2 files changed, 33 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/EigenBase.h b/Eigen/src/Core/EigenBase.h index 1a577c2dc..986e2a196 100644 --- a/Eigen/src/Core/EigenBase.h +++ b/Eigen/src/Core/EigenBase.h @@ -121,7 +121,11 @@ template template Derived& DenseBase::operator=(const EigenBase &other) { +#ifndef EIGEN_TEST_EVALUATORS other.derived().evalTo(derived()); +#else + call_assignment(derived(), other.derived()); +#endif return derived(); } @@ -129,7 +133,11 @@ template template Derived& DenseBase::operator+=(const EigenBase &other) { +#ifndef EIGEN_TEST_EVALUATORS other.derived().addTo(derived()); +#else + call_assignment(derived(), other.derived(), internal::add_assign_op()); +#endif return derived(); } @@ -137,7 +145,11 @@ template template Derived& DenseBase::operator-=(const EigenBase &other) { +#ifndef EIGEN_TEST_EVALUATORS other.derived().subTo(derived()); +#else + call_assignment(derived(), other.derived(), internal::sub_assign_op()); +#endif return derived(); } diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index c2091ee49..0a29cb32f 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -172,9 +172,9 @@ struct storage_kind_to_shape { struct Sparse2Sparse {}; struct Sparse2Dense {}; -template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; -template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; -template<> struct AssignmentKind { typedef Sparse2Dense Kind; }; +template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; +template<> struct AssignmentKind { typedef Sparse2Sparse Kind; }; +template<> struct AssignmentKind { typedef Sparse2Dense Kind; }; template @@ -252,7 +252,24 @@ struct Assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> struct Assignment { - static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) + static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + { + eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); + typedef typename SrcXprType::Index Index; + + typename internal::evaluator::type srcEval(src); + typename internal::evaluator::type dstEval(dst); + const Index outerEvaluationSize = (internal::evaluator::Flags&RowMajorBit) ? src.rows() : src.cols(); + for (Index j=0; j::InnerIterator i(srcEval,j); i; ++i) + func.assignCoeff(dstEval.coeffRef(i.row(),i.col()), i.value()); + } +}; + +template< typename DstXprType, typename SrcXprType, typename Scalar> +struct Assignment, Sparse2Dense, Scalar> +{ + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); typedef typename SrcXprType::Index Index; From 107bb308c376a5a4b7f11f2792396d2fa0d12904 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Aug 2014 16:24:23 +0200 Subject: [PATCH 0682/1103] Fix various small issues detected by gcc --- Eigen/src/Core/CoreEvaluators.h | 2 +- Eigen/src/Geometry/Transform.h | 2 +- Eigen/src/SparseCore/SparseBlock.h | 6 +++--- Eigen/src/SparseCore/SparseTriangularView.h | 2 +- Eigen/src/SparseQR/SparseQR.h | 4 ++-- 5 files changed, 8 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 66984e378..8c9190a0e 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -965,7 +965,7 @@ struct evaluator > typedef PartialReduxExpr XprType; typedef typename XprType::Scalar InputScalar; enum { - TraversalSize = Direction==Vertical ? ArgType::RowsAtCompileTime : XprType::ColsAtCompileTime + TraversalSize = Direction==int(Vertical) ? int(ArgType::RowsAtCompileTime) : int(XprType::ColsAtCompileTime) }; typedef typename MemberOp::template Cost CostOpType; enum { diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index a62548f12..bcf3e2723 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -375,7 +375,7 @@ public: #endif #ifdef EIGEN_TEST_EVALUATORS - Index rows() const { return Mode==Projective ? m_matrix.cols() : m_matrix.cols()-1; } + Index rows() const { return int(Mode)==int(Projective) ? m_matrix.cols() : (m_matrix.cols()-1); } Index cols() const { return m_matrix.cols(); } #endif diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 39e636437..21445b645 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -577,7 +577,7 @@ namespace internal { #ifdef EIGEN_TEST_EVALUATORS // -template +template struct unary_evaluator, IteratorBased > : public evaluator_base > { @@ -615,7 +615,7 @@ struct unary_evaluator, IteratorBa const XprType &m_block; }; -template +template class unary_evaluator, IteratorBased>::InnerVectorInnerIterator : public EvalIterator { @@ -640,7 +640,7 @@ public: inline operator bool() const { return EvalIterator::operator bool() && EvalIterator::index() < m_end; } }; -template +template class unary_evaluator, IteratorBased>::OuterVectorInnerIterator { const unary_evaluator& m_eval; diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index 7586a0a6e..87f4ab18d 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -200,7 +200,7 @@ public: public: EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& xprEval, Index outer) - : Base(xprEval.m_argImpl,outer) + : Base(xprEval.m_argImpl,outer), m_returnOne(false) { if(SkipFirst) { diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 4c6553bf2..e8d7b8607 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -178,7 +178,7 @@ class SparseQR y.resize((std::max)(cols(),Index(y.rows())),y.cols()); y.topRows(rank) = this->matrixR().topLeftCorner(rank, rank).template triangularView().solve(b.topRows(rank)); y.bottomRows(y.rows()-rank).setZero(); - + // Apply the column permutation if (m_perm_c.size()) dest = colsPermutation() * y.topRows(cols()); else dest = y.topRows(cols()); @@ -447,7 +447,7 @@ void SparseQR::factorize(const MatrixType& mat) } } // End update current column - Scalar tau; + Scalar tau = 0; RealScalar beta = 0; if(nonzeroCol < diagSize) From 51357a6622d30aa8cb799cf2ed69c96cf94007b7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Aug 2014 16:26:23 +0200 Subject: [PATCH 0683/1103] Fix geo_orthomethods unit test for complexes --- test/geo_orthomethods.cpp | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/test/geo_orthomethods.cpp b/test/geo_orthomethods.cpp index 2b50c16e7..7f8beb205 100644 --- a/test/geo_orthomethods.cpp +++ b/test/geo_orthomethods.cpp @@ -49,11 +49,11 @@ template void orthomethods_3() mcross = mat3.colwise().cross(vec3); VERIFY_IS_APPROX(mcross.col(i), mat3.col(i).cross(vec3)); - VERIFY_IS_MUCH_SMALLER_THAN((mat3.transpose() * mat3.colwise().cross(vec3)).diagonal().cwiseAbs().sum(), Scalar(1)); - VERIFY_IS_MUCH_SMALLER_THAN((mat3.transpose() * mat3.colwise().cross(Vector3::Random())).diagonal().cwiseAbs().sum(), Scalar(1)); + VERIFY_IS_MUCH_SMALLER_THAN((mat3.adjoint() * mat3.colwise().cross(vec3)).diagonal().cwiseAbs().sum(), Scalar(1)); + VERIFY_IS_MUCH_SMALLER_THAN((mat3.adjoint() * mat3.colwise().cross(Vector3::Random())).diagonal().cwiseAbs().sum(), Scalar(1)); - VERIFY_IS_MUCH_SMALLER_THAN((vec3.transpose() * mat3.colwise().cross(vec3)).cwiseAbs().sum(), Scalar(1)); - VERIFY_IS_MUCH_SMALLER_THAN((vec3.transpose() * Matrix3::Random().colwise().cross(vec3)).cwiseAbs().sum(), Scalar(1)); + VERIFY_IS_MUCH_SMALLER_THAN((vec3.adjoint() * mat3.colwise().cross(vec3)).cwiseAbs().sum(), Scalar(1)); + VERIFY_IS_MUCH_SMALLER_THAN((vec3.adjoint() * Matrix3::Random().colwise().cross(vec3)).cwiseAbs().sum(), Scalar(1)); mcross = mat3.rowwise().cross(vec3); VERIFY_IS_APPROX(mcross.row(i), mat3.row(i).cross(vec3)); From f25338f4d76dff13a9810996165ad7b465a15c88 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Aug 2014 16:49:44 +0200 Subject: [PATCH 0684/1103] Fix nesting of Homogenous evaluator --- Eigen/src/Geometry/Homogeneous.h | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index 7a3b5de80..8f3e5f419 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -351,6 +351,9 @@ struct unary_evaluator, IndexBased> typedef Homogeneous XprType; typedef typename XprType::PlainObject PlainObject; typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; unary_evaluator(const XprType& op) : Base(), m_temp(op) @@ -367,10 +370,6 @@ template< typename DstXprType, typename ArgType, typename Scalar> struct Assignment, internal::assign_op, Dense2Dense, Scalar> { typedef Homogeneous SrcXprType; - // TODO clang generates garbage if this function is inlined. no valgrind error though. -#ifdef __clang__ - EIGEN_DONT_INLINE -#endif static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { dst.template topRows(src.nestedExpression().rows()) = src.nestedExpression(); @@ -383,10 +382,6 @@ template< typename DstXprType, typename ArgType, typename Scalar> struct Assignment, internal::assign_op, Dense2Dense, Scalar> { typedef Homogeneous SrcXprType; - // TODO clang generates garbage if this function is inlined. no valgrind error though. -#ifdef __clang__ - EIGEN_DONT_INLINE -#endif static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { dst.template leftCols(src.nestedExpression().cols()) = src.nestedExpression(); From 4dd55a2958318f295cfc050563dedfe5dca384f0 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 1 Aug 2014 17:00:20 +0200 Subject: [PATCH 0685/1103] Optimize reduxions for Homogeneous --- Eigen/src/Core/Redux.h | 10 +++++----- Eigen/src/Geometry/Homogeneous.h | 7 +++++++ test/geo_homogeneous.cpp | 4 ++++ 3 files changed, 16 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 136cd165e..438ad018a 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -449,7 +449,7 @@ template EIGEN_STRONG_INLINE typename internal::traits::Scalar DenseBase::minCoeff() const { - return this->redux(Eigen::internal::scalar_min_op()); + return derived().redux(Eigen::internal::scalar_min_op()); } /** \returns the maximum of all coefficients of \c *this. @@ -459,7 +459,7 @@ template EIGEN_STRONG_INLINE typename internal::traits::Scalar DenseBase::maxCoeff() const { - return this->redux(Eigen::internal::scalar_max_op()); + return derived().redux(Eigen::internal::scalar_max_op()); } /** \returns the sum of all coefficients of *this @@ -472,7 +472,7 @@ DenseBase::sum() const { if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0)) return Scalar(0); - return this->redux(Eigen::internal::scalar_sum_op()); + return derived().redux(Eigen::internal::scalar_sum_op()); } /** \returns the mean of all coefficients of *this @@ -483,7 +483,7 @@ template EIGEN_STRONG_INLINE typename internal::traits::Scalar DenseBase::mean() const { - return Scalar(this->redux(Eigen::internal::scalar_sum_op())) / Scalar(this->size()); + return Scalar(derived().redux(Eigen::internal::scalar_sum_op())) / Scalar(this->size()); } /** \returns the product of all coefficients of *this @@ -499,7 +499,7 @@ DenseBase::prod() const { if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0)) return Scalar(1); - return this->redux(Eigen::internal::scalar_product_op()); + return derived().redux(Eigen::internal::scalar_product_op()); } /** \returns the trace of \c *this, i.e. the sum of the coefficients on the main diagonal. diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index 8f3e5f419..96909bb7e 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -138,6 +138,13 @@ template class Homogeneous } #endif + template + EIGEN_STRONG_INLINE typename internal::result_of::type + redux(const Func& func) const + { + return func(m_matrix.redux(func), Scalar(1)); + } + protected: typename MatrixType::Nested m_matrix; }; diff --git a/test/geo_homogeneous.cpp b/test/geo_homogeneous.cpp index 078894035..2f9d18c0f 100644 --- a/test/geo_homogeneous.cpp +++ b/test/geo_homogeneous.cpp @@ -38,6 +38,10 @@ template void homogeneous(void) hv0 << v0, 1; VERIFY_IS_APPROX(v0.homogeneous(), hv0); VERIFY_IS_APPROX(v0, hv0.hnormalized()); + + VERIFY_IS_APPROX(v0.homogeneous().sum(), hv0.sum()); + VERIFY_IS_APPROX(v0.homogeneous().minCoeff(), hv0.minCoeff()); + VERIFY_IS_APPROX(v0.homogeneous().maxCoeff(), hv0.maxCoeff()); hm0 << m0, ones.transpose(); VERIFY_IS_APPROX(m0.colwise().homogeneous(), hm0); From 3e59163a24c5ff1a8009887cb5d5e091ae6df540 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 2 Aug 2014 02:47:30 +0200 Subject: [PATCH 0686/1103] Fix bug #850: workaround MSVC 2008 weird compilation bug --- Eigen/src/Core/PlainObjectBase.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index f8292c793..69f34bd3e 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -708,7 +708,9 @@ class PlainObjectBase : public internal::dense_xpr_base::type EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if::value,T>::type* = 0) { - EIGEN_STATIC_ASSERT(bool(NumTraits::IsInteger), + // NOTE MSVC 2008 complains if we directly put bool(NumTraits::IsInteger) as the EIGEN_STATIC_ASSERT argument. + const bool is_integer = NumTraits::IsInteger; + EIGEN_STATIC_ASSERT(is_integer, FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) resize(size); } From 953ec08089ac7c5fc3c8f11236f3c546c88b5853 Mon Sep 17 00:00:00 2001 From: Kolja Brix Date: Sat, 2 Aug 2014 18:39:15 +0200 Subject: [PATCH 0687/1103] Correct GMRES: * Fix error in calculation of residual at restart. * Use relative residual as stopping criterion. * Improve documentation. --- .../Eigen/src/IterativeSolvers/GMRES.h | 156 +++++++++--------- 1 file changed, 80 insertions(+), 76 deletions(-) diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index c8c84069e..67498705b 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -11,7 +11,7 @@ #ifndef EIGEN_GMRES_H #define EIGEN_GMRES_H -namespace Eigen { +namespace Eigen { namespace internal { @@ -27,11 +27,11 @@ namespace internal { * \param iters on input: maximum number of iterations to perform * on output: number of iterations performed * \param restart number of iterations for a restart - * \param tol_error on input: residual tolerance + * \param tol_error on input: relative residual tolerance * on output: residuum achieved * - * \sa IterativeMethods::bicgstab() - * + * \sa IterativeMethods::bicgstab() + * * * For references, please see: * @@ -70,18 +70,24 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition const int m = mat.rows(); - VectorType p0 = rhs - mat*x; + // residual and preconditioned residual + const VectorType p0 = rhs - mat*x; VectorType r0 = precond.solve(p0); - + + const RealScalar r0Norm = r0.norm(); + // is initial guess already good enough? - if(abs(r0.norm()) < tol) { - return true; + if(r0Norm == 0) { + tol_error=0; + return true; } + // storage for Hessenberg matrix and Householder data + FMatrixType H = FMatrixType::Zero(m, restart + 1); VectorType w = VectorType::Zero(restart + 1); - - FMatrixType H = FMatrixType::Zero(m, restart + 1); // Hessenberg matrix VectorType tau = VectorType::Zero(restart + 1); + + // storage for Jacobi rotations std::vector < JacobiRotation < Scalar > > G(restart); // generate first Householder vector @@ -112,11 +118,10 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition } if (v.tail(m - k).norm() != 0.0) { - if (k <= restart) { // generate new Householder vector - VectorType e(m - k - 1); + VectorType e(m - k - 1); RealScalar beta; v.tail(m - k).makeHouseholder(e, tau.coeffRef(k), beta); H.col(k).tail(m - k - 1) = e; @@ -125,78 +130,77 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition v.tail(m - k).applyHouseholderOnTheLeft(H.col(k).tail(m - k - 1), tau.coeffRef(k), workspace.data()); } - } + } - if (k > 1) { - for (int i = 0; i < k - 1; ++i) { - // apply old Givens rotations to v - v.applyOnTheLeft(i, i + 1, G[i].adjoint()); - } - } + if (k > 1) { + for (int i = 0; i < k - 1; ++i) { + // apply old Givens rotations to v + v.applyOnTheLeft(i, i + 1, G[i].adjoint()); + } + } - if (k ().solveInPlace(y); + // solve upper triangular system + VectorType y = w.head(k); + H.topLeftCorner(k, k).template triangularView < Eigen::Upper > ().solveInPlace(y); - // use Horner-like scheme to calculate solution vector - VectorType x_new = y(k - 1) * VectorType::Unit(m, k - 1); + // use Horner-like scheme to calculate solution vector + VectorType x_new = y(k - 1) * VectorType::Unit(m, k - 1); - // apply Householder reflection H_{k} to x_new - x_new.tail(m - k + 1).applyHouseholderOnTheLeft(H.col(k - 1).tail(m - k), tau.coeffRef(k - 1), workspace.data()); + // apply Householder reflection H_{k} to x_new + x_new.tail(m - k + 1).applyHouseholderOnTheLeft(H.col(k - 1).tail(m - k), tau.coeffRef(k - 1), workspace.data()); - for (int i = k - 2; i >= 0; --i) { - x_new += y(i) * VectorType::Unit(m, i); - // apply Householder reflection H_{i} to x_new - x_new.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data()); - } + for (int i = k - 2; i >= 0; --i) { + x_new += y(i) * VectorType::Unit(m, i); + // apply Householder reflection H_{i} to x_new + x_new.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data()); + } - x += x_new; + x += x_new; - if (stop) { - return true; - } else { - k=0; + if (stop) { + return true; + } else { + k=0; - // reset data for a restart r0 = rhs - mat * x; - VectorType p0=mat*x; - VectorType p1=precond.solve(p0); - r0 = rhs - p1; -// r0_sqnorm = r0.squaredNorm(); - w = VectorType::Zero(restart + 1); - H = FMatrixType::Zero(m, restart + 1); - tau = VectorType::Zero(restart + 1); + // reset data for restart + const VectorType p0 = rhs - mat*x; + r0 = precond.solve(p0); - // generate first Householder vector - RealScalar beta; - r0.makeHouseholder(e, tau.coeffRef(0), beta); - w(0)=(Scalar) beta; - H.bottomLeftCorner(m - 1, 1) = e; + // clear Hessenberg matrix and Householder data + H = FMatrixType::Zero(m, restart + 1); + w = VectorType::Zero(restart + 1); + tau = VectorType::Zero(restart + 1); - } + // generate first Householder vector + RealScalar beta; + r0.makeHouseholder(e, tau.coeffRef(0), beta); + w(0)=(Scalar) beta; + H.bottomLeftCorner(m - 1, 1) = e; - } + } + } } - + return false; } @@ -230,7 +234,7 @@ struct traits > * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations() * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations * and NumTraits::epsilon() for the tolerance. - * + * * This class can be used as the direct solver classes. Here is a typical usage example: * \code * int n = 10000; @@ -244,7 +248,7 @@ struct traits > * // update b, and solve again * x = solver.solve(b); * \endcode - * + * * By default the iterations start with x=0 as an initial guess of the solution. * One can control the start using the solveWithGuess() method. Here is a step by * step execution example starting with a random guess and printing the evolution @@ -260,7 +264,7 @@ struct traits > * } while (solver.info()!=Success && i<100); * \endcode * Note that such a step by step excution is slightly slower. - * + * * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner */ template< typename _MatrixType, typename _Preconditioner> @@ -272,10 +276,10 @@ class GMRES : public IterativeSolverBase > using Base::m_iterations; using Base::m_info; using Base::m_isInitialized; - + private: int m_restart; - + public: typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; @@ -289,10 +293,10 @@ public: GMRES() : Base(), m_restart(30) {} /** Initialize the solver with matrix \a A for further \c Ax=b solving. - * + * * This constructor is a shortcut for the default constructor followed * by a call to compute(). - * + * * \warning this class stores a reference to the matrix A as well as some * precomputed values that depend on it. Therefore, if \a A is changed * this class becomes invalid. Call compute() to update it with the new @@ -301,16 +305,16 @@ public: GMRES(const MatrixType& A) : Base(A), m_restart(30) {} ~GMRES() {} - + /** Get the number of iterations after that a restart is performed. */ int get_restart() { return m_restart; } - + /** Set the number of iterations after that a restart is performed. * \param restart number of iterations for a restarti, default is 30. */ void set_restart(const int restart) { m_restart=restart; } - + /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A * \a x0 as an initial solution. * @@ -326,17 +330,17 @@ public: return internal::solve_retval_with_guess (*this, b.derived(), x0); } - + /** \internal */ template void _solveWithGuess(const Rhs& b, Dest& x) const - { + { bool failed = false; for(int j=0; j Date: Mon, 4 Aug 2014 12:46:00 +0200 Subject: [PATCH 0688/1103] Memory allocated on the stack is freed at the function exit, so reduce iteration count to avoid stack overflow --- test/dynalloc.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/dynalloc.cpp b/test/dynalloc.cpp index 4f53ea6f0..1190eb9cd 100644 --- a/test/dynalloc.cpp +++ b/test/dynalloc.cpp @@ -55,7 +55,7 @@ void check_aligned_new() void check_aligned_stack_alloc() { - for(int i = 1; i < 1000; i++) + for(int i = 1; i < 400; i++) { ei_declare_aligned_stack_constructed_variable(float,p,i,0); VERIFY(size_t(p)%ALIGNMENT==0); From 50085d2c28bcd4e3b8e1eb2e11dec60312cb8ac3 Mon Sep 17 00:00:00 2001 From: Silvio Traversaro Date: Thu, 7 Aug 2014 15:48:53 +0000 Subject: [PATCH 0689/1103] FindEigen3.cmake: Add reading hints of Eigen directory location from environment variables EIGEN3_ROOT and EIGEN3_ROOT_DIR . --- cmake/FindEigen3.cmake | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/cmake/FindEigen3.cmake b/cmake/FindEigen3.cmake index 9c546a05d..cea1afeab 100644 --- a/cmake/FindEigen3.cmake +++ b/cmake/FindEigen3.cmake @@ -9,6 +9,12 @@ # EIGEN3_FOUND - system has eigen lib with correct version # EIGEN3_INCLUDE_DIR - the eigen include directory # EIGEN3_VERSION - eigen version +# +# This module reads hints about search locations from +# the following enviroment variables: +# +# EIGEN3_ROOT +# EIGEN3_ROOT_DIR # Copyright (c) 2006, 2007 Montel Laurent, # Copyright (c) 2008, 2009 Gael Guennebaud, @@ -62,6 +68,9 @@ if (EIGEN3_INCLUDE_DIR) else (EIGEN3_INCLUDE_DIR) find_path(EIGEN3_INCLUDE_DIR NAMES signature_of_eigen3_matrix_library + HINTS + ENV EIGEN3_ROOT + ENV EIGEN3_ROOT_DIR PATHS ${CMAKE_INSTALL_PREFIX}/include ${KDE4_INCLUDE_DIR} From 305aa1f9c570db60a92e49020f21e70311bbed36 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 10:47:40 +0200 Subject: [PATCH 0690/1103] Add examples for hnormalized and homogenous (fix bug #846) --- Eigen/src/Geometry/Homogeneous.h | 4 ++-- doc/snippets/DirectionWise_hnormalized.cpp | 7 +++++++ doc/snippets/MatrixBase_hnormalized.cpp | 6 ++++++ doc/snippets/MatrixBase_homogeneous.cpp | 6 ++++++ doc/snippets/VectorwiseOp_homogeneous.cpp | 7 +++++++ 5 files changed, 28 insertions(+), 2 deletions(-) create mode 100644 doc/snippets/DirectionWise_hnormalized.cpp create mode 100644 doc/snippets/MatrixBase_hnormalized.cpp create mode 100644 doc/snippets/MatrixBase_homogeneous.cpp create mode 100644 doc/snippets/VectorwiseOp_homogeneous.cpp diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index 00e71d190..97dd21d15 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -120,7 +120,7 @@ template class Homogeneous * Example: \include MatrixBase_homogeneous.cpp * Output: \verbinclude MatrixBase_homogeneous.out * - * \sa class Homogeneous + * \sa VectorwiseOp::homogeneous(), class Homogeneous */ template inline typename MatrixBase::HomogeneousReturnType @@ -137,7 +137,7 @@ MatrixBase::homogeneous() const * Example: \include VectorwiseOp_homogeneous.cpp * Output: \verbinclude VectorwiseOp_homogeneous.out * - * \sa MatrixBase::homogeneous() */ + * \sa MatrixBase::homogeneous(), class Homogeneous */ template inline Homogeneous VectorwiseOp::homogeneous() const diff --git a/doc/snippets/DirectionWise_hnormalized.cpp b/doc/snippets/DirectionWise_hnormalized.cpp new file mode 100644 index 000000000..3410790a8 --- /dev/null +++ b/doc/snippets/DirectionWise_hnormalized.cpp @@ -0,0 +1,7 @@ +typedef Matrix Matrix4Xd; +Matrix4Xd M = Matrix4Xd::Random(4,5); +Projective3d P(Matrix4d::Random()); +cout << "The matrix M is:" << endl << M << endl << endl; +cout << "M.colwise().hnormalized():" << endl << M.colwise().hnormalized() << endl << endl; +cout << "P*M:" << endl << P*M << endl << endl; +cout << "(P*M).colwise().hnormalized():" << endl << (P*M).colwise().hnormalized() << endl << endl; \ No newline at end of file diff --git a/doc/snippets/MatrixBase_hnormalized.cpp b/doc/snippets/MatrixBase_hnormalized.cpp new file mode 100644 index 000000000..652cd77c0 --- /dev/null +++ b/doc/snippets/MatrixBase_hnormalized.cpp @@ -0,0 +1,6 @@ +Vector4d v = Vector4d::Random(); +Projective3d P(Matrix4d::Random()); +cout << "v = " << v.transpose() << "]^T" << endl; +cout << "v.hnormalized() = " << v.hnormalized().transpose() << "]^T" << endl; +cout << "P*v = " << (P*v).transpose() << "]^T" << endl; +cout << "(P*v).hnormalized() = " << (P*v).hnormalized().transpose() << "]^T" << endl; \ No newline at end of file diff --git a/doc/snippets/MatrixBase_homogeneous.cpp b/doc/snippets/MatrixBase_homogeneous.cpp new file mode 100644 index 000000000..457c28f91 --- /dev/null +++ b/doc/snippets/MatrixBase_homogeneous.cpp @@ -0,0 +1,6 @@ +Vector3d v = Vector3d::Random(), w; +Projective3d P(Matrix4d::Random()); +cout << "v = [" << v.transpose() << "]^T" << endl; +cout << "h.homogeneous() = [" << v.homogeneous().transpose() << "]^T" << endl; +cout << "(P * v.homogeneous()) = [" << (P * v.homogeneous()).transpose() << "]^T" << endl; +cout << "(P * v.homogeneous()).hnormalized() = [" << (P * v.homogeneous()).eval().hnormalized().transpose() << "]^T" << endl; \ No newline at end of file diff --git a/doc/snippets/VectorwiseOp_homogeneous.cpp b/doc/snippets/VectorwiseOp_homogeneous.cpp new file mode 100644 index 000000000..aba4fed0e --- /dev/null +++ b/doc/snippets/VectorwiseOp_homogeneous.cpp @@ -0,0 +1,7 @@ +typedef Matrix Matrix3Xd; +Matrix3Xd M = Matrix3Xd::Random(3,5); +Projective3d P(Matrix4d::Random()); +cout << "The matrix M is:" << endl << M << endl << endl; +cout << "M.colwise().homogeneous():" << endl << M.colwise().homogeneous() << endl << endl; +cout << "P * M.colwise().homogeneous():" << endl << P * M.colwise().homogeneous() << endl << endl; +cout << "P * M.colwise().homogeneous().hnormalized(): " << endl << (P * M.colwise().homogeneous()).colwise().hnormalized() << endl << endl; \ No newline at end of file From 3eb5253ca1352391f4ea31c4a2dd06c34c4a33e7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 14:41:14 +0200 Subject: [PATCH 0691/1103] Optimization: "matrix * real" did not call the special path and the real was converted to a complex. Add respective unit test to avoid future regression. --- Eigen/src/Core/MatrixBase.h | 1 + Eigen/src/Core/util/XprHelper.h | 10 +++++++++- test/linearstructure.cpp | 26 ++++++++++++++++++++++++++ 3 files changed, 36 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index f5987d194..3cb5e04fd 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -81,6 +81,7 @@ template class MatrixBase using Base::operator-=; using Base::operator*=; using Base::operator/=; + using Base::operator*; typedef typename Base::CoeffReturnType CoeffReturnType; typedef typename Base::ConstTransposeReturnType ConstTransposeReturnType; diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 1b3e122e1..7c77b2263 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -383,13 +383,21 @@ struct special_scalar_op_base : public DenseCo const CwiseUnaryOp, Derived> operator*(const OtherScalar& scalar) const { +#ifdef EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN + EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN +#endif return CwiseUnaryOp, Derived> (*static_cast(this), scalar_multiple2_op(scalar)); } inline friend const CwiseUnaryOp, Derived> operator*(const OtherScalar& scalar, const Derived& matrix) - { return static_cast(matrix).operator*(scalar); } + { +#ifdef EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN + EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN +#endif + return static_cast(matrix).operator*(scalar); + } }; template struct cast_return_type diff --git a/test/linearstructure.cpp b/test/linearstructure.cpp index 618984d5c..b627915ce 100644 --- a/test/linearstructure.cpp +++ b/test/linearstructure.cpp @@ -2,11 +2,16 @@ // for linear algebra. // // Copyright (C) 2006-2008 Benoit Jacob +// Copyright (C) 2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +static bool g_called; + +#define EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN { g_called = true; } + #include "main.h" template void linearStructure(const MatrixType& m) @@ -68,6 +73,24 @@ template void linearStructure(const MatrixType& m) VERIFY_IS_APPROX(m1.block(0,0,rows,cols) * s1, m1 * s1); } +// Make sure that complex * real and real * complex are properly optimized +template void real_complex(DenseIndex rows = MatrixType::RowsAtCompileTime, DenseIndex cols = MatrixType::ColsAtCompileTime) +{ + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + + RealScalar s = internal::random(); + MatrixType m1 = MatrixType::Random(rows, cols); + + g_called = false; + VERIFY_IS_APPROX(s*m1, Scalar(s)*m1); + VERIFY(g_called && "real * matrix not properly optimized"); + + g_called = false; + VERIFY_IS_APPROX(m1*s, m1*Scalar(s)); + VERIFY(g_called && "matrix * real not properly optimized"); +} + void test_linearstructure() { for(int i = 0; i < g_repeat; i++) { @@ -80,5 +103,8 @@ void test_linearstructure() CALL_SUBTEST_7( linearStructure(MatrixXi (internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); CALL_SUBTEST_8( linearStructure(MatrixXcd(internal::random(1,EIGEN_TEST_MAX_SIZE/2), internal::random(1,EIGEN_TEST_MAX_SIZE/2))) ); CALL_SUBTEST_9( linearStructure(ArrayXXf (internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); + + CALL_SUBTEST_10( real_complex() ); + CALL_SUBTEST_10( real_complex(10,10) ); } } From 18fbe7e7d4cdb737ef5775bbb32fe62b6f8ef70e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 14:49:23 +0200 Subject: [PATCH 0692/1103] Fix stableNorm() with respect to NaN and inf, and add respective unit tests. blueNorm() and hypotNorm() are broken wrt to NaN/inf --- Eigen/src/Core/StableNorm.h | 11 +++++- test/stable_norm.cpp | 69 ++++++++++++++++++++++++++++++++++++- 2 files changed, 78 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h index 525620bad..07a021505 100644 --- a/Eigen/src/Core/StableNorm.h +++ b/Eigen/src/Core/StableNorm.h @@ -20,7 +20,7 @@ inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& sc using std::max; Scalar maxCoeff = bl.cwiseAbs().maxCoeff(); - if (maxCoeff>scale) + if(maxCoeff>scale) { ssq = ssq * numext::abs2(scale/maxCoeff); Scalar tmp = Scalar(1)/maxCoeff; @@ -29,12 +29,21 @@ inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& sc invScale = NumTraits::highest(); scale = Scalar(1)/invScale; } + else if(maxCoeff>NumTraits::highest()) // we got a INF + { + invScale = Scalar(1); + scale = maxCoeff; + } else { scale = maxCoeff; invScale = tmp; } } + else if(maxCoeff!=maxCoeff) // we got a NaN + { + scale = maxCoeff; + } // TODO if the maxCoeff is much much smaller than the current scale, // then we can neglect this sub vector diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp index 549f91fbf..88cab7aa3 100644 --- a/test/stable_norm.cpp +++ b/test/stable_norm.cpp @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -14,6 +14,21 @@ template bool isNotNaN(const T& x) return x==x; } +template bool isNaN(const T& x) +{ + return x!=x; +} + +template bool isInf(const T& x) +{ + return x > NumTraits::highest(); +} + +template bool isMinusInf(const T& x) +{ + return x < NumTraits::lowest(); +} + // workaround aggressive optimization in ICC template EIGEN_DONT_INLINE T sub(T a, T b) { return a - b; } @@ -106,6 +121,58 @@ template void stable_norm(const MatrixType& m) VERIFY_IS_APPROX(vrand.rowwise().stableNorm(), vrand.rowwise().norm()); VERIFY_IS_APPROX(vrand.rowwise().blueNorm(), vrand.rowwise().norm()); VERIFY_IS_APPROX(vrand.rowwise().hypotNorm(), vrand.rowwise().norm()); + + // test NaN, +inf, -inf + MatrixType v; + Index i = internal::random(0,rows-1); + Index j = internal::random(0,cols-1); + + // NaN + { + v = vrand; + v(i,j) = RealScalar(0)/RealScalar(0); + VERIFY(!isFinite(v.squaredNorm())); VERIFY(isNaN(v.squaredNorm())); + VERIFY(!isFinite(v.norm())); VERIFY(isNaN(v.norm())); + VERIFY(!isFinite(v.stableNorm())); VERIFY(isNaN(v.stableNorm())); + VERIFY(!isFinite(v.blueNorm())); VERIFY(isNaN(v.blueNorm())); +// VERIFY(!isFinite(v.hypotNorm())); //VERIFY(isNaN(v.hypotNorm())); + } + + // +inf + { + v = vrand; + v(i,j) = RealScalar(1)/RealScalar(0); + VERIFY(!isFinite(v.squaredNorm())); VERIFY(isInf(v.squaredNorm())); + VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); + VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); +// VERIFY(!isFinite(v.blueNorm())); //VERIFY(isInf(v.blueNorm())); +// VERIFY(!isFinite(v.hypotNorm())); //VERIFY(isInf(v.hypotNorm())); + } + + // -inf + { + v = vrand; + v(i,j) = RealScalar(-1)/RealScalar(0); + VERIFY(!isFinite(v.squaredNorm())); VERIFY(isInf(v.squaredNorm())); + VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); + VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); +// VERIFY(!isFinite(v.blueNorm())); VERIFY(isInf(v.blueNorm())); +// VERIFY(!isFinite(v.hypotNorm())); VERIFY(isInf(v.hypotNorm())); + } + + // mix + { + Index i2 = internal::random(0,rows-1); + Index j2 = internal::random(0,cols-1); + v = vrand; + v(i,j) = RealScalar(-1)/RealScalar(0); + v(i2,j2) = RealScalar(0)/RealScalar(0); + VERIFY(!isFinite(v.squaredNorm())); VERIFY(isNaN(v.squaredNorm())); + VERIFY(!isFinite(v.norm())); VERIFY(isNaN(v.norm())); + VERIFY(!isFinite(v.stableNorm())); VERIFY(isNaN(v.stableNorm())); +// VERIFY(!isFinite(v.blueNorm())); //VERIFY(isNaN(v.blueNorm())); +// VERIFY(!isFinite(v.hypotNorm())); VERIFY(isNaN(v.hypotNorm())); + } } void test_stable_norm() From a44a343f034eb915f9ad6dcd69e4be534d48bbe5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 15:06:24 +0200 Subject: [PATCH 0693/1103] Fix blueNorm wrt NaN/INF. --- Eigen/src/Core/StableNorm.h | 11 +++++------ test/stable_norm.cpp | 6 +++--- 2 files changed, 8 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h index 07a021505..64d43e1b1 100644 --- a/Eigen/src/Core/StableNorm.h +++ b/Eigen/src/Core/StableNorm.h @@ -64,7 +64,7 @@ blueNorm_impl(const EigenBase& _vec) using std::abs; const Derived& vec(_vec.derived()); static bool initialized = false; - static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr; + static RealScalar b1, b2, s1m, s2m, rbig, relerr; if(!initialized) { int ibeta, it, iemin, iemax, iexp; @@ -93,7 +93,6 @@ blueNorm_impl(const EigenBase& _vec) iexp = - ((iemax+it)/2); s2m = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp))); // scaling factor for upper range - overfl = rbig*s2m; // overflow boundary for abig eps = RealScalar(pow(double(ibeta), 1-it)); relerr = sqrt(eps); // tolerance for neglecting asml initialized = true; @@ -110,13 +109,13 @@ blueNorm_impl(const EigenBase& _vec) else if(ax < b1) asml += numext::abs2(ax*s1m); else amed += numext::abs2(ax); } + if(amed!=amed) + return amed; // we got a NaN if(abig > RealScalar(0)) { abig = sqrt(abig); - if(abig > overfl) - { - return rbig; - } + if(abig > rbig) // overflow, or *this contains INF values + return abig; // return INF if(amed > RealScalar(0)) { abig = abig/s2m; diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp index 88cab7aa3..119b5b424 100644 --- a/test/stable_norm.cpp +++ b/test/stable_norm.cpp @@ -145,7 +145,7 @@ template void stable_norm(const MatrixType& m) VERIFY(!isFinite(v.squaredNorm())); VERIFY(isInf(v.squaredNorm())); VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); -// VERIFY(!isFinite(v.blueNorm())); //VERIFY(isInf(v.blueNorm())); + VERIFY(!isFinite(v.blueNorm())); VERIFY(isInf(v.blueNorm())); // VERIFY(!isFinite(v.hypotNorm())); //VERIFY(isInf(v.hypotNorm())); } @@ -156,7 +156,7 @@ template void stable_norm(const MatrixType& m) VERIFY(!isFinite(v.squaredNorm())); VERIFY(isInf(v.squaredNorm())); VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); -// VERIFY(!isFinite(v.blueNorm())); VERIFY(isInf(v.blueNorm())); + VERIFY(!isFinite(v.blueNorm())); VERIFY(isInf(v.blueNorm())); // VERIFY(!isFinite(v.hypotNorm())); VERIFY(isInf(v.hypotNorm())); } @@ -170,7 +170,7 @@ template void stable_norm(const MatrixType& m) VERIFY(!isFinite(v.squaredNorm())); VERIFY(isNaN(v.squaredNorm())); VERIFY(!isFinite(v.norm())); VERIFY(isNaN(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isNaN(v.stableNorm())); -// VERIFY(!isFinite(v.blueNorm())); //VERIFY(isNaN(v.blueNorm())); + VERIFY(!isFinite(v.blueNorm())); VERIFY(isNaN(v.blueNorm())); // VERIFY(!isFinite(v.hypotNorm())); VERIFY(isNaN(v.hypotNorm())); } } From 42e27d41a2014043799b44b68b0d5644629279ba Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 16:09:39 +0200 Subject: [PATCH 0694/1103] Fix hypot() and hypotNorm() wrt NaN and INF values. --- Eigen/src/Core/MathFunctions.h | 15 +++++++++++---- Eigen/src/Core/functors/BinaryFunctors.h | 14 +++++++++++--- test/stable_norm.cpp | 8 ++++---- 3 files changed, 26 insertions(+), 11 deletions(-) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 20fc2be74..5df1fbfec 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -308,10 +308,17 @@ struct hypot_impl using std::sqrt; RealScalar _x = abs(x); RealScalar _y = abs(y); - RealScalar p = (max)(_x, _y); - if(p==RealScalar(0)) return 0; - RealScalar q = (min)(_x, _y); - RealScalar qp = q/p; + Scalar p, qp; + if(_x>_y) + { + p = _x; + qp = _y / p; + } + else + { + p = _y; + qp = _x / p; + } return p * sqrt(RealScalar(1) + qp*qp); } }; diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h index ba094f5d1..157d075a7 100644 --- a/Eigen/src/Core/functors/BinaryFunctors.h +++ b/Eigen/src/Core/functors/BinaryFunctors.h @@ -167,9 +167,17 @@ template struct scalar_hypot_op { EIGEN_USING_STD_MATH(max); EIGEN_USING_STD_MATH(min); using std::sqrt; - Scalar p = (max)(_x, _y); - Scalar q = (min)(_x, _y); - Scalar qp = q/p; + Scalar p, qp; + if(_x>_y) + { + p = _x; + qp = _y / p; + } + else + { + p = _y; + qp = _x / p; + } return p * sqrt(Scalar(1) + qp*qp); } }; diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp index 119b5b424..f76919af4 100644 --- a/test/stable_norm.cpp +++ b/test/stable_norm.cpp @@ -135,7 +135,7 @@ template void stable_norm(const MatrixType& m) VERIFY(!isFinite(v.norm())); VERIFY(isNaN(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isNaN(v.stableNorm())); VERIFY(!isFinite(v.blueNorm())); VERIFY(isNaN(v.blueNorm())); -// VERIFY(!isFinite(v.hypotNorm())); //VERIFY(isNaN(v.hypotNorm())); + VERIFY(!isFinite(v.hypotNorm())); VERIFY(isNaN(v.hypotNorm())); } // +inf @@ -146,7 +146,7 @@ template void stable_norm(const MatrixType& m) VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); VERIFY(!isFinite(v.blueNorm())); VERIFY(isInf(v.blueNorm())); -// VERIFY(!isFinite(v.hypotNorm())); //VERIFY(isInf(v.hypotNorm())); + VERIFY(!isFinite(v.hypotNorm())); VERIFY(isInf(v.hypotNorm())); } // -inf @@ -157,7 +157,7 @@ template void stable_norm(const MatrixType& m) VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); VERIFY(!isFinite(v.blueNorm())); VERIFY(isInf(v.blueNorm())); -// VERIFY(!isFinite(v.hypotNorm())); VERIFY(isInf(v.hypotNorm())); + VERIFY(!isFinite(v.hypotNorm())); VERIFY(isInf(v.hypotNorm())); } // mix @@ -171,7 +171,7 @@ template void stable_norm(const MatrixType& m) VERIFY(!isFinite(v.norm())); VERIFY(isNaN(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isNaN(v.stableNorm())); VERIFY(!isFinite(v.blueNorm())); VERIFY(isNaN(v.blueNorm())); -// VERIFY(!isFinite(v.hypotNorm())); VERIFY(isNaN(v.hypotNorm())); + VERIFY(!isFinite(v.hypotNorm())); VERIFY(isNaN(v.hypotNorm())); } } From ff9bfc45f74249a8aabf10df6be8ac1bfcc9ee5e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 17:10:17 +0200 Subject: [PATCH 0695/1103] relax some LM unit tests --- unsupported/test/NonLinearOptimization.cpp | 17 +++++---- unsupported/test/levenberg_marquardt.cpp | 44 ++++++++++++---------- 2 files changed, 33 insertions(+), 28 deletions(-) diff --git a/unsupported/test/NonLinearOptimization.cpp b/unsupported/test/NonLinearOptimization.cpp index d7376b0f5..702bb7c60 100644 --- a/unsupported/test/NonLinearOptimization.cpp +++ b/unsupported/test/NonLinearOptimization.cpp @@ -1262,8 +1262,8 @@ void testNistBoxBOD(void) // check return value VERIFY_IS_EQUAL(info, 1); - VERIFY_IS_EQUAL(lm.nfev, 31); - VERIFY_IS_EQUAL(lm.njev, 25); + VERIFY(lm.nfev < 31); // 31 + VERIFY(lm.njev < 25); // 25 // check norm^2 VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 1.1680088766E+03); // check x @@ -1342,10 +1342,6 @@ void testNistMGH17(void) lm.parameters.maxfev = 1000; info = lm.minimize(x); - // check return value - VERIFY_IS_EQUAL(info, 2); - VERIFY_IS_EQUAL(lm.nfev, 602 ); - VERIFY_IS_EQUAL(lm.njev, 545 ); // check norm^2 VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 5.4648946975E-05); // check x @@ -1354,6 +1350,11 @@ void testNistMGH17(void) VERIFY_IS_APPROX(x[2], -1.4646871366E+00); VERIFY_IS_APPROX(x[3], 1.2867534640E-02); VERIFY_IS_APPROX(x[4], 2.2122699662E-02); + + // check return value + VERIFY_IS_EQUAL(info, 2); + VERIFY(lm.nfev < 650); // 602 + VERIFY(lm.njev < 600); // 545 /* * Second try @@ -1832,8 +1833,8 @@ void test_NonLinearOptimization() // NIST tests, level of difficulty = "Average" CALL_SUBTEST/*_5*/(testNistHahn1()); CALL_SUBTEST/*_6*/(testNistMisra1d()); -// CALL_SUBTEST/*_7*/(testNistMGH17()); -// CALL_SUBTEST/*_8*/(testNistLanczos1()); + CALL_SUBTEST/*_7*/(testNistMGH17()); + CALL_SUBTEST/*_8*/(testNistLanczos1()); // // NIST tests, level of difficulty = "Higher" CALL_SUBTEST/*_9*/(testNistRat42()); diff --git a/unsupported/test/levenberg_marquardt.cpp b/unsupported/test/levenberg_marquardt.cpp index 04464727d..1fa1c3c22 100644 --- a/unsupported/test/levenberg_marquardt.cpp +++ b/unsupported/test/levenberg_marquardt.cpp @@ -787,16 +787,17 @@ void testNistMGH10(void) LevenbergMarquardt lm(functor); info = lm.minimize(x); - // check return value - VERIFY_IS_EQUAL(info, 1); - VERIFY_IS_EQUAL(lm.nfev(), 284 ); - VERIFY_IS_EQUAL(lm.njev(), 249 ); // check norm^2 VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 8.7945855171E+01); // check x VERIFY_IS_APPROX(x[0], 5.6096364710E-03); VERIFY_IS_APPROX(x[1], 6.1813463463E+03); VERIFY_IS_APPROX(x[2], 3.4522363462E+02); + + // check return value + //VERIFY_IS_EQUAL(info, 1); + VERIFY_IS_EQUAL(lm.nfev(), 284 ); + VERIFY_IS_EQUAL(lm.njev(), 249 ); /* * Second try @@ -805,16 +806,17 @@ void testNistMGH10(void) // do the computation info = lm.minimize(x); - // check return value - VERIFY_IS_EQUAL(info, 1); - VERIFY_IS_EQUAL(lm.nfev(), 126); - VERIFY_IS_EQUAL(lm.njev(), 116); // check norm^2 VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 8.7945855171E+01); // check x VERIFY_IS_APPROX(x[0], 5.6096364710E-03); VERIFY_IS_APPROX(x[1], 6.1813463463E+03); VERIFY_IS_APPROX(x[2], 3.4522363462E+02); + + // check return value + //VERIFY_IS_EQUAL(info, 1); + VERIFY_IS_EQUAL(lm.nfev(), 126); + VERIFY_IS_EQUAL(lm.njev(), 116); } @@ -866,15 +868,16 @@ void testNistBoxBOD(void) lm.setFactor(10); info = lm.minimize(x); - // check return value - VERIFY_IS_EQUAL(info, 1); - VERIFY_IS_EQUAL(lm.nfev(), 31); - VERIFY_IS_EQUAL(lm.njev(), 25); // check norm^2 VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 1.1680088766E+03); // check x VERIFY_IS_APPROX(x[0], 2.1380940889E+02); VERIFY_IS_APPROX(x[1], 5.4723748542E-01); + + // check return value + VERIFY_IS_EQUAL(info, 1); + VERIFY(lm.nfev() < 31); // 31 + VERIFY(lm.njev() < 25); // 25 /* * Second try @@ -948,10 +951,6 @@ void testNistMGH17(void) lm.setMaxfev(1000); info = lm.minimize(x); - // check return value -// VERIFY_IS_EQUAL(info, 2); //FIXME Use (lm.info() == Success) -// VERIFY_IS_EQUAL(lm.nfev(), 602 ); - VERIFY_IS_EQUAL(lm.njev(), 545 ); // check norm^2 VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 5.4648946975E-05); // check x @@ -960,6 +959,11 @@ void testNistMGH17(void) VERIFY_IS_APPROX(x[2], -1.4646871366E+00); VERIFY_IS_APPROX(x[3], 1.2867534640E-02); VERIFY_IS_APPROX(x[4], 2.2122699662E-02); + + // check return value +// VERIFY_IS_EQUAL(info, 2); //FIXME Use (lm.info() == Success) + VERIFY(lm.nfev() < 700 ); // 602 + VERIFY(lm.njev() < 600 ); // 545 /* * Second try @@ -1035,10 +1039,6 @@ void testNistMGH09(void) lm.setMaxfev(1000); info = lm.minimize(x); - // check return value - VERIFY_IS_EQUAL(info, 1); - VERIFY_IS_EQUAL(lm.nfev(), 490 ); - VERIFY_IS_EQUAL(lm.njev(), 376 ); // check norm^2 VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 3.0750560385E-04); // check x @@ -1046,6 +1046,10 @@ void testNistMGH09(void) VERIFY_IS_APPROX(x[1], 0.19126423573); // should be 1.9128232873E-01 VERIFY_IS_APPROX(x[2], 0.12305309914); // should be 1.2305650693E-01 VERIFY_IS_APPROX(x[3], 0.13605395375); // should be 1.3606233068E-01 + // check return value + VERIFY_IS_EQUAL(info, 1); + VERIFY(lm.nfev() < 510 ); // 490 + VERIFY(lm.njev() < 400 ); // 376 /* * Second try From 280661e67d053b4e7f3358b160bcc11f76704be9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 17:29:06 +0200 Subject: [PATCH 0696/1103] Remove LM::sqrt_() member function in favor of a shortcut for sqrt(epsilon()) --- .../LevenbergMarquardt.h | 19 ++++++++++++------- 1 file changed, 12 insertions(+), 7 deletions(-) diff --git a/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h b/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h index ecb8dccf4..69106ddc5 100644 --- a/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h +++ b/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h @@ -45,18 +45,24 @@ namespace LevenbergMarquardtSpace { template class LevenbergMarquardt { + static Scalar sqrt_epsilon() + { + using std::sqrt; + return sqrt(NumTraits::epsilon()); + } + public: LevenbergMarquardt(FunctorType &_functor) : functor(_functor) { nfev = njev = iter = 0; fnorm = gnorm = 0.; useExternalScaling=false; } typedef DenseIndex Index; - + struct Parameters { Parameters() : factor(Scalar(100.)) , maxfev(400) - , ftol(sqrt_(NumTraits::epsilon())) - , xtol(sqrt_(NumTraits::epsilon())) + , ftol(sqrt_epsilon()) + , xtol(sqrt_epsilon()) , gtol(Scalar(0.)) , epsfcn(Scalar(0.)) {} Scalar factor; @@ -72,7 +78,7 @@ public: LevenbergMarquardtSpace::Status lmder1( FVectorType &x, - const Scalar tol = sqrt_(NumTraits::epsilon()) + const Scalar tol = sqrt_epsilon() ); LevenbergMarquardtSpace::Status minimize(FVectorType &x); @@ -83,12 +89,12 @@ public: FunctorType &functor, FVectorType &x, Index *nfev, - const Scalar tol = sqrt_(NumTraits::epsilon()) + const Scalar tol = sqrt_epsilon() ); LevenbergMarquardtSpace::Status lmstr1( FVectorType &x, - const Scalar tol = sqrt_(NumTraits::epsilon()) + const Scalar tol = sqrt_epsilon() ); LevenbergMarquardtSpace::Status minimizeOptimumStorage(FVectorType &x); @@ -109,7 +115,6 @@ public: Scalar lm_param(void) { return par; } private: - static Scalar sqrt_(const Scalar& x) { using std::sqrt; return sqrt(x); } FunctorType &functor; Index n; From 60314beb38919a87b2b605784a5c3a33dec3b895 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 17:35:11 +0200 Subject: [PATCH 0697/1103] Update reference value for testNistLanczos1 test --- unsupported/test/NonLinearOptimization.cpp | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/unsupported/test/NonLinearOptimization.cpp b/unsupported/test/NonLinearOptimization.cpp index 702bb7c60..75974f84f 100644 --- a/unsupported/test/NonLinearOptimization.cpp +++ b/unsupported/test/NonLinearOptimization.cpp @@ -1022,7 +1022,8 @@ void testNistLanczos1(void) VERIFY_IS_EQUAL(lm.nfev, 79); VERIFY_IS_EQUAL(lm.njev, 72); // check norm^2 - VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 1.430899764097e-25); // should be 1.4307867721E-25, but nist results are on 128-bit floats + std::cout.precision(30); + VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 1.4290986055242372e-25); // should be 1.4307867721E-25, but nist results are on 128-bit floats // check x VERIFY_IS_APPROX(x[0], 9.5100000027E-02); VERIFY_IS_APPROX(x[1], 1.0000000001E+00); @@ -1043,7 +1044,7 @@ void testNistLanczos1(void) VERIFY_IS_EQUAL(lm.nfev, 9); VERIFY_IS_EQUAL(lm.njev, 8); // check norm^2 - VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 1.428595533845e-25); // should be 1.4307867721E-25, but nist results are on 128-bit floats + VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 1.430571737783119393e-25); // should be 1.4307867721E-25, but nist results are on 128-bit floats // check x VERIFY_IS_APPROX(x[0], 9.5100000027E-02); VERIFY_IS_APPROX(x[1], 1.0000000001E+00); From 25bceefb4e64123af88bb8ec6c74b5436fa4130b Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sat, 6 Sep 2014 11:47:24 +0100 Subject: [PATCH 0698/1103] Replace asm by __asm__ (bug #873) --- Eigen/src/Core/arch/AVX/PacketMath.h | 4 ++-- Eigen/src/Core/arch/NEON/PacketMath.h | 2 +- Eigen/src/Core/util/Macros.h | 2 +- 3 files changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 66b97bd69..01730c5ee 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -141,7 +141,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate // the result of the product. Packet8f res = c; - asm("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); + __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); return res; #else return _mm256_fmadd_ps(a,b,c); @@ -151,7 +151,7 @@ template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& #if defined(__clang__) || defined(__GNUC__) // see above Packet4d res = c; - asm("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); + __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); return res; #else return _mm256_fmadd_pd(a,b,c); diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 7c4509585..0504c095c 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -57,7 +57,7 @@ typedef uint32x4_t Packet4ui; #elif defined __pld #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR) #elif !defined(__aarch64__) - #define EIGEN_ARM_PREFETCH(ADDR) asm volatile ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); + #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" ); #else // by default no explicit prefetching #define EIGEN_ARM_PREFETCH(ADDR) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 99e682653..197288e09 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -279,7 +279,7 @@ namespace Eigen { #if !defined(EIGEN_ASM_COMMENT) #if (defined __GNUC__) && ( defined(__i386__) || defined(__x86_64__) ) - #define EIGEN_ASM_COMMENT(X) asm("#" X) + #define EIGEN_ASM_COMMENT(X) __asm__("#" X) #else #define EIGEN_ASM_COMMENT(X) #endif From abb33258ce52a8cc3b540fb7cafb1812d9b71dd9 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sat, 6 Sep 2014 14:59:44 +0100 Subject: [PATCH 0699/1103] Doc: difference between array and matrix cosine etc (bug #830) --- Eigen/src/plugins/ArrayCwiseUnaryOps.h | 18 ++++++++++++++++++ unsupported/Eigen/MatrixFunctions | 21 +++++++++++++++++---- 2 files changed, 35 insertions(+), 4 deletions(-) diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index ce462e951..f6d7d8944 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -29,6 +29,9 @@ abs2() const } /** \returns an expression of the coefficient-wise exponential of *this. + * + * This function computes the coefficient-wise exponential. The function MatrixBase::exp() in the + * unsupported module MatrixFunctions computes the matrix exponential. * * Example: \include Cwise_exp.cpp * Output: \verbinclude Cwise_exp.out @@ -43,6 +46,9 @@ exp() const } /** \returns an expression of the coefficient-wise logarithm of *this. + * + * This function computes the coefficient-wise logarithm. The function MatrixBase::log() in the + * unsupported module MatrixFunctions computes the matrix logarithm. * * Example: \include Cwise_log.cpp * Output: \verbinclude Cwise_log.out @@ -57,6 +63,9 @@ log() const } /** \returns an expression of the coefficient-wise square root of *this. + * + * This function computes the coefficient-wise square root. The function MatrixBase::sqrt() in the + * unsupported module MatrixFunctions computes the matrix square root. * * Example: \include Cwise_sqrt.cpp * Output: \verbinclude Cwise_sqrt.out @@ -71,6 +80,9 @@ sqrt() const } /** \returns an expression of the coefficient-wise cosine of *this. + * + * This function computes the coefficient-wise cosine. The function MatrixBase::cos() in the + * unsupported module MatrixFunctions computes the matrix cosine. * * Example: \include Cwise_cos.cpp * Output: \verbinclude Cwise_cos.out @@ -86,6 +98,9 @@ cos() const /** \returns an expression of the coefficient-wise sine of *this. + * + * This function computes the coefficient-wise sine. The function MatrixBase::sin() in the + * unsupported module MatrixFunctions computes the matrix sine. * * Example: \include Cwise_sin.cpp * Output: \verbinclude Cwise_sin.out @@ -155,6 +170,9 @@ atan() const } /** \returns an expression of the coefficient-wise power of *this to the given exponent. + * + * This function computes the coefficient-wise power. The function MatrixBase::pow() in the + * unsupported module MatrixFunctions computes the matrix power. * * Example: \include Cwise_pow.cpp * Output: \verbinclude Cwise_pow.out diff --git a/unsupported/Eigen/MatrixFunctions b/unsupported/Eigen/MatrixFunctions index 0b12aaffb..0320606c1 100644 --- a/unsupported/Eigen/MatrixFunctions +++ b/unsupported/Eigen/MatrixFunctions @@ -82,7 +82,9 @@ const MatrixFunctionReturnValue MatrixBase::cos() const \param[in] M a square matrix. \returns expression representing \f$ \cos(M) \f$. -This function calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::cos(). +This function computes the matrix cosine. Use ArrayBase::cos() for computing the entry-wise cosine. + +The implementation calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::cos(). \sa \ref matrixbase_sin "sin()" for an example. @@ -123,6 +125,9 @@ differential equations: the solution of \f$ y' = My \f$ with the initial condition \f$ y(0) = y_0 \f$ is given by \f$ y(t) = \exp(M) y_0 \f$. +The matrix exponential is different from applying the exp function to all the entries in the matrix. +Use ArrayBase::exp() if you want to do the latter. + The cost of the computation is approximately \f$ 20 n^3 \f$ for matrices of size \f$ n \f$. The number 20 depends weakly on the norm of the matrix. @@ -177,6 +182,9 @@ the scalar logarithm, the equation \f$ \exp(X) = M \f$ may have multiple solutions; this function returns a matrix whose eigenvalues have imaginary part in the interval \f$ (-\pi,\pi] \f$. +The matrix logarithm is different from applying the log function to all the entries in the matrix. +Use ArrayBase::log() if you want to do the latter. + In the real case, the matrix \f$ M \f$ should be invertible and it should have no eigenvalues which are real and negative (pairs of complex conjugate eigenvalues are allowed). In the complex case, it @@ -232,7 +240,8 @@ const MatrixPowerReturnValue MatrixBase::pow(RealScalar p) con The matrix power \f$ M^p \f$ is defined as \f$ \exp(p \log(M)) \f$, where exp denotes the matrix exponential, and log denotes the matrix -logarithm. +logarithm. This is different from raising all the entries in the matrix +to the p-th power. Use ArrayBase::pow() if you want to do the latter. If \p p is complex, the scalar type of \p M should be the type of \p p . \f$ M^p \f$ simply evaluates into \f$ \exp(p \log(M)) \f$. @@ -391,7 +400,9 @@ const MatrixFunctionReturnValue MatrixBase::sin() const \param[in] M a square matrix. \returns expression representing \f$ \sin(M) \f$. -This function calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::sin(). +This function computes the matrix sine. Use ArrayBase::sin() for computing the entry-wise sine. + +The implementation calls \ref matrixbase_matrixfunction "matrixFunction()" with StdStemFunctions::sin(). Example: \include MatrixSine.cpp Output: \verbinclude MatrixSine.out @@ -428,7 +439,9 @@ const MatrixSquareRootReturnValue MatrixBase::sqrt() const The matrix square root of \f$ M \f$ is the matrix \f$ M^{1/2} \f$ whose square is the original matrix; so if \f$ S = M^{1/2} \f$ then -\f$ S^2 = M \f$. +\f$ S^2 = M \f$. This is different from taking the square root of all +the entries in the matrix; use ArrayBase::sqrt() if you want to do the +latter. In the real case, the matrix \f$ M \f$ should be invertible and it should have no eigenvalues which are real and negative (pairs of From fafc829424894d8cbb88f52e56cf074a351d92a5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 7 Sep 2014 22:38:09 +0200 Subject: [PATCH 0700/1103] bug #804: copy group__TopicUnalignedArrayAssert.html to TopicUnalignedArrayAssert.html as the second is linked to by old Eigen versions. --- doc/CMakeLists.txt | 1 + 1 file changed, 1 insertion(+) diff --git a/doc/CMakeLists.txt b/doc/CMakeLists.txt index 1b8aaf9aa..46e5fc9d7 100644 --- a/doc/CMakeLists.txt +++ b/doc/CMakeLists.txt @@ -97,6 +97,7 @@ add_dependencies(doc-unsupported-prerequisites unsupported_snippets unsupported_ add_custom_target(doc ALL COMMAND doxygen COMMAND doxygen Doxyfile-unsupported + COMMAND ${CMAKE_COMMAND} -E copy ${Eigen_BINARY_DIR}/doc/html/group__TopicUnalignedArrayAssert.html ${Eigen_BINARY_DIR}/doc/html/TopicUnalignedArrayAssert.html COMMAND ${CMAKE_COMMAND} -E rename html eigen-doc COMMAND ${CMAKE_COMMAND} -E remove eigen-doc/eigen-doc.tgz COMMAND ${CMAKE_COMMAND} -E tar cfz eigen-doc/eigen-doc.tgz eigen-doc From e54898f53e73b0cb3cc0d993f7816d14efe59fa6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 7 Sep 2014 23:02:30 +0200 Subject: [PATCH 0701/1103] bug #619: workaround MSVC 2008 implementing std::abs for int only on WINCE --- Eigen/src/Core/MathFunctions.h | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 5df1fbfec..e9fed2e52 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -12,6 +12,15 @@ namespace Eigen { +// On WINCE, std::abs is defined for int only, so let's defined our own overloads: +// This issue has been confirmed with MSVC 2008 only, but the issue might exist for more recent versions too. +#if defined(_WIN32_WCE) && defined(_MSC_VER) && _MSC_VER<=1500 +long abs(long x) { return (labs(x)); } +double abs(double x) { return (fabs(x)); } +float abs(float x) { return (fabsf(x)); } +long double abs(long double x) { return (fabsl(x)); } +#endif + namespace internal { /** \internal \struct global_math_functions_filtering_base From 6162672dc55dc891a1e2ab1c93d503d7f41cfdac Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 8 Sep 2014 10:04:26 +0200 Subject: [PATCH 0702/1103] Runtime alignement is not possible if AlignedOnScalar is not true (e.g., for complex) --- Eigen/src/Core/Redux.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 5b82c9a65..c626946ba 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -207,7 +207,7 @@ struct redux_impl const Index packetSize = packet_traits::size; const Index alignedStart = internal::first_aligned(mat); enum { - alignment = bool(Derived::Flags & DirectAccessBit) || bool(Derived::Flags & AlignedBit) + alignment = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits::AlignedOnScalar)) || bool(Derived::Flags & AlignedBit) ? Aligned : Unaligned }; const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize); From 51b3f558bb76c11149fc64971db786798f1b692c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 8 Sep 2014 10:21:22 +0200 Subject: [PATCH 0703/1103] Fix bug #822: outer products needed linear access, and add respective unit tests --- Eigen/src/Core/GeneralProduct.h | 4 ++-- test/product.h | 8 ++++++++ 2 files changed, 10 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 624b8b6e8..7179eb124 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -231,7 +231,7 @@ EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& // FIXME not very good if rhs is real and lhs complex while alpha is real too const Index cols = dest.cols(); for (Index j=0; j diff --git a/test/product.h b/test/product.h index 856b234ac..0b3abe402 100644 --- a/test/product.h +++ b/test/product.h @@ -139,4 +139,12 @@ template void product(const MatrixType& m) // inner product Scalar x = square2.row(c) * square2.col(c2); VERIFY_IS_APPROX(x, square2.row(c).transpose().cwiseProduct(square2.col(c2)).sum()); + + // outer product + VERIFY_IS_APPROX(m1.col(c) * m1.row(r), m1.block(0,c,rows,1) * m1.block(r,0,1,cols)); + VERIFY_IS_APPROX(m1.row(r).transpose() * m1.col(c).transpose(), m1.block(r,0,1,cols).transpose() * m1.block(0,c,rows,1).transpose()); + VERIFY_IS_APPROX(m1.block(0,c,rows,1) * m1.row(r), m1.block(0,c,rows,1) * m1.block(r,0,1,cols)); + VERIFY_IS_APPROX(m1.col(c) * m1.block(r,0,1,cols), m1.block(0,c,rows,1) * m1.block(r,0,1,cols)); + VERIFY_IS_APPROX(m1.leftCols(1) * m1.row(r), m1.block(0,0,rows,1) * m1.block(r,0,1,cols)); + VERIFY_IS_APPROX(m1.col(c) * m1.topRows(1), m1.block(0,c,rows,1) * m1.block(0,0,1,cols)); } From 4b678b96eb6cacf59ed19940bcafa1fa4e6f55c0 Mon Sep 17 00:00:00 2001 From: Yan Zhou Date: Mon, 8 Sep 2014 17:37:58 +0800 Subject: [PATCH 0704/1103] fix for MKL_BLAS not defined in MKL 11.2 --- .../products/TriangularMatrixMatrix_MKL.h | 4 +-- Eigen/src/Core/util/MKL_support.h | 32 +++++++++++++++++++ 2 files changed, 34 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h b/Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h index ba41a1c99..4cc56a42f 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h @@ -109,7 +109,7 @@ struct product_triangular_matrix_matrix_trmm #define EIGEN_MKL_VML_THRESHOLD 128 +/* MKL_DOMAIN_BLAS, etc are defined only in 10.3 update 7 */ +/* MKL_BLAS, etc are not defined in 11.2 */ +#ifdef MKL_DOMAIN_ALL +#define EIGEN_MKL_DOMAIN_ALL MKL_DOMAIN_ALL +#else +#define EIGEN_MKL_DOMAIN_ALL MKL_ALL +#endif + +#ifdef MKL_DOMAIN_BLAS +#define EIGEN_MKL_DOMAIN_BLAS MKL_DOMAIN_BLAS +#else +#define EIGEN_MKL_DOMAIN_BLAS MKL_BLAS +#endif + +#ifdef MKL_DOMAIN_FFT +#define EIGEN_MKL_DOMAIN_FFT MKL_DOMAIN_FFT +#else +#define EIGEN_MKL_DOMAIN_FFT MKL_FFT +#endif + +#ifdef MKL_DOMAIN_VML +#define EIGEN_MKL_DOMAIN_VML MKL_DOMAIN_VML +#else +#define EIGEN_MKL_DOMAIN_VML MKL_VML +#endif + +#ifdef MKL_DOMAIN_PARDISO +#define EIGEN_MKL_DOMAIN_PARDISO MKL_DOMAIN_PARDISO +#else +#define EIGEN_MKL_DOMAIN_PARDISO MKL_PARDISO +#endif + namespace Eigen { typedef std::complex dcomplex; From 921a645481aa8825549960c19db2c1bee8375f8f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 10 Sep 2014 10:33:19 +0200 Subject: [PATCH 0705/1103] ArrayWrapper and MatrixWrapper classes should not be nested by reference. --- Eigen/src/Core/ArrayWrapper.h | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/Eigen/src/Core/ArrayWrapper.h b/Eigen/src/Core/ArrayWrapper.h index 4bb648024..28d7b7bd5 100644 --- a/Eigen/src/Core/ArrayWrapper.h +++ b/Eigen/src/Core/ArrayWrapper.h @@ -29,6 +29,11 @@ struct traits > : public traits::type > { typedef ArrayXpr XprKind; + // Let's remove NestByRefBit + enum { + Flags0 = traits::type >::Flags, + Flags = Flags0 & ~NestByRefBit + }; }; } @@ -166,6 +171,11 @@ struct traits > : public traits::type > { typedef MatrixXpr XprKind; + // Let's remove NestByRefBit + enum { + Flags0 = traits::type >::Flags, + Flags = Flags0 & ~NestByRefBit + }; }; } From d6236d3b26f6b652c452d884c440099892fdcdba Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 10 Sep 2014 11:54:20 +0200 Subject: [PATCH 0706/1103] Fix bug #791: infinite loop in JacobiSVD in the presence of NaN. --- Eigen/src/SVD/JacobiSVD.h | 3 ++- test/jacobisvd.cpp | 16 ++++++++++++---- 2 files changed, 14 insertions(+), 5 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 6f3907f5d..6ff689de3 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -741,7 +741,8 @@ JacobiSVD::compute(const MatrixType& matrix, unsig EIGEN_USING_STD_MATH(max); RealScalar threshold = (max)(considerAsZero, precision * (max)(abs(m_workMatrix.coeff(p,p)), abs(m_workMatrix.coeff(q,q)))); - if((max)(abs(m_workMatrix.coeff(p,q)),abs(m_workMatrix.coeff(q,p))) > threshold) + // We compare both values to threshold instead of calling max to be robust to NaN (See bug 791) + if(abs(m_workMatrix.coeff(p,q))>threshold || abs(m_workMatrix.coeff(q,p)) > threshold) { finished = false; diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index 36721b496..422d46695 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -315,16 +315,23 @@ void jacobisvd_inf_nan() VERIFY(sub(some_inf, some_inf) != sub(some_inf, some_inf)); svd.compute(MatrixType::Constant(10,10,some_inf), ComputeFullU | ComputeFullV); - Scalar some_nan = zero() / zero(); - VERIFY(some_nan != some_nan); - svd.compute(MatrixType::Constant(10,10,some_nan), ComputeFullU | ComputeFullV); + Scalar nan = std::numeric_limits::quiet_NaN(); + VERIFY(nan != nan); + svd.compute(MatrixType::Constant(10,10,nan), ComputeFullU | ComputeFullV); MatrixType m = MatrixType::Zero(10,10); m(internal::random(0,9), internal::random(0,9)) = some_inf; svd.compute(m, ComputeFullU | ComputeFullV); m = MatrixType::Zero(10,10); - m(internal::random(0,9), internal::random(0,9)) = some_nan; + m(internal::random(0,9), internal::random(0,9)) = nan; + svd.compute(m, ComputeFullU | ComputeFullV); + + // regression test for bug 791 + m.resize(3,3); + m << 0, 2*NumTraits::epsilon(), 0.5, + 0, -0.5, 0, + nan, 0, 0; svd.compute(m, ComputeFullU | ComputeFullV); } @@ -437,6 +444,7 @@ void test_jacobisvd() // Test on inf/nan matrix CALL_SUBTEST_7( jacobisvd_inf_nan() ); + CALL_SUBTEST_10( jacobisvd_inf_nan() ); } CALL_SUBTEST_7(( jacobisvd(MatrixXf(internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2), internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2))) )); From 84a7ead059917964cc8719f372d7d153bb2cad53 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 10 Sep 2014 11:59:45 +0200 Subject: [PATCH 0707/1103] Add one more regression test for bug #791. --- test/jacobisvd.cpp | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index 422d46695..7dbb29c59 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -333,6 +333,13 @@ void jacobisvd_inf_nan() 0, -0.5, 0, nan, 0, 0; svd.compute(m, ComputeFullU | ComputeFullV); + + m.resize(4,4); + m << 1, 0, 0, 0, + 0, 3, 1, 2e-308, + 1, 0, 1, nan, + 0, nan, nan, 0; + svd.compute(m, ComputeFullU | ComputeFullV); } // Regression test for bug 286: JacobiSVD loops indefinitely with some From 2d90484450f3934db3f5db39ef37967fb9444263 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 10 Sep 2014 23:10:01 +0200 Subject: [PATCH 0708/1103] mat/=scalar was transformed into mat*=(1/scalar) thus laking accuracy. This was also inconsistent with mat = mat/scalar. --- Eigen/src/Core/SelfCwiseBinaryOp.h | 10 ++-------- 1 file changed, 2 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h index 65864adf8..8abdca4a5 100644 --- a/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -209,15 +209,9 @@ inline Derived& ArrayBase::operator-=(const Scalar& other) template inline Derived& DenseBase::operator/=(const Scalar& other) { - typedef typename internal::conditional::IsInteger, - internal::scalar_quotient_op, - internal::scalar_product_op >::type BinOp; typedef typename Derived::PlainObject PlainObject; - SelfCwiseBinaryOp tmp(derived()); - Scalar actual_other; - if(NumTraits::IsInteger) actual_other = other; - else actual_other = Scalar(1)/other; - tmp = PlainObject::Constant(rows(),cols(), actual_other); + SelfCwiseBinaryOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived()); + tmp = PlainObject::Constant(rows(),cols(), other); return derived(); } From 5e890d3ad78a7e5c491a43202993d617fffb964a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 10 Sep 2014 23:11:58 +0200 Subject: [PATCH 0709/1103] Improve further the accuracy of JacobiSVD wrt under/overflow while improving speed for small matrices (hypot is very slow). --- Eigen/src/SVD/JacobiSVD.h | 17 +++++++++-------- test/jacobisvd.cpp | 29 ++++++++++++++++++++++++++--- 2 files changed, 35 insertions(+), 11 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 6ff689de3..3ab8a4c8a 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -425,18 +425,19 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, JacobiRotation rot1; RealScalar t = m.coeff(0,0) + m.coeff(1,1); RealScalar d = m.coeff(1,0) - m.coeff(0,1); - if(t == RealScalar(0)) + + if(d == RealScalar(0)) { - rot1.c() = RealScalar(0); - rot1.s() = d > RealScalar(0) ? RealScalar(1) : RealScalar(-1); + rot1.s() = RealScalar(0); + rot1.c() = RealScalar(1); } else { - RealScalar t2d2 = numext::hypot(t,d); - rot1.c() = abs(t)/t2d2; - rot1.s() = d/t2d2; - if(tmakeJacobi(m,0,1); diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index 7dbb29c59..cd04db5be 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -354,11 +354,33 @@ void jacobisvd_underoverflow() Matrix2d M; M << -7.90884e-313, -4.94e-324, 0, 5.60844e-313; + JacobiSVD svd; + svd.compute(M,ComputeFullU|ComputeFullV); + jacobisvd_check_full(M,svd); + + VectorXd value_set(9); + value_set << 0, 1, -1, 5.60844e-313, -5.60844e-313, 4.94e-324, -4.94e-324, -4.94e-223, 4.94e-223; + Array4i id(0,0,0,0); + int k = 0; + do + { + M << value_set(id(0)), value_set(id(1)), value_set(id(2)), value_set(id(3)); + svd.compute(M,ComputeFullU|ComputeFullV); + jacobisvd_check_full(M,svd); + + id(k)++; + if(id(k)>=value_set.size()) + { + while(k<3 && id(k)>=value_set.size()) id(++k)++; + id.head(k).setZero(); + k=0; + } + + } while((id svd; - svd.compute(M); // just check we don't loop indefinitely // Check for overflow: Matrix3d M3; @@ -367,7 +389,8 @@ void jacobisvd_underoverflow() -8.7190887618028355e+307, -7.3453213709232193e+307, -2.4367363684472105e+307; JacobiSVD svd3; - svd3.compute(M3); // just check we don't loop indefinitely + svd3.compute(M3,ComputeFullU|ComputeFullV); // just check we don't loop indefinitely + jacobisvd_check_full(M3,svd3); } void jacobisvd_preallocate() From 57f71a5552025320f64e330b8e67326097bbce92 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 11 Sep 2014 10:27:46 +0200 Subject: [PATCH 0710/1103] Update bench_norm utility --- bench/bench_norm.cpp | 33 ++++++++++++++++++++++----------- 1 file changed, 22 insertions(+), 11 deletions(-) diff --git a/bench/bench_norm.cpp b/bench/bench_norm.cpp index 398fef835..129afcfb2 100644 --- a/bench/bench_norm.cpp +++ b/bench/bench_norm.cpp @@ -6,19 +6,25 @@ using namespace Eigen; using namespace std; template -EIGEN_DONT_INLINE typename T::Scalar sqsumNorm(const T& v) +EIGEN_DONT_INLINE typename T::Scalar sqsumNorm(T& v) { return v.norm(); } template -EIGEN_DONT_INLINE typename T::Scalar hypotNorm(const T& v) +EIGEN_DONT_INLINE typename T::Scalar stableNorm(T& v) +{ + return v.stableNorm(); +} + +template +EIGEN_DONT_INLINE typename T::Scalar hypotNorm(T& v) { return v.hypotNorm(); } template -EIGEN_DONT_INLINE typename T::Scalar blueNorm(const T& v) +EIGEN_DONT_INLINE typename T::Scalar blueNorm(T& v) { return v.blueNorm(); } @@ -217,20 +223,21 @@ EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v) } #define BENCH_PERF(NRM) { \ + float af = 0; double ad = 0; std::complex ac = 0; \ Eigen::BenchTimer tf, td, tcf; tf.reset(); td.reset(); tcf.reset();\ for (int k=0; k Date: Fri, 8 Aug 2014 04:05:28 -0400 Subject: [PATCH 0711/1103] Fix bug #852: define Traits type in general_matrix_matrix_product when EIGEN_USE_BLAS is defined --- Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h b/Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h index 060af328e..b6ae729b2 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h @@ -53,6 +53,8 @@ template< \ int RhsStorageOrder, bool ConjugateRhs> \ struct general_matrix_matrix_product \ { \ +typedef gebp_traits Traits; \ +\ static void run(Index rows, Index cols, Index depth, \ const EIGTYPE* _lhs, Index lhsStride, \ const EIGTYPE* _rhs, Index rhsStride, \ From 439feca139a093292923e14c085352e5dd2239a2 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 13 Aug 2014 08:22:05 -0700 Subject: [PATCH 0712/1103] Reworked the TensorExecutor code to support in place evaluation. --- .../Eigen/CXX11/src/Tensor/TensorExecutor.h | 106 ++++++++++-------- 1 file changed, 58 insertions(+), 48 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index f50f839fc..d6e2ab1a2 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -11,7 +11,7 @@ #define EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H #ifdef EIGEN_USE_THREADS -#include +#include " #endif namespace Eigen { @@ -28,45 +28,49 @@ namespace internal { // Default strategy: the expression is evaluated with a single cpu thread. template::PacketAccess> -struct TensorExecutor +class TensorExecutor { + public: typedef typename Expression::Index Index; EIGEN_DEVICE_FUNC static inline void run(const Expression& expr, const Device& device = Device()) { TensorEvaluator evaluator(expr, device); - evaluator.evalSubExprsIfNeeded(); - - const Index size = evaluator.dimensions().TotalSize(); - for (Index i = 0; i < size; ++i) { - evaluator.evalScalar(i); + const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); + if (needs_assign) + { + const Index size = evaluator.dimensions().TotalSize(); + for (Index i = 0; i < size; ++i) { + evaluator.evalScalar(i); + } } - evaluator.cleanup(); } }; template -struct TensorExecutor +class TensorExecutor { + public: typedef typename Expression::Index Index; static inline void run(const Expression& expr, const DefaultDevice& device = DefaultDevice()) { TensorEvaluator evaluator(expr, device); - evaluator.evalSubExprsIfNeeded(); + const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); + if (needs_assign) + { + const Index size = evaluator.dimensions().TotalSize(); + static const int PacketSize = unpacket_traits::PacketReturnType>::size; + const int VectorizedSize = (size / PacketSize) * PacketSize; - const Index size = evaluator.dimensions().TotalSize(); - static const int PacketSize = unpacket_traits::PacketReturnType>::size; - const int VectorizedSize = (size / PacketSize) * PacketSize; - - for (Index i = 0; i < VectorizedSize; i += PacketSize) { - evaluator.evalPacket(i); + for (Index i = 0; i < VectorizedSize; i += PacketSize) { + evaluator.evalPacket(i); + } + for (Index i = VectorizedSize; i < size; ++i) { + evaluator.evalScalar(i); + } } - for (Index i = VectorizedSize; i < size; ++i) { - evaluator.evalScalar(i); - } - evaluator.cleanup(); } }; @@ -107,38 +111,40 @@ struct EvalRange { }; template -struct TensorExecutor +class TensorExecutor { + public: typedef typename Expression::Index Index; static inline void run(const Expression& expr, const ThreadPoolDevice& device) { typedef TensorEvaluator Evaluator; Evaluator evaluator(expr, device); - evaluator.evalSubExprsIfNeeded(); + const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); + if (needs_assign) + { + const Index size = evaluator.dimensions().TotalSize(); - const Index size = evaluator.dimensions().TotalSize(); + static const int PacketSize = Vectorizable ? unpacket_traits::size : 1; - static const int PacketSize = Vectorizable ? unpacket_traits::size : 1; + int blocksz = std::ceil(static_cast(size)/device.numThreads()) + PacketSize - 1; + const Index blocksize = std::max(PacketSize, (blocksz - (blocksz % PacketSize))); + const Index numblocks = size / blocksize; - int blocksz = std::ceil(static_cast(size)/device.numThreads()) + PacketSize - 1; - const Index blocksize = std::max(PacketSize, (blocksz - (blocksz % PacketSize))); - const Index numblocks = size / blocksize; + Index i = 0; + vector > results; + results.reserve(numblocks); + for (int i = 0; i < numblocks; ++i) { + results.push_back(std::async(std::launch::async, &EvalRange::run, &evaluator, i*blocksize, (i+1)*blocksize)); + } - Index i = 0; - vector > results; - results.reserve(numblocks); - for (int i = 0; i < numblocks; ++i) { - results.push_back(std::async(std::launch::async, &EvalRange::run, &evaluator, i*blocksize, (i+1)*blocksize)); + for (int i = 0; i < numblocks; ++i) { + results[i].get(); + } + + if (numblocks * blocksize < size) { + EvalRange::run(&evaluator, numblocks * blocksize, size); + } } - - for (int i = 0; i < numblocks; ++i) { - results[i].get(); - } - - if (numblocks * blocksize < size) { - EvalRange::run(&evaluator, numblocks * blocksize, size); - } - evaluator.cleanup(); } }; @@ -157,19 +163,23 @@ __global__ void EigenMetaKernel(Evaluator eval, unsigned int size) { } template -struct TensorExecutor +class TensorExecutor { + public: typedef typename Expression::Index Index; static inline void run(const Expression& expr, const GpuDevice& device) { TensorEvaluator evaluator(expr, device); - evaluator.evalSubExprsIfNeeded(); - const int num_blocks = getNumCudaMultiProcessors() * maxCudaThreadsPerMultiProcessor() / maxCudaThreadsPerBlock(); - const int block_size = maxCudaThreadsPerBlock(); + const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); + if (needs_assign) + { + const int num_blocks = getNumCudaMultiProcessors() * maxCudaThreadsPerMultiProcessor() / maxCudaThreadsPerBlock(); + const int block_size = maxCudaThreadsPerBlock(); - const Index size = evaluator.dimensions().TotalSize(); - EigenMetaKernel > <<>>(evaluator, size); - eigen_assert(cudaGetLastError() == cudaSuccess); + const Index size = evaluator.dimensions().TotalSize(); + EigenMetaKernel > <<>>(evaluator, size); + assert(cudaGetLastError() == cudaSuccess); + } evaluator.cleanup(); } }; From b1892ab14d8ac94bef233d0cef0ef7df1e9a592e Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 13 Aug 2014 08:26:44 -0700 Subject: [PATCH 0713/1103] Added suppor for in place evaluation to simple tensor expressions. Use mempy to speedup tensor copies whenever possible. --- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 12 +++- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 55 ++++++++++++++----- 2 files changed, 50 insertions(+), 17 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index a2a925775..3bfe80c9e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -102,6 +102,7 @@ struct TensorEvaluator, Device> { } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; @@ -112,9 +113,14 @@ struct TensorEvaluator, Device> return m_rightImpl.dimensions(); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_leftImpl.evalSubExprsIfNeeded(); - m_rightImpl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + eigen_assert(internal::dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions())); + m_leftImpl.evalSubExprsIfNeeded(NULL); + // If the lhs provides raw access to its storage area (i.e. if m_leftImpl.data() returns a non + // null value), attempt to evaluate the rhs expression in place. Returns true iff in place + // evaluation isn't supported and the caller still needs to manually assign the values generated + // by the rhs to the lhs. + return m_rightImpl.evalSubExprsIfNeeded(m_leftImpl.data()); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_leftImpl.cleanup(); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index ac9829ce9..0f969036c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -39,13 +39,20 @@ struct TensorEvaluator PacketAccess = Derived::PacketAccess, }; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device&) - : m_data(const_cast(m.data())), m_dims(m.dimensions()) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const Derived& m, const Device& device) + : m_data(const_cast(m.data())), m_dims(m.dimensions()), m_device(device) { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dims; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* dest) { + if (dest) { + m_device.memcpy((void*)dest, m_data, sizeof(Scalar) * m_dims.TotalSize()); + return false; + } + return true; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { @@ -70,9 +77,12 @@ struct TensorEvaluator return internal::pstoret(m_data + index, x); } + Scalar* data() const { return m_data; } + protected: Scalar* m_data; Dimensions m_dims; + const Device& m_device; }; @@ -98,7 +108,7 @@ struct TensorEvaluator EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dims; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { return true; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { @@ -112,6 +122,8 @@ struct TensorEvaluator return internal::ploadt(m_data + index); } + const Scalar* data() const { return m_data; } + protected: const Scalar* m_data; Dimensions m_dims; @@ -138,13 +150,14 @@ struct TensorEvaluator, Device> { } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { return true; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { } EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const @@ -158,6 +171,8 @@ struct TensorEvaluator, Device> return m_functor.packetOp(index); } + Scalar* data() const { return NULL; } + private: const NullaryOp m_functor; TensorEvaluator m_argImpl; @@ -183,14 +198,16 @@ struct TensorEvaluator, Device> { } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_argImpl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + m_argImpl.evalSubExprsIfNeeded(NULL); + return true; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_argImpl.cleanup(); @@ -207,6 +224,8 @@ struct TensorEvaluator, Device> return m_functor.packetOp(m_argImpl.template packet(index)); } + Scalar* data() const { return NULL; } + private: const UnaryOp m_functor; TensorEvaluator m_argImpl; @@ -233,6 +252,7 @@ struct TensorEvaluator::Dimensions Dimensions; @@ -243,9 +263,10 @@ struct TensorEvaluator(index), m_rightImpl.template packet(index)); } + Scalar* data() const { return NULL; } + private: const BinaryOp m_functor; TensorEvaluator m_leftImpl; @@ -289,6 +312,7 @@ struct TensorEvaluator { } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; @@ -299,10 +323,11 @@ struct TensorEvaluator return m_condImpl.dimensions(); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_condImpl.evalSubExprsIfNeeded(); - m_thenImpl.evalSubExprsIfNeeded(); - m_elseImpl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + m_condImpl.evalSubExprsIfNeeded(NULL); + m_thenImpl.evalSubExprsIfNeeded(NULL); + m_elseImpl.evalSubExprsIfNeeded(NULL); + return true; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_condImpl.cleanup(); @@ -327,6 +352,8 @@ struct TensorEvaluator m_elseImpl.template packet(index)); } + Scalar* data() const { return NULL; } + private: TensorEvaluator m_condImpl; TensorEvaluator m_thenImpl; From 1aa2bf82741f2f51fbf0a29ff95e0d017f6962a3 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 13 Aug 2014 08:27:58 -0700 Subject: [PATCH 0714/1103] Support for in place evaluation of expressions containing slicing and reshaping operations --- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 30 ++++++++++++++----- 1 file changed, 22 insertions(+), 8 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 3b42c8514..2b1b503cf 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -103,13 +103,14 @@ struct TensorEvaluator, Device> { } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_impl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + return m_impl.evalSubExprsIfNeeded(data); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_impl.cleanup(); @@ -126,6 +127,8 @@ struct TensorEvaluator, Device> return m_impl.template packet(index); } + Scalar* data() const { return NULL; } + protected: NewDimensions m_dimensions; TensorEvaluator m_impl; @@ -150,13 +153,14 @@ struct TensorEvaluator, Device> { } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_impl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + return m_impl.evalSubExprsIfNeeded(data); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_impl.cleanup(); @@ -182,6 +186,8 @@ struct TensorEvaluator, Device> return m_impl.template packet(index); } + Scalar* data() const { return NULL; } + private: NewDimensions m_dimensions; TensorEvaluator m_impl; @@ -306,14 +312,16 @@ struct TensorEvaluator, Devi } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; typedef Sizes Dimensions; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_impl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_impl.cleanup(); @@ -366,6 +374,8 @@ struct TensorEvaluator, Devi } } + Scalar* data() const { return NULL; } + private: Dimensions m_dimensions; array m_outputStrides; @@ -415,14 +425,16 @@ struct TensorEvaluator, Device> } typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; typedef Sizes Dimensions; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_impl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_impl.cleanup(); @@ -515,6 +527,8 @@ struct TensorEvaluator, Device> } } + Scalar* data() const { return NULL; } + private: Dimensions m_dimensions; array m_outputStrides; From 72e75297089e7c141108696195763c024571974d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 13 Aug 2014 08:29:40 -0700 Subject: [PATCH 0715/1103] Fixed a typo. --- unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index d6e2ab1a2..faf965df8 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -11,7 +11,7 @@ #define EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H #ifdef EIGEN_USE_THREADS -#include " +#include #endif namespace Eigen { From f8fad09301106c574ed88ffde52e15483d14673f Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 13 Aug 2014 08:33:18 -0700 Subject: [PATCH 0716/1103] Updated the convolution and contraction evaluators to follow the new EvalSubExprsIfNeeded apu. --- .../Eigen/CXX11/src/Tensor/TensorContraction.h | 7 ++++--- .../Eigen/CXX11/src/Tensor/TensorConvolution.h | 11 ++++++----- 2 files changed, 10 insertions(+), 8 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index b2e12fd15..8d7a1351e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -184,9 +184,10 @@ struct TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); @@ -151,11 +151,12 @@ struct TensorEvaluator Date: Wed, 13 Aug 2014 08:36:33 -0700 Subject: [PATCH 0717/1103] Added missing apis. --- unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 4 +++- unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h | 3 +++ 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index 8d7a1351e..b2969337f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -173,7 +173,7 @@ struct TensorEvaluator::value> m_leftOffsets; array::value> m_rightOffsets; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 8864c5329..e3068dcae 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -148,6 +148,7 @@ struct TensorEvaluator m_inputStride; array m_outputStride; From f1d8c13dbcbe38938dcd727f9b50339a981197c3 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 13 Aug 2014 08:40:26 -0700 Subject: [PATCH 0718/1103] Fixed misc typos. --- unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index b2969337f..897d73806 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -184,7 +184,7 @@ struct TensorEvaluator Date: Wed, 13 Aug 2014 08:44:47 -0700 Subject: [PATCH 0719/1103] Added ability to get the nth element from an abstract array type. --- .../Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h index 1d3164d6a..4c6b95773 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h @@ -181,6 +181,15 @@ array repeat(t v) { return array; } +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Head::type array_get(type_list& a) { + return get >::value; +} +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Head::type array_get(const type_list& a) { + return get >::value; +} + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const array& a) { t prod = 1; @@ -196,8 +205,8 @@ template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T& array_get(array& a) { return a[I]; } -template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE -const T& array_get(const array& a) { +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T& array_get(const array& a) { return a[I]; } From eeb43f9e2b7ac56af685d8fc494685df8227a53f Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 14 Aug 2014 00:22:47 -0700 Subject: [PATCH 0720/1103] Added support for padding, stridding, and shuffling --- unsupported/Eigen/CXX11/Tensor | 3 + .../Eigen/CXX11/src/Tensor/TensorBase.h | 15 ++ .../src/Tensor/TensorForwardDeclarations.h | 3 + .../Eigen/CXX11/src/Tensor/TensorPadding.h | 163 +++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorShuffling.h | 168 +++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorStriding.h | 172 ++++++++++++++++++ 6 files changed, 524 insertions(+) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 7e504b302..0775d440a 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -42,6 +42,9 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 527d47c57..0295fcdbc 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -215,6 +215,21 @@ class TensorBase slice(const StartIndices& startIndices, const Sizes& sizes) const { return TensorSlicingOp(derived(), startIndices, sizes); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorPaddingOp + pad(const PaddingDimensions& padding) const { + return TensorPaddingOp(derived(), padding); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorShufflingOp + shuffle(const Shuffle& shuffle) const { + return TensorShufflingOp(derived(), shuffle); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorStridingOp + stride(const Strides& strides) const { + return TensorStridingOp(derived(), strides); + } // Force the evaluation of the expression. EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 5d6e7776a..baa5968bc 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -26,6 +26,9 @@ template class template class TensorConvolutionOp; template class TensorReshapingOp; template class TensorSlicingOp; +template class TensorPaddingOp; +template class TensorShufflingOp; +template class TensorStridingOp; template class TensorAssignOp; template class TensorEvalToOp; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h new file mode 100644 index 000000000..45558d7dd --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h @@ -0,0 +1,163 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_PADDING_H +#define EIGEN_CXX11_TENSOR_TENSOR_PADDING_H + +namespace Eigen { + +/** \class TensorPadding + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor padding class. + * At the moment only 0-padding is supported. + * + */ +namespace internal { +template +struct traits > : public traits +{ + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorPaddingOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorPaddingOp type; +}; + +} // end namespace internal + + + +template +class TensorPaddingOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorPaddingOp(const XprType& expr, const PaddingDimensions& padding_dims) + : m_xpr(expr), m_padding_dims(padding_dims) {} + + EIGEN_DEVICE_FUNC + const PaddingDimensions& padding() const { return m_padding_dims; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + protected: + typename XprType::Nested m_xpr; + const PaddingDimensions m_padding_dims; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorPaddingOp XprType; + typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::value; + typedef DSizes Dimensions; + + enum { + IsAligned = false, + PacketAccess = /*TensorEvaluator::PacketAccess*/false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_padding(op.padding()) + { + // Compute dimensions + m_dimensions = m_impl.dimensions(); + for (int i = 0; i < NumDims; ++i) { + m_dimensions[i] += m_padding[i].first + m_padding[i].second; + } + + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } else { + m_inputStrides[0] = 1; + m_outputStrides[0] = 1; + } + } + } + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i >= 0; --i) { + const Index idx = index / m_outputStrides[i]; + if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { + return Scalar(0); + } + inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + return m_impl.coeff(inputIndex); + } + + /* template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return m_impl.template packet(index); + }*/ + + Scalar* data() const { return NULL; } + + protected: + PaddingDimensions m_padding; + Dimensions m_dimensions; + array m_outputStrides; + array m_inputStrides; + TensorEvaluator m_impl; +}; + + + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_PADDING_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h new file mode 100644 index 000000000..4dfc99203 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h @@ -0,0 +1,168 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H +#define EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H + +namespace Eigen { + +/** \class TensorShuffling + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor shuffling class. + * + * + */ +namespace internal { +template +struct traits > : public traits +{ + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorShufflingOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorShufflingOp type; +}; + +} // end namespace internal + + + +template +class TensorShufflingOp : public TensorBase, WriteAccessors> +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp(const XprType& expr, const Shuffle& shuffle) + : m_xpr(expr), m_shuffle(shuffle) {} + + EIGEN_DEVICE_FUNC + const Shuffle& shuffle() const { return m_shuffle; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorShufflingOp& operator = (const OtherDerived& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + + protected: + typename XprType::Nested m_xpr; + const Shuffle m_shuffle; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorShufflingOp XprType; + typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::Dimensions>::value; + typedef DSizes Dimensions; + + enum { + IsAligned = /*TensorEvaluator::IsAligned*/false, + PacketAccess = /*TensorEvaluator::PacketAccess*/false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_shuffle(op.shuffle()) + { + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + for (int i = 0; i < NumDims; ++i) { + m_dimensions[i] = input_dims[m_shuffle[i]]; + } + + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } else { + m_inputStrides[0] = 1; + m_outputStrides[0] = 1; + } + } + } + + // typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += idx * m_inputStrides[m_shuffle[i]]; + index -= idx * m_outputStrides[i]; + } + inputIndex += index * m_inputStrides[m_shuffle[0]]; + return m_impl.coeff(inputIndex); + } + + /* template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return m_impl.template packet(index); + }*/ + + Scalar* data() const { return NULL; } + + protected: + Dimensions m_dimensions; + Shuffle m_shuffle; + array m_outputStrides; + array m_inputStrides; + TensorEvaluator m_impl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h new file mode 100644 index 000000000..7acdbfc72 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h @@ -0,0 +1,172 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H +#define EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H + +namespace Eigen { + +/** \class TensorStriding + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor striding class. + * + * + */ +namespace internal { +template +struct traits > : public traits +{ + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorStridingOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorStridingOp type; +}; + +} // end namespace internal + + + +template +class TensorStridingOp : public TensorBase, WriteAccessors> +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorStridingOp(const XprType& expr, const Strides& dims) + : m_xpr(expr), m_dims(dims) {} + + EIGEN_DEVICE_FUNC + const Strides& strides() const { return m_dims; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorStridingOp& operator = (const OtherDerived& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + + protected: + typename XprType::Nested m_xpr; + const Strides m_dims; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorStridingOp XprType; + typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::Dimensions>::value; + typedef DSizes Dimensions; + + enum { + IsAligned = /*TensorEvaluator::IsAligned*/false, + PacketAccess = /*TensorEvaluator::PacketAccess*/false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device) + { + m_dimensions = m_impl.dimensions(); + for (int i = 0; i < NumDims; ++i) { + m_dimensions[i] = ceilf(static_cast(m_dimensions[i]) / op.strides()[i]); + } + + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } else { + m_inputStrides[0] = 1; + m_outputStrides[0] = 1; + } + } + for (int i = 0; i < NumDims; ++i) { + m_inputStrides[i] *= op.strides()[i]; + } + } + + // typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += idx * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + inputIndex += index * m_inputStrides[0]; + return m_impl.coeff(inputIndex); + } + + /* template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + return m_impl.template packet(index); + }*/ + + Scalar* data() const { return NULL; } + + protected: + // Strides m_strides; + Dimensions m_dimensions; + array m_outputStrides; + array m_inputStrides; + TensorEvaluator m_impl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H From 8c8db49331a89236be7fdf045279504dd7d1797a Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 14 Aug 2014 00:25:22 -0700 Subject: [PATCH 0721/1103] Added a few regression tests --- unsupported/test/CMakeLists.txt | 3 + unsupported/test/cxx11_tensor_padding.cpp | 54 +++++++++ unsupported/test/cxx11_tensor_shuffling.cpp | 116 ++++++++++++++++++++ unsupported/test/cxx11_tensor_striding.cpp | 71 ++++++++++++ 4 files changed, 244 insertions(+) create mode 100644 unsupported/test/cxx11_tensor_padding.cpp create mode 100644 unsupported/test/cxx11_tensor_shuffling.cpp create mode 100644 unsupported/test/cxx11_tensor_striding.cpp diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 406564673..cd2063848 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -109,6 +109,9 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") # ei_add_test(cxx11_tensor_morphing "-std=c++0x") + ei_add_test(cxx11_tensor_padding "-std=c++0x") + ei_add_test(cxx11_tensor_shuffling "-std=c++0x") + ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") # ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_padding.cpp b/unsupported/test/cxx11_tensor_padding.cpp new file mode 100644 index 000000000..d93bb1883 --- /dev/null +++ b/unsupported/test/cxx11_tensor_padding.cpp @@ -0,0 +1,54 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +static void test_simple_padding() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + + array, 4> paddings; + paddings[0] = make_pair(0, 0); + paddings[1] = make_pair(2, 1); + paddings[2] = make_pair(3, 4); + paddings[3] = make_pair(0, 0); + + Tensor padded; + padded = tensor.pad(paddings); + + VERIFY_IS_EQUAL(padded.dimension(0), 2+0); + VERIFY_IS_EQUAL(padded.dimension(1), 3+3); + VERIFY_IS_EQUAL(padded.dimension(2), 5+7); + VERIFY_IS_EQUAL(padded.dimension(3), 7+0); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 6; ++j) { + for (int k = 0; k < 12; ++k) { + for (int l = 0; l < 7; ++l) { + if (j >= 2 && j < 5 && k >= 3 && k < 8) { + VERIFY_IS_EQUAL(tensor(i,j-2,k-3,l), padded(i,j,k,l)); + } else { + VERIFY_IS_EQUAL(0.0f, padded(i,j,k,l)); + } + } + } + } + } +} + + +void test_cxx11_tensor_padding() +{ + CALL_SUBTEST(test_simple_padding()); +} diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp new file mode 100644 index 000000000..92dd01a52 --- /dev/null +++ b/unsupported/test/cxx11_tensor_shuffling.cpp @@ -0,0 +1,116 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +static void test_simple_shuffling() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + array shuffles; + shuffles[0] = 0; + shuffles[1] = 1; + shuffles[2] = 2; + shuffles[3] = 3; + + Tensor no_shuffle; + no_shuffle = tensor.shuffle(shuffles); + + VERIFY_IS_EQUAL(no_shuffle.dimension(0), 2); + VERIFY_IS_EQUAL(no_shuffle.dimension(1), 3); + VERIFY_IS_EQUAL(no_shuffle.dimension(2), 5); + VERIFY_IS_EQUAL(no_shuffle.dimension(3), 7); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(tensor(i,j,k,l), no_shuffle(i,j,k,l)); + } + } + } + } + + shuffles[0] = 2; + shuffles[1] = 3; + shuffles[2] = 1; + shuffles[3] = 0; + Tensor shuffle; + shuffle = tensor.shuffle(shuffles); + + VERIFY_IS_EQUAL(shuffle.dimension(0), 5); + VERIFY_IS_EQUAL(shuffle.dimension(1), 7); + VERIFY_IS_EQUAL(shuffle.dimension(2), 3); + VERIFY_IS_EQUAL(shuffle.dimension(3), 2); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(tensor(i,j,k,l), shuffle(k,l,j,i)); + } + } + } + } +} + + +static void test_expr_shuffling() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + + array shuffles; + shuffles[0] = 2; + shuffles[1] = 3; + shuffles[2] = 1; + shuffles[3] = 0; + Tensor expected; + expected = tensor.shuffle(shuffles); + + Tensor result(5,7,3,2); + + array src_slice_dim(Eigen::array(2,3,1,7)); + array src_slice_start(Eigen::array(0,0,0,0)); + array dst_slice_dim(Eigen::array(1,7,3,2)); + array dst_slice_start(Eigen::array(0,0,0,0)); + + for (int i = 0; i < 5; ++i) { + result.slice(dst_slice_start, dst_slice_dim) = + tensor.slice(src_slice_start, src_slice_dim).shuffle(shuffles); + src_slice_start[2] += 1; + dst_slice_start[0] += 1; + } + + VERIFY_IS_EQUAL(result.dimension(0), 5); + VERIFY_IS_EQUAL(result.dimension(1), 7); + VERIFY_IS_EQUAL(result.dimension(2), 3); + VERIFY_IS_EQUAL(result.dimension(3), 2); + + for (int i = 0; i < expected.dimension(0); ++i) { + for (int j = 0; j < expected.dimension(1); ++j) { + for (int k = 0; k < expected.dimension(2); ++k) { + for (int l = 0; l < expected.dimension(3); ++l) { + VERIFY_IS_EQUAL(result(i,j,k,l), expected(i,j,k,l)); + } + } + } + } +} + + +void test_cxx11_tensor_shuffling() +{ + CALL_SUBTEST(test_simple_shuffling()); + CALL_SUBTEST(test_expr_shuffling()); +} diff --git a/unsupported/test/cxx11_tensor_striding.cpp b/unsupported/test/cxx11_tensor_striding.cpp new file mode 100644 index 000000000..502569d1d --- /dev/null +++ b/unsupported/test/cxx11_tensor_striding.cpp @@ -0,0 +1,71 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +static void test_simple_striding() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + array strides; + strides[0] = 1; + strides[1] = 1; + strides[2] = 1; + strides[3] = 1; + + Tensor no_stride; + no_stride = tensor.stride(strides); + + VERIFY_IS_EQUAL(no_stride.dimension(0), 2); + VERIFY_IS_EQUAL(no_stride.dimension(1), 3); + VERIFY_IS_EQUAL(no_stride.dimension(2), 5); + VERIFY_IS_EQUAL(no_stride.dimension(3), 7); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(tensor(i,j,k,l), no_stride(i,j,k,l)); + } + } + } + } + + strides[0] = 2; + strides[1] = 4; + strides[2] = 2; + strides[3] = 3; + Tensor stride; + stride = tensor.stride(strides); + + VERIFY_IS_EQUAL(stride.dimension(0), 1); + VERIFY_IS_EQUAL(stride.dimension(1), 1); + VERIFY_IS_EQUAL(stride.dimension(2), 3); + VERIFY_IS_EQUAL(stride.dimension(3), 3); + + for (int i = 0; i < 1; ++i) { + for (int j = 0; j < 1; ++j) { + for (int k = 0; k < 3; ++k) { + for (int l = 0; l < 3; ++l) { + VERIFY_IS_EQUAL(tensor(2*i,4*j,2*k,3*l), stride(i,j,k,l)); + } + } + } + } +} + + +void test_cxx11_tensor_striding() +{ + CALL_SUBTEST(test_simple_striding()); +} From 756292f8aa124c842d1e6d9beeb0c416c0d9a7f3 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 14 Aug 2014 00:32:59 -0700 Subject: [PATCH 0722/1103] Fixed compilation errors --- unsupported/test/CMakeLists.txt | 2 +- unsupported/test/cxx11_tensor_padding.cpp | 10 +++++----- 2 files changed, 6 insertions(+), 6 deletions(-) diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index cd2063848..520935105 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -110,7 +110,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_map "-std=c++0x") # ei_add_test(cxx11_tensor_morphing "-std=c++0x") ei_add_test(cxx11_tensor_padding "-std=c++0x") - ei_add_test(cxx11_tensor_shuffling "-std=c++0x") +# ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_padding.cpp b/unsupported/test/cxx11_tensor_padding.cpp index d93bb1883..cb010f512 100644 --- a/unsupported/test/cxx11_tensor_padding.cpp +++ b/unsupported/test/cxx11_tensor_padding.cpp @@ -18,11 +18,11 @@ static void test_simple_padding() Tensor tensor(2,3,5,7); tensor.setRandom(); - array, 4> paddings; - paddings[0] = make_pair(0, 0); - paddings[1] = make_pair(2, 1); - paddings[2] = make_pair(3, 4); - paddings[3] = make_pair(0, 0); + array, 4> paddings; + paddings[0] = std::make_pair(0, 0); + paddings[1] = std::make_pair(2, 1); + paddings[2] = std::make_pair(3, 4); + paddings[3] = std::make_pair(0, 0); Tensor padded; padded = tensor.pad(paddings); From 33c702c79fe227a5b22229c26af276d359a6cb1d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 14 Aug 2014 22:13:21 -0700 Subject: [PATCH 0723/1103] Added support for fast integer divisions by a constant Sped up tensor slicing by a factor of 3 by using these fast integer divisions. --- unsupported/Eigen/CXX11/Tensor | 1 + .../Eigen/CXX11/src/Tensor/TensorIntDiv.h | 82 +++++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 26 +++--- unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_intdiv.cpp | 77 +++++++++++++++++ 5 files changed, 177 insertions(+), 10 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h create mode 100644 unsupported/test/cxx11_tensor_intdiv.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 0775d440a..82552c3c2 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -34,6 +34,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBase.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h new file mode 100644 index 000000000..cf97031be --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h @@ -0,0 +1,82 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H +#define EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H + + +namespace Eigen { + +/** \internal + * + * \class TensorIntDiv + * \ingroup CXX11_Tensor_Module + * + * \brief Fast integer division by a constant. + * + * See the paper from Granlund and Montgomery for explanation. + * (at http://dx.doi.org/10.1145/773473.178249) + * + * \sa Tensor + */ + +namespace internal { + +template +struct TensorIntDivisor { + public: + TensorIntDivisor() { + multiplier = 0; + shift1 = 0; + shift2 = 0; + } + + // Must have 1 <= divider <= 2^31-1 + TensorIntDivisor(const T divider) { + static const int N = 32; + eigen_assert(divider > 0); + eigen_assert(divider <= (1<<(N-1)) - 1); + + // fast ln2 + const int leading_zeros = __builtin_clz(divider); + const int l = N - (leading_zeros+1); + + multiplier = (static_cast(1) << (N+l)) / divider - (static_cast(1) << N) + 1; + shift1 = (std::min)(1, l); + shift2 = (std::max)(0, l-1); + } + + // Must have 0 <= numerator <= 2^32-1 + T divide(const T numerator) const { + static const int N = 32; + eigen_assert(numerator >= 0); + eigen_assert(numerator <= (1ull<> 32; + uint32_t t = (static_cast(numerator) - t1) >> shift1; + return (t1 + t) >> shift2; + } + + private: + uint64_t multiplier; + int32_t shift1; + int32_t shift2; +}; + + +template +static T operator / (const T& numerator, const TensorIntDivisor& divisor) { + return divisor.divide(numerator); +} + + +} // end namespace internal +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 2b1b503cf..ca3735d64 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -305,8 +305,10 @@ struct TensorEvaluator, Devi for (int i = 0; i < NumDims; ++i) { if (i > 0) { m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1]; + m_fastOutputStrides[i] = internal::TensorIntDivisor(m_outputStrides[i]); } else { m_outputStrides[0] = 1; + m_fastOutputStrides[0] = 1; } } } @@ -331,7 +333,7 @@ struct TensorEvaluator, Devi { Index inputIndex = 0; for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_outputStrides[i]; + const Index idx = index / m_fastOutputStrides[i]; inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; index -= idx * m_outputStrides[i]; } @@ -349,8 +351,8 @@ struct TensorEvaluator, Devi Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / m_outputStrides[i]; - const Index idx1 = indices[1] / m_outputStrides[i]; + const Index idx0 = indices[0] / m_fastOutputStrides[i]; + const Index idx1 = indices[1] / m_fastOutputStrides[i]; inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; indices[0] -= idx0 * m_outputStrides[i]; @@ -379,6 +381,7 @@ struct TensorEvaluator, Devi private: Dimensions m_dimensions; array m_outputStrides; + array, NumDims> m_fastOutputStrides; array m_inputStrides; const StartIndices m_offsets; TensorEvaluator m_impl; @@ -418,9 +421,11 @@ struct TensorEvaluator, Device> for (int i = 0; i < NumDims; ++i) { if (i > 0) { m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1]; + m_fastOutputStrides[i] = internal::TensorIntDivisor(m_outputStrides[i]); } else { m_outputStrides[0] = 1; - } + m_fastOutputStrides[0] = 1; + } } } @@ -444,7 +449,7 @@ struct TensorEvaluator, Device> { Index inputIndex = 0; for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_outputStrides[i]; + const Index idx = index / m_fastOutputStrides[i]; inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; index -= idx * m_outputStrides[i]; } @@ -460,8 +465,8 @@ struct TensorEvaluator, Device> Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / m_outputStrides[i]; - const Index idx1 = indices[1] / m_outputStrides[i]; + const Index idx0 = indices[0] / m_fastOutputStrides[i]; + const Index idx1 = indices[1] / m_fastOutputStrides[i]; inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; indices[0] -= idx0 * m_outputStrides[i]; @@ -489,7 +494,7 @@ struct TensorEvaluator, Device> { Index inputIndex = 0; for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_outputStrides[i]; + const Index idx = index / m_fastOutputStrides[i]; inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; index -= idx * m_outputStrides[i]; } @@ -504,8 +509,8 @@ struct TensorEvaluator, Device> Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / m_outputStrides[i]; - const Index idx1 = indices[1] / m_outputStrides[i]; + const Index idx0 = indices[0] / m_fastOutputStrides[i]; + const Index idx1 = indices[1] / m_fastOutputStrides[i]; inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; indices[0] -= idx0 * m_outputStrides[i]; @@ -532,6 +537,7 @@ struct TensorEvaluator, Device> private: Dimensions m_dimensions; array m_outputStrides; + array, NumDims> m_fastOutputStrides; array m_inputStrides; const StartIndices m_offsets; TensorEvaluator m_impl; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 520935105..e2204827e 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -106,6 +106,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_convolution "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") + ei_add_test(cxx11_tensor_intdiv "-std=c++0x") ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") # ei_add_test(cxx11_tensor_morphing "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_intdiv.cpp b/unsupported/test/cxx11_tensor_intdiv.cpp new file mode 100644 index 000000000..a510dc695 --- /dev/null +++ b/unsupported/test/cxx11_tensor_intdiv.cpp @@ -0,0 +1,77 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + + +static void test_signed_32bit() +{ + for (int32_t i = 1; i < 25000; ++i) { + const Eigen::internal::TensorIntDivisor div(i); + + for (int32_t j = 0; j < 25000; ++j) { + const int32_t fast_div = j / div; + const int32_t slow_div = j / i; + VERIFY_IS_EQUAL(fast_div, slow_div); + } + } +} + + +static void test_unsigned_32bit() +{ + for (uint32_t i = 1; i < 25000; ++i) { + const Eigen::internal::TensorIntDivisor div(i); + + for (uint32_t j = 0; j < 25000; ++j) { + const uint32_t fast_div = j / div; + const uint32_t slow_div = j / i; + VERIFY_IS_EQUAL(fast_div, slow_div); + } + } +} + + +static void test_signed_64bit() +{ + for (int64_t i = 2; i < 25000; ++i) { + const Eigen::internal::TensorIntDivisor div(i); + + for (int64_t j = 0; j < 25000; ++j) { + const int64_t fast_div = j / div; + const int64_t slow_div = j / i; + VERIFY_IS_EQUAL(fast_div, slow_div); + } + } +} + + +static void test_unsigned_64bit() +{ + for (uint64_t i = 2; i < 25000; ++i) { + const Eigen::internal::TensorIntDivisor div(i); + + for (uint64_t j = 0; j < 25000; ++j) { + const uint64_t fast_div = j / div; + const uint64_t slow_div = j / i; + VERIFY_IS_EQUAL(fast_div, slow_div); + } + } +} + + +void test_cxx11_tensor_intdiv() +{ + CALL_SUBTEST(test_signed_32bit()); + CALL_SUBTEST(test_unsigned_32bit()); + CALL_SUBTEST(test_signed_64bit()); + CALL_SUBTEST(test_unsigned_64bit()); +} From 9ac3c821ea3b956634116bcdf80bfab7d9a00d91 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 19 Aug 2014 16:57:10 -0700 Subject: [PATCH 0724/1103] Improved the speed of convolutions when running on cuda devices --- .../CXX11/src/Tensor/TensorConvolution.h | 632 +++++++++++++++++- 1 file changed, 622 insertions(+), 10 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 4158271c3..7d0a21c3b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -20,6 +20,126 @@ namespace Eigen { * */ namespace internal { + + +template class IndexMapper { + public: + IndexMapper(const InputDims& input_dims, const array& kernel_dims, + const array& indices) { + + array dimensions = input_dims; + for (int i = 0; i < NumKernelDims; ++i) { + const Index index = indices[i]; + const Index input_dim = input_dims[index]; + const Index kernel_dim = kernel_dims[i]; + const Index result_dim = input_dim - kernel_dim + 1; + dimensions[index] = result_dim; + } + + array inputStrides; + array outputStrides; + for (int i = 0; i < NumDims; ++i) { + if (i > 0) { + inputStrides[i] = inputStrides[i-1] * input_dims[i-1]; + outputStrides[i] = outputStrides[i-1] * dimensions[i-1]; + } else { + inputStrides[0] = 1; + outputStrides[0] = 1; + } + } + + array cudaInputDimensions; + array cudaOutputDimensions; + array tmp = dimensions; + array ordering; + for (int i = 0; i < NumKernelDims; ++i) { + ordering[i] = indices[i]; + tmp[indices[i]] = -1; + cudaInputDimensions[i] = input_dims[ordering[i]]; + cudaOutputDimensions[i] = dimensions[ordering[i]]; + } + int written = NumKernelDims; + for (int i = 0; i < NumDims; ++i) { + if (tmp[i] >= 0) { + ordering[written] = i; + cudaInputDimensions[written] = input_dims[i]; + cudaOutputDimensions[written] = dimensions[i]; + ++written; + } + } + + for (int i = 0; i < NumDims; ++i) { + m_inputStrides[i] = inputStrides[ordering[i]]; + m_outputStrides[i] = outputStrides[ordering[i]]; + } + + for (int i = 0; i < NumDims; ++i) { + if (i > NumKernelDims) { + m_cudaInputStrides[i] = m_cudaInputStrides[i-1] * cudaInputDimensions[i-1]; + m_cudaOutputStrides[i] = m_cudaOutputStrides[i-1] * cudaOutputDimensions[i-1]; + } else { + m_cudaInputStrides[i] = 1; + m_cudaOutputStrides[i] = 1; + } + } + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputPlaneToTensorInputOffset(Index p) const { + Index inputIndex = 0; + for (int d = NumDims - 1; d > NumKernelDims; --d) { + const Index idx = p / m_cudaInputStrides[d]; + inputIndex += idx * m_inputStrides[d]; + p -= idx * m_cudaInputStrides[d]; + } + inputIndex += p * m_inputStrides[NumKernelDims]; + return inputIndex; + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputPlaneToTensorOutputOffset(Index p) const { + Index outputIndex = 0; + for (int d = NumDims - 1; d > NumKernelDims; --d) { + const Index idx = p / m_cudaOutputStrides[d]; + outputIndex += idx * m_outputStrides[d]; + p -= idx * m_cudaOutputStrides[d]; + } + outputIndex += p * m_outputStrides[NumKernelDims]; + return outputIndex; + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i) const { + return i * m_inputStrides[0]; + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i) const { + return i * m_outputStrides[0]; + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i, Index j) const { + return i * m_inputStrides[0] + j*m_inputStrides[1]; + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i, Index j) const { + return i * m_outputStrides[0] + j * m_outputStrides[1]; + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaInputKernelToTensorInputOffset(Index i, Index j, Index k) const { + return i * m_inputStrides[0] + j*m_inputStrides[1] + k*m_inputStrides[2]; + } + + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Index mapCudaOutputKernelToTensorOutputOffset(Index i, Index j, Index k) const { + return i * m_outputStrides[0] + j*m_outputStrides[1] + k*m_outputStrides[2]; + } + + private: + static const size_t NumDims = internal::array_size::value; + array m_inputStrides; + array m_outputStrides; + array m_cudaInputStrides; + array m_cudaOutputStrides; +}; + + + template struct traits > { @@ -75,15 +195,15 @@ class TensorConvolutionOp : public TensorBase::type& inputExpression() const { return m_input_xpr; } - EIGEN_DEVICE_FUNC + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename internal::remove_all::type& kernelExpression() const { return m_kernel_xpr; } @@ -99,8 +219,8 @@ struct TensorEvaluator XprType; - static const int NumDims = TensorEvaluator::Dimensions::count; - static const int KernelDims = internal::array_size::value; + static const int NumDims = internal::array_size::Dimensions>::value; + static const int NumKernelDims = internal::array_size::value; typedef typename XprType::Index Index; typedef DSizes Dimensions; @@ -111,7 +231,7 @@ struct TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); @@ -124,7 +244,8 @@ struct TensorEvaluator m_inputStride; array m_outputStride; - array m_indexStride; - array m_kernelStride; + array m_indexStride; + array m_kernelStride; TensorEvaluator m_inputImpl; TensorEvaluator m_kernelImpl; Dimensions m_dimensions; }; + + +// Use an optimized implementation of the evaluation code for GPUs whenever possible. +#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) + +template +struct GetKernelSize { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int /*kernelSize*/) const { + return StaticKernelSize; + } +}; +template <> +struct GetKernelSize { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int kernelSize) const { + return kernelSize; + } +}; + + + + +template +__global__ void EigenConvolutionKernel1D(InputEvaluator eval, const internal::IndexMapper indexMapper, const float* __restrict kernel, const int numPlanes, const int numX, const int maxX, const int kernelSize, float* buffer) { + extern __shared__ float s[]; + + const int first_x = blockIdx.x * maxX; + const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1; + const int num_x_input = last_x - first_x + GetKernelSize()(kernelSize); + const int num_x_output = last_x - first_x + 1; + + const int first_plane = blockIdx.y * blockDim.y; + const int plane_stride = blockDim.y * gridDim.y; + + for (int p = first_plane + threadIdx.y; p < numPlanes; p += plane_stride) { + // Load inputs to shared memory + const int plane_input_offset = indexMapper.mapCudaInputPlaneToTensorInputOffset(p); + const int plane_kernel_offset = threadIdx.y * num_x_input; + #pragma unroll + for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) { + const int tensor_index = plane_input_offset + indexMapper.mapCudaInputKernelToTensorInputOffset(i+first_x); + s[i + plane_kernel_offset] = eval.coeff(tensor_index); + } + + __syncthreads(); + + // Compute the convolution + const int plane_output_offset = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(p); + + #pragma unroll + for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) { + const int kernel_offset = plane_kernel_offset + i; + float result = 0.0f; + #pragma unroll + for (int k = 0; k < GetKernelSize()(kernelSize); ++k) { + result += s[k + kernel_offset] * kernel[k]; + } + const int tensor_index = plane_output_offset + indexMapper.mapCudaOutputKernelToTensorOutputOffset(i+first_x); + buffer[tensor_index] = result; + } + __syncthreads(); + } +}; + + +template +__global__ void EigenConvolutionKernel2D(InputEvaluator eval, const internal::IndexMapper indexMapper, const float* __restrict kernel, const int numPlanes, const int numX, const int maxX, const int numY, const int maxY, const int kernelSizeX, const int kernelSizeY, float* buffer) { + extern __shared__ float s[]; + + const int first_x = blockIdx.x * maxX; + const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1; + const int num_x_input = last_x - first_x + GetKernelSize()(kernelSizeX); + const int num_x_output = last_x - first_x + 1; + + const int first_y = blockIdx.y * maxY; + const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1; + const int num_y_input = last_y - first_y + GetKernelSize()(kernelSizeY); + const int num_y_output = last_y - first_y + 1; + + const int first_plane = blockIdx.z * blockDim.z; + const int plane_stride = blockDim.z * gridDim.z; + + for (int p = first_plane + threadIdx.z; p < numPlanes; p += plane_stride) { + + const int plane_input_offset = indexMapper.mapCudaInputPlaneToTensorInputOffset(p); + const int plane_kernel_offset = threadIdx.z * num_y_input; + + // Load inputs to shared memory + #pragma unroll + for (int j = threadIdx.y; j < num_y_input; j += blockDim.y) { + const int input_offset = num_x_input * (j + plane_kernel_offset); + #pragma unroll + for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) { + const int tensor_index = plane_input_offset + indexMapper.mapCudaInputKernelToTensorInputOffset(i+first_x, j+first_y); + s[i + input_offset] = eval.coeff(tensor_index); + } + } + + __syncthreads(); + + // Convolution + const int plane_output_offset = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(p); + + #pragma unroll + for (int j = threadIdx.y; j < num_y_output; j += blockDim.y) { + #pragma unroll + for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) { + float result = 0.0f; + #pragma unroll + for (int l = 0; l < GetKernelSize()(kernelSizeY); ++l) { + const int kernel_offset = kernelSizeX * l; + const int input_offset = i + num_x_input * (j + l + plane_kernel_offset); + #pragma unroll + for (int k = 0; k < GetKernelSize()(kernelSizeX); ++k) { + result += s[k + input_offset] * kernel[k + kernel_offset]; + } + } + const int tensor_index = plane_output_offset + indexMapper.mapCudaOutputKernelToTensorOutputOffset(i+first_x, j+first_y); + buffer[tensor_index] = result; + } + } + + __syncthreads(); + } +}; + + +template +__global__ void EigenConvolutionKernel3D(InputEvaluator eval, const internal::IndexMapper indexMapper, const float* __restrict kernel, const size_t numPlanes, const size_t numX, const size_t maxX, const size_t numY, const size_t maxY, const size_t numZ, const size_t maxZ, const size_t kernelSizeX, const size_t kernelSizeY, const size_t kernelSizeZ, float* buffer) { + extern __shared__ float s[]; + + // Load inputs to shared memory + const int first_x = blockIdx.x * maxX; + const int last_x = (first_x + maxX < numX ? first_x + maxX : numX) - 1; + const int num_x_input = last_x - first_x + kernelSizeX; + + const int first_y = blockIdx.y * maxY; + const int last_y = (first_y + maxY < numY ? first_y + maxY : numY) - 1; + const int num_y_input = last_y - first_y + kernelSizeY; + + const int first_z = blockIdx.z * maxZ; + const int last_z = (first_z + maxZ < numZ ? first_z + maxZ : numZ) - 1; + const int num_z_input = last_z - first_z + kernelSizeZ; + + for (int p = 0; p < numPlanes; ++p) { + + const int plane_input_offset = indexMapper.mapCudaInputPlaneToTensorInputOffset(p); + const int plane_kernel_offset = 0; + + for (int k = threadIdx.z; k < num_z_input; k += blockDim.z) { + for (int j = threadIdx.y; j < num_y_input; j += blockDim.y) { + for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) { + const int tensor_index = plane_input_offset + indexMapper.mapCudaInputKernelToTensorInputOffset(i+first_x, j+first_y, k+first_z); + s[i + num_x_input * (j + num_y_input * (k + plane_kernel_offset))] = eval.coeff(tensor_index); + } + } + } + + __syncthreads(); + + // Convolution + const int num_z_output = last_z - first_z + 1; + const int num_y_output = last_y - first_y + 1; + const int num_x_output = last_x - first_x + 1; + const int plane_output_offset = indexMapper.mapCudaOutputPlaneToTensorOutputOffset(p); + + for (int k = threadIdx.z; k < num_z_output; k += blockDim.z) { + for (int j = threadIdx.y; j < num_y_output; j += blockDim.y) { + for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) { + float result = 0.0f; + for (int n = 0; n < kernelSizeZ; ++n) { + for (int m = 0; m < kernelSizeY; ++m) { + for (int l = 0; l < kernelSizeX; ++l) { + result += s[i + l + num_x_input * (j + m + num_y_input * (k + n + plane_kernel_offset))] * kernel[l + kernelSizeX * (m + kernelSizeY * n)]; + } + } + } + const int tensor_index = plane_output_offset + indexMapper.mapCudaOutputKernelToTensorOutputOffset(i+first_x, j+first_y, k+first_z); + buffer[tensor_index] = result; + } + } + } + __syncthreads(); + } +}; + + + +template +struct TensorEvaluator, GpuDevice> +{ + typedef TensorConvolutionOp XprType; + + static const int NumDims = internal::array_size::Dimensions>::value; + static const int NumKernelDims = internal::array_size::value; + typedef typename XprType::Index Index; + typedef DSizes Dimensions; + typedef typename TensorEvaluator::Dimensions KernelDimensions; + + enum { + IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, + PacketAccess = false, + }; + + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const GpuDevice& device) + : m_inputImpl(op.inputExpression(), device), m_kernelArg(op.kernelExpression()), m_kernelImpl(op.kernelExpression(), device), m_indices(op.indices()), m_buf(NULL), m_kernel(NULL), m_local_kernel(false), m_device(device) + { + const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); + const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); + + m_dimensions = m_inputImpl.dimensions(); + for (int i = 0; i < NumKernelDims; ++i) { + const Index index = op.indices()[i]; + const Index input_dim = input_dims[index]; + const Index kernel_dim = kernel_dims[i]; + const Index result_dim = input_dim - kernel_dim + 1; + m_dimensions[index] = result_dim; + } + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename InputArgType::Scalar Scalar; + + EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + preloadKernel(); + m_inputImpl.evalSubExprsIfNeeded(NULL); + if (data) { + executeEval(data); + return false; + } else { + m_buf = (Scalar*)m_device.allocate(dimensions().TotalSize() * sizeof(Scalar)); + executeEval(m_buf); + return true; + } + } + + EIGEN_STRONG_INLINE void cleanup() { + m_inputImpl.cleanup(); + if (m_buf) { + m_device.deallocate(m_buf); + m_buf = NULL; + } + if (m_local_kernel) { + m_device.deallocate((void*)m_kernel); + m_local_kernel = false; + } + m_kernel = NULL; + } + + EIGEN_STRONG_INLINE void preloadKernel() { + // Don't make a local copy of the kernel unless we have to (i.e. it's an + // expression that needs to be evaluated) + const Scalar* in_place = m_kernelImpl.data(); + if (in_place) { + m_kernel = in_place; + m_local_kernel = false; + } else { + size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar); + Scalar* local = (Scalar*)m_device.allocate(kernel_sz); + typedef TensorEvalToOp EvalTo; + EvalTo evalToTmp(local, m_kernelArg); + internal::TensorExecutor::PacketAccess>::run(evalToTmp, m_device); + + m_kernel = local; + m_local_kernel = true; + } + } + + static unsigned int ceil(unsigned int num, unsigned int denom) { + const unsigned int rounded_toward_zero = num / denom; + if (num > rounded_toward_zero * denom) { + return rounded_toward_zero + 1; + } + return rounded_toward_zero; + } + + void executeEval(Scalar* data) const { + typedef typename TensorEvaluator::Dimensions InputDims; + + const int maxSharedMem = sharedMemPerBlock(); + const int maxThreadsPerBlock = maxCudaThreadsPerBlock(); + const int maxBlocksPerProcessor = maxCudaThreadsPerMultiProcessor() / maxThreadsPerBlock; + const int numMultiProcessors = getNumCudaMultiProcessors(); + const int warpSize = 32; + + switch (NumKernelDims) { + case 1: { + const int kernel_size = m_kernelImpl.dimensions().TotalSize(); + + const int numX = dimensions()[m_indices[0]]; + const int numP = dimensions().TotalSize() / numX; + + int maxX; + dim3 block_size; + if (m_indices[0] == 0) { + // Maximum the reuse + const int inner_dim = ((maxSharedMem / (sizeof(Scalar)) - kernel_size + 1 + 31) / 32) * 32; + maxX = (std::min)(inner_dim, numX); + const int maxP = (std::min)(maxSharedMem / ((kernel_size - 1 + maxX) * sizeof(Scalar)), numP); + block_size.x = (std::min)(maxThreadsPerBlock, maxX); + block_size.y = (std::min)(maxThreadsPerBlock / block_size.x, maxP); + } + else { + // Read as much as possible alongside the inner most dimension, that is the plane + const int inner_dim = maxSharedMem / ((warpSize + kernel_size) * sizeof(Scalar)); + const int maxP = (std::min)(inner_dim, numP); + maxX = (std::min)(maxSharedMem / (inner_dim * sizeof(Scalar)) - kernel_size + 1, numX); + + block_size.x = (std::min)(warpSize, maxX); + block_size.y = (std::min)(maxThreadsPerBlock/block_size.x, maxP); + } + + const int shared_mem = block_size.y * (maxX + kernel_size - 1) * sizeof(Scalar); + assert(shared_mem <= maxSharedMem); + + const int num_x_blocks = ceil(numX, maxX); + const int blocksPerProcessor = (std::min)(maxBlocksPerProcessor, maxSharedMem / shared_mem); + const int num_y_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks); + + dim3 num_blocks(num_x_blocks, min(num_y_blocks, ceil(numP, block_size.y))); + + + //cout << "launching 1D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " maxX: " << maxX << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl; + + const array indices(m_indices[0]); + const array kernel_dims(m_kernelImpl.dimensions()[0]); + internal::IndexMapper indexMapper(m_inputImpl.dimensions(), kernel_dims, indices); + switch(kernel_size) { + case 4: { + EigenConvolutionKernel1D, Index, InputDims, 4> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data); + break; + } + case 7: { + EigenConvolutionKernel1D, Index, InputDims, 7> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data); + break; + } + default: { + EigenConvolutionKernel1D, Index, InputDims, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data); + } + } + cudaError_t error = cudaGetLastError(); + assert(error == cudaSuccess); + break; + } + + case 2: { + const int kernel_size_x = m_kernelImpl.dimensions()[0]; + const int kernel_size_y = m_kernelImpl.dimensions()[1]; + + const int numX = dimensions()[m_indices[0]]; + const int numY = dimensions()[m_indices[1]]; + const int numP = dimensions().TotalSize() / (numX*numY); + + const float scaling_factor = sqrtf(static_cast(maxSharedMem) / (sizeof(Scalar) * kernel_size_y * kernel_size_x)); + + // Snap maxX to warp size + int inner_dim = ((static_cast(scaling_factor * kernel_size_x) - kernel_size_x + 1 + 32) / 32) * 32; + const int maxX = (std::min)(inner_dim, numX); + const int maxY = (std::min)(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1)) - kernel_size_y + 1, numY); + const int maxP = (std::min)(maxSharedMem / ((kernel_size_x - 1 + maxX) * (kernel_size_y - 1 + maxY) * sizeof(Scalar)), numP); + + dim3 block_size; + block_size.x = (std::min)(1024, maxX); + block_size.y = (std::min)(1024/block_size.x, maxY); + block_size.z = (std::min)(1024/(block_size.x*block_size.y), maxP); + + const int shared_mem = block_size.z * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * sizeof(Scalar); + assert(shared_mem <= maxSharedMem); + + const int num_x_blocks = ceil(numX, maxX); + const int num_y_blocks = ceil(numY, maxY); + const int blocksPerProcessor = (std::min)(maxBlocksPerProcessor, maxSharedMem / shared_mem); + const int num_z_blocks = ceil(numMultiProcessors * blocksPerProcessor, num_x_blocks * num_y_blocks); + + dim3 num_blocks(num_x_blocks, num_y_blocks, min(num_z_blocks, ceil(numP, block_size.z))); + + + //cout << "launching 2D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z << " maxX: " << maxX << " maxY: " << maxY << " maxP: " << maxP << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl; + + const array indices(m_indices[0], m_indices[1]); + const array kernel_dims(m_kernelImpl.dimensions()[0], m_kernelImpl.dimensions()[1]); + internal::IndexMapper indexMapper(m_inputImpl.dimensions(), kernel_dims, indices); + switch (kernel_size_x) { + case 4: { + switch (kernel_size_y) { + case 7: { + EigenConvolutionKernel2D, Index, InputDims, 4, 7> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data); + break; + } + default: { + EigenConvolutionKernel2D, Index, InputDims, 4, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data); + break; + } + } + break; + } + case 7: { + switch (kernel_size_y) { + case 4: { + EigenConvolutionKernel2D, Index, InputDims, 7, 4> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data); + break; + } + default: { + EigenConvolutionKernel2D, Index, InputDims, 7, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data); + break; + } + } + break; + } + default: { + EigenConvolutionKernel2D, Index, InputDims, Eigen::Dynamic, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data); + break; + } + } + cudaError_t error = cudaGetLastError(); + assert(error == cudaSuccess); + break; + } + + case 3: { + const int kernel_size_x = m_kernelImpl.dimensions()[0]; + const int kernel_size_y = m_kernelImpl.dimensions()[1]; + const int kernel_size_z = m_kernelImpl.dimensions()[2]; + + const int numX = dimensions()[m_indices[0]]; + const int numY = dimensions()[m_indices[1]]; + const int numZ = dimensions()[m_indices[2]]; + const int numP = dimensions().TotalSize() / (numX*numY*numZ); + + const int maxX = (std::min)(128, (std::min)(maxSharedMem / (sizeof(Scalar) * kernel_size_y * kernel_size_z) - kernel_size_x + 1, numX)); + const int maxY = (std::min)(128, (std::min)(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * kernel_size_z) - kernel_size_y + 1, numY)); + const int maxZ = (std::min)(128, (std::min)(maxSharedMem / (sizeof(Scalar) * (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1)) - kernel_size_z + 1, numZ)); + + dim3 block_size; + block_size.x = (std::min)(32, maxX); + block_size.y = (std::min)(32, maxY); + block_size.z = (std::min)(1024/(block_size.x*block_size.y), maxZ); + dim3 num_blocks(ceil(numX, maxX), ceil(numY, maxY), ceil(numZ, maxZ)); + + const int shared_mem = (maxX + kernel_size_x - 1) * (maxY + kernel_size_y - 1) * (maxZ + kernel_size_z - 1) * sizeof(Scalar); + assert(shared_mem <= maxSharedMem); + + //cout << "launching 3D kernel with block_size.x: " << block_size.x << " block_size.y: " << block_size.y << " block_size.z: " << block_size.z << " num_blocks.x: " << num_blocks.x << " num_blocks.y: " << num_blocks.y << " num_blocks.z: " << num_blocks.z << " shared_mem: " << shared_mem << " in stream " << m_device.stream() << endl; + const array indices(m_indices[0], m_indices[1], m_indices[2]); + const array kernel_dims(m_kernelImpl.dimensions()[0], m_kernelImpl.dimensions()[1], m_kernelImpl.dimensions()[2]); + internal::IndexMapper indexMapper(m_inputImpl.dimensions(), kernel_dims, indices); + + EigenConvolutionKernel3D, Index, InputDims> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data); + cudaError_t error = cudaGetLastError(); + assert(error == cudaSuccess); + break; + } + + default: { + assert(false && "not supported yet"); + } + } + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + assert(m_buf); + assert(index < m_dimensions.TotalSize()); + return m_buf[index]; + } + + private: + // No assignment (copies are needed by the kernels) + TensorEvaluator& operator = (const TensorEvaluator&); + + TensorEvaluator m_inputImpl; + TensorEvaluator m_kernelImpl; + KernelArgType m_kernelArg; + Indices m_indices; + Dimensions m_dimensions; + Scalar* m_buf; + const Scalar* m_kernel; + bool m_local_kernel; + + const GpuDevice& m_device; +}; +#endif + + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_CONVOLUTION_H From 6a3423da800234645ea3a192507a0ee2fe317d6c Mon Sep 17 00:00:00 2001 From: Vladimir Chalupecky Date: Wed, 20 Aug 2014 05:42:22 +0000 Subject: [PATCH 0725/1103] Fix uninitialized variable warning in SparseQR --- Eigen/src/SparseQR/SparseQR.h | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 4c6553bf2..66dccfc43 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -447,7 +447,7 @@ void SparseQR::factorize(const MatrixType& mat) } } // End update current column - Scalar tau; + Scalar tau = RealScalar(0); RealScalar beta = 0; if(nonzeroCol < diagSize) @@ -461,7 +461,6 @@ void SparseQR::factorize(const MatrixType& mat) for (Index itq = 1; itq < nzcolQ; ++itq) sqrNorm += numext::abs2(tval(Qidx(itq))); if(sqrNorm == RealScalar(0) && numext::imag(c0) == RealScalar(0)) { - tau = RealScalar(0); beta = numext::real(c0); tval(Qidx(0)) = 1; } From eeadc06e838a737492625cc3e4d5d5555bb40ff7 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 20 Aug 2014 16:39:25 +0200 Subject: [PATCH 0726/1103] EIGEN_EXCEPTIONS was not defined in test/main.h, therefore all VERIFY_RAISES_ASSERT tests were not enabled --- Eigen/Core | 2 +- test/main.h | 6 +++++- 2 files changed, 6 insertions(+), 2 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 9a73fe37b..776b7faf3 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -42,7 +42,7 @@ #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC; #endif -#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) +#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(EIGEN_EXCEPTIONS) #define EIGEN_EXCEPTIONS #endif diff --git a/test/main.h b/test/main.h index 3295dcb71..376232cf2 100644 --- a/test/main.h +++ b/test/main.h @@ -76,6 +76,10 @@ namespace Eigen #define EIGEN_DEFAULT_IO_FORMAT IOFormat(4, 0, " ", "\n", "", "", "", "") +#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) + #define EIGEN_EXCEPTIONS +#endif + #ifndef EIGEN_NO_ASSERTION_CHECKING namespace Eigen @@ -172,7 +176,7 @@ namespace Eigen #ifndef VERIFY_RAISES_ASSERT #define VERIFY_RAISES_ASSERT(a) \ - std::cout << "Can't VERIFY_RAISES_ASSERT( " #a " ) with exceptions disabled"; + std::cout << "Can't VERIFY_RAISES_ASSERT( " #a " ) with exceptions disabled\n"; #endif #if !defined(__CUDACC__) From 3d298da2696ac956a430f6fbef93bf65ada0d304 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 20 Aug 2014 17:00:50 -0700 Subject: [PATCH 0727/1103] Added support for broadcasting --- unsupported/Eigen/CXX11/Tensor | 1 + .../Eigen/CXX11/src/Tensor/TensorBase.h | 6 + .../CXX11/src/Tensor/TensorBroadcasting.h | 186 ++++++++++++++++++ .../src/Tensor/TensorForwardDeclarations.h | 1 + unsupported/test/CMakeLists.txt | 1 + .../test/cxx11_tensor_broadcasting.cpp | 114 +++++++++++ 6 files changed, 309 insertions(+) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h create mode 100644 unsupported/test/cxx11_tensor_broadcasting.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 82552c3c2..ebe6419e8 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -42,6 +42,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 0295fcdbc..da5148a5b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -204,6 +204,12 @@ class TensorBase return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorBroadcastingOp + broadcast(const Broadcast& broadcast) const { + return TensorBroadcastingOp(derived(), broadcast); + } + // Morphing operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorReshapingOp diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h new file mode 100644 index 000000000..3b2a9c8b9 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h @@ -0,0 +1,186 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H +#define EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H + +namespace Eigen { + +/** \class TensorBroadcasting + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor broadcasting class. + * + * + */ +namespace internal { +template +struct traits > : public traits +{ + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorBroadcastingOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorBroadcastingOp type; +}; + +} // end namespace internal + + + +template +class TensorBroadcastingOp : public TensorBase, WriteAccessors> +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBroadcastingOp(const XprType& expr, const Broadcast& broadcast) + : m_xpr(expr), m_broadcast(broadcast) {} + + EIGEN_DEVICE_FUNC + const Broadcast& broadcast() const { return m_broadcast; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + protected: + typename XprType::Nested m_xpr; + const Broadcast m_broadcast; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorBroadcastingOp XprType; + typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::Dimensions>::value; + typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; + + enum { + IsAligned = false, + PacketAccess = TensorEvaluator::PacketAccess, + }; + +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device) + { + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + const Broadcast& broadcast = op.broadcast(); + for (int i = 0; i < NumDims; ++i) { + eigen_assert(input_dims[i] > 0); + m_dimensions[i] = input_dims[i] * broadcast[i]; + } + + m_inputStrides[0] = 1; + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + // TODO: attempt to speed this up. The integer divisions and modulo are slow + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + inputIndex += (index % m_impl.dimensions()[0]); + return m_impl.coeff(inputIndex); + } + + // Ignore the LoadMode and always use unaligned loads since we can't guarantee + // the alignment at compile time. + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + static const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); + + const Index originalIndex = index; + + Index inputIndex = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + const Index innermostLoc = index % m_impl.dimensions()[0]; + inputIndex += innermostLoc; + + // Todo: this could be extended to the second dimension if we're not + // broadcasting alongside the first dimension, and so on. + if (innermostLoc + packetSize <= m_impl.dimensions()[0]) { + return m_impl.template packet(inputIndex); + } else { + EIGEN_ALIGN_DEFAULT CoeffReturnType values[packetSize]; + values[0] = m_impl.coeff(inputIndex); + for (int i = 1; i < packetSize; ++i) { + values[i] = coeff(originalIndex+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + } + + Scalar* data() const { return NULL; } + + protected: + Dimensions m_dimensions; + array m_outputStrides; + array m_inputStrides; + TensorEvaluator m_impl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index baa5968bc..afbcc9486 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -22,6 +22,7 @@ template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; template class TensorSelectOp; template class TensorReductionOp; +template class TensorBroadcastingOp; template class TensorContractionOp; template class TensorConvolutionOp; template class TensorReshapingOp; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index e2204827e..164388746 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -109,6 +109,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_intdiv "-std=c++0x") ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") + ei_add_test(cxx11_tensor_broadcasting "-std=c++0x") # ei_add_test(cxx11_tensor_morphing "-std=c++0x") ei_add_test(cxx11_tensor_padding "-std=c++0x") # ei_add_test(cxx11_tensor_shuffling "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_broadcasting.cpp b/unsupported/test/cxx11_tensor_broadcasting.cpp new file mode 100644 index 000000000..9663912a4 --- /dev/null +++ b/unsupported/test/cxx11_tensor_broadcasting.cpp @@ -0,0 +1,114 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +static void test_simple_broadcasting() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + array broadcasts; + broadcasts[0] = 1; + broadcasts[1] = 1; + broadcasts[2] = 1; + broadcasts[3] = 1; + + Tensor no_broadcast; + no_broadcast = tensor.broadcast(broadcasts); + + VERIFY_IS_EQUAL(no_broadcast.dimension(0), 2); + VERIFY_IS_EQUAL(no_broadcast.dimension(1), 3); + VERIFY_IS_EQUAL(no_broadcast.dimension(2), 5); + VERIFY_IS_EQUAL(no_broadcast.dimension(3), 7); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(tensor(i,j,k,l), no_broadcast(i,j,k,l)); + } + } + } + } + + broadcasts[0] = 2; + broadcasts[1] = 3; + broadcasts[2] = 1; + broadcasts[3] = 4; + Tensor broadcast; + broadcast = tensor.broadcast(broadcasts); + + VERIFY_IS_EQUAL(broadcast.dimension(0), 4); + VERIFY_IS_EQUAL(broadcast.dimension(1), 9); + VERIFY_IS_EQUAL(broadcast.dimension(2), 5); + VERIFY_IS_EQUAL(broadcast.dimension(3), 28); + + for (int i = 0; i < 4; ++i) { + for (int j = 0; j < 9; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 28; ++l) { + VERIFY_IS_EQUAL(tensor(i%2,j%3,k%5,l%7), broadcast(i,j,k,l)); + } + } + } + } +} + + +static void test_vectorized_broadcasting() +{ + Tensor tensor(8,3,5); + tensor.setRandom(); + array broadcasts; + broadcasts[0] = 2; + broadcasts[1] = 3; + broadcasts[2] = 4; + + Tensor broadcast; + broadcast = tensor.broadcast(broadcasts); + + VERIFY_IS_EQUAL(broadcast.dimension(0), 16); + VERIFY_IS_EQUAL(broadcast.dimension(1), 9); + VERIFY_IS_EQUAL(broadcast.dimension(2), 20); + + for (int i = 0; i < 16; ++i) { + for (int j = 0; j < 9; ++j) { + for (int k = 0; k < 20; ++k) { + VERIFY_IS_EQUAL(tensor(i%8,j%3,k%5), broadcast(i,j,k)); + } + } + } + + tensor.resize(11,3,5); + tensor.setRandom(); + broadcast = tensor.broadcast(broadcasts); + + VERIFY_IS_EQUAL(broadcast.dimension(0), 22); + VERIFY_IS_EQUAL(broadcast.dimension(1), 9); + VERIFY_IS_EQUAL(broadcast.dimension(2), 20); + + for (int i = 0; i < 22; ++i) { + for (int j = 0; j < 9; ++j) { + for (int k = 0; k < 20; ++k) { + VERIFY_IS_EQUAL(tensor(i%11,j%3,k%5), broadcast(i,j,k)); + } + } + } +} + + +void test_cxx11_tensor_broadcasting() +{ + CALL_SUBTEST(test_simple_broadcasting()); + CALL_SUBTEST(test_vectorized_broadcasting()); +} From 9c0aa81fbfc4abd0dc297d75305b9d579dad2754 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 21 Aug 2014 10:49:09 +0200 Subject: [PATCH 0728/1103] bug #854: fix numerical issue in SelfAdjointEigenSolver::computeDirect for 3x3 matrices. The tolerance to detect stable cross products was too optimistic. Add respective unit tests. --- .../src/Eigenvalues/SelfAdjointEigenSolver.h | 18 ++++++-- test/eigensolver_selfadjoint.cpp | 43 +++++++++++++------ 2 files changed, 44 insertions(+), 17 deletions(-) diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index fc8ecaa6f..a6bbdac6b 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -605,7 +605,6 @@ template struct direct_selfadjoint_eigenvalues::epsilon(); - safeNorm2 *= safeNorm2; if((eivals(2)-eivals(0))<=Eigen::NumTraits::epsilon()) { eivecs.setIdentity(); @@ -619,7 +618,7 @@ template struct direct_selfadjoint_eigenvalues d1 ? 2 : 0; - d0 = d0 > d1 ? d1 : d0; + d0 = d0 > d1 ? d0 : d1; tmp.diagonal().array () -= eivals(k); VectorType cross; @@ -627,19 +626,25 @@ template struct direct_selfadjoint_eigenvaluessafeNorm2) + { eivecs.col(k) = cross / sqrt(n); + } else { n = (cross = tmp.row(0).cross(tmp.row(2))).squaredNorm(); if(n>safeNorm2) + { eivecs.col(k) = cross / sqrt(n); + } else { n = (cross = tmp.row(1).cross(tmp.row(2))).squaredNorm(); if(n>safeNorm2) + { eivecs.col(k) = cross / sqrt(n); + } else { // the input matrix and/or the eigenvaues probably contains some inf/NaN, @@ -659,12 +664,16 @@ template struct direct_selfadjoint_eigenvalues::epsilon()) + { eivecs.col(1) = eivecs.col(k).unitOrthogonal(); + } else { - n = (cross = eivecs.col(k).cross(tmp.row(0).normalized())).squaredNorm(); + n = (cross = eivecs.col(k).cross(tmp.row(0))).squaredNorm(); if(n>safeNorm2) + { eivecs.col(1) = cross / sqrt(n); + } else { n = (cross = eivecs.col(k).cross(tmp.row(1))).squaredNorm(); @@ -678,13 +687,14 @@ template struct direct_selfadjoint_eigenvalues void selfadjointeigensolver(const MatrixType& m) MatrixType a = MatrixType::Random(rows,cols); MatrixType a1 = MatrixType::Random(rows,cols); MatrixType symmA = a.adjoint() * a + a1.adjoint() * a1; + MatrixType symmC = symmA; + + // randomly nullify some rows/columns + { + Index count = 1;//internal::random(-cols,cols); + for(Index k=0; k(0,cols-1); + symmA.row(i).setZero(); + symmA.col(i).setZero(); + } + } + symmA.template triangularView().setZero(); + symmC.template triangularView().setZero(); MatrixType b = MatrixType::Random(rows,cols); MatrixType b1 = MatrixType::Random(rows,cols); @@ -40,7 +54,7 @@ template void selfadjointeigensolver(const MatrixType& m) SelfAdjointEigenSolver eiDirect; eiDirect.computeDirect(symmA); // generalized eigen pb - GeneralizedSelfAdjointEigenSolver eiSymmGen(symmA, symmB); + GeneralizedSelfAdjointEigenSolver eiSymmGen(symmC, symmB); VERIFY_IS_EQUAL(eiSymm.info(), Success); VERIFY((symmA.template selfadjointView() * eiSymm.eigenvectors()).isApprox( @@ -57,27 +71,28 @@ template void selfadjointeigensolver(const MatrixType& m) VERIFY_IS_APPROX(eiSymm.eigenvalues(), eiSymmNoEivecs.eigenvalues()); // generalized eigen problem Ax = lBx - eiSymmGen.compute(symmA, symmB,Ax_lBx); + eiSymmGen.compute(symmC, symmB,Ax_lBx); VERIFY_IS_EQUAL(eiSymmGen.info(), Success); - VERIFY((symmA.template selfadjointView() * eiSymmGen.eigenvectors()).isApprox( + VERIFY((symmC.template selfadjointView() * eiSymmGen.eigenvectors()).isApprox( symmB.template selfadjointView() * (eiSymmGen.eigenvectors() * eiSymmGen.eigenvalues().asDiagonal()), largerEps)); // generalized eigen problem BAx = lx - eiSymmGen.compute(symmA, symmB,BAx_lx); + eiSymmGen.compute(symmC, symmB,BAx_lx); VERIFY_IS_EQUAL(eiSymmGen.info(), Success); - VERIFY((symmB.template selfadjointView() * (symmA.template selfadjointView() * eiSymmGen.eigenvectors())).isApprox( + VERIFY((symmB.template selfadjointView() * (symmC.template selfadjointView() * eiSymmGen.eigenvectors())).isApprox( (eiSymmGen.eigenvectors() * eiSymmGen.eigenvalues().asDiagonal()), largerEps)); // generalized eigen problem ABx = lx - eiSymmGen.compute(symmA, symmB,ABx_lx); + eiSymmGen.compute(symmC, symmB,ABx_lx); VERIFY_IS_EQUAL(eiSymmGen.info(), Success); - VERIFY((symmA.template selfadjointView() * (symmB.template selfadjointView() * eiSymmGen.eigenvectors())).isApprox( + VERIFY((symmC.template selfadjointView() * (symmB.template selfadjointView() * eiSymmGen.eigenvectors())).isApprox( (eiSymmGen.eigenvectors() * eiSymmGen.eigenvalues().asDiagonal()), largerEps)); + eiSymm.compute(symmC); MatrixType sqrtSymmA = eiSymm.operatorSqrt(); - VERIFY_IS_APPROX(MatrixType(symmA.template selfadjointView()), sqrtSymmA*sqrtSymmA); - VERIFY_IS_APPROX(sqrtSymmA, symmA.template selfadjointView()*eiSymm.operatorInverseSqrt()); + VERIFY_IS_APPROX(MatrixType(symmC.template selfadjointView()), sqrtSymmA*sqrtSymmA); + VERIFY_IS_APPROX(sqrtSymmA, symmC.template selfadjointView()*eiSymm.operatorInverseSqrt()); MatrixType id = MatrixType::Identity(rows, cols); VERIFY_IS_APPROX(id.template selfadjointView().operatorNorm(), RealScalar(1)); @@ -95,9 +110,9 @@ template void selfadjointeigensolver(const MatrixType& m) VERIFY_RAISES_ASSERT(eiSymmUninitialized.operatorInverseSqrt()); // test Tridiagonalization's methods - Tridiagonalization tridiag(symmA); + Tridiagonalization tridiag(symmC); // FIXME tridiag.matrixQ().adjoint() does not work - VERIFY_IS_APPROX(MatrixType(symmA.template selfadjointView()), tridiag.matrixQ() * tridiag.matrixT().eval() * MatrixType(tridiag.matrixQ()).adjoint()); + VERIFY_IS_APPROX(MatrixType(symmC.template selfadjointView()), tridiag.matrixQ() * tridiag.matrixT().eval() * MatrixType(tridiag.matrixQ()).adjoint()); // Test computation of eigenvalues from tridiagonal matrix if(rows > 1) @@ -111,8 +126,8 @@ template void selfadjointeigensolver(const MatrixType& m) if (rows > 1) { // Test matrix with NaN - symmA(0,0) = std::numeric_limits::quiet_NaN(); - SelfAdjointEigenSolver eiSymmNaN(symmA); + symmC(0,0) = std::numeric_limits::quiet_NaN(); + SelfAdjointEigenSolver eiSymmNaN(symmC); VERIFY_IS_EQUAL(eiSymmNaN.info(), NoConvergence); } } @@ -122,8 +137,10 @@ void test_eigensolver_selfadjoint() int s = 0; for(int i = 0; i < g_repeat; i++) { // very important to test 3x3 and 2x2 matrices since we provide special paths for them + CALL_SUBTEST_1( selfadjointeigensolver(Matrix2f()) ); CALL_SUBTEST_1( selfadjointeigensolver(Matrix2d()) ); CALL_SUBTEST_1( selfadjointeigensolver(Matrix3f()) ); + CALL_SUBTEST_1( selfadjointeigensolver(Matrix3d()) ); CALL_SUBTEST_2( selfadjointeigensolver(Matrix4d()) ); s = internal::random(1,EIGEN_TEST_MAX_SIZE/4); CALL_SUBTEST_3( selfadjointeigensolver(MatrixXf(s,s)) ); From b49ef99617b11404007db283ce68f5dd5b2eff0e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 21 Aug 2014 22:14:25 +0200 Subject: [PATCH 0729/1103] Do not apply the preconditioner before starting the iterations as this might destroy a very good initial guess. --- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 1 - 1 file changed, 1 deletion(-) diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index dc524c225..27824b9d5 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -39,7 +39,6 @@ bool bicgstab(const MatrixType& mat, const Rhs& rhs, Dest& x, int maxIters = iters; int n = mat.cols(); - x = precond.solve(x); VectorType r = rhs - mat * x; VectorType r0 = r; From 2e50289ba323740ef89eb51a74e00e645a8eb9be Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 26 Aug 2014 12:54:19 +0200 Subject: [PATCH 0730/1103] bug #861: enable posix_memalign with PGI --- Eigen/src/Core/util/Memory.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 810ee786b..30133ba67 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -64,7 +64,7 @@ // Currently, let's include it only on unix systems: #if defined(__unix__) || defined(__unix) #include - #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0) + #if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || (defined __PGI) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0) #define EIGEN_HAS_POSIX_MEMALIGN 1 #endif #endif From be3477e2069cbfe7158a00c11e265d81d7e88e7e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 26 Aug 2014 12:52:29 +0200 Subject: [PATCH 0731/1103] bug #857: workaround MSVC compilation issue. --- Eigen/src/SparseCore/SparseDenseProduct.h | 9 --------- Eigen/src/SparseCore/SparseMatrixBase.h | 3 ++- 2 files changed, 2 insertions(+), 10 deletions(-) diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 4a7813296..d40e966c1 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -318,15 +318,6 @@ class DenseTimeSparseProduct DenseTimeSparseProduct& operator=(const DenseTimeSparseProduct&); }; -// sparse * dense -template -template -inline const typename SparseDenseProductReturnType::Type -SparseMatrixBase::operator*(const MatrixBase &other) const -{ - return typename SparseDenseProductReturnType::Type(derived(), other.derived()); -} - } // end namespace Eigen #endif // EIGEN_SPARSEDENSEPRODUCT_H diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index 1050cf3f1..fb5025049 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -358,7 +358,8 @@ template class SparseMatrixBase : public EigenBase /** sparse * dense (returns a dense object unless it is an outer product) */ template const typename SparseDenseProductReturnType::Type - operator*(const MatrixBase &other) const; + operator*(const MatrixBase &other) const + { return typename SparseDenseProductReturnType::Type(derived(), other.derived()); } /** \returns an expression of P H P^-1 where H is the matrix represented by \c *this */ SparseSymmetricPermutationProduct twistedBy(const PermutationMatrix& perm) const From 25a3e65a685f66f04b83003f12cd97b91e646c20 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 26 Aug 2014 23:32:32 +0200 Subject: [PATCH 0732/1103] In SparseQR, calling factorize() without analyzePattern() was broken. --- Eigen/src/SparseQR/SparseQR.h | 16 ++++++++++++++-- test/sparseqr.cpp | 2 ++ 2 files changed, 16 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 66dccfc43..002b4824b 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -75,7 +75,7 @@ class SparseQR typedef Matrix ScalarVector; typedef PermutationMatrix PermutationType; public: - SparseQR () : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false) + SparseQR () : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) { } /** Construct a QR factorization of the matrix \a mat. @@ -84,7 +84,7 @@ class SparseQR * * \sa compute() */ - SparseQR(const MatrixType& mat) : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false) + SparseQR(const MatrixType& mat) : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) { compute(mat); } @@ -262,6 +262,7 @@ class SparseQR IndexVector m_etree; // Column elimination tree IndexVector m_firstRowElt; // First element in each row bool m_isQSorted; // whether Q is sorted or not + bool m_isEtreeOk; // whether the elimination tree match the initial input matrix template friend struct SparseQR_QProduct; template friend struct SparseQRMatrixQReturnType; @@ -297,6 +298,7 @@ void SparseQR::analyzePattern(const MatrixType& mat) // Compute the column elimination tree of the permuted matrix m_outputPerm_c = m_perm_c.inverse(); internal::coletree(mat, m_etree, m_firstRowElt, m_outputPerm_c.indices().data()); + m_isEtreeOk = true; m_R.resize(m, n); m_Q.resize(m, diagSize); @@ -330,6 +332,15 @@ void SparseQR::factorize(const MatrixType& mat) Index nzcolR, nzcolQ; // Number of nonzero for the current column of R and Q ScalarVector tval(m); // The dense vector used to compute the current column RealScalar pivotThreshold = m_threshold; + + m_R.setZero(); + m_Q.setZero(); + if(!m_isEtreeOk) + { + m_outputPerm_c = m_perm_c.inverse(); + internal::coletree(mat, m_etree, m_firstRowElt, m_outputPerm_c.indices().data()); + m_isEtreeOk = true; + } m_pmat = mat; m_pmat.uncompress(); // To have the innerNonZeroPtr allocated @@ -513,6 +524,7 @@ void SparseQR::factorize(const MatrixType& mat) // Recompute the column elimination tree internal::coletree(m_pmat, m_etree, m_firstRowElt, m_pivotperm.indices().data()); + m_isEtreeOk = false; } } diff --git a/test/sparseqr.cpp b/test/sparseqr.cpp index 1fe4a98ee..8e6887dd3 100644 --- a/test/sparseqr.cpp +++ b/test/sparseqr.cpp @@ -54,6 +54,8 @@ template void test_sparseqr_scalar() b = dA * DenseVector::Random(A.cols()); solver.compute(A); + if(internal::random(0,1)>0.5) + solver.factorize(A); // this checks that calling analyzePattern is not needed if the pattern do not change. if (solver.info() != Success) { std::cerr << "sparse QR factorization failed\n"; From 0ba490cf80d9c389de410beaa3551b2a2a72a801 Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Fri, 22 Aug 2014 12:13:07 +0200 Subject: [PATCH 0733/1103] Fixed CMakeLists.txt files to prevent CMake 3.0.0 warnings about deprecated LOCATION target property. Small whitespace cleanup in CMakelLists.txt. --- doc/examples/CMakeLists.txt | 4 +--- doc/snippets/CMakeLists.txt | 6 ++---- doc/special_examples/CMakeLists.txt | 16 +++++----------- unsupported/doc/examples/CMakeLists.txt | 4 +--- unsupported/doc/snippets/CMakeLists.txt | 4 +--- 5 files changed, 10 insertions(+), 24 deletions(-) diff --git a/doc/examples/CMakeLists.txt b/doc/examples/CMakeLists.txt index 71b272a15..08cf8efd7 100644 --- a/doc/examples/CMakeLists.txt +++ b/doc/examples/CMakeLists.txt @@ -6,12 +6,10 @@ foreach(example_src ${examples_SRCS}) if(EIGEN_STANDARD_LIBRARIES_TO_LINK_TO) target_link_libraries(${example} ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO}) endif() - get_target_property(example_executable - ${example} LOCATION) add_custom_command( TARGET ${example} POST_BUILD - COMMAND ${example_executable} + COMMAND ${example} ARGS >${CMAKE_CURRENT_BINARY_DIR}/${example}.out ) add_dependencies(all_examples ${example}) diff --git a/doc/snippets/CMakeLists.txt b/doc/snippets/CMakeLists.txt index 92a22ea61..1135900cf 100644 --- a/doc/snippets/CMakeLists.txt +++ b/doc/snippets/CMakeLists.txt @@ -14,12 +14,10 @@ foreach(snippet_src ${snippets_SRCS}) if(EIGEN_STANDARD_LIBRARIES_TO_LINK_TO) target_link_libraries(${compile_snippet_target} ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO}) endif() - get_target_property(compile_snippet_executable - ${compile_snippet_target} LOCATION) add_custom_command( TARGET ${compile_snippet_target} POST_BUILD - COMMAND ${compile_snippet_executable} + COMMAND ${compile_snippet_target} ARGS >${CMAKE_CURRENT_BINARY_DIR}/${snippet}.out ) add_dependencies(all_snippets ${compile_snippet_target}) @@ -27,4 +25,4 @@ foreach(snippet_src ${snippets_SRCS}) PROPERTIES OBJECT_DEPENDS ${snippet_src}) endforeach(snippet_src) -ei_add_target_property(compile_tut_arithmetic_transpose_aliasing COMPILE_FLAGS -DEIGEN_NO_DEBUG) \ No newline at end of file +ei_add_target_property(compile_tut_arithmetic_transpose_aliasing COMPILE_FLAGS -DEIGEN_NO_DEBUG) diff --git a/doc/special_examples/CMakeLists.txt b/doc/special_examples/CMakeLists.txt index 45e2339e3..aab80a55d 100644 --- a/doc/special_examples/CMakeLists.txt +++ b/doc/special_examples/CMakeLists.txt @@ -1,4 +1,3 @@ - if(NOT EIGEN_TEST_NOQT) find_package(Qt4) if(QT4_FOUND) @@ -6,17 +5,16 @@ if(NOT EIGEN_TEST_NOQT) endif() endif(NOT EIGEN_TEST_NOQT) - if(QT4_FOUND) add_executable(Tutorial_sparse_example Tutorial_sparse_example.cpp Tutorial_sparse_example_details.cpp) target_link_libraries(Tutorial_sparse_example ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO} ${QT_QTCORE_LIBRARY} ${QT_QTGUI_LIBRARY}) - + add_custom_command( TARGET Tutorial_sparse_example POST_BUILD COMMAND Tutorial_sparse_example ARGS ${CMAKE_CURRENT_BINARY_DIR}/../html/Tutorial_sparse_example.jpeg ) - + add_dependencies(all_examples Tutorial_sparse_example) endif(QT4_FOUND) @@ -26,15 +24,11 @@ if(EIGEN_COMPILER_SUPPORT_CPP11) target_link_libraries(random_cpp11 ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO}) add_dependencies(all_examples random_cpp11) ei_add_target_property(random_cpp11 COMPILE_FLAGS "-std=c++11") - - get_target_property(random_cpp11_exec - random_cpp11 LOCATION) + add_custom_command( TARGET random_cpp11 POST_BUILD - COMMAND ${random_cpp11_exec} + COMMAND random_cpp11 ARGS >${CMAKE_CURRENT_BINARY_DIR}/random_cpp11.out ) - - -endif() \ No newline at end of file +endif() diff --git a/unsupported/doc/examples/CMakeLists.txt b/unsupported/doc/examples/CMakeLists.txt index 978f9afd8..c47646dfc 100644 --- a/unsupported/doc/examples/CMakeLists.txt +++ b/unsupported/doc/examples/CMakeLists.txt @@ -10,12 +10,10 @@ FOREACH(example_src ${examples_SRCS}) if(EIGEN_STANDARD_LIBRARIES_TO_LINK_TO) target_link_libraries(example_${example} ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO}) endif() - GET_TARGET_PROPERTY(example_executable - example_${example} LOCATION) ADD_CUSTOM_COMMAND( TARGET example_${example} POST_BUILD - COMMAND ${example_executable} + COMMAND example_${example} ARGS >${CMAKE_CURRENT_BINARY_DIR}/${example}.out ) ADD_DEPENDENCIES(unsupported_examples example_${example}) diff --git a/unsupported/doc/snippets/CMakeLists.txt b/unsupported/doc/snippets/CMakeLists.txt index 4a4157933..f0c5cc2a8 100644 --- a/unsupported/doc/snippets/CMakeLists.txt +++ b/unsupported/doc/snippets/CMakeLists.txt @@ -14,12 +14,10 @@ FOREACH(snippet_src ${snippets_SRCS}) if(EIGEN_STANDARD_LIBRARIES_TO_LINK_TO) target_link_libraries(${compile_snippet_target} ${EIGEN_STANDARD_LIBRARIES_TO_LINK_TO}) endif() - GET_TARGET_PROPERTY(compile_snippet_executable - ${compile_snippet_target} LOCATION) ADD_CUSTOM_COMMAND( TARGET ${compile_snippet_target} POST_BUILD - COMMAND ${compile_snippet_executable} + COMMAND ${compile_snippet_target} ARGS >${CMAKE_CURRENT_BINARY_DIR}/${snippet}.out ) ADD_DEPENDENCIES(unsupported_snippets ${compile_snippet_target}) From fb5c1e9097886616d40a0988af5ca706292e54eb Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Sat, 23 Aug 2014 13:18:30 -0700 Subject: [PATCH 0734/1103] Optimized and cleaned up the tensor morphing code --- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 218 +++++------------- 1 file changed, 63 insertions(+), 155 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index ca3735d64..d9a6b3f1b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -127,7 +127,7 @@ struct TensorEvaluator, Device> return m_impl.template packet(index); } - Scalar* data() const { return NULL; } + Scalar* data() const { return m_impl.data(); } protected: NewDimensions m_dimensions; @@ -136,10 +136,12 @@ struct TensorEvaluator, Device> // Eval as lvalue -// TODO(bsteiner): share the code with the evaluator for rvalue reshapes. template -struct TensorEvaluator, Device> + struct TensorEvaluator, Device> + : public TensorEvaluator, Device> + { + typedef TensorEvaluator, Device> Base; typedef TensorReshapingOp XprType; typedef NewDimensions Dimensions; @@ -149,7 +151,7 @@ struct TensorEvaluator, Device> }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) - : m_impl(op.expression(), device), m_dimensions(op.dimensions()) + : Base(op, device) { } typedef typename XprType::Index Index; @@ -157,40 +159,15 @@ struct TensorEvaluator, Device> typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { - return m_impl.evalSubExprsIfNeeded(data); - } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { - m_impl.cleanup(); - } - - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const - { - return m_impl.coeff(index); - } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) { - return m_impl.coeffRef(index); + return this->m_impl.coeffRef(index); } template EIGEN_STRONG_INLINE void writePacket(Index index, const PacketReturnType& x) { - m_impl.template writePacket(index, x); + this->m_impl.template writePacket(index, x); } - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const - { - return m_impl.template packet(index); - } - - Scalar* data() const { return NULL; } - - private: - NewDimensions m_dimensions; - TensorEvaluator m_impl; }; @@ -286,7 +263,7 @@ struct TensorEvaluator, Devi }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) - : m_impl(op.expression(), device), m_dimensions(op.sizes()), m_offsets(op.startIndices()) + : m_impl(op.expression(), device), m_device(device), m_dimensions(op.sizes()), m_offsets(op.startIndices()) { for (int i = 0; i < internal::array_size::value; ++i) { eigen_assert(m_impl.dimensions()[i] >= op.sizes()[i] + op.startIndices()[i]); @@ -321,24 +298,37 @@ struct TensorEvaluator, Devi EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { m_impl.evalSubExprsIfNeeded(NULL); + if (data && m_impl.data()) { + Index contiguous_values = 1; + for (int i = 0; i < NumDims; ++i) { + contiguous_values *= dimensions()[i]; + if (dimensions()[i] != m_impl.dimensions()[i]) { + break; + } + } + // Use memcpy if it's going to be faster than using the regular evaluation. + if (contiguous_values > 2 * m_device.numThreads()) { + Scalar* src = m_impl.data(); + for (int i = 0; i < internal::array_prod(dimensions()); i += contiguous_values) { + Index offset = srcCoeff(i); + m_device.memcpy(data+i, src+offset, contiguous_values * sizeof(Scalar)); + } + return false; + } + } return true; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { m_impl.cleanup(); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { - Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_fastOutputStrides[i]; - inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; - index -= idx * m_outputStrides[i]; - } - inputIndex += (index + m_offsets[0]); - return m_impl.coeff(inputIndex); + return m_impl.coeff(srcCoeff(index)); } template @@ -376,23 +366,37 @@ struct TensorEvaluator, Devi } } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { return NULL; } + + protected: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_fastOutputStrides[i]; + inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + inputIndex += (index + m_offsets[0]); + return inputIndex; + } - private: Dimensions m_dimensions; array m_outputStrides; array, NumDims> m_fastOutputStrides; array m_inputStrides; const StartIndices m_offsets; TensorEvaluator m_impl; + const Device& m_device; }; // Eval as lvalue -// TODO(bsteiner): share the code with the evaluator for rvalue slices. template struct TensorEvaluator, Device> + : public TensorEvaluator, Device> { + typedef TensorEvaluator, Device> Base; typedef TensorSlicingOp XprType; static const int NumDims = internal::array_size::value; @@ -402,32 +406,8 @@ struct TensorEvaluator, Device> }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) - : m_impl(op.expression(), device), m_dimensions(op.sizes()), m_offsets(op.startIndices()) - { - for (int i = 0; i < internal::array_size::value; ++i) { - eigen_assert(m_impl.dimensions()[i] >= op.sizes()[i] + op.startIndices()[i]); - } - - const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; - } else { - m_inputStrides[0] = 1; - } - } - - const Sizes& output_dims = op.sizes(); - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { - m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1]; - m_fastOutputStrides[i] = internal::TensorIntDivisor(m_outputStrides[i]); - } else { - m_outputStrides[0] = 1; - m_fastOutputStrides[0] = 1; - } - } - } + : Base(op, device) + { } typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -435,71 +415,9 @@ struct TensorEvaluator, Device> typedef typename XprType::PacketReturnType PacketReturnType; typedef Sizes Dimensions; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { - m_impl.evalSubExprsIfNeeded(NULL); - return true; - } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { - m_impl.cleanup(); - } - - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const - { - Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_fastOutputStrides[i]; - inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; - index -= idx * m_outputStrides[i]; - } - inputIndex += (index + m_offsets[0]); - return m_impl.coeff(inputIndex); - } - - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const - { - static const int packetSize = internal::unpacket_traits::size; - EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) - Index inputIndices[] = {0, 0}; - Index indices[] = {index, index + packetSize - 1}; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / m_fastOutputStrides[i]; - const Index idx1 = indices[1] / m_fastOutputStrides[i]; - inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; - inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; - indices[0] -= idx0 * m_outputStrides[i]; - indices[1] -= idx1 * m_outputStrides[i]; - } - inputIndices[0] += (indices[0] + m_offsets[0]); - inputIndices[1] += (indices[1] + m_offsets[0]); - if (inputIndices[1] - inputIndices[0] == packetSize - 1) { - PacketReturnType rslt = m_impl.template packet(inputIndices[0]); - return rslt; - } - else { - CoeffReturnType values[packetSize]; - values[0] = m_impl.coeff(inputIndices[0]); - values[packetSize-1] = m_impl.coeff(inputIndices[1]); - for (int i = 1; i < packetSize-1; ++i) { - values[i] = coeff(index+i); - } - PacketReturnType rslt = internal::pload(values); - return rslt; - } - } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) { - Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_fastOutputStrides[i]; - inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; - index -= idx * m_outputStrides[i]; - } - inputIndex += (index + m_offsets[0]); - return m_impl.coeffRef(inputIndex); + return this->m_impl.coeffRef(this->srcCoeff(index)); } template EIGEN_STRONG_INLINE @@ -509,38 +427,28 @@ struct TensorEvaluator, Device> Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / m_fastOutputStrides[i]; - const Index idx1 = indices[1] / m_fastOutputStrides[i]; - inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; - inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; - indices[0] -= idx0 * m_outputStrides[i]; - indices[1] -= idx1 * m_outputStrides[i]; + const Index idx0 = indices[0] / this->m_fastOutputStrides[i]; + const Index idx1 = indices[1] / this->m_fastOutputStrides[i]; + inputIndices[0] += (idx0 + this->m_offsets[i]) * this->m_inputStrides[i]; + inputIndices[1] += (idx1 + this->m_offsets[i]) * this->m_inputStrides[i]; + indices[0] -= idx0 * this->m_outputStrides[i]; + indices[1] -= idx1 * this->m_outputStrides[i]; } - inputIndices[0] += (indices[0] + m_offsets[0]); - inputIndices[1] += (indices[1] + m_offsets[0]); + inputIndices[0] += (indices[0] + this->m_offsets[0]); + inputIndices[1] += (indices[1] + this->m_offsets[0]); if (inputIndices[1] - inputIndices[0] == packetSize - 1) { - m_impl.template writePacket(inputIndices[0], x); + this->m_impl.template writePacket(inputIndices[0], x); } else { CoeffReturnType values[packetSize]; internal::pstore(values, x); - m_impl.coeffRef(inputIndices[0]) = values[0]; - m_impl.coeffRef(inputIndices[1]) = values[packetSize-1]; + this->m_impl.coeffRef(inputIndices[0]) = values[0]; + this->m_impl.coeffRef(inputIndices[1]) = values[packetSize-1]; for (int i = 1; i < packetSize-1; ++i) { - coeffRef(index+i) = values[i]; + this->coeffRef(index+i) = values[i]; } } } - - Scalar* data() const { return NULL; } - - private: - Dimensions m_dimensions; - array m_outputStrides; - array, NumDims> m_fastOutputStrides; - array m_inputStrides; - const StartIndices m_offsets; - TensorEvaluator m_impl; }; From 36fffe48f7231e07915ec231d33cf46faa0fa918 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Sat, 23 Aug 2014 14:35:41 -0700 Subject: [PATCH 0735/1103] Misc api improvements and cleanups --- .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 9 +++++ .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 26 +++++++++++++ unsupported/test/CMakeLists.txt | 2 +- unsupported/test/cxx11_tensor_morphing.cpp | 37 ++++++++++--------- 4 files changed, 55 insertions(+), 19 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h index b9c8c19fe..ef5e11537 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -27,6 +27,10 @@ struct DefaultDevice { EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { ::memset(buffer, c, n); } + + EIGEN_STRONG_INLINE size_t numThreads() const { + return 1; + } }; @@ -115,6 +119,11 @@ struct GpuDevice { cudaMemsetAsync(buffer, c, n, *stream_); } + EIGEN_STRONG_INLINE size_t numThreads() const { + // Fixme: + return 32; + } + private: // TODO: multigpu. const cudaStream_t* stream_; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 3b169a06f..5a113dc19 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -195,6 +195,32 @@ struct DSizes : array { } EIGEN_DEVICE_FUNC explicit DSizes(const array& a) : Base(a) { } + EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0) { + (*this)[0] = i0; + } + EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1) { + (*this)[0] = i0; + (*this)[1] = i1; + } + EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2) { + (*this)[0] = i0; + (*this)[1] = i1; + (*this)[2] = i2; + } + EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3) { + (*this)[0] = i0; + (*this)[1] = i1; + (*this)[2] = i2; + (*this)[3] = i3; + } + EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3, const DenseIndex i4) { + (*this)[0] = i0; + (*this)[1] = i1; + (*this)[2] = i2; + (*this)[3] = i3; + (*this)[4] = i4; + } + DSizes& operator = (const array& other) { *static_cast(this) = other; return *this; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 164388746..615ff3e6d 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -110,7 +110,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") ei_add_test(cxx11_tensor_broadcasting "-std=c++0x") -# ei_add_test(cxx11_tensor_morphing "-std=c++0x") + ei_add_test(cxx11_tensor_morphing "-std=c++0x") ei_add_test(cxx11_tensor_padding "-std=c++0x") # ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp index fbfdaadb7..2a6a97856 100644 --- a/unsupported/test/cxx11_tensor_morphing.cpp +++ b/unsupported/test/cxx11_tensor_morphing.cpp @@ -52,7 +52,7 @@ static void test_reshape_in_expr() { TensorMap> tensor2(m2.data(), 3,5,7,11,13); Tensor::Dimensions newDims1{{2,3*5*7*11}}; Tensor::Dimensions newDims2{{3*5*7*11,13}}; - array::DimensionPair, 1> contract_along{{1, 0}}; + array::DimensionPair, 1> contract_along{{std::make_pair(1, 0)}}; Tensor tensor3(2,13); tensor3 = tensor1.reshape(newDims1).contract(tensor2.reshape(newDims2), contract_along); @@ -74,7 +74,8 @@ static void test_reshape_as_lvalue() Tensor::Dimensions dim{{2,3,7}}; tensor2d.reshape(dim) = tensor; - Tensor tensor5d(2,3,1,7,1); + float scratch[2*3*1*7*1]; + TensorMap> tensor5d(scratch, 2,3,1,7,1); tensor5d.reshape(dim).device(Eigen::DefaultDevice()) = tensor; for (int i = 0; i < 2; ++i) { @@ -94,14 +95,14 @@ static void test_simple_slice() tensor.setRandom(); Tensor slice1(1,1,1,1,1); - Eigen::DSizes indices(Eigen::array(1,2,3,4,5)); - Eigen::DSizes sizes(Eigen::array(1,1,1,1,1)); + Eigen::DSizes indices(1,2,3,4,5); + Eigen::DSizes sizes(1,1,1,1,1); slice1 = tensor.slice(indices, sizes); VERIFY_IS_EQUAL(slice1(0,0,0,0,0), tensor(1,2,3,4,5)); Tensor slice2(1,1,2,2,3); - Eigen::DSizes indices2(Eigen::array(1,1,3,4,5)); - Eigen::DSizes sizes2(Eigen::array(1,1,2,2,3)); + Eigen::DSizes indices2(1,1,3,4,5); + Eigen::DSizes sizes2(1,1,2,2,3); slice2 = tensor.slice(indices2, sizes2); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 2; ++j) { @@ -124,12 +125,12 @@ static void test_slice_in_expr() { TensorMap> tensor1(m1.data(), 7, 7); TensorMap> tensor2(m2.data(), 3, 3); Tensor tensor3(3,1); - array::DimensionPair, 1> contract_along{{1, 0}}; + array::DimensionPair, 1> contract_along{{std::make_pair(1, 0)}}; - Eigen::DSizes indices1(Eigen::array(1,2)); - Eigen::DSizes sizes1(Eigen::array(3,3)); - Eigen::DSizes indices2(Eigen::array(0,2)); - Eigen::DSizes sizes2(Eigen::array(3,1)); + Eigen::DSizes indices1(1,2); + Eigen::DSizes sizes1(3,3); + Eigen::DSizes indices2(0,2); + Eigen::DSizes sizes2(3,1); tensor3 = tensor1.slice(indices1, sizes1).contract(tensor2.slice(indices2, sizes2), contract_along); Map res(tensor3.data(), 3, 1); @@ -153,18 +154,18 @@ static void test_slice_as_lvalue() tensor4.setRandom(); Tensor result(4,5,7); - Eigen::DSizes sizes12(Eigen::array(2,2,7)); - Eigen::DSizes first_slice(Eigen::array(0,0,0)); + Eigen::DSizes sizes12(2,2,7); + Eigen::DSizes first_slice(0,0,0); result.slice(first_slice, sizes12) = tensor1; - Eigen::DSizes second_slice(Eigen::array(2,0,0)); + Eigen::DSizes second_slice(2,0,0); result.slice(second_slice, sizes12).device(Eigen::DefaultDevice()) = tensor2; - Eigen::DSizes sizes3(Eigen::array(4,3,5)); - Eigen::DSizes third_slice(Eigen::array(0,2,0)); + Eigen::DSizes sizes3(4,3,5); + Eigen::DSizes third_slice(0,2,0); result.slice(third_slice, sizes3) = tensor3; - Eigen::DSizes sizes4(Eigen::array(4,3,2)); - Eigen::DSizes fourth_slice(Eigen::array(0,2,5)); + Eigen::DSizes sizes4(4,3,2); + Eigen::DSizes fourth_slice(0,2,5); result.slice(fourth_slice, sizes4) = tensor4; for (int j = 0; j < 2; ++j) { From 2959045f2fe111f93b23517fd6f7afe49720a290 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 26 Aug 2014 09:47:18 -0700 Subject: [PATCH 0736/1103] Optimized the tensor padding code. --- .../Eigen/CXX11/src/Tensor/TensorPadding.h | 95 ++++++++++++++++--- 1 file changed, 81 insertions(+), 14 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h index 45558d7dd..4482c0992 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h @@ -87,7 +87,7 @@ struct TensorEvaluator, Device enum { IsAligned = false, - PacketAccess = /*TensorEvaluator::PacketAccess*/false, + PacketAccess = TensorEvaluator::PacketAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -100,15 +100,13 @@ struct TensorEvaluator, Device } const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; - m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; - } else { - m_inputStrides[0] = 1; - m_outputStrides[0] = 1; - } + m_inputStrides[0] = 1; + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; } + m_outputStrides[NumDims] = m_outputStrides[NumDims-1] * m_dimensions[NumDims-1]; } typedef typename XprType::Scalar Scalar; @@ -128,7 +126,7 @@ struct TensorEvaluator, Device EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { Index inputIndex = 0; - for (int i = NumDims - 1; i >= 0; --i) { + for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { return Scalar(0); @@ -136,21 +134,90 @@ struct TensorEvaluator, Device inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; index -= idx * m_outputStrides[i]; } + if (index < m_padding[0].first || index >= m_dimensions[0] - m_padding[0].second) { + return Scalar(0); + } + inputIndex += (index - m_padding[0].first); return m_impl.coeff(inputIndex); } - /* template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - return m_impl.template packet(index); - }*/ + static const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); + + const Index initialIndex = index; + Index inputIndex = 0; + for (int i = NumDims - 1; i > 0; --i) { + const int first = index; + const int last = index + packetSize - 1; + const int lastPaddedLeft = m_padding[i].first * m_outputStrides[i]; + const int firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i]; + const int lastPaddedRight = m_outputStrides[i+1]; + + if (last < lastPaddedLeft) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= firstPaddedRight && last < lastPaddedRight) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= lastPaddedLeft && last < firstPaddedRight) { + // all the coefficient are between the 2 padding zones. + const Index idx = index / m_outputStrides[i]; + inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + else { + // Every other case + return packetWithPossibleZero(initialIndex); + } + } + + const Index last = index + packetSize - 1; + const Index first = index; + const int lastPaddedLeft = m_padding[0].first; + const int firstPaddedRight = (m_dimensions[0] - m_padding[0].second); + const int lastPaddedRight = m_outputStrides[1]; + + if (last < lastPaddedLeft) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= firstPaddedRight && last < lastPaddedRight) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= lastPaddedLeft && last < firstPaddedRight) { + // all the coefficient are between the 2 padding zones. + inputIndex += (index - m_padding[0].first); + return m_impl.template packet(inputIndex); + } + // Every other case + return packetWithPossibleZero(initialIndex); + } Scalar* data() const { return NULL; } protected: + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const + { + static const int packetSize = internal::unpacket_traits::size; + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + for (int i = 0; i < packetSize; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + PaddingDimensions m_padding; Dimensions m_dimensions; - array m_outputStrides; + array m_outputStrides; array m_inputStrides; TensorEvaluator m_impl; }; From c3e408047427a12669720b64397e080956786829 Mon Sep 17 00:00:00 2001 From: Freddie Witherden Date: Wed, 27 Aug 2014 15:24:51 +0100 Subject: [PATCH 0737/1103] Allow LevenbergMarquardt to work with non-standard types. --- .../src/LevenbergMarquardt/LevenbergMarquardt.h | 8 +++++--- .../src/NonLinearOptimization/LevenbergMarquardt.h | 12 +++++++----- 2 files changed, 12 insertions(+), 8 deletions(-) diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h index 51dd1d3c4..7cebe4e06 100644 --- a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h +++ b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h @@ -144,11 +144,13 @@ class LevenbergMarquardt : internal::no_assignment_operator /** Sets the default parameters */ void resetParameters() - { + { + using std::sqrt; + m_factor = 100.; m_maxfev = 400; - m_ftol = std::sqrt(NumTraits::epsilon()); - m_xtol = std::sqrt(NumTraits::epsilon()); + m_ftol = sqrt(NumTraits::epsilon()); + m_xtol = sqrt(NumTraits::epsilon()); m_gtol = 0. ; m_epsfcn = 0. ; } diff --git a/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h b/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h index bfeb26fc9..ecb8dccf4 100644 --- a/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h +++ b/unsupported/Eigen/src/NonLinearOptimization/LevenbergMarquardt.h @@ -55,8 +55,8 @@ public: Parameters() : factor(Scalar(100.)) , maxfev(400) - , ftol(std::sqrt(NumTraits::epsilon())) - , xtol(std::sqrt(NumTraits::epsilon())) + , ftol(sqrt_(NumTraits::epsilon())) + , xtol(sqrt_(NumTraits::epsilon())) , gtol(Scalar(0.)) , epsfcn(Scalar(0.)) {} Scalar factor; @@ -72,7 +72,7 @@ public: LevenbergMarquardtSpace::Status lmder1( FVectorType &x, - const Scalar tol = std::sqrt(NumTraits::epsilon()) + const Scalar tol = sqrt_(NumTraits::epsilon()) ); LevenbergMarquardtSpace::Status minimize(FVectorType &x); @@ -83,12 +83,12 @@ public: FunctorType &functor, FVectorType &x, Index *nfev, - const Scalar tol = std::sqrt(NumTraits::epsilon()) + const Scalar tol = sqrt_(NumTraits::epsilon()) ); LevenbergMarquardtSpace::Status lmstr1( FVectorType &x, - const Scalar tol = std::sqrt(NumTraits::epsilon()) + const Scalar tol = sqrt_(NumTraits::epsilon()) ); LevenbergMarquardtSpace::Status minimizeOptimumStorage(FVectorType &x); @@ -109,6 +109,8 @@ public: Scalar lm_param(void) { return par; } private: + static Scalar sqrt_(const Scalar& x) { using std::sqrt; return sqrt(x); } + FunctorType &functor; Index n; Index m; From e49e84d97940a4201e2b50a2ffe01e66f19ad783 Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Fri, 29 Aug 2014 10:41:05 +0200 Subject: [PATCH 0738/1103] Added missing STL include of in main.h Removed duplicated include of Added comments on the background of min/max macro definitions and STL header includes --- test/main.h | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-) diff --git a/test/main.h b/test/main.h index 376232cf2..773873a0d 100644 --- a/test/main.h +++ b/test/main.h @@ -17,13 +17,36 @@ #include #include #include + +// The following includes of STL headers have to be done _before_ the +// definition of macros min() and max(). The reason is that many STL +// implementations will not work properly as the min and max symbols collide +// with the STL functions std:min() and std::max(). The STL headers may check +// for the macro definition of min/max and issue a warning or undefine the +// macros. +// +// Still, Windows defines min() and max() in windef.h as part of the regular +// Windows system interfaces and many other Windows APIs depend on these +// macros being available. To prevent the macro expansion of min/max and to +// make Eigen compatible with the Windows environment all function calls of +// std::min() and std::max() have to be written with parenthesis around the +// function name. +// +// All STL headers used by Eigen should be included here. Because main.h is +// included before any Eigen header and because the STL headers are guarded +// against multiple inclusions, no STL header will see our own min/max macro +// definitions. #include #include -#include #include #include #include +#include +// To test that all calls from Eigen code to std::min() and std::max() are +// protected by parenthesis against macro expansion, the min()/max() macros +// are defined here and any not-parenthesized min/max call will cause a +// compiler error. #define min(A,B) please_protect_your_min_with_parentheses #define max(A,B) please_protect_your_max_with_parentheses From 1ed9e2d0044a5c35052965ea53e4f905279a06f1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Aug 2014 11:55:03 +0200 Subject: [PATCH 0739/1103] In sparse matrix product, enable sorted insertion when doing two transposition is defenitely not optimal. --- .../ConservativeSparseSparseProduct.h | 82 ++++++++++--------- 1 file changed, 45 insertions(+), 37 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 5c320e2d2..608044a95 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -15,7 +15,7 @@ namespace Eigen { namespace internal { template -static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res) +static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res, bool sortedInsertion = false) { typedef typename remove_all::type::Scalar Scalar; typedef typename remove_all::type::Index Index; @@ -64,53 +64,51 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r values[i] += x * y; } } - - // unordered insertion - for(Index k=0; k use a quick sort - // otherwise => loop through the entire vector - // In order to avoid to perform an expensive log2 when the - // result is clearly very sparse we use a linear bound up to 200. - //if((nnz<200 && nnz1) std::sort(indices.data(),indices.data()+nnz); + // unordered insertion for(Index k=0; k use a quick sort + // otherwise => loop through the entire vector + // In order to avoid to perform an expensive log2 when the + // result is clearly very sparse we use a linear bound up to 200. + if((nnz<200 && nnz1) std::sort(indices.data(),indices.data()+nnz); + for(Index k=0; k RowMajorMatrix; typedef SparseMatrix ColMajorMatrix; ColMajorMatrix resCol(lhs.rows(),rhs.cols()); - internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol); - // sort the non zeros: - RowMajorMatrix resRow(resCol); - res = resRow; + // FIXME, the following heuristic is probably not very good. + if(lhs.rows()>=rhs.cols()) + { + // perform sorted insertion + internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, true); + res = resCol; + } + else + { + // ressort to transpose to sort the entries + internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, false); + RowMajorMatrix resRow(resCol); + res = resRow; + } } }; From 460662cbcc89c378b9ea097220b77d0eea9551ff Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Aug 2014 14:18:23 +0200 Subject: [PATCH 0740/1103] Fix SparseVector::coeffRef(i,j) and add missing SparseVector::insert*Unordered --- Eigen/src/SparseCore/SparseVector.h | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index 7e15c814b..0b1b389ce 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -109,7 +109,7 @@ class SparseVector inline Scalar& coeffRef(Index row, Index col) { eigen_assert(IsColVector ? (col==0 && row>=0 && row=0 && col Date: Fri, 29 Aug 2014 14:19:03 +0200 Subject: [PATCH 0741/1103] Optimization in sparse-sparse matrix products for small ones --- .../ConservativeSparseSparseProduct.h | 18 ++++++++++-------- 1 file changed, 10 insertions(+), 8 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 608044a95..8067565f9 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -24,10 +24,10 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r Index rows = lhs.innerSize(); Index cols = rhs.outerSize(); eigen_assert(lhs.outerSize() == rhs.innerSize()); - - std::vector mask(rows,false); - Matrix values(rows); - Matrix indices(rows); + + ei_declare_aligned_stack_constructed_variable(bool, mask, rows, 0); + ei_declare_aligned_stack_constructed_variable(Scalar, values, rows, 0); + ei_declare_aligned_stack_constructed_variable(Index, indices, rows, 0); // estimate the number of non zero entries // given a rhs column containing Y non zeros, we assume that the respective Y columns @@ -77,7 +77,7 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r else { // alternative ordered insertion code: - const Index t200 = rows/(log2(200)*1.39); + const Index t200 = rows/11; // 11 == (log2(200)*1.39) const Index t = (rows*100)/139; // FIXME reserve nnz non zeros @@ -88,7 +88,7 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r // result is clearly very sparse we use a linear bound up to 200. if((nnz<200 && nnz1) std::sort(indices.data(),indices.data()+nnz); + if(nnz>1) std::sort(indices,indices+nnz); for(Index k=0; k RowMajorMatrix; - typedef SparseMatrix ColMajorMatrix; + typedef SparseMatrix ColMajorMatrixAux; + typedef typename sparse_eval::type ColMajorMatrix; + ColMajorMatrix resCol(lhs.rows(),rhs.cols()); // FIXME, the following heuristic is probably not very good. if(lhs.rows()>=rhs.cols()) { // perform sorted insertion internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, true); - res = resCol; + res.swap(resCol); } else { From f29dbec321617d46287c4415889c4485ad70bea3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Aug 2014 15:12:03 +0200 Subject: [PATCH 0742/1103] undef Unsable macro --- Eigen/src/Core/ReturnByValue.h | 1 + 1 file changed, 1 insertion(+) diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index f15601d99..9d53fba27 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -81,6 +81,7 @@ template class ReturnByValue const Unusable& coeff(Index,Index) const { return *reinterpret_cast(this); } Unusable& coeffRef(Index) { return *reinterpret_cast(this); } Unusable& coeffRef(Index,Index) { return *reinterpret_cast(this); } +#undef Unusable #endif }; From c1d0f15bde4614c3efb84248e5d2ceb9cb995779 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 29 Aug 2014 15:31:32 +0200 Subject: [PATCH 0743/1103] Enable evaluators by default --- Eigen/Core | 11 +++++++---- test/CMakeLists.txt | 8 ++++---- test/evaluators.cpp | 4 ++++ 3 files changed, 15 insertions(+), 8 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 239bc6990..47c4f6299 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -12,10 +12,13 @@ #define EIGEN_CORE_H // EIGEN_TEST_EVALUATORS => EIGEN_ENABLE_EVALUATORS -#ifdef EIGEN_TEST_EVALUATORS -#ifndef EIGEN_ENABLE_EVALUATORS -#define EIGEN_ENABLE_EVALUATORS -#endif +#ifndef EIGEN_TEST_NO_EVALUATORS + #ifndef EIGEN_TEST_EVALUATORS + #define EIGEN_TEST_EVALUATORS + #endif + #ifndef EIGEN_ENABLE_EVALUATORS + #define EIGEN_ENABLE_EVALUATORS + #endif #endif // first thing Eigen does: stop the compiler from committing suicide diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 6446b8bb0..075b8d3de 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -139,10 +139,10 @@ endif(TEST_LIB) set_property(GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT "Official") add_custom_target(BuildOfficial) -option(EIGEN_TEST_EVALUATORS "Enable work in progress evaluators" OFF) -if(EIGEN_TEST_EVALUATORS) - add_definitions("-DEIGEN_TEST_EVALUATORS=1") -endif(EIGEN_TEST_EVALUATORS) +option(EIGEN_TEST_NO_EVALUATORS "Disable evaluators in unit tests" OFF) +if(EIGEN_TEST_NO_EVALUATORS) + add_definitions("-DEIGEN_TEST_NO_EVALUATORS=1") +endif(EIGEN_TEST_NO_EVALUATORS) ei_add_test(meta) ei_add_test(sizeof) diff --git a/test/evaluators.cpp b/test/evaluators.cpp index 305047a6a..2ca453b1c 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -7,6 +7,10 @@ #undef EIGEN_TEST_EVALUATORS #endif +#ifdef EIGEN_TEST_NO_EVALUATORS +#undef EIGEN_TEST_NO_EVALUATORS +#endif + #include "main.h" namespace Eigen { From 7ff266e3ce592ec1a6284cf16811965eec775f25 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Fri, 29 Aug 2014 20:03:49 +0000 Subject: [PATCH 0744/1103] Initial VSX commit --- CMakeLists.txt | 6 + Eigen/Core | 17 +- Eigen/src/Core/arch/AltiVec/Complex.h | 186 ++++++++++++++++++ Eigen/src/Core/arch/AltiVec/PacketMath.h | 64 +++++- .../Core/products/GeneralBlockPanelKernel.h | 4 +- Eigen/src/Core/util/Constants.h | 6 + bench/btl/libs/eigen2/eigen2_interface.hh | 2 +- cmake/EigenTesting.cmake | 6 + test/main.h | 2 +- test/packetmath.cpp | 2 +- 10 files changed, 284 insertions(+), 11 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 96d6c8701..ea42cc8db 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -219,6 +219,12 @@ if(NOT MSVC) message(STATUS "Enabling AltiVec in tests/examples") endif() + option(EIGEN_TEST_VSX "Enable/Disable VSX in tests/examples" OFF) + if(EIGEN_TEST_VSX) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64 -mvsx") + message(STATUS "Enabling VSX in tests/examples") + endif() + option(EIGEN_TEST_NEON "Enable/Disable Neon in tests/examples" OFF) if(EIGEN_TEST_NEON) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfpu=neon -mcpu=cortex-a8") diff --git a/Eigen/Core b/Eigen/Core index 776b7faf3..8ea165d5b 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -171,6 +171,15 @@ #undef bool #undef vector #undef pixel + #elif defined __VSX__ + #define EIGEN_VECTORIZE + #define EIGEN_VECTORIZE_VSX + #include + // We need to #undef all these ugly tokens defined in + // => use __vector instead of vector + #undef bool + #undef vector + #undef pixel #elif defined __ARM_NEON__ #define EIGEN_VECTORIZE #define EIGEN_VECTORIZE_NEON @@ -235,6 +244,8 @@ inline static const char *SimdInstructionSetsInUse(void) { return "SSE, SSE2"; #elif defined(EIGEN_VECTORIZE_ALTIVEC) return "AltiVec"; +#elif defined(EIGEN_VECTORIZE_VSX) + return "VSX"; #elif defined(EIGEN_VECTORIZE_NEON) return "ARM NEON"; #else @@ -286,9 +297,13 @@ using std::ptrdiff_t; #include "src/Core/arch/SSE/PacketMath.h" #include "src/Core/arch/SSE/MathFunctions.h" #include "src/Core/arch/SSE/Complex.h" -#elif defined EIGEN_VECTORIZE_ALTIVEC +#elif defined(EIGEN_VECTORIZE_ALTIVEC) #include "src/Core/arch/AltiVec/PacketMath.h" #include "src/Core/arch/AltiVec/Complex.h" +#elif defined(EIGEN_VECTORIZE_ALTIVEC) + #include "src/Core/arch/AltiVec/PacketMath.h" + #include "src/Core/arch/AltiVec/VSX.h" + #include "src/Core/arch/AltiVec/Complex.h" #elif defined EIGEN_VECTORIZE_NEON #include "src/Core/arch/NEON/PacketMath.h" #include "src/Core/arch/NEON/Complex.h" diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 13b874d0c..064341a3b 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -14,7 +14,13 @@ namespace Eigen { namespace internal { +#ifdef _BIG_ENDIAN static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; +#else +static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4f_ZERO_, (Packet4ui)p4i_ZERO);//{ 0x80000000, 0x00000000, 0x80000000, 0x00000000 }; +static Packet2ul p2ul_CONJ_XOR = (Packet2ul) vec_sld((Packet4ui)p2d_ZERO_, (Packet4ui)p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +#endif + static Packet16uc p16uc_COMPLEX_RE = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; static Packet16uc p16uc_COMPLEX_IM = vec_sld(p16uc_DUPLICATE, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; @@ -237,6 +243,186 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) kernel.packet[0].v = tmp; } +//---------- double ---------- +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + Packet2d v; +}; + +template<> struct packet_traits > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits { typedef std::complex type; enum {size=1}; typedef Packet1cd half; }; + +template<> EIGEN_STRONG_INLINE Packet1cd padd(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd psub(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2ul)a.v, p2ul_CONJ_XOR)); } + +template<> EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) +{ + Packet2d v1, v2; + + // Permute and multiply the real parts of a and b + v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE); + // Get the imaginary parts of a + v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM); + // multiply a_re * b + v1 = vec_madd(v1, b.v, p4f_ZERO); + // multiply a_im * b and get the conjugate result + v2 = vec_madd(v2, b.v, p4f_ZERO); + v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR); + // permute back to a proper order + v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV); + + return Packet2cf(vec_add(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pand (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd por (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pxor (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pandnot(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); } + +template<> EIGEN_STRONG_INLINE Packet1cd pload (const std::complex* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload((const double*)from)); } +template<> EIGEN_STRONG_INLINE Packet1cd ploadu(const std::complex* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu((const double*)from)); } + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex* from) +{ + return pset1(*from); +} + +template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { vec_dstt((double *)addr, DST_CTRL(2,2,32), DST_CHAN); } + +template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet1cd& a) +{ + std::complex EIGEN_ALIGN16 res[2]; + pstore((float *)&res, a.v); + + return res[0]; +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) +{ + Packet2d rev_a; + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2); + return Packet1cd(rev_a); +} + +template<> EIGEN_STRONG_INLINE std::complex predux(const Packet1cd& a) +{ + Packet2d b; + b = (Packet2d) vec_sld(a.v, a.v, 8); + b = padd(a.v, b); + return pfirst(Packet1cd(b)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd preduxp(const Packet1cd* vecs) +{ + Packet2d b1, b2; + + b1 = (Packet2d) vec_sld(vecs[0].v, vecs[1].v, 8); + b2 = (Packet2d) vec_sld(vecs[1].v, vecs[0].v, 8); + b2 = (Packet2d) vec_sld(b2, b2, 8); + b2 = padd(b1, b2); + + return Packet1cd(b2); +} + +template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet1cd& a) +{ + Packet2d b; + Packet1cd prod; + b = (Packet2d) vec_sld(a.v, a.v, 8); + prod = pmul(a, Packet1cd(b)); + + return pfirst(prod); +} + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet1cd& first, const Packet1cd& second) + { } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for AltiVec + Packet1cd res = conj_helper().pmul(a,b); + Packet2d s = vec_madd(b.v, b.v, p4f_ZERO); + return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV)))); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pcplxflip(const Packet1cd& x) +{ + return Packet1cd(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) +{ + Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); + kernel.packet[0].v = tmp; +} + } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index b43e8ace3..e70039ae2 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -31,6 +31,12 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 #endif +#ifdef __VSX__ +typedef __vector double Packet2d; +typedef __vector unsigned long Packet2ul; +typedef __vector long Packet2l; +#endif // __VSX__ + typedef __vector float Packet4f; typedef __vector int Packet4i; typedef __vector unsigned int Packet4ui; @@ -50,22 +56,37 @@ typedef __vector unsigned char Packet16uc; #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ Packet4f p4f_##NAME = pset1(X) -#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ - Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1(X)) - #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ Packet4i p4i_##NAME = pset1(X) +#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \ + Packet2d p2d_##NAME = pset1(X) + +#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \ + Packet2l p2l_##NAME = pset1(X) + #define DST_CHAN 1 #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride)) +// Handle endianness properly while loading constants // Define global static constants: +#ifdef _BIG_ENDIAN static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; static Packet16uc p16uc_REVERSE = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); //{ 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15} static Packet16uc p16uc_DUPLICATE = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; +#else +static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 }; +static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 }; +static Packet16uc p16uc_REVERSE = { 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15 }; +static Packet16uc p16uc_FORWARD = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; +static Packet16uc p16uc_DUPLICATE = { 4,5,6,7, 4,5,6,7, 0,1,2,3, 0,1,2,3 }; +static Packet2d p2d_ZERO_ = (Packet2d) { 0x8000000000000000, 0x8000000000000000 }; +static Packet2l p2l_ZERO = (Packet2l) { 0x0, 0x0 }; +#endif // _BIG_ENDIAN +// These constants are endian-agnostic static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0} static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,} static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1} @@ -93,6 +114,24 @@ template<> struct packet_traits : default_packet_traits HasSqrt = 0 }; }; +#ifdef __VSX__ +template<> struct packet_traits : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 0, + + HasDiv = 1, + HasExp = 0, + HasSqrt = 0 + }; +}; +#endif // __VSX__ + template<> struct packet_traits : default_packet_traits { typedef Packet4i type; @@ -105,6 +144,10 @@ template<> struct packet_traits : default_packet_traits }; }; +#ifdef __VSX__ +template<> struct unpacket_traits { typedef double type; enum {size=2}; typedef Packet2d half; }; +#endif // __VSX__ + template<> struct unpacket_traits { typedef float type; enum {size=4}; typedef Packet4f half; }; template<> struct unpacket_traits { typedef int type; enum {size=4}; typedef Packet4i half; }; /* @@ -311,7 +354,6 @@ template<> EIGEN_STRONG_INLINE Packet4i pmin(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pmax(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); } template<> EIGEN_STRONG_INLINE Packet4i pmax(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } -// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics template<> EIGEN_STRONG_INLINE Packet4f pand(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); } template<> EIGEN_STRONG_INLINE Packet4i pand(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); } @@ -327,10 +369,10 @@ template<> EIGEN_STRONG_INLINE Packet4i pandnot(const Packet4i& a, con template<> EIGEN_STRONG_INLINE Packet4f pload(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } template<> EIGEN_STRONG_INLINE Packet4i pload(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } +#ifndef __VSX__ template<> EIGEN_STRONG_INLINE Packet4f ploadu(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD - // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html Packet16uc MSQ, LSQ; Packet16uc mask; MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword @@ -350,6 +392,18 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) mask = vec_lvsl(0, from); // create the permute mask return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data } +#else +template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + return (Packet4i) vec_vsx_ld((long)from & 15, from); // align the data +} +template<> EIGEN_STRONG_INLINE Packet4f ploadu(const float* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + return (Packet4f) vec_vsx_ld((long)from & 15, from); // align the data +} +#endif template<> EIGEN_STRONG_INLINE Packet4f ploaddup(const float* from) { diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 7da52c2e8..7b2ed6728 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -182,7 +182,7 @@ public: nr = 4, // register block size along the M direction (currently, this one cannot be modified) -#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) +#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -290,7 +290,7 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, nr = 4, -#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) +#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) // we assume 16 registers mr = 3*LhsPacketSize, #else diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 05107fdfe..31073b990 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -413,10 +413,16 @@ namespace Architecture Generic = 0x0, SSE = 0x1, AltiVec = 0x2, + VSX = 0x3, + NEON = 0x4, #if defined EIGEN_VECTORIZE_SSE Target = SSE #elif defined EIGEN_VECTORIZE_ALTIVEC Target = AltiVec +#elif defined EIGEN_VECTORIZE_VSX + Target = VSX +#elif defined EIGEN_VECTORIZE_NEON + Target = NEON #else Target = Generic #endif diff --git a/bench/btl/libs/eigen2/eigen2_interface.hh b/bench/btl/libs/eigen2/eigen2_interface.hh index 47fe58135..1deabdae2 100644 --- a/bench/btl/libs/eigen2/eigen2_interface.hh +++ b/bench/btl/libs/eigen2/eigen2_interface.hh @@ -47,7 +47,7 @@ public : { #if defined(EIGEN_VECTORIZE_SSE) if (SIZE==Dynamic) return "eigen2"; else return "tiny_eigen2"; - #elif defined(EIGEN_VECTORIZE_ALTIVEC) + #elif defined(EIGEN_VECTORIZE_ALTIVEC) || defined(EIGEN_VECTORIZE_VSX) if (SIZE==Dynamic) return "eigen2"; else return "tiny_eigen2"; #else if (SIZE==Dynamic) return "eigen2_novec"; else return "tiny_eigen2_novec"; diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index 9b9776894..6383c6eac 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -282,6 +282,12 @@ macro(ei_testing_print_summary) message(STATUS "Altivec: Using architecture defaults") endif() + if(EIGEN_TEST_VSX) + message(STATUS "VSX: ON") + else() + message(STATUS "VSX: Using architecture defaults") + endif() + if(EIGEN_TEST_NEON) message(STATUS "ARM NEON: ON") else() diff --git a/test/main.h b/test/main.h index 773873a0d..7667eaa18 100644 --- a/test/main.h +++ b/test/main.h @@ -76,7 +76,7 @@ #endif // bounds integer values for AltiVec -#ifdef __ALTIVEC__ +#if defined(__ALTIVEC__) || defined(__VSX__) #define EIGEN_MAKING_DOCS #endif diff --git a/test/packetmath.cpp b/test/packetmath.cpp index e5dc473c2..e716d6d9a 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -156,7 +156,7 @@ template void packetmath() CHECK_CWISE2(REF_ADD, internal::padd); CHECK_CWISE2(REF_SUB, internal::psub); CHECK_CWISE2(REF_MUL, internal::pmul); - #ifndef EIGEN_VECTORIZE_ALTIVEC + #if !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) if (!internal::is_same::value) CHECK_CWISE2(REF_DIV, internal::pdiv); #endif From 0369db12af70fe5e63416a18c8236d419b0597c4 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 10:52:58 +0200 Subject: [PATCH 0745/1103] bug #871: fix compilation on ARM/Neon regarding __has_builtin usage --- Eigen/src/Core/arch/NEON/PacketMath.h | 2 +- Eigen/src/Core/util/Macros.h | 7 +++++++ 2 files changed, 8 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 380b76ae9..7c4509585 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -52,7 +52,7 @@ typedef uint32x4_t Packet4ui; // arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function // which available on LLVM and GCC (at least) -#if (defined(__has_builtin) && __has_builtin(__builtin_prefetch)) || defined(__GNUC__) +#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || defined(__GNUC__) #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR); #elif defined __pld #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 5e9b0a112..99e682653 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -107,6 +107,13 @@ #define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t #endif +// Cross compiler wrapper around LLVM's __has_builtin +#ifdef __has_builtin +# define EIGEN_HAS_BUILTIN(x) __has_builtin(x) +#else +# define EIGEN_HAS_BUILTIN(x) 0 +#endif + // A Clang feature extension to determine compiler features. // We use it to determine 'cxx_rvalue_references' #ifndef __has_feature From e6cc24cbd60113c704d8f88293d4efdd1eef1b98 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 14:20:11 +0200 Subject: [PATCH 0746/1103] Fix compilation in legacy mode --- Eigen/src/SparseCore/SparseDenseProduct.h | 8 -------- 1 file changed, 8 deletions(-) diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index a715b8bde..c8a515e8c 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -362,14 +362,6 @@ class DenseTimeSparseProduct DenseTimeSparseProduct& operator=(const DenseTimeSparseProduct&); }; -// sparse * dense -template -template -inline const typename SparseDenseProductReturnType::Type -SparseMatrixBase::operator*(const MatrixBase &other) const -{ - return typename SparseDenseProductReturnType::Type(derived(), other.derived()); -} #endif // EIGEN_TEST_EVALUATORS #ifdef EIGEN_TEST_EVALUATORS From bc065c75d2a8df928cb368d0352b8fcb25791fa8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 14:50:59 +0200 Subject: [PATCH 0747/1103] Implement the missing bits to make Solve compatible with sparse rhs --- Eigen/src/Core/Solve.h | 9 +++++++++ Eigen/src/SparseCore/SparseAssign.h | 12 ++++++++++++ Eigen/src/SparseCore/SparseUtil.h | 8 ++++++++ 3 files changed, 29 insertions(+) diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index c28789968..a1501c259 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -99,6 +99,15 @@ private: Scalar coeff(Index i) const; }; +#ifdef EIGEN_TEST_EVALUATORS +// Generic API dispatcher +template +class SolveImpl : public internal::generic_xpr_base, MatrixXpr, StorageKind>::type +{ + public: + typedef typename internal::generic_xpr_base, MatrixXpr, StorageKind>::type Base; +}; +#endif namespace internal { diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index 0a29cb32f..36c9fe845 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -284,6 +284,18 @@ struct Assignment +struct Assignment, internal::assign_op, Sparse2Sparse, Scalar> +{ + typedef Solve SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + src.dec()._solve_impl(src.rhs(), dst); + } +}; + } // end namespace internal #endif // EIGEN_TEST_EVALUATORS diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 0183907a1..28bf89bca 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -156,6 +156,14 @@ template struct plain_matrix_type typedef SparseMatrix<_Scalar, _Options, _Index> type; }; +#ifdef EIGEN_TEST_EVALUATORS +template +struct solve_traits +{ + typedef typename sparse_eval::type PlainObject; +}; +#endif + template struct generic_xpr_base { From 85c765957418cd2fd24b46ca3a14e6fcbad43f05 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 15:00:19 +0200 Subject: [PATCH 0748/1103] Refactoring of sparse solvers through a SparseSolverBase class and usage of the Solve<> expression. Introduce a SolveWithGuess expression on top of Solve. --- Eigen/IterativeLinearSolvers | 4 + Eigen/SparseCholesky | 3 + Eigen/SparseCore | 1 + Eigen/src/CholmodSupport/CholmodSupport.h | 30 +++-- .../BasicPreconditioners.h | 20 ++- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 41 +++++-- .../ConjugateGradient.h | 31 ++++- .../IterativeLinearSolvers/IncompleteLUT.h | 20 ++- .../IterativeSolverBase.h | 39 +++++- .../IterativeLinearSolvers/SolveWithGuess.h | 114 ++++++++++++++++++ Eigen/src/PaStiXSupport/PaStiXSupport.h | 32 ++--- Eigen/src/PardisoSupport/PardisoSupport.h | 44 +++---- Eigen/src/QR/HouseholderQR.h | 4 +- Eigen/src/SPQRSupport/SuiteSparseQRSupport.h | 33 ++--- Eigen/src/SparseCholesky/SimplicialCholesky.h | 41 ++++++- Eigen/src/SparseCore/SparseSolverBase.h | 113 +++++++++++++++++ Eigen/src/SparseLU/SparseLU.h | 35 ++++-- Eigen/src/SparseQR/SparseQR.h | 39 ++++-- Eigen/src/SuperLUSupport/SuperLUSupport.h | 31 +++-- Eigen/src/UmfPackSupport/UmfPackSupport.h | 21 ++-- 20 files changed, 565 insertions(+), 131 deletions(-) create mode 100644 Eigen/src/IterativeLinearSolvers/SolveWithGuess.h create mode 100644 Eigen/src/SparseCore/SparseSolverBase.h diff --git a/Eigen/IterativeLinearSolvers b/Eigen/IterativeLinearSolvers index 0f4159dc1..4df2c5a14 100644 --- a/Eigen/IterativeLinearSolvers +++ b/Eigen/IterativeLinearSolvers @@ -26,8 +26,12 @@ * \endcode */ +#ifndef EIGEN_TEST_EVALUATORS #include "src/misc/Solve.h" #include "src/misc/SparseSolve.h" +#else +#include "src/IterativeLinearSolvers/SolveWithGuess.h" +#endif #include "src/IterativeLinearSolvers/IterativeSolverBase.h" #include "src/IterativeLinearSolvers/BasicPreconditioners.h" diff --git a/Eigen/SparseCholesky b/Eigen/SparseCholesky index 9f5056aa1..2c4f66105 100644 --- a/Eigen/SparseCholesky +++ b/Eigen/SparseCholesky @@ -34,8 +34,11 @@ #error The SparseCholesky module has nothing to offer in MPL2 only mode #endif +#ifndef EIGEN_TEST_EVALUATORS #include "src/misc/Solve.h" #include "src/misc/SparseSolve.h" +#endif + #include "src/SparseCholesky/SimplicialCholesky.h" #ifndef EIGEN_MPL2_ONLY diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 41cae58b0..b68c8fa8a 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -58,6 +58,7 @@ struct Sparse {}; #include "src/SparseCore/TriangularSolver.h" #include "src/SparseCore/SparsePermutation.h" #include "src/SparseCore/SparseFuzzy.h" +#include "src/SparseCore/SparseSolverBase.h" #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h index c449960de..4416c5308 100644 --- a/Eigen/src/CholmodSupport/CholmodSupport.h +++ b/Eigen/src/CholmodSupport/CholmodSupport.h @@ -157,8 +157,12 @@ enum CholmodMode { * \sa class CholmodSupernodalLLT, class CholmodSimplicialLDLT, class CholmodSimplicialLLT */ template -class CholmodBase : internal::noncopyable +class CholmodBase : public SparseSolverBase { + protected: + typedef SparseSolverBase Base; + using Base::derived; + using Base::m_isInitialized; public: typedef _MatrixType MatrixType; enum { UpLo = _UpLo }; @@ -170,14 +174,14 @@ class CholmodBase : internal::noncopyable public: CholmodBase() - : m_cholmodFactor(0), m_info(Success), m_isInitialized(false) + : m_cholmodFactor(0), m_info(Success) { m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0); cholmod_start(&m_cholmod); } CholmodBase(const MatrixType& matrix) - : m_cholmodFactor(0), m_info(Success), m_isInitialized(false) + : m_cholmodFactor(0), m_info(Success) { m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0); cholmod_start(&m_cholmod); @@ -194,9 +198,6 @@ class CholmodBase : internal::noncopyable inline Index cols() const { return m_cholmodFactor->n; } inline Index rows() const { return m_cholmodFactor->n; } - Derived& derived() { return *static_cast(this); } - const Derived& derived() const { return *static_cast(this); } - /** \brief Reports whether previous computation was successful. * * \returns \c Success if computation was succesful, @@ -216,6 +217,7 @@ class CholmodBase : internal::noncopyable return derived(); } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -235,14 +237,15 @@ class CholmodBase : internal::noncopyable * \sa compute() */ template - inline const internal::sparse_solve_retval + inline const internal::sparse_solve_retval solve(const SparseMatrixBase& b) const { eigen_assert(m_isInitialized && "LLT is not initialized."); eigen_assert(rows()==b.rows() && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); + return internal::sparse_solve_retval(derived(), b.derived()); } +#endif // EIGEN_TEST_EVALUATORS /** Performs a symbolic decomposition on the sparsity pattern of \a matrix. * @@ -290,7 +293,7 @@ class CholmodBase : internal::noncopyable #ifndef EIGEN_PARSED_BY_DOXYGEN /** \internal */ template - void _solve(const MatrixBase &b, MatrixBase &dest) const + void _solve_impl(const MatrixBase &b, MatrixBase &dest) const { eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()"); const Index size = m_cholmodFactor->n; @@ -312,7 +315,7 @@ class CholmodBase : internal::noncopyable /** \internal */ template - void _solve(const SparseMatrix &b, SparseMatrix &dest) const + void _solve_impl(const SparseMatrix &b, SparseMatrix &dest) const { eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()"); const Index size = m_cholmodFactor->n; @@ -357,7 +360,6 @@ class CholmodBase : internal::noncopyable cholmod_factor* m_cholmodFactor; RealScalar m_shiftOffset[2]; mutable ComputationInfo m_info; - bool m_isInitialized; int m_factorizationIsOk; int m_analysisIsOk; }; @@ -572,6 +574,7 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom } }; +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -583,7 +586,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; @@ -596,11 +599,12 @@ struct sparse_solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal +#endif } // end namespace Eigen diff --git a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h index 1f3c060d0..92af28cc8 100644 --- a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h +++ b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -80,12 +80,14 @@ class DiagonalPreconditioner return factorize(mat); } + /** \internal */ template - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const Rhs& b, Dest& x) const { x = m_invdiag.array() * b.array() ; } +#ifndef EIGEN_TEST_EVALUATORS template inline const internal::solve_retval solve(const MatrixBase& b) const { @@ -94,12 +96,23 @@ class DiagonalPreconditioner && "DiagonalPreconditioner::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } +#else + template inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "DiagonalPreconditioner is not initialized."); + eigen_assert(m_invdiag.size()==b.rows() + && "DiagonalPreconditioner::solve(): invalid number of rows of the right hand side matrix b"); + return Solve(*this, b.derived()); + } +#endif protected: Vector m_invdiag; bool m_isInitialized; }; +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -111,11 +124,12 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } +#endif // EIGEN_TEST_EVALUATORS /** \ingroup IterativeLinearSolvers_Module * \brief A naive preconditioner which approximates any matrix as the identity matrix diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index 27824b9d5..2cad946d3 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // Copyright (C) 2012 Désiré Nuentsa-Wakam // // This Source Code Form is subject to the terms of the Mozilla @@ -181,7 +181,8 @@ public: BiCGSTAB(const MatrixType& A) : Base(A) {} ~BiCGSTAB() {} - + +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A * \a x0 as an initial solution. * @@ -197,10 +198,26 @@ public: return internal::solve_retval_with_guess (*this, b.derived(), x0); } - +#else + /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A + * \a x0 as an initial solution. + * + * \sa compute() + */ + template + inline const SolveWithGuess + solveWithGuess(const MatrixBase& b, const Guess& x0) const + { + eigen_assert(m_isInitialized && "BiCGSTAB is not initialized."); + eigen_assert(Base::rows()==b.rows() + && "BiCGSTAB::solve(): invalid number of rows of the right hand side matrix b"); + return SolveWithGuess(*this, b.derived(), x0); + } +#endif + /** \internal */ template - void _solveWithGuess(const Rhs& b, Dest& x) const + void _solve_with_guess_impl(const Rhs& b, Dest& x) const { bool failed = false; for(int j=0; j - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const MatrixBase& b, Dest& x) const { -// x.setZero(); - x = b; - _solveWithGuess(b,x); + // x.setZero(); + x = b; + _solve_with_guess_impl(b,x); } protected: }; - +#ifndef EIGEN_TEST_EVALUATORS namespace internal { - template +template struct solve_retval, Rhs> : solve_retval_base, Rhs> { @@ -243,11 +261,12 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal +#endif } // end namespace Eigen diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index 3ce517940..000780eff 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -193,6 +193,7 @@ public: ~ConjugateGradient() {} +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A * \a x0 as an initial solution. * @@ -208,10 +209,26 @@ public: return internal::solve_retval_with_guess (*this, b.derived(), x0); } +#else + /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A + * \a x0 as an initial solution. + * + * \sa compute() + */ + template + inline const SolveWithGuess + solveWithGuess(const MatrixBase& b, const Guess& x0) const + { + eigen_assert(m_isInitialized && "ConjugateGradient is not initialized."); + eigen_assert(Base::rows()==b.rows() + && "ConjugateGradient::solve(): invalid number of rows of the right hand side matrix b"); + return SolveWithGuess(*this, b.derived(), x0); + } +#endif /** \internal */ template - void _solveWithGuess(const Rhs& b, Dest& x) const + void _solve_with_guess_impl(const Rhs& b, Dest& x) const { m_iterations = Base::maxIterations(); m_error = Base::m_tolerance; @@ -231,18 +248,19 @@ public: } /** \internal */ + using Base::_solve_impl; template - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const MatrixBase& b, Dest& x) const { x.setOnes(); - _solveWithGuess(b,x); + _solve_with_guess_impl(b.derived(),x); } protected: }; - +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -254,11 +272,12 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal +#endif } // end namespace Eigen diff --git a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h index b55afc136..a39ed7748 100644 --- a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h +++ b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h @@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2012 Désiré Nuentsa-Wakam +// Copyright (C) 2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -158,7 +159,11 @@ class IncompleteLUT : internal::noncopyable void setFillfactor(int fillfactor); template +#ifndef EIGEN_TEST_EVALUATORS void _solve(const Rhs& b, Dest& x) const +#else + void _solve_impl(const Rhs& b, Dest& x) const +#endif { x = m_Pinv * b; x = m_lu.template triangularView().solve(x); @@ -166,14 +171,25 @@ class IncompleteLUT : internal::noncopyable x = m_P * x; } +#ifndef EIGEN_TEST_EVALUATORS template inline const internal::solve_retval - solve(const MatrixBase& b) const + solve(const MatrixBase& b) const { eigen_assert(m_isInitialized && "IncompleteLUT is not initialized."); eigen_assert(cols()==b.rows() && "IncompleteLUT::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } +#else + template inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "IncompleteLUT is not initialized."); + eigen_assert(cols()==b.rows() + && "IncompleteLUT::solve(): invalid number of rows of the right hand side matrix b"); + return Solve(*this, b.derived()); + } +#endif protected: @@ -445,6 +461,7 @@ void IncompleteLUT::factorize(const _MatrixType& amat) m_info = Success; } +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -461,6 +478,7 @@ struct solve_retval, Rhs> }; } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h index 2036922d6..2fc1a511b 100644 --- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -18,8 +18,12 @@ namespace Eigen { * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner */ template< typename Derived> -class IterativeSolverBase : internal::noncopyable +class IterativeSolverBase : public SparseSolverBase { +protected: + typedef SparseSolverBase Base; + using Base::m_isInitialized; + public: typedef typename internal::traits::MatrixType MatrixType; typedef typename internal::traits::Preconditioner Preconditioner; @@ -29,8 +33,7 @@ public: public: - Derived& derived() { return *static_cast(this); } - const Derived& derived() const { return *static_cast(this); } + using Base::derived; /** Default constructor. */ IterativeSolverBase() @@ -159,6 +162,7 @@ public: return m_error; } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -171,7 +175,9 @@ public: && "IterativeSolverBase::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(derived(), b.derived()); } +#endif +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -185,6 +191,7 @@ public: && "IterativeSolverBase::solve(): invalid number of rows of the right hand side matrix b"); return internal::sparse_solve_retval(*this, b.derived()); } +#endif // EIGEN_TEST_EVALUATORS /** \returns Success if the iterations converged, and NoConvergence otherwise. */ ComputationInfo info() const @@ -193,6 +200,7 @@ public: return m_info; } +#ifndef EIGEN_TEST_EVALUATORS /** \internal */ template void _solve_sparse(const Rhs& b, SparseMatrix &dest) const @@ -210,6 +218,25 @@ public: dest.col(k) = tx.sparseView(0); } } +#else + /** \internal */ + template + void _solve_impl(const Rhs& b, SparseMatrix &dest) const + { + eigen_assert(rows()==b.rows()); + + int rhsCols = b.cols(); + int size = b.rows(); + Eigen::Matrix tb(size); + Eigen::Matrix tx(size); + for(int k=0; k @@ -248,6 +276,7 @@ struct sparse_solve_retval, Rhs> }; } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h new file mode 100644 index 000000000..46dd5babe --- /dev/null +++ b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h @@ -0,0 +1,114 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SOLVEWITHGUESS_H +#define EIGEN_SOLVEWITHGUESS_H + +namespace Eigen { + +template class SolveWithGuess; + +/** \class SolveWithGuess + * \ingroup IterativeLinearSolvers_Module + * + * \brief Pseudo expression representing a solving operation + * + * \tparam Decomposition the type of the matrix or decomposion object + * \tparam Rhstype the type of the right-hand side + * + * This class represents an expression of A.solve(B) + * and most of the time this is the only way it is used. + * + */ +namespace internal { + + +template +struct traits > + : traits > +{}; + +} + + +template +class SolveWithGuess : public internal::generic_xpr_base, MatrixXpr, typename internal::traits::StorageKind>::type +{ +public: + typedef typename RhsType::Index Index; + typedef typename internal::traits::PlainObject PlainObject; + typedef typename internal::generic_xpr_base, MatrixXpr, typename internal::traits::StorageKind>::type Base; + + SolveWithGuess(const Decomposition &dec, const RhsType &rhs, const GuessType &guess) + : m_dec(dec), m_rhs(rhs), m_guess(guess) + {} + + EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); } + + EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; } + EIGEN_DEVICE_FUNC const RhsType& rhs() const { return m_rhs; } + EIGEN_DEVICE_FUNC const GuessType& guess() const { return m_guess; } + +protected: + const Decomposition &m_dec; + const RhsType &m_rhs; + const GuessType &m_guess; + + typedef typename internal::traits::Scalar Scalar; + +private: + Scalar coeff(Index row, Index col) const; + Scalar coeff(Index i) const; +}; + +namespace internal { + +// Evaluator of SolveWithGuess -> eval into a temporary +template +struct evaluator > + : public evaluator::PlainObject>::type +{ + typedef SolveWithGuess SolveType; + typedef typename SolveType::PlainObject PlainObject; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const SolveType& solve) + : m_result(solve.rows(), solve.cols()) + { + ::new (static_cast(this)) Base(m_result); + solve.dec()._solve_with_guess_impl(solve.rhs(), m_result, solve().guess()); + } + +protected: + PlainObject m_result; +}; + +// Specialization for "dst = dec.solve(rhs)" +// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere +template +struct Assignment, internal::assign_op, Dense2Dense, Scalar> +{ + typedef Solve SrcXprType; + static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) + { + // FIXME shall we resize dst here? + dst = src.guess(); + src.dec()._solve_with_guess_impl(src.rhs(), dst/*, src.guess()*/); + } +}; + +} // end namepsace internal + +} // end namespace Eigen + +#endif // EIGEN_SOLVEWITHGUESS_H diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h index 8a546dc2f..0dc5c6a6f 100644 --- a/Eigen/src/PaStiXSupport/PaStiXSupport.h +++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h @@ -125,9 +125,15 @@ namespace internal // This is the base class to interface with PaStiX functions. // Users should not used this class directly. template -class PastixBase : internal::noncopyable +class PastixBase : public SparseSolverBase { + protected: + typedef SparseSolverBase Base; + using Base::derived; + using Base::m_isInitialized; public: + using Base::_solve_impl; + typedef typename internal::pastix_traits::MatrixType _MatrixType; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; @@ -138,7 +144,7 @@ class PastixBase : internal::noncopyable public: - PastixBase() : m_initisOk(false), m_analysisIsOk(false), m_factorizationIsOk(false), m_isInitialized(false), m_pastixdata(0), m_size(0) + PastixBase() : m_initisOk(false), m_analysisIsOk(false), m_factorizationIsOk(false), m_pastixdata(0), m_size(0) { init(); } @@ -148,6 +154,7 @@ class PastixBase : internal::noncopyable clean(); } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -161,19 +168,11 @@ class PastixBase : internal::noncopyable && "PastixBase::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } +#endif template - bool _solve (const MatrixBase &b, MatrixBase &x) const; + bool _solve_impl(const MatrixBase &b, MatrixBase &x) const; - Derived& derived() - { - return *static_cast(this); - } - const Derived& derived() const - { - return *static_cast(this); - } - /** Returns a reference to the integer vector IPARM of PaStiX parameters * to modify the default parameters. * The statistics related to the different phases of factorization and solve are saved here as well @@ -228,6 +227,7 @@ class PastixBase : internal::noncopyable return m_info; } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -241,6 +241,7 @@ class PastixBase : internal::noncopyable && "PastixBase::solve(): invalid number of rows of the right hand side matrix b"); return internal::sparse_solve_retval(*this, b.derived()); } +#endif // EIGEN_TEST_EVALUATORS protected: @@ -268,7 +269,6 @@ class PastixBase : internal::noncopyable int m_initisOk; int m_analysisIsOk; int m_factorizationIsOk; - bool m_isInitialized; mutable ComputationInfo m_info; mutable pastix_data_t *m_pastixdata; // Data structure for pastix mutable int m_comm; // The MPI communicator identifier @@ -393,7 +393,7 @@ void PastixBase::factorize(ColSpMatrix& mat) /* Solve the system */ template template -bool PastixBase::_solve (const MatrixBase &b, MatrixBase &x) const +bool PastixBase::_solve_impl(const MatrixBase &b, MatrixBase &x) const { eigen_assert(m_isInitialized && "The matrix should be factorized first"); EIGEN_STATIC_ASSERT((Dest::Flags&RowMajorBit)==0, @@ -694,6 +694,7 @@ class PastixLDLT : public PastixBase< PastixLDLT<_MatrixType, _UpLo> > } }; +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -705,7 +706,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; @@ -723,6 +724,7 @@ struct sparse_solve_retval, Rhs> }; } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/PardisoSupport/PardisoSupport.h b/Eigen/src/PardisoSupport/PardisoSupport.h index b6571069e..e3b1c5bc0 100644 --- a/Eigen/src/PardisoSupport/PardisoSupport.h +++ b/Eigen/src/PardisoSupport/PardisoSupport.h @@ -96,10 +96,17 @@ namespace internal } template -class PardisoImpl : internal::noncopyable +class PardisoImpl : public SparseSolveBase { + protected: + typedef SparseSolveBase Base; + using Base::derived; + using Base::m_isInitialized; + typedef internal::pardiso_traits Traits; public: + using base::_solve_impl; + typedef typename Traits::MatrixType MatrixType; typedef typename Traits::Scalar Scalar; typedef typename Traits::RealScalar RealScalar; @@ -118,7 +125,7 @@ class PardisoImpl : internal::noncopyable eigen_assert((sizeof(Index) >= sizeof(_INTEGER_t) && sizeof(Index) <= 8) && "Non-supported index type"); m_iparm.setZero(); m_msglvl = 0; // No output - m_initialized = false; + m_isInitialized = false; } ~PardisoImpl() @@ -136,7 +143,7 @@ class PardisoImpl : internal::noncopyable */ ComputationInfo info() const { - eigen_assert(m_initialized && "Decomposition is not initialized."); + eigen_assert(m_isInitialized && "Decomposition is not initialized."); return m_info; } @@ -166,6 +173,7 @@ class PardisoImpl : internal::noncopyable Derived& compute(const MatrixType& matrix); +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -174,7 +182,7 @@ class PardisoImpl : internal::noncopyable inline const internal::solve_retval solve(const MatrixBase& b) const { - eigen_assert(m_initialized && "Pardiso solver is not initialized."); + eigen_assert(m_isInitialized && "Pardiso solver is not initialized."); eigen_assert(rows()==b.rows() && "PardisoImpl::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); @@ -188,28 +196,20 @@ class PardisoImpl : internal::noncopyable inline const internal::sparse_solve_retval solve(const SparseMatrixBase& b) const { - eigen_assert(m_initialized && "Pardiso solver is not initialized."); + eigen_assert(m_isInitialized && "Pardiso solver is not initialized."); eigen_assert(rows()==b.rows() && "PardisoImpl::solve(): invalid number of rows of the right hand side matrix b"); return internal::sparse_solve_retval(*this, b.derived()); } - - Derived& derived() - { - return *static_cast(this); - } - const Derived& derived() const - { - return *static_cast(this); - } +#endif template - bool _solve(const MatrixBase &b, MatrixBase& x) const; + bool _solve_impl(const MatrixBase &b, MatrixBase& x) const; protected: void pardisoRelease() { - if(m_initialized) // Factorization ran at least once + if(m_isInitialized) // Factorization ran at least once { internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, -1, m_size, 0, 0, 0, m_perm.data(), 0, m_iparm.data(), m_msglvl, 0, 0); @@ -270,7 +270,7 @@ class PardisoImpl : internal::noncopyable mutable SparseMatrixType m_matrix; ComputationInfo m_info; - bool m_initialized, m_analysisIsOk, m_factorizationIsOk; + bool m_analysisIsOk, m_factorizationIsOk; Index m_type, m_msglvl; mutable void *m_pt[64]; mutable ParameterType m_iparm; @@ -298,7 +298,7 @@ Derived& PardisoImpl::compute(const MatrixType& a) manageErrorCode(error); m_analysisIsOk = true; m_factorizationIsOk = true; - m_initialized = true; + m_isInitialized = true; return derived(); } @@ -321,7 +321,7 @@ Derived& PardisoImpl::analyzePattern(const MatrixType& a) manageErrorCode(error); m_analysisIsOk = true; m_factorizationIsOk = false; - m_initialized = true; + m_isInitialized = true; return derived(); } @@ -345,7 +345,7 @@ Derived& PardisoImpl::factorize(const MatrixType& a) template template -bool PardisoImpl::_solve(const MatrixBase &b, MatrixBase& x) const +bool PardisoImpl::_solve_impl(const MatrixBase &b, MatrixBase& x) const { if(m_iparm[0] == 0) // Factorization was not computed return false; @@ -546,6 +546,7 @@ class PardisoLDLT : public PardisoImpl< PardisoLDLT > } }; +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -557,7 +558,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; @@ -575,6 +576,7 @@ struct sparse_solve_retval, Rhs> }; } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index 8808e6c0d..136e80ce9 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -350,7 +350,8 @@ void HouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) c dst.bottomRows(cols()-rank).setZero(); } #endif - + +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -366,6 +367,7 @@ struct solve_retval, Rhs> }; } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS /** Performs the QR factorization of the given matrix \a matrix. The result of * the factorization is stored into \c *this, and a reference to \c *this diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h index a2cc2a9e2..483aaef07 100644 --- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h @@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2012 Desire Nuentsa +// Copyright (C) 2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -54,8 +55,11 @@ namespace Eigen { * */ template -class SPQR +class SPQR : public SparseSolverBase > { + protected: + typedef SparseSolverBase > Base; + using Base::m_isInitialized; public: typedef typename _MatrixType::Scalar Scalar; typedef typename _MatrixType::RealScalar RealScalar; @@ -64,19 +68,13 @@ class SPQR typedef PermutationMatrix PermutationType; public: SPQR() - : m_isInitialized(false), - m_ordering(SPQR_ORDERING_DEFAULT), - m_allow_tol(SPQR_DEFAULT_TOL), - m_tolerance (NumTraits::epsilon()) + : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) { cholmod_l_start(&m_cc); } SPQR(const _MatrixType& matrix) - : m_isInitialized(false), - m_ordering(SPQR_ORDERING_DEFAULT), - m_allow_tol(SPQR_DEFAULT_TOL), - m_tolerance (NumTraits::epsilon()) + : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) { cholmod_l_start(&m_cc); compute(matrix); @@ -127,10 +125,11 @@ class SPQR */ inline Index cols() const { return m_cR->ncol; } - /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. - * - * \sa compute() - */ +#ifndef EIGEN_TEST_EVALUATORS + /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. + * + * \sa compute() + */ template inline const internal::solve_retval solve(const MatrixBase& B) const { @@ -139,9 +138,10 @@ class SPQR && "SPQR::solve(): invalid number of rows of the right hand side matrix B"); return internal::solve_retval(*this, B.derived()); } +#endif // EIGEN_TEST_EVALUATORS template - void _solve(const MatrixBase &b, MatrixBase &dest) const + void _solve_impl(const MatrixBase &b, MatrixBase &dest) const { eigen_assert(m_isInitialized && " The QR factorization should be computed first, call compute()"); eigen_assert(b.cols()==1 && "This method is for vectors only"); @@ -214,7 +214,6 @@ class SPQR return m_info; } protected: - bool m_isInitialized; bool m_analysisIsOk; bool m_factorizationIsOk; mutable bool m_isRUpToDate; @@ -293,6 +292,7 @@ struct SPQRMatrixQTransposeReturnType{ const SPQRType& m_spqr; }; +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -304,11 +304,12 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal +#endif }// End namespace Eigen #endif diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index 1abd31304..ac8cd29b0 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -33,8 +33,11 @@ enum SimplicialCholeskyMode { * */ template -class SimplicialCholeskyBase : internal::noncopyable +class SimplicialCholeskyBase : public SparseSolverBase { + typedef SparseSolverBase Base; + using Base::m_isInitialized; + public: typedef typename internal::traits::MatrixType MatrixType; typedef typename internal::traits::OrderingType OrderingType; @@ -46,14 +49,16 @@ class SimplicialCholeskyBase : internal::noncopyable typedef Matrix VectorType; public: + + using Base::derived; /** Default constructor */ SimplicialCholeskyBase() - : m_info(Success), m_isInitialized(false), m_shiftOffset(0), m_shiftScale(1) + : m_info(Success), m_shiftOffset(0), m_shiftScale(1) {} SimplicialCholeskyBase(const MatrixType& matrix) - : m_info(Success), m_isInitialized(false), m_shiftOffset(0), m_shiftScale(1) + : m_info(Success), m_shiftOffset(0), m_shiftScale(1) { derived().compute(matrix); } @@ -79,6 +84,7 @@ class SimplicialCholeskyBase : internal::noncopyable return m_info; } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -106,6 +112,7 @@ class SimplicialCholeskyBase : internal::noncopyable && "SimplicialCholesky::solve(): invalid number of rows of the right hand side matrix b"); return internal::sparse_solve_retval(*this, b.derived()); } +#endif // EIGEN_TEST_EVALUATORS /** \returns the permutation P * \sa permutationPinv() */ @@ -150,7 +157,11 @@ class SimplicialCholeskyBase : internal::noncopyable /** \internal */ template +#ifndef EIGEN_TEST_EVALUATORS void _solve(const MatrixBase &b, MatrixBase &dest) const +#else + void _solve_impl(const MatrixBase &b, MatrixBase &dest) const +#endif { eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()"); eigen_assert(m_matrix.rows()==b.rows()); @@ -176,6 +187,14 @@ class SimplicialCholeskyBase : internal::noncopyable dest = m_Pinv * dest; } +#ifdef EIGEN_TEST_EVALUATORS + template + void _solve_impl(const SparseMatrixBase &b, SparseMatrixBase &dest) const + { + internal::solve_sparse_through_dense_panels(derived(), b, dest); + } +#endif + #endif // EIGEN_PARSED_BY_DOXYGEN protected: @@ -226,7 +245,6 @@ class SimplicialCholeskyBase : internal::noncopyable }; mutable ComputationInfo m_info; - bool m_isInitialized; bool m_factorizationIsOk; bool m_analysisIsOk; @@ -560,7 +578,11 @@ public: /** \internal */ template +#ifndef EIGEN_TEST_EVALUATORS void _solve(const MatrixBase &b, MatrixBase &dest) const +#else + void _solve_impl(const MatrixBase &b, MatrixBase &dest) const +#endif { eigen_assert(Base::m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()"); eigen_assert(Base::m_matrix.rows()==b.rows()); @@ -596,6 +618,15 @@ public: dest = Base::m_Pinv * dest; } +#ifdef EIGEN_TEST_EVALUATORS + /** \internal */ + template + void _solve_impl(const SparseMatrixBase &b, SparseMatrixBase &dest) const + { + internal::solve_sparse_through_dense_panels(*this, b, dest); + } +#endif + Scalar determinant() const { if(m_LDLT) @@ -636,6 +667,7 @@ void SimplicialCholeskyBase::ordering(const MatrixType& a, CholMatrixTy ap.template selfadjointView() = a.template selfadjointView().twistedBy(m_P); } +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -665,6 +697,7 @@ struct sparse_solve_retval, Rhs> }; } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/SparseCore/SparseSolverBase.h b/Eigen/src/SparseCore/SparseSolverBase.h new file mode 100644 index 000000000..9a6ed1292 --- /dev/null +++ b/Eigen/src/SparseCore/SparseSolverBase.h @@ -0,0 +1,113 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SPARSESOLVERBASE_H +#define EIGEN_SPARSESOLVERBASE_H + +namespace Eigen { + +namespace internal { + + /** \internal + * Helper functions to solve with a sparse right-hand-side and result. + * The rhs is decomposed into small vertical panels which are solved through dense temporaries. + */ +template +void solve_sparse_through_dense_panels(const Decomposition &dec, const Rhs& rhs, Dest &dest) +{ + EIGEN_STATIC_ASSERT((Dest::Flags&RowMajorBit)==0,THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES); + typedef typename Dest::Scalar DestScalar; + // we process the sparse rhs per block of NbColsAtOnce columns temporarily stored into a dense matrix. + static const int NbColsAtOnce = 4; + int rhsCols = rhs.cols(); + int size = rhs.rows(); + // the temporary matrices do not need more columns than NbColsAtOnce: + int tmpCols = (std::min)(rhsCols, NbColsAtOnce); + Eigen::Matrix tmp(size,tmpCols); + Eigen::Matrix tmpX(size,tmpCols); + for(int k=0; k(rhsCols-k, NbColsAtOnce); + tmp.leftCols(actualCols) = rhs.middleCols(k,actualCols); + tmpX.leftCols(actualCols) = dec.solve(tmp.leftCols(actualCols)); + dest.middleCols(k,actualCols) = tmpX.leftCols(actualCols).sparseView(); + } +} + +} // end namespace internal + +/** \class SparseSolverBase + * \ingroup SparseCore_Module + * \brief A base class for sparse solvers + * + * \tparam Derived the actual type of the solver. + * + */ +template +class SparseSolverBase : internal::noncopyable +{ + public: + + /** Default constructor */ + SparseSolverBase() + : m_isInitialized(false) + {} + + ~SparseSolverBase() + {} + + Derived& derived() { return *static_cast(this); } + const Derived& derived() const { return *static_cast(this); } + +#ifdef EIGEN_TEST_EVALUATORS + /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A. + * + * \sa compute() + */ + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "Solver is not initialized."); + eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b"); + return Solve(derived(), b.derived()); + } + + /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A. + * + * \sa compute() + */ + template + inline const Solve + solve(const SparseMatrixBase& b) const + { + eigen_assert(m_isInitialized && "Solver is not initialized."); + eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b"); + return Solve(derived(), b.derived()); + } +#endif + + +#ifndef EIGEN_PARSED_BY_DOXYGEN + /** \internal default implementation of solving with a sparse rhs */ + template + void _solve_impl(const SparseMatrixBase &b, SparseMatrixBase &dest) const + { + internal::solve_sparse_through_dense_panels(derived(), b.derived(), dest.derived()); + } +#endif // EIGEN_PARSED_BY_DOXYGEN + + protected: + + mutable bool m_isInitialized; +}; + +} // end namespace Eigen + +#endif // EIGEN_SPARSESOLVERBASE_H diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 7a9aeec2d..eb61fe3d9 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2012 Désiré Nuentsa-Wakam -// Copyright (C) 2012 Gael Guennebaud +// Copyright (C) 2012-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -70,9 +70,14 @@ template struct SparseLUMatrixURetu * \sa \ref OrderingMethods_Module */ template -class SparseLU : public internal::SparseLUImpl +class SparseLU : public SparseSolverBase >, public internal::SparseLUImpl { + protected: + typedef SparseSolverBase > APIBase; + using APIBase::m_isInitialized; public: + using APIBase::_solve_impl; + typedef _MatrixType MatrixType; typedef _OrderingType OrderingType; typedef typename MatrixType::Scalar Scalar; @@ -86,11 +91,11 @@ class SparseLU : public internal::SparseLUImpl Base; public: - SparseLU():m_isInitialized(true),m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1) + SparseLU():m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1) { initperfvalues(); } - SparseLU(const MatrixType& matrix):m_isInitialized(true),m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1) + SparseLU(const MatrixType& matrix):m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1) { initperfvalues(); compute(matrix); @@ -168,6 +173,7 @@ class SparseLU : public internal::SparseLUImplsolve(B) must be colmun-major. @@ -195,6 +201,20 @@ class SparseLU : public internal::SparseLUImpl(*this, B.derived()); } +#else + +#ifdef EIGEN_PARSED_BY_DOXYGEN + /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. + * + * \warning the destination matrix X in X = this->solve(B) must be colmun-major. + * + * \sa compute() + */ + template + inline const Solve solve(const MatrixBase& B) const; +#endif // EIGEN_PARSED_BY_DOXYGEN + +#endif // EIGEN_TEST_EVALUATORS /** \brief Reports whether previous computation was successful. * @@ -219,7 +239,7 @@ class SparseLU : public internal::SparseLUImpl - bool _solve(const MatrixBase &B, MatrixBase &X_base) const + bool _solve_impl(const MatrixBase &B, MatrixBase &X_base) const { Dest& X(X_base.derived()); eigen_assert(m_factorizationIsOk && "The matrix should be factorized first"); @@ -332,7 +352,6 @@ class SparseLU : public internal::SparseLUImpl @@ -739,7 +759,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; @@ -756,6 +776,7 @@ struct sparse_solve_retval, Rhs> } }; } // end namespace internal +#endif } // End namespace Eigen diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 6be569533..8e946b045 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -62,9 +62,13 @@ namespace internal { * */ template -class SparseQR +class SparseQR : public SparseSolverBase > { + protected: + typedef SparseSolverBase > Base; + using Base::m_isInitialized; public: + using Base::_solve_impl; typedef _MatrixType MatrixType; typedef _OrderingType OrderingType; typedef typename MatrixType::Scalar Scalar; @@ -75,7 +79,7 @@ class SparseQR typedef Matrix ScalarVector; typedef PermutationMatrix PermutationType; public: - SparseQR () : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) + SparseQR () : m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) { } /** Construct a QR factorization of the matrix \a mat. @@ -84,7 +88,7 @@ class SparseQR * * \sa compute() */ - SparseQR(const MatrixType& mat) : m_isInitialized(false), m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) + SparseQR(const MatrixType& mat) : m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) { compute(mat); } @@ -162,7 +166,7 @@ class SparseQR /** \internal */ template - bool _solve(const MatrixBase &B, MatrixBase &dest) const + bool _solve_impl(const MatrixBase &B, MatrixBase &dest) const { eigen_assert(m_isInitialized && "The factorization should be called first, use compute()"); eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix"); @@ -186,7 +190,6 @@ class SparseQR m_info = Success; return true; } - /** Sets the threshold that is used to determine linearly dependent columns during the factorization. * @@ -199,6 +202,7 @@ class SparseQR m_threshold = threshold; } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. * * \sa compute() @@ -217,6 +221,26 @@ class SparseQR eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix"); return internal::sparse_solve_retval(*this, B.derived()); } +#else + /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. + * + * \sa compute() + */ + template + inline const Solve solve(const MatrixBase& B) const + { + eigen_assert(m_isInitialized && "The factorization should be called first, use compute()"); + eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix"); + return Solve(*this, B.derived()); + } + template + inline const Solve solve(const SparseMatrixBase& B) const + { + eigen_assert(m_isInitialized && "The factorization should be called first, use compute()"); + eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix"); + return Solve(*this, B.derived()); + } +#endif // EIGEN_TEST_EVALUATORS /** \brief Reports whether previous computation was successful. * @@ -244,7 +268,6 @@ class SparseQR protected: - bool m_isInitialized; bool m_analysisIsok; bool m_factorizationIsok; mutable ComputationInfo m_info; @@ -554,6 +577,7 @@ void SparseQR::factorize(const MatrixType& mat) m_info = Success; } +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -565,7 +589,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; template @@ -581,6 +605,7 @@ struct sparse_solve_retval, Rhs> } }; } // end namespace internal +#endif // EIGEN_TEST_EVALUATORS template struct SparseQR_QProduct : ReturnByValue > diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h index bcb355760..fcecd4fcf 100644 --- a/Eigen/src/SuperLUSupport/SuperLUSupport.h +++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2011 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -288,8 +288,12 @@ MappedSparseMatrix map_superlu(SluMatrix& sluMat) * \brief The base class for the direct and incomplete LU factorization of SuperLU */ template -class SuperLUBase : internal::noncopyable +class SuperLUBase : public SparseSolverBase { + protected: + typedef SparseSolverBase Base; + using Base::derived; + using Base::m_isInitialized; public: typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; @@ -309,9 +313,6 @@ class SuperLUBase : internal::noncopyable clearFactors(); } - Derived& derived() { return *static_cast(this); } - const Derived& derived() const { return *static_cast(this); } - inline Index rows() const { return m_matrix.rows(); } inline Index cols() const { return m_matrix.cols(); } @@ -336,6 +337,7 @@ class SuperLUBase : internal::noncopyable derived().factorize(matrix); } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -361,7 +363,8 @@ class SuperLUBase : internal::noncopyable && "SuperLU::solve(): invalid number of rows of the right hand side matrix b"); return internal::sparse_solve_retval(*this, b.derived()); } - +#endif // EIGEN_TEST_EVALUATORS + /** Performs a symbolic decomposition on the sparcity of \a matrix. * * This function is particularly useful when solving for several problems having the same structure. @@ -453,7 +456,6 @@ class SuperLUBase : internal::noncopyable mutable char m_sluEqued; mutable ComputationInfo m_info; - bool m_isInitialized; int m_factorizationIsOk; int m_analysisIsOk; mutable bool m_extractedDataAreDirty; @@ -491,6 +493,7 @@ class SuperLU : public SuperLUBase<_MatrixType,SuperLU<_MatrixType> > typedef TriangularView UMatrixType; public: + using Base::_solve_impl; SuperLU() : Base() { init(); } @@ -528,7 +531,7 @@ class SuperLU : public SuperLUBase<_MatrixType,SuperLU<_MatrixType> > #ifndef EIGEN_PARSED_BY_DOXYGEN /** \internal */ template - void _solve(const MatrixBase &b, MatrixBase &dest) const; + void _solve_impl(const MatrixBase &b, MatrixBase &dest) const; #endif // EIGEN_PARSED_BY_DOXYGEN inline const LMatrixType& matrixL() const @@ -637,7 +640,7 @@ void SuperLU::factorize(const MatrixType& a) template template -void SuperLU::_solve(const MatrixBase &b, MatrixBase& x) const +void SuperLU::_solve_impl(const MatrixBase &b, MatrixBase& x) const { eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or analyzePattern()/factorize()"); @@ -828,6 +831,7 @@ class SuperILU : public SuperLUBase<_MatrixType,SuperILU<_MatrixType> > typedef typename Base::Index Index; public: + using Base::_solve_impl; SuperILU() : Base() { init(); } @@ -863,7 +867,7 @@ class SuperILU : public SuperLUBase<_MatrixType,SuperILU<_MatrixType> > #ifndef EIGEN_PARSED_BY_DOXYGEN /** \internal */ template - void _solve(const MatrixBase &b, MatrixBase &dest) const; + void _solve_impl(const MatrixBase &b, MatrixBase &dest) const; #endif // EIGEN_PARSED_BY_DOXYGEN protected: @@ -948,7 +952,7 @@ void SuperILU::factorize(const MatrixType& a) template template -void SuperILU::_solve(const MatrixBase &b, MatrixBase& x) const +void SuperILU::_solve_impl(const MatrixBase &b, MatrixBase& x) const { eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or analyzePattern()/factorize()"); @@ -991,6 +995,7 @@ void SuperILU::_solve(const MatrixBase &b, MatrixBase& x) } #endif +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -1002,7 +1007,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec().derived()._solve(rhs(),dst); + dec().derived()._solve_impl(rhs(),dst); } }; @@ -1020,7 +1025,7 @@ struct sparse_solve_retval, Rhs> }; } // end namespace internal - +#endif } // end namespace Eigen #endif // EIGEN_SUPERLUSUPPORT_H diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h index 3a48cecf7..845c8076a 100644 --- a/Eigen/src/UmfPackSupport/UmfPackSupport.h +++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h @@ -121,9 +121,13 @@ inline int umfpack_get_determinant(std::complex *Mx, double *Ex, void *N * \sa \ref TutorialSparseDirectSolvers */ template -class UmfPackLU : internal::noncopyable +class UmfPackLU : public SparseSolverBase > { + protected: + typedef SparseSolverBase > Base; + using Base::m_isInitialized; public: + using Base::_solve_impl; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; @@ -197,7 +201,8 @@ class UmfPackLU : internal::noncopyable analyzePattern(matrix); factorize(matrix); } - + +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -223,6 +228,7 @@ class UmfPackLU : internal::noncopyable && "UmfPackLU::solve(): invalid number of rows of the right hand side matrix b"); return internal::sparse_solve_retval(*this, b.derived()); } +#endif /** Performs a symbolic decomposition on the sparcity of \a matrix. * @@ -274,7 +280,7 @@ class UmfPackLU : internal::noncopyable #ifndef EIGEN_PARSED_BY_DOXYGEN /** \internal */ template - bool _solve(const MatrixBase &b, MatrixBase &x) const; + bool _solve_impl(const MatrixBase &b, MatrixBase &x) const; #endif Scalar determinant() const; @@ -328,7 +334,6 @@ class UmfPackLU : internal::noncopyable void* m_symbolic; mutable ComputationInfo m_info; - bool m_isInitialized; int m_factorizationIsOk; int m_analysisIsOk; mutable bool m_extractedDataAreDirty; @@ -376,7 +381,7 @@ typename UmfPackLU::Scalar UmfPackLU::determinant() cons template template -bool UmfPackLU::_solve(const MatrixBase &b, MatrixBase &x) const +bool UmfPackLU::_solve_impl(const MatrixBase &b, MatrixBase &x) const { const int rhsCols = b.cols(); eigen_assert((BDerived::Flags&RowMajorBit)==0 && "UmfPackLU backend does not support non col-major rhs yet"); @@ -396,7 +401,7 @@ bool UmfPackLU::_solve(const MatrixBase &b, MatrixBase @@ -408,7 +413,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; @@ -426,7 +431,7 @@ struct sparse_solve_retval, Rhs> }; } // end namespace internal - +#endif } // end namespace Eigen #endif // EIGEN_UMFPACKSUPPORT_H From fbb53b6cbb7f1a7cce2166c9df981b76fdd37f87 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 16:53:52 +0200 Subject: [PATCH 0749/1103] Fix sparse matrix times sparse vector. --- Eigen/src/SparseCore/ConservativeSparseSparseProduct.h | 6 ++++-- test/sparse_vector.cpp | 1 + 2 files changed, 5 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 815fdb6d8..535713ec5 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -28,6 +28,8 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r ei_declare_aligned_stack_constructed_variable(bool, mask, rows, 0); ei_declare_aligned_stack_constructed_variable(Scalar, values, rows, 0); ei_declare_aligned_stack_constructed_variable(Index, indices, rows, 0); + + std::memset(mask,0,sizeof(bool)*rows); // estimate the number of non zero entries // given a rhs column containing Y non zeros, we assume that the respective Y columns @@ -155,14 +157,14 @@ struct conservative_sparse_sparse_product_selector(lhs, rhs, resCol, true); - res.swap(resCol); + res = resCol.markAsRValue(); } else { // ressort to transpose to sort the entries internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, false); RowMajorMatrix resRow(resCol); - res = resRow; + res = resRow.markAsRValue(); } } }; diff --git a/test/sparse_vector.cpp b/test/sparse_vector.cpp index 0c9476803..5eea9edfd 100644 --- a/test/sparse_vector.cpp +++ b/test/sparse_vector.cpp @@ -71,6 +71,7 @@ template void sparse_vector(int rows, int cols) VERIFY_IS_APPROX(v1.dot(v2), refV1.dot(refV2)); VERIFY_IS_APPROX(v1.dot(refV2), refV1.dot(refV2)); + VERIFY_IS_APPROX(m1*v2, refM1*refV2); VERIFY_IS_APPROX(v1.dot(m1*v2), refV1.dot(refM1*refV2)); int i = internal::random(0,rows-1); VERIFY_IS_APPROX(v1.dot(m1.col(i)), refV1.dot(refM1.col(i))); From f9580a3473b20b97cf3e939154f85004e232d96e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:14:30 +0200 Subject: [PATCH 0750/1103] Fix Cholmod support without evaluators --- Eigen/src/CholmodSupport/CholmodSupport.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h index 4416c5308..d3db51d0b 100644 --- a/Eigen/src/CholmodSupport/CholmodSupport.h +++ b/Eigen/src/CholmodSupport/CholmodSupport.h @@ -237,13 +237,13 @@ class CholmodBase : public SparseSolverBase * \sa compute() */ template - inline const internal::sparse_solve_retval + inline const internal::sparse_solve_retval solve(const SparseMatrixBase& b) const { eigen_assert(m_isInitialized && "LLT is not initialized."); eigen_assert(rows()==b.rows() && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(derived(), b.derived()); + return internal::sparse_solve_retval(*this, b.derived()); } #endif // EIGEN_TEST_EVALUATORS From 1bf3b3484975ead812c4615f0d54f25d859320c3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:15:08 +0200 Subject: [PATCH 0751/1103] Fix regression in sparse-sparse product --- Eigen/src/SparseCore/ConservativeSparseSparseProduct.h | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 535713ec5..ae8fc75e5 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -151,18 +151,19 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrixAux; typedef typename sparse_eval::type ColMajorMatrix; - ColMajorMatrix resCol(lhs.rows(),rhs.cols()); // FIXME, the following heuristic is probably not very good. if(lhs.rows()>=rhs.cols()) { + ColMajorMatrix resCol(lhs.rows(),rhs.cols()); // perform sorted insertion internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, true); res = resCol.markAsRValue(); } else { + ColMajorMatrixAux resCol(lhs.rows(),rhs.cols()); // ressort to transpose to sort the entries - internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, false); + internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, false); RowMajorMatrix resRow(resCol); res = resRow.markAsRValue(); } From 863b7362bc6cbd6f5c3ba96b4d6935557f9d9486 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:16:32 +0200 Subject: [PATCH 0752/1103] Fix usage of m_isInitialized in SparseLU and Pastix support. --- Eigen/src/PaStiXSupport/PaStiXSupport.h | 1 - Eigen/src/SparseLU/SparseLU.h | 2 ++ 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h index 0dc5c6a6f..95d53c850 100644 --- a/Eigen/src/PaStiXSupport/PaStiXSupport.h +++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h @@ -328,7 +328,6 @@ void PastixBase::compute(ColSpMatrix& mat) factorize(mat); m_iparm(IPARM_MATRIX_VERIFICATION) = API_NO; - m_isInitialized = m_factorizationIsOk; } diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index eb61fe3d9..fb01f99cd 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -482,6 +482,8 @@ void SparseLU::factorize(const MatrixType& matrix) typedef typename IndexVector::Scalar Index; + m_isInitialized = true; + // Apply the column permutation computed in analyzepattern() // m_mat = matrix * m_perm_c.inverse(); From 8a74ce922cbee1c416675e57e48393dd6b399e4e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:19:16 +0200 Subject: [PATCH 0753/1103] Make IncompleteLUT use SparseSolverBase. --- .../IterativeLinearSolvers/IncompleteLUT.h | 26 ++++++------------- 1 file changed, 8 insertions(+), 18 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h index a39ed7748..f337c5fb0 100644 --- a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h +++ b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h @@ -94,8 +94,12 @@ Index QuickSplit(VectorV &row, VectorI &ind, Index ncut) * http://comments.gmane.org/gmane.comp.lib.eigen/3302 */ template -class IncompleteLUT : internal::noncopyable +class IncompleteLUT : public SparseSolverBase > { + protected: + typedef SparseSolverBase > Base; + using Base::m_isInitialized; + public: typedef _Scalar Scalar; typedef typename NumTraits::Real RealScalar; typedef Matrix Vector; @@ -108,13 +112,13 @@ class IncompleteLUT : internal::noncopyable IncompleteLUT() : m_droptol(NumTraits::dummy_precision()), m_fillfactor(10), - m_analysisIsOk(false), m_factorizationIsOk(false), m_isInitialized(false) + m_analysisIsOk(false), m_factorizationIsOk(false) {} template IncompleteLUT(const MatrixType& mat, const RealScalar& droptol=NumTraits::dummy_precision(), int fillfactor = 10) : m_droptol(droptol),m_fillfactor(fillfactor), - m_analysisIsOk(false),m_factorizationIsOk(false),m_isInitialized(false) + m_analysisIsOk(false),m_factorizationIsOk(false) { eigen_assert(fillfactor != 0); compute(mat); @@ -159,11 +163,7 @@ class IncompleteLUT : internal::noncopyable void setFillfactor(int fillfactor); template -#ifndef EIGEN_TEST_EVALUATORS - void _solve(const Rhs& b, Dest& x) const -#else void _solve_impl(const Rhs& b, Dest& x) const -#endif { x = m_Pinv * b; x = m_lu.template triangularView().solve(x); @@ -180,15 +180,6 @@ class IncompleteLUT : internal::noncopyable && "IncompleteLUT::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } -#else - template inline const Solve - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "IncompleteLUT is not initialized."); - eigen_assert(cols()==b.rows() - && "IncompleteLUT::solve(): invalid number of rows of the right hand side matrix b"); - return Solve(*this, b.derived()); - } #endif protected: @@ -208,7 +199,6 @@ protected: int m_fillfactor; bool m_analysisIsOk; bool m_factorizationIsOk; - bool m_isInitialized; ComputationInfo m_info; PermutationMatrix m_P; // Fill-reducing permutation PermutationMatrix m_Pinv; // Inverse permutation @@ -473,7 +463,7 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; From 1c4b69c5fb4ec6a8b71a64f39ea82dacf50a8bfd Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:19:51 +0200 Subject: [PATCH 0754/1103] Factorize solveWithGuess in IterativeSolverBase --- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 15 --------------- .../IterativeLinearSolvers/ConjugateGradient.h | 15 --------------- .../IterativeLinearSolvers/IterativeSolverBase.h | 16 ++++++++++++++++ 3 files changed, 16 insertions(+), 30 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index 2cad946d3..b4d1d2a79 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -198,21 +198,6 @@ public: return internal::solve_retval_with_guess (*this, b.derived(), x0); } -#else - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A - * \a x0 as an initial solution. - * - * \sa compute() - */ - template - inline const SolveWithGuess - solveWithGuess(const MatrixBase& b, const Guess& x0) const - { - eigen_assert(m_isInitialized && "BiCGSTAB is not initialized."); - eigen_assert(Base::rows()==b.rows() - && "BiCGSTAB::solve(): invalid number of rows of the right hand side matrix b"); - return SolveWithGuess(*this, b.derived(), x0); - } #endif /** \internal */ diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index 000780eff..b0273aaaf 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -209,21 +209,6 @@ public: return internal::solve_retval_with_guess (*this, b.derived(), x0); } -#else - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A - * \a x0 as an initial solution. - * - * \sa compute() - */ - template - inline const SolveWithGuess - solveWithGuess(const MatrixBase& b, const Guess& x0) const - { - eigen_assert(m_isInitialized && "ConjugateGradient is not initialized."); - eigen_assert(Base::rows()==b.rows() - && "ConjugateGradient::solve(): invalid number of rows of the right hand side matrix b"); - return SolveWithGuess(*this, b.derived(), x0); - } #endif /** \internal */ diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h index 2fc1a511b..26487dbb2 100644 --- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h @@ -193,6 +193,22 @@ public: } #endif // EIGEN_TEST_EVALUATORS +#ifdef EIGEN_TEST_EVALUATORS + /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A + * and \a x0 as an initial solution. + * + * \sa solve(), compute() + */ + template + inline const SolveWithGuess + solveWithGuess(const MatrixBase& b, const Guess& x0) const + { + eigen_assert(m_isInitialized && "Solver is not initialized."); + eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b"); + return SolveWithGuess(derived(), b.derived(), x0); + } +#endif // EIGEN_TEST_EVALUATORS + /** \returns Success if the iterations converged, and NoConvergence otherwise. */ ComputationInfo info() const { From b3d63b4db21e746ef3ef260caa28773c8d3ae77b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:21:05 +0200 Subject: [PATCH 0755/1103] Add evaluator for DynamicSparseMatrix --- .../src/SparseExtra/DynamicSparseMatrix.h | 32 +++++++++++++++++++ 1 file changed, 32 insertions(+) diff --git a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h index dec16df28..4210df68a 100644 --- a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h +++ b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h @@ -352,6 +352,38 @@ class DynamicSparseMatrix::ReverseInnerIterator : public const Index m_outer; }; +#ifdef EIGEN_ENABLE_EVALUATORS +namespace internal { + +template +struct evaluator > + : evaluator_base > +{ + typedef _Scalar Scalar; + typedef _Index Index; + typedef DynamicSparseMatrix<_Scalar,_Options,_Index> SparseMatrixType; + typedef typename SparseMatrixType::InnerIterator InnerIterator; + typedef typename SparseMatrixType::ReverseInnerIterator ReverseInnerIterator; + + enum { + CoeffReadCost = NumTraits<_Scalar>::ReadCost, + Flags = SparseMatrixType::Flags + }; + + evaluator() : m_matrix(0) {} + evaluator(const SparseMatrixType &mat) : m_matrix(&mat) {} + + operator SparseMatrixType&() { return m_matrix->const_cast_derived(); } + operator const SparseMatrixType&() const { return *m_matrix; } + + Scalar coeff(Index row, Index col) const { return m_matrix->coeff(row,col); } + + const SparseMatrixType *m_matrix; +}; + +} +#endif + } // end namespace Eigen #endif // EIGEN_DYNAMIC_SPARSEMATRIX_H From b051bbd64fcde21a352ff35d23adcd00afaf845d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:21:47 +0200 Subject: [PATCH 0756/1103] Make unsupport sparse solvers use SparseSolverBase --- .../Eigen/src/IterativeSolvers/DGMRES.h | 13 +++++++++---- .../Eigen/src/IterativeSolvers/GMRES.h | 15 +++++++++------ .../src/IterativeSolvers/IncompleteCholesky.h | 15 ++++++++++++--- .../Eigen/src/IterativeSolvers/IncompleteLU.h | 19 +++++++++++++------ .../Eigen/src/IterativeSolvers/MINRES.h | 19 ++++++++++++------- 5 files changed, 55 insertions(+), 26 deletions(-) diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h index 9fcc8a8d9..fe0bfd948 100644 --- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h @@ -108,6 +108,7 @@ class DGMRES : public IterativeSolverBase > using Base::m_isInitialized; using Base::m_tolerance; public: + using Base::_solve_impl; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::Index Index; @@ -138,6 +139,7 @@ class DGMRES : public IterativeSolverBase > ~DGMRES() {} +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A * \a x0 as an initial solution. * @@ -153,10 +155,11 @@ class DGMRES : public IterativeSolverBase > return internal::solve_retval_with_guess (*this, b.derived(), x0); } +#endif /** \internal */ template - void _solveWithGuess(const Rhs& b, Dest& x) const + void _solve_with_guess_impl(const Rhs& b, Dest& x) const { bool failed = false; for(int j=0; j > /** \internal */ template - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const Rhs& b, MatrixBase& x) const { x = b; - _solveWithGuess(b,x); + _solve_with_guess_impl(b,x.derived()); } /** * Get the restart value @@ -522,6 +525,7 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresApplyDeflation(const RhsType &x, return 0; } +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -533,10 +537,11 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal +#endif } // end namespace Eigen #endif diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index 67498705b..fd76a9d2c 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -281,6 +281,7 @@ private: int m_restart; public: + using Base::_solve_impl; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::Index Index; @@ -315,6 +316,7 @@ public: */ void set_restart(const int restart) { m_restart=restart; } +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A * \a x0 as an initial solution. * @@ -330,10 +332,11 @@ public: return internal::solve_retval_with_guess (*this, b.derived(), x0); } +#endif /** \internal */ template - void _solveWithGuess(const Rhs& b, Dest& x) const + void _solve_with_guess_impl(const Rhs& b, Dest& x) const { bool failed = false; for(int j=0; j - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const Rhs& b, MatrixBase &x) const { x = b; if(x.squaredNorm() == 0) return; // Check Zero right hand side - _solveWithGuess(b,x); + _solve_with_guess_impl(b,x.derived()); } protected: }; - +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -376,12 +379,12 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal - +#endif } // end namespace Eigen #endif // EIGEN_GMRES_H diff --git a/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h b/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h index 661c1f2e0..1ee1c89b2 100644 --- a/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h +++ b/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h @@ -27,8 +27,11 @@ namespace Eigen { */ template > -class IncompleteCholesky : internal::noncopyable +class IncompleteCholesky : public SparseSolverBase > { + protected: + typedef SparseSolverBase > Base; + using Base::m_isInitialized; public: typedef SparseMatrix MatrixType; typedef _OrderingType OrderingType; @@ -89,7 +92,7 @@ class IncompleteCholesky : internal::noncopyable } template - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const Rhs& b, Dest& x) const { eigen_assert(m_factorizationIsOk && "factorize() should be called first"); if (m_perm.rows() == b.rows()) @@ -103,6 +106,8 @@ class IncompleteCholesky : internal::noncopyable x = m_perm * x; x = m_scal.asDiagonal() * x; } + +#ifndef EIGEN_TEST_EVALUATORS template inline const internal::solve_retval solve(const MatrixBase& b) const { @@ -112,13 +117,14 @@ class IncompleteCholesky : internal::noncopyable && "IncompleteLLT::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } +#endif + protected: SparseMatrix m_L; // The lower part stored in CSC ScalarType m_scal; // The vector for scaling the matrix Scalar m_shift; //The initial shift parameter bool m_analysisIsOk; bool m_factorizationIsOk; - bool m_isInitialized; ComputationInfo m_info; PermutationType m_perm; @@ -256,6 +262,8 @@ inline void IncompleteCholesky::updateList(const Idx listCol[rowIdx(jk)].push_back(col); } } + +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -272,6 +280,7 @@ struct solve_retval, Rhs> }; } // end namespace internal +#endif } // end namespace Eigen diff --git a/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h b/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h index 67e780181..e86f65644 100644 --- a/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h +++ b/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h @@ -13,8 +13,12 @@ namespace Eigen { template -class IncompleteLU +class IncompleteLU : public SparseSolverBase > { + protected: + typedef SparseSolverBase > Base; + using Base::m_isInitialized; + typedef _Scalar Scalar; typedef Matrix Vector; typedef typename Vector::Index Index; @@ -23,10 +27,10 @@ class IncompleteLU public: typedef Matrix MatrixType; - IncompleteLU() : m_isInitialized(false) {} + IncompleteLU() {} template - IncompleteLU(const MatrixType& mat) : m_isInitialized(false) + IncompleteLU(const MatrixType& mat) { compute(mat); } @@ -71,12 +75,13 @@ class IncompleteLU } template - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const Rhs& b, Dest& x) const { x = m_lu.template triangularView().solve(b); x = m_lu.template triangularView().solve(x); } +#ifndef EIGEN_TEST_EVALUATORS template inline const internal::solve_retval solve(const MatrixBase& b) const { @@ -85,12 +90,13 @@ class IncompleteLU && "IncompleteLU::solve(): invalid number of rows of the right hand side matrix b"); return internal::solve_retval(*this, b.derived()); } +#endif protected: FactorType m_lu; - bool m_isInitialized; }; +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -102,11 +108,12 @@ struct solve_retval, Rhs> template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal +#endif } // end namespace Eigen diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index 98f9ecc17..28d5c692d 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -2,7 +2,7 @@ // for linear algebra. // // Copyright (C) 2012 Giacomo Po -// Copyright (C) 2011 Gael Guennebaud +// Copyright (C) 2011-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -217,6 +217,7 @@ namespace Eigen { using Base::m_info; using Base::m_isInitialized; public: + using Base::_solve_impl; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::Index Index; @@ -244,7 +245,8 @@ namespace Eigen { /** Destructor. */ ~MINRES(){} - + +#ifndef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A * \a x0 as an initial solution. * @@ -260,10 +262,11 @@ namespace Eigen { return internal::solve_retval_with_guess (*this, b.derived(), x0); } +#endif /** \internal */ template - void _solveWithGuess(const Rhs& b, Dest& x) const + void _solve_with_guess_impl(const Rhs& b, Dest& x) const { m_iterations = Base::maxIterations(); m_error = Base::m_tolerance; @@ -284,16 +287,17 @@ namespace Eigen { /** \internal */ template - void _solve(const Rhs& b, Dest& x) const + void _solve_impl(const Rhs& b, MatrixBase &x) const { x.setZero(); - _solveWithGuess(b,x); + _solve_with_guess_impl(b,x.derived()); } protected: }; +#ifndef EIGEN_TEST_EVALUATORS namespace internal { template @@ -305,12 +309,13 @@ namespace Eigen { template void evalTo(Dest& dst) const { - dec()._solve(rhs(),dst); + dec()._solve_impl(rhs(),dst); } }; } // end namespace internal - +#endif + } // end namespace Eigen #endif // EIGEN_MINRES_H From daad9585a3635223b273cb45ba2a965d91060ff8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:24:07 +0200 Subject: [PATCH 0757/1103] Fix Kronecker product in legacy mode. --- unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index b8f2cba17..ca66d4d89 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -215,7 +215,7 @@ struct traits > typedef typename remove_all<_Lhs>::type Lhs; typedef typename remove_all<_Rhs>::type Rhs; typedef typename scalar_product_traits::ReturnType Scalar; - typedef typename promote_storage_type::StorageKind, typename traits::StorageKind>::ret StorageKind; + typedef typename cwise_promote_storage_type::StorageKind, typename traits::StorageKind, scalar_product_op >::ret StorageKind; typedef typename promote_index_type::type Index; enum { From 8754341848b125b1f720b9b210875aa1543666d8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:25:13 +0200 Subject: [PATCH 0758/1103] Fix remaining garbage during a merge. --- unsupported/Eigen/MPRealSupport | 4 ---- 1 file changed, 4 deletions(-) diff --git a/unsupported/Eigen/MPRealSupport b/unsupported/Eigen/MPRealSupport index 630e0aa77..8e42965a3 100644 --- a/unsupported/Eigen/MPRealSupport +++ b/unsupported/Eigen/MPRealSupport @@ -159,11 +159,7 @@ int main() { if(rows==0 || cols==0 || depth==0) return; -<<<<<<< local -======= - ->>>>>>> other mpreal acc1(0,mpfr_get_prec(blockA[0].mpfr_srcptr())), tmp (0,mpfr_get_prec(blockA[0].mpfr_srcptr())); From b121eecf606b584d72d264d2864460e963050687 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:34:55 +0200 Subject: [PATCH 0759/1103] Fix regression is sparse-sparse product --- .../src/SparseCore/ConservativeSparseSparseProduct.h | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-) diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 8067565f9..67bc33a93 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -28,6 +28,8 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r ei_declare_aligned_stack_constructed_variable(bool, mask, rows, 0); ei_declare_aligned_stack_constructed_variable(Scalar, values, rows, 0); ei_declare_aligned_stack_constructed_variable(Index, indices, rows, 0); + + std::memset(mask,0,sizeof(bool)*rows); // estimate the number of non zero entries // given a rhs column containing Y non zeros, we assume that the respective Y columns @@ -136,20 +138,21 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrixAux; typedef typename sparse_eval::type ColMajorMatrix; - ColMajorMatrix resCol(lhs.rows(),rhs.cols()); // FIXME, the following heuristic is probably not very good. if(lhs.rows()>=rhs.cols()) { + ColMajorMatrix resCol(lhs.rows(),rhs.cols()); // perform sorted insertion internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, true); - res.swap(resCol); + res = resCol.markAsRValue(); } else { + ColMajorMatrixAux resCol(lhs.rows(),rhs.cols()); // ressort to transpose to sort the entries - internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, false); + internal::conservative_sparse_sparse_product_impl(lhs, rhs, resCol, false); RowMajorMatrix resRow(resCol); - res = resRow; + res = resRow.markAsRValue(); } } }; From 72c4f8ca8fc64a812801295babfb6d56ca96f427 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 17:35:58 +0200 Subject: [PATCH 0760/1103] Disable a few unit tests in unsupported --- unsupported/test/CMakeLists.txt | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-) diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 0a6c56c19..5e9e74050 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -5,6 +5,11 @@ add_custom_target(BuildUnsupported) include_directories(../../test ../../unsupported ../../Eigen ${CMAKE_CURRENT_BINARY_DIR}/../../test) + +if(EIGEN_TEST_NO_EVALUATORS) + add_definitions("-DEIGEN_TEST_NO_EVALUATORS=1") +endif(EIGEN_TEST_NO_EVALUATORS) + find_package(GoogleHash) if(GOOGLEHASH_FOUND) add_definitions("-DEIGEN_GOOGLEHASH_SUPPORT") @@ -28,18 +33,23 @@ endif(ADOLC_FOUND) ei_add_test(NonLinearOptimization) ei_add_test(NumericalDiff) +if(EIGEN_TEST_NO_EVALUATORS) ei_add_test(autodiff_scalar) ei_add_test(autodiff) +endif() if (NOT CMAKE_CXX_COMPILER MATCHES "clang\\+\\+$") ei_add_test(BVH) endif() +if(EIGEN_TEST_NO_EVALUATORS) ei_add_test(matrix_exponential) ei_add_test(matrix_function) ei_add_test(matrix_power) ei_add_test(matrix_square_root) ei_add_test(alignedvector3) +endif() + ei_add_test(FFT) find_package(MPFR 2.3.0) @@ -86,12 +96,14 @@ endif() ei_add_test(polynomialsolver) ei_add_test(polynomialutils) -ei_add_test(kronecker_product) ei_add_test(splines) ei_add_test(gmres) ei_add_test(minres) ei_add_test(levenberg_marquardt) +if(EIGEN_TEST_NO_EVALUATORS) ei_add_test(bdcsvd) +ei_add_test(kronecker_product) +endif() option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." OFF) if(EIGEN_TEST_CXX11) From eb392960285c2645d45118d424786a73768ff50a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 18:16:20 +0200 Subject: [PATCH 0761/1103] Reafctoring in D&C SVD unsupported module: clean and merge the SVDBase class to Eigen/SVD, rm copy/pasted JacobiSVD.h file --- Eigen/SVD | 1 + Eigen/src/SVD/JacobiSVD.h | 182 +--- .../Eigen => Eigen}/src/SVD/SVDBase.h | 139 ++-- Eigen/src/SVD/UpperBidiagonalization.h | 1 + unsupported/Eigen/BDCSVD | 26 + unsupported/Eigen/SVD | 35 - .../Eigen/src/{SVD => BDCSVD}/BDCSVD.h | 48 +- unsupported/Eigen/src/BDCSVD/CMakeLists.txt | 6 + .../Eigen/src/{SVD => BDCSVD}/TODOBdcsvd.txt | 0 .../src/{SVD => BDCSVD}/doneInBDCSVD.txt | 0 unsupported/Eigen/src/CMakeLists.txt | 1 + unsupported/Eigen/src/SVD/CMakeLists.txt | 6 - unsupported/Eigen/src/SVD/JacobiSVD.h | 782 ------------------ unsupported/test/svd_common.h | 2 +- 14 files changed, 180 insertions(+), 1049 deletions(-) rename {unsupported/Eigen => Eigen}/src/SVD/SVDBase.h (65%) create mode 100644 unsupported/Eigen/BDCSVD delete mode 100644 unsupported/Eigen/SVD rename unsupported/Eigen/src/{SVD => BDCSVD}/BDCSVD.h (96%) create mode 100644 unsupported/Eigen/src/BDCSVD/CMakeLists.txt rename unsupported/Eigen/src/{SVD => BDCSVD}/TODOBdcsvd.txt (100%) rename unsupported/Eigen/src/{SVD => BDCSVD}/doneInBDCSVD.txt (100%) delete mode 100644 unsupported/Eigen/src/SVD/CMakeLists.txt delete mode 100644 unsupported/Eigen/src/SVD/JacobiSVD.h diff --git a/Eigen/SVD b/Eigen/SVD index 5eee46df5..c3d24286c 100644 --- a/Eigen/SVD +++ b/Eigen/SVD @@ -21,6 +21,7 @@ */ #include "src/misc/Solve.h" +#include "src/SVD/SVDBase.h" #include "src/SVD/JacobiSVD.h" #if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT) #include "src/SVD/JacobiSVD_MKL.h" diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 412daa746..6f3907f5d 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2009-2010 Benoit Jacob +// Copyright (C) 2013-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -442,6 +443,12 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, *j_left = rot1 * j_right->transpose(); } +template +struct traits > +{ + typedef _MatrixType MatrixType; +}; + } // end namespace internal /** \ingroup SVD_Module @@ -498,7 +505,9 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, * \sa MatrixBase::jacobiSvd() */ template class JacobiSVD + : public SVDBase > { + typedef SVDBase Base; public: typedef _MatrixType MatrixType; @@ -515,13 +524,10 @@ template class JacobiSVD MatrixOptions = MatrixType::Options }; - typedef Matrix - MatrixUType; - typedef Matrix - MatrixVType; - typedef typename internal::plain_diag_type::type SingularValuesType; + typedef typename Base::MatrixUType MatrixUType; + typedef typename Base::MatrixVType MatrixVType; + typedef typename Base::SingularValuesType SingularValuesType; + typedef typename internal::plain_row_type::type RowType; typedef typename internal::plain_col_type::type ColType; typedef Matrix class JacobiSVD * perform decompositions via JacobiSVD::compute(const MatrixType&). */ JacobiSVD() - : m_isInitialized(false), - m_isAllocated(false), - m_usePrescribedThreshold(false), - m_computationOptions(0), - m_rows(-1), m_cols(-1), m_diagSize(0) {} @@ -549,11 +550,6 @@ template class JacobiSVD * \sa JacobiSVD() */ JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0) - : m_isInitialized(false), - m_isAllocated(false), - m_usePrescribedThreshold(false), - m_computationOptions(0), - m_rows(-1), m_cols(-1) { allocate(rows, cols, computationOptions); } @@ -569,11 +565,6 @@ template class JacobiSVD * available with the (non-default) FullPivHouseholderQR preconditioner. */ JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0) - : m_isInitialized(false), - m_isAllocated(false), - m_usePrescribedThreshold(false), - m_computationOptions(0), - m_rows(-1), m_cols(-1) { compute(matrix, computationOptions); } @@ -601,54 +592,6 @@ template class JacobiSVD return compute(matrix, m_computationOptions); } - /** \returns the \a U matrix. - * - * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, - * the U matrix is n-by-n if you asked for #ComputeFullU, and is n-by-m if you asked for #ComputeThinU. - * - * The \a m first columns of \a U are the left singular vectors of the matrix being decomposed. - * - * This method asserts that you asked for \a U to be computed. - */ - const MatrixUType& matrixU() const - { - eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); - eigen_assert(computeU() && "This JacobiSVD decomposition didn't compute U. Did you ask for it?"); - return m_matrixU; - } - - /** \returns the \a V matrix. - * - * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, - * the V matrix is p-by-p if you asked for #ComputeFullV, and is p-by-m if you asked for ComputeThinV. - * - * The \a m first columns of \a V are the right singular vectors of the matrix being decomposed. - * - * This method asserts that you asked for \a V to be computed. - */ - const MatrixVType& matrixV() const - { - eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); - eigen_assert(computeV() && "This JacobiSVD decomposition didn't compute V. Did you ask for it?"); - return m_matrixV; - } - - /** \returns the vector of singular values. - * - * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, the - * returned vector has size \a m. Singular values are always sorted in decreasing order. - */ - const SingularValuesType& singularValues() const - { - eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); - return m_singularValues; - } - - /** \returns true if \a U (full or thin) is asked for in this SVD decomposition */ - inline bool computeU() const { return m_computeFullU || m_computeThinU; } - /** \returns true if \a V (full or thin) is asked for in this SVD decomposition */ - inline bool computeV() const { return m_computeFullV || m_computeThinV; } - /** \returns a (least squares) solution of \f$ A x = b \f$ using the current SVD decomposition of A. * * \param b the right-hand-side of the equation to solve. @@ -666,94 +609,31 @@ template class JacobiSVD eigen_assert(computeU() && computeV() && "JacobiSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); return internal::solve_retval(*this, b.derived()); } - - /** \returns the number of singular values that are not exactly 0 */ - Index nonzeroSingularValues() const - { - eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); - return m_nonzeroSingularValues; - } - /** \returns the rank of the matrix of which \c *this is the SVD. - * - * \note This method has to determine which singular values should be considered nonzero. - * For that, it uses the threshold value that you can control by calling - * setThreshold(const RealScalar&). - */ - inline Index rank() const - { - using std::abs; - eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); - if(m_singularValues.size()==0) return 0; - RealScalar premultiplied_threshold = m_singularValues.coeff(0) * threshold(); - Index i = m_nonzeroSingularValues-1; - while(i>=0 && m_singularValues.coeff(i) < premultiplied_threshold) --i; - return i+1; - } + using Base::computeU; + using Base::computeV; - /** Allows to prescribe a threshold to be used by certain methods, such as rank() and solve(), - * which need to determine when singular values are to be considered nonzero. - * This is not used for the SVD decomposition itself. - * - * When it needs to get the threshold value, Eigen calls threshold(). - * The default is \c NumTraits::epsilon() - * - * \param threshold The new value to use as the threshold. - * - * A singular value will be considered nonzero if its value is strictly greater than - * \f$ \vert singular value \vert \leqslant threshold \times \vert max singular value \vert \f$. - * - * If you want to come back to the default behavior, call setThreshold(Default_t) - */ - JacobiSVD& setThreshold(const RealScalar& threshold) - { - m_usePrescribedThreshold = true; - m_prescribedThreshold = threshold; - return *this; - } - - /** Allows to come back to the default behavior, letting Eigen use its default formula for - * determining the threshold. - * - * You should pass the special object Eigen::Default as parameter here. - * \code svd.setThreshold(Eigen::Default); \endcode - * - * See the documentation of setThreshold(const RealScalar&). - */ - JacobiSVD& setThreshold(Default_t) - { - m_usePrescribedThreshold = false; - return *this; - } - - /** Returns the threshold that will be used by certain methods such as rank(). - * - * See the documentation of setThreshold(const RealScalar&). - */ - RealScalar threshold() const - { - eigen_assert(m_isInitialized || m_usePrescribedThreshold); - return m_usePrescribedThreshold ? m_prescribedThreshold - : (std::max)(1,m_diagSize)*NumTraits::epsilon(); - } - - inline Index rows() const { return m_rows; } - inline Index cols() const { return m_cols; } - private: void allocate(Index rows, Index cols, unsigned int computationOptions); protected: - MatrixUType m_matrixU; - MatrixVType m_matrixV; - SingularValuesType m_singularValues; + using Base::m_matrixU; + using Base::m_matrixV; + using Base::m_singularValues; + using Base::m_isInitialized; + using Base::m_isAllocated; + using Base::m_usePrescribedThreshold; + using Base::m_computeFullU; + using Base::m_computeThinU; + using Base::m_computeFullV; + using Base::m_computeThinV; + using Base::m_computationOptions; + using Base::m_nonzeroSingularValues; + using Base::m_rows; + using Base::m_cols; + using Base::m_diagSize; + using Base::m_prescribedThreshold; WorkMatrixType m_workMatrix; - bool m_isInitialized, m_isAllocated, m_usePrescribedThreshold; - bool m_computeFullU, m_computeThinU; - bool m_computeFullV, m_computeThinV; - unsigned int m_computationOptions; - Index m_nonzeroSingularValues, m_rows, m_cols, m_diagSize; - RealScalar m_prescribedThreshold; template friend struct internal::svd_precondition_2x2_block_to_be_real; diff --git a/unsupported/Eigen/src/SVD/SVDBase.h b/Eigen/src/SVD/SVDBase.h similarity index 65% rename from unsupported/Eigen/src/SVD/SVDBase.h rename to Eigen/src/SVD/SVDBase.h index fd8af3b8c..61b01fb8a 100644 --- a/unsupported/Eigen/src/SVD/SVDBase.h +++ b/Eigen/src/SVD/SVDBase.h @@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2009-2010 Benoit Jacob +// Copyright (C) 2014 Gael Guennebaud // // Copyright (C) 2013 Gauthier Brun // Copyright (C) 2013 Nicolas Carre @@ -12,8 +13,8 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. -#ifndef EIGEN_SVD_H -#define EIGEN_SVD_H +#ifndef EIGEN_SVDBASE_H +#define EIGEN_SVDBASE_H namespace Eigen { /** \ingroup SVD_Module @@ -21,9 +22,10 @@ namespace Eigen { * * \class SVDBase * - * \brief Mother class of SVD classes algorithms + * \brief Base class of SVD algorithms + * + * \tparam Derived the type of the actual SVD decomposition * - * \param MatrixType the type of the matrix of which we are computing the SVD decomposition * SVD decomposition consists in decomposing any n-by-p matrix \a A as a product * \f[ A = U S V^* \f] * where \a U is a n-by-n unitary, \a V is a p-by-p unitary, and \a S is a n-by-p real positive matrix which is zero outside of its main diagonal; @@ -42,12 +44,12 @@ namespace Eigen { * terminate in finite (and reasonable) time. * \sa MatrixBase::genericSvd() */ -template +template class SVDBase { public: - typedef _MatrixType MatrixType; + typedef typename internal::traits::MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; typedef typename MatrixType::Index Index; @@ -61,46 +63,16 @@ public: MatrixOptions = MatrixType::Options }; - typedef Matrix - MatrixUType; - typedef Matrix - MatrixVType; + typedef Matrix MatrixUType; + typedef Matrix MatrixVType; typedef typename internal::plain_diag_type::type SingularValuesType; - typedef typename internal::plain_row_type::type RowType; - typedef typename internal::plain_col_type::type ColType; - typedef Matrix - WorkMatrixType; - - - - - /** \brief Method performing the decomposition of given matrix using custom options. - * - * \param matrix the matrix to decompose - * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed. - * By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU, - * #ComputeFullV, #ComputeThinV. - * - * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not - * available with the (non-default) FullPivHouseholderQR preconditioner. - */ - SVDBase& compute(const MatrixType& matrix, unsigned int computationOptions); - - /** \brief Method performing the decomposition of given matrix using current options. - * - * \param matrix the matrix to decompose - * - * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int). - */ - //virtual SVDBase& compute(const MatrixType& matrix) = 0; - SVDBase& compute(const MatrixType& matrix); + + Derived& derived() { return *static_cast(this); } + const Derived& derived() const { return *static_cast(this); } /** \returns the \a U matrix. * - * For the SVDBase decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, + * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, * the U matrix is n-by-n if you asked for #ComputeFullU, and is n-by-m if you asked for #ComputeThinU. * * The \a m first columns of \a U are the left singular vectors of the matrix being decomposed. @@ -141,26 +113,84 @@ public: return m_singularValues; } - - /** \returns the number of singular values that are not exactly 0 */ Index nonzeroSingularValues() const { eigen_assert(m_isInitialized && "SVD is not initialized."); return m_nonzeroSingularValues; } + + /** \returns the rank of the matrix of which \c *this is the SVD. + * + * \note This method has to determine which singular values should be considered nonzero. + * For that, it uses the threshold value that you can control by calling + * setThreshold(const RealScalar&). + */ + inline Index rank() const + { + using std::abs; + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + if(m_singularValues.size()==0) return 0; + RealScalar premultiplied_threshold = m_singularValues.coeff(0) * threshold(); + Index i = m_nonzeroSingularValues-1; + while(i>=0 && m_singularValues.coeff(i) < premultiplied_threshold) --i; + return i+1; + } + + /** Allows to prescribe a threshold to be used by certain methods, such as rank() and solve(), + * which need to determine when singular values are to be considered nonzero. + * This is not used for the SVD decomposition itself. + * + * When it needs to get the threshold value, Eigen calls threshold(). + * The default is \c NumTraits::epsilon() + * + * \param threshold The new value to use as the threshold. + * + * A singular value will be considered nonzero if its value is strictly greater than + * \f$ \vert singular value \vert \leqslant threshold \times \vert max singular value \vert \f$. + * + * If you want to come back to the default behavior, call setThreshold(Default_t) + */ + Derived& setThreshold(const RealScalar& threshold) + { + m_usePrescribedThreshold = true; + m_prescribedThreshold = threshold; + return derived(); + } + /** Allows to come back to the default behavior, letting Eigen use its default formula for + * determining the threshold. + * + * You should pass the special object Eigen::Default as parameter here. + * \code svd.setThreshold(Eigen::Default); \endcode + * + * See the documentation of setThreshold(const RealScalar&). + */ + Derived& setThreshold(Default_t) + { + m_usePrescribedThreshold = false; + return derived(); + } + + /** Returns the threshold that will be used by certain methods such as rank(). + * + * See the documentation of setThreshold(const RealScalar&). + */ + RealScalar threshold() const + { + eigen_assert(m_isInitialized || m_usePrescribedThreshold); + return m_usePrescribedThreshold ? m_prescribedThreshold + : (std::max)(1,m_diagSize)*NumTraits::epsilon(); + } /** \returns true if \a U (full or thin) is asked for in this SVD decomposition */ inline bool computeU() const { return m_computeFullU || m_computeThinU; } /** \returns true if \a V (full or thin) is asked for in this SVD decomposition */ inline bool computeV() const { return m_computeFullV || m_computeThinV; } - inline Index rows() const { return m_rows; } inline Index cols() const { return m_cols; } - protected: // return true if already allocated bool allocate(Index rows, Index cols, unsigned int computationOptions) ; @@ -168,12 +198,12 @@ protected: MatrixUType m_matrixU; MatrixVType m_matrixV; SingularValuesType m_singularValues; - bool m_isInitialized, m_isAllocated; + bool m_isInitialized, m_isAllocated, m_usePrescribedThreshold; bool m_computeFullU, m_computeThinU; bool m_computeFullV, m_computeThinV; unsigned int m_computationOptions; Index m_nonzeroSingularValues, m_rows, m_cols, m_diagSize; - + RealScalar m_prescribedThreshold; /** \brief Default Constructor. * @@ -182,8 +212,9 @@ protected: SVDBase() : m_isInitialized(false), m_isAllocated(false), + m_usePrescribedThreshold(false), m_computationOptions(0), - m_rows(-1), m_cols(-1) + m_rows(-1), m_cols(-1), m_diagSize(0) {} @@ -220,17 +251,13 @@ bool SVDBase::allocate(Index rows, Index cols, unsigned int computat m_diagSize = (std::min)(m_rows, m_cols); m_singularValues.resize(m_diagSize); if(RowsAtCompileTime==Dynamic) - m_matrixU.resize(m_rows, m_computeFullU ? m_rows - : m_computeThinU ? m_diagSize - : 0); + m_matrixU.resize(m_rows, m_computeFullU ? m_rows : m_computeThinU ? m_diagSize : 0); if(ColsAtCompileTime==Dynamic) - m_matrixV.resize(m_cols, m_computeFullV ? m_cols - : m_computeThinV ? m_diagSize - : 0); + m_matrixV.resize(m_cols, m_computeFullV ? m_cols : m_computeThinV ? m_diagSize : 0); return false; } }// end namespace -#endif // EIGEN_SVD_H +#endif // EIGEN_SVDBASE_H diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 40067682c..225b19e3c 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -2,6 +2,7 @@ // for linear algebra. // // Copyright (C) 2010 Benoit Jacob +// Copyright (C) 2013-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/unsupported/Eigen/BDCSVD b/unsupported/Eigen/BDCSVD new file mode 100644 index 000000000..44649dbd0 --- /dev/null +++ b/unsupported/Eigen/BDCSVD @@ -0,0 +1,26 @@ +#ifndef EIGEN_BDCSVD_MODULE_H +#define EIGEN_BDCSVD_MODULE_H + +#include + +#include "../../Eigen/src/Core/util/DisableStupidWarnings.h" + +/** \defgroup BDCSVD_Module BDCSVD module + * + * + * + * This module provides Divide & Conquer SVD decomposition for matrices (both real and complex). + * This decomposition is accessible via the following MatrixBase method: + * - MatrixBase::bdcSvd() + * + * \code + * #include + * \endcode + */ + +#include "src/BDCSVD/BDCSVD.h" + +#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h" + +#endif // EIGEN_BDCSVD_MODULE_H +/* vim: set filetype=cpp et sw=2 ts=2 ai: */ diff --git a/unsupported/Eigen/SVD b/unsupported/Eigen/SVD deleted file mode 100644 index 900a8aa60..000000000 --- a/unsupported/Eigen/SVD +++ /dev/null @@ -1,35 +0,0 @@ -#ifndef EIGEN_SVD_MODULE_H -#define EIGEN_SVD_MODULE_H - -#include -#include -#include - -#include "../../Eigen/src/Core/util/DisableStupidWarnings.h" - -/** \defgroup SVD_Module SVD module - * - * - * - * This module provides SVD decomposition for matrices (both real and complex). - * This decomposition is accessible via the following MatrixBase method: - * - MatrixBase::jacobiSvd() - * - * \code - * #include - * \endcode - */ - -#include "../../Eigen/src/misc/Solve.h" -#include "../../Eigen/src/SVD/UpperBidiagonalization.h" -#include "src/SVD/SVDBase.h" -#include "src/SVD/JacobiSVD.h" -#include "src/SVD/BDCSVD.h" -#if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT) -#include "../../Eigen/src/SVD/JacobiSVD_MKL.h" -#endif - -#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h" - -#endif // EIGEN_SVD_MODULE_H -/* vim: set filetype=cpp et sw=2 ts=2 ai: */ diff --git a/unsupported/Eigen/src/SVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h similarity index 96% rename from unsupported/Eigen/src/SVD/BDCSVD.h rename to unsupported/Eigen/src/BDCSVD/BDCSVD.h index 80006fd60..a7c369633 100644 --- a/unsupported/Eigen/src/SVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -24,6 +24,20 @@ #define ALGOSWAP 16 namespace Eigen { + +template class BDCSVD; + +namespace internal { + +template +struct traits > +{ + typedef _MatrixType MatrixType; +}; + +} // end namespace internal + + /** \ingroup SVD_Module * * @@ -36,9 +50,9 @@ namespace Eigen { * It should be used to speed up the calcul of SVD for big matrices. */ template -class BDCSVD : public SVDBase<_MatrixType> +class BDCSVD : public SVDBase > { - typedef SVDBase<_MatrixType> Base; + typedef SVDBase Base; public: using Base::rows; @@ -77,9 +91,7 @@ public: * The default constructor is useful in cases in which the user intends to * perform decompositions via BDCSVD::compute(const MatrixType&). */ - BDCSVD() - : SVDBase<_MatrixType>::SVDBase(), - algoswap(ALGOSWAP), m_numIters(0) + BDCSVD() : algoswap(ALGOSWAP), m_numIters(0) {} @@ -90,8 +102,7 @@ public: * \sa BDCSVD() */ BDCSVD(Index rows, Index cols, unsigned int computationOptions = 0) - : SVDBase<_MatrixType>::SVDBase(), - algoswap(ALGOSWAP), m_numIters(0) + : algoswap(ALGOSWAP), m_numIters(0) { allocate(rows, cols, computationOptions); } @@ -107,8 +118,7 @@ public: * available with the (non - default) FullPivHouseholderQR preconditioner. */ BDCSVD(const MatrixType& matrix, unsigned int computationOptions = 0) - : SVDBase<_MatrixType>::SVDBase(), - algoswap(ALGOSWAP), m_numIters(0) + : algoswap(ALGOSWAP), m_numIters(0) { compute(matrix, computationOptions); } @@ -116,6 +126,7 @@ public: ~BDCSVD() { } + /** \brief Method performing the decomposition of given matrix using custom options. * * \param matrix the matrix to decompose @@ -126,7 +137,7 @@ public: * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not * available with the (non - default) FullPivHouseholderQR preconditioner. */ - SVDBase& compute(const MatrixType& matrix, unsigned int computationOptions); + BDCSVD& compute(const MatrixType& matrix, unsigned int computationOptions); /** \brief Method performing the decomposition of given matrix using current options. * @@ -134,7 +145,7 @@ public: * * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int). */ - SVDBase& compute(const MatrixType& matrix) + BDCSVD& compute(const MatrixType& matrix) { return compute(matrix, this->m_computationOptions); } @@ -160,8 +171,8 @@ public: solve(const MatrixBase& b) const { eigen_assert(this->m_isInitialized && "BDCSVD is not initialized."); - eigen_assert(SVDBase<_MatrixType>::computeU() && SVDBase<_MatrixType>::computeV() && - "BDCSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); + eigen_assert(computeU() && computeV() && + "BDCSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); return internal::solve_retval(*this, b.derived()); } @@ -195,6 +206,9 @@ public: return this->m_matrixV; } } + + using Base::computeU; + using Base::computeV; private: void allocate(Index rows, Index cols, unsigned int computationOptions); @@ -229,7 +243,7 @@ template void BDCSVD::allocate(Index rows, Index cols, unsigned int computationOptions) { isTranspose = (cols > rows); - if (SVDBase::allocate(rows, cols, computationOptions)) return; + if (Base::allocate(rows, cols, computationOptions)) return; m_computed = MatrixXr::Zero(this->m_diagSize + 1, this->m_diagSize ); if (isTranspose){ compU = this->computeU(); @@ -262,8 +276,7 @@ void BDCSVD::allocate(Index rows, Index cols, unsigned int computati // Methode which compute the BDCSVD for the int template<> -SVDBase >& -BDCSVD >::compute(const MatrixType& matrix, unsigned int computationOptions) { +BDCSVD >& BDCSVD >::compute(const MatrixType& matrix, unsigned int computationOptions) { allocate(matrix.rows(), matrix.cols(), computationOptions); this->m_nonzeroSingularValues = 0; m_computed = Matrix::Zero(rows(), cols()); @@ -279,8 +292,7 @@ BDCSVD >::compute(const MatrixType& matrix, unsign // Methode which compute the BDCSVD template -SVDBase& -BDCSVD::compute(const MatrixType& matrix, unsigned int computationOptions) +BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsigned int computationOptions) { allocate(matrix.rows(), matrix.cols(), computationOptions); using std::abs; diff --git a/unsupported/Eigen/src/BDCSVD/CMakeLists.txt b/unsupported/Eigen/src/BDCSVD/CMakeLists.txt new file mode 100644 index 000000000..73b89ea18 --- /dev/null +++ b/unsupported/Eigen/src/BDCSVD/CMakeLists.txt @@ -0,0 +1,6 @@ +FILE(GLOB Eigen_BDCSVD_SRCS "*.h") + +INSTALL(FILES + ${Eigen_BDCSVD_SRCS} + DESTINATION ${INCLUDE_INSTALL_DIR}unsupported/Eigen/src/BDCSVD COMPONENT Devel + ) diff --git a/unsupported/Eigen/src/SVD/TODOBdcsvd.txt b/unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt similarity index 100% rename from unsupported/Eigen/src/SVD/TODOBdcsvd.txt rename to unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt diff --git a/unsupported/Eigen/src/SVD/doneInBDCSVD.txt b/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt similarity index 100% rename from unsupported/Eigen/src/SVD/doneInBDCSVD.txt rename to unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt diff --git a/unsupported/Eigen/src/CMakeLists.txt b/unsupported/Eigen/src/CMakeLists.txt index 8eb2808e3..654a2327f 100644 --- a/unsupported/Eigen/src/CMakeLists.txt +++ b/unsupported/Eigen/src/CMakeLists.txt @@ -12,3 +12,4 @@ ADD_SUBDIRECTORY(Skyline) ADD_SUBDIRECTORY(SparseExtra) ADD_SUBDIRECTORY(KroneckerProduct) ADD_SUBDIRECTORY(Splines) +ADD_SUBDIRECTORY(BDCSVD) diff --git a/unsupported/Eigen/src/SVD/CMakeLists.txt b/unsupported/Eigen/src/SVD/CMakeLists.txt deleted file mode 100644 index b40baf092..000000000 --- a/unsupported/Eigen/src/SVD/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_SVD_SRCS "*.h") - -INSTALL(FILES - ${Eigen_SVD_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}unsupported/Eigen/src/SVD COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/SVD/JacobiSVD.h b/unsupported/Eigen/src/SVD/JacobiSVD.h deleted file mode 100644 index 02fac409e..000000000 --- a/unsupported/Eigen/src/SVD/JacobiSVD.h +++ /dev/null @@ -1,782 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2009-2010 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_JACOBISVD_H -#define EIGEN_JACOBISVD_H - -namespace Eigen { - -namespace internal { -// forward declaration (needed by ICC) -// the empty body is required by MSVC -template::IsComplex> -struct svd_precondition_2x2_block_to_be_real {}; - -/*** QR preconditioners (R-SVD) - *** - *** Their role is to reduce the problem of computing the SVD to the case of a square matrix. - *** This approach, known as R-SVD, is an optimization for rectangular-enough matrices, and is a requirement for - *** JacobiSVD which by itself is only able to work on square matrices. - ***/ - -enum { PreconditionIfMoreColsThanRows, PreconditionIfMoreRowsThanCols }; - -template -struct qr_preconditioner_should_do_anything -{ - enum { a = MatrixType::RowsAtCompileTime != Dynamic && - MatrixType::ColsAtCompileTime != Dynamic && - MatrixType::ColsAtCompileTime <= MatrixType::RowsAtCompileTime, - b = MatrixType::RowsAtCompileTime != Dynamic && - MatrixType::ColsAtCompileTime != Dynamic && - MatrixType::RowsAtCompileTime <= MatrixType::ColsAtCompileTime, - ret = !( (QRPreconditioner == NoQRPreconditioner) || - (Case == PreconditionIfMoreColsThanRows && bool(a)) || - (Case == PreconditionIfMoreRowsThanCols && bool(b)) ) - }; -}; - -template::ret -> struct qr_preconditioner_impl {}; - -template -class qr_preconditioner_impl -{ -public: - typedef typename MatrixType::Index Index; - void allocate(const JacobiSVD&) {} - bool run(JacobiSVD&, const MatrixType&) - { - return false; - } -}; - -/*** preconditioner using FullPivHouseholderQR ***/ - -template -class qr_preconditioner_impl -{ -public: - typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - enum - { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime - }; - typedef Matrix WorkspaceType; - - void allocate(const JacobiSVD& svd) - { - if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) - { - m_qr.~QRType(); - ::new (&m_qr) QRType(svd.rows(), svd.cols()); - } - if (svd.m_computeFullU) m_workspace.resize(svd.rows()); - } - - bool run(JacobiSVD& svd, const MatrixType& matrix) - { - if(matrix.rows() > matrix.cols()) - { - m_qr.compute(matrix); - svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView(); - if(svd.m_computeFullU) m_qr.matrixQ().evalTo(svd.m_matrixU, m_workspace); - if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation(); - return true; - } - return false; - } -private: - typedef FullPivHouseholderQR QRType; - QRType m_qr; - WorkspaceType m_workspace; -}; - -template -class qr_preconditioner_impl -{ -public: - typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - enum - { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime, - MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, - MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, - Options = MatrixType::Options - }; - typedef Matrix - TransposeTypeWithSameStorageOrder; - - void allocate(const JacobiSVD& svd) - { - if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols()) - { - m_qr.~QRType(); - ::new (&m_qr) QRType(svd.cols(), svd.rows()); - } - m_adjoint.resize(svd.cols(), svd.rows()); - if (svd.m_computeFullV) m_workspace.resize(svd.cols()); - } - - bool run(JacobiSVD& svd, const MatrixType& matrix) - { - if(matrix.cols() > matrix.rows()) - { - m_adjoint = matrix.adjoint(); - m_qr.compute(m_adjoint); - svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView().adjoint(); - if(svd.m_computeFullV) m_qr.matrixQ().evalTo(svd.m_matrixV, m_workspace); - if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation(); - return true; - } - else return false; - } -private: - typedef FullPivHouseholderQR QRType; - QRType m_qr; - TransposeTypeWithSameStorageOrder m_adjoint; - typename internal::plain_row_type::type m_workspace; -}; - -/*** preconditioner using ColPivHouseholderQR ***/ - -template -class qr_preconditioner_impl -{ -public: - typedef typename MatrixType::Index Index; - - void allocate(const JacobiSVD& svd) - { - if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) - { - m_qr.~QRType(); - ::new (&m_qr) QRType(svd.rows(), svd.cols()); - } - if (svd.m_computeFullU) m_workspace.resize(svd.rows()); - else if (svd.m_computeThinU) m_workspace.resize(svd.cols()); - } - - bool run(JacobiSVD& svd, const MatrixType& matrix) - { - if(matrix.rows() > matrix.cols()) - { - m_qr.compute(matrix); - svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView(); - if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace); - else if(svd.m_computeThinU) - { - svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols()); - m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace); - } - if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation(); - return true; - } - return false; - } - -private: - typedef ColPivHouseholderQR QRType; - QRType m_qr; - typename internal::plain_col_type::type m_workspace; -}; - -template -class qr_preconditioner_impl -{ -public: - typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - enum - { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime, - MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, - MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, - Options = MatrixType::Options - }; - - typedef Matrix - TransposeTypeWithSameStorageOrder; - - void allocate(const JacobiSVD& svd) - { - if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols()) - { - m_qr.~QRType(); - ::new (&m_qr) QRType(svd.cols(), svd.rows()); - } - if (svd.m_computeFullV) m_workspace.resize(svd.cols()); - else if (svd.m_computeThinV) m_workspace.resize(svd.rows()); - m_adjoint.resize(svd.cols(), svd.rows()); - } - - bool run(JacobiSVD& svd, const MatrixType& matrix) - { - if(matrix.cols() > matrix.rows()) - { - m_adjoint = matrix.adjoint(); - m_qr.compute(m_adjoint); - - svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView().adjoint(); - if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace); - else if(svd.m_computeThinV) - { - svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows()); - m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace); - } - if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation(); - return true; - } - else return false; - } - -private: - typedef ColPivHouseholderQR QRType; - QRType m_qr; - TransposeTypeWithSameStorageOrder m_adjoint; - typename internal::plain_row_type::type m_workspace; -}; - -/*** preconditioner using HouseholderQR ***/ - -template -class qr_preconditioner_impl -{ -public: - typedef typename MatrixType::Index Index; - - void allocate(const JacobiSVD& svd) - { - if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) - { - m_qr.~QRType(); - ::new (&m_qr) QRType(svd.rows(), svd.cols()); - } - if (svd.m_computeFullU) m_workspace.resize(svd.rows()); - else if (svd.m_computeThinU) m_workspace.resize(svd.cols()); - } - - bool run(JacobiSVD& svd, const MatrixType& matrix) - { - if(matrix.rows() > matrix.cols()) - { - m_qr.compute(matrix); - svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView(); - if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace); - else if(svd.m_computeThinU) - { - svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols()); - m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace); - } - if(svd.computeV()) svd.m_matrixV.setIdentity(matrix.cols(), matrix.cols()); - return true; - } - return false; - } -private: - typedef HouseholderQR QRType; - QRType m_qr; - typename internal::plain_col_type::type m_workspace; -}; - -template -class qr_preconditioner_impl -{ -public: - typedef typename MatrixType::Index Index; - typedef typename MatrixType::Scalar Scalar; - enum - { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime, - MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, - MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, - Options = MatrixType::Options - }; - - typedef Matrix - TransposeTypeWithSameStorageOrder; - - void allocate(const JacobiSVD& svd) - { - if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols()) - { - m_qr.~QRType(); - ::new (&m_qr) QRType(svd.cols(), svd.rows()); - } - if (svd.m_computeFullV) m_workspace.resize(svd.cols()); - else if (svd.m_computeThinV) m_workspace.resize(svd.rows()); - m_adjoint.resize(svd.cols(), svd.rows()); - } - - bool run(JacobiSVD& svd, const MatrixType& matrix) - { - if(matrix.cols() > matrix.rows()) - { - m_adjoint = matrix.adjoint(); - m_qr.compute(m_adjoint); - - svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView().adjoint(); - if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace); - else if(svd.m_computeThinV) - { - svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows()); - m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace); - } - if(svd.computeU()) svd.m_matrixU.setIdentity(matrix.rows(), matrix.rows()); - return true; - } - else return false; - } - -private: - typedef HouseholderQR QRType; - QRType m_qr; - TransposeTypeWithSameStorageOrder m_adjoint; - typename internal::plain_row_type::type m_workspace; -}; - -/*** 2x2 SVD implementation - *** - *** JacobiSVD consists in performing a series of 2x2 SVD subproblems - ***/ - -template -struct svd_precondition_2x2_block_to_be_real -{ - typedef JacobiSVD SVD; - typedef typename SVD::Index Index; - static void run(typename SVD::WorkMatrixType&, SVD&, Index, Index) {} -}; - -template -struct svd_precondition_2x2_block_to_be_real -{ - typedef JacobiSVD SVD; - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::RealScalar RealScalar; - typedef typename SVD::Index Index; - static void run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q) - { - using std::sqrt; - Scalar z; - JacobiRotation rot; - RealScalar n = sqrt(numext::abs2(work_matrix.coeff(p,p)) + numext::abs2(work_matrix.coeff(q,p))); - if(n==0) - { - z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q); - work_matrix.row(p) *= z; - if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z); - z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q); - work_matrix.row(q) *= z; - if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z); - } - else - { - rot.c() = conj(work_matrix.coeff(p,p)) / n; - rot.s() = work_matrix.coeff(q,p) / n; - work_matrix.applyOnTheLeft(p,q,rot); - if(svd.computeU()) svd.m_matrixU.applyOnTheRight(p,q,rot.adjoint()); - if(work_matrix.coeff(p,q) != Scalar(0)) - { - Scalar z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q); - work_matrix.col(q) *= z; - if(svd.computeV()) svd.m_matrixV.col(q) *= z; - } - if(work_matrix.coeff(q,q) != Scalar(0)) - { - z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q); - work_matrix.row(q) *= z; - if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z); - } - } - } -}; - -template -void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, - JacobiRotation *j_left, - JacobiRotation *j_right) -{ - using std::sqrt; - Matrix m; - m << numext::real(matrix.coeff(p,p)), numext::real(matrix.coeff(p,q)), - numext::real(matrix.coeff(q,p)), numext::real(matrix.coeff(q,q)); - JacobiRotation rot1; - RealScalar t = m.coeff(0,0) + m.coeff(1,1); - RealScalar d = m.coeff(1,0) - m.coeff(0,1); - if(t == RealScalar(0)) - { - rot1.c() = RealScalar(0); - rot1.s() = d > RealScalar(0) ? RealScalar(1) : RealScalar(-1); - } - else - { - RealScalar u = d / t; - rot1.c() = RealScalar(1) / sqrt(RealScalar(1) + numext::abs2(u)); - rot1.s() = rot1.c() * u; - } - m.applyOnTheLeft(0,1,rot1); - j_right->makeJacobi(m,0,1); - *j_left = rot1 * j_right->transpose(); -} - -} // end namespace internal - -/** \ingroup SVD_Module - * - * - * \class JacobiSVD - * - * \brief Two-sided Jacobi SVD decomposition of a rectangular matrix - * - * \param MatrixType the type of the matrix of which we are computing the SVD decomposition - * \param QRPreconditioner this optional parameter allows to specify the type of QR decomposition that will be used internally - * for the R-SVD step for non-square matrices. See discussion of possible values below. - * - * SVD decomposition consists in decomposing any n-by-p matrix \a A as a product - * \f[ A = U S V^* \f] - * where \a U is a n-by-n unitary, \a V is a p-by-p unitary, and \a S is a n-by-p real positive matrix which is zero outside of its main diagonal; - * the diagonal entries of S are known as the \em singular \em values of \a A and the columns of \a U and \a V are known as the left - * and right \em singular \em vectors of \a A respectively. - * - * Singular values are always sorted in decreasing order. - * - * This JacobiSVD decomposition computes only the singular values by default. If you want \a U or \a V, you need to ask for them explicitly. - * - * You can ask for only \em thin \a U or \a V to be computed, meaning the following. In case of a rectangular n-by-p matrix, letting \a m be the - * smaller value among \a n and \a p, there are only \a m singular vectors; the remaining columns of \a U and \a V do not correspond to actual - * singular vectors. Asking for \em thin \a U or \a V means asking for only their \a m first columns to be formed. So \a U is then a n-by-m matrix, - * and \a V is then a p-by-m matrix. Notice that thin \a U and \a V are all you need for (least squares) solving. - * - * Here's an example demonstrating basic usage: - * \include JacobiSVD_basic.cpp - * Output: \verbinclude JacobiSVD_basic.out - * - * This JacobiSVD class is a two-sided Jacobi R-SVD decomposition, ensuring optimal reliability and accuracy. The downside is that it's slower than - * bidiagonalizing SVD algorithms for large square matrices; however its complexity is still \f$ O(n^2p) \f$ where \a n is the smaller dimension and - * \a p is the greater dimension, meaning that it is still of the same order of complexity as the faster bidiagonalizing R-SVD algorithms. - * In particular, like any R-SVD, it takes advantage of non-squareness in that its complexity is only linear in the greater dimension. - * - * If the input matrix has inf or nan coefficients, the result of the computation is undefined, but the computation is guaranteed to - * terminate in finite (and reasonable) time. - * - * The possible values for QRPreconditioner are: - * \li ColPivHouseholderQRPreconditioner is the default. In practice it's very safe. It uses column-pivoting QR. - * \li FullPivHouseholderQRPreconditioner, is the safest and slowest. It uses full-pivoting QR. - * Contrary to other QRs, it doesn't allow computing thin unitaries. - * \li HouseholderQRPreconditioner is the fastest, and less safe and accurate than the pivoting variants. It uses non-pivoting QR. - * This is very similar in safety and accuracy to the bidiagonalization process used by bidiagonalizing SVD algorithms (since bidiagonalization - * is inherently non-pivoting). However the resulting SVD is still more reliable than bidiagonalizing SVDs because the Jacobi-based iterarive - * process is more reliable than the optimized bidiagonal SVD iterations. - * \li NoQRPreconditioner allows not to use a QR preconditioner at all. This is useful if you know that you will only be computing - * JacobiSVD decompositions of square matrices. Non-square matrices require a QR preconditioner. Using this option will result in - * faster compilation and smaller executable code. It won't significantly speed up computation, since JacobiSVD is always checking - * if QR preconditioning is needed before applying it anyway. - * - * \sa MatrixBase::jacobiSvd() - */ -template -class JacobiSVD : public SVDBase<_MatrixType> -{ - public: - - typedef _MatrixType MatrixType; - typedef typename MatrixType::Scalar Scalar; - typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime, - DiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime,ColsAtCompileTime), - MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, - MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, - MaxDiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(MaxRowsAtCompileTime,MaxColsAtCompileTime), - MatrixOptions = MatrixType::Options - }; - - typedef Matrix - MatrixUType; - typedef Matrix - MatrixVType; - typedef typename internal::plain_diag_type::type SingularValuesType; - typedef typename internal::plain_row_type::type RowType; - typedef typename internal::plain_col_type::type ColType; - typedef Matrix - WorkMatrixType; - - /** \brief Default Constructor. - * - * The default constructor is useful in cases in which the user intends to - * perform decompositions via JacobiSVD::compute(const MatrixType&). - */ - JacobiSVD() - : SVDBase<_MatrixType>::SVDBase() - {} - - - /** \brief Default Constructor with memory preallocation - * - * Like the default constructor but with preallocation of the internal data - * according to the specified problem size. - * \sa JacobiSVD() - */ - JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0) - : SVDBase<_MatrixType>::SVDBase() - { - allocate(rows, cols, computationOptions); - } - - /** \brief Constructor performing the decomposition of given matrix. - * - * \param matrix the matrix to decompose - * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed. - * By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU, - * #ComputeFullV, #ComputeThinV. - * - * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not - * available with the (non-default) FullPivHouseholderQR preconditioner. - */ - JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0) - : SVDBase<_MatrixType>::SVDBase() - { - compute(matrix, computationOptions); - } - - /** \brief Method performing the decomposition of given matrix using custom options. - * - * \param matrix the matrix to decompose - * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed. - * By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU, - * #ComputeFullV, #ComputeThinV. - * - * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not - * available with the (non-default) FullPivHouseholderQR preconditioner. - */ - SVDBase& compute(const MatrixType& matrix, unsigned int computationOptions); - - /** \brief Method performing the decomposition of given matrix using current options. - * - * \param matrix the matrix to decompose - * - * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int). - */ - SVDBase& compute(const MatrixType& matrix) - { - return compute(matrix, this->m_computationOptions); - } - - /** \returns a (least squares) solution of \f$ A x = b \f$ using the current SVD decomposition of A. - * - * \param b the right-hand-side of the equation to solve. - * - * \note Solving requires both U and V to be computed. Thin U and V are enough, there is no need for full U or V. - * - * \note SVD solving is implicitly least-squares. Thus, this method serves both purposes of exact solving and least-squares solving. - * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$. - */ - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(this->m_isInitialized && "JacobiSVD is not initialized."); - eigen_assert(SVDBase::computeU() && SVDBase::computeV() && "JacobiSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); - return internal::solve_retval(*this, b.derived()); - } - - - - private: - void allocate(Index rows, Index cols, unsigned int computationOptions); - - protected: - WorkMatrixType m_workMatrix; - - template - friend struct internal::svd_precondition_2x2_block_to_be_real; - template - friend struct internal::qr_preconditioner_impl; - - internal::qr_preconditioner_impl m_qr_precond_morecols; - internal::qr_preconditioner_impl m_qr_precond_morerows; -}; - -template -void JacobiSVD::allocate(Index rows, Index cols, unsigned int computationOptions) -{ - if (SVDBase::allocate(rows, cols, computationOptions)) return; - - if (QRPreconditioner == FullPivHouseholderQRPreconditioner) - { - eigen_assert(!(this->m_computeThinU || this->m_computeThinV) && - "JacobiSVD: can't compute thin U or thin V with the FullPivHouseholderQR preconditioner. " - "Use the ColPivHouseholderQR preconditioner instead."); - } - - m_workMatrix.resize(this->m_diagSize, this->m_diagSize); - - if(this->m_cols>this->m_rows) m_qr_precond_morecols.allocate(*this); - if(this->m_rows>this->m_cols) m_qr_precond_morerows.allocate(*this); -} - -template -SVDBase& -JacobiSVD::compute(const MatrixType& matrix, unsigned int computationOptions) -{ - using std::abs; - allocate(matrix.rows(), matrix.cols(), computationOptions); - - // currently we stop when we reach precision 2*epsilon as the last bit of precision can require an unreasonable number of iterations, - // only worsening the precision of U and V as we accumulate more rotations - const RealScalar precision = RealScalar(2) * NumTraits::epsilon(); - - // limit for very small denormal numbers to be considered zero in order to avoid infinite loops (see bug 286) - const RealScalar considerAsZero = RealScalar(2) * std::numeric_limits::denorm_min(); - - /*** step 1. The R-SVD step: we use a QR decomposition to reduce to the case of a square matrix */ - - if(!m_qr_precond_morecols.run(*this, matrix) && !m_qr_precond_morerows.run(*this, matrix)) - { - m_workMatrix = matrix.block(0,0,this->m_diagSize,this->m_diagSize); - if(this->m_computeFullU) this->m_matrixU.setIdentity(this->m_rows,this->m_rows); - if(this->m_computeThinU) this->m_matrixU.setIdentity(this->m_rows,this->m_diagSize); - if(this->m_computeFullV) this->m_matrixV.setIdentity(this->m_cols,this->m_cols); - if(this->m_computeThinV) this->m_matrixV.setIdentity(this->m_cols, this->m_diagSize); - } - - /*** step 2. The main Jacobi SVD iteration. ***/ - - bool finished = false; - while(!finished) - { - finished = true; - - // do a sweep: for all index pairs (p,q), perform SVD of the corresponding 2x2 sub-matrix - - for(Index p = 1; p < this->m_diagSize; ++p) - { - for(Index q = 0; q < p; ++q) - { - // if this 2x2 sub-matrix is not diagonal already... - // notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't - // keep us iterating forever. Similarly, small denormal numbers are considered zero. - using std::max; - RealScalar threshold = (max)(considerAsZero, precision * (max)(abs(m_workMatrix.coeff(p,p)), - abs(m_workMatrix.coeff(q,q)))); - if((max)(abs(m_workMatrix.coeff(p,q)),abs(m_workMatrix.coeff(q,p))) > threshold) - { - finished = false; - - // perform SVD decomposition of 2x2 sub-matrix corresponding to indices p,q to make it diagonal - internal::svd_precondition_2x2_block_to_be_real::run(m_workMatrix, *this, p, q); - JacobiRotation j_left, j_right; - internal::real_2x2_jacobi_svd(m_workMatrix, p, q, &j_left, &j_right); - - // accumulate resulting Jacobi rotations - m_workMatrix.applyOnTheLeft(p,q,j_left); - if(SVDBase::computeU()) this->m_matrixU.applyOnTheRight(p,q,j_left.transpose()); - - m_workMatrix.applyOnTheRight(p,q,j_right); - if(SVDBase::computeV()) this->m_matrixV.applyOnTheRight(p,q,j_right); - } - } - } - } - - /*** step 3. The work matrix is now diagonal, so ensure it's positive so its diagonal entries are the singular values ***/ - - for(Index i = 0; i < this->m_diagSize; ++i) - { - RealScalar a = abs(m_workMatrix.coeff(i,i)); - this->m_singularValues.coeffRef(i) = a; - if(SVDBase::computeU() && (a!=RealScalar(0))) this->m_matrixU.col(i) *= this->m_workMatrix.coeff(i,i)/a; - } - - /*** step 4. Sort singular values in descending order and compute the number of nonzero singular values ***/ - - this->m_nonzeroSingularValues = this->m_diagSize; - for(Index i = 0; i < this->m_diagSize; i++) - { - Index pos; - RealScalar maxRemainingSingularValue = this->m_singularValues.tail(this->m_diagSize-i).maxCoeff(&pos); - if(maxRemainingSingularValue == RealScalar(0)) - { - this->m_nonzeroSingularValues = i; - break; - } - if(pos) - { - pos += i; - std::swap(this->m_singularValues.coeffRef(i), this->m_singularValues.coeffRef(pos)); - if(SVDBase::computeU()) this->m_matrixU.col(pos).swap(this->m_matrixU.col(i)); - if(SVDBase::computeV()) this->m_matrixV.col(pos).swap(this->m_matrixV.col(i)); - } - } - - this->m_isInitialized = true; - return *this; -} - -namespace internal { -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef JacobiSVD<_MatrixType, QRPreconditioner> JacobiSVDType; - EIGEN_MAKE_SOLVE_HELPERS(JacobiSVDType,Rhs) - - template void evalTo(Dest& dst) const - { - eigen_assert(rhs().rows() == dec().rows()); - - // A = U S V^* - // So A^{-1} = V S^{-1} U^* - - Index diagSize = (std::min)(dec().rows(), dec().cols()); - typename JacobiSVDType::SingularValuesType invertedSingVals(diagSize); - - Index nonzeroSingVals = dec().nonzeroSingularValues(); - invertedSingVals.head(nonzeroSingVals) = dec().singularValues().head(nonzeroSingVals).array().inverse(); - invertedSingVals.tail(diagSize - nonzeroSingVals).setZero(); - - dst = dec().matrixV().leftCols(diagSize) - * invertedSingVals.asDiagonal() - * dec().matrixU().leftCols(diagSize).adjoint() - * rhs(); - } -}; -} // end namespace internal - -/** \svd_module - * - * \return the singular value decomposition of \c *this computed by two-sided - * Jacobi transformations. - * - * \sa class JacobiSVD - */ -template -JacobiSVD::PlainObject> -MatrixBase::jacobiSvd(unsigned int computationOptions) const -{ - return JacobiSVD(*this, computationOptions); -} - -} // end namespace Eigen - -#endif // EIGEN_JACOBISVD_H diff --git a/unsupported/test/svd_common.h b/unsupported/test/svd_common.h index b40c23a2b..6a96e7c74 100644 --- a/unsupported/test/svd_common.h +++ b/unsupported/test/svd_common.h @@ -18,7 +18,7 @@ #define EIGEN_RUNTIME_NO_MALLOC #include "main.h" -#include +#include #include From 1f398dfc826c2427375e6aa6a15bfe23383d76f7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 18:31:54 +0200 Subject: [PATCH 0762/1103] Factorize *SVD::solve to SVDBase --- Eigen/src/SVD/JacobiSVD.h | 71 --------------------------------- Eigen/src/SVD/SVDBase.h | 70 ++++++++++++++++++++++++++++++++ unsupported/test/CMakeLists.txt | 2 +- 3 files changed, 71 insertions(+), 72 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 3d778516a..2b9d03c2b 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -592,47 +592,12 @@ template class JacobiSVD return compute(matrix, m_computationOptions); } - /** \returns a (least squares) solution of \f$ A x = b \f$ using the current SVD decomposition of A. - * - * \param b the right-hand-side of the equation to solve. - * - * \note Solving requires both U and V to be computed. Thin U and V are enough, there is no need for full U or V. - * - * \note SVD solving is implicitly least-squares. Thus, this method serves both purposes of exact solving and least-squares solving. - * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$. - */ -#ifdef EIGEN_TEST_EVALUATORS - template - inline const Solve - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); - eigen_assert(computeU() && computeV() && "JacobiSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); - return Solve(*this, b.derived()); - } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); - eigen_assert(computeU() && computeV() && "JacobiSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); - return internal::solve_retval(*this, b.derived()); - } -#endif - using Base::computeU; using Base::computeV; using Base::rows; using Base::cols; using Base::rank; - #ifndef EIGEN_PARSED_BY_DOXYGEN - template - EIGEN_DEVICE_FUNC - void _solve_impl(const RhsType &rhs, DstType &dst) const; - #endif - private: void allocate(Index rows, Index cols, unsigned int computationOptions); @@ -817,42 +782,6 @@ JacobiSVD::compute(const MatrixType& matrix, unsig return *this; } -#ifndef EIGEN_PARSED_BY_DOXYGEN -template -template -void JacobiSVD<_MatrixType,QRPreconditioner>::_solve_impl(const RhsType &rhs, DstType &dst) const -{ - eigen_assert(rhs.rows() == rows()); - - // A = U S V^* - // So A^{-1} = V S^{-1} U^* - - Matrix tmp; - Index l_rank = rank(); - - tmp.noalias() = m_matrixU.leftCols(l_rank).adjoint() * rhs; - tmp = m_singularValues.head(l_rank).asDiagonal().inverse() * tmp; - dst = m_matrixV.leftCols(l_rank) * tmp; -} -#endif - -namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef JacobiSVD<_MatrixType, QRPreconditioner> JacobiSVDType; - EIGEN_MAKE_SOLVE_HELPERS(JacobiSVDType,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(), dst); - } -}; -#endif -} // end namespace internal - #ifndef __CUDACC__ /** \svd_module * diff --git a/Eigen/src/SVD/SVDBase.h b/Eigen/src/SVD/SVDBase.h index 61b01fb8a..a4bb97f4b 100644 --- a/Eigen/src/SVD/SVDBase.h +++ b/Eigen/src/SVD/SVDBase.h @@ -190,6 +190,41 @@ public: inline Index rows() const { return m_rows; } inline Index cols() const { return m_cols; } + + /** \returns a (least squares) solution of \f$ A x = b \f$ using the current SVD decomposition of A. + * + * \param b the right-hand-side of the equation to solve. + * + * \note Solving requires both U and V to be computed. Thin U and V are enough, there is no need for full U or V. + * + * \note SVD solving is implicitly least-squares. Thus, this method serves both purposes of exact solving and least-squares solving. + * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$. + */ +#ifdef EIGEN_TEST_EVALUATORS + template + inline const Solve + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "SVD is not initialized."); + eigen_assert(computeU() && computeV() && "SVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); + return Solve(derived(), b.derived()); + } +#else + template + inline const internal::solve_retval + solve(const MatrixBase& b) const + { + eigen_assert(m_isInitialized && "SVD is not initialized."); + eigen_assert(computeU() && computeV() && "SVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); + return internal::solve_retval(*this, b.derived()); + } +#endif + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + EIGEN_DEVICE_FUNC + void _solve_impl(const RhsType &rhs, DstType &dst) const; + #endif protected: // return true if already allocated @@ -220,6 +255,41 @@ protected: }; +#ifndef EIGEN_PARSED_BY_DOXYGEN +template +template +void SVDBase::_solve_impl(const RhsType &rhs, DstType &dst) const +{ + eigen_assert(rhs.rows() == rows()); + + // A = U S V^* + // So A^{-1} = V S^{-1} U^* + + Matrix tmp; + Index l_rank = rank(); + + tmp.noalias() = m_matrixU.leftCols(l_rank).adjoint() * rhs; + tmp = m_singularValues.head(l_rank).asDiagonal().inverse() * tmp; + dst = m_matrixV.leftCols(l_rank) * tmp; +} +#endif + +namespace internal { +#ifndef EIGEN_TEST_EVALUATORS +template +struct solve_retval, Rhs> + : solve_retval_base, Rhs> +{ + typedef SVDBase SVDType; + EIGEN_MAKE_SOLVE_HELPERS(SVDType,Rhs) + + template void evalTo(Dest& dst) const + { + dec().derived()._solve_impl(rhs(), dst); + } +}; +#endif +} // end namespace internal template bool SVDBase::allocate(Index rows, Index cols, unsigned int computationOptions) diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 5e9e74050..970a05bbd 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -100,8 +100,8 @@ ei_add_test(splines) ei_add_test(gmres) ei_add_test(minres) ei_add_test(levenberg_marquardt) -if(EIGEN_TEST_NO_EVALUATORS) ei_add_test(bdcsvd) +if(EIGEN_TEST_NO_EVALUATORS) ei_add_test(kronecker_product) endif() From 47829e2d16c2bc4fba6aef8fd1b7712a9db839eb Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 1 Sep 2014 18:32:59 +0200 Subject: [PATCH 0763/1103] Disable solve_ret_val like mechanism with evaluator enabled --- Eigen/src/misc/Solve.h | 4 +++- Eigen/src/misc/SparseSolve.h | 4 ++++ 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/Eigen/src/misc/Solve.h b/Eigen/src/misc/Solve.h index 7f70d60af..ebdd981d0 100644 --- a/Eigen/src/misc/Solve.h +++ b/Eigen/src/misc/Solve.h @@ -10,6 +10,8 @@ #ifndef EIGEN_MISC_SOLVE_H #define EIGEN_MISC_SOLVE_H +#ifndef EIGEN_TEST_EVALUATORS + namespace Eigen { namespace internal { @@ -72,5 +74,5 @@ template struct solve_retval_base : Base(dec, rhs) {} } // end namespace Eigen - +#endif // EIGEN_TEST_EVALUATORS #endif // EIGEN_MISC_SOLVE_H diff --git a/Eigen/src/misc/SparseSolve.h b/Eigen/src/misc/SparseSolve.h index 05caa9266..2396dc8e8 100644 --- a/Eigen/src/misc/SparseSolve.h +++ b/Eigen/src/misc/SparseSolve.h @@ -10,6 +10,8 @@ #ifndef EIGEN_SPARSE_SOLVE_H #define EIGEN_SPARSE_SOLVE_H +#ifndef EIGEN_TEST_EVALUATORS + namespace Eigen { namespace internal { @@ -127,4 +129,6 @@ template struct solve_ } // end namespace Eigen +#endif // EIGEN_TEST_EVALUATORS + #endif // EIGEN_SPARSE_SOLVE_H From a96f3d629cfd5e562430f49c9c4d632c365e8020 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 2 Sep 2014 22:30:23 +0200 Subject: [PATCH 0764/1103] Clean bdcsvd --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 73 +++------------------------ 1 file changed, 7 insertions(+), 66 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index a7c369633..829446911 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -57,6 +57,8 @@ class BDCSVD : public SVDBase > public: using Base::rows; using Base::cols; + using Base::computeU; + using Base::computeV; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; @@ -72,15 +74,10 @@ public: MatrixOptions = MatrixType::Options }; - typedef Matrix - MatrixUType; - typedef Matrix - MatrixVType; - typedef typename internal::plain_diag_type::type SingularValuesType; - typedef typename internal::plain_row_type::type RowType; - typedef typename internal::plain_col_type::type ColType; + typedef typename Base::MatrixUType MatrixUType; + typedef typename Base::MatrixVType MatrixVType; + typedef typename Base::SingularValuesType SingularValuesType; + typedef Matrix MatrixX; typedef Matrix MatrixXr; typedef Matrix VectorType; @@ -155,27 +152,6 @@ public: eigen_assert(s>3 && "BDCSVD the size of the algo switch has to be greater than 3"); algoswap = s; } - - - /** \returns a (least squares) solution of \f$ A x = b \f$ using the current SVD decomposition of A. - * - * \param b the right - hand - side of the equation to solve. - * - * \note Solving requires both U and V to be computed. Thin U and V are enough, there is no need for full U or V. - * - * \note SVD solving is implicitly least - squares. Thus, this method serves both purposes of exact solving and least - squares solving. - * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$. - */ - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(this->m_isInitialized && "BDCSVD is not initialized."); - eigen_assert(computeU() && computeV() && - "BDCSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); - return internal::solve_retval(*this, b.derived()); - } - const MatrixUType& matrixU() const { @@ -188,11 +164,9 @@ public: { eigen_assert(this->computeU() && "This SVD decomposition didn't compute U. Did you ask for it?"); return this->m_matrixU; - } - + } } - const MatrixVType& matrixV() const { eigen_assert(this->m_isInitialized && "SVD is not initialized."); @@ -206,9 +180,6 @@ public: return this->m_matrixV; } } - - using Base::computeU; - using Base::computeV; private: void allocate(Index rows, Index cols, unsigned int computationOptions); @@ -898,36 +869,6 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index }//end deflation -namespace internal{ - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef BDCSVD<_MatrixType> BDCSVDType; - EIGEN_MAKE_SOLVE_HELPERS(BDCSVDType, Rhs) - - template void evalTo(Dest& dst) const - { - eigen_assert(rhs().rows() == dec().rows()); - // A = U S V^* - // So A^{ - 1} = V S^{ - 1} U^* - Index diagSize = (std::min)(dec().rows(), dec().cols()); - typename BDCSVDType::SingularValuesType invertedSingVals(diagSize); - Index nonzeroSingVals = dec().nonzeroSingularValues(); - invertedSingVals.head(nonzeroSingVals) = dec().singularValues().head(nonzeroSingVals).array().inverse(); - invertedSingVals.tail(diagSize - nonzeroSingVals).setZero(); - - dst = dec().matrixV().leftCols(diagSize) - * invertedSingVals.asDiagonal() - * dec().matrixU().leftCols(diagSize).adjoint() - * rhs(); - return; - } -}; - -} //end namespace internal - /** \svd_module * * \return the singular value decomposition of \c *this computed by From c82dc227f19e75571ff4d8c47dfbd66765c8dbc5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 3 Sep 2014 10:15:24 +0200 Subject: [PATCH 0765/1103] Cleaning in BDCSVD (formating, handling of transpose case, remove some for loops) --- Eigen/src/SVD/SVDBase.h | 1 - unsupported/Eigen/src/BDCSVD/BDCSVD.h | 441 ++++++++++++-------------- 2 files changed, 202 insertions(+), 240 deletions(-) diff --git a/Eigen/src/SVD/SVDBase.h b/Eigen/src/SVD/SVDBase.h index a4bb97f4b..60a5aabb6 100644 --- a/Eigen/src/SVD/SVDBase.h +++ b/Eigen/src/SVD/SVDBase.h @@ -267,7 +267,6 @@ void SVDBase::_solve_impl(const RhsType &rhs, DstType &dst) const Matrix tmp; Index l_rank = rank(); - tmp.noalias() = m_matrixU.leftCols(l_rank).adjoint() * rhs; tmp = m_singularValues.head(l_rank).asDiagonal().inverse() * tmp; dst = m_matrixV.leftCols(l_rank) * tmp; diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index 829446911..64cee029b 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -19,10 +19,6 @@ #ifndef EIGEN_BDCSVD_H #define EIGEN_BDCSVD_H -#define EPSILON 0.0000000000000001 - -#define ALGOSWAP 16 - namespace Eigen { template class BDCSVD; @@ -88,7 +84,7 @@ public: * The default constructor is useful in cases in which the user intends to * perform decompositions via BDCSVD::compute(const MatrixType&). */ - BDCSVD() : algoswap(ALGOSWAP), m_numIters(0) + BDCSVD() : m_algoswap(16), m_numIters(0) {} @@ -99,7 +95,7 @@ public: * \sa BDCSVD() */ BDCSVD(Index rows, Index cols, unsigned int computationOptions = 0) - : algoswap(ALGOSWAP), m_numIters(0) + : m_algoswap(16), m_numIters(0) { allocate(rows, cols, computationOptions); } @@ -115,7 +111,7 @@ public: * available with the (non - default) FullPivHouseholderQR preconditioner. */ BDCSVD(const MatrixType& matrix, unsigned int computationOptions = 0) - : algoswap(ALGOSWAP), m_numIters(0) + : m_algoswap(16), m_numIters(0) { compute(matrix, computationOptions); } @@ -150,35 +146,7 @@ public: void setSwitchSize(int s) { eigen_assert(s>3 && "BDCSVD the size of the algo switch has to be greater than 3"); - algoswap = s; - } - - const MatrixUType& matrixU() const - { - eigen_assert(this->m_isInitialized && "SVD is not initialized."); - if (isTranspose){ - eigen_assert(this->computeV() && "This SVD decomposition didn't compute U. Did you ask for it?"); - return this->m_matrixV; - } - else - { - eigen_assert(this->computeU() && "This SVD decomposition didn't compute U. Did you ask for it?"); - return this->m_matrixU; - } - } - - const MatrixVType& matrixV() const - { - eigen_assert(this->m_isInitialized && "SVD is not initialized."); - if (isTranspose){ - eigen_assert(this->computeU() && "This SVD decomposition didn't compute V. Did you ask for it?"); - return this->m_matrixU; - } - else - { - eigen_assert(this->computeV() && "This SVD decomposition didn't compute V. Did you ask for it?"); - return this->m_matrixV; - } + m_algoswap = s; } private: @@ -194,15 +162,26 @@ private: void deflation43(Index firstCol, Index shift, Index i, Index size); void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size); void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift); - void copyUV(const typename internal::UpperBidiagonalization::HouseholderUSequenceType& householderU, - const typename internal::UpperBidiagonalization::HouseholderVSequenceType& householderV); + template + void copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naivev); protected: MatrixXr m_naiveU, m_naiveV; MatrixXr m_computed; - Index nRec; - int algoswap; - bool isTranspose, compU, compV; + Index m_nRec; + int m_algoswap; + bool m_isTranspose, m_compU, m_compV; + + using Base::m_singularValues; + using Base::m_diagSize; + using Base::m_computeFullU; + using Base::m_computeFullV; + using Base::m_computeThinU; + using Base::m_computeThinV; + using Base::m_matrixU; + using Base::m_matrixV; + using Base::m_isInitialized; + using Base::m_nonzeroSingularValues; public: int m_numIters; @@ -213,50 +192,35 @@ public: template void BDCSVD::allocate(Index rows, Index cols, unsigned int computationOptions) { - isTranspose = (cols > rows); - if (Base::allocate(rows, cols, computationOptions)) return; - m_computed = MatrixXr::Zero(this->m_diagSize + 1, this->m_diagSize ); - if (isTranspose){ - compU = this->computeU(); - compV = this->computeV(); - } - else - { - compV = this->computeU(); - compU = this->computeV(); - } - if (compU) m_naiveU = MatrixXr::Zero(this->m_diagSize + 1, this->m_diagSize + 1 ); - else m_naiveU = MatrixXr::Zero(2, this->m_diagSize + 1 ); + m_isTranspose = (cols > rows); + if (Base::allocate(rows, cols, computationOptions)) + return; - if (compV) m_naiveV = MatrixXr::Zero(this->m_diagSize, this->m_diagSize); + m_computed = MatrixXr::Zero(m_diagSize + 1, m_diagSize ); + m_compU = computeV(); + m_compV = computeU(); + if (m_isTranspose) + std::swap(m_compU, m_compV); - - //should be changed for a cleaner implementation - if (isTranspose){ - bool aux; - if (this->computeU()||this->computeV()){ - aux = this->m_computeFullU; - this->m_computeFullU = this->m_computeFullV; - this->m_computeFullV = aux; - aux = this->m_computeThinU; - this->m_computeThinU = this->m_computeThinV; - this->m_computeThinV = aux; - } - } + if (m_compU) m_naiveU = MatrixXr::Zero(m_diagSize + 1, m_diagSize + 1 ); + else m_naiveU = MatrixXr::Zero(2, m_diagSize + 1 ); + + if (m_compV) m_naiveV = MatrixXr::Zero(m_diagSize, m_diagSize); }// end allocate // Methode which compute the BDCSVD for the int template<> -BDCSVD >& BDCSVD >::compute(const MatrixType& matrix, unsigned int computationOptions) { +BDCSVD >& BDCSVD >::compute(const MatrixType& matrix, unsigned int computationOptions) +{ allocate(matrix.rows(), matrix.cols(), computationOptions); - this->m_nonzeroSingularValues = 0; + m_nonzeroSingularValues = 0; m_computed = Matrix::Zero(rows(), cols()); - for (int i=0; im_diagSize; i++) { - this->m_singularValues.coeffRef(i) = 0; - } - if (this->m_computeFullU) this->m_matrixU = Matrix::Zero(rows(), rows()); - if (this->m_computeFullV) this->m_matrixV = Matrix::Zero(cols(), cols()); - this->m_isInitialized = true; + + m_singularValues.head(m_diagSize).setZero(); + + if (m_computeFullU) m_matrixU.setZero(rows(), rows()); + if (m_computeFullV) m_matrixV.setZero(cols(), cols()); + m_isInitialized = true; return *this; } @@ -268,59 +232,62 @@ BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsign allocate(matrix.rows(), matrix.cols(), computationOptions); using std::abs; - //**** step 1 Bidiagonalization isTranspose = (matrix.cols()>matrix.rows()) ; + //**** step 1 Bidiagonalization m_isTranspose = (matrix.cols()>matrix.rows()) ; MatrixType copy; - if (isTranspose) copy = matrix.adjoint(); - else copy = matrix; + if (m_isTranspose) copy = matrix.adjoint(); + else copy = matrix; internal::UpperBidiagonalization bid(copy); //**** step 2 Divide - m_computed.topRows(this->m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose(); + m_computed.topRows(m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose(); m_computed.template bottomRows<1>().setZero(); - divide(0, this->m_diagSize - 1, 0, 0, 0); + divide(0, m_diagSize - 1, 0, 0, 0); //**** step 3 copy - for (int i=0; im_diagSize; i++) { + for (int i=0; im_singularValues.coeffRef(i) = a; - if (a == 0){ - this->m_nonzeroSingularValues = i; - this->m_singularValues.tail(this->m_diagSize - i - 1).setZero(); + m_singularValues.coeffRef(i) = a; + if (a == 0) + { + m_nonzeroSingularValues = i; + m_singularValues.tail(m_diagSize - i - 1).setZero(); break; } - else if (i == this->m_diagSize - 1) + else if (i == m_diagSize - 1) { - this->m_nonzeroSingularValues = i + 1; + m_nonzeroSingularValues = i + 1; break; } } - copyUV(bid.householderU(), bid.householderV()); - this->m_isInitialized = true; + if(m_isTranspose) copyUV(bid.householderV(), bid.householderU(), m_naiveV, m_naiveU); + else copyUV(bid.householderU(), bid.householderV(), m_naiveU, m_naiveV); + m_isInitialized = true; return *this; }// end compute template -void BDCSVD::copyUV(const typename internal::UpperBidiagonalization::HouseholderUSequenceType& householderU, - const typename internal::UpperBidiagonalization::HouseholderVSequenceType& householderV) +template +void BDCSVD::copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naiveV) { // Note exchange of U and V: m_matrixU is set from m_naiveV and vice versa - if (this->computeU()){ - Index Ucols = this->m_computeThinU ? this->m_nonzeroSingularValues : householderU.cols(); - this->m_matrixU = MatrixX::Identity(householderU.cols(), Ucols); - Index blockCols = this->m_computeThinU ? this->m_nonzeroSingularValues : this->m_diagSize; - this->m_matrixU.block(0, 0, this->m_diagSize, blockCols) = - m_naiveV.template cast().block(0, 0, this->m_diagSize, blockCols); - this->m_matrixU = householderU * this->m_matrixU; + if (computeU()) + { + Index Ucols = m_computeThinU ? m_nonzeroSingularValues : householderU.cols(); + m_matrixU = MatrixX::Identity(householderU.cols(), Ucols); + Index blockCols = m_computeThinU ? m_nonzeroSingularValues : m_diagSize; + m_matrixU.topLeftCorner(m_diagSize, blockCols) = naiveV.template cast().topLeftCorner(m_diagSize, blockCols); + m_matrixU = householderU * m_matrixU; } - if (this->computeV()){ - Index Vcols = this->m_computeThinV ? this->m_nonzeroSingularValues : householderV.cols(); - this->m_matrixV = MatrixX::Identity(householderV.cols(), Vcols); - Index blockCols = this->m_computeThinV ? this->m_nonzeroSingularValues : this->m_diagSize; - this->m_matrixV.block(0, 0, this->m_diagSize, blockCols) = - m_naiveU.template cast().block(0, 0, this->m_diagSize, blockCols); - this->m_matrixV = householderV * this->m_matrixV; + if (computeV()) + { + Index Vcols = m_computeThinV ? m_nonzeroSingularValues : householderV.cols(); + m_matrixV = MatrixX::Identity(householderV.cols(), Vcols); + Index blockCols = m_computeThinV ? m_nonzeroSingularValues : m_diagSize; + m_matrixV.topLeftCorner(m_diagSize, blockCols) = naiveU.template cast().topLeftCorner(m_diagSize, blockCols); + m_matrixV = householderV * m_matrixV; } } @@ -335,8 +302,7 @@ void BDCSVD::copyUV(const typename internal::UpperBidiagonalization< //@param shift : Each time one takes the left submatrix, one must add 1 to the shift. Why? Because! We actually want the last column of the U submatrix // to become the first column (*coeff) and to shift all the other columns to the right. There are more details on the reference paper. template -void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, - Index firstColW, Index shift) +void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift) { // requires nbRows = nbCols + 1; using std::pow; @@ -351,21 +317,19 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, MatrixXr l, f; // We use the other algorithm which is more efficient for small // matrices. - if (n < algoswap){ - JacobiSVD b(m_computed.block(firstCol, firstCol, n + 1, n), - ComputeFullU | (ComputeFullV * compV)) ; - if (compU) m_naiveU.block(firstCol, firstCol, n + 1, n + 1).real() << b.matrixU(); + if (n < m_algoswap) + { + JacobiSVD b(m_computed.block(firstCol, firstCol, n + 1, n), ComputeFullU | (m_compV ? ComputeFullV : 0)) ; + if (m_compU) + m_naiveU.block(firstCol, firstCol, n + 1, n + 1).real() = b.matrixU(); else { - m_naiveU.row(0).segment(firstCol, n + 1).real() << b.matrixU().row(0); - m_naiveU.row(1).segment(firstCol, n + 1).real() << b.matrixU().row(n); + m_naiveU.row(0).segment(firstCol, n + 1).real() = b.matrixU().row(0); + m_naiveU.row(1).segment(firstCol, n + 1).real() = b.matrixU().row(n); } - if (compV) m_naiveV.block(firstRowW, firstColW, n, n).real() << b.matrixV(); + if (m_compV) m_naiveV.block(firstRowW, firstColW, n, n).real() = b.matrixV(); m_computed.block(firstCol + shift, firstCol + shift, n + 1, n).setZero(); - for (int i=0; i::divide (Index firstCol, Index lastCol, Index firstRowW, // right submatrix before the left one. divide(k + 1 + firstCol, lastCol, k + 1 + firstRowW, k + 1 + firstColW, shift); divide(firstCol, k - 1 + firstCol, firstRowW, firstColW + 1, shift + 1); - if (compU) + if (m_compU) { lambda = m_naiveU(firstCol + k, firstCol + k); phi = m_naiveU(firstCol + k + 1, lastCol + 1); @@ -386,9 +350,8 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, lambda = m_naiveU(1, firstCol + k); phi = m_naiveU(0, lastCol + 1); } - r0 = sqrt((abs(alphaK * lambda) * abs(alphaK * lambda)) - + abs(betaK * phi) * abs(betaK * phi)); - if (compU) + r0 = sqrt((abs(alphaK * lambda) * abs(alphaK * lambda)) + abs(betaK * phi) * abs(betaK * phi)); + if (m_compU) { l = m_naiveU.row(firstCol + k).segment(firstCol, k); f = m_naiveU.row(firstCol + k + 1).segment(firstCol + k + 1, n - k - 1); @@ -398,7 +361,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, l = m_naiveU.row(1).segment(firstCol, k); f = m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1); } - if (compV) m_naiveV(firstRowW+k, firstColW) = 1; + if (m_compV) m_naiveV(firstRowW+k, firstColW) = 1; if (r0 == 0) { c0 = 1; @@ -409,21 +372,18 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, c0 = alphaK * lambda / r0; s0 = betaK * phi / r0; } - if (compU) + if (m_compU) { MatrixXr q1 (m_naiveU.col(firstCol + k).segment(firstCol, k + 1)); // we shiftW Q1 to the right for (Index i = firstCol + k - 1; i >= firstCol; i--) - { - m_naiveU.col(i + 1).segment(firstCol, k + 1) << m_naiveU.col(i).segment(firstCol, k + 1); - } + m_naiveU.col(i + 1).segment(firstCol, k + 1) = m_naiveU.col(i).segment(firstCol, k + 1); // we shift q1 at the left with a factor c0 - m_naiveU.col(firstCol).segment( firstCol, k + 1) << (q1 * c0); + m_naiveU.col(firstCol).segment( firstCol, k + 1) = (q1 * c0); // last column = q1 * - s0 - m_naiveU.col(lastCol + 1).segment(firstCol, k + 1) << (q1 * ( - s0)); + m_naiveU.col(lastCol + 1).segment(firstCol, k + 1) = (q1 * ( - s0)); // first column = q2 * s0 - m_naiveU.col(firstCol).segment(firstCol + k + 1, n - k) << - m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) *s0; + m_naiveU.col(firstCol).segment(firstCol + k + 1, n - k) = m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) * s0; // q2 *= c0 m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) *= c0; } @@ -432,9 +392,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, RealScalar q1 = (m_naiveU(0, firstCol + k)); // we shift Q1 to the right for (Index i = firstCol + k - 1; i >= firstCol; i--) - { m_naiveU(0, i + 1) = m_naiveU(0, i); - } // we shift q1 at the left with a factor c0 m_naiveU(0, firstCol) = (q1 * c0); // last column = q1 * - s0 @@ -447,8 +405,8 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1).setZero(); } m_computed(firstCol + shift, firstCol + shift) = r0; - m_computed.col(firstCol + shift).segment(firstCol + shift + 1, k) << alphaK * l.transpose().real(); - m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) << betaK * f.transpose().real(); + m_computed.col(firstCol + shift).segment(firstCol + shift + 1, k) = alphaK * l.transpose().real(); + m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) = betaK * f.transpose().real(); // Second part: try to deflate singular values in combined matrix @@ -458,9 +416,9 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, MatrixXr UofSVD, VofSVD; VectorType singVals; computeSVDofM(firstCol + shift, n, UofSVD, singVals, VofSVD); - if (compU) m_naiveU.block(firstCol, firstCol, n + 1, n + 1) *= UofSVD; - else m_naiveU.block(0, firstCol, 2, n + 1) *= UofSVD; - if (compV) m_naiveV.block(firstRowW, firstColW, n, n) *= VofSVD; + if (m_compU) m_naiveU.block(firstCol, firstCol, n + 1, n + 1) *= UofSVD; // FIXME this requires a temporary + else m_naiveU.block(0, firstCol, 2, n + 1) *= UofSVD; // FIXME this requires a temporary, and exploit that there are 2 rows at compile time + if (m_compV) m_naiveV.block(firstRowW, firstColW, n, n) *= VofSVD; // FIXME this requires a temporary m_computed.block(firstCol + shift, firstCol + shift, n, n).setZero(); m_computed.block(firstCol + shift, firstCol + shift, n, n).diagonal() = singVals; }// end divide @@ -468,7 +426,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, // Compute SVD of m_computed.block(firstCol, firstCol, n + 1, n); this block only has non-zeros in // the first column and on the diagonal and has undergone deflation, so diagonal is in increasing // order except for possibly the (0,0) entry. The computed SVD is stored U, singVals and V, except -// that if compV is false, then V is not computed. Singular values are sorted in decreasing order. +// that if m_compV is false, then V is not computed. Singular values are sorted in decreasing order. // // TODO Opportunities for optimization: better root finding algo, better stopping criterion, better // handling of round-off errors, be consistent in ordering @@ -483,7 +441,7 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec // compute singular values and vectors (in decreasing order) singVals.resize(n); U.resize(n+1, n+1); - if (compV) V.resize(n, n); + if (m_compV) V.resize(n, n); if (col0.hasNaN() || diag.hasNaN()) return; @@ -495,7 +453,7 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec // Reverse order so that singular values in increased order singVals.reverseInPlace(); U.leftCols(n) = U.leftCols(n).rowwise().reverse().eval(); - if (compV) V = V.rowwise().reverse().eval(); + if (m_compV) V = V.rowwise().reverse().eval(); } template @@ -504,10 +462,13 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia { using std::abs; using std::swap; + using std::max; Index n = col0.size(); - for (Index k = 0; k < n; ++k) { - if (col0(k) == 0) { + for (Index k = 0; k < n; ++k) + { + if (col0(k) == 0) + { // entry is deflated, so singular value is on diagonal singVals(k) = diag(k); mus(k) = 0; @@ -523,27 +484,29 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia RealScalar mid = left + (right-left) / 2; RealScalar fMid = 1 + (col0.square() / ((diag + mid) * (diag - mid))).sum(); - RealScalar shift; - if (k == n-1 || fMid > 0) shift = left; - else shift = right; + RealScalar shift = (k == n-1 || fMid > 0) ? left : right; // measure everything relative to shift ArrayXr diagShifted = diag - shift; // initial guess RealScalar muPrev, muCur; - if (shift == left) { + if (shift == left) + { muPrev = (right - left) * 0.1; if (k == n-1) muCur = right - left; - else muCur = (right - left) * 0.5; - } else { + else muCur = (right - left) * 0.5; + } + else + { muPrev = -(right - left) * 0.1; muCur = -(right - left) * 0.5; } RealScalar fPrev = 1 + (col0.square() / ((diagShifted - muPrev) * (diag + shift + muPrev))).sum(); RealScalar fCur = 1 + (col0.square() / ((diagShifted - muCur) * (diag + shift + muCur))).sum(); - if (abs(fPrev) < abs(fCur)) { + if (abs(fPrev) < abs(fCur)) + { swap(fPrev, fCur); swap(muPrev, muCur); } @@ -551,7 +514,8 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // rational interpolation: fit a function of the form a / mu + b through the two previous // iterates and use its zero to compute the next iterate bool useBisection = false; - while (abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (std::max)(abs(muCur), abs(muPrev)) && fCur != fPrev && !useBisection) { + while (abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (max)(abs(muCur), abs(muPrev)) && fCur != fPrev && !useBisection) + { ++m_numIters; RealScalar a = (fCur - fPrev) / (1/muCur - 1/muPrev); @@ -567,13 +531,17 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia } // fall back on bisection method if rational interpolation did not work - if (useBisection) { + if (useBisection) + { RealScalar leftShifted, rightShifted; - if (shift == left) { + if (shift == left) + { leftShifted = 1e-30; if (k == 0) rightShifted = right - left; - else rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe - } else { + else rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe + } + else + { leftShifted = -(right - left) * 0.6; rightShifted = -1e-30; } @@ -582,13 +550,17 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia RealScalar fRight = 1 + (col0.square() / ((diagShifted - rightShifted) * (diag + shift + rightShifted))).sum(); assert(fLeft * fRight < 0); - while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (std::max)(abs(leftShifted), abs(rightShifted))) { + while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (max)(abs(leftShifted), abs(rightShifted))) + { RealScalar midShifted = (leftShifted + rightShifted) / 2; RealScalar fMid = 1 + (col0.square() / ((diagShifted - midShifted) * (diag + shift + midShifted))).sum(); - if (fLeft * fMid < 0) { + if (fLeft * fMid < 0) + { rightShifted = midShifted; fRight = fMid; - } else { + } + else + { leftShifted = midShifted; fLeft = fMid; } @@ -615,13 +587,15 @@ void BDCSVD::perturbCol0 (const ArrayXr& col0, const ArrayXr& diag, const VectorType& singVals, const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat) { + using std::sqrt; Index n = col0.size(); - for (Index k = 0; k < n; ++k) { + for (Index k = 0; k < n; ++k) + { if (col0(k) == 0) zhat(k) = 0; - else { + else + { // see equation (3.6) - using std::sqrt; RealScalar tmp = sqrt( (singVals(n-1) + diag(k)) * (mus(n-1) + (shifts(n-1) - diag(k))) @@ -647,16 +621,21 @@ void BDCSVD::computeSingVecs const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V) { Index n = zhat.size(); - for (Index k = 0; k < n; ++k) { - if (zhat(k) == 0) { + for (Index k = 0; k < n; ++k) + { + if (zhat(k) == 0) + { U.col(k) = VectorType::Unit(n+1, k); - if (compV) V.col(k) = VectorType::Unit(n, k); - } else { + if (m_compV) V.col(k) = VectorType::Unit(n, k); + } + else + { U.col(k).head(n) = zhat / (((diag - shifts(k)) - mus(k)) * (diag + singVals[k])); U(n,k) = 0; U.col(k).normalize(); - if (compV) { + if (m_compV) + { V.col(k).tail(n-1) = (diag * zhat / (((diag - shifts(k)) - mus(k)) * (diag + singVals[k]))).tail(n-1); V(0,k) = -1; V.col(k).normalize(); @@ -671,15 +650,17 @@ void BDCSVD::computeSingVecs // i >= 1, di almost null and zi non null. // We use a rotation to zero out zi applied to the left of M template -void BDCSVD::deflation43(Index firstCol, Index shift, Index i, Index size){ +void BDCSVD::deflation43(Index firstCol, Index shift, Index i, Index size) +{ using std::abs; using std::sqrt; using std::pow; RealScalar c = m_computed(firstCol + shift, firstCol + shift); RealScalar s = m_computed(i, firstCol + shift); RealScalar r = sqrt(pow(abs(c), 2) + pow(abs(s), 2)); - if (r == 0){ - m_computed(i, i)=0; + if (r == 0) + { + m_computed(i, i) = 0; return; } c/=r; @@ -687,7 +668,8 @@ void BDCSVD::deflation43(Index firstCol, Index shift, Index i, Index m_computed(firstCol + shift, firstCol + shift) = r; m_computed(i, firstCol + shift) = 0; m_computed(i, i) = 0; - if (compU){ + if (m_compU) + { m_naiveU.col(firstCol).segment(firstCol,size) = c * m_naiveU.col(firstCol).segment(firstCol, size) - s * m_naiveU.col(i).segment(firstCol, size) ; @@ -703,7 +685,8 @@ void BDCSVD::deflation43(Index firstCol, Index shift, Index i, Index // i,j >= 1, i != j and |di - dj| < epsilon * norm2(M) // We apply two rotations to have zj = 0; template -void BDCSVD::deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size){ +void BDCSVD::deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size) +{ using std::abs; using std::sqrt; using std::conj; @@ -711,7 +694,8 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi RealScalar c = m_computed(firstColm, firstColm + j - 1); RealScalar s = m_computed(firstColm, firstColm + i - 1); RealScalar r = sqrt(pow(abs(c), 2) + pow(abs(s), 2)); - if (r==0){ + if (r==0) + { m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j); return; } @@ -720,7 +704,8 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi m_computed(firstColm + i, firstColm) = r; m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j); m_computed(firstColm + j, firstColm) = 0; - if (compU){ + if (m_compU) + { m_naiveU.col(firstColu + i).segment(firstColu, size) = c * m_naiveU.col(firstColu + i).segment(firstColu, size) - s * m_naiveU.col(firstColu + j).segment(firstColu, size) ; @@ -729,7 +714,8 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi (c + s*s/c) * m_naiveU.col(firstColu + j).segment(firstColu, size) + (s/c) * m_naiveU.col(firstColu + i).segment(firstColu, size); } - if (compV){ + if (m_compV) + { m_naiveV.col(firstColW + i).segment(firstRowW, size - 1) = c * m_naiveV.col(firstColW + i).segment(firstRowW, size - 1) + s * m_naiveV.col(firstColW + j).segment(firstRowW, size - 1) ; @@ -743,72 +729,56 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi // acts on block from (firstCol+shift, firstCol+shift) to (lastCol+shift, lastCol+shift) [inclusive] template -void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift){ +void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift) +{ //condition 4.1 using std::sqrt; + using std::abs; const Index length = lastCol + 1 - firstCol; RealScalar norm1 = m_computed.block(firstCol+shift, firstCol+shift, length, 1).squaredNorm(); RealScalar norm2 = m_computed.block(firstCol+shift, firstCol+shift, length, length).diagonal().squaredNorm(); - RealScalar EPS = 10 * NumTraits::epsilon() * sqrt(norm1 + norm2); - if (m_computed(firstCol + shift, firstCol + shift) < EPS){ - m_computed(firstCol + shift, firstCol + shift) = EPS; - } + RealScalar epsilon = 10 * NumTraits::epsilon() * sqrt(norm1 + norm2); + if (m_computed(firstCol + shift, firstCol + shift) < epsilon) + m_computed(firstCol + shift, firstCol + shift) = epsilon; //condition 4.2 - for (Index i=firstCol + shift + 1;i<=lastCol + shift;i++){ - if (std::abs(m_computed(i, firstCol + shift)) < EPS){ + for (Index i=firstCol + shift + 1;i<=lastCol + shift;i++) + if (abs(m_computed(i, firstCol + shift)) < epsilon) m_computed(i, firstCol + shift) = 0; - } - } //condition 4.3 - for (Index i=firstCol + shift + 1;i<=lastCol + shift; i++){ - if (m_computed(i, i) < EPS){ + for (Index i=firstCol + shift + 1;i<=lastCol + shift; i++) + if (m_computed(i, i) < epsilon) deflation43(firstCol, shift, i, length); - } - } //condition 4.4 Index i=firstCol + shift + 1, j=firstCol + shift + k + 1; //we stock the final place of each line - Index *permutation = new Index[length]; + Index *permutation = new Index[length]; // FIXME avoid repeated dynamic memory allocation - for (Index p =1; p < length; p++) { - if (i> firstCol + shift + k){ - permutation[p] = j; - j++; - } else if (j> lastCol + shift) - { - permutation[p] = i; - i++; - } - else - { - if (m_computed(i, i) < m_computed(j, j)){ - permutation[p] = j; - j++; - } - else - { - permutation[p] = i; - i++; - } - } + for (Index p =1; p < length; p++) + { + if (i> firstCol + shift + k) permutation[p] = j++; + else if (j> lastCol + shift) permutation[p] = i++; + else if (m_computed(i, i) < m_computed(j, j)) permutation[p] = j++; + else permutation[p] = i++; } //we do the permutation RealScalar aux; //we stock the current index of each col //and the column of each index - Index *realInd = new Index[length]; - Index *realCol = new Index[length]; - for (int pos = 0; pos< length; pos++){ + Index *realInd = new Index[length]; // FIXME avoid repeated dynamic memory allocation + Index *realCol = new Index[length]; // FIXME avoid repeated dynamic memory allocation + for (int pos = 0; pos< length; pos++) + { realCol[pos] = pos + firstCol + shift; realInd[pos] = pos; } const Index Zero = firstCol + shift; VectorType temp; - for (int i = 1; i < length - 1; i++){ + for (int i = 1; i < length - 1; i++) + { const Index I = i + Zero; const Index realI = realInd[i]; const Index j = permutation[length - i] - Zero; @@ -825,25 +795,25 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index m_computed(J, Zero) = aux; // change columns - if (compU) { + if (m_compU) + { temp = m_naiveU.col(I - shift).segment(firstCol, length + 1); - m_naiveU.col(I - shift).segment(firstCol, length + 1) << - m_naiveU.col(J - shift).segment(firstCol, length + 1); - m_naiveU.col(J - shift).segment(firstCol, length + 1) << temp; + m_naiveU.col(I - shift).segment(firstCol, length + 1) = m_naiveU.col(J - shift).segment(firstCol, length + 1); + m_naiveU.col(J - shift).segment(firstCol, length + 1) = temp; } else { temp = m_naiveU.col(I - shift).segment(0, 2); - m_naiveU.col(I - shift).segment(0, 2) << - m_naiveU.col(J - shift).segment(0, 2); - m_naiveU.col(J - shift).segment(0, 2) << temp; + m_naiveU.col(I - shift).template head<2>() = m_naiveU.col(J - shift).segment(0, 2); + m_naiveU.col(J - shift).template head<2>() = temp; } - if (compV) { + if (m_compV) + { const Index CWI = I + firstColW - Zero; const Index CWJ = J + firstColW - Zero; temp = m_naiveV.col(CWI).segment(firstRowW, length); - m_naiveV.col(CWI).segment(firstRowW, length) << m_naiveV.col(CWJ).segment(firstRowW, length); - m_naiveV.col(CWJ).segment(firstRowW, length) << temp; + m_naiveV.col(CWI).segment(firstRowW, length) = m_naiveV.col(CWJ).segment(firstRowW, length); + m_naiveV.col(CWJ).segment(firstRowW, length) = temp; } //update real pos @@ -852,20 +822,13 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index realInd[J - Zero] = realI; realInd[I - Zero] = j; } - for (Index i = firstCol + shift + 1; i Date: Wed, 3 Sep 2014 11:38:13 -0700 Subject: [PATCH 0766/1103] Improved the performance of the tensor convolution code by a factor of about 4. --- .../CXX11/src/Tensor/TensorConvolution.h | 147 +++++++++++++----- .../Eigen/CXX11/src/Tensor/TensorEvalTo.h | 7 +- 2 files changed, 111 insertions(+), 43 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 7d0a21c3b..4a5fd9c79 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -226,22 +226,18 @@ struct TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, - PacketAccess = /*TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess */ - false, + PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) - : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device) + : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_kernel(NULL), m_kernelArg(op.kernelExpression()), m_local_kernel(false), m_device(device) { const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { - m_inputStride[i] = m_inputStride[i-1] * input_dims[i-1]; - } else { - m_inputStride[0] = 1; - } + m_inputStride[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_inputStride[i] = m_inputStride[i-1] * input_dims[i-1]; } m_dimensions = m_inputImpl.dimensions(); @@ -251,7 +247,6 @@ struct TensorEvaluator 0) { m_kernelStride[i] = m_kernelStride[i-1] * kernel_dims[i-1]; } else { @@ -260,16 +255,12 @@ struct TensorEvaluator 0) { - m_outputStride[i] = m_outputStride[i-1] * m_dimensions[i-1]; - } else { - m_outputStride[0] = 1; - } + m_outputStride[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_outputStride[i] = m_outputStride[i-1] * m_dimensions[i-1]; } } - typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; @@ -278,57 +269,126 @@ struct TensorEvaluator= 0; --i) { - const Index idx = index / m_outputStride[i]; - startInput += idx * m_inputStride[i]; - index -= idx * m_outputStride[i]; - } - CoeffReturnType result = CoeffReturnType(0); - convolve(startInput, 0, 0, result); + convolve(firstInput(index), 0, NumKernelDims-1, result); return result; } - /* TODO: vectorization template - EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const + EIGEN_DEVICE_FUNC PacketReturnType packet(const Index index) const { - assert(false); - }*/ + const int PacketSize = internal::unpacket_traits::size; + Index indices[2] = {index, index+PacketSize-1}; + Index startInputs[2] = {0, 0}; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / m_outputStride[i]; + const Index idx1 = indices[1] / m_outputStride[i]; + startInputs[0] += idx0 * m_inputStride[i]; + startInputs[1] += idx1 * m_inputStride[i]; + indices[0] -= idx0 * m_outputStride[i]; + indices[1] -= idx1 * m_outputStride[i]; + } + startInputs[0] += indices[0]; + startInputs[1] += indices[1]; - EIGEN_DEVICE_FUNC void convolve(Index firstIndex, Index firstKernel, int DimIndex, CoeffReturnType& accum) const { - for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) { - const Index input = firstIndex + j * m_indexStride[DimIndex]; - const Index kernel = firstKernel + j * m_kernelStride[DimIndex]; - if (DimIndex < NumKernelDims-1) { - convolve(input, kernel, DimIndex+1, accum); - } else { - - accum += m_inputImpl.coeff(input) * m_kernelImpl.coeff(kernel); + if (startInputs[1]-startInputs[0] == PacketSize-1) { + PacketReturnType result = internal::pset1(0); + convolvePacket(startInputs[0], 0, NumKernelDims-1, result); + return result; + } else { + EIGEN_ALIGN_DEFAULT Scalar data[PacketSize]; + data[0] = Scalar(0); + convolve(startInputs[0], 0, NumKernelDims-1, data[0]); + for (int i = 1; i < PacketSize-1; ++i) { + data[i] = Scalar(0); + convolve(firstInput(index+i), 0, NumKernelDims-1, data[i]); } + data[PacketSize-1] = Scalar(0); + convolve(startInputs[1], 0, NumKernelDims-1, data[PacketSize-1]); + return internal::pload(data); } } Scalar* data() const { return NULL; } private: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const { + Index startInput = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStride[i]; + startInput += idx * m_inputStride[i]; + index -= idx * m_outputStride[i]; + } + startInput += index; + return startInput; + } + + EIGEN_DEVICE_FUNC void convolve(Index firstIndex, Index firstKernel, int DimIndex, CoeffReturnType& accum) const { + for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) { + const Index input = firstIndex + j * m_indexStride[DimIndex]; + const Index kernel = firstKernel + j * m_kernelStride[DimIndex]; + if (DimIndex > 0) { + convolve(input, kernel, DimIndex-1, accum); + } else { + accum += m_inputImpl.coeff(input) * m_kernel[kernel]; + } + } + } + + template + EIGEN_DEVICE_FUNC void convolvePacket(Index firstIndex, Index firstKernel, int DimIndex, Packet& accum) const { + for (int j = 0; j < m_kernelImpl.dimensions()[DimIndex]; ++j) { + const Index input = firstIndex + j * m_indexStride[DimIndex]; + const Index kernel = firstKernel + j * m_kernelStride[DimIndex]; + if (DimIndex > 0) { + convolvePacket(input, kernel, DimIndex-1, accum); + } else { + accum = internal::pmadd(m_inputImpl.template packet(input), internal::pset1(m_kernel[kernel]), accum); + } + } + } + + EIGEN_STRONG_INLINE void preloadKernel() { + // Don't make a local copy of the kernel unless we have to (i.e. it's an + // expression that needs to be evaluated) + const Scalar* in_place = m_kernelImpl.data(); + if (in_place) { + m_kernel = in_place; + m_local_kernel = false; + } else { + size_t kernel_sz = m_kernelImpl.dimensions().TotalSize() * sizeof(Scalar); + Scalar* local = (Scalar*)m_device.allocate(kernel_sz); + typedef TensorEvalToOp EvalTo; + EvalTo evalToTmp(local, m_kernelArg); + internal::TensorExecutor::PacketAccess>::run(evalToTmp, m_device); + + m_kernel = local; + m_local_kernel = true; + } + } + // No copy, no assignment TensorEvaluator(const TensorEvaluator&); TensorEvaluator& operator = (const TensorEvaluator&); @@ -341,6 +401,11 @@ struct TensorEvaluator m_inputImpl; TensorEvaluator m_kernelImpl; Dimensions m_dimensions; + + KernelArgType m_kernelArg; + const Scalar* m_kernel; + bool m_local_kernel; + const Device& m_device; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h index db716a80e..587cbd5ca 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h @@ -108,8 +108,9 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_impl.evalSubExprsIfNeeded(); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalScalar(Index i) { @@ -134,6 +135,8 @@ struct TensorEvaluator, Device> return internal::ploadt(m_buffer + index); } + Scalar* data() const { return NULL; } + private: TensorEvaluator m_impl; const Device& m_device; From 80993b95d3be6dbc11b149b253f72d22f41c586e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 3 Sep 2014 22:56:39 +0200 Subject: [PATCH 0767/1103] Disable a test which had never worked without evalautors --- test/sparse_vector.cpp | 3 +++ 1 file changed, 3 insertions(+) diff --git a/test/sparse_vector.cpp b/test/sparse_vector.cpp index 5eea9edfd..6cd5a9a8c 100644 --- a/test/sparse_vector.cpp +++ b/test/sparse_vector.cpp @@ -71,7 +71,10 @@ template void sparse_vector(int rows, int cols) VERIFY_IS_APPROX(v1.dot(v2), refV1.dot(refV2)); VERIFY_IS_APPROX(v1.dot(refV2), refV1.dot(refV2)); +#ifdef EIGEN_TEST_EVALUATORS + // the following did not compiled without evaluators VERIFY_IS_APPROX(m1*v2, refM1*refV2); +#endif VERIFY_IS_APPROX(v1.dot(m1*v2), refV1.dot(refM1*refV2)); int i = internal::random(0,rows-1); VERIFY_IS_APPROX(v1.dot(m1.col(i)), refV1.dot(refM1.col(i))); From 8846aa6d1b68fb0951c70bc339af93b418117ae2 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 4 Sep 2014 09:15:59 +0200 Subject: [PATCH 0768/1103] Optimization: enable cache-efficient application of HouseholderSequence. --- Eigen/src/Householder/BlockHouseholder.h | 73 ++++++++++++++++----- Eigen/src/Householder/HouseholderSequence.h | 34 ++++++++-- Eigen/src/QR/HouseholderQR.h | 6 +- Eigen/src/SVD/UpperBidiagonalization.h | 4 +- 4 files changed, 89 insertions(+), 28 deletions(-) diff --git a/Eigen/src/Householder/BlockHouseholder.h b/Eigen/src/Householder/BlockHouseholder.h index 60dbea5f5..35dbf80a1 100644 --- a/Eigen/src/Householder/BlockHouseholder.h +++ b/Eigen/src/Householder/BlockHouseholder.h @@ -16,48 +16,85 @@ namespace Eigen { namespace internal { + +/** \internal */ +// template +// void make_block_householder_triangular_factor(TriangularFactorType& triFactor, const VectorsType& vectors, const CoeffsType& hCoeffs) +// { +// typedef typename TriangularFactorType::Index Index; +// typedef typename VectorsType::Scalar Scalar; +// const Index nbVecs = vectors.cols(); +// eigen_assert(triFactor.rows() == nbVecs && triFactor.cols() == nbVecs && vectors.rows()>=nbVecs); +// +// for(Index i = 0; i < nbVecs; i++) +// { +// Index rs = vectors.rows() - i; +// // Warning, note that hCoeffs may alias with vectors. +// // It is then necessary to copy it before modifying vectors(i,i). +// typename CoeffsType::Scalar h = hCoeffs(i); +// // This hack permits to pass trough nested Block<> and Transpose<> expressions. +// Scalar *Vii_ptr = const_cast(vectors.data() + vectors.outerStride()*i + vectors.innerStride()*i); +// Scalar Vii = *Vii_ptr; +// *Vii_ptr = Scalar(1); +// triFactor.col(i).head(i).noalias() = -h * vectors.block(i, 0, rs, i).adjoint() +// * vectors.col(i).tail(rs); +// *Vii_ptr = Vii; +// // FIXME add .noalias() once the triangular product can work inplace +// triFactor.col(i).head(i) = triFactor.block(0,0,i,i).template triangularView() +// * triFactor.col(i).head(i); +// triFactor(i,i) = hCoeffs(i); +// } +// } /** \internal */ +// This variant avoid modifications in vectors template void make_block_householder_triangular_factor(TriangularFactorType& triFactor, const VectorsType& vectors, const CoeffsType& hCoeffs) { typedef typename TriangularFactorType::Index Index; - typedef typename VectorsType::Scalar Scalar; const Index nbVecs = vectors.cols(); eigen_assert(triFactor.rows() == nbVecs && triFactor.cols() == nbVecs && vectors.rows()>=nbVecs); - for(Index i = 0; i < nbVecs; i++) + for(Index i = nbVecs-1; i >=0 ; --i) { - Index rs = vectors.rows() - i; - Scalar Vii = vectors(i,i); - vectors.const_cast_derived().coeffRef(i,i) = Scalar(1); - triFactor.col(i).head(i).noalias() = -hCoeffs(i) * vectors.block(i, 0, rs, i).adjoint() - * vectors.col(i).tail(rs); - vectors.const_cast_derived().coeffRef(i, i) = Vii; - // FIXME add .noalias() once the triangular product can work inplace - triFactor.col(i).head(i) = triFactor.block(0,0,i,i).template triangularView() - * triFactor.col(i).head(i); + Index rs = vectors.rows() - i - 1; + Index rt = nbVecs-i-1; + + if(rt>0) + { + triFactor.row(i).tail(rt).noalias() = -hCoeffs(i) * vectors.col(i).tail(rs).adjoint() + * vectors.bottomRightCorner(rs, rt).template triangularView(); + + // FIXME add .noalias() once the triangular product can work inplace + triFactor.row(i).tail(rt) = triFactor.row(i).tail(rt) * triFactor.bottomRightCorner(rt,rt).template triangularView(); + + } triFactor(i,i) = hCoeffs(i); } } -/** \internal */ +/** \internal + * if forward then perform mat = H0 * H1 * H2 * mat + * otherwise perform mat = H2 * H1 * H0 * mat + */ template -void apply_block_householder_on_the_left(MatrixType& mat, const VectorsType& vectors, const CoeffsType& hCoeffs) +void apply_block_householder_on_the_left(MatrixType& mat, const VectorsType& vectors, const CoeffsType& hCoeffs, bool forward) { typedef typename MatrixType::Index Index; enum { TFactorSize = MatrixType::ColsAtCompileTime }; Index nbVecs = vectors.cols(); - Matrix T(nbVecs,nbVecs); - make_block_householder_triangular_factor(T, vectors, hCoeffs); - - const TriangularView& V(vectors); + Matrix T(nbVecs,nbVecs); + + if(forward) make_block_householder_triangular_factor(T, vectors, hCoeffs); + else make_block_householder_triangular_factor(T, vectors, hCoeffs.conjugate()); + const TriangularView V(vectors); // A -= V T V^* A Matrix tmp = V.adjoint() * mat; // FIXME add .noalias() once the triangular product can work inplace - tmp = T.template triangularView().adjoint() * tmp; + if(forward) tmp = T.template triangularView() * tmp; + else tmp = T.template triangularView().adjoint() * tmp; mat.noalias() -= V * tmp; } diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h index 99d3eb21f..111936f0e 100644 --- a/Eigen/src/Householder/HouseholderSequence.h +++ b/Eigen/src/Householder/HouseholderSequence.h @@ -319,12 +319,36 @@ template class HouseholderS template inline void applyThisOnTheLeft(Dest& dst, Workspace& workspace) const { - workspace.resize(dst.cols()); - for(Index k = 0; k < m_length; ++k) + const Index BlockSize = 1; + // if the entries are large enough, then apply the reflectors by block + if(m_length>BlockSize && dst.cols()>1) { - Index actual_k = m_trans ? k : m_length-k-1; - dst.bottomRows(rows()-m_shift-actual_k) - .applyHouseholderOnTheLeft(essentialVector(actual_k), m_coeffs.coeff(actual_k), workspace.data()); + for(Index i = 0; i < m_length; i+=BlockSize) + { + Index end = m_trans ? (std::min)(m_length,i+BlockSize) : m_length-i; + Index k = m_trans ? i : (std::max)(Index(0),end-BlockSize); + Index bs = end-k; + Index start = k + m_shift; + + typedef Block::type,Dynamic,Dynamic> SubVectorsType; + SubVectorsType sub_vecs1(m_vectors.const_cast_derived(), Side==OnTheRight ? k : start, + Side==OnTheRight ? start : k, + Side==OnTheRight ? bs : m_vectors.rows()-start, + Side==OnTheRight ? m_vectors.cols()-start : bs); + typename internal::conditional, SubVectorsType&>::type sub_vecs(sub_vecs1); + Block sub_dst(dst,dst.rows()-rows()+m_shift+k,0, rows()-m_shift-k,dst.cols()); + apply_block_householder_on_the_left(sub_dst, sub_vecs, m_coeffs.segment(k, bs), !m_trans); + } + } + else + { + workspace.resize(dst.cols()); + for(Index k = 0; k < m_length; ++k) + { + Index actual_k = m_trans ? k : m_length-k-1; + dst.bottomRows(rows()-m_shift-actual_k) + .applyHouseholderOnTheLeft(essentialVector(actual_k), m_coeffs.coeff(actual_k), workspace.data()); + } } } diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index 136e80ce9..8604fff6f 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -299,8 +299,8 @@ struct householder_qr_inplace_blocked for (k = 0; k < size; k += blockSize) { Index bs = (std::min)(size-k,blockSize); // actual size of the block - Index tcols = cols - k - bs; // trailing columns - Index brows = rows-k; // rows of the block + Index tcols = cols - k - bs; // trailing columns + Index brows = rows-k; // rows of the block // partition the matrix: // A00 | A01 | A02 @@ -318,7 +318,7 @@ struct householder_qr_inplace_blocked if(tcols) { BlockType A21_22 = mat.block(k,k+bs,brows,tcols); - apply_block_householder_on_the_left(A21_22,A11_21,hCoeffsSegment.adjoint()); + apply_block_householder_on_the_left(A21_22,A11_21,hCoeffsSegment, false); // false == backward } } } diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 225b19e3c..0e081254d 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -215,10 +215,10 @@ void upperbidiagonalization_blocked_helper(MatrixType& A, if(k) u_k -= U_k1.adjoint() * X.row(k).head(k).adjoint(); } - // 5 - construct right Householder transform in-placecols + // 5 - construct right Householder transform in-place u_k.makeHouseholderInPlace(tau_u, upper_diagonal[k]); - // this eases the application of Householder transforAions + // this eases the application of Householder transformations // A(k,k+1) will store tau_u later A(k,k+1) = Scalar(1); From 15bad3670b7e9261764fb65037c148ab87ad9d07 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 4 Sep 2014 09:17:01 +0200 Subject: [PATCH 0769/1103] Apply Householder U and V in-place. --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index 64cee029b..fcad104c0 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -279,7 +279,7 @@ void BDCSVD::copyUV(const HouseholderU &householderU, const Househol m_matrixU = MatrixX::Identity(householderU.cols(), Ucols); Index blockCols = m_computeThinU ? m_nonzeroSingularValues : m_diagSize; m_matrixU.topLeftCorner(m_diagSize, blockCols) = naiveV.template cast().topLeftCorner(m_diagSize, blockCols); - m_matrixU = householderU * m_matrixU; + householderU.applyThisOnTheLeft(m_matrixU); } if (computeV()) { @@ -287,7 +287,7 @@ void BDCSVD::copyUV(const HouseholderU &householderU, const Househol m_matrixV = MatrixX::Identity(householderV.cols(), Vcols); Index blockCols = m_computeThinV ? m_nonzeroSingularValues : m_diagSize; m_matrixV.topLeftCorner(m_diagSize, blockCols) = naiveU.template cast().topLeftCorner(m_diagSize, blockCols); - m_matrixV = householderV * m_matrixV; + householderV.applyThisOnTheLeft(m_matrixV); } } @@ -314,7 +314,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, RealScalar betaK; RealScalar r0; RealScalar lambda, phi, c0, s0; - MatrixXr l, f; + VectorType l, f; // We use the other algorithm which is more efficient for small // matrices. if (n < m_algoswap) @@ -385,7 +385,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, // first column = q2 * s0 m_naiveU.col(firstCol).segment(firstCol + k + 1, n - k) = m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) * s0; // q2 *= c0 - m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) *= c0; + m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) *= c0; } else { @@ -408,7 +408,6 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, m_computed.col(firstCol + shift).segment(firstCol + shift + 1, k) = alphaK * l.transpose().real(); m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) = betaK * f.transpose().real(); - // Second part: try to deflate singular values in combined matrix deflation(firstCol, lastCol, k, firstRowW, firstColW, shift); @@ -417,7 +416,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, VectorType singVals; computeSVDofM(firstCol + shift, n, UofSVD, singVals, VofSVD); if (m_compU) m_naiveU.block(firstCol, firstCol, n + 1, n + 1) *= UofSVD; // FIXME this requires a temporary - else m_naiveU.block(0, firstCol, 2, n + 1) *= UofSVD; // FIXME this requires a temporary, and exploit that there are 2 rows at compile time + else m_naiveU.middleCols(firstCol, n + 1) *= UofSVD; // FIXME this requires a temporary, and exploit that there are 2 rows at compile time if (m_compV) m_naiveV.block(firstRowW, firstColW, n, n) *= VofSVD; // FIXME this requires a temporary m_computed.block(firstCol + shift, firstCol + shift, n, n).setZero(); m_computed.block(firstCol + shift, firstCol + shift, n, n).diagonal() = singVals; @@ -434,7 +433,8 @@ template void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V) { // TODO Get rid of these copies (?) - ArrayXr col0 = m_computed.block(firstCol, firstCol, n, 1); + // FIXME at least preallocate them + ArrayXr col0 = m_computed.col(firstCol).segment(firstCol, n); ArrayXr diag = m_computed.block(firstCol, firstCol, n, n).diagonal(); diag(0) = 0; @@ -446,14 +446,15 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec if (col0.hasNaN() || diag.hasNaN()) return; ArrayXr shifts(n), mus(n), zhat(n); + computeSingVals(col0, diag, singVals, shifts, mus); perturbCol0(col0, diag, singVals, shifts, mus, zhat); computeSingVecs(zhat, diag, singVals, shifts, mus, U, V); // Reverse order so that singular values in increased order singVals.reverseInPlace(); - U.leftCols(n) = U.leftCols(n).rowwise().reverse().eval(); - if (m_compV) V = V.rowwise().reverse().eval(); + U.leftCols(n) = U.leftCols(n).rowwise().reverse().eval(); // FIXME this requires a temporary + if (m_compV) V = V.rowwise().reverse().eval(); // FIXME this requires a temporary } template From f50548e86af75fd8e0d1689a9fb4184cf1fec509 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 4 Sep 2014 19:50:27 -0700 Subject: [PATCH 0770/1103] Added missing tensor copy constructors. As a result it is now possible to declare and initialize a tensor on the same line, as in: Tensor T = A + B; or Tensor T(A.reshape(new_shape)); --- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 27 ++++++++++++++++++--- 1 file changed, 23 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index fdbe8df4c..879057f38 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -55,7 +55,7 @@ namespace Eigen { * change dramatically. * * - * \ref TopicStorageOrders + * \ref TopicStorageOrders */ template @@ -75,7 +75,7 @@ class Tensor : public TensorBase > enum { IsAligned = bool(EIGEN_ALIGN) & !(Options_&DontAlign), - PacketAccess = true, + PacketAccess = (internal::packet_traits::size > 1), }; static const int Options = Options_; @@ -224,12 +224,31 @@ class Tensor : public TensorBase > } #endif - inline Tensor(const array& dimensions) - : m_storage(internal::array_prod(dimensions), dimensions) + inline explicit Tensor(const array& dimensions) + : m_storage(internal::array_prod(dimensions), dimensions) { EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Tensor(const TensorBase& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other.derived()); + resize(TensorEvaluator(assign, DefaultDevice()).dimensions()); + internal::TensorExecutor::run(assign, DefaultDevice()); + } + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Tensor(const TensorBase& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other.derived()); + resize(TensorEvaluator(assign, DefaultDevice()).dimensions()); + internal::TensorExecutor::run(assign, DefaultDevice()); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Tensor& operator=(const Tensor& other) { From d43f737b4ad52e84a3b4d954d9bfb4c40cf9e819 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 4 Sep 2014 20:02:28 -0700 Subject: [PATCH 0771/1103] Added support for evaluation of tensor shuffling operations as lvalues --- .../Eigen/CXX11/src/Tensor/TensorBase.h | 19 +++- .../Eigen/CXX11/src/Tensor/TensorShuffling.h | 104 ++++++++++++++---- 2 files changed, 96 insertions(+), 27 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index da5148a5b..2da8f8cc8 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -222,19 +222,19 @@ class TensorBase return TensorSlicingOp(derived(), startIndices, sizes); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - TensorPaddingOp + const TensorPaddingOp pad(const PaddingDimensions& padding) const { - return TensorPaddingOp(derived(), padding); + return TensorPaddingOp(derived(), padding); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - TensorShufflingOp + const TensorShufflingOp shuffle(const Shuffle& shuffle) const { - return TensorShufflingOp(derived(), shuffle); + return TensorShufflingOp(derived(), shuffle); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - TensorStridingOp + const TensorStridingOp stride(const Strides& strides) const { - return TensorStridingOp(derived(), strides); + return TensorStridingOp(derived(), strides); } // Force the evaluation of the expression. @@ -244,6 +244,7 @@ class TensorBase } protected: + template friend class Tensor; template friend class TensorBase; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Derived& derived() const { return *static_cast(this); } @@ -258,6 +259,7 @@ class TensorBase : public TensorBase::type PacketReturnType; + template friend class Tensor; template friend class TensorBase; EIGEN_DEVICE_FUNC @@ -293,6 +295,11 @@ class TensorBase : public TensorBase(derived(), startIndices, sizes); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorShufflingOp + shuffle(const Shuffle& shuffle) const { + return TensorShufflingOp(derived(), shuffle); + } // Select the device on which to evaluate the expression. template diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h index 4dfc99203..f7e7fc107 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h @@ -48,7 +48,7 @@ struct nested, 1, typename eval -class TensorShufflingOp : public TensorBase, WriteAccessors> +class TensorShufflingOp : public TensorBase > { public: typedef typename Eigen::internal::traits::Scalar Scalar; @@ -94,33 +94,38 @@ struct TensorEvaluator, Device> typedef typename XprType::Index Index; static const int NumDims = internal::array_size::Dimensions>::value; typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; enum { - IsAligned = /*TensorEvaluator::IsAligned*/false, - PacketAccess = /*TensorEvaluator::PacketAccess*/false, + IsAligned = true, + PacketAccess = (internal::packet_traits::size > 1), }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) - : m_impl(op.expression(), device), m_shuffle(op.shuffle()) + : m_impl(op.expression(), device) { const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + const Shuffle& shuffle = op.shuffle(); for (int i = 0; i < NumDims; ++i) { - m_dimensions[i] = input_dims[m_shuffle[i]]; + m_dimensions[i] = input_dims[shuffle[i]]; } + array inputStrides; + for (int i = 0; i < NumDims; ++i) { if (i > 0) { - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + inputStrides[i] = inputStrides[i-1] * input_dims[i-1]; m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; } else { - m_inputStrides[0] = 1; + inputStrides[0] = 1; m_outputStrides[0] = 1; } } + for (int i = 0; i < NumDims; ++i) { + m_inputStrides[i] = inputStrides[shuffle[i]]; + } } - // typedef typename XprType::Index Index; - typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; @@ -136,33 +141,90 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { - Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_outputStrides[i]; - inputIndex += idx * m_inputStrides[m_shuffle[i]]; - index -= idx * m_outputStrides[i]; - } - inputIndex += index * m_inputStrides[m_shuffle[0]]; - return m_impl.coeff(inputIndex); + return m_impl.coeff(srcCoeff(index)); } - /* template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - return m_impl.template packet(index); - }*/ + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); + + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + for (int i = 0; i < packetSize; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } Scalar* data() const { return NULL; } protected: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const + { + Index inputIndex = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += idx * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + return inputIndex + index * m_inputStrides[0]; + } + Dimensions m_dimensions; - Shuffle m_shuffle; array m_outputStrides; array m_inputStrides; TensorEvaluator m_impl; }; +// Eval as lvalue +template +struct TensorEvaluator, Device> + : public TensorEvaluator, Device> +{ + typedef TensorEvaluator, Device> Base; + + typedef TensorShufflingOp XprType; + typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::Dimensions>::value; + typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; + + enum { + IsAligned = true, + PacketAccess = (internal::packet_traits::size > 1), + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : Base(op, device) + { } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) + { + return this->m_impl.coeffRef(this->srcCoeff(index)); + } + + template EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketReturnType& x) + { + static const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + internal::pstore(values, x); + for (int i = 0; i < packetSize; ++i) { + this->coeffRef(index+i) = values[i]; + } + } +}; + + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_SHUFFLING_H From 1abe4ed14c0012d85e833c5f507f282cf26edc36 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 4 Sep 2014 20:27:28 -0700 Subject: [PATCH 0772/1103] Created more regression tests --- test/main.h | 1 + unsupported/test/cxx11_tensor_assign.cpp | 26 ++ unsupported/test/cxx11_tensor_contraction.cpp | 166 +++++++++++ unsupported/test/cxx11_tensor_device.cpp | 281 +++++++++++++++--- unsupported/test/cxx11_tensor_shuffling.cpp | 47 +++ unsupported/test/cxx11_tensor_simple.cpp | 26 ++ 6 files changed, 511 insertions(+), 36 deletions(-) diff --git a/test/main.h b/test/main.h index 3295dcb71..763cec8f9 100644 --- a/test/main.h +++ b/test/main.h @@ -207,6 +207,7 @@ inline void verify_impl(bool condition, const char *testname, const char *file, #define VERIFY(a) ::verify_impl(a, g_test_stack.back().c_str(), __FILE__, __LINE__, EI_PP_MAKE_STRING(a)) #define VERIFY_IS_EQUAL(a, b) VERIFY(test_is_equal(a, b)) +#define VERIFY_IS_NOT_EQUAL(a, b) VERIFY(!test_is_equal(a, b)) #define VERIFY_IS_APPROX(a, b) VERIFY(test_isApprox(a, b)) #define VERIFY_IS_NOT_APPROX(a, b) VERIFY(!test_isApprox(a, b)) #define VERIFY_IS_MUCH_SMALLER_THAN(a, b) VERIFY(test_isMuchSmallerThan(a, b)) diff --git a/unsupported/test/cxx11_tensor_assign.cpp b/unsupported/test/cxx11_tensor_assign.cpp index b024bed19..f2b126413 100644 --- a/unsupported/test/cxx11_tensor_assign.cpp +++ b/unsupported/test/cxx11_tensor_assign.cpp @@ -228,6 +228,30 @@ static void test_same_type() } } +static void test_auto_resize() +{ + Tensor tensor1; + Tensor tensor2(3); + Tensor tensor3(5); + Tensor tensor4(7); + + Tensor new_tensor(5); + new_tensor.setRandom(); + + tensor1 = tensor2 = tensor3 = tensor4 = new_tensor; + + VERIFY_IS_EQUAL(tensor1.dimension(0), new_tensor.dimension(0)); + VERIFY_IS_EQUAL(tensor2.dimension(0), new_tensor.dimension(0)); + VERIFY_IS_EQUAL(tensor3.dimension(0), new_tensor.dimension(0)); + VERIFY_IS_EQUAL(tensor4.dimension(0), new_tensor.dimension(0)); + for (int i = 0; i < new_tensor.dimension(0); ++i) { + VERIFY_IS_EQUAL(tensor1(i), new_tensor(i)); + VERIFY_IS_EQUAL(tensor2(i), new_tensor(i)); + VERIFY_IS_EQUAL(tensor3(i), new_tensor(i)); + VERIFY_IS_EQUAL(tensor4(i), new_tensor(i)); + } +} + void test_cxx11_tensor_assign() { @@ -235,4 +259,6 @@ void test_cxx11_tensor_assign() CALL_SUBTEST(test_2d()); CALL_SUBTEST(test_3d()); CALL_SUBTEST(test_same_type()); + CALL_SUBTEST(test_auto_resize()); + } diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp index fc67d500b..a37fcd967 100644 --- a/unsupported/test/cxx11_tensor_contraction.cpp +++ b/unsupported/test/cxx11_tensor_contraction.cpp @@ -141,6 +141,66 @@ static void test_multidims() } +static void test_holes() { + Tensor t1(2, 5, 7, 3); + Tensor t2(2, 7, 11, 13, 3); + t1.setRandom(); + t2.setRandom(); + + Eigen::array dims({{DimPair(0, 0), DimPair(3, 4)}}); + Tensor result = t1.contract(t2, dims); + VERIFY_IS_EQUAL(result.dimension(0), 5); + VERIFY_IS_EQUAL(result.dimension(1), 7); + VERIFY_IS_EQUAL(result.dimension(2), 7); + VERIFY_IS_EQUAL(result.dimension(3), 11); + VERIFY_IS_EQUAL(result.dimension(4), 13); + + for (int i = 0; i < 5; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 5; ++l) { + for (int m = 0; m < 5; ++m) { + VERIFY_IS_APPROX(result(i, j, k, l, m), + t1(0, i, j, 0) * t2(0, k, l, m, 0) + + t1(1, i, j, 0) * t2(1, k, l, m, 0) + + t1(0, i, j, 1) * t2(0, k, l, m, 1) + + t1(1, i, j, 1) * t2(1, k, l, m, 1) + + t1(0, i, j, 2) * t2(0, k, l, m, 2) + + t1(1, i, j, 2) * t2(1, k, l, m, 2)); + } + } + } + } + } +} + + +static void test_full_redux() +{ + Tensor t1(2, 2); + Tensor t2(2, 2, 2); + t1.setRandom(); + t2.setRandom(); + + Eigen::array dims({{DimPair(0, 0), DimPair(1, 1)}}); + Tensor result = t1.contract(t2, dims); + VERIFY_IS_EQUAL(result.dimension(0), 2); + VERIFY_IS_APPROX(result(0), t1(0, 0) * t2(0, 0, 0) + t1(1, 0) * t2(1, 0, 0) + + t1(0, 1) * t2(0, 1, 0) + t1(1, 1) * t2(1, 1, 0)); + VERIFY_IS_APPROX(result(1), t1(0, 0) * t2(0, 0, 1) + t1(1, 0) * t2(1, 0, 1) + + t1(0, 1) * t2(0, 1, 1) + t1(1, 1) * t2(1, 1, 1)); + + dims[0] = DimPair(1, 0); + dims[1] = DimPair(2, 1); + result = t2.contract(t1, dims); + VERIFY_IS_EQUAL(result.dimension(0), 2); + VERIFY_IS_APPROX(result(0), t1(0, 0) * t2(0, 0, 0) + t1(1, 0) * t2(0, 1, 0) + + t1(0, 1) * t2(0, 0, 1) + t1(1, 1) * t2(0, 1, 1)); + VERIFY_IS_APPROX(result(1), t1(0, 0) * t2(1, 0, 0) + t1(1, 0) * t2(1, 1, 0) + + t1(0, 1) * t2(1, 0, 1) + t1(1, 1) * t2(1, 1, 1)); +} + + static void test_expr() { Tensor mat1(2, 3); @@ -160,10 +220,116 @@ static void test_expr() } +static void test_out_of_order_contraction() +{ + Tensor mat1(2, 2, 2); + Tensor mat2(2, 2, 2); + + mat1.setRandom(); + mat2.setRandom(); + + Tensor mat3(2, 2); + + Eigen::array dims({{DimPair(2, 0), DimPair(0, 2)}}); + mat3 = mat1.contract(mat2, dims); + + VERIFY_IS_APPROX(mat3(0, 0), + mat1(0,0,0)*mat2(0,0,0) + mat1(1,0,0)*mat2(0,0,1) + + mat1(0,0,1)*mat2(1,0,0) + mat1(1,0,1)*mat2(1,0,1)); + VERIFY_IS_APPROX(mat3(1, 0), + mat1(0,1,0)*mat2(0,0,0) + mat1(1,1,0)*mat2(0,0,1) + + mat1(0,1,1)*mat2(1,0,0) + mat1(1,1,1)*mat2(1,0,1)); + VERIFY_IS_APPROX(mat3(0, 1), + mat1(0,0,0)*mat2(0,1,0) + mat1(1,0,0)*mat2(0,1,1) + + mat1(0,0,1)*mat2(1,1,0) + mat1(1,0,1)*mat2(1,1,1)); + VERIFY_IS_APPROX(mat3(1, 1), + mat1(0,1,0)*mat2(0,1,0) + mat1(1,1,0)*mat2(0,1,1) + + mat1(0,1,1)*mat2(1,1,0) + mat1(1,1,1)*mat2(1,1,1)); + + Eigen::array dims2({{DimPair(0, 2), DimPair(2, 0)}}); + mat3 = mat1.contract(mat2, dims2); + + VERIFY_IS_APPROX(mat3(0, 0), + mat1(0,0,0)*mat2(0,0,0) + mat1(1,0,0)*mat2(0,0,1) + + mat1(0,0,1)*mat2(1,0,0) + mat1(1,0,1)*mat2(1,0,1)); + VERIFY_IS_APPROX(mat3(1, 0), + mat1(0,1,0)*mat2(0,0,0) + mat1(1,1,0)*mat2(0,0,1) + + mat1(0,1,1)*mat2(1,0,0) + mat1(1,1,1)*mat2(1,0,1)); + VERIFY_IS_APPROX(mat3(0, 1), + mat1(0,0,0)*mat2(0,1,0) + mat1(1,0,0)*mat2(0,1,1) + + mat1(0,0,1)*mat2(1,1,0) + mat1(1,0,1)*mat2(1,1,1)); + VERIFY_IS_APPROX(mat3(1, 1), + mat1(0,1,0)*mat2(0,1,0) + mat1(1,1,0)*mat2(0,1,1) + + mat1(0,1,1)*mat2(1,1,0) + mat1(1,1,1)*mat2(1,1,1)); + +} + + +static void test_consistency() +{ + // this does something like testing (A*B)^T = (B^T * A^T) + + Tensor mat1(4, 3, 5); + Tensor mat2(3, 2, 1, 5, 4); + mat1.setRandom(); + mat2.setRandom(); + + Tensor mat3(5, 2, 1, 5); + Tensor mat4(2, 1, 5, 5); + + // contract on dimensions of size 4 and 3 + Eigen::array dims1({{DimPair(0, 4), DimPair(1, 0)}}); + Eigen::array dims2({{DimPair(4, 0), DimPair(0, 1)}}); + + mat3 = mat1.contract(mat2, dims1); + mat4 = mat2.contract(mat1, dims2); + + // check that these are equal except for ordering of dimensions + for (size_t i = 0; i < 5; i++) { + for (size_t j = 0; j < 10; j++) { + VERIFY_IS_APPROX(mat3.data()[i + 5 * j], mat4.data()[j + 10 * i]); + } + } +} + + +static void test_large_contraction() +{ + Tensor t_left(30, 50, 8, 31); + Tensor t_right(8, 31, 7, 20, 10); + Tensor t_result(30, 50, 7, 20, 10); + + t_left.setRandom(); + t_right.setRandom(); + + typedef Map MapXf; + MapXf m_left(t_left.data(), 1500, 248); + MapXf m_right(t_right.data(), 248, 1400); + MatrixXf m_result(1500, 1400); + + // this contraction should be equivalent to a single matrix multiplication + Eigen::array dims({{DimPair(2, 0), DimPair(3, 1)}}); + + // compute results by separate methods + t_result = t_left.contract(t_right, dims); + m_result = m_left * m_right; + + for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { + VERIFY(&t_result.data()[i] != &m_result.data()[i]); + VERIFY_IS_APPROX(t_result.data()[i], m_result.data()[i]); + } +} + + void test_cxx11_tensor_contraction() { CALL_SUBTEST(test_evals()); CALL_SUBTEST(test_scalar()); CALL_SUBTEST(test_multidims()); + CALL_SUBTEST(test_holes()); + CALL_SUBTEST(test_full_redux()); CALL_SUBTEST(test_expr()); + CALL_SUBTEST(test_out_of_order_contraction()); + CALL_SUBTEST(test_consistency()); + CALL_SUBTEST(test_large_contraction()); } diff --git a/unsupported/test/cxx11_tensor_device.cpp b/unsupported/test/cxx11_tensor_device.cpp index caf2e9735..f331cb481 100644 --- a/unsupported/test/cxx11_tensor_device.cpp +++ b/unsupported/test/cxx11_tensor_device.cpp @@ -22,17 +22,43 @@ using Eigen::RowMajor; // Context for evaluation on cpu struct CPUContext { - CPUContext(const Eigen::Tensor& in1, Eigen::Tensor& in2, Eigen::Tensor& out) : in1_(in1), in2_(in2), out_(out) { } + CPUContext(const Eigen::Tensor& in1, Eigen::Tensor& in2, Eigen::Tensor& out) : in1_(in1), in2_(in2), out_(out), kernel_1d_(2), kernel_2d_(Eigen::array(2,2)), kernel_3d_(Eigen::array(2,2,2)) { + kernel_1d_(0) = 3.14f; + kernel_1d_(1) = 2.7f; + + kernel_2d_(Eigen::array(0,0)) = 3.14f; + kernel_2d_(Eigen::array(1,0)) = 2.7f; + kernel_2d_(Eigen::array(0,1)) = 0.2f; + kernel_2d_(Eigen::array(1,1)) = 7.0f; + + kernel_3d_(Eigen::array(0,0,0)) = 3.14f; + kernel_3d_(Eigen::array(0,1,0)) = 2.7f; + kernel_3d_(Eigen::array(0,0,1)) = 0.2f; + kernel_3d_(Eigen::array(0,1,1)) = 7.0f; + kernel_3d_(Eigen::array(1,0,0)) = -1.0f; + kernel_3d_(Eigen::array(1,1,0)) = -0.3f; + kernel_3d_(Eigen::array(1,0,1)) = -0.7f; + kernel_3d_(Eigen::array(1,1,1)) = -0.5f; + } + + const Eigen::DefaultDevice& device() const { return cpu_device_; } const Eigen::Tensor& in1() const { return in1_; } const Eigen::Tensor& in2() const { return in2_; } - Eigen::TensorDevice, Eigen::DefaultDevice> out() { return TensorDevice, Eigen::DefaultDevice>(cpu_device_, out_); } + Eigen::Tensor& out() { return out_; } + const Eigen::Tensor& kernel1d() const { return kernel_1d_; } + const Eigen::Tensor& kernel2d() const { return kernel_2d_; } + const Eigen::Tensor& kernel3d() const { return kernel_3d_; } private: const Eigen::Tensor& in1_; const Eigen::Tensor& in2_; Eigen::Tensor& out_; + Eigen::Tensor kernel_1d_; + Eigen::Tensor kernel_2d_; + Eigen::Tensor kernel_3d_; + Eigen::DefaultDevice cpu_device_; }; @@ -40,19 +66,45 @@ struct CPUContext { // Context for evaluation on GPU struct GPUContext { GPUContext(const Eigen::TensorMap >& in1, Eigen::TensorMap >& in2, Eigen::TensorMap >& out) : in1_(in1), in2_(in2), out_(out), gpu_device_(&stream_) { - cudaStreamCreate(&stream_); + assert(cudaMalloc((void**)(&kernel_1d_), 2*sizeof(float)) == cudaSuccess); + float kernel_1d_val[] = {3.14f, 2.7f}; + assert(cudaMemcpy(kernel_1d_, kernel_1d_val, 2*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess); + + assert(cudaMalloc((void**)(&kernel_2d_), 4*sizeof(float)) == cudaSuccess); + float kernel_2d_val[] = {3.14f, 2.7f, 0.2f, 7.0f}; + assert(cudaMemcpy(kernel_2d_, kernel_2d_val, 4*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess); + + assert(cudaMalloc((void**)(&kernel_3d_), 8*sizeof(float)) == cudaSuccess); + float kernel_3d_val[] = {3.14f, -1.0f, 2.7f, -0.3f, 0.2f, -0.7f, 7.0f, -0.5f}; + assert(cudaMemcpy(kernel_3d_, kernel_3d_val, 8*sizeof(float), cudaMemcpyHostToDevice) == cudaSuccess); + + assert(cudaStreamCreate(&stream_) == cudaSuccess); } ~GPUContext() { - cudaStreamDestroy(stream_); + assert(cudaFree(kernel_1d_) == cudaSuccess); + assert(cudaFree(kernel_2d_) == cudaSuccess); + assert(cudaFree(kernel_3d_) == cudaSuccess); + assert(cudaStreamDestroy(stream_) == cudaSuccess); } + + const Eigen::GpuDevice& device() const { return gpu_device_; } + const Eigen::TensorMap >& in1() const { return in1_; } const Eigen::TensorMap >& in2() const { return in2_; } - Eigen::TensorDevice >, Eigen::GpuDevice> out() { return TensorDevice >, Eigen::GpuDevice>(gpu_device_, out_); } + Eigen::TensorMap >& out() { return out_; } + Eigen::TensorMap > kernel1d() const { return Eigen::TensorMap >(kernel_1d_, 2); } + Eigen::TensorMap > kernel2d() const { return Eigen::TensorMap >(kernel_2d_, Eigen::array(2, 2)); } + Eigen::TensorMap > kernel3d() const { return Eigen::TensorMap >(kernel_3d_, Eigen::array(2, 2, 2)); } private: const Eigen::TensorMap >& in1_; const Eigen::TensorMap >& in2_; Eigen::TensorMap >& out_; + + float* kernel_1d_; + float* kernel_2d_; + float* kernel_3d_; + cudaStream_t stream_; Eigen::GpuDevice gpu_device_; }; @@ -62,49 +114,151 @@ struct GPUContext { template static void test_contextual_eval(Context* context) { - context->out() = context->in1() + context->in2() * 3.14f + context->in1().constant(2.718f); + context->out().device(context->device()) = context->in1() + context->in2() * 3.14f + context->in1().constant(2.718f); } template static void test_forced_contextual_eval(Context* context) { - context->out() = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f); + context->out().device(context->device()) = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f); } -static void test_cpu() { - Eigen::Tensor in1(Eigen::array(2,3,7)); - Eigen::Tensor in2(Eigen::array(2,3,7)); - Eigen::Tensor out(Eigen::array(2,3,7)); +template +static void test_contraction(Context* context) +{ + Eigen::array, 2> dims; + dims[0] = std::make_pair(1, 1); + dims[1] = std::make_pair(2, 2); - in1.setRandom(); - in2.setRandom(); + Eigen::array shape(40, 50*70); + + Eigen::DSizes indices(0,0); + Eigen::DSizes sizes(40,40); + + context->out().reshape(shape).slice(indices, sizes).device(context->device()) = context->in1().contract(context->in2(), dims); +} + + +template +static void test_1d_convolution(Context* context) +{ + Eigen::DSizes indices(Eigen::array(0,0,0)); + Eigen::DSizes sizes(Eigen::array(40,49,70)); + + Eigen::array dims(1); + context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel1d(), dims); +} + +template +static void test_2d_convolution(Context* context) +{ + Eigen::DSizes indices(Eigen::array(0,0,0)); + Eigen::DSizes sizes(Eigen::array(40,49,69)); + + Eigen::array dims(1,2); + context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel2d(), dims); +} + +template +static void test_3d_convolution(Context* context) +{ + Eigen::DSizes indices(Eigen::array(0,0,0)); + Eigen::DSizes sizes(Eigen::array(39,49,69)); + + Eigen::array dims(0,1,2); + context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel3d(), dims); +} + + +static void test_cpu() { + Eigen::Tensor in1(Eigen::array(40,50,70)); + Eigen::Tensor in2(Eigen::array(40,50,70)); + Eigen::Tensor out(Eigen::array(40,50,70)); + + in1 = in1.random() + in1.constant(10.0f); + in2 = in2.random() + in2.constant(10.0f); CPUContext context(in1, in2, out); test_contextual_eval(&context); - for (int i = 0; i < 2; ++i) { - for (int j = 0; j < 3; ++j) { - for (int k = 0; k < 7; ++k) { + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 50; ++j) { + for (int k = 0; k < 70; ++k) { VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); } } } test_forced_contextual_eval(&context); - for (int i = 0; i < 2; ++i) { - for (int j = 0; j < 3; ++j) { - for (int k = 0; k < 7; ++k) { + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 50; ++j) { + for (int k = 0; k < 70; ++k) { VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k))) * 3.14f + 2.718f); } } } + + test_contraction(&context); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 40; ++j) { + const float result = out(Eigen::array(i,j,0)); + float expected = 0; + for (int k = 0; k < 50; ++k) { + for (int l = 0; l < 70; ++l) { + expected += in1(Eigen::array(i, k, l)) * in2(Eigen::array(j, k, l)); + } + } + VERIFY_IS_APPROX(expected, result); + } + } + + test_1d_convolution(&context); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 49; ++j) { + for (int k = 0; k < 70; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f)); + } + } + } + + test_2d_convolution(&context); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 49; ++j) { + for (int k = 0; k < 69; ++k) { + const float result = out(Eigen::array(i,j,k)); + const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f) + + (in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f); + if (fabs(expected) < 1e-4 && fabs(result) < 1e-4) { + continue; + } + VERIFY_IS_APPROX(expected, result); + } + } + } + + test_3d_convolution(&context); + for (int i = 0; i < 39; ++i) { + for (int j = 0; j < 49; ++j) { + for (int k = 0; k < 69; ++k) { + const float result = out(Eigen::array(i,j,k)); + const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f + + in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f) + + (in1(Eigen::array(i+1,j,k)) * -1.0f + in1(Eigen::array(i+1,j+1,k)) * -0.3f + + in1(Eigen::array(i+1,j,k+1)) * -0.7f + in1(Eigen::array(i+1,j+1,k+1)) * -0.5f); + if (fabs(expected) < 1e-4 && fabs(result) < 1e-4) { + continue; + } + VERIFY_IS_APPROX(expected, result); + } + } + } } static void test_gpu() { - Eigen::Tensor in1(Eigen::array(2,3,7)); - Eigen::Tensor in2(Eigen::array(2,3,7)); - Eigen::Tensor out(Eigen::array(2,3,7)); - in1.setRandom(); - in2.setRandom(); + Eigen::Tensor in1(Eigen::array(40,50,70)); + Eigen::Tensor in2(Eigen::array(40,50,70)); + Eigen::Tensor out(Eigen::array(40,50,70)); + in1 = in1.random() + in1.constant(10.0f); + in2 = in2.random() + in2.constant(10.0f); std::size_t in1_bytes = in1.size() * sizeof(float); std::size_t in2_bytes = in2.size() * sizeof(float); @@ -120,32 +274,87 @@ static void test_gpu() { cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice); cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice); - Eigen::TensorMap > gpu_in1(d_in1, Eigen::array(2,3,7)); - Eigen::TensorMap > gpu_in2(d_in2, Eigen::array(2,3,7)); - Eigen::TensorMap > gpu_out(d_out, Eigen::array(2,3,7)); + Eigen::TensorMap > gpu_in1(d_in1, Eigen::array(40,50,70)); + Eigen::TensorMap > gpu_in2(d_in2, Eigen::array(40,50,70)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(40,50,70)); GPUContext context(gpu_in1, gpu_in2, gpu_out); test_contextual_eval(&context); - cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost); - for (int i = 0; i < 2; ++i) { - for (int j = 0; j < 3; ++j) { - for (int k = 0; k < 7; ++k) { + assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 50; ++j) { + for (int k = 0; k < 70; ++k) { VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); } } } test_forced_contextual_eval(&context); - cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost); - for (int i = 0; i < 2; ++i) { - for (int j = 0; j < 3; ++j) { - for (int k = 0; k < 7; ++k) { + assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 50; ++j) { + for (int k = 0; k < 70; ++k) { VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k))) * 3.14f + 2.718f); } } } -} + test_contraction(&context); + assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 40; ++j) { + const float result = out(Eigen::array(i,j,0)); + float expected = 0; + for (int k = 0; k < 50; ++k) { + for (int l = 0; l < 70; ++l) { + expected += in1(Eigen::array(i, k, l)) * in2(Eigen::array(j, k, l)); + } + } + VERIFY_IS_APPROX(expected, result); + } + } + + test_1d_convolution(&context); + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess); + assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 49; ++j) { + for (int k = 0; k < 70; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f)); + } + } + } + + test_2d_convolution(&context); + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess); + assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 49; ++j) { + for (int k = 0; k < 69; ++k) { + const float result = out(Eigen::array(i,j,k)); + const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f + + in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f); + VERIFY_IS_APPROX(expected, result); + } + } + } + + test_3d_convolution(&context); + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, context.device().stream()) == cudaSuccess); + assert(cudaStreamSynchronize(context.device().stream()) == cudaSuccess); + for (int i = 0; i < 39; ++i) { + for (int j = 0; j < 49; ++j) { + for (int k = 0; k < 69; ++k) { + const float result = out(Eigen::array(i,j,k)); + const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f + + in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f + + in1(Eigen::array(i+1,j,k)) * -1.0f + in1(Eigen::array(i+1,j+1,k)) * -0.3f + + in1(Eigen::array(i+1,j,k+1)) * -0.7f + in1(Eigen::array(i+1,j+1,k+1)) * -0.5f); + VERIFY_IS_APPROX(expected, result); + } + } + } +} void test_cxx11_tensor_device() diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp index 92dd01a52..5ab8b6821 100644 --- a/unsupported/test/cxx11_tensor_shuffling.cpp +++ b/unsupported/test/cxx11_tensor_shuffling.cpp @@ -106,11 +106,58 @@ static void test_expr_shuffling() } } } + + dst_slice_start[0] = 0; + result.setRandom(); + for (int i = 0; i < 5; ++i) { + result.slice(dst_slice_start, dst_slice_dim) = + tensor.shuffle(shuffles).slice(dst_slice_start, dst_slice_dim); + dst_slice_start[0] += 1; + } + + for (int i = 0; i < expected.dimension(0); ++i) { + for (int j = 0; j < expected.dimension(1); ++j) { + for (int k = 0; k < expected.dimension(2); ++k) { + for (int l = 0; l < expected.dimension(3); ++l) { + VERIFY_IS_EQUAL(result(i,j,k,l), expected(i,j,k,l)); + } + } + } + } } +static void test_shuffling_as_value() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + array shuffles; + shuffles[2] = 0; + shuffles[3] = 1; + shuffles[1] = 2; + shuffles[0] = 3; + Tensor shuffle(5,7,3,2); + shuffle.shuffle(shuffles) = tensor; + + VERIFY_IS_EQUAL(shuffle.dimension(0), 5); + VERIFY_IS_EQUAL(shuffle.dimension(1), 7); + VERIFY_IS_EQUAL(shuffle.dimension(2), 3); + VERIFY_IS_EQUAL(shuffle.dimension(3), 2); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(tensor(i,j,k,l), shuffle(k,l,j,i)); + } + } + } + } +} + void test_cxx11_tensor_shuffling() { CALL_SUBTEST(test_simple_shuffling()); CALL_SUBTEST(test_expr_shuffling()); + CALL_SUBTEST(test_shuffling_as_value()); } diff --git a/unsupported/test/cxx11_tensor_simple.cpp b/unsupported/test/cxx11_tensor_simple.cpp index 1455f2a4c..a70591c82 100644 --- a/unsupported/test/cxx11_tensor_simple.cpp +++ b/unsupported/test/cxx11_tensor_simple.cpp @@ -257,12 +257,38 @@ static void test_simple_assign() VERIFY_IS_EQUAL((e2(1,0,2)), -1); } +static void test_resize() +{ + Tensor epsilon; + epsilon.resize(2,3,7); + VERIFY_IS_EQUAL(epsilon.dimension(0), 2); + VERIFY_IS_EQUAL(epsilon.dimension(1), 3); + VERIFY_IS_EQUAL(epsilon.dimension(2), 7); + VERIFY_IS_EQUAL(epsilon.dimensions().TotalSize(), 2ul*3*7); + + const int* old_data = epsilon.data(); + epsilon.resize(3,2,7); + VERIFY_IS_EQUAL(epsilon.dimension(0), 3); + VERIFY_IS_EQUAL(epsilon.dimension(1), 2); + VERIFY_IS_EQUAL(epsilon.dimension(2), 7); + VERIFY_IS_EQUAL(epsilon.dimensions().TotalSize(), 2ul*3*7); + VERIFY_IS_EQUAL(epsilon.data(), old_data); + + epsilon.resize(3,5,7); + VERIFY_IS_EQUAL(epsilon.dimension(0), 3); + VERIFY_IS_EQUAL(epsilon.dimension(1), 5); + VERIFY_IS_EQUAL(epsilon.dimension(2), 7); + VERIFY_IS_EQUAL(epsilon.dimensions().TotalSize(), 3ul*5*7); + VERIFY_IS_NOT_EQUAL(epsilon.data(), old_data); +} + void test_cxx11_tensor_simple() { CALL_SUBTEST(test_1d()); CALL_SUBTEST(test_2d()); CALL_SUBTEST(test_3d()); CALL_SUBTEST(test_simple_assign()); + CALL_SUBTEST(test_resize()); } /* From b23556bbbd796c078e4310c806feceacc8e8c3e8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 5 Sep 2014 08:50:50 +0200 Subject: [PATCH 0773/1103] Oops, a block size of 1 is not very useful, set it to 48 as in HouseholderQR --- Eigen/src/Householder/HouseholderSequence.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h index 111936f0e..ab32f7d3f 100644 --- a/Eigen/src/Householder/HouseholderSequence.h +++ b/Eigen/src/Householder/HouseholderSequence.h @@ -319,7 +319,7 @@ template class HouseholderS template inline void applyThisOnTheLeft(Dest& dst, Workspace& workspace) const { - const Index BlockSize = 1; + const Index BlockSize = 48; // if the entries are large enough, then apply the reflectors by block if(m_length>BlockSize && dst.cols()>1) { From 74db22455ae0172faaae91321da0b303bb82369d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 5 Sep 2014 07:47:43 -0700 Subject: [PATCH 0774/1103] Misc fixes. --- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 12 +++--- .../Eigen/CXX11/src/Tensor/TensorPadding.h | 2 +- unsupported/test/cxx11_tensor_padding.cpp | 38 ++++++++++++++++++- 3 files changed, 43 insertions(+), 9 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index d9a6b3f1b..28ae7b3c6 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -163,7 +163,7 @@ template { return this->m_impl.coeffRef(index); } - template EIGEN_STRONG_INLINE + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writePacket(Index index, const PacketReturnType& x) { this->m_impl.template writePacket(index, x); @@ -314,7 +314,7 @@ struct TensorEvaluator, Devi Scalar* src = m_impl.data(); for (int i = 0; i < internal::array_prod(dimensions()); i += contiguous_values) { Index offset = srcCoeff(i); - m_device.memcpy(data+i, src+offset, contiguous_values * sizeof(Scalar)); + m_device.memcpy((void*)(data+i), src+offset, contiguous_values * sizeof(Scalar)); } return false; } @@ -334,7 +334,7 @@ struct TensorEvaluator, Devi template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - static const int packetSize = internal::unpacket_traits::size; + const int packetSize = internal::unpacket_traits::size; EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) eigen_assert(index+packetSize-1 < dimensions().TotalSize()); @@ -355,7 +355,7 @@ struct TensorEvaluator, Devi return rslt; } else { - CoeffReturnType values[packetSize]; + typename internal::remove_const::type values[packetSize]; values[0] = m_impl.coeff(inputIndices[0]); values[packetSize-1] = m_impl.coeff(inputIndices[1]); for (int i = 1; i < packetSize-1; ++i) { @@ -420,10 +420,10 @@ struct TensorEvaluator, Device> return this->m_impl.coeffRef(this->srcCoeff(index)); } - template EIGEN_STRONG_INLINE + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writePacket(Index index, const PacketReturnType& x) { - static const int packetSize = internal::unpacket_traits::size; + const int packetSize = internal::unpacket_traits::size; Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; for (int i = NumDims - 1; i > 0; --i) { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h index 4482c0992..7da89458f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h @@ -48,7 +48,7 @@ struct nested, 1, typename eval -class TensorPaddingOp : public TensorBase > +class TensorPaddingOp : public TensorBase, ReadOnlyAccessors> { public: typedef typename Eigen::internal::traits::Scalar Scalar; diff --git a/unsupported/test/cxx11_tensor_padding.cpp b/unsupported/test/cxx11_tensor_padding.cpp index cb010f512..6f74216dd 100644 --- a/unsupported/test/cxx11_tensor_padding.cpp +++ b/unsupported/test/cxx11_tensor_padding.cpp @@ -37,9 +37,42 @@ static void test_simple_padding() for (int k = 0; k < 12; ++k) { for (int l = 0; l < 7; ++l) { if (j >= 2 && j < 5 && k >= 3 && k < 8) { - VERIFY_IS_EQUAL(tensor(i,j-2,k-3,l), padded(i,j,k,l)); + VERIFY_IS_EQUAL(padded(i,j,k,l), tensor(i,j-2,k-3,l)); } else { - VERIFY_IS_EQUAL(0.0f, padded(i,j,k,l)); + VERIFY_IS_EQUAL(padded(i,j,k,l), 0.0f); + } + } + } + } + } +} + +static void test_padded_expr() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + + array, 4> paddings; + paddings[0] = std::make_pair(0, 0); + paddings[1] = std::make_pair(2, 1); + paddings[2] = std::make_pair(3, 4); + paddings[3] = std::make_pair(0, 0); + + Eigen::DSizes reshape_dims; + reshape_dims[0] = 12; + reshape_dims[1] = 84; + + Tensor result; + result = tensor.pad(paddings).reshape(reshape_dims); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 6; ++j) { + for (int k = 0; k < 12; ++k) { + for (int l = 0; l < 7; ++l) { + if (j >= 2 && j < 5 && k >= 3 && k < 8) { + VERIFY_IS_EQUAL(result(i+2*j,k+12*l), tensor(i,j-2,k-3,l)); + } else { + VERIFY_IS_EQUAL(result(i+2*j,k+12*l), 0.0f); } } } @@ -51,4 +84,5 @@ static void test_simple_padding() void test_cxx11_tensor_padding() { CALL_SUBTEST(test_simple_padding()); + CALL_SUBTEST(test_padded_expr()); } From dacd39ea76d488133392b3abecf1c5061ba568d7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 5 Sep 2014 17:51:46 +0200 Subject: [PATCH 0775/1103] Exploit sparse structure in naiveU and naiveV when updating them. --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 54 +++++++++++++++++++++++++-- 1 file changed, 51 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index fcad104c0..0167872af 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -164,6 +164,7 @@ private: void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift); template void copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naivev); + static void structured_update(Block A, const MatrixXr &B, Index n1); protected: MatrixXr m_naiveU, m_naiveV; @@ -240,6 +241,8 @@ BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsign internal::UpperBidiagonalization bid(copy); //**** step 2 Divide + m_naiveU.setZero(); + m_naiveV.setZero(); m_computed.topRows(m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose(); m_computed.template bottomRows<1>().setZero(); divide(0, m_diagSize - 1, 0, 0, 0); @@ -291,6 +294,48 @@ void BDCSVD::copyUV(const HouseholderU &householderU, const Househol } } +/** \internal + * Performs A = A * B exploiting the special structure of the matrix A. Splitting A as: + * A = [A1] + * [A2] + * such that A1.rows()==n1, then we assume that at least half of the columns of A1 and A2 are zeros. + * We can thus pack them prior to the the matrix product. However, this is only worth the effort if the matrix is large + * enough. + */ +template +void BDCSVD::structured_update(Block A, const MatrixXr &B, Index n1) +{ + Index n = A.rows(); + if(n>100) + { + // If the matrices are large enough, let's exploit the sparse strucure of A by + // splitting it in half (wrt n1), and packing the non-zero columns. + DenseIndex n2 = n - n1; + MatrixXr A1(n1,n), A2(n2,n), B1(n,n), B2(n,n); + Index k1=0, k2=0; + for(Index j=0; j::divide (Index firstCol, Index lastCol, Index firstRowW, MatrixXr UofSVD, VofSVD; VectorType singVals; computeSVDofM(firstCol + shift, n, UofSVD, singVals, VofSVD); - if (m_compU) m_naiveU.block(firstCol, firstCol, n + 1, n + 1) *= UofSVD; // FIXME this requires a temporary - else m_naiveU.middleCols(firstCol, n + 1) *= UofSVD; // FIXME this requires a temporary, and exploit that there are 2 rows at compile time - if (m_compV) m_naiveV.block(firstRowW, firstColW, n, n) *= VofSVD; // FIXME this requires a temporary + + if (m_compU) structured_update(m_naiveU.block(firstCol, firstCol, n + 1, n + 1), UofSVD, (n+2)/2); + else m_naiveU.middleCols(firstCol, n + 1) *= UofSVD; // FIXME this requires a temporary, and exploit that there are 2 rows at compile time + + if (m_compV) structured_update(m_naiveV.block(firstRowW, firstColW, n, n), VofSVD, (n+1)/2); + m_computed.block(firstCol + shift, firstCol + shift, n, n).setZero(); m_computed.block(firstCol + shift, firstCol + shift, n, n).diagonal() = singVals; }// end divide From efdff157493826bbcc023a85e08596fd58d7997a Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Sat, 6 Sep 2014 13:28:24 -0700 Subject: [PATCH 0776/1103] Fixed a typo in the contraction code --- unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index 897d73806..46624724c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -168,7 +168,7 @@ struct TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * internal::array_size::value) { + if (TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * internal::array_size::value) { m_dimensions[0] = 1; } } From 188a13f9fe235e1579a1e2da048b3595d60e37e8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 8 Sep 2014 09:50:03 +0200 Subject: [PATCH 0777/1103] Fix compilation of coeff(Index) on sub-inner-panels --- Eigen/src/Core/Block.h | 3 ++- Eigen/src/Core/CoreEvaluators.h | 2 +- Eigen/src/Core/MapBase.h | 6 ++++++ test/block.cpp | 8 ++++++++ 4 files changed, 17 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 0f3ab304a..2d62f7a46 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -101,7 +101,8 @@ struct traits > : traits::value ? LvalueBit : 0, FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, - Flags = (traits::Flags & DirectAccessBit) | FlagsLvalueBit | FlagsRowMajorBit // FIXME DirectAccessBit should not be handled by expressions + Flags = (traits::Flags & DirectAccessBit) | FlagsLvalueBit | FlagsRowMajorBit + // FIXME DirectAccessBit should not be handled by expressions #endif }; }; diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 8c9190a0e..b8dc05d98 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -722,7 +722,7 @@ struct evaluator > ? PacketAccessBit : 0, MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) ? AlignedBit : 0, - FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (traits::Flags&LinearAccessBit))) ? LinearAccessBit : 0, + FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (evaluator::Flags&LinearAccessBit))) ? LinearAccessBit : 0, FlagsRowMajorBit = XprType::Flags&RowMajorBit, Flags0 = evaluator::Flags & ( (HereditaryBits & ~RowMajorBit) | DirectAccessBit | diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index 927a59c50..591ea26fb 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -11,9 +11,15 @@ #ifndef EIGEN_MAPBASE_H #define EIGEN_MAPBASE_H +#ifndef EIGEN_TEST_EVALUATORS #define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \ EIGEN_STATIC_ASSERT((int(internal::traits::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \ YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT) +#else +#define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \ + EIGEN_STATIC_ASSERT((int(internal::evaluator::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \ + YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT) +#endif namespace Eigen { diff --git a/test/block.cpp b/test/block.cpp index 269acd28e..3b77b704a 100644 --- a/test/block.cpp +++ b/test/block.cpp @@ -130,6 +130,14 @@ template void block(const MatrixType& m) VERIFY(numext::real(ones.col(c1).dot(ones.col(c2))) == RealScalar(rows)); VERIFY(numext::real(ones.row(r1).dot(ones.row(r2))) == RealScalar(cols)); + + // chekc that linear acccessors works on blocks + m1 = m1_copy; + if((MatrixType::Flags&RowMajorBit)==0) + VERIFY_IS_EQUAL(m1.leftCols(c1).coeff(r1+c1*rows), m1(r1,c1)); + else + VERIFY_IS_EQUAL(m1.topRows(r1).coeff(c1+r1*cols), m1(r1,c1)); + // now test some block-inside-of-block. From 470aa15c35539a7b79dd053c74578ac9e769f6e5 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Tue, 9 Sep 2014 16:58:48 +0000 Subject: [PATCH 0778/1103] First time it compiles, but fails to pass the tests. --- Eigen/Core | 6 +- Eigen/src/Core/arch/AltiVec/Complex.h | 74 ++++-- Eigen/src/Core/arch/AltiVec/PacketMath.h | 309 +++++++++++++++++++---- 3 files changed, 313 insertions(+), 76 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index 8ea165d5b..ac3cbd0c7 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -297,13 +297,9 @@ using std::ptrdiff_t; #include "src/Core/arch/SSE/PacketMath.h" #include "src/Core/arch/SSE/MathFunctions.h" #include "src/Core/arch/SSE/Complex.h" -#elif defined(EIGEN_VECTORIZE_ALTIVEC) +#elif defined(EIGEN_VECTORIZE_ALTIVEC) || defined(EIGEN_VECTORIZE_VSX) #include "src/Core/arch/AltiVec/PacketMath.h" #include "src/Core/arch/AltiVec/Complex.h" -#elif defined(EIGEN_VECTORIZE_ALTIVEC) - #include "src/Core/arch/AltiVec/PacketMath.h" - #include "src/Core/arch/AltiVec/VSX.h" - #include "src/Core/arch/AltiVec/Complex.h" #elif defined EIGEN_VECTORIZE_NEON #include "src/Core/arch/NEON/PacketMath.h" #include "src/Core/arch/NEON/Complex.h" diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 064341a3b..d90cfe0de 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -18,7 +18,7 @@ namespace internal { static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; #else static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4f_ZERO_, (Packet4ui)p4i_ZERO);//{ 0x80000000, 0x00000000, 0x80000000, 0x00000000 }; -static Packet2ul p2ul_CONJ_XOR = (Packet2ul) vec_sld((Packet4ui)p2d_ZERO_, (Packet4ui)p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui)p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; #endif static Packet16uc p16uc_COMPLEX_RE = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; @@ -276,28 +276,58 @@ template<> struct packet_traits > : default_packet_traits template<> struct unpacket_traits { typedef std::complex type; enum {size=1}; typedef Packet1cd half; }; +template<> EIGEN_STRONG_INLINE Packet1cd pload (const std::complex* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload((const double*)from)); } +template<> EIGEN_STRONG_INLINE Packet1cd ploadu(const std::complex* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu((const double*)from)); } +template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE Packet1cd pset1(const std::complex& from) +{ + Packet1cd res; + /* On AltiVec we cannot load 64-bit registers, so wa have to take care of alignment */ + if((ptrdiff_t(&from) % 16) == 0) + res.v = pload((const double *)&from); + else + res.v = ploadu((const double *)&from); + res.v = vec_perm(res.v, res.v, p16uc_PSET_HI); + return res; +} +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, DenseIndex stride) +{ + std::complex EIGEN_ALIGN16 af[2]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return pload(af); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet1cd>(std::complex* to, const Packet1cd& from, DenseIndex stride) +{ + std::complex EIGEN_ALIGN16 af[2]; + pstore >(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} + template<> EIGEN_STRONG_INLINE Packet1cd padd(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd psub(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } -template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2ul)a.v, p2ul_CONJ_XOR)); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR)); } template<> EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) { - Packet2d v1, v2; + Packet2d a_re, a_im, v1, v2; // Permute and multiply the real parts of a and b - v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE); + a_re = vec_perm(a.v, a.v, p16uc_PSET_HI); // Get the imaginary parts of a - v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM); + a_im = vec_perm(a.v, a.v, p16uc_PSET_HI); // multiply a_re * b - v1 = vec_madd(v1, b.v, p4f_ZERO); + v1 = vec_madd(a_re, b.v, p2d_ZERO_); // multiply a_im * b and get the conjugate result - v2 = vec_madd(v2, b.v, p4f_ZERO); - v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR); - // permute back to a proper order - v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV); + v2 = vec_madd(a_im, b.v, p2d_ZERO_); + v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8); + v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR); - return Packet2cf(vec_add(v1, v2)); + return Packet1cd(vec_add(v1, v2)); } template<> EIGEN_STRONG_INLINE Packet1cd pand (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); } @@ -305,23 +335,17 @@ template<> EIGEN_STRONG_INLINE Packet1cd por (const Packet1cd& a, template<> EIGEN_STRONG_INLINE Packet1cd pxor (const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pandnot(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); } -template<> EIGEN_STRONG_INLINE Packet1cd pload (const std::complex* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload((const double*)from)); } -template<> EIGEN_STRONG_INLINE Packet1cd ploadu(const std::complex* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu((const double*)from)); } - template<> EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex* from) { return pset1(*from); } -template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } -template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } - -template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { vec_dstt((double *)addr, DST_CTRL(2,2,32), DST_CHAN); } +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { vec_dstt((long *)addr, DST_CTRL(2,2,32), DST_CHAN); } template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet1cd& a) { std::complex EIGEN_ALIGN16 res[2]; - pstore((float *)&res, a.v); + pstore >(res, a); return res[0]; } @@ -336,7 +360,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) template<> EIGEN_STRONG_INLINE std::complex predux(const Packet1cd& a) { Packet2d b; - b = (Packet2d) vec_sld(a.v, a.v, 8); + b = (Packet2d) vec_sld((Packet4ui) a.v, (Packet4ui) a.v, 8); b = padd(a.v, b); return pfirst(Packet1cd(b)); } @@ -345,9 +369,9 @@ template<> EIGEN_STRONG_INLINE Packet1cd preduxp(const Packet1cd* vec { Packet2d b1, b2; - b1 = (Packet2d) vec_sld(vecs[0].v, vecs[1].v, 8); - b2 = (Packet2d) vec_sld(vecs[1].v, vecs[0].v, 8); - b2 = (Packet2d) vec_sld(b2, b2, 8); + b1 = (Packet2d) vec_sld((Packet4ui) vecs[0].v, (Packet4ui) vecs[1].v, 8); + b2 = (Packet2d) vec_sld((Packet4ui) vecs[1].v, (Packet4ui) vecs[0].v, 8); + b2 = (Packet2d) vec_sld((Packet4ui) b2, (Packet4ui) b2, 8); b2 = padd(b1, b2); return Packet1cd(b2); @@ -357,7 +381,7 @@ template<> EIGEN_STRONG_INLINE std::complex predux_mul(const { Packet2d b; Packet1cd prod; - b = (Packet2d) vec_sld(a.v, a.v, 8); + b = (Packet2d) vec_sld((Packet4ui) a.v, (Packet4ui) a.v, 8); prod = pmul(a, Packet1cd(b)); return pfirst(prod); @@ -407,7 +431,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, con { // TODO optimize it for AltiVec Packet1cd res = conj_helper().pmul(a,b); - Packet2d s = vec_madd(b.v, b.v, p4f_ZERO); + Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_); return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV)))); } diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index e70039ae2..0489ad886 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -31,12 +31,6 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 #endif -#ifdef __VSX__ -typedef __vector double Packet2d; -typedef __vector unsigned long Packet2ul; -typedef __vector long Packet2l; -#endif // __VSX__ - typedef __vector float Packet4f; typedef __vector int Packet4i; typedef __vector unsigned int Packet4ui; @@ -82,8 +76,6 @@ static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 }; static Packet16uc p16uc_REVERSE = { 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15 }; static Packet16uc p16uc_FORWARD = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; static Packet16uc p16uc_DUPLICATE = { 4,5,6,7, 4,5,6,7, 0,1,2,3, 0,1,2,3 }; -static Packet2d p2d_ZERO_ = (Packet2d) { 0x8000000000000000, 0x8000000000000000 }; -static Packet2l p2l_ZERO = (Packet2l) { 0x0, 0x0 }; #endif // _BIG_ENDIAN // These constants are endian-agnostic @@ -114,24 +106,6 @@ template<> struct packet_traits : default_packet_traits HasSqrt = 0 }; }; -#ifdef __VSX__ -template<> struct packet_traits : default_packet_traits -{ - typedef Packet2d type; - typedef Packet2d half; - enum { - Vectorizable = 1, - AlignedOnScalar = 1, - size=2, - HasHalfPacket = 0, - - HasDiv = 1, - HasExp = 0, - HasSqrt = 0 - }; -}; -#endif // __VSX__ - template<> struct packet_traits : default_packet_traits { typedef Packet4i type; @@ -144,9 +118,6 @@ template<> struct packet_traits : default_packet_traits }; }; -#ifdef __VSX__ -template<> struct unpacket_traits { typedef double type; enum {size=2}; typedef Packet2d half; }; -#endif // __VSX__ template<> struct unpacket_traits { typedef float type; enum {size=4}; typedef Packet4f half; }; template<> struct unpacket_traits { typedef int type; enum {size=4}; typedef Packet4i half; }; @@ -196,11 +167,18 @@ inline std::ostream & operator <<(std::ostream & s, const Packetbi & v) } */ +// Need to define them first or we get specialization after instantiation errors +template<> EIGEN_STRONG_INLINE Packet4f pload(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } +template<> EIGEN_STRONG_INLINE Packet4i pload(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } + +template<> EIGEN_STRONG_INLINE void pstore(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } +template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } + template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html float EIGEN_ALIGN16 af[4]; af[0] = from; - Packet4f vc = vec_ld(0, af); + Packet4f vc = pload(af); vc = vec_splat(vc, 0); return vc; } @@ -208,17 +186,15 @@ template<> EIGEN_STRONG_INLINE Packet4f pset1(const float& from) { template<> EIGEN_STRONG_INLINE Packet4i pset1(const int& from) { int EIGEN_ALIGN16 ai[4]; ai[0] = from; - Packet4i vc = vec_ld(0, ai); + Packet4i vc = pload(ai); vc = vec_splat(vc, 0); return vc; } - - template<> EIGEN_STRONG_INLINE void pbroadcast4(const float *a, Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) { - a3 = vec_ld(0,a); + a3 = pload(a); a0 = vec_splat(a3, 0); a1 = vec_splat(a3, 1); a2 = vec_splat(a3, 2); @@ -228,7 +204,7 @@ template<> EIGEN_STRONG_INLINE void pbroadcast4(const int *a, Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3) { - a3 = vec_ld(0,a); + a3 = pload(a); a0 = vec_splat(a3, 0); a1 = vec_splat(a3, 1); a2 = vec_splat(a3, 2); @@ -242,7 +218,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const floa af[1] = from[1*stride]; af[2] = from[2*stride]; af[3] = from[3*stride]; - return vec_ld(0, af); + return pload(af); } template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, DenseIndex stride) { @@ -251,12 +227,12 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* f ai[1] = from[1*stride]; ai[2] = from[2*stride]; ai[3] = from[3*stride]; - return vec_ld(0, ai); + return pload(ai); } template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, DenseIndex stride) { float EIGEN_ALIGN16 af[4]; - vec_st(from, 0, af); + pstore(af, from); to[0*stride] = af[0]; to[1*stride] = af[1]; to[2*stride] = af[2]; @@ -265,7 +241,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, co template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, DenseIndex stride) { int EIGEN_ALIGN16 ai[4]; - vec_st(from, 0, ai); + pstore((int *)ai, from); to[0*stride] = ai[0]; to[1*stride] = ai[1]; to[2*stride] = ai[2]; @@ -366,10 +342,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pxor(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pandnot(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); } template<> EIGEN_STRONG_INLINE Packet4i pandnot(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); } -template<> EIGEN_STRONG_INLINE Packet4f pload(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } -template<> EIGEN_STRONG_INLINE Packet4i pload(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } - -#ifndef __VSX__ +#ifdef _BIG_ENDIAN template<> EIGEN_STRONG_INLINE Packet4f ploadu(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD @@ -393,6 +366,7 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data } #else +// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD @@ -420,9 +394,6 @@ template<> EIGEN_STRONG_INLINE Packet4i ploaddup(const int* from) return vec_perm(p, p, p16uc_DUPLICATE); } -template<> EIGEN_STRONG_INLINE void pstore(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } -template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } - template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE @@ -642,6 +613,252 @@ ptranspose(PacketBlock& kernel) { kernel.packet[3] = vec_mergel(t1, t3); } + +//---------- double ---------- +#ifdef __VSX__ +typedef __vector double Packet2d; +typedef __vector unsigned long long Packet2ul; +typedef __vector long long Packet2l; + +static Packet2l p2l_ZERO = { 0x0, 0x0 }; +static Packet2d p2l_ONE = { 1, 1 }; +static Packet2d p2d_ONE = { 1.0, 1.0 }; +static Packet2d p2d_ZERO = { 0.0, 0.0 }; +static Packet2d p2d_ZERO_ = { (double) 0x8000000000000000, (double) 0x8000000000000000 }; +static Packet2d p2d_COUNTDOWN = { 1.0, 0.0 }; + +#ifdef _BIG_ENDIAN +static Packet16uc p16uc_SPLATQ1 = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_SPLATQ2 = { 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; +#else +static Packet16uc p16uc_SPLATQ1 = { 4,5,6,7, 0,1,2,3, 4,5,6,7, 0,1,3,4 }; +static Packet16uc p16uc_SPLATQ2 = { 12,13,14,15, 8,9,10,11, 12,13,14,15, 8,9,10,11 }; +#endif + + +static EIGEN_STRONG_INLINE Packet2d vec_splat_dbl(Packet2d& a, int index) +{ + switch (index) { + case 0: + return (Packet2d) vec_perm(a, a, p16uc_SPLATQ1); + case 1: + return (Packet2d) vec_perm(a, a, p16uc_SPLATQ2); + } + return a; +} + +template<> struct packet_traits : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 0, + + HasDiv = 1, + HasExp = 0, + HasSqrt = 0 + }; +}; + +template<> struct unpacket_traits { typedef double type; enum {size=2}; typedef Packet2d half; }; + +/* +inline std::ostream & operator <<(std::ostream & s, const Packet2d & v) +{ + union { + Packetdf v; + double n[2]; + } vt; + vt.v = v; + s << vt.n[0] << ", " << vt.n[1]; + return s; +}*/ + +// Need to define them first or we get specialization after instantiation errors +template<> EIGEN_STRONG_INLINE Packet2d pload(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return (Packet2d) vec_ld(0, (const float *) from); } //FIXME + +template<> EIGEN_STRONG_INLINE void pstore(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st((Packet4f)from, 0, (float *)to); } + +template<> EIGEN_STRONG_INLINE Packet2d pset1(const double& from) { + double EIGEN_ALIGN16 af[2]; + af[0] = from; + Packet2d vc = pload(af); + vc = vec_splat_dbl(vc, 0); + return vc; +} +template<> EIGEN_STRONG_INLINE void +pbroadcast4(const double *a, + Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) +{ + a1 = pload(a); + a0 = vec_splat_dbl(a1, 0); + a1 = vec_splat_dbl(a1, 1); + a3 = pload(a+2); + a2 = vec_splat_dbl(a3, 0); + a3 = vec_splat_dbl(a1, 1); +} +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, DenseIndex stride) +{ + double EIGEN_ALIGN16 af[2]; + af[0] = from[0*stride]; + af[1] = from[1*stride]; + return pload(af); +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, DenseIndex stride) +{ + double EIGEN_ALIGN16 af[2]; + pstore(af, from); + to[0*stride] = af[0]; + to[1*stride] = af[1]; +} +template<> EIGEN_STRONG_INLINE Packet2d plset(const double& a) { return vec_add(pset1(a), p2d_COUNTDOWN); } + +template<> EIGEN_STRONG_INLINE Packet2d padd(const Packet2d& a, const Packet2d& b) { return vec_add(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d psub(const Packet2d& a, const Packet2d& b) { return vec_sub(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return psub(p2d_ZERO, a); } + +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet2d pmul(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); } +template<> EIGEN_STRONG_INLINE Packet2d pdiv(const Packet2d& a, const Packet2d& b) +{ + Packet2d t, y_0, y_1, res; + + // Altivec does not offer a divide instruction, we have to do a reciprocal approximation + y_0 = vec_re(b); + + // Do one Newton-Raphson iteration to get the needed accuracy + t = vec_nmsub(y_0, b, p2d_ONE); + y_1 = vec_madd(y_0, t, y_0); + + res = vec_madd(a, y_1, p2d_ZERO); + return res; +} + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } + +template<> EIGEN_STRONG_INLINE Packet2d pmin(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pmax(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pand(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d por(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pxor(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); } + +template<> EIGEN_STRONG_INLINE Packet2d pandnot(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); } + + +template<> EIGEN_STRONG_INLINE Packet2d ploadu(const double* from) +{ + EIGEN_DEBUG_ALIGNED_LOAD + return (Packet2d) vec_vsx_ld((int)((size_t)(from) & 15), (const float *)from); // align the data +} +template<> EIGEN_STRONG_INLINE Packet2d ploaddup(const double* from) +{ + Packet2d p; + if((ptrdiff_t(from) % 16) == 0) p = pload(from); + else p = ploadu(from); + return vec_perm(p, p, p16uc_DUPLICATE); +} + +template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& from) +{ + EIGEN_DEBUG_UNALIGNED_STORE + // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html + // Warning: not thread safe! + Packet16uc MSQ, LSQ, edges; + Packet16uc edgeAlign, align; + + MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword + LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword + edgeAlign = vec_lvsl(0, to); // permute map to extract edges + edges=vec_perm(LSQ,MSQ,edgeAlign); // extract the edges + align = vec_lvsr( 0, to ); // permute map to misalign data + MSQ = vec_perm(edges,(Packet16uc)from,align); // misalign the data (MSQ) + LSQ = vec_perm((Packet16uc)from,edges,align); // misalign the data (LSQ) + vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first + vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part +} +template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { vec_dstt((const float *) addr, DST_CTRL(2,2,32), DST_CHAN); } + +template<> EIGEN_STRONG_INLINE double pfirst(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; } + +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return (Packet2d)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } + +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); } + +template<> EIGEN_STRONG_INLINE double predux(const Packet2d& a) +{ + Packet2d b, sum; + b = (Packet2d) vec_sld((Packet4ui) a, (Packet4ui)a, 8); + sum = vec_add(a, b); + return pfirst(sum); +} + +template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) +{ + Packet2d v[2], sum; + + v[0] = (Packet2d) vec_sld((Packet4ui) vecs[0], (Packet4ui) vecs[1], 8); + v[1] = (Packet2d) vec_sld((Packet4ui) vecs[1], (Packet4ui) vecs[0], 8); + sum = vec_add(v[0], v[1]); + + return sum; +} +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE double predux_mul(const Packet2d& a) +{ + Packet2d prod; + prod = pmul(a, (Packet2d)vec_sld((Packet4ui) a, (Packet4ui)a, 8)); + return pfirst(pmul(prod, (Packet2d)vec_sld((Packet4ui) prod, (Packet4ui) prod, 4))); +} + +// min +template<> EIGEN_STRONG_INLINE double predux_min(const Packet2d& a) +{ + Packet2d b, res; + b = vec_min(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)); + res = vec_min(b, (Packet2d) vec_sld((Packet4ui) b, (Packet4ui) b, 4)); + return pfirst(res); +} + +// max +template<> EIGEN_STRONG_INLINE double predux_max(const Packet2d& a) +{ + Packet2d res; + res = vec_max(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)); + return pfirst(res); +} + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) + { + if (Offset!=0) + first = (Packet2d) vec_sld((Packet4ui) first, (Packet4ui) second, Offset*8); + } +}; + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { + Packet2d t0, t1; + t0 = (Packet2d) vec_sld((Packet4ui) kernel.packet[0], (Packet4ui) kernel.packet[1], 8); + t1 = (Packet2d) vec_sld((Packet4ui) kernel.packet[1], (Packet4ui) kernel.packet[0], 8); + kernel.packet[0] = t0; + kernel.packet[1] = t1; +} + +#endif // __VSX__ } // end namespace internal } // end namespace Eigen From 9452eb38f812194a676edc1b9eb9d08b7bc0f297 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Fri, 12 Sep 2014 14:52:35 +0100 Subject: [PATCH 0779/1103] Make UpperBidiagonalization accept row-major matrices (bug #769) * Give temporary workspace the same storage order as original matrix * Take storage order into account when determining inner stride of rows and columns * Change one test to use a row-major matrix. --- Eigen/src/SVD/UpperBidiagonalization.h | 29 +++++++++++++++++++------- test/upperbidiagonalization.cpp | 2 +- 2 files changed, 23 insertions(+), 8 deletions(-) diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 225b19e3c..64906bf0c 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -154,14 +154,19 @@ void upperbidiagonalization_blocked_helper(MatrixType& A, typename MatrixType::RealScalar *diagonal, typename MatrixType::RealScalar *upper_diagonal, typename MatrixType::Index bs, - Ref > X, - Ref > Y) + Ref::Flags & RowMajorBit> > X, + Ref::Flags & RowMajorBit> > Y) { typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; - typedef Ref > SubColumnType; - typedef Ref, 0, InnerStride<> > SubRowType; - typedef Ref > SubMatType; + enum { StorageOrder = traits::Flags & RowMajorBit }; + typedef InnerStride ColInnerStride; + typedef InnerStride RowInnerStride; + typedef Ref, 0, ColInnerStride> SubColumnType; + typedef Ref, 0, RowInnerStride> SubRowType; + typedef Ref > SubMatType; Index brows = A.rows(); Index bcols = A.cols(); @@ -288,8 +293,18 @@ void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagona Index cols = A.cols(); Index size = (std::min)(rows, cols); - Matrix X(rows,maxBlockSize); - Matrix Y(cols,maxBlockSize); + // X and Y are work space + enum { StorageOrder = traits::Flags & RowMajorBit }; + Matrix X(rows,maxBlockSize); + Matrix Y(cols,maxBlockSize); Index blockSize = (std::min)(maxBlockSize,size); Index k = 0; diff --git a/test/upperbidiagonalization.cpp b/test/upperbidiagonalization.cpp index d15bf588b..847b34b55 100644 --- a/test/upperbidiagonalization.cpp +++ b/test/upperbidiagonalization.cpp @@ -35,7 +35,7 @@ void test_upperbidiagonalization() CALL_SUBTEST_1( upperbidiag(MatrixXf(3,3)) ); CALL_SUBTEST_2( upperbidiag(MatrixXd(17,12)) ); CALL_SUBTEST_3( upperbidiag(MatrixXcf(20,20)) ); - CALL_SUBTEST_4( upperbidiag(MatrixXcd(16,15)) ); + CALL_SUBTEST_4( upperbidiag(Matrix,Dynamic,Dynamic,RowMajor>(16,15)) ); CALL_SUBTEST_5( upperbidiag(Matrix()) ); CALL_SUBTEST_6( upperbidiag(Matrix()) ); CALL_SUBTEST_7( upperbidiag(Matrix()) ); From af9c9f7706bb7701a7224827e981ecc2f3bd9ac7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 14 Sep 2014 17:33:39 +0200 Subject: [PATCH 0780/1103] Fix comparison to block size --- Eigen/src/Householder/HouseholderSequence.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h index ab32f7d3f..0c6a09861 100644 --- a/Eigen/src/Householder/HouseholderSequence.h +++ b/Eigen/src/Householder/HouseholderSequence.h @@ -321,7 +321,7 @@ template class HouseholderS { const Index BlockSize = 48; // if the entries are large enough, then apply the reflectors by block - if(m_length>BlockSize && dst.cols()>1) + if(m_length>=BlockSize && dst.cols()>1) { for(Index i = 0; i < m_length; i+=BlockSize) { From dfc54e1bbf99a069146025677b70e38755383df9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 14 Sep 2014 18:27:48 +0200 Subject: [PATCH 0781/1103] Fix /= when using evaluator as in changeset 2d90484450f3934db3f5db39ef37967fb9444263 --- Eigen/src/Core/SelfCwiseBinaryOp.h | 12 +----------- test/linearstructure.cpp | 16 +++++++++++++++- 2 files changed, 16 insertions(+), 12 deletions(-) diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h index 87fcde323..bec6f4968 100644 --- a/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -212,17 +212,7 @@ template inline Derived& DenseBase::operator/=(const Scalar& other) { typedef typename Derived::PlainObject PlainObject; - - typedef typename internal::conditional::IsInteger, - internal::div_assign_op, - internal::mul_assign_op >::type AssignOp; - - Scalar actual_other; - if(NumTraits::IsInteger) actual_other = other; - else actual_other = Scalar(1)/other; - - internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),actual_other), AssignOp()); - + internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op()); return derived(); } #else diff --git a/test/linearstructure.cpp b/test/linearstructure.cpp index b627915ce..87dfa1b6b 100644 --- a/test/linearstructure.cpp +++ b/test/linearstructure.cpp @@ -9,7 +9,6 @@ // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. static bool g_called; - #define EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN { g_called = true; } #include "main.h" @@ -107,4 +106,19 @@ void test_linearstructure() CALL_SUBTEST_10( real_complex() ); CALL_SUBTEST_10( real_complex(10,10) ); } + +#ifdef EIGEN_TEST_PART_4 + { + // make sure that /=scalar and /scalar do not overflow + // rational: 1.0/4.94e-320 overflow, but m/4.94e-320 should not + Matrix4d m2, m3; + m3 = m2 = Matrix4d::Random()*1e-20; + m2 = m2 / 4.9e-320; + VERIFY_IS_APPROX(m2.cwiseQuotient(m2), Matrix4d::Ones()); + m3 /= 4.9e-320; + VERIFY_IS_APPROX(m3.cwiseQuotient(m3), Matrix4d::Ones()); + + + } +#endif } From fda680f9cf8e81f62ae815b700698958b7b0f027 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 14 Sep 2014 18:31:29 +0200 Subject: [PATCH 0782/1103] Adapt changeset 51b3f558bb76c11149fc64971db786798f1b692c to evaluators: (Fix bug #822: outer products needed linear access, and add respective unit tests) --- Eigen/src/Core/ProductEvaluators.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 8a63384a7..f880e7696 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -215,7 +215,7 @@ EIGEN_DONT_INLINE void outer_product_selector_run(Dst& dst, const Lhs &lhs, cons // FIXME we should probably build an evaluator for dst and rhs const Index cols = dst.cols(); for (Index j=0; j From 26db954776c0e8f0230ca1542ad303e23ede4db1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 14 Sep 2014 19:06:08 +0200 Subject: [PATCH 0783/1103] Re-enable aliasing checks when using evaluators --- Eigen/src/Core/Assign.h | 7 ++++--- Eigen/src/Core/AssignEvaluator.h | 7 +++++++ Eigen/src/Core/DenseBase.h | 2 ++ Eigen/src/Core/Transpose.h | 10 +++++++++- 4 files changed, 22 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index 071cfbf7e..5395e5436 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -521,12 +521,13 @@ EIGEN_STRONG_INLINE Derived& DenseBase eigen_assert(rows() == other.rows() && cols() == other.cols()); internal::assign_impl::Traversal) : int(InvalidTraversal)>::run(derived(),other.derived()); - -#endif // EIGEN_TEST_EVALUATORS - + #ifndef EIGEN_NO_DEBUG checkTransposeAliasing(other.derived()); #endif + +#endif // EIGEN_TEST_EVALUATORS + return derived(); } diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 1727a6a4a..521b8a93a 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -763,6 +763,9 @@ void call_assignment_no_alias(Dst& dst, const Src& src) call_assignment_no_alias(dst, src, internal::assign_op()); } +// forxard declaration +template void check_for_aliasing(const Dst &dst, const Src &src); + // Generic Dense to Dense assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> struct Assignment @@ -771,6 +774,10 @@ struct Assignment { eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); +#ifndef EIGEN_NO_DEBUG + internal::check_for_aliasing(dst, src); +#endif + call_dense_assignment_loop(dst, src, func); } }; diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 14643b5a8..f35c2edc4 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -314,11 +314,13 @@ template class DenseBase EIGEN_DEVICE_FUNC void transposeInPlace(); #ifndef EIGEN_NO_DEBUG +#ifndef EIGEN_TEST_EVALUATORS protected: template void checkTransposeAliasing(const OtherDerived& other) const; public: #endif +#endif EIGEN_DEVICE_FUNC static const ConstantReturnType diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index f5148221d..b22b50a8f 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -437,15 +437,23 @@ struct checkTransposeAliasing_impl } }; +template +void check_for_aliasing(const Dst &dst, const Src &src) +{ + internal::checkTransposeAliasing_impl::run(dst, src); +} + } // end namespace internal +#ifndef EIGEN_TEST_EVALUATORS template template void DenseBase::checkTransposeAliasing(const OtherDerived& other) const { internal::checkTransposeAliasing_impl::run(derived(), other); } -#endif +#endif // EIGEN_TEST_EVALUATORS +#endif // EIGEN_NO_DEBUG } // end namespace Eigen From c83e01f2d6113fb2f7dcc591b3364eb87b7a74aa Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 14 Sep 2014 19:38:49 +0200 Subject: [PATCH 0784/1103] Favor column major storage for inner products --- Eigen/src/Core/AssignEvaluator.h | 4 ++-- Eigen/src/Core/Product.h | 2 +- test/vectorization_logic.cpp | 16 +++++++++------- 3 files changed, 12 insertions(+), 10 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 521b8a93a..4e35e432e 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -55,8 +55,8 @@ private: }; enum { - DstIsRowMajor = DstEvaluator::Flags&RowMajorBit, - SrcIsRowMajor = SrcEvaluator::Flags&RowMajorBit, + DstIsRowMajor = DstFlags&RowMajorBit, + SrcIsRowMajor = SrcFlags&RowMajorBit, StorageOrdersAgree = (int(DstIsRowMajor) == int(SrcIsRowMajor)), MightVectorize = StorageOrdersAgree && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit) diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 0cf20f2e2..6825873d5 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -85,7 +85,7 @@ struct traits > #endif // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator. - Flags = ( MaxRowsAtCompileTime==1 + Flags = ( (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) || ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit)) || ((RhsTraits::Flags&NoPreferredStorageOrderBit) && (LhsTraits::Flags&RowMajorBit)) ) ? RowMajorBit : (MaxColsAtCompileTime==1 ? 0 : NoPreferredStorageOrderBit) diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp index 42015e21b..303eb6cf0 100644 --- a/test/vectorization_logic.cpp +++ b/test/vectorization_logic.cpp @@ -30,13 +30,15 @@ std::string demangle_unrolling(int t) std::string demangle_flags(int f) { std::string res; - if(f&RowMajorBit) res += " | RowMajor"; - if(f&PacketAccessBit) res += " | Packet"; - if(f&LinearAccessBit) res += " | Linear"; - if(f&LvalueBit) res += " | Lvalue"; - if(f&DirectAccessBit) res += " | Direct"; - if(f&AlignedBit) res += " | Aligned"; - if(f&NestByRefBit) res += " | NestByRef"; + if(f&RowMajorBit) res += " | RowMajor"; + if(f&PacketAccessBit) res += " | Packet"; + if(f&LinearAccessBit) res += " | Linear"; + if(f&LvalueBit) res += " | Lvalue"; + if(f&DirectAccessBit) res += " | Direct"; + if(f&AlignedBit) res += " | Aligned"; + if(f&NestByRefBit) res += " | NestByRef"; + if(f&NoPreferredStorageOrderBit) res += " | NoPreferredStorageOrderBit"; + return res; } From 0403d49006818079c79c038a6b55c48bc3839c22 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 14 Sep 2014 20:12:07 +0200 Subject: [PATCH 0785/1103] Fix inverse unit test making sure we try to invert an invertible matrix --- test/inverse.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/test/inverse.cpp b/test/inverse.cpp index 1195bcc76..1e7b20958 100644 --- a/test/inverse.cpp +++ b/test/inverse.cpp @@ -73,6 +73,7 @@ template void inverse(const MatrixType& m) { Matrix m3; m3.setRandom(); + m3.topLeftCorner(rows,rows) = m1; m2 = m3.template topLeftCorner().inverse(); VERIFY_IS_APPROX( (m3.template topLeftCorner()), m2.inverse() ); } From 8514179aa31a6a42523c66f66a94723ed758cd5f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 15 Sep 2014 13:53:52 +0200 Subject: [PATCH 0786/1103] Fix traits::IsAligned when using evaluators --- Eigen/src/Geometry/Quaternion.h | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index 11e5398d4..850940302 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -217,7 +217,11 @@ struct traits > typedef _Scalar Scalar; typedef Matrix<_Scalar,4,1,_Options> Coefficients; enum{ +#ifndef EIGEN_TEST_EVALUATORS IsAligned = internal::traits::Flags & AlignedBit, +#else + IsAligned = (internal::traits::EvaluatorFlags & AlignedBit) != 0, +#endif Flags = IsAligned ? (AlignedBit | LvalueBit) : LvalueBit }; }; From 466d6d41c6dfda5a2814d0ee6d02439ccda4c8d0 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 15 Sep 2014 17:40:17 +0200 Subject: [PATCH 0787/1103] Avoid a potential risk of recursive definition using traits to get he scalar type --- Eigen/src/Core/CoreEvaluators.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index b8dc05d98..09a83a382 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -49,8 +49,8 @@ template<> struct storage_kind_to_shape { typedef DenseShape Shape; }; template< typename T, typename LhsKind = typename evaluator_traits::Kind, typename RhsKind = typename evaluator_traits::Kind, - typename LhsScalar = typename T::Lhs::Scalar, - typename RhsScalar = typename T::Rhs::Scalar> struct binary_evaluator; + typename LhsScalar = typename traits::Scalar, + typename RhsScalar = typename traits::Scalar> struct binary_evaluator; template< typename T, typename Kind = typename evaluator_traits::Kind, From 486ca277a002b888a13fe14b136ff54b3979d475 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 16 Sep 2014 10:29:29 -0700 Subject: [PATCH 0788/1103] Workaround MSVC ICE --- Eigen/src/Core/TriangularMatrix.h | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 0383ca9f5..cc585bc6c 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -277,7 +277,16 @@ template class TriangularView solve(const MatrixBase& other) const { return Solve(*this, other.derived()); } + // workaround MSVC ICE + #ifdef _MSC_VER + template + EIGEN_DEVICE_FUNC + inline const internal::triangular_solve_retval + solve(const MatrixBase& other) const + { return Base::template solve(other); } + #else using Base::solve; + #endif #endif // EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC @@ -575,7 +584,7 @@ template class TriangularViewImpl<_Mat EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const { if(!(internal::is_same::value && internal::extract_data(dst) == internal::extract_data(rhs))) dst = rhs; - this->template solveInPlace(dst); + this->solveInPlace(dst); } template From 0f0580b97c4209faeb01f415523fab5863e742b1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 17 Sep 2014 09:55:44 +0200 Subject: [PATCH 0789/1103] Remove not needed template keyword. --- Eigen/src/SparseCore/SparseTriangularView.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index 87f4ab18d..3df7a4cd4 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -50,7 +50,7 @@ protected: EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const { if(!(internal::is_same::value && internal::extract_data(dst) == internal::extract_data(rhs))) dst = rhs; - this->template solveInPlace(dst); + this->solveInPlace(dst); } #endif // EIGEN_TEST_EVALUATORS From fc23e937076d01e565fac5b771f22099dee5ad6c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 17 Sep 2014 09:56:07 +0200 Subject: [PATCH 0790/1103] Add a portable log2 function for integers --- Eigen/src/Core/MathFunctions.h | 15 +++++++++++++++ .../SparseCore/ConservativeSparseSparseProduct.h | 2 +- 2 files changed, 16 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index e9fed2e52..8834f71fa 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -694,6 +694,21 @@ bool (isfinite)(const std::complex& x) return isfinite(real(x)) && isfinite(imag(x)); } +// Log base 2 for 32 bits positive integers. +// Conveniently returns 0 for x==0. +int log2(int x) +{ + eigen_assert(x>=0); + unsigned int v(x); + static const int table[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 }; + v |= v >> 1; + v |= v >> 2; + v |= v >> 4; + v |= v >> 8; + v |= v >> 16; + return table[(v * 0x07C4ACDDU) >> 27]; +} + } // end namespace numext namespace internal { diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index ae8fc75e5..398008597 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -101,7 +101,7 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r // otherwise => loop through the entire vector // In order to avoid to perform an expensive log2 when the // result is clearly very sparse we use a linear bound up to 200. - if((nnz<200 && nnz1) std::sort(indices,indices+nnz); for(Index k=0; k Date: Tue, 16 Sep 2014 16:05:06 -0700 Subject: [PATCH 0791/1103] avoid division by 0 --- test/stable_norm.cpp | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/test/stable_norm.cpp b/test/stable_norm.cpp index f76919af4..6cd65c64a 100644 --- a/test/stable_norm.cpp +++ b/test/stable_norm.cpp @@ -130,7 +130,7 @@ template void stable_norm(const MatrixType& m) // NaN { v = vrand; - v(i,j) = RealScalar(0)/RealScalar(0); + v(i,j) = std::numeric_limits::quiet_NaN(); VERIFY(!isFinite(v.squaredNorm())); VERIFY(isNaN(v.squaredNorm())); VERIFY(!isFinite(v.norm())); VERIFY(isNaN(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isNaN(v.stableNorm())); @@ -141,7 +141,7 @@ template void stable_norm(const MatrixType& m) // +inf { v = vrand; - v(i,j) = RealScalar(1)/RealScalar(0); + v(i,j) = std::numeric_limits::infinity(); VERIFY(!isFinite(v.squaredNorm())); VERIFY(isInf(v.squaredNorm())); VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); @@ -152,7 +152,7 @@ template void stable_norm(const MatrixType& m) // -inf { v = vrand; - v(i,j) = RealScalar(-1)/RealScalar(0); + v(i,j) = -std::numeric_limits::infinity(); VERIFY(!isFinite(v.squaredNorm())); VERIFY(isInf(v.squaredNorm())); VERIFY(!isFinite(v.norm())); VERIFY(isInf(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isInf(v.stableNorm())); @@ -165,8 +165,8 @@ template void stable_norm(const MatrixType& m) Index i2 = internal::random(0,rows-1); Index j2 = internal::random(0,cols-1); v = vrand; - v(i,j) = RealScalar(-1)/RealScalar(0); - v(i2,j2) = RealScalar(0)/RealScalar(0); + v(i,j) = -std::numeric_limits::infinity(); + v(i2,j2) = std::numeric_limits::quiet_NaN(); VERIFY(!isFinite(v.squaredNorm())); VERIFY(isNaN(v.squaredNorm())); VERIFY(!isFinite(v.norm())); VERIFY(isNaN(v.norm())); VERIFY(!isFinite(v.stableNorm())); VERIFY(isNaN(v.stableNorm())); From 125619146bc3db3074c8f7632869c7f85b857b8a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 16 Sep 2014 16:06:32 -0700 Subject: [PATCH 0792/1103] workaround weird MSVC compilation issue: a typdedef in a base class shadows a template parameter of a derived class --- Eigen/src/SparseCore/SparseDenseProduct.h | 37 +++++++++++------------ 1 file changed, 18 insertions(+), 19 deletions(-) diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index c8a515e8c..803d98e2d 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -402,30 +402,30 @@ struct generic_product_impl } }; -template +template struct sparse_dense_outer_product_evaluator { protected: - typedef typename conditional::type Lhs1; - typedef typename conditional::type Rhs; - typedef Product ProdXprType; + typedef typename conditional::type Lhs1; + typedef typename conditional::type ActualRhs; + typedef Product ProdXprType; // if the actual left-hand side is a dense vector, // then build a sparse-view so that we can seamlessly iterator over it. typedef typename conditional::StorageKind,Sparse>::value, - Lhs1, SparseView >::type Lhs; + Lhs1, SparseView >::type ActualLhs; typedef typename conditional::StorageKind,Sparse>::value, Lhs1 const&, SparseView >::type LhsArg; - typedef typename evaluator::type LhsEval; - typedef typename evaluator::type RhsEval; - typedef typename evaluator::InnerIterator LhsIterator; + typedef typename evaluator::type LhsEval; + typedef typename evaluator::type RhsEval; + typedef typename evaluator::InnerIterator LhsIterator; typedef typename ProdXprType::Scalar Scalar; typedef typename ProdXprType::Index Index; public: enum { - Flags = Transpose ? RowMajorBit : 0, + Flags = NeedToTranspose ? RowMajorBit : 0, CoeffReadCost = Dynamic }; @@ -436,17 +436,16 @@ public: : LhsIterator(xprEval.m_lhsXprImpl, 0), m_outer(outer), m_empty(false), - m_factor(get(xprEval.m_rhsXprImpl, outer, typename internal::traits::StorageKind() )) + m_factor(get(xprEval.m_rhsXprImpl, outer, typename internal::traits::StorageKind() )) {} EIGEN_STRONG_INLINE Index outer() const { return m_outer; } - EIGEN_STRONG_INLINE Index row() const { return Transpose ? m_outer : LhsIterator::index(); } - EIGEN_STRONG_INLINE Index col() const { return Transpose ? LhsIterator::index() : m_outer; } + EIGEN_STRONG_INLINE Index row() const { return NeedToTranspose ? m_outer : LhsIterator::index(); } + EIGEN_STRONG_INLINE Index col() const { return NeedToTranspose ? LhsIterator::index() : m_outer; } EIGEN_STRONG_INLINE Scalar value() const { return LhsIterator::value() * m_factor; } EIGEN_STRONG_INLINE operator bool() const { return LhsIterator::operator bool() && (!m_empty); } - protected: Scalar get(const RhsEval &rhs, Index outer, Dense = Dense()) const { @@ -467,19 +466,19 @@ public: Scalar m_factor; }; - sparse_dense_outer_product_evaluator(const Lhs &lhs, const Rhs &rhs) - : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs) + sparse_dense_outer_product_evaluator(const ActualLhs &lhs, const ActualRhs &rhs) + : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs) {} // transpose case - sparse_dense_outer_product_evaluator(const Rhs &rhs, const Lhs1 &lhs) - : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs) + sparse_dense_outer_product_evaluator(const ActualRhs &rhs, const Lhs1 &lhs) + : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs) {} protected: const LhsArg m_lhs; - typename evaluator::nestedType m_lhsXprImpl; - typename evaluator::nestedType m_rhsXprImpl; + typename evaluator::nestedType m_lhsXprImpl; + typename evaluator::nestedType m_rhsXprImpl; }; // sparse * dense outer product From c2f66c65aaef44c698ba141a08b5940baf4c56c8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 16 Sep 2014 16:23:45 -0700 Subject: [PATCH 0793/1103] workaround MSVC compilation issue (shadow issue) --- Eigen/src/Core/Inverse.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 1d167867f..84abd258e 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -101,11 +101,11 @@ namespace internal { * * \sa class Inverse */ -template -struct unary_evaluator > - : public evaluator::PlainObject>::type +template +struct unary_evaluator > + : public evaluator::PlainObject>::type { - typedef Inverse InverseType; + typedef Inverse InverseType; typedef typename InverseType::PlainObject PlainObject; typedef typename evaluator::type Base; From e44d78dab30809f359043b8d09787b7435526671 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 16 Sep 2014 17:10:25 -0700 Subject: [PATCH 0794/1103] workaround ambiguous call --- test/product_mmtr.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/test/product_mmtr.cpp b/test/product_mmtr.cpp index 7d6746800..92e6b668f 100644 --- a/test/product_mmtr.cpp +++ b/test/product_mmtr.cpp @@ -13,7 +13,8 @@ ref2 = ref1 = DEST; \ DEST.template triangularView() OP; \ ref1 OP; \ - ref2.template triangularView() = ref1; \ + ref2.template triangularView() \ + = ref1.template triangularView(); \ VERIFY_IS_APPROX(DEST,ref2); \ } From 0bf589486132e011a40d0a657a719980562b026f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 16 Sep 2014 18:21:39 -0700 Subject: [PATCH 0795/1103] workaround one more shadowing issue with MSVC --- Eigen/src/Core/Transpose.h | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index b22b50a8f..dd6180a8f 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -29,9 +29,10 @@ namespace Eigen { namespace internal { template -struct traits > : traits +struct traits > { - typedef typename MatrixType::Scalar Scalar; + typedef typename traits::Scalar Scalar; + typedef typename traits::Index Index; typedef typename nested::type MatrixTypeNested; typedef typename remove_reference::type MatrixTypeNestedPlain; typedef typename traits::StorageKind StorageKind; From 8b3be4907da2cbf47ec15734b7f364e6e66bf4c3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 18 Sep 2014 10:53:53 +0200 Subject: [PATCH 0796/1103] log2(int) must be inlined. --- Eigen/src/Core/MathFunctions.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 8834f71fa..73859b0ee 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -696,7 +696,7 @@ bool (isfinite)(const std::complex& x) // Log base 2 for 32 bits positive integers. // Conveniently returns 0 for x==0. -int log2(int x) +inline int log2(int x) { eigen_assert(x>=0); unsigned int v(x); From 0ca43f7e9a654e32da0066163a8656415961e266 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 18 Sep 2014 15:15:27 +0200 Subject: [PATCH 0797/1103] Remove deprecated code not used by evaluators --- Eigen/Cholesky | 1 - Eigen/CholmodSupport | 4 - Eigen/Core | 25 +- Eigen/IterativeLinearSolvers | 6 - Eigen/LU | 1 - Eigen/PaStiXSupport | 4 - Eigen/QR | 1 - Eigen/SPQRSupport | 2 - Eigen/SVD | 1 - Eigen/SparseCholesky | 5 - Eigen/SparseLU | 3 - Eigen/SparseQR | 3 - Eigen/SuperLUSupport | 4 - Eigen/UmfPackSupport | 3 - Eigen/src/Cholesky/LDLT.h | 29 - Eigen/src/Cholesky/LLT.h | 30 - Eigen/src/CholmodSupport/CholmodSupport.h | 62 -- Eigen/src/Core/ArrayBase.h | 65 -- Eigen/src/Core/Assign.h | 587 ------------------ Eigen/src/Core/AssignEvaluator.h | 6 - Eigen/src/Core/BandMatrix.h | 4 - Eigen/src/Core/Block.h | 15 - Eigen/src/Core/BooleanRedux.h | 51 -- Eigen/src/Core/CwiseBinaryOp.h | 72 --- Eigen/src/Core/CwiseNullaryOp.h | 9 - Eigen/src/Core/CwiseUnaryOp.h | 47 -- Eigen/src/Core/CwiseUnaryView.h | 31 - Eigen/src/Core/DenseBase.h | 56 -- Eigen/src/Core/DenseCoeffsBase.h | 164 ----- Eigen/src/Core/Diagonal.h | 6 - Eigen/src/Core/DiagonalMatrix.h | 42 -- Eigen/src/Core/DiagonalProduct.h | 117 ---- Eigen/src/Core/Dot.h | 11 - Eigen/src/Core/EigenBase.h | 12 - Eigen/src/Core/Fuzzy.h | 5 - Eigen/src/Core/GeneralProduct.h | 458 -------------- Eigen/src/Core/Inverse.h | 4 - Eigen/src/Core/Map.h | 18 - Eigen/src/Core/MapBase.h | 14 - Eigen/src/Core/Matrix.h | 4 - Eigen/src/Core/MatrixBase.h | 33 +- Eigen/src/Core/NoAlias.h | 63 -- Eigen/src/Core/PermutationMatrix.h | 73 --- Eigen/src/Core/PlainObjectBase.h | 25 - Eigen/src/Core/Product.h | 7 - Eigen/src/Core/ProductBase.h | 252 -------- Eigen/src/Core/Redux.h | 14 - Eigen/src/Core/Ref.h | 4 - Eigen/src/Core/Replicate.h | 5 - Eigen/src/Core/ReturnByValue.h | 11 +- Eigen/src/Core/Reverse.h | 10 - Eigen/src/Core/Select.h | 7 - Eigen/src/Core/SelfAdjointView.h | 137 ---- Eigen/src/Core/SelfCwiseBinaryOp.h | 209 ------- Eigen/src/Core/Solve.h | 2 - Eigen/src/Core/Swap.h | 131 ---- Eigen/src/Core/Transpose.h | 71 --- Eigen/src/Core/TriangularMatrix.h | 415 ------------- Eigen/src/Core/VectorwiseOp.h | 16 - Eigen/src/Core/Visitor.h | 13 - Eigen/src/Core/products/CoeffBasedProduct.h | 460 -------------- Eigen/src/Core/products/GeneralMatrixMatrix.h | 82 --- .../products/GeneralMatrixMatrixTriangular.h | 13 - .../Core/products/SelfadjointMatrixMatrix.h | 54 -- .../Core/products/SelfadjointMatrixVector.h | 113 ---- .../Core/products/TriangularMatrixMatrix.h | 51 -- .../Core/products/TriangularMatrixVector.h | 51 -- Eigen/src/Core/util/ForwardDeclarations.h | 4 - Eigen/src/Core/util/Macros.h | 42 -- Eigen/src/Core/util/XprHelper.h | 110 +--- Eigen/src/Geometry/AlignedBox.h | 4 - Eigen/src/Geometry/Homogeneous.h | 41 -- Eigen/src/Geometry/OrthoMethods.h | 15 - Eigen/src/Geometry/Quaternion.h | 4 - Eigen/src/Geometry/Transform.h | 4 - Eigen/src/Householder/HouseholderSequence.h | 3 - .../BasicPreconditioners.h | 29 - Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 37 -- .../ConjugateGradient.h | 37 -- .../IterativeLinearSolvers/IncompleteLUT.h | 30 - .../IterativeSolverBase.h | 72 --- Eigen/src/LU/Determinant.h | 4 - Eigen/src/LU/FullPivLU.h | 35 -- Eigen/src/LU/InverseImpl.h | 58 -- Eigen/src/LU/PartialPivLU.h | 35 -- Eigen/src/PaStiXSupport/PaStiXSupport.h | 64 -- Eigen/src/PardisoSupport/PardisoSupport.h | 62 -- Eigen/src/QR/ColPivHouseholderQR.h | 37 -- Eigen/src/QR/FullPivHouseholderQR.h | 40 -- Eigen/src/QR/HouseholderQR.h | 28 - Eigen/src/SPQRSupport/SuiteSparseQRSupport.h | 34 - Eigen/src/SVD/SVDBase.h | 28 - Eigen/src/SparseCholesky/SimplicialCholesky.h | 76 +-- .../ConservativeSparseSparseProduct.h | 10 - Eigen/src/SparseCore/MappedSparseMatrix.h | 2 - Eigen/src/SparseCore/SparseAssign.h | 113 ---- Eigen/src/SparseCore/SparseBlock.h | 95 --- Eigen/src/SparseCore/SparseCwiseBinaryOp.h | 254 -------- Eigen/src/SparseCore/SparseCwiseUnaryOp.h | 132 ---- Eigen/src/SparseCore/SparseDenseProduct.h | 248 +------- Eigen/src/SparseCore/SparseDiagonalProduct.h | 178 ------ Eigen/src/SparseCore/SparseDot.h | 18 +- Eigen/src/SparseCore/SparseMatrix.h | 93 --- Eigen/src/SparseCore/SparseMatrixBase.h | 55 -- Eigen/src/SparseCore/SparsePermutation.h | 44 -- Eigen/src/SparseCore/SparseProduct.h | 178 ------ Eigen/src/SparseCore/SparseRedux.h | 5 - Eigen/src/SparseCore/SparseSelfAdjointView.h | 149 +---- Eigen/src/SparseCore/SparseSolverBase.h | 9 +- .../SparseSparseProductWithPruning.h | 15 - Eigen/src/SparseCore/SparseTranspose.h | 51 -- Eigen/src/SparseCore/SparseTriangularView.h | 8 - Eigen/src/SparseCore/SparseUtil.h | 24 - Eigen/src/SparseCore/SparseVector.h | 25 - Eigen/src/SparseCore/SparseView.h | 43 -- Eigen/src/SparseCore/TriangularSolver.h | 156 ----- Eigen/src/SparseLU/SparseLU.h | 63 -- Eigen/src/SparseQR/SparseQR.h | 51 -- Eigen/src/SuperLUSupport/SuperLUSupport.h | 59 -- Eigen/src/UmfPackSupport/UmfPackSupport.h | 59 -- Eigen/src/misc/Solve.h | 78 --- Eigen/src/misc/SparseSolve.h | 134 ---- test/CMakeLists.txt | 5 - test/evaluators.cpp | 12 - test/mixingtypes.cpp | 10 +- test/nesting_ops.cpp | 5 - test/sparse_vector.cpp | 3 - test/vectorization_logic.cpp | 22 - unsupported/Eigen/IterativeSolvers | 3 - unsupported/Eigen/SparseExtra | 3 - .../Eigen/src/IterativeSolvers/DGMRES.h | 36 -- .../Eigen/src/IterativeSolvers/GMRES.h | 36 -- .../src/IterativeSolvers/IncompleteCholesky.h | 31 - .../Eigen/src/IterativeSolvers/IncompleteLU.h | 30 - .../Eigen/src/IterativeSolvers/MINRES.h | 37 -- .../src/SparseExtra/DynamicSparseMatrix.h | 2 - unsupported/test/CMakeLists.txt | 26 +- 137 files changed, 41 insertions(+), 7806 deletions(-) delete mode 100644 Eigen/src/Core/products/CoeffBasedProduct.h delete mode 100644 Eigen/src/misc/Solve.h delete mode 100644 Eigen/src/misc/SparseSolve.h diff --git a/Eigen/Cholesky b/Eigen/Cholesky index 7314d326c..dd0ca911c 100644 --- a/Eigen/Cholesky +++ b/Eigen/Cholesky @@ -21,7 +21,6 @@ * \endcode */ -#include "src/misc/Solve.h" #include "src/Cholesky/LLT.h" #include "src/Cholesky/LDLT.h" #ifdef EIGEN_USE_LAPACKE diff --git a/Eigen/CholmodSupport b/Eigen/CholmodSupport index 745b884e7..687cd9777 100644 --- a/Eigen/CholmodSupport +++ b/Eigen/CholmodSupport @@ -33,12 +33,8 @@ extern "C" { * */ -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" - #include "src/CholmodSupport/CholmodSupport.h" - #include "src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_CHOLMODSUPPORT_MODULE_H diff --git a/Eigen/Core b/Eigen/Core index 47c4f6299..2510e4898 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -11,16 +11,6 @@ #ifndef EIGEN_CORE_H #define EIGEN_CORE_H -// EIGEN_TEST_EVALUATORS => EIGEN_ENABLE_EVALUATORS -#ifndef EIGEN_TEST_NO_EVALUATORS - #ifndef EIGEN_TEST_EVALUATORS - #define EIGEN_TEST_EVALUATORS - #endif - #ifndef EIGEN_ENABLE_EVALUATORS - #define EIGEN_ENABLE_EVALUATORS - #endif -#endif - // first thing Eigen does: stop the compiler from committing suicide #include "src/Core/util/DisableStupidWarnings.h" @@ -317,11 +307,9 @@ using std::ptrdiff_t; #include "src/Core/MatrixBase.h" #include "src/Core/EigenBase.h" -#ifdef EIGEN_ENABLE_EVALUATORS #include "src/Core/Product.h" #include "src/Core/CoreEvaluators.h" #include "src/Core/AssignEvaluator.h" -#endif #ifndef EIGEN_PARSED_BY_DOXYGEN // work around Doxygen bug triggered by Assign.h r814874 // at least confirmed with Doxygen 1.5.5 and 1.5.6 @@ -332,10 +320,10 @@ using std::ptrdiff_t; #include "src/Core/util/BlasUtil.h" #include "src/Core/DenseStorage.h" #include "src/Core/NestByValue.h" -#ifndef EIGEN_ENABLE_EVALUATORS -#include "src/Core/ForceAlignedAccess.h" -#include "src/Core/Flagged.h" -#endif + +// #include "src/Core/ForceAlignedAccess.h" +// #include "src/Core/Flagged.h" + #include "src/Core/ReturnByValue.h" #include "src/Core/NoAlias.h" #include "src/Core/PlainObjectBase.h" @@ -368,18 +356,13 @@ using std::ptrdiff_t; #include "src/Core/CommaInitializer.h" #include "src/Core/ProductBase.h" #include "src/Core/GeneralProduct.h" -#ifdef EIGEN_ENABLE_EVALUATORS #include "src/Core/Solve.h" #include "src/Core/Inverse.h" -#endif #include "src/Core/TriangularMatrix.h" #include "src/Core/SelfAdjointView.h" #include "src/Core/products/GeneralBlockPanelKernel.h" #include "src/Core/products/Parallelizer.h" -#include "src/Core/products/CoeffBasedProduct.h" -#ifdef EIGEN_ENABLE_EVALUATORS #include "src/Core/ProductEvaluators.h" -#endif #include "src/Core/products/GeneralMatrixVector.h" #include "src/Core/products/GeneralMatrixMatrix.h" #include "src/Core/SolveTriangular.h" diff --git a/Eigen/IterativeLinearSolvers b/Eigen/IterativeLinearSolvers index 4df2c5a14..c06668bd2 100644 --- a/Eigen/IterativeLinearSolvers +++ b/Eigen/IterativeLinearSolvers @@ -26,13 +26,7 @@ * \endcode */ -#ifndef EIGEN_TEST_EVALUATORS -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" -#else #include "src/IterativeLinearSolvers/SolveWithGuess.h" -#endif - #include "src/IterativeLinearSolvers/IterativeSolverBase.h" #include "src/IterativeLinearSolvers/BasicPreconditioners.h" #include "src/IterativeLinearSolvers/ConjugateGradient.h" diff --git a/Eigen/LU b/Eigen/LU index 38e9067c7..132ecc42c 100644 --- a/Eigen/LU +++ b/Eigen/LU @@ -16,7 +16,6 @@ * \endcode */ -#include "src/misc/Solve.h" #include "src/misc/Kernel.h" #include "src/misc/Image.h" #include "src/LU/FullPivLU.h" diff --git a/Eigen/PaStiXSupport b/Eigen/PaStiXSupport index 7c616ee5e..e7d275f97 100644 --- a/Eigen/PaStiXSupport +++ b/Eigen/PaStiXSupport @@ -35,12 +35,8 @@ extern "C" { * */ -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" - #include "src/PaStiXSupport/PaStiXSupport.h" - #include "src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_PASTIXSUPPORT_MODULE_H diff --git a/Eigen/QR b/Eigen/QR index 4c2533610..230cb079a 100644 --- a/Eigen/QR +++ b/Eigen/QR @@ -24,7 +24,6 @@ * \endcode */ -#include "src/misc/Solve.h" #include "src/QR/HouseholderQR.h" #include "src/QR/FullPivHouseholderQR.h" #include "src/QR/ColPivHouseholderQR.h" diff --git a/Eigen/SPQRSupport b/Eigen/SPQRSupport index 77016442e..e3f49bb5a 100644 --- a/Eigen/SPQRSupport +++ b/Eigen/SPQRSupport @@ -21,8 +21,6 @@ * */ -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" #include "src/CholmodSupport/CholmodSupport.h" #include "src/SPQRSupport/SuiteSparseQRSupport.h" diff --git a/Eigen/SVD b/Eigen/SVD index c3d24286c..c13472e82 100644 --- a/Eigen/SVD +++ b/Eigen/SVD @@ -20,7 +20,6 @@ * \endcode */ -#include "src/misc/Solve.h" #include "src/SVD/SVDBase.h" #include "src/SVD/JacobiSVD.h" #if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT) diff --git a/Eigen/SparseCholesky b/Eigen/SparseCholesky index 2c4f66105..b6a320c40 100644 --- a/Eigen/SparseCholesky +++ b/Eigen/SparseCholesky @@ -34,11 +34,6 @@ #error The SparseCholesky module has nothing to offer in MPL2 only mode #endif -#ifndef EIGEN_TEST_EVALUATORS -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" -#endif - #include "src/SparseCholesky/SimplicialCholesky.h" #ifndef EIGEN_MPL2_ONLY diff --git a/Eigen/SparseLU b/Eigen/SparseLU index 8527a49bd..38b38b531 100644 --- a/Eigen/SparseLU +++ b/Eigen/SparseLU @@ -20,9 +20,6 @@ * Please, see the documentation of the SparseLU class for more details. */ -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" - // Ordering interface #include "OrderingMethods" diff --git a/Eigen/SparseQR b/Eigen/SparseQR index 4ee42065e..efb2695ba 100644 --- a/Eigen/SparseQR +++ b/Eigen/SparseQR @@ -21,9 +21,6 @@ * */ -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" - #include "OrderingMethods" #include "src/SparseCore/SparseColEtree.h" #include "src/SparseQR/SparseQR.h" diff --git a/Eigen/SuperLUSupport b/Eigen/SuperLUSupport index 575e14fbc..d1eac9464 100644 --- a/Eigen/SuperLUSupport +++ b/Eigen/SuperLUSupport @@ -48,12 +48,8 @@ namespace Eigen { struct SluMatrix; } * */ -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" - #include "src/SuperLUSupport/SuperLUSupport.h" - #include "src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_SUPERLUSUPPORT_MODULE_H diff --git a/Eigen/UmfPackSupport b/Eigen/UmfPackSupport index 984f64a84..0efad5dee 100644 --- a/Eigen/UmfPackSupport +++ b/Eigen/UmfPackSupport @@ -26,9 +26,6 @@ extern "C" { * */ -#include "src/misc/Solve.h" -#include "src/misc/SparseSolve.h" - #include "src/UmfPackSupport/UmfPackSupport.h" #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index f621323a3..32c770654 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -174,7 +174,6 @@ template class LDLT * * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt() */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -184,17 +183,6 @@ template class LDLT && "LDLT::solve(): invalid number of rows of the right hand side matrix b"); return Solve(*this, b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "LDLT is not initialized."); - eigen_assert(m_matrix.rows()==b.rows() - && "LDLT::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } -#endif template bool solveInPlace(MatrixBase &bAndX) const; @@ -524,23 +512,6 @@ void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) cons dst = m_transpositions.transpose() * dst; } #endif - -namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef LDLT<_MatrixType,_UpLo> LDLTType; - EIGEN_MAKE_SOLVE_HELPERS(LDLTType,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; -#endif -} /** \internal use x = ldlt_object.solve(x); * diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index 89fb9a011..cb9e0eb7b 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -117,7 +117,6 @@ template class LLT * * \sa solveInPlace(), MatrixBase::llt(), SelfAdjointView::llt() */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -127,17 +126,6 @@ template class LLT && "LLT::solve(): invalid number of rows of the right hand side matrix b"); return Solve(*this, b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "LLT is not initialized."); - eigen_assert(m_matrix.rows()==b.rows() - && "LLT::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } -#endif template void solveInPlace(MatrixBase &bAndX) const; @@ -433,24 +421,6 @@ void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const } #endif -namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef LLT<_MatrixType,UpLo> LLTType; - EIGEN_MAKE_SOLVE_HELPERS(LLTType,Rhs) - - template void evalTo(Dest& dst) const - { - dst = rhs(); - dec().solveInPlace(dst); - } -}; -#endif -} - /** \internal use x = llt_object.solve(x); * * This is the \em in-place version of solve(). diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h index d3db51d0b..3524ffb2d 100644 --- a/Eigen/src/CholmodSupport/CholmodSupport.h +++ b/Eigen/src/CholmodSupport/CholmodSupport.h @@ -217,36 +217,6 @@ class CholmodBase : public SparseSolverBase return derived(); } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "LLT is not initialized."); - eigen_assert(rows()==b.rows() - && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } - - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval - solve(const SparseMatrixBase& b) const - { - eigen_assert(m_isInitialized && "LLT is not initialized."); - eigen_assert(rows()==b.rows() - && "CholmodDecomposition::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); - } -#endif // EIGEN_TEST_EVALUATORS - /** Performs a symbolic decomposition on the sparsity pattern of \a matrix. * * This function is particularly useful when solving for several problems having the same structure. @@ -574,38 +544,6 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom } }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef CholmodBase<_MatrixType,_UpLo,Derived> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef CholmodBase<_MatrixType,_UpLo,Derived> Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} // end namespace internal -#endif - } // end namespace Eigen #endif // EIGEN_CHOLMODSUPPORT_H diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index f5bae6357..4e80634b9 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -64,9 +64,6 @@ template class ArrayBase using Base::MaxSizeAtCompileTime; using Base::IsVectorAtCompileTime; using Base::Flags; -#ifndef EIGEN_TEST_EVALUATORS - using Base::CoeffReadCost; -#endif using Base::derived; using Base::const_cast_derived; @@ -123,11 +120,7 @@ template class ArrayBase EIGEN_DEVICE_FUNC Derived& operator=(const ArrayBase& other) { -#ifndef EIGEN_TEST_EVALUATORS - return internal::assign_selector::run(derived(), other.derived()); -#else internal::call_assignment(derived(), other.derived()); -#endif } EIGEN_DEVICE_FUNC @@ -183,7 +176,6 @@ template class ArrayBase {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;} }; -#ifdef EIGEN_TEST_EVALUATORS /** replaces \c *this by \c *this - \a other. * * \returns a reference to \c *this @@ -235,63 +227,6 @@ ArrayBase::operator/=(const ArrayBase& other) call_assignment(derived(), other.derived(), internal::div_assign_op()); return derived(); } -#else // EIGEN_TEST_EVALUATORS -/** replaces \c *this by \c *this - \a other. - * - * \returns a reference to \c *this - */ -template -template -EIGEN_STRONG_INLINE Derived & -ArrayBase::operator-=(const ArrayBase &other) -{ - SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); - tmp = other.derived(); - return derived(); -} - -/** replaces \c *this by \c *this + \a other. - * - * \returns a reference to \c *this - */ -template -template -EIGEN_STRONG_INLINE Derived & -ArrayBase::operator+=(const ArrayBase& other) -{ - SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); - tmp = other.derived(); - return derived(); -} - -/** replaces \c *this by \c *this * \a other coefficient wise. - * - * \returns a reference to \c *this - */ -template -template -EIGEN_STRONG_INLINE Derived & -ArrayBase::operator*=(const ArrayBase& other) -{ - SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); - tmp = other.derived(); - return derived(); -} - -/** replaces \c *this by \c *this / \a other coefficient wise. - * - * \returns a reference to \c *this - */ -template -template -EIGEN_STRONG_INLINE Derived & -ArrayBase::operator/=(const ArrayBase& other) -{ - SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); - tmp = other.derived(); - return derived(); -} -#endif } // end namespace Eigen diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h index 5395e5436..53806ba33 100644 --- a/Eigen/src/Core/Assign.h +++ b/Eigen/src/Core/Assign.h @@ -13,487 +13,6 @@ #define EIGEN_ASSIGN_H namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -/*************************************************************************** -* Part 1 : the logic deciding a strategy for traversal and unrolling * -***************************************************************************/ - -template -struct assign_traits -{ -public: - enum { - DstIsAligned = Derived::Flags & AlignedBit, - DstHasDirectAccess = Derived::Flags & DirectAccessBit, - SrcIsAligned = OtherDerived::Flags & AlignedBit, - JointAlignment = bool(DstIsAligned) && bool(SrcIsAligned) ? Aligned : Unaligned - }; - -private: - enum { - InnerSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::SizeAtCompileTime) - : int(Derived::Flags)&RowMajorBit ? int(Derived::ColsAtCompileTime) - : int(Derived::RowsAtCompileTime), - InnerMaxSize = int(Derived::IsVectorAtCompileTime) ? int(Derived::MaxSizeAtCompileTime) - : int(Derived::Flags)&RowMajorBit ? int(Derived::MaxColsAtCompileTime) - : int(Derived::MaxRowsAtCompileTime), - MaxSizeAtCompileTime = Derived::SizeAtCompileTime, - PacketSize = packet_traits::size - }; - - enum { - StorageOrdersAgree = (int(Derived::IsRowMajor) == int(OtherDerived::IsRowMajor)), - MightVectorize = StorageOrdersAgree - && (int(Derived::Flags) & int(OtherDerived::Flags) & ActualPacketAccessBit), - MayInnerVectorize = MightVectorize && int(InnerSize)!=Dynamic && int(InnerSize)%int(PacketSize)==0 - && int(DstIsAligned) && int(SrcIsAligned), - MayLinearize = StorageOrdersAgree && (int(Derived::Flags) & int(OtherDerived::Flags) & LinearAccessBit), - MayLinearVectorize = MightVectorize && MayLinearize && DstHasDirectAccess - && (DstIsAligned || MaxSizeAtCompileTime == Dynamic), - /* If the destination isn't aligned, we have to do runtime checks and we don't unroll, - so it's only good for large enough sizes. */ - MaySliceVectorize = MightVectorize && DstHasDirectAccess - && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=3*PacketSize) - /* slice vectorization can be slow, so we only want it if the slices are big, which is - indicated by InnerMaxSize rather than InnerSize, think of the case of a dynamic block - in a fixed-size matrix */ - }; - -public: - enum { - Traversal = int(MayInnerVectorize) ? int(InnerVectorizedTraversal) - : int(MayLinearVectorize) ? int(LinearVectorizedTraversal) - : int(MaySliceVectorize) ? int(SliceVectorizedTraversal) - : int(MayLinearize) ? int(LinearTraversal) - : int(DefaultTraversal), - Vectorized = int(Traversal) == InnerVectorizedTraversal - || int(Traversal) == LinearVectorizedTraversal - || int(Traversal) == SliceVectorizedTraversal - }; - -private: - enum { - UnrollingLimit = EIGEN_UNROLLING_LIMIT * (Vectorized ? int(PacketSize) : 1), - MayUnrollCompletely = int(Derived::SizeAtCompileTime) != Dynamic - && int(OtherDerived::CoeffReadCost) != Dynamic - && int(Derived::SizeAtCompileTime) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit), - MayUnrollInner = int(InnerSize) != Dynamic - && int(OtherDerived::CoeffReadCost) != Dynamic - && int(InnerSize) * int(OtherDerived::CoeffReadCost) <= int(UnrollingLimit) - }; - -public: - enum { - Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal)) - ? ( - int(MayUnrollCompletely) ? int(CompleteUnrolling) - : int(MayUnrollInner) ? int(InnerUnrolling) - : int(NoUnrolling) - ) - : int(Traversal) == int(LinearVectorizedTraversal) - ? ( bool(MayUnrollCompletely) && bool(DstIsAligned) ? int(CompleteUnrolling) : int(NoUnrolling) ) - : int(Traversal) == int(LinearTraversal) - ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) : int(NoUnrolling) ) - : int(NoUnrolling) - }; - -#ifdef EIGEN_DEBUG_ASSIGN - static void debug() - { - EIGEN_DEBUG_VAR(DstIsAligned) - EIGEN_DEBUG_VAR(SrcIsAligned) - EIGEN_DEBUG_VAR(JointAlignment) - EIGEN_DEBUG_VAR(Derived::SizeAtCompileTime) - EIGEN_DEBUG_VAR(OtherDerived::CoeffReadCost) - EIGEN_DEBUG_VAR(InnerSize) - EIGEN_DEBUG_VAR(InnerMaxSize) - EIGEN_DEBUG_VAR(PacketSize) - EIGEN_DEBUG_VAR(StorageOrdersAgree) - EIGEN_DEBUG_VAR(MightVectorize) - EIGEN_DEBUG_VAR(MayLinearize) - EIGEN_DEBUG_VAR(MayInnerVectorize) - EIGEN_DEBUG_VAR(MayLinearVectorize) - EIGEN_DEBUG_VAR(MaySliceVectorize) - EIGEN_DEBUG_VAR(Traversal) - EIGEN_DEBUG_VAR(UnrollingLimit) - EIGEN_DEBUG_VAR(MayUnrollCompletely) - EIGEN_DEBUG_VAR(MayUnrollInner) - EIGEN_DEBUG_VAR(Unrolling) - } -#endif -}; - -/*************************************************************************** -* Part 2 : meta-unrollers -***************************************************************************/ - -/************************ -*** Default traversal *** -************************/ - -template -struct assign_DefaultTraversal_CompleteUnrolling -{ - enum { - outer = Index / Derived1::InnerSizeAtCompileTime, - inner = Index % Derived1::InnerSizeAtCompileTime - }; - - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - dst.copyCoeffByOuterInner(outer, inner, src); - assign_DefaultTraversal_CompleteUnrolling::run(dst, src); - } -}; - -template -struct assign_DefaultTraversal_CompleteUnrolling -{ - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {} -}; - -template -struct assign_DefaultTraversal_InnerUnrolling -{ - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer) - { - dst.copyCoeffByOuterInner(outer, Index, src); - assign_DefaultTraversal_InnerUnrolling::run(dst, src, outer); - } -}; - -template -struct assign_DefaultTraversal_InnerUnrolling -{ - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {} -}; - -/*********************** -*** Linear traversal *** -***********************/ - -template -struct assign_LinearTraversal_CompleteUnrolling -{ - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - dst.copyCoeff(Index, src); - assign_LinearTraversal_CompleteUnrolling::run(dst, src); - } -}; - -template -struct assign_LinearTraversal_CompleteUnrolling -{ - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {} -}; - -/************************** -*** Inner vectorization *** -**************************/ - -template -struct assign_innervec_CompleteUnrolling -{ - enum { - outer = Index / Derived1::InnerSizeAtCompileTime, - inner = Index % Derived1::InnerSizeAtCompileTime, - JointAlignment = assign_traits::JointAlignment - }; - - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - dst.template copyPacketByOuterInner(outer, inner, src); - assign_innervec_CompleteUnrolling::size, Stop>::run(dst, src); - } -}; - -template -struct assign_innervec_CompleteUnrolling -{ - static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &) {} -}; - -template -struct assign_innervec_InnerUnrolling -{ - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src, typename Derived1::Index outer) - { - dst.template copyPacketByOuterInner(outer, Index, src); - assign_innervec_InnerUnrolling::size, Stop>::run(dst, src, outer); - } -}; - -template -struct assign_innervec_InnerUnrolling -{ - static EIGEN_STRONG_INLINE void run(Derived1 &, const Derived2 &, typename Derived1::Index) {} -}; - -/*************************************************************************** -* Part 3 : implementation of all cases -***************************************************************************/ - -template::Traversal, - int Unrolling = assign_traits::Unrolling, - int Version = Specialized> -struct assign_impl; - -/************************ -*** Default traversal *** -************************/ - -template -struct assign_impl -{ - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &, const Derived2 &) { } -}; - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - const Index innerSize = dst.innerSize(); - const Index outerSize = dst.outerSize(); - for(Index outer = 0; outer < outerSize; ++outer) - for(Index inner = 0; inner < innerSize; ++inner) - dst.copyCoeffByOuterInner(outer, inner, src); - } -}; - -template -struct assign_impl -{ - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - assign_DefaultTraversal_CompleteUnrolling - ::run(dst, src); - } -}; - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - const Index outerSize = dst.outerSize(); - for(Index outer = 0; outer < outerSize; ++outer) - assign_DefaultTraversal_InnerUnrolling - ::run(dst, src, outer); - } -}; - -/*********************** -*** Linear traversal *** -***********************/ - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - const Index size = dst.size(); - for(Index i = 0; i < size; ++i) - dst.copyCoeff(i, src); - } -}; - -template -struct assign_impl -{ - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - assign_LinearTraversal_CompleteUnrolling - ::run(dst, src); - } -}; - -/************************** -*** Inner vectorization *** -**************************/ - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - static inline void run(Derived1 &dst, const Derived2 &src) - { - const Index innerSize = dst.innerSize(); - const Index outerSize = dst.outerSize(); - const Index packetSize = packet_traits::size; - for(Index outer = 0; outer < outerSize; ++outer) - for(Index inner = 0; inner < innerSize; inner+=packetSize) - dst.template copyPacketByOuterInner(outer, inner, src); - } -}; - -template -struct assign_impl -{ - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - assign_innervec_CompleteUnrolling - ::run(dst, src); - } -}; - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - const Index outerSize = dst.outerSize(); - for(Index outer = 0; outer < outerSize; ++outer) - assign_innervec_InnerUnrolling - ::run(dst, src, outer); - } -}; - -/*************************** -*** Linear vectorization *** -***************************/ - -template -struct unaligned_assign_impl -{ - template - static EIGEN_STRONG_INLINE void run(const Derived&, OtherDerived&, typename Derived::Index, typename Derived::Index) {} -}; - -template <> -struct unaligned_assign_impl -{ - // MSVC must not inline this functions. If it does, it fails to optimize the - // packet access path. -#ifdef _MSC_VER - template - static EIGEN_DONT_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end) -#else - template - static EIGEN_STRONG_INLINE void run(const Derived& src, OtherDerived& dst, typename Derived::Index start, typename Derived::Index end) -#endif - { - for (typename Derived::Index index = start; index < end; ++index) - dst.copyCoeff(index, src); - } -}; - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - const Index size = dst.size(); - typedef packet_traits PacketTraits; - enum { - packetSize = PacketTraits::size, - dstAlignment = PacketTraits::AlignedOnScalar ? Aligned : int(assign_traits::DstIsAligned) , - srcAlignment = assign_traits::JointAlignment - }; - const Index alignedStart = assign_traits::DstIsAligned ? 0 - : internal::first_aligned(&dst.coeffRef(0), size); - const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize; - - unaligned_assign_impl::DstIsAligned!=0>::run(src,dst,0,alignedStart); - - for(Index index = alignedStart; index < alignedEnd; index += packetSize) - { - dst.template copyPacket(index, src); - } - - unaligned_assign_impl<>::run(src,dst,alignedEnd,size); - } -}; - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - static EIGEN_STRONG_INLINE void run(Derived1 &dst, const Derived2 &src) - { - enum { size = Derived1::SizeAtCompileTime, - packetSize = packet_traits::size, - alignedSize = (size/packetSize)*packetSize }; - - assign_innervec_CompleteUnrolling::run(dst, src); - assign_DefaultTraversal_CompleteUnrolling::run(dst, src); - } -}; - -/************************** -*** Slice vectorization *** -***************************/ - -template -struct assign_impl -{ - typedef typename Derived1::Index Index; - static inline void run(Derived1 &dst, const Derived2 &src) - { - typedef packet_traits PacketTraits; - enum { - packetSize = PacketTraits::size, - alignable = PacketTraits::AlignedOnScalar, - dstAlignment = alignable ? Aligned : int(assign_traits::DstIsAligned) , - srcAlignment = assign_traits::JointAlignment - }; - const Index packetAlignedMask = packetSize - 1; - const Index innerSize = dst.innerSize(); - const Index outerSize = dst.outerSize(); - const Index alignedStep = alignable ? (packetSize - dst.outerStride() % packetSize) & packetAlignedMask : 0; - Index alignedStart = ((!alignable) || assign_traits::DstIsAligned) ? 0 - : internal::first_aligned(&dst.coeffRef(0,0), innerSize); - - for(Index outer = 0; outer < outerSize; ++outer) - { - const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask); - // do the non-vectorizable part of the assignment - for(Index inner = 0; inner(outer, inner, src); - - // do the non-vectorizable part of the assignment - for(Index inner = alignedEnd; inner((alignedStart+alignedStep)%packetSize, innerSize); - } - } -}; - -} // end namespace internal - -#endif // EIGEN_TEST_EVALUATORS - -/*************************************************************************** -* Part 4 : implementation of DenseBase methods -***************************************************************************/ template template @@ -508,71 +27,12 @@ EIGEN_STRONG_INLINE Derived& DenseBase EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Derived,OtherDerived) EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) -#ifdef EIGEN_TEST_EVALUATORS - eigen_assert(rows() == other.rows() && cols() == other.cols()); internal::call_assignment_no_alias(derived(),other.derived()); -#else // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_DEBUG_ASSIGN - internal::assign_traits::debug(); -#endif - eigen_assert(rows() == other.rows() && cols() == other.cols()); - internal::assign_impl::Traversal) - : int(InvalidTraversal)>::run(derived(),other.derived()); - -#ifndef EIGEN_NO_DEBUG - checkTransposeAliasing(other.derived()); -#endif - -#endif // EIGEN_TEST_EVALUATORS - return derived(); } -namespace internal { - -#ifndef EIGEN_TEST_EVALUATORS -template::Flags) & EvalBeforeAssigningBit) != 0, - bool NeedToTranspose = ((int(Derived::RowsAtCompileTime) == 1 && int(OtherDerived::ColsAtCompileTime) == 1) - | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&". - // revert to || as soon as not needed anymore. - (int(Derived::ColsAtCompileTime) == 1 && int(OtherDerived::RowsAtCompileTime) == 1)) - && int(Derived::SizeAtCompileTime) != 1> -struct assign_selector; - -template -struct assign_selector { - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); } - template - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; } -}; -template -struct assign_selector { - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.eval()); } -}; -template -struct assign_selector { - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); } - template - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { Transpose dstTrans(dst); other.evalTo(dstTrans); return dst; } -}; -template -struct assign_selector { - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose().eval()); } -}; -#endif // EIGEN_TEST_EVALUATORS -} // end namespace internal - -#ifdef EIGEN_TEST_EVALUATORS template template EIGEN_DEVICE_FUNC @@ -624,53 +84,6 @@ EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const ReturnByValue< other.derived().evalTo(derived()); return derived(); } -#else // EIGEN_TEST_EVALUATORS -template -template -EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE Derived& DenseBase::operator=(const DenseBase& other) -{ - return internal::assign_selector::run(derived(), other.derived()); -} - -template -EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE Derived& DenseBase::operator=(const DenseBase& other) -{ - return internal::assign_selector::run(derived(), other.derived()); -} - -template -EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const MatrixBase& other) -{ - return internal::assign_selector::run(derived(), other.derived()); -} - -template -template -EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const DenseBase& other) -{ - return internal::assign_selector::run(derived(), other.derived()); -} - -template -template -EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const EigenBase& other) -{ - return internal::assign_selector::evalTo(derived(), other.derived()); -} - -template -template -EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE Derived& MatrixBase::operator=(const ReturnByValue& other) -{ - return internal::assign_selector::evalTo(derived(), other.derived()); -} -#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 4e35e432e..8ab71446c 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -695,13 +695,7 @@ void call_assignment(const Dst& dst, const Src& src) template void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) { -#ifdef EIGEN_TEST_EVALUATORS typename plain_matrix_type::type tmp(src); -#else - typename Src::PlainObject tmp(src.rows(), src.cols()); - call_assignment_no_alias(tmp, src, internal::assign_op()); -#endif - call_assignment_no_alias(dst, tmp, func); } diff --git a/Eigen/src/Core/BandMatrix.h b/Eigen/src/Core/BandMatrix.h index ba4749707..b0ebe1160 100644 --- a/Eigen/src/Core/BandMatrix.h +++ b/Eigen/src/Core/BandMatrix.h @@ -328,8 +328,6 @@ class TridiagonalMatrix : public BandMatrix @@ -347,8 +345,6 @@ struct evaluator_traits struct AssignmentKind { typedef EigenBase2EigenBase Kind; }; -#endif - } // end namespace internal diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 2d62f7a46..737e5dc24 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -84,26 +84,11 @@ struct traits > : traits::size) == 0) - && (InnerStrideAtCompileTime == 1) - ? PacketAccessBit : 0, - MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) ? AlignedBit : 0, - FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (traits::Flags&LinearAccessBit))) ? LinearAccessBit : 0, - FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, - FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, - Flags0 = traits::Flags & ( (HereditaryBits & ~RowMajorBit) | - DirectAccessBit | - MaskPacketAccessBit | - MaskAlignedBit), - Flags = Flags0 | FlagsLinearAccessBit | FlagsLvalueBit | FlagsRowMajorBit -#else // FIXME, this traits is rather specialized for dense object and it needs to be cleaned further FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, Flags = (traits::Flags & DirectAccessBit) | FlagsLvalueBit | FlagsRowMajorBit // FIXME DirectAccessBit should not be handled by expressions -#endif }; }; diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h index 192e1db53..dac1887e0 100644 --- a/Eigen/src/Core/BooleanRedux.h +++ b/Eigen/src/Core/BooleanRedux.h @@ -17,18 +17,11 @@ namespace internal { template struct all_unroller { -#ifdef EIGEN_TEST_EVALUATORS typedef typename Derived::ExpressionTraits Traits; enum { col = (UnrollCount-1) / Traits::RowsAtCompileTime, row = (UnrollCount-1) % Traits::RowsAtCompileTime }; -#else - enum { - col = (UnrollCount-1) / Derived::RowsAtCompileTime, - row = (UnrollCount-1) % Derived::RowsAtCompileTime - }; -#endif static inline bool run(const Derived &mat) { @@ -51,18 +44,11 @@ struct all_unroller template struct any_unroller { -#ifdef EIGEN_TEST_EVALUATORS typedef typename Derived::ExpressionTraits Traits; enum { col = (UnrollCount-1) / Traits::RowsAtCompileTime, row = (UnrollCount-1) % Traits::RowsAtCompileTime }; -#else - enum { - col = (UnrollCount-1) / Derived::RowsAtCompileTime, - row = (UnrollCount-1) % Derived::RowsAtCompileTime - }; -#endif static inline bool run(const Derived &mat) { @@ -94,7 +80,6 @@ struct any_unroller template inline bool DenseBase::all() const { -#ifdef EIGEN_TEST_EVALUATORS typedef typename internal::evaluator::type Evaluator; enum { unroll = SizeAtCompileTime != Dynamic @@ -112,24 +97,6 @@ inline bool DenseBase::all() const if (!evaluator.coeff(i, j)) return false; return true; } -#else - enum { - unroll = SizeAtCompileTime != Dynamic - && CoeffReadCost != Dynamic - && NumTraits::AddCost != Dynamic - && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT - }; - - if(unroll) - return internal::all_unroller::run(derived()); - else - { - for(Index j = 0; j < cols(); ++j) - for(Index i = 0; i < rows(); ++i) - if (!coeff(i, j)) return false; - return true; - } -#endif } /** \returns true if at least one coefficient is true @@ -139,7 +106,6 @@ inline bool DenseBase::all() const template inline bool DenseBase::any() const { -#ifdef EIGEN_TEST_EVALUATORS typedef typename internal::evaluator::type Evaluator; enum { unroll = SizeAtCompileTime != Dynamic @@ -157,23 +123,6 @@ inline bool DenseBase::any() const if (evaluator.coeff(i, j)) return true; return false; } -#else - enum { - unroll = SizeAtCompileTime != Dynamic - && CoeffReadCost != Dynamic - && NumTraits::AddCost != Dynamic - && SizeAtCompileTime * (CoeffReadCost + NumTraits::AddCost) <= EIGEN_UNROLLING_LIMIT - }; - if(unroll) - return internal::any_unroller::run(derived()); - else - { - for(Index j = 0; j < cols(); ++j) - for(Index i = 0; i < rows(); ++i) - if (coeff(i, j)) return true; - return false; - } -#endif } /** \returns the number of coefficients which evaluate to true diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index b4662907e..de9109e53 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -66,28 +66,7 @@ struct traits > typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; enum { -#ifndef EIGEN_TEST_EVALUATORS - LhsFlags = _LhsNested::Flags, - RhsFlags = _RhsNested::Flags, - SameType = is_same::value, - StorageOrdersAgree = (int(Lhs::Flags)&RowMajorBit)==(int(Rhs::Flags)&RowMajorBit), - Flags0 = (int(LhsFlags) | int(RhsFlags)) & ( - HereditaryBits - | (int(LhsFlags) & int(RhsFlags) & - ( AlignedBit - | (StorageOrdersAgree ? LinearAccessBit : 0) - | (functor_traits::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0) - ) - ) - ), - Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit), - - LhsCoeffReadCost = _LhsNested::CoeffReadCost, - RhsCoeffReadCost = _RhsNested::CoeffReadCost, - CoeffReadCost = LhsCoeffReadCost + RhsCoeffReadCost + functor_traits::Cost -#else Flags = _LhsNested::Flags & RowMajorBit -#endif }; }; } // end namespace internal @@ -163,46 +142,6 @@ class CwiseBinaryOp : internal::no_assignment_operator, const BinaryOp m_functor; }; -#ifndef EIGEN_TEST_EVALUATORS -template -class CwiseBinaryOpImpl - : public internal::dense_xpr_base >::type -{ - typedef CwiseBinaryOp Derived; - public: - - typedef typename internal::dense_xpr_base >::type Base; - EIGEN_DENSE_PUBLIC_INTERFACE( Derived ) - - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const - { - return derived().functor()(derived().lhs().coeff(rowId, colId), - derived().rhs().coeff(rowId, colId)); - } - - template - EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const - { - return derived().functor().packetOp(derived().lhs().template packet(rowId, colId), - derived().rhs().template packet(rowId, colId)); - } - - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index index) const - { - return derived().functor()(derived().lhs().coeff(index), - derived().rhs().coeff(index)); - } - - template - EIGEN_STRONG_INLINE PacketScalar packet(Index index) const - { - return derived().functor().packetOp(derived().lhs().template packet(index), - derived().rhs().template packet(index)); - } -}; -#else // Generic API dispatcher template class CwiseBinaryOpImpl @@ -211,7 +150,6 @@ class CwiseBinaryOpImpl public: typedef typename internal::generic_xpr_base >::type Base; }; -#endif /** replaces \c *this by \c *this - \a other. * @@ -222,12 +160,7 @@ template EIGEN_STRONG_INLINE Derived & MatrixBase::operator-=(const MatrixBase &other) { -#ifdef EIGEN_TEST_EVALUATORS call_assignment(derived(), other.derived(), internal::sub_assign_op()); -#else - SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); - tmp = other.derived(); -#endif return derived(); } @@ -240,12 +173,7 @@ template EIGEN_STRONG_INLINE Derived & MatrixBase::operator+=(const MatrixBase& other) { -#ifdef EIGEN_TEST_EVALUATORS call_assignment(derived(), other.derived(), internal::add_assign_op()); -#else - SelfCwiseBinaryOp, Derived, OtherDerived> tmp(derived()); - tmp = other.derived(); -#endif return derived(); } diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index f9f127cc2..8b8397da6 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -35,16 +35,7 @@ template struct traits > : traits { enum { -#ifndef EIGEN_TEST_EVALUATORS - Flags = (traits::Flags - & ( HereditaryBits - | (functor_has_linear_access::ret ? LinearAccessBit : 0) - | (functor_traits::PacketAccess ? PacketAccessBit : 0))) - | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit), - CoeffReadCost = functor_traits::Cost -#else Flags = traits::Flags & RowMajorBit -#endif }; }; } diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index c2bc47c93..79a872934 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -44,14 +44,7 @@ struct traits > typedef typename XprType::Nested XprTypeNested; typedef typename remove_reference::type _XprTypeNested; enum { -#ifndef EIGEN_TEST_EVALUATORS - Flags = _XprTypeNested::Flags & ( - HereditaryBits | LinearAccessBit | AlignedBit - | (functor_traits::PacketAccess ? PacketAccessBit : 0)), - CoeffReadCost = _XprTypeNested::CoeffReadCost + functor_traits::Cost -#else Flags = _XprTypeNested::Flags & RowMajorBit -#endif }; }; } @@ -97,45 +90,6 @@ class CwiseUnaryOp : internal::no_assignment_operator, const UnaryOp m_functor; }; -#ifndef EIGEN_TEST_EVALUATORS -// This is the generic implementation for dense storage. -// It can be used for any expression types implementing the dense concept. -template -class CwiseUnaryOpImpl - : public internal::dense_xpr_base >::type -{ - public: - - typedef CwiseUnaryOp Derived; - typedef typename internal::dense_xpr_base >::type Base; - EIGEN_DENSE_PUBLIC_INTERFACE(Derived) - - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index rowId, Index colId) const - { - return derived().functor()(derived().nestedExpression().coeff(rowId, colId)); - } - - template - EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const - { - return derived().functor().packetOp(derived().nestedExpression().template packet(rowId, colId)); - } - - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index index) const - { - return derived().functor()(derived().nestedExpression().coeff(index)); - } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE PacketScalar packet(Index index) const - { - return derived().functor().packetOp(derived().nestedExpression().template packet(index)); - } -}; -#else // EIGEN_TEST_EVALUATORS // Generic API dispatcher template class CwiseUnaryOpImpl @@ -144,7 +98,6 @@ class CwiseUnaryOpImpl public: typedef typename internal::generic_xpr_base >::type Base; }; -#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index 99cc03ac1..71249a39c 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -37,12 +37,7 @@ struct traits > typedef typename MatrixType::Nested MatrixTypeNested; typedef typename remove_all::type _MatrixTypeNested; enum { -#ifndef EIGEN_TEST_EVALUATORS - Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)), - CoeffReadCost = traits<_MatrixTypeNested>::CoeffReadCost + functor_traits::Cost, -#else Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | LvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions -#endif MatrixTypeInnerStride = inner_stride_at_compile_time::ret, // need to cast the sizeof's from size_t to int explicitly, otherwise: // "error: no integral type can represent all of the enumerator values @@ -93,7 +88,6 @@ class CwiseUnaryView : public CwiseUnaryViewImpl class CwiseUnaryViewImpl @@ -102,7 +96,6 @@ class CwiseUnaryViewImpl public: typedef typename internal::generic_xpr_base >::type Base; }; -#endif template class CwiseUnaryViewImpl @@ -128,30 +121,6 @@ class CwiseUnaryViewImpl { return derived().nestedExpression().outerStride() * sizeof(typename internal::traits::Scalar) / sizeof(Scalar); } - -#ifndef EIGEN_TEST_EVALUATORS - - EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const - { - return derived().functor()(derived().nestedExpression().coeff(row, col)); - } - - EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const - { - return derived().functor()(derived().nestedExpression().coeff(index)); - } - - EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col) - { - return derived().functor()(const_cast_derived().nestedExpression().coeffRef(row, col)); - } - - EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) - { - return derived().functor()(const_cast_derived().nestedExpression().coeffRef(index)); - } - -#endif }; } // end namespace Eigen diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index f35c2edc4..6078af553 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -74,18 +74,6 @@ template class DenseBase using Base::colIndexByOuterInner; using Base::coeff; using Base::coeffByOuterInner; -#ifndef EIGEN_TEST_EVALUATORS - using Base::packet; - using Base::packetByOuterInner; - using Base::writePacket; - using Base::writePacketByOuterInner; - using Base::coeffRef; - using Base::coeffRefByOuterInner; - using Base::copyCoeff; - using Base::copyCoeffByOuterInner; - using Base::copyPacket; - using Base::copyPacketByOuterInner; -#endif using Base::operator(); using Base::operator[]; using Base::x; @@ -171,13 +159,6 @@ template class DenseBase InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime) : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime), -#ifndef EIGEN_TEST_EVALUATORS - CoeffReadCost = internal::traits::CoeffReadCost, - /**< This is a rough measure of how expensive it is to read one coefficient from - * this expression. - */ -#endif - InnerStrideAtCompileTime = internal::inner_stride_at_compile_time::ret, OuterStrideAtCompileTime = internal::outer_stride_at_compile_time::ret }; @@ -292,15 +273,10 @@ template class DenseBase EIGEN_DEVICE_FUNC CommaInitializer operator<< (const Scalar& s); -#ifndef EIGEN_TEST_EVALUATORS - template - const Flagged flagged() const; -#else // TODO flagged is temporarly disabled. It seems useless now template const Derived& flagged() const { return derived(); } -#endif template EIGEN_DEVICE_FUNC @@ -313,15 +289,6 @@ template class DenseBase ConstTransposeReturnType transpose() const; EIGEN_DEVICE_FUNC void transposeInPlace(); -#ifndef EIGEN_NO_DEBUG -#ifndef EIGEN_TEST_EVALUATORS - protected: - template - void checkTransposeAliasing(const OtherDerived& other) const; - public: -#endif -#endif - EIGEN_DEVICE_FUNC static const ConstantReturnType Constant(Index rows, Index cols, const Scalar& value); @@ -402,7 +369,6 @@ template class DenseBase return typename internal::eval::type(derived()); } -#ifdef EIGEN_TEST_EVALUATORS /** swaps *this with the expression \a other. * */ @@ -425,28 +391,6 @@ template class DenseBase eigen_assert(rows()==other.rows() && cols()==other.cols()); call_assignment(derived(), other.derived(), internal::swap_assign_op()); } -#else // EIGEN_TEST_EVALUATORS - /** swaps *this with the expression \a other. - * - */ - template - EIGEN_DEVICE_FUNC - void swap(const DenseBase& other, - int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase) - { - SwapWrapper(derived()).lazyAssign(other.derived()); - } - - /** swaps *this with the matrix or array \a other. - * - */ - template - EIGEN_DEVICE_FUNC - void swap(PlainObjectBase& other) - { - SwapWrapper(derived()).lazyAssign(other.derived()); - } -#endif // EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC inline const NestByValue nestByValue() const; EIGEN_DEVICE_FUNC inline const ForceAlignedAccess forceAlignedAccess() const; diff --git a/Eigen/src/Core/DenseCoeffsBase.h b/Eigen/src/Core/DenseCoeffsBase.h index de31b8df2..a9e4dbaf9 100644 --- a/Eigen/src/Core/DenseCoeffsBase.h +++ b/Eigen/src/Core/DenseCoeffsBase.h @@ -98,11 +98,7 @@ class DenseCoeffsBase : public EigenBase { eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols()); -#ifndef EIGEN_TEST_EVALUATORS - return derived().coeff(row, col); -#else return typename internal::evaluator::type(derived()).coeff(row,col); -#endif } EIGEN_DEVICE_FUNC @@ -144,11 +140,7 @@ class DenseCoeffsBase : public EigenBase coeff(Index index) const { eigen_internal_assert(index >= 0 && index < size()); -#ifndef EIGEN_TEST_EVALUATORS - return derived().coeff(index); -#else return typename internal::evaluator::type(derived()).coeff(index); -#endif } @@ -226,11 +218,7 @@ class DenseCoeffsBase : public EigenBase EIGEN_STRONG_INLINE PacketReturnType packet(Index row, Index col) const { eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols()); -#ifndef EIGEN_TEST_EVALUATORS - return derived().template packet(row,col); -#else return typename internal::evaluator::type(derived()).template packet(row,col); -#endif } @@ -256,11 +244,7 @@ class DenseCoeffsBase : public EigenBase EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { eigen_internal_assert(index >= 0 && index < size()); -#ifndef EIGEN_TEST_EVALUATORS - return derived().template packet(index); -#else return typename internal::evaluator::type(derived()).template packet(index); -#endif } protected: @@ -341,11 +325,7 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && row < rows() && col >= 0 && col < cols()); -#ifndef EIGEN_TEST_EVALUATORS - return derived().coeffRef(row, col); -#else return typename internal::evaluator::type(derived()).coeffRef(row,col); -#endif } EIGEN_DEVICE_FUNC @@ -391,11 +371,7 @@ class DenseCoeffsBase : public DenseCoeffsBase= 0 && index < size()); -#ifndef EIGEN_TEST_EVALUATORS - return derived().coeffRef(index); -#else return typename internal::evaluator::type(derived()).coeffRef(index); -#endif } /** \returns a reference to the coefficient at given index. @@ -455,146 +431,6 @@ class DenseCoeffsBase : public DenseCoeffsBase - EIGEN_STRONG_INLINE void writePacket - (Index row, Index col, const typename internal::packet_traits::type& val) - { - eigen_internal_assert(row >= 0 && row < rows() - && col >= 0 && col < cols()); - derived().template writePacket(row,col,val); - } - - - /** \internal */ - template - EIGEN_STRONG_INLINE void writePacketByOuterInner - (Index outer, Index inner, const typename internal::packet_traits::type& val) - { - writePacket(rowIndexByOuterInner(outer, inner), - colIndexByOuterInner(outer, inner), - val); - } - - /** \internal - * Stores the given packet of coefficients, at the given index in this expression. It is your responsibility - * to ensure that a packet really starts there. This method is only available on expressions having the - * PacketAccessBit and the LinearAccessBit. - * - * The \a LoadMode parameter may have the value \a Aligned or \a Unaligned. Its effect is to select - * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets - * starting at an address which is a multiple of the packet size. - */ - template - EIGEN_STRONG_INLINE void writePacket - (Index index, const typename internal::packet_traits::type& val) - { - eigen_internal_assert(index >= 0 && index < size()); - derived().template writePacket(index,val); - } - -#ifndef EIGEN_PARSED_BY_DOXYGEN - - /** \internal Copies the coefficient at position (row,col) of other into *this. - * - * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code - * with usual assignments. - * - * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox. - */ - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, const DenseBase& other) - { - eigen_internal_assert(row >= 0 && row < rows() - && col >= 0 && col < cols()); - derived().coeffRef(row, col) = other.derived().coeff(row, col); - } - - /** \internal Copies the coefficient at the given index of other into *this. - * - * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code - * with usual assignments. - * - * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox. - */ - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void copyCoeff(Index index, const DenseBase& other) - { - eigen_internal_assert(index >= 0 && index < size()); - derived().coeffRef(index) = other.derived().coeff(index); - } - - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void copyCoeffByOuterInner(Index outer, Index inner, const DenseBase& other) - { - const Index row = rowIndexByOuterInner(outer,inner); - const Index col = colIndexByOuterInner(outer,inner); - // derived() is important here: copyCoeff() may be reimplemented in Derived! - derived().copyCoeff(row, col, other); - } - - /** \internal Copies the packet at position (row,col) of other into *this. - * - * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code - * with usual assignments. - * - * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox. - */ - - template - EIGEN_STRONG_INLINE void copyPacket(Index row, Index col, const DenseBase& other) - { - eigen_internal_assert(row >= 0 && row < rows() - && col >= 0 && col < cols()); - derived().template writePacket(row, col, - other.derived().template packet(row, col)); - } - - /** \internal Copies the packet at the given index of other into *this. - * - * This method is overridden in SwapWrapper, allowing swap() assignments to share 99% of their code - * with usual assignments. - * - * Outside of this internal usage, this method has probably no usefulness. It is hidden in the public API dox. - */ - - template - EIGEN_STRONG_INLINE void copyPacket(Index index, const DenseBase& other) - { - eigen_internal_assert(index >= 0 && index < size()); - derived().template writePacket(index, - other.derived().template packet(index)); - } - - /** \internal */ - template - EIGEN_STRONG_INLINE void copyPacketByOuterInner(Index outer, Index inner, const DenseBase& other) - { - const Index row = rowIndexByOuterInner(outer,inner); - const Index col = colIndexByOuterInner(outer,inner); - // derived() is important here: copyCoeff() may be reimplemented in Derived! - derived().template copyPacket< OtherDerived, StoreMode, LoadMode>(row, col, other); - } -#endif -#endif // EIGEN_TEST_EVALUATORS - }; /** \brief Base class providing direct read-only coefficient access to matrices and arrays. diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index 3ff6a3e66..1ffcd97f9 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -51,14 +51,8 @@ struct traits > : (EIGEN_PLAIN_ENUM_MIN(MatrixType::MaxRowsAtCompileTime - EIGEN_PLAIN_ENUM_MAX(-DiagIndex, 0), MatrixType::MaxColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))), MaxColsAtCompileTime = 1, -#ifndef EIGEN_TEST_EVALUATORS - MaskLvalueBit = is_lvalue::value ? LvalueBit : 0, - Flags = (unsigned int)_MatrixTypeNested::Flags & (HereditaryBits | LinearAccessBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, - CoeffReadCost = _MatrixTypeNested::CoeffReadCost, -#else MaskLvalueBit = is_lvalue::value ? LvalueBit : 0, Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions -#endif MatrixTypeOuterStride = outer_stride_at_compile_time::ret, InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1, OuterStrideAtCompileTime = 0 diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index 801131b54..44c249aa6 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -45,20 +45,6 @@ class DiagonalBase : public EigenBase EIGEN_DEVICE_FUNC DenseMatrixType toDenseMatrix() const { return derived(); } -#ifndef EIGEN_TEST_EVALUATORS - template - EIGEN_DEVICE_FUNC - void evalTo(MatrixBase &other) const; - template - EIGEN_DEVICE_FUNC - void addTo(MatrixBase &other) const - { other.diagonal() += diagonal(); } - template - EIGEN_DEVICE_FUNC - void subTo(MatrixBase &other) const - { other.diagonal() -= diagonal(); } -#endif // EIGEN_TEST_EVALUATORS - EIGEN_DEVICE_FUNC inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); } EIGEN_DEVICE_FUNC @@ -69,17 +55,6 @@ class DiagonalBase : public EigenBase EIGEN_DEVICE_FUNC inline Index cols() const { return diagonal().size(); } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the diagonal matrix product of \c *this by the matrix \a matrix. - */ - template - EIGEN_DEVICE_FUNC - const DiagonalProduct - operator*(const MatrixBase &matrix) const - { - return DiagonalProduct(matrix.derived(), derived()); - } -#else template EIGEN_DEVICE_FUNC const Product @@ -87,7 +62,6 @@ class DiagonalBase : public EigenBase { return Product(derived(),matrix.derived()); } -#endif // EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC inline const DiagonalWrapper, const DiagonalVectorType> > @@ -110,16 +84,6 @@ class DiagonalBase : public EigenBase } }; -#ifndef EIGEN_TEST_EVALUATORS -template -template -void DiagonalBase::evalTo(MatrixBase &other) const -{ - other.setZero(); - other.diagonal() = diagonal(); -} -#endif // EIGEN_TEST_EVALUATORS - #endif /** \class DiagonalMatrix @@ -273,10 +237,6 @@ struct traits > MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime, Flags = (traits::Flags & LvalueBit) | NoPreferredStorageOrderBit -#ifndef EIGEN_TEST_EVALUATORS - , - CoeffReadCost = traits<_DiagonalVectorType>::CoeffReadCost -#endif }; }; } @@ -347,7 +307,6 @@ bool MatrixBase::isDiagonal(const RealScalar& prec) const return true; } -#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template<> struct storage_kind_to_shape { typedef DiagonalShape Shape; }; @@ -368,7 +327,6 @@ struct Assignment }; } // namespace internal -#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/DiagonalProduct.h b/Eigen/src/Core/DiagonalProduct.h index c6dafdddc..d372b938f 100644 --- a/Eigen/src/Core/DiagonalProduct.h +++ b/Eigen/src/Core/DiagonalProduct.h @@ -13,122 +13,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { -template -struct traits > - : traits -{ - typedef typename scalar_product_traits::ReturnType Scalar; - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime, - MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, - MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, - -#ifndef EIGEN_TEST_EVALUATORS - _StorageOrder = MatrixType::Flags & RowMajorBit ? RowMajor : ColMajor, - _ScalarAccessOnDiag = !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft) - ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)), - _SameTypes = is_same::value, - // FIXME currently we need same types, but in the future the next rule should be the one - //_Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))), - _Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))), - _LinearAccessMask = (RowsAtCompileTime==1 || ColsAtCompileTime==1) ? LinearAccessBit : 0, - Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit, //(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit), - CoeffReadCost = NumTraits::MulCost + MatrixType::CoeffReadCost + DiagonalType::DiagonalVectorType::CoeffReadCost -#else - Flags = RowMajorBit & (unsigned int)(MatrixType::Flags) -#endif - }; -}; -} - -template -class DiagonalProduct : internal::no_assignment_operator, - public MatrixBase > -{ - public: - - typedef MatrixBase Base; - EIGEN_DENSE_PUBLIC_INTERFACE(DiagonalProduct) - - inline DiagonalProduct(const MatrixType& matrix, const DiagonalType& diagonal) - : m_matrix(matrix), m_diagonal(diagonal) - { - eigen_assert(diagonal.diagonal().size() == (ProductOrder == OnTheLeft ? matrix.rows() : matrix.cols())); - } - - EIGEN_STRONG_INLINE Index rows() const { return m_matrix.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return m_matrix.cols(); } - - EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const - { - return m_diagonal.diagonal().coeff(ProductOrder == OnTheLeft ? row : col) * m_matrix.coeff(row, col); - } - - EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const - { - enum { - StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor - }; - return coeff(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); - } - - template - EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const - { - enum { - StorageOrder = Flags & RowMajorBit ? RowMajor : ColMajor - }; - const Index indexInDiagonalVector = ProductOrder == OnTheLeft ? row : col; - return packet_impl(row,col,indexInDiagonalVector,typename internal::conditional< - ((int(StorageOrder) == RowMajor && int(ProductOrder) == OnTheLeft) - ||(int(StorageOrder) == ColMajor && int(ProductOrder) == OnTheRight)), internal::true_type, internal::false_type>::type()); - } - - template - EIGEN_STRONG_INLINE PacketScalar packet(Index idx) const - { - enum { - StorageOrder = int(MatrixType::Flags) & RowMajorBit ? RowMajor : ColMajor - }; - return packet(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); - } - - protected: - template - EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::true_type) const - { - return internal::pmul(m_matrix.template packet(row, col), - internal::pset1(m_diagonal.diagonal().coeff(id))); - } - - template - EIGEN_STRONG_INLINE PacketScalar packet_impl(Index row, Index col, Index id, internal::false_type) const - { - enum { - InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime, - DiagonalVectorPacketLoadMode = (LoadMode == Aligned && (((InnerSize%16) == 0) || (int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit)==AlignedBit) ? Aligned : Unaligned) - }; - return internal::pmul(m_matrix.template packet(row, col), - m_diagonal.diagonal().template packet(id)); - } - - typename MatrixType::Nested m_matrix; - typename DiagonalType::Nested m_diagonal; -}; - -/** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal. - */ -template -template -inline const DiagonalProduct -MatrixBase::operator*(const DiagonalBase &a_diagonal) const -{ - return DiagonalProduct(derived(), a_diagonal.derived()); -} -#else // EIGEN_TEST_EVALUATORS /** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal. */ template @@ -138,7 +22,6 @@ MatrixBase::operator*(const DiagonalBase &a_diagonal) { return Product(derived(),a_diagonal.derived()); } -#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h index d6441c6a5..68e9c2660 100644 --- a/Eigen/src/Core/Dot.h +++ b/Eigen/src/Core/Dot.h @@ -113,13 +113,7 @@ template inline const typename MatrixBase::PlainObject MatrixBase::normalized() const { -#ifndef EIGEN_TEST_EVALUATORS - typedef typename internal::nested::type Nested; - typedef typename internal::remove_reference::type _Nested; -#else typedef typename internal::nested_eval::type _Nested; -// typedef typename internal::remove_reference::type _Nested; -#endif // EIGEN_TEST_EVALUATORS _Nested n(derived()); return n / n.norm(); } @@ -211,13 +205,8 @@ template bool MatrixBase::isOrthogonal (const MatrixBase& other, const RealScalar& prec) const { -#ifndef EIGEN_TEST_EVALUATORS - typename internal::nested::type nested(derived()); - typename internal::nested::type otherNested(other.derived()); -#else typename internal::nested_eval::type nested(derived()); typename internal::nested_eval::type otherNested(other.derived()); -#endif return numext::abs2(nested.dot(otherNested)) <= prec * prec * nested.squaredNorm() * otherNested.squaredNorm(); } diff --git a/Eigen/src/Core/EigenBase.h b/Eigen/src/Core/EigenBase.h index 986e2a196..52b66e6dc 100644 --- a/Eigen/src/Core/EigenBase.h +++ b/Eigen/src/Core/EigenBase.h @@ -121,11 +121,7 @@ template template Derived& DenseBase::operator=(const EigenBase &other) { -#ifndef EIGEN_TEST_EVALUATORS - other.derived().evalTo(derived()); -#else call_assignment(derived(), other.derived()); -#endif return derived(); } @@ -133,11 +129,7 @@ template template Derived& DenseBase::operator+=(const EigenBase &other) { -#ifndef EIGEN_TEST_EVALUATORS - other.derived().addTo(derived()); -#else call_assignment(derived(), other.derived(), internal::add_assign_op()); -#endif return derived(); } @@ -145,11 +137,7 @@ template template Derived& DenseBase::operator-=(const EigenBase &other) { -#ifndef EIGEN_TEST_EVALUATORS - other.derived().subTo(derived()); -#else call_assignment(derived(), other.derived(), internal::sub_assign_op()); -#endif return derived(); } diff --git a/Eigen/src/Core/Fuzzy.h b/Eigen/src/Core/Fuzzy.h index 9c8d10683..8cd069a0d 100644 --- a/Eigen/src/Core/Fuzzy.h +++ b/Eigen/src/Core/Fuzzy.h @@ -23,13 +23,8 @@ struct isApprox_selector static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) { EIGEN_USING_STD_MATH(min); -#ifdef EIGEN_TEST_EVALUATORS typename internal::nested_eval::type nested(x); typename internal::nested_eval::type otherNested(y); -#else - typename internal::nested::type nested(x); - typename internal::nested::type otherNested(y); -#endif return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum()); } }; diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index abbf69549..e05ff8dce 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -13,31 +13,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS -/** \class GeneralProduct - * \ingroup Core_Module - * - * \brief Expression of the product of two general matrices or vectors - * - * \param LhsNested the type used to store the left-hand side - * \param RhsNested the type used to store the right-hand side - * \param ProductMode the type of the product - * - * This class represents an expression of the product of two general matrices. - * We call a general matrix, a dense matrix with full storage. For instance, - * This excludes triangular, selfadjoint, and sparse matrices. - * It is the return type of the operator* between general matrices. Its template - * arguments are determined automatically by ProductReturnType. Therefore, - * GeneralProduct should never be used direclty. To determine the result type of a - * function which involves a matrix product, use ProductReturnType::Type. - * - * \sa ProductReturnType, MatrixBase::operator*(const MatrixBase&) - */ -template::value> -class GeneralProduct; -#endif // EIGEN_TEST_EVALUATORS - - enum { Large = 2, Small = 3 @@ -156,56 +131,6 @@ template<> struct product_type_selector { enum } // end namespace internal -#ifndef EIGEN_TEST_EVALUATORS -/** \class ProductReturnType - * \ingroup Core_Module - * - * \brief Helper class to get the correct and optimized returned type of operator* - * - * \param Lhs the type of the left-hand side - * \param Rhs the type of the right-hand side - * \param ProductMode the type of the product (determined automatically by internal::product_mode) - * - * This class defines the typename Type representing the optimized product expression - * between two matrix expressions. In practice, using ProductReturnType::Type - * is the recommended way to define the result type of a function returning an expression - * which involve a matrix product. The class Product should never be - * used directly. - * - * \sa class Product, MatrixBase::operator*(const MatrixBase&) - */ -template -struct ProductReturnType -{ - // TODO use the nested type to reduce instanciations ???? -// typedef typename internal::nested::type LhsNested; -// typedef typename internal::nested::type RhsNested; - - typedef GeneralProduct Type; -}; - -template -struct ProductReturnType -{ - typedef typename internal::nested::type >::type LhsNested; - typedef typename internal::nested::type >::type RhsNested; - typedef CoeffBasedProduct Type; -}; - -template -struct ProductReturnType -{ - typedef typename internal::nested::type >::type LhsNested; - typedef typename internal::nested::type >::type RhsNested; - typedef CoeffBasedProduct Type; -}; - -// this is a workaround for sun CC -template -struct LazyProductReturnType : public ProductReturnType -{}; -#endif - /*********************************************************************** * Implementation of Inner Vector Vector Product ***********************************************************************/ @@ -216,124 +141,11 @@ struct LazyProductReturnType : public ProductReturnType with: operator=(Scalar x); -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct traits > - : traits::ReturnType,1,1> > -{}; - -} - -template -class GeneralProduct - : internal::no_assignment_operator, - public Matrix::ReturnType,1,1> -{ - typedef Matrix::ReturnType,1,1> Base; - public: - GeneralProduct(const Lhs& lhs, const Rhs& rhs) - { - EIGEN_STATIC_ASSERT((internal::is_same::value), - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - - Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum(); - } - - /** Convertion to scalar */ - operator const typename Base::Scalar() const { - return Base::coeff(0,0); - } -}; -#endif // EIGEN_TEST_EVALUATORS /*********************************************************************** * Implementation of Outer Vector Vector Product ***********************************************************************/ -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -// Column major -template -EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const false_type&) -{ - typedef typename Dest::Index Index; - // FIXME make sure lhs is sequentially stored - // FIXME not very good if rhs is real and lhs complex while alpha is real too - const Index cols = dest.cols(); - for (Index j=0; j -EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest& dest, const Func& func, const true_type&) { - typedef typename Dest::Index Index; - // FIXME make sure rhs is sequentially stored - // FIXME not very good if lhs is real and rhs complex while alpha is real too - const Index rows = dest.rows(); - for (Index i=0; i -struct traits > - : traits, Lhs, Rhs> > -{}; - -} - -template -class GeneralProduct - : public ProductBase, Lhs, Rhs> -{ - template struct IsRowMajor : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {}; - - public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) - - GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) - { - EIGEN_STATIC_ASSERT((internal::is_same::value), - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - } - - struct set { template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() = src; } }; - struct add { template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } }; - struct sub { template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } }; - struct adds { - Scalar m_scale; - adds(const Scalar& s) : m_scale(s) {} - template void operator()(const Dst& dst, const Src& src) const { - dst.const_cast_derived() += m_scale * src; - } - }; - - template - inline void evalTo(Dest& dest) const { - internal::outer_product_selector_run(*this, dest, set(), IsRowMajor()); - } - - template - inline void addTo(Dest& dest) const { - internal::outer_product_selector_run(*this, dest, add(), IsRowMajor()); - } - - template - inline void subTo(Dest& dest) const { - internal::outer_product_selector_run(*this, dest, sub(), IsRowMajor()); - } - - template void scaleAndAddTo(Dest& dest, const Scalar& alpha) const - { - internal::outer_product_selector_run(*this, dest, adds(alpha), IsRowMajor()); - } -}; - -#endif // EIGEN_TEST_EVALUATORS - /*********************************************************************** * Implementation of General Matrix Vector Product ***********************************************************************/ @@ -347,50 +159,11 @@ class GeneralProduct */ namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct traits > - : traits, Lhs, Rhs> > -{}; -template -struct gemv_selector; -#endif -#ifdef EIGEN_ENABLE_EVALUATORS template struct gemv_dense_sense_selector; -#endif } // end namespace internal -#ifndef EIGEN_TEST_EVALUATORS -template -class GeneralProduct - : public ProductBase, Lhs, Rhs> -{ - public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) - - typedef typename Lhs::Scalar LhsScalar; - typedef typename Rhs::Scalar RhsScalar; - - GeneralProduct(const Lhs& a_lhs, const Rhs& a_rhs) : Base(a_lhs,a_rhs) - { -// EIGEN_STATIC_ASSERT((internal::is_same::value), -// YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - } - - enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight }; - typedef typename internal::conditional::type MatrixType; - - template void scaleAndAddTo(Dest& dst, const Scalar& alpha) const - { - eigen_assert(m_lhs.rows() == dst.rows() && m_rhs.cols() == dst.cols()); - internal::gemv_selector::HasUsableDirectAccess)>::run(*this, dst, alpha); - } -}; -#endif - namespace internal { template struct gemv_static_vector_if; @@ -429,177 +202,6 @@ struct gemv_static_vector_if #endif }; -#ifndef EIGEN_TEST_EVALUATORS - -// The vector is on the left => transposition -template -struct gemv_selector -{ - template - static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha) - { - Transpose destT(dest); - enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor }; - gemv_selector - ::run(GeneralProduct,Transpose, GemvProduct> - (prod.rhs().transpose(), prod.lhs().transpose()), destT, alpha); - } -}; - -template<> struct gemv_selector -{ - template - static inline void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha) - { - typedef typename ProductType::Index Index; - typedef typename ProductType::LhsScalar LhsScalar; - typedef typename ProductType::RhsScalar RhsScalar; - typedef typename ProductType::Scalar ResScalar; - typedef typename ProductType::RealScalar RealScalar; - typedef typename ProductType::ActualLhsType ActualLhsType; - typedef typename ProductType::ActualRhsType ActualRhsType; - typedef typename ProductType::LhsBlasTraits LhsBlasTraits; - typedef typename ProductType::RhsBlasTraits RhsBlasTraits; - typedef Map, Aligned> MappedDest; - - ActualLhsType actualLhs = LhsBlasTraits::extract(prod.lhs()); - ActualRhsType actualRhs = RhsBlasTraits::extract(prod.rhs()); - - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) - * RhsBlasTraits::extractScalarFactor(prod.rhs()); - - enum { - // FIXME find a way to allow an inner stride on the result if packet_traits::size==1 - // on, the other hand it is good for the cache to pack the vector anyways... - EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1, - ComplexByReal = (NumTraits::IsComplex) && (!NumTraits::IsComplex), - MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal - }; - - gemv_static_vector_if static_dest; - - bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0)); - bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible; - - RhsScalar compatibleAlpha = get_factor::run(actualAlpha); - - ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), - evalToDest ? dest.data() : static_dest.data()); - - if(!evalToDest) - { - #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - Index size = dest.size(); - EIGEN_DENSE_STORAGE_CTOR_PLUGIN - #endif - if(!alphaIsCompatible) - { - MappedDest(actualDestPtr, dest.size()).setZero(); - compatibleAlpha = RhsScalar(1); - } - else - MappedDest(actualDestPtr, dest.size()) = dest; - } - - general_matrix_vector_product - ::run( - actualLhs.rows(), actualLhs.cols(), - actualLhs.data(), actualLhs.outerStride(), - actualRhs.data(), actualRhs.innerStride(), - actualDestPtr, 1, - compatibleAlpha); - - if (!evalToDest) - { - if(!alphaIsCompatible) - dest += actualAlpha * MappedDest(actualDestPtr, dest.size()); - else - dest = MappedDest(actualDestPtr, dest.size()); - } - } -}; - -template<> struct gemv_selector -{ - template - static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha) - { - typedef typename ProductType::LhsScalar LhsScalar; - typedef typename ProductType::RhsScalar RhsScalar; - typedef typename ProductType::Scalar ResScalar; - typedef typename ProductType::Index Index; - typedef typename ProductType::ActualLhsType ActualLhsType; - typedef typename ProductType::ActualRhsType ActualRhsType; - typedef typename ProductType::_ActualRhsType _ActualRhsType; - typedef typename ProductType::LhsBlasTraits LhsBlasTraits; - typedef typename ProductType::RhsBlasTraits RhsBlasTraits; - - typename add_const::type actualLhs = LhsBlasTraits::extract(prod.lhs()); - typename add_const::type actualRhs = RhsBlasTraits::extract(prod.rhs()); - - ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs()) - * RhsBlasTraits::extractScalarFactor(prod.rhs()); - - enum { - // FIXME find a way to allow an inner stride on the result if packet_traits::size==1 - // on, the other hand it is good for the cache to pack the vector anyways... - DirectlyUseRhs = _ActualRhsType::InnerStrideAtCompileTime==1 - }; - - gemv_static_vector_if static_rhs; - - ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(), - DirectlyUseRhs ? const_cast(actualRhs.data()) : static_rhs.data()); - - if(!DirectlyUseRhs) - { - #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - Index size = actualRhs.size(); - EIGEN_DENSE_STORAGE_CTOR_PLUGIN - #endif - Map(actualRhsPtr, actualRhs.size()) = actualRhs; - } - - general_matrix_vector_product - ::run( - actualLhs.rows(), actualLhs.cols(), - actualLhs.data(), actualLhs.outerStride(), - actualRhsPtr, 1, - dest.data(), dest.innerStride(), - actualAlpha); - } -}; - -template<> struct gemv_selector -{ - template - static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha) - { - typedef typename Dest::Index Index; - // TODO makes sure dest is sequentially stored in memory, otherwise use a temp - const Index size = prod.rhs().rows(); - for(Index k=0; k struct gemv_selector -{ - template - static void run(const ProductType& prod, Dest& dest, const typename ProductType::Scalar& alpha) - { - typedef typename Dest::Index Index; - // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp - const Index rows = prod.rows(); - for(Index i=0; i transposition template struct gemv_dense_sense_selector @@ -767,8 +369,6 @@ template<> struct gemv_dense_sense_selector } }; -#endif // EIGEN_ENABLE_EVALUATORS - } // end namespace internal /*************************************************************************** @@ -783,7 +383,6 @@ template<> struct gemv_dense_sense_selector */ #ifndef __CUDACC__ -#ifdef EIGEN_TEST_EVALUATORS template template inline const Product @@ -814,37 +413,6 @@ MatrixBase::operator*(const MatrixBase &other) const return Product(derived(), other.derived()); } -#else // EIGEN_TEST_EVALUATORS -template -template -inline const typename ProductReturnType::Type -MatrixBase::operator*(const MatrixBase &other) const -{ - // A note regarding the function declaration: In MSVC, this function will sometimes - // not be inlined since DenseStorage is an unwindable object for dynamic - // matrices and product types are holding a member to store the result. - // Thus it does not help tagging this function with EIGEN_STRONG_INLINE. - enum { - ProductIsValid = Derived::ColsAtCompileTime==Dynamic - || OtherDerived::RowsAtCompileTime==Dynamic - || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), - AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, - SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) - }; - // note to the lost user: - // * for a dot product use: v1.dot(v2) - // * for a coeff-wise product use: v1.cwiseProduct(v2) - EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), - INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) - EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), - INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) - EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) -#ifdef EIGEN_DEBUG_PRODUCT - internal::product_type::debug(); -#endif - return typename ProductReturnType::Type(derived(), other.derived()); -} -#endif // EIGEN_TEST_EVALUATORS #endif // __CUDACC__ @@ -859,7 +427,6 @@ MatrixBase::operator*(const MatrixBase &other) const * * \sa operator*(const MatrixBase&) */ -#ifdef EIGEN_TEST_EVALUATORS template template const Product @@ -883,31 +450,6 @@ MatrixBase::lazyProduct(const MatrixBase &other) const return Product(derived(), other.derived()); } -#else // EIGEN_TEST_EVALUATORS -template -template -const typename LazyProductReturnType::Type -MatrixBase::lazyProduct(const MatrixBase &other) const -{ - enum { - ProductIsValid = Derived::ColsAtCompileTime==Dynamic - || OtherDerived::RowsAtCompileTime==Dynamic - || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime), - AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime, - SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived) - }; - // note to the lost user: - // * for a dot product use: v1.dot(v2) - // * for a coeff-wise product use: v1.cwiseProduct(v2) - EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), - INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) - EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), - INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) - EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) - - return typename LazyProductReturnType::Type(derived(), other.derived()); -} -#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 84abd258e..5cfa7e50c 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -12,8 +12,6 @@ namespace Eigen { -#ifdef EIGEN_TEST_EVALUATORS - // TODO move the general declaration in Core, and rename this file DenseInverseImpl.h, or something like this... template class InverseImpl; @@ -127,8 +125,6 @@ protected: } // end namespace internal -#endif - } // end namespace Eigen #endif // EIGEN_INVERSE_H diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index 7dfdc3d59..87c1787bf 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -80,26 +80,8 @@ struct traits > ? int(PlainObjectType::OuterStrideAtCompileTime) : int(StrideType::OuterStrideAtCompileTime), IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned), -#ifndef EIGEN_TEST_EVALUATORS - HasNoInnerStride = InnerStrideAtCompileTime == 1, - HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0, - HasNoStride = HasNoInnerStride && HasNoOuterStride, - IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic, - KeepsPacketAccess = bool(HasNoInnerStride) - && ( bool(IsDynamicSize) - || HasNoOuterStride - || ( OuterStrideAtCompileTime!=Dynamic - && ((static_cast(sizeof(Scalar))*OuterStrideAtCompileTime)%EIGEN_ALIGN_BYTES)==0 ) ), - Flags0 = TraitsBase::Flags & (~NestByRefBit), - Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit), - Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime)) - ? int(Flags1) : int(Flags1 & ~LinearAccessBit), - Flags3 = is_lvalue::value ? int(Flags2) : (int(Flags2) & ~LvalueBit), - Flags = KeepsPacketAccess ? int(Flags3) : (int(Flags3) & ~PacketAccessBit) -#else Flags0 = TraitsBase::Flags & (~NestByRefBit), Flags = is_lvalue::value ? int(Flags0) : (int(Flags0) & ~LvalueBit) -#endif }; private: enum { Options }; // Expressions don't have Options diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index 591ea26fb..6d3b344e8 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -11,15 +11,9 @@ #ifndef EIGEN_MAPBASE_H #define EIGEN_MAPBASE_H -#ifndef EIGEN_TEST_EVALUATORS -#define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \ - EIGEN_STATIC_ASSERT((int(internal::traits::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \ - YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT) -#else #define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \ EIGEN_STATIC_ASSERT((int(internal::evaluator::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \ YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT) -#endif namespace Eigen { @@ -167,15 +161,7 @@ template class MapBase EIGEN_DEVICE_FUNC void checkSanity() const { -#ifndef EIGEN_TEST_EVALUATORS - // moved to evaluator - EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(internal::traits::Flags&PacketAccessBit, - internal::inner_stride_at_compile_time::ret==1), - PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1); - eigen_assert(EIGEN_IMPLIES(internal::traits::Flags&AlignedBit, (size_t(m_data) % EIGEN_ALIGN_BYTES) == 0) && "data is not aligned"); -#else eigen_assert(EIGEN_IMPLIES(internal::traits::IsAligned, (size_t(m_data) % EIGEN_ALIGN_BYTES) == 0) && "data is not aligned"); -#endif } PointerType m_data; diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 1daaabb07..8a5821548 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -115,12 +115,8 @@ struct traits > MaxRowsAtCompileTime = _MaxRows, MaxColsAtCompileTime = _MaxCols, Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret, -#ifndef EIGEN_TEST_EVALUATORS - CoeffReadCost = NumTraits::ReadCost, -#else // FIXME, the following flag in only used to define NeedsToAlign in PlainObjectBase EvaluatorFlags = compute_matrix_evaluator_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret, -#endif Options = _Options, InnerStrideAtCompileTime = 1, OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index bb49b9e84..9dbbd6fb5 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -66,9 +66,6 @@ template class MatrixBase using Base::MaxSizeAtCompileTime; using Base::IsVectorAtCompileTime; using Base::Flags; -#ifndef EIGEN_TEST_EVALUATORS - using Base::CoeffReadCost; -#endif using Base::derived; using Base::const_cast_derived; @@ -188,25 +185,15 @@ template class MatrixBase { return this->lazyProduct(other); } #else -#ifdef EIGEN_TEST_EVALUATORS template const Product operator*(const MatrixBase &other) const; -#else - template - const typename ProductReturnType::Type - operator*(const MatrixBase &other) const; -#endif #endif template EIGEN_DEVICE_FUNC -#ifdef EIGEN_TEST_EVALUATORS const Product -#else - const typename LazyProductReturnType::Type -#endif lazyProduct(const MatrixBase &other) const; template @@ -218,17 +205,10 @@ template class MatrixBase template void applyOnTheRight(const EigenBase& other); -#ifndef EIGEN_TEST_EVALUATORS - template - EIGEN_DEVICE_FUNC - const DiagonalProduct - operator*(const DiagonalBase &diagonal) const; -#else // EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC const Product operator*(const DiagonalBase &diagonal) const; -#endif // EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC @@ -347,19 +327,12 @@ template class MatrixBase NoAlias noalias(); -#ifndef EIGEN_TEST_EVALUATORS - inline const ForceAlignedAccess forceAlignedAccess() const; - inline ForceAlignedAccess forceAlignedAccess(); - template inline typename internal::add_const_on_value_type,Derived&>::type>::type forceAlignedAccessIf() const; - template inline typename internal::conditional,Derived&>::type forceAlignedAccessIf(); -#else // TODO forceAlignedAccess is temporarly disabled // Need to find a nicer workaround. inline const Derived& forceAlignedAccess() const { return derived(); } inline Derived& forceAlignedAccess() { return derived(); } template inline const Derived& forceAlignedAccessIf() const { return derived(); } template inline Derived& forceAlignedAccessIf() { return derived(); } -#endif Scalar trace() const; @@ -382,13 +355,9 @@ template class MatrixBase const PartialPivLU lu() const; - #ifdef EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC const Inverse inverse() const; - #else - EIGEN_DEVICE_FUNC - const internal::inverse_impl inverse() const; - #endif + template void computeInverseAndDetWithCheck( ResultType& inverse, diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index fe6dded60..097c9c062 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -35,7 +35,6 @@ class NoAlias NoAlias(ExpressionType& expression) : m_expression(expression) {} -#ifdef EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase& other) @@ -59,68 +58,6 @@ class NoAlias call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op()); return m_expression; } - -#else - - /** Behaves like MatrixBase::lazyAssign(other) - * \sa MatrixBase::lazyAssign() */ - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase& other) - { return internal::assign_selector::run(m_expression,other.derived()); } - - /** \sa MatrixBase::operator+= */ - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase& other) - { - typedef SelfCwiseBinaryOp, ExpressionType, OtherDerived> SelfAdder; - SelfAdder tmp(m_expression); - typedef typename internal::nested::type OtherDerivedNested; - typedef typename internal::remove_all::type _OtherDerivedNested; - internal::assign_selector::run(tmp,OtherDerivedNested(other.derived())); - return m_expression; - } - - /** \sa MatrixBase::operator-= */ - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase& other) - { - typedef SelfCwiseBinaryOp, ExpressionType, OtherDerived> SelfAdder; - SelfAdder tmp(m_expression); - typedef typename internal::nested::type OtherDerivedNested; - typedef typename internal::remove_all::type _OtherDerivedNested; - internal::assign_selector::run(tmp,OtherDerivedNested(other.derived())); - return m_expression; - } - -#ifndef EIGEN_PARSED_BY_DOXYGEN - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE ExpressionType& operator+=(const ProductBase& other) - { other.derived().addTo(m_expression); return m_expression; } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE ExpressionType& operator-=(const ProductBase& other) - { other.derived().subTo(m_expression); return m_expression; } - - template - EIGEN_STRONG_INLINE ExpressionType& operator+=(const CoeffBasedProduct& other) - { return m_expression.derived() += CoeffBasedProduct(other.lhs(), other.rhs()); } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE ExpressionType& operator-=(const CoeffBasedProduct& other) - { return m_expression.derived() -= CoeffBasedProduct(other.lhs(), other.rhs()); } - - template - ExpressionType& operator=(const ReturnByValue& func) - { return m_expression = func; } -#endif - -#endif EIGEN_DEVICE_FUNC ExpressionType& expression() const diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 31e0697a1..200518173 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -61,9 +61,6 @@ class PermutationBase : public EigenBase typedef typename Traits::IndicesType IndicesType; enum { Flags = Traits::Flags, -#ifndef EIGEN_TEST_EVALUATORS - CoeffReadCost = Traits::CoeffReadCost, -#endif RowsAtCompileTime = Traits::RowsAtCompileTime, ColsAtCompileTime = Traits::ColsAtCompileTime, MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime, @@ -291,9 +288,7 @@ class PermutationMatrix : public PermutationBase Traits; public: -#ifdef EIGEN_TEST_EVALUATORS typedef const PermutationMatrix& Nested; -#endif #ifndef EIGEN_PARSED_BY_DOXYGEN typedef typename Traits::IndicesType IndicesType; @@ -484,18 +479,9 @@ struct traits > enum { RowsAtCompileTime = _IndicesType::SizeAtCompileTime, ColsAtCompileTime = _IndicesType::SizeAtCompileTime, -#ifdef EIGEN_TEST_EVALUATORS MaxRowsAtCompileTime = IndicesType::MaxSizeAtCompileTime, MaxColsAtCompileTime = IndicesType::MaxSizeAtCompileTime, -#else - MaxRowsAtCompileTime = IndicesType::MaxRowsAtCompileTime, // is this a bug in Eigen 2.2 ? - MaxColsAtCompileTime = IndicesType::MaxColsAtCompileTime, -#endif Flags = 0 -#ifndef EIGEN_TEST_EVALUATORS - , - CoeffReadCost = _IndicesType::CoeffReadCost -#endif }; }; } @@ -524,7 +510,6 @@ class PermutationWrapper : public PermutationBase &permutation, (permutation.derived(), matrix.derived()); } -#else // EIGEN_TEST_EVALUATORS - -/** \returns the matrix with the permutation applied to the columns. - */ -template -inline const internal::permut_matrix_product_retval -operator*(const MatrixBase& matrix, - const PermutationBase &permutation) -{ - return internal::permut_matrix_product_retval - - (permutation.derived(), matrix.derived()); -} - -/** \returns the matrix with the permutation applied to the rows. - */ -template -inline const internal::permut_matrix_product_retval - -operator*(const PermutationBase &permutation, - const MatrixBase& matrix) -{ - return internal::permut_matrix_product_retval - - (permutation.derived(), matrix.derived()); -} - -#endif // EIGEN_TEST_EVALUATORS - namespace internal { template @@ -682,9 +638,6 @@ class Transpose > typedef typename Derived::DenseMatrixType DenseMatrixType; enum { Flags = Traits::Flags, -#ifndef EIGEN_TEST_EVALUATORS - CoeffReadCost = Traits::CoeffReadCost, -#endif RowsAtCompileTime = Traits::RowsAtCompileTime, ColsAtCompileTime = Traits::ColsAtCompileTime, MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime, @@ -713,8 +666,6 @@ class Transpose > DenseMatrixType toDenseMatrix() const { return *this; } -#ifdef EIGEN_TEST_EVALUATORS - /** \returns the matrix with the inverse permutation applied to the columns. */ template friend @@ -733,28 +684,6 @@ class Transpose > return Product(*this, matrix.derived()); } -#else // EIGEN_TEST_EVALUATORS - - /** \returns the matrix with the inverse permutation applied to the columns. - */ - template friend - inline const internal::permut_matrix_product_retval - operator*(const MatrixBase& matrix, const Transpose& trPerm) - { - return internal::permut_matrix_product_retval(trPerm.m_permutation, matrix.derived()); - } - - /** \returns the matrix with the inverse permutation applied to the rows. - */ - template - inline const internal::permut_matrix_product_retval - operator*(const MatrixBase& matrix) const - { - return internal::permut_matrix_product_retval(m_permutation, matrix.derived()); - } - -#endif // EIGEN_TEST_EVALUATORS - const PermutationType& nestedPermutation() const { return m_permutation; } protected: @@ -767,7 +696,6 @@ const PermutationWrapper MatrixBase::asPermutation() con return derived(); } -#ifdef EIGEN_TEST_EVALUATORS namespace internal { // TODO currently a permutation matrix expression has the form PermutationMatrix or PermutationWrapper @@ -799,7 +727,6 @@ struct evaluator_traits > > template<> struct AssignmentKind { typedef EigenBase2EigenBase Kind; }; } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 3637b6256..11aec1552 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -128,11 +128,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type DenseStorage m_storage; public: -#ifndef EIGEN_TEST_EVALUATORS - enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits::Flags & AlignedBit) != 0 }; -#else enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits::EvaluatorFlags & AlignedBit) != 0 }; -#endif EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) EIGEN_DEVICE_FUNC @@ -643,7 +639,6 @@ class PlainObjectBase : public internal::dense_xpr_base::type * * \internal */ -#ifdef EIGEN_TEST_EVALUATORS // aliasing is dealt once in internall::call_assignment // so at this stage we have to assume aliasing... and resising has to be done later. template @@ -654,23 +649,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type return this->derived(); return this->derived(); } -#else - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Derived& _set(const DenseBase& other) - { - _set_selector(other.derived(), typename internal::conditional(int(OtherDerived::Flags) & EvalBeforeAssigningBit), internal::true_type, internal::false_type>::type()); - return this->derived(); - } - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::true_type&) { _set_noalias(other.eval()); } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void _set_selector(const OtherDerived& other, const internal::false_type&) { _set_noalias(other); } -#endif /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which * is the case when creating a new matrix) so one can enforce lazy evaluation. * @@ -685,12 +664,8 @@ class PlainObjectBase : public internal::dense_xpr_base::type //_resize_to_match(other); // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because // it wouldn't allow to copy a row-vector into a column-vector. -#ifdef EIGEN_TEST_EVALUATORS internal::call_assignment_no_alias(this->derived(), other.derived(), internal::assign_op()); return this->derived(); -#else - return internal::assign_selector::run(this->derived(), other.derived()); -#endif } template diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 6825873d5..ae64d5200 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -79,11 +79,6 @@ struct traits > // FIXME: only needed by GeneralMatrixMatrixTriangular InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsTraits::ColsAtCompileTime, RhsTraits::RowsAtCompileTime), -#ifndef EIGEN_TEST_EVALUATORS - // dummy, for evaluators unit test only - CoeffReadCost = Dynamic, -#endif - // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator. Flags = ( (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) || ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit)) @@ -164,7 +159,6 @@ public: } // namespace internal -#ifdef EIGEN_TEST_EVALUATORS // Generic API dispatcher template class ProductImpl : public internal::generic_xpr_base, MatrixXpr, StorageKind>::type @@ -172,7 +166,6 @@ class ProductImpl : public internal::generic_xpr_base, M public: typedef typename internal::generic_xpr_base, MatrixXpr, StorageKind>::type Base; }; -#endif template class ProductImpl diff --git a/Eigen/src/Core/ProductBase.h b/Eigen/src/Core/ProductBase.h index 4b1fe356b..050343b2d 100644 --- a/Eigen/src/Core/ProductBase.h +++ b/Eigen/src/Core/ProductBase.h @@ -12,258 +12,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS - -/** \class ProductBase - * \ingroup Core_Module - * - */ - -namespace internal { -template -struct traits > -{ - typedef MatrixXpr XprKind; - typedef typename remove_all<_Lhs>::type Lhs; - typedef typename remove_all<_Rhs>::type Rhs; - typedef typename scalar_product_traits::ReturnType Scalar; - typedef typename product_promote_storage_type::StorageKind, - typename traits::StorageKind, - 0>::ret StorageKind; - typedef typename promote_index_type::Index, - typename traits::Index>::type Index; - enum { - RowsAtCompileTime = traits::RowsAtCompileTime, - ColsAtCompileTime = traits::ColsAtCompileTime, - MaxRowsAtCompileTime = traits::MaxRowsAtCompileTime, - MaxColsAtCompileTime = traits::MaxColsAtCompileTime, - Flags = (MaxRowsAtCompileTime==1 ? RowMajorBit : 0) - | EvalBeforeNestingBit | EvalBeforeAssigningBit | NestByRefBit, - // Note that EvalBeforeNestingBit and NestByRefBit - // are not used in practice because nested is overloaded for products - CoeffReadCost = 0 // FIXME why is it needed ? - }; -}; -} - -#define EIGEN_PRODUCT_PUBLIC_INTERFACE(Derived) \ - typedef ProductBase Base; \ - EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \ - typedef typename Base::LhsNested LhsNested; \ - typedef typename Base::_LhsNested _LhsNested; \ - typedef typename Base::LhsBlasTraits LhsBlasTraits; \ - typedef typename Base::ActualLhsType ActualLhsType; \ - typedef typename Base::_ActualLhsType _ActualLhsType; \ - typedef typename Base::RhsNested RhsNested; \ - typedef typename Base::_RhsNested _RhsNested; \ - typedef typename Base::RhsBlasTraits RhsBlasTraits; \ - typedef typename Base::ActualRhsType ActualRhsType; \ - typedef typename Base::_ActualRhsType _ActualRhsType; \ - using Base::m_lhs; \ - using Base::m_rhs; - -template -class ProductBase : public MatrixBase -{ - public: - typedef MatrixBase Base; - EIGEN_DENSE_PUBLIC_INTERFACE(ProductBase) - - typedef typename Lhs::Nested LhsNested; - typedef typename internal::remove_all::type _LhsNested; - typedef internal::blas_traits<_LhsNested> LhsBlasTraits; - typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; - typedef typename internal::remove_all::type _ActualLhsType; - typedef typename internal::traits::Scalar LhsScalar; - - typedef typename Rhs::Nested RhsNested; - typedef typename internal::remove_all::type _RhsNested; - typedef internal::blas_traits<_RhsNested> RhsBlasTraits; - typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType; - typedef typename internal::remove_all::type _ActualRhsType; - typedef typename internal::traits::Scalar RhsScalar; - - // Diagonal of a product: no need to evaluate the arguments because they are going to be evaluated only once - typedef CoeffBasedProduct FullyLazyCoeffBaseProductType; - - public: - - typedef typename Base::PlainObject PlainObject; - - ProductBase(const Lhs& a_lhs, const Rhs& a_rhs) - : m_lhs(a_lhs), m_rhs(a_rhs) - { - eigen_assert(a_lhs.cols() == a_rhs.rows() - && "invalid matrix product" - && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); - } - - inline Index rows() const { return m_lhs.rows(); } - inline Index cols() const { return m_rhs.cols(); } - - template - inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst,Scalar(1)); } - - template - inline void addTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(1)); } - - template - inline void subTo(Dest& dst) const { scaleAndAddTo(dst,Scalar(-1)); } - - template - inline void scaleAndAddTo(Dest& dst, const Scalar& alpha) const { derived().scaleAndAddTo(dst,alpha); } - - const _LhsNested& lhs() const { return m_lhs; } - const _RhsNested& rhs() const { return m_rhs; } - - // Implicit conversion to the nested type (trigger the evaluation of the product) - operator const PlainObject& () const - { - m_result.resize(m_lhs.rows(), m_rhs.cols()); - derived().evalTo(m_result); - return m_result; - } - - const Diagonal diagonal() const - { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); } - - template - const Diagonal diagonal() const - { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs); } - - const Diagonal diagonal(Index index) const - { return FullyLazyCoeffBaseProductType(m_lhs, m_rhs).diagonal(index); } - - // restrict coeff accessors to 1x1 expressions. No need to care about mutators here since this isn't an Lvalue expression - typename Base::CoeffReturnType coeff(Index row, Index col) const - { - EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) - eigen_assert(this->rows() == 1 && this->cols() == 1); - Matrix result = *this; - return result.coeff(row,col); - } - - typename Base::CoeffReturnType coeff(Index i) const - { - EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) - eigen_assert(this->rows() == 1 && this->cols() == 1); - Matrix result = *this; - return result.coeff(i); - } - - const Scalar& coeffRef(Index row, Index col) const - { - EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) - eigen_assert(this->rows() == 1 && this->cols() == 1); - return derived().coeffRef(row,col); - } - - const Scalar& coeffRef(Index i) const - { - EIGEN_STATIC_ASSERT_SIZE_1x1(Derived) - eigen_assert(this->rows() == 1 && this->cols() == 1); - return derived().coeffRef(i); - } - - protected: - - LhsNested m_lhs; - RhsNested m_rhs; - - mutable PlainObject m_result; -}; - -// here we need to overload the nested rule for products -// such that the nested type is a const reference to a plain matrix -namespace internal { -template -struct nested, N, PlainObject> -{ - typedef PlainObject const& type; -}; -} - -template -class ScaledProduct; - -// Note that these two operator* functions are not defined as member -// functions of ProductBase, because, otherwise we would have to -// define all overloads defined in MatrixBase. Furthermore, Using -// "using Base::operator*" would not work with MSVC. -// -// Also note that here we accept any compatible scalar types -template -const ScaledProduct -operator*(const ProductBase& prod, const typename Derived::Scalar& x) -{ return ScaledProduct(prod.derived(), x); } - -template -typename internal::enable_if::value, - const ScaledProduct >::type -operator*(const ProductBase& prod, const typename Derived::RealScalar& x) -{ return ScaledProduct(prod.derived(), x); } - - -template -const ScaledProduct -operator*(const typename Derived::Scalar& x,const ProductBase& prod) -{ return ScaledProduct(prod.derived(), x); } - -template -typename internal::enable_if::value, - const ScaledProduct >::type -operator*(const typename Derived::RealScalar& x,const ProductBase& prod) -{ return ScaledProduct(prod.derived(), x); } - -namespace internal { -template -struct traits > - : traits, - typename NestedProduct::_LhsNested, - typename NestedProduct::_RhsNested> > -{ - typedef typename traits::StorageKind StorageKind; -}; -} - -template -class ScaledProduct - : public ProductBase, - typename NestedProduct::_LhsNested, - typename NestedProduct::_RhsNested> -{ - public: - typedef ProductBase, - typename NestedProduct::_LhsNested, - typename NestedProduct::_RhsNested> Base; - typedef typename Base::Scalar Scalar; - typedef typename Base::PlainObject PlainObject; -// EIGEN_PRODUCT_PUBLIC_INTERFACE(ScaledProduct) - - ScaledProduct(const NestedProduct& prod, const Scalar& x) - : Base(prod.lhs(),prod.rhs()), m_prod(prod), m_alpha(x) {} - - template - inline void evalTo(Dest& dst) const { dst.setZero(); scaleAndAddTo(dst, Scalar(1)); } - - template - inline void addTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(1)); } - - template - inline void subTo(Dest& dst) const { scaleAndAddTo(dst, Scalar(-1)); } - - template - inline void scaleAndAddTo(Dest& dst, const Scalar& a_alpha) const { m_prod.derived().scaleAndAddTo(dst,a_alpha * m_alpha); } - - const Scalar& alpha() const { return m_alpha; } - - protected: - const NestedProduct& m_prod; - Scalar m_alpha; -}; - -#endif // EIGEN_TEST_EVALUATORS - /** \internal * Overloaded to perform an efficient C = (A*B).lazy() */ template diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 0aeae88bc..c6c355d43 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -69,11 +69,7 @@ public: #ifdef EIGEN_DEBUG_ASSIGN static void debug() { -#ifdef EIGEN_TEST_EVALUATORS std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl; -#else - std::cerr << "Xpr: " << typeid(Derived).name() << std::endl; -#endif std::cerr.setf(std::ios::hex, std::ios::basefield); EIGEN_DEBUG_VAR(Derived::Flags) std::cerr.unsetf(std::ios::hex); @@ -338,7 +334,6 @@ struct redux_impl } }; -#ifdef EIGEN_ENABLE_EVALUATORS // evaluator adaptor template class redux_evaluator @@ -395,7 +390,6 @@ protected: typename internal::evaluator::nestedType m_evaluator; const XprType &m_xpr; }; -#endif } // end namespace internal @@ -417,7 +411,6 @@ EIGEN_STRONG_INLINE typename internal::result_of::redux(const Func& func) const { eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix"); -#ifdef EIGEN_TEST_EVALUATORS // FIXME, eval_nest should be handled by redux_evaluator, however: // - it is currently difficult to provide the right Flags since they are still handled by the expressions @@ -433,13 +426,6 @@ DenseBase::redux(const Func& func) const ThisEvaluator thisEval(derived()); return internal::redux_impl::run(thisEval, func); - -#else - typedef typename internal::remove_all::type ThisNested; - - return internal::redux_impl - ::run(derived(), func); -#endif } /** \returns the minimum of all coefficients of \c *this. diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index 6390a8b64..09921c9e7 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -243,11 +243,7 @@ template class Ref< template void construct(const Expression& expr, internal::false_type) { -#ifdef EIGEN_TEST_EVALUATORS internal::call_assignment_no_alias(m_object,expr,internal::assign_op()); -#else - m_object.lazyAssign(expr); -#endif Base::construct(m_object); } diff --git a/Eigen/src/Core/Replicate.h b/Eigen/src/Core/Replicate.h index e63f0d421..3777049ee 100644 --- a/Eigen/src/Core/Replicate.h +++ b/Eigen/src/Core/Replicate.h @@ -54,13 +54,8 @@ struct traits > : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0 : (MatrixType::Flags & RowMajorBit) ? 1 : 0, -#ifndef EIGEN_TEST_EVALUATORS - Flags = (_MatrixTypeNested::Flags & HereditaryBits & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0), - CoeffReadCost = _MatrixTypeNested::CoeffReadCost -#else // FIXME enable DirectAccess with negative strides? Flags = IsRowMajor ? RowMajorBit : 0 -#endif }; }; } diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index 9d53fba27..f4e12a93b 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -33,25 +33,18 @@ struct traits > }; }; -#ifndef EIGEN_TEST_EVALUATORS /* The ReturnByValue object doesn't even have a coeff() method. * So the only way that nesting it in an expression can work, is by evaluating it into a plain matrix. * So internal::nested always gives the plain return matrix type. * * FIXME: I don't understand why we need this specialization: isn't this taken care of by the EvalBeforeNestingBit ?? + * Answer: EvalBeforeNestingBit should be deprecated since we have the evaluators */ template -struct nested, n, PlainObject> -{ - typedef typename traits::ReturnType type; -}; -#else -template struct nested_eval, n, PlainObject> { typedef typename traits::ReturnType type; }; -#endif } // end namespace internal @@ -93,7 +86,6 @@ Derived& DenseBase::operator=(const ReturnByValue& other) return derived(); } -#ifdef EIGEN_TEST_EVALUATORS namespace internal { // Expression is evaluated in a temporary; default implementation of Assignment is bypassed so that @@ -123,7 +115,6 @@ protected: }; } // end namespace internal -#endif } // end namespace Eigen diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h index ceb6e4701..01de90800 100644 --- a/Eigen/src/Core/Reverse.h +++ b/Eigen/src/Core/Reverse.h @@ -44,17 +44,7 @@ struct traits > ColsAtCompileTime = MatrixType::ColsAtCompileTime, MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, - -#ifndef EIGEN_TEST_EVALUATORS - // let's enable LinearAccess only with vectorization because of the product overhead - LinearAccess = ( (Direction==BothDirections) && (int(_MatrixTypeNested::Flags)&PacketAccessBit) ) - ? LinearAccessBit : 0, - - Flags = int(_MatrixTypeNested::Flags) & (HereditaryBits | LvalueBit | PacketAccessBit | LinearAccess), - CoeffReadCost = _MatrixTypeNested::CoeffReadCost -#else Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit) -#endif }; }; diff --git a/Eigen/src/Core/Select.h b/Eigen/src/Core/Select.h index d4fd88e62..0cb85a4ad 100644 --- a/Eigen/src/Core/Select.h +++ b/Eigen/src/Core/Select.h @@ -43,14 +43,7 @@ struct traits > ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime, MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime, -#ifndef EIGEN_TEST_EVALUATORS - Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & HereditaryBits, - CoeffReadCost = traits::type>::CoeffReadCost - + EIGEN_SIZE_MAX(traits::type>::CoeffReadCost, - traits::type>::CoeffReadCost) -#else Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & RowMajorBit -#endif }; }; } diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 546f61252..19cb232c9 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -40,20 +40,10 @@ struct traits > : traits Mode = UpLo | SelfAdjoint, Flags = MatrixTypeNestedCleaned::Flags & (HereditaryBits) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved -#ifndef EIGEN_TEST_EVALUATORS - , - CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost -#endif }; }; } -#ifndef EIGEN_TEST_EVALUATORS -template -struct SelfadjointProductMatrix; -#endif - // FIXME could also be called SelfAdjointWrapper to be consistent with DiagonalWrapper ?? template class SelfAdjointView : public TriangularBase > @@ -118,8 +108,6 @@ template class SelfAdjointView EIGEN_DEVICE_FUNC MatrixTypeNestedCleaned& nestedExpression() { return *const_cast(&m_matrix); } -#ifdef EIGEN_TEST_EVALUATORS - /** Efficient triangular matrix times vector/matrix product */ template EIGEN_DEVICE_FUNC @@ -145,31 +133,6 @@ template class SelfAdjointView return (s*mat.nestedExpression()).template selfadjointView(); } -#else // EIGEN_TEST_EVALUATORS - - /** Efficient self-adjoint matrix times vector/matrix product */ - template - EIGEN_DEVICE_FUNC - SelfadjointProductMatrix - operator*(const MatrixBase& rhs) const - { - return SelfadjointProductMatrix - - (m_matrix, rhs.derived()); - } - - /** Efficient vector/matrix times self-adjoint matrix product */ - template friend - EIGEN_DEVICE_FUNC - SelfadjointProductMatrix - operator*(const MatrixBase& lhs, const SelfAdjointView& rhs) - { - return SelfadjointProductMatrix - - (lhs.derived(),rhs.m_matrix); - } -#endif - /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this: * \f$ this = this + \alpha u v^* + conj(\alpha) v u^* \f$ * \returns a reference to \c *this @@ -231,104 +194,6 @@ template class SelfAdjointView namespace internal { -#ifndef EIGEN_TEST_EVALUATORS - -template -struct triangular_assignment_selector -{ - enum { - col = (UnrollCount-1) / Derived1::RowsAtCompileTime, - row = (UnrollCount-1) % Derived1::RowsAtCompileTime - }; - - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - triangular_assignment_selector::run(dst, src); - - if(row == col) - dst.coeffRef(row, col) = numext::real(src.coeff(row, col)); - else if(row < col) - dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col)); - } -}; - -template -struct triangular_assignment_selector -{ - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &, const Derived2 &) {} -}; - -template -struct triangular_assignment_selector -{ - enum { - col = (UnrollCount-1) / Derived1::RowsAtCompileTime, - row = (UnrollCount-1) % Derived1::RowsAtCompileTime - }; - - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - triangular_assignment_selector::run(dst, src); - - if(row == col) - dst.coeffRef(row, col) = numext::real(src.coeff(row, col)); - else if(row > col) - dst.coeffRef(col, row) = numext::conj(dst.coeffRef(row, col) = src.coeff(row, col)); - } -}; - -template -struct triangular_assignment_selector -{ - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &, const Derived2 &) {} -}; - -template -struct triangular_assignment_selector -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - for(Index j = 0; j < dst.cols(); ++j) - { - for(Index i = 0; i < j; ++i) - { - dst.copyCoeff(i, j, src); - dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j)); - } - dst.copyCoeff(j, j, src); - } - } -}; - -template -struct triangular_assignment_selector -{ - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - typedef typename Derived1::Index Index; - for(Index i = 0; i < dst.rows(); ++i) - { - for(Index j = 0; j < i; ++j) - { - dst.copyCoeff(i, j, src); - dst.coeffRef(j,i) = numext::conj(dst.coeff(i,j)); - } - dst.copyCoeff(i, i, src); - } - } -}; - -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_ENABLE_EVALUATORS - // TODO currently a selfadjoint expression has the form SelfAdjointView<.,.> // in the future selfadjoint-ness should be defined by the expression traits // such that Transpose > is valid. (currently TriangularBase::transpose() is overloaded to make it work) @@ -382,8 +247,6 @@ public: { eigen_internal_assert(false && "should never be called"); } }; -#endif // EIGEN_ENABLE_EVALUATORS - } // end namespace internal /*************************************************************************** diff --git a/Eigen/src/Core/SelfCwiseBinaryOp.h b/Eigen/src/Core/SelfCwiseBinaryOp.h index bec6f4968..38185d9d7 100644 --- a/Eigen/src/Core/SelfCwiseBinaryOp.h +++ b/Eigen/src/Core/SelfCwiseBinaryOp.h @@ -12,178 +12,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS - -/** \class SelfCwiseBinaryOp - * \ingroup Core_Module - * - * \internal - * - * \brief Internal helper class for optimizing operators like +=, -= - * - * This is a pseudo expression class re-implementing the copyCoeff/copyPacket - * method to directly performs a +=/-= operations in an optimal way. In particular, - * this allows to make sure that the input/output data are loaded only once using - * aligned packet loads. - * - * \sa class SwapWrapper for a similar trick. - */ - -namespace internal { -template -struct traits > - : traits > -{ - enum { - // Note that it is still a good idea to preserve the DirectAccessBit - // so that assign can correctly align the data. - Flags = traits >::Flags | (Lhs::Flags&AlignedBit) | (Lhs::Flags&DirectAccessBit) | (Lhs::Flags&LvalueBit), - OuterStrideAtCompileTime = Lhs::OuterStrideAtCompileTime, - InnerStrideAtCompileTime = Lhs::InnerStrideAtCompileTime - }; -}; -} - -template class SelfCwiseBinaryOp - : public internal::dense_xpr_base< SelfCwiseBinaryOp >::type -{ - public: - - typedef typename internal::dense_xpr_base::type Base; - EIGEN_DENSE_PUBLIC_INTERFACE(SelfCwiseBinaryOp) - - typedef typename internal::packet_traits::type Packet; - - EIGEN_DEVICE_FUNC - inline SelfCwiseBinaryOp(Lhs& xpr, const BinaryOp& func = BinaryOp()) : m_matrix(xpr), m_functor(func) {} - - EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } - EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } - EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_matrix.outerStride(); } - EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.innerStride(); } - EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_matrix.data(); } - - // note that this function is needed by assign to correctly align loads/stores - // TODO make Assign use .data() - EIGEN_DEVICE_FUNC - inline Scalar& coeffRef(Index row, Index col) - { - EIGEN_STATIC_ASSERT_LVALUE(Lhs) - return m_matrix.const_cast_derived().coeffRef(row, col); - } - EIGEN_DEVICE_FUNC - inline const Scalar& coeffRef(Index row, Index col) const - { - return m_matrix.coeffRef(row, col); - } - - // note that this function is needed by assign to correctly align loads/stores - // TODO make Assign use .data() - EIGEN_DEVICE_FUNC - inline Scalar& coeffRef(Index index) - { - EIGEN_STATIC_ASSERT_LVALUE(Lhs) - return m_matrix.const_cast_derived().coeffRef(index); - } - EIGEN_DEVICE_FUNC - inline const Scalar& coeffRef(Index index) const - { - return m_matrix.const_cast_derived().coeffRef(index); - } - - template - EIGEN_DEVICE_FUNC - void copyCoeff(Index row, Index col, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(row >= 0 && row < rows() - && col >= 0 && col < cols()); - Scalar& tmp = m_matrix.coeffRef(row,col); - tmp = m_functor(tmp, _other.coeff(row,col)); - } - - template - EIGEN_DEVICE_FUNC - void copyCoeff(Index index, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(index >= 0 && index < m_matrix.size()); - Scalar& tmp = m_matrix.coeffRef(index); - tmp = m_functor(tmp, _other.coeff(index)); - } - - template - void copyPacket(Index row, Index col, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(row >= 0 && row < rows() - && col >= 0 && col < cols()); - m_matrix.template writePacket(row, col, - m_functor.packetOp(m_matrix.template packet(row, col),_other.template packet(row, col)) ); - } - - template - void copyPacket(Index index, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(index >= 0 && index < m_matrix.size()); - m_matrix.template writePacket(index, - m_functor.packetOp(m_matrix.template packet(index),_other.template packet(index)) ); - } - - // reimplement lazyAssign to handle complex *= real - // see CwiseBinaryOp ctor for details - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE SelfCwiseBinaryOp& lazyAssign(const DenseBase& rhs) - { - EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs,RhsDerived) - EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename RhsDerived::Scalar); - - #ifdef EIGEN_DEBUG_ASSIGN - internal::assign_traits::debug(); - #endif - eigen_assert(rows() == rhs.rows() && cols() == rhs.cols()); - internal::assign_impl::run(*this,rhs.derived()); - #ifndef EIGEN_NO_DEBUG - this->checkTransposeAliasing(rhs.derived()); - #endif - return *this; - } - - // overloaded to honor evaluation of special matrices - // maybe another solution would be to not use SelfCwiseBinaryOp - // at first... - EIGEN_DEVICE_FUNC - SelfCwiseBinaryOp& operator=(const Rhs& _rhs) - { - typename internal::nested::type rhs(_rhs); - return Base::operator=(rhs); - } - - EIGEN_DEVICE_FUNC - Lhs& expression() const - { - return m_matrix; - } - - EIGEN_DEVICE_FUNC - const BinaryOp& functor() const - { - return m_functor; - } - - protected: - Lhs& m_matrix; - const BinaryOp& m_functor; - - private: - SelfCwiseBinaryOp& operator=(const SelfCwiseBinaryOp&); -}; - -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_TEST_EVALUATORS template inline Derived& DenseBase::operator*=(const Scalar& other) { @@ -215,43 +43,6 @@ inline Derived& DenseBase::operator/=(const Scalar& other) internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op()); return derived(); } -#else -template -inline Derived& DenseBase::operator*=(const Scalar& other) -{ - typedef typename Derived::PlainObject PlainObject; - SelfCwiseBinaryOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived()); - tmp = PlainObject::Constant(rows(),cols(),other); - return derived(); -} - -template -inline Derived& ArrayBase::operator+=(const Scalar& other) -{ - typedef typename Derived::PlainObject PlainObject; - SelfCwiseBinaryOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived()); - tmp = PlainObject::Constant(rows(),cols(),other); - return derived(); -} - -template -inline Derived& ArrayBase::operator-=(const Scalar& other) -{ - typedef typename Derived::PlainObject PlainObject; - SelfCwiseBinaryOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived()); - tmp = PlainObject::Constant(rows(),cols(),other); - return derived(); -} - -template -inline Derived& DenseBase::operator/=(const Scalar& other) -{ - typedef typename Derived::PlainObject PlainObject; - SelfCwiseBinaryOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived()); - tmp = PlainObject::Constant(rows(),cols(), other); - return derived(); -} -#endif } // end namespace Eigen diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index a1501c259..7b12be1e6 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -99,7 +99,6 @@ private: Scalar coeff(Index i) const; }; -#ifdef EIGEN_TEST_EVALUATORS // Generic API dispatcher template class SolveImpl : public internal::generic_xpr_base, MatrixXpr, StorageKind>::type @@ -107,7 +106,6 @@ class SolveImpl : public internal::generic_xpr_base public: typedef typename internal::generic_xpr_base, MatrixXpr, StorageKind>::type Base; }; -#endif namespace internal { diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index 9a1c5f4f8..3277cb5ba 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -12,135 +12,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS - -/** \class SwapWrapper - * \ingroup Core_Module - * - * \internal - * - * \brief Internal helper class for swapping two expressions - */ -namespace internal { -template -struct traits > : traits {}; -} - -template class SwapWrapper - : public internal::dense_xpr_base >::type -{ - public: - - typedef typename internal::dense_xpr_base::type Base; - EIGEN_DENSE_PUBLIC_INTERFACE(SwapWrapper) - typedef typename internal::packet_traits::type Packet; - - EIGEN_DEVICE_FUNC - inline SwapWrapper(ExpressionType& xpr) : m_expression(xpr) {} - - EIGEN_DEVICE_FUNC - inline Index rows() const { return m_expression.rows(); } - EIGEN_DEVICE_FUNC - inline Index cols() const { return m_expression.cols(); } - EIGEN_DEVICE_FUNC - inline Index outerStride() const { return m_expression.outerStride(); } - EIGEN_DEVICE_FUNC - inline Index innerStride() const { return m_expression.innerStride(); } - - typedef typename internal::conditional< - internal::is_lvalue::value, - Scalar, - const Scalar - >::type ScalarWithConstIfNotLvalue; - - EIGEN_DEVICE_FUNC - inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); } - EIGEN_DEVICE_FUNC - inline const Scalar* data() const { return m_expression.data(); } - - EIGEN_DEVICE_FUNC - inline Scalar& coeffRef(Index rowId, Index colId) - { - return m_expression.const_cast_derived().coeffRef(rowId, colId); - } - - EIGEN_DEVICE_FUNC - inline Scalar& coeffRef(Index index) - { - return m_expression.const_cast_derived().coeffRef(index); - } - - EIGEN_DEVICE_FUNC - inline Scalar& coeffRef(Index rowId, Index colId) const - { - return m_expression.coeffRef(rowId, colId); - } - - EIGEN_DEVICE_FUNC - inline Scalar& coeffRef(Index index) const - { - return m_expression.coeffRef(index); - } - - template - EIGEN_DEVICE_FUNC - void copyCoeff(Index rowId, Index colId, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(rowId >= 0 && rowId < rows() - && colId >= 0 && colId < cols()); - Scalar tmp = m_expression.coeff(rowId, colId); - m_expression.coeffRef(rowId, colId) = _other.coeff(rowId, colId); - _other.coeffRef(rowId, colId) = tmp; - } - - template - EIGEN_DEVICE_FUNC - void copyCoeff(Index index, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(index >= 0 && index < m_expression.size()); - Scalar tmp = m_expression.coeff(index); - m_expression.coeffRef(index) = _other.coeff(index); - _other.coeffRef(index) = tmp; - } - - template - void copyPacket(Index rowId, Index colId, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(rowId >= 0 && rowId < rows() - && colId >= 0 && colId < cols()); - Packet tmp = m_expression.template packet(rowId, colId); - m_expression.template writePacket(rowId, colId, - _other.template packet(rowId, colId) - ); - _other.template writePacket(rowId, colId, tmp); - } - - template - void copyPacket(Index index, const DenseBase& other) - { - OtherDerived& _other = other.const_cast_derived(); - eigen_internal_assert(index >= 0 && index < m_expression.size()); - Packet tmp = m_expression.template packet(index); - m_expression.template writePacket(index, - _other.template packet(index) - ); - _other.template writePacket(index, tmp); - } - - EIGEN_DEVICE_FUNC - ExpressionType& expression() const { return m_expression; } - - protected: - ExpressionType& m_expression; -}; - -#endif - -#ifdef EIGEN_ENABLE_EVALUATORS - namespace internal { // Overload default assignPacket behavior for swapping them @@ -189,8 +60,6 @@ public: } // namespace internal -#endif - } // end namespace Eigen #endif // EIGEN_SWAP_H diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index dd6180a8f..144bb2c01 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -42,18 +42,10 @@ struct traits > ColsAtCompileTime = MatrixType::RowsAtCompileTime, MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime, MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime, -#ifndef EIGEN_TEST_EVALUATORS FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit), Flags1 = Flags0 | FlagsLvalueBit, Flags = Flags1 ^ RowMajorBit, - CoeffReadCost = MatrixTypeNestedPlain::CoeffReadCost, -#else - FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, - Flags0 = MatrixTypeNestedPlain::Flags & ~(LvalueBit | NestByRefBit), - Flags1 = Flags0 | FlagsLvalueBit, - Flags = Flags1 ^ RowMajorBit, -#endif InnerStrideAtCompileTime = inner_stride_at_compile_time::ret, OuterStrideAtCompileTime = outer_stride_at_compile_time::ret }; @@ -109,7 +101,6 @@ struct TransposeImpl_base } // end namespace internal -#ifdef EIGEN_TEST_EVALUATORS // Generic API dispatcher template class TransposeImpl @@ -118,7 +109,6 @@ class TransposeImpl public: typedef typename internal::generic_xpr_base >::type Base; }; -#endif template class TransposeImpl : public internal::TransposeImpl_base::type @@ -141,59 +131,6 @@ template class TransposeImpl inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } inline const Scalar* data() const { return derived().nestedExpression().data(); } - -#ifndef EIGEN_TEST_EVALUATORS - - EIGEN_DEVICE_FUNC - inline ScalarWithConstIfNotLvalue& coeffRef(Index rowId, Index colId) - { - EIGEN_STATIC_ASSERT_LVALUE(MatrixType) - return derived().nestedExpression().const_cast_derived().coeffRef(colId, rowId); - } - - EIGEN_DEVICE_FUNC - inline ScalarWithConstIfNotLvalue& coeffRef(Index index) - { - EIGEN_STATIC_ASSERT_LVALUE(MatrixType) - return derived().nestedExpression().const_cast_derived().coeffRef(index); - } - - EIGEN_DEVICE_FUNC - inline CoeffReturnType coeff(Index rowId, Index colId) const - { - return derived().nestedExpression().coeff(colId, rowId); - } - - EIGEN_DEVICE_FUNC - inline CoeffReturnType coeff(Index index) const - { - return derived().nestedExpression().coeff(index); - } - - template - inline const PacketScalar packet(Index rowId, Index colId) const - { - return derived().nestedExpression().template packet(colId, rowId); - } - - template - inline void writePacket(Index rowId, Index colId, const PacketScalar& x) - { - derived().nestedExpression().const_cast_derived().template writePacket(colId, rowId, x); - } - - template - inline const PacketScalar packet(Index index) const - { - return derived().nestedExpression().template packet(index); - } - - template - inline void writePacket(Index index, const PacketScalar& x) - { - derived().nestedExpression().const_cast_derived().template writePacket(index, x); - } -#endif // FIXME: shall we keep the const version of coeffRef? EIGEN_DEVICE_FUNC @@ -446,14 +383,6 @@ void check_for_aliasing(const Dst &dst, const Src &src) } // end namespace internal -#ifndef EIGEN_TEST_EVALUATORS -template -template -void DenseBase::checkTransposeAliasing(const OtherDerived& other) const -{ - internal::checkTransposeAliasing_impl::run(derived(), other); -} -#endif // EIGEN_TEST_EVALUATORS #endif // EIGEN_NO_DEBUG } // end namespace Eigen diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index cc585bc6c..0d315dd50 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -32,9 +32,6 @@ template class TriangularBase : public EigenBase enum { Mode = internal::traits::Mode, -#ifndef EIGEN_TEST_EVALUATORS - CoeffReadCost = internal::traits::CoeffReadCost, -#endif RowsAtCompileTime = internal::traits::RowsAtCompileTime, ColsAtCompileTime = internal::traits::ColsAtCompileTime, MaxRowsAtCompileTime = internal::traits::MaxRowsAtCompileTime, @@ -177,21 +174,10 @@ struct traits > : traits enum { Mode = _Mode, Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) -#ifndef EIGEN_TEST_EVALUATORS - , - CoeffReadCost = MatrixTypeNestedCleaned::CoeffReadCost -#endif }; }; } -#ifndef EIGEN_TEST_EVALUATORS -template -struct TriangularProduct; -#endif - template class TriangularViewImpl; template class TriangularView @@ -270,7 +256,6 @@ template class TriangularView return m_matrix.transpose(); } -#ifdef EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC inline const Solve @@ -287,7 +272,6 @@ template class TriangularView #else using Base::solve; #endif -#endif // EIGEN_TEST_EVALUATORS EIGEN_DEVICE_FUNC const SelfAdjointView selfadjointView() const @@ -348,8 +332,6 @@ template class TriangularViewImpl<_Mat EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride(); } -#ifdef EIGEN_TEST_EVALUATORS - /** \sa MatrixBase::operator+=() */ template EIGEN_DEVICE_FUNC @@ -365,16 +347,6 @@ template class TriangularViewImpl<_Mat return derived(); } -#else - /** \sa MatrixBase::operator+=() */ - template - EIGEN_DEVICE_FUNC - TriangularViewType& operator+=(const DenseBase& other) { return *this = derived().nestedExpression() + other.derived(); } - /** \sa MatrixBase::operator-=() */ - template - EIGEN_DEVICE_FUNC - TriangularViewType& operator-=(const DenseBase& other) { return *this = derived().nestedExpression() - other.derived(); } -#endif /** \sa MatrixBase::operator*=() */ EIGEN_DEVICE_FUNC TriangularViewType& operator*=(const typename internal::traits::Scalar& other) { return *this = derived().nestedExpression() * other; } @@ -437,8 +409,6 @@ template class TriangularViewImpl<_Mat EIGEN_DEVICE_FUNC void lazyAssign(const MatrixBase& other); -#ifdef EIGEN_TEST_EVALUATORS - /** Efficient triangular matrix times vector/matrix product */ template EIGEN_DEVICE_FUNC @@ -456,30 +426,6 @@ template class TriangularViewImpl<_Mat { return Product(lhs.derived(),rhs.derived()); } - -#else // EIGEN_TEST_EVALUATORS - /** Efficient triangular matrix times vector/matrix product */ - template - EIGEN_DEVICE_FUNC - TriangularProduct - operator*(const MatrixBase& rhs) const - { - return TriangularProduct - - (derived().nestedExpression(), rhs.derived()); - } - - /** Efficient vector/matrix times triangular matrix product */ - template friend - EIGEN_DEVICE_FUNC - TriangularProduct - operator*(const MatrixBase& lhs, const TriangularViewImpl& rhs) - { - return TriangularProduct - - (lhs.derived(),rhs.derived().nestedExpression()); - } -#endif template EIGEN_DEVICE_FUNC @@ -490,14 +436,6 @@ template class TriangularViewImpl<_Mat EIGEN_DEVICE_FUNC void solveInPlace(const MatrixBase& other) const; -#ifndef EIGEN_TEST_EVALUATORS - template - EIGEN_DEVICE_FUNC - inline const internal::triangular_solve_retval - solve(const MatrixBase& other) const - { return solve(other); } -#endif // EIGEN_TEST_EVALUATORS - template EIGEN_DEVICE_FUNC void solveInPlace(const MatrixBase& other) const @@ -507,11 +445,7 @@ template class TriangularViewImpl<_Mat EIGEN_DEVICE_FUNC void swap(TriangularBase const & other) { - #ifdef EIGEN_TEST_EVALUATORS call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op()); - #else - TriangularView,Mode>(const_cast(derived().nestedExpression())).lazyAssign(other.const_cast_derived().nestedExpression()); - #endif } // TODO: this overload is ambiguous and it should be deprecated (Gael) @@ -519,65 +453,8 @@ template class TriangularViewImpl<_Mat EIGEN_DEVICE_FUNC void swap(MatrixBase const & other) { - #ifdef EIGEN_TEST_EVALUATORS call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op()); - #else - SwapWrapper swaper(const_cast(derived().nestedExpression())); - TriangularView,Mode>(swaper).lazyAssign(other.derived()); - #endif } - -#ifndef EIGEN_TEST_EVALUATORS - - // TODO simplify the following: - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularViewType& operator=(const ProductBase& other) - { - setZero(); - return assignProduct(other,1); - } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularViewType& operator+=(const ProductBase& other) - { - return assignProduct(other,1); - } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularViewType& operator-=(const ProductBase& other) - { - return assignProduct(other,-1); - } - - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularViewType& operator=(const ScaledProduct& other) - { - setZero(); - return assignProduct(other,other.alpha()); - } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularViewType& operator+=(const ScaledProduct& other) - { - return assignProduct(other,other.alpha()); - } - - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularViewType& operator-=(const ScaledProduct& other) - { - return assignProduct(other,-other.alpha()); - } - -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC @@ -590,194 +467,12 @@ template class TriangularViewImpl<_Mat template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TriangularViewType& _assignProduct(const ProductType& prod, const Scalar& alpha); - - protected: -#else - protected: - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TriangularViewType& assignProduct(const ProductBase& prod, const Scalar& alpha); -#endif }; /*************************************************************************** * Implementation of triangular evaluation/assignment ***************************************************************************/ -namespace internal { - -#ifndef EIGEN_TEST_EVALUATORS - -template -struct triangular_assignment_selector -{ - enum { - col = (UnrollCount-1) / Derived1::RowsAtCompileTime, - row = (UnrollCount-1) % Derived1::RowsAtCompileTime - }; - - typedef typename Derived1::Scalar Scalar; - - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - triangular_assignment_selector::run(dst, src); - - eigen_assert( Mode == Upper || Mode == Lower - || Mode == StrictlyUpper || Mode == StrictlyLower - || Mode == UnitUpper || Mode == UnitLower); - if((Mode == Upper && row <= col) - || (Mode == Lower && row >= col) - || (Mode == StrictlyUpper && row < col) - || (Mode == StrictlyLower && row > col) - || (Mode == UnitUpper && row < col) - || (Mode == UnitLower && row > col)) - dst.copyCoeff(row, col, src); - else if(ClearOpposite) - { - if (Mode&UnitDiag && row==col) - dst.coeffRef(row, col) = Scalar(1); - else - dst.coeffRef(row, col) = Scalar(0); - } - } -}; - -// prevent buggy user code from causing an infinite recursion -template -struct triangular_assignment_selector -{ - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &, const Derived2 &) {} -}; - -template -struct triangular_assignment_selector -{ - typedef typename Derived1::Index Index; - typedef typename Derived1::Scalar Scalar; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - for(Index j = 0; j < dst.cols(); ++j) - { - Index maxi = (std::min)(j, dst.rows()-1); - for(Index i = 0; i <= maxi; ++i) - dst.copyCoeff(i, j, src); - if (ClearOpposite) - for(Index i = maxi+1; i < dst.rows(); ++i) - dst.coeffRef(i, j) = Scalar(0); - } - } -}; - -template -struct triangular_assignment_selector -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - for(Index j = 0; j < dst.cols(); ++j) - { - for(Index i = j; i < dst.rows(); ++i) - dst.copyCoeff(i, j, src); - Index maxi = (std::min)(j, dst.rows()); - if (ClearOpposite) - for(Index i = 0; i < maxi; ++i) - dst.coeffRef(i, j) = static_cast(0); - } - } -}; - -template -struct triangular_assignment_selector -{ - typedef typename Derived1::Index Index; - typedef typename Derived1::Scalar Scalar; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - for(Index j = 0; j < dst.cols(); ++j) - { - Index maxi = (std::min)(j, dst.rows()); - for(Index i = 0; i < maxi; ++i) - dst.copyCoeff(i, j, src); - if (ClearOpposite) - for(Index i = maxi; i < dst.rows(); ++i) - dst.coeffRef(i, j) = Scalar(0); - } - } -}; - -template -struct triangular_assignment_selector -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - for(Index j = 0; j < dst.cols(); ++j) - { - for(Index i = j+1; i < dst.rows(); ++i) - dst.copyCoeff(i, j, src); - Index maxi = (std::min)(j, dst.rows()-1); - if (ClearOpposite) - for(Index i = 0; i <= maxi; ++i) - dst.coeffRef(i, j) = static_cast(0); - } - } -}; - -template -struct triangular_assignment_selector -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - for(Index j = 0; j < dst.cols(); ++j) - { - Index maxi = (std::min)(j, dst.rows()); - for(Index i = 0; i < maxi; ++i) - dst.copyCoeff(i, j, src); - if (ClearOpposite) - { - for(Index i = maxi+1; i < dst.rows(); ++i) - dst.coeffRef(i, j) = 0; - } - } - dst.diagonal().setOnes(); - } -}; -template -struct triangular_assignment_selector -{ - typedef typename Derived1::Index Index; - EIGEN_DEVICE_FUNC - static inline void run(Derived1 &dst, const Derived2 &src) - { - for(Index j = 0; j < dst.cols(); ++j) - { - Index maxi = (std::min)(j, dst.rows()); - for(Index i = maxi+1; i < dst.rows(); ++i) - dst.copyCoeff(i, j, src); - if (ClearOpposite) - { - for(Index i = 0; i < maxi; ++i) - dst.coeffRef(i, j) = 0; - } - } - dst.diagonal().setOnes(); - } -}; - -#endif // EIGEN_TEST_EVALUATORS - -} // end namespace internal - -#ifdef EIGEN_TEST_EVALUATORS - // FIXME should we keep that possibility template template @@ -816,84 +511,6 @@ void TriangularViewImpl::lazyAssign(const TriangularBas internal::call_assignment(derived().noalias(), other.derived()); } -#else - -// FIXME should we keep that possibility -template -template -inline TriangularView& -TriangularViewImpl::operator=(const MatrixBase& other) -{ - if(OtherDerived::Flags & EvalBeforeAssigningBit) - { - typename internal::plain_matrix_type::type other_evaluated(other.rows(), other.cols()); - other_evaluated.template triangularView().lazyAssign(other.derived()); - lazyAssign(other_evaluated); - } - else - lazyAssign(other.derived()); - return derived(); -} - -// FIXME should we keep that possibility -template -template -void TriangularViewImpl::lazyAssign(const MatrixBase& other) -{ - enum { - unroll = MatrixType::SizeAtCompileTime != Dynamic - && internal::traits::CoeffReadCost != Dynamic - && MatrixType::SizeAtCompileTime*internal::traits::CoeffReadCost/2 <= EIGEN_UNROLLING_LIMIT - }; - eigen_assert(derived().rows() == other.rows() && derived().cols() == other.cols()); - - internal::triangular_assignment_selector - ::run(derived().nestedExpression().const_cast_derived(), other.derived()); -} - - - -template -template -inline TriangularView& -TriangularViewImpl::operator=(const TriangularBase& other) -{ - eigen_assert(Mode == int(OtherDerived::Mode)); - if(internal::traits::Flags & EvalBeforeAssigningBit) - { - typename OtherDerived::DenseMatrixType other_evaluated(other.rows(), other.cols()); - other_evaluated.template triangularView().lazyAssign(other.derived().nestedExpression()); - lazyAssign(other_evaluated); - } - else - lazyAssign(other.derived().nestedExpression()); - return derived(); -} - -template -template -void TriangularViewImpl::lazyAssign(const TriangularBase& other) -{ - enum { - unroll = MatrixType::SizeAtCompileTime != Dynamic - && internal::traits::CoeffReadCost != Dynamic - && MatrixType::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 - <= EIGEN_UNROLLING_LIMIT - }; - eigen_assert(derived().rows() == other.rows() && derived().cols() == other.cols()); - - internal::triangular_assignment_selector - ::run(derived().nestedExpression().const_cast_derived(), other.derived().nestedExpression()); -} - -#endif // EIGEN_TEST_EVALUATORS - /*************************************************************************** * Implementation of TriangularBase methods ***************************************************************************/ @@ -914,31 +531,6 @@ void TriangularBase::evalTo(MatrixBase &other) const evalToLazy(other.derived()); } -#ifndef EIGEN_TEST_EVALUATORS - -/** Assigns a triangular or selfadjoint matrix to a dense matrix. - * If the matrix is triangular, the opposite part is set to zero. */ -template -template -void TriangularBase::evalToLazy(MatrixBase &other) const -{ - enum { - unroll = DenseDerived::SizeAtCompileTime != Dynamic - && internal::traits::CoeffReadCost != Dynamic - && DenseDerived::SizeAtCompileTime * internal::traits::CoeffReadCost / 2 - <= EIGEN_UNROLLING_LIMIT - }; - other.derived().resize(this->rows(), this->cols()); - - internal::triangular_assignment_selector - ::MatrixTypeNestedCleaned, Derived::Mode, - unroll ? int(DenseDerived::SizeAtCompileTime) : Dynamic, - true // clear the opposite triangular part - >::run(other.derived(), derived().nestedExpression()); -} - -#endif // EIGEN_TEST_EVALUATORS - /*************************************************************************** * Implementation of TriangularView methods ***************************************************************************/ @@ -1028,8 +620,6 @@ bool MatrixBase::isLowerTriangular(const RealScalar& prec) const } -#ifdef EIGEN_ENABLE_EVALUATORS - /*************************************************************************** **************************************************************************** * Evaluators and Assignment of triangular expressions @@ -1268,7 +858,6 @@ struct triangular_assignment_loop } // end namespace internal -#ifdef EIGEN_TEST_EVALUATORS /** Assigns a triangular or selfadjoint matrix to a dense matrix. * If the matrix is triangular, the opposite part is set to zero. */ template @@ -1315,11 +904,7 @@ struct Assignment, internal::sub_ass } }; - } // end namespace internal -#endif - -#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index 1a9eead43..a8130e902 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -48,25 +48,9 @@ struct traits > ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime, MaxRowsAtCompileTime = Direction==Vertical ? 1 : MatrixType::MaxRowsAtCompileTime, MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime, -#ifndef EIGEN_TEST_EVALUATORS - Flags0 = (unsigned int)_MatrixTypeNested::Flags & HereditaryBits, - Flags = (Flags0 & ~RowMajorBit) | (RowsAtCompileTime == 1 ? RowMajorBit : 0), -#else Flags = RowsAtCompileTime == 1 ? RowMajorBit : 0, -#endif TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime : MatrixType::ColsAtCompileTime }; -#ifndef EIGEN_TEST_EVALUATORS - #if EIGEN_GNUC_AT_LEAST(3,4) - typedef typename MemberOp::template Cost CostOpType; - #else - typedef typename MemberOp::template Cost CostOpType; - #endif - enum { - CoeffReadCost = TraversalSize==Dynamic ? Dynamic - : TraversalSize * traits<_MatrixTypeNested>::CoeffReadCost + int(CostOpType::value) - }; -#endif }; } diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h index 76d452d9a..810ec28e7 100644 --- a/Eigen/src/Core/Visitor.h +++ b/Eigen/src/Core/Visitor.h @@ -53,7 +53,6 @@ struct visitor_impl } }; -#ifdef EIGEN_ENABLE_EVALUATORS // evaluator adaptor template class visitor_evaluator @@ -81,7 +80,6 @@ protected: typename internal::evaluator::nestedType m_evaluator; const XprType &m_xpr; }; -#endif } // end namespace internal /** Applies the visitor \a visitor to the whole coefficients of the matrix or vector. @@ -105,7 +103,6 @@ template template void DenseBase::visit(Visitor& visitor) const { -#ifdef EIGEN_TEST_EVALUATORS typedef typename internal::visitor_evaluator ThisEvaluator; ThisEvaluator thisEval(derived()); @@ -117,16 +114,6 @@ void DenseBase::visit(Visitor& visitor) const return internal::visitor_impl::run(thisEval, visitor); -#else - enum { unroll = SizeAtCompileTime != Dynamic - && CoeffReadCost != Dynamic - && (SizeAtCompileTime == 1 || internal::functor_traits::Cost != Dynamic) - && SizeAtCompileTime * CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits::Cost - <= EIGEN_UNROLLING_LIMIT }; - return internal::visitor_impl::run(derived(), visitor); -#endif } namespace internal { diff --git a/Eigen/src/Core/products/CoeffBasedProduct.h b/Eigen/src/Core/products/CoeffBasedProduct.h deleted file mode 100644 index 76806fd62..000000000 --- a/Eigen/src/Core/products/CoeffBasedProduct.h +++ /dev/null @@ -1,460 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2006-2008 Benoit Jacob -// Copyright (C) 2008-2010 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_COEFFBASED_PRODUCT_H -#define EIGEN_COEFFBASED_PRODUCT_H - -namespace Eigen { - -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -/********************************************************************************* -* Coefficient based product implementation. -* It is designed for the following use cases: -* - small fixed sizes -* - lazy products -*********************************************************************************/ - -/* Since the all the dimensions of the product are small, here we can rely - * on the generic Assign mechanism to evaluate the product per coeff (or packet). - * - * Note that here the inner-loops should always be unrolled. - */ - -template -struct product_coeff_impl; - -template -struct product_packet_impl; - -template -struct traits > -{ - typedef MatrixXpr XprKind; - typedef typename remove_all::type _LhsNested; - typedef typename remove_all::type _RhsNested; - typedef typename scalar_product_traits::ReturnType Scalar; - typedef typename product_promote_storage_type::StorageKind, - typename traits<_RhsNested>::StorageKind, - 0>::ret StorageKind; - typedef typename promote_index_type::Index, - typename traits<_RhsNested>::Index>::type Index; - - enum { - LhsFlags = traits<_LhsNested>::Flags, - RhsFlags = traits<_RhsNested>::Flags, - - RowsAtCompileTime = _LhsNested::RowsAtCompileTime, - ColsAtCompileTime = _RhsNested::ColsAtCompileTime, - InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(_LhsNested::ColsAtCompileTime, _RhsNested::RowsAtCompileTime), - - MaxRowsAtCompileTime = _LhsNested::MaxRowsAtCompileTime, - MaxColsAtCompileTime = _RhsNested::MaxColsAtCompileTime, - - LhsRowMajor = LhsFlags & RowMajorBit, - RhsRowMajor = RhsFlags & RowMajorBit, - - SameType = is_same::value, - - CanVectorizeRhs = RhsRowMajor && (RhsFlags & PacketAccessBit) - && (ColsAtCompileTime == Dynamic - || ( (ColsAtCompileTime % packet_traits::size) == 0 - && (RhsFlags&AlignedBit) - ) - ), - - CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit) - && (RowsAtCompileTime == Dynamic - || ( (RowsAtCompileTime % packet_traits::size) == 0 - && (LhsFlags&AlignedBit) - ) - ), - - EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1 - : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0 - : (RhsRowMajor && !CanVectorizeLhs), - - Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit) - | (EvalToRowMajor ? RowMajorBit : 0) - | NestingFlags - | (CanVectorizeLhs ? (LhsFlags & AlignedBit) : 0) - | (CanVectorizeRhs ? (RhsFlags & AlignedBit) : 0) - // TODO enable vectorization for mixed types - | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0), -#ifndef EIGEN_TEST_EVALUATORS - LhsCoeffReadCost = traits<_LhsNested>::CoeffReadCost, - RhsCoeffReadCost = traits<_RhsNested>::CoeffReadCost, - CoeffReadCost = (InnerSize == Dynamic || LhsCoeffReadCost==Dynamic || RhsCoeffReadCost==Dynamic || NumTraits::AddCost==Dynamic || NumTraits::MulCost==Dynamic) ? Dynamic - : InnerSize * (NumTraits::MulCost + LhsCoeffReadCost + RhsCoeffReadCost) - + (InnerSize - 1) * NumTraits::AddCost, -#endif - /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside - * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner - * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect - * the Flags, it is safe to make this value depend on ActualPacketAccessBit, that doesn't affect the ABI. - */ - CanVectorizeInner = SameType - && LhsRowMajor - && (!RhsRowMajor) - && (LhsFlags & RhsFlags & ActualPacketAccessBit) - && (LhsFlags & RhsFlags & AlignedBit) - && (InnerSize % packet_traits::size == 0) - }; -}; - -} // end namespace internal - -#ifndef EIGEN_TEST_EVALUATORS - -template -class CoeffBasedProduct - : internal::no_assignment_operator, - public MatrixBase > -{ - public: - - typedef MatrixBase Base; - EIGEN_DENSE_PUBLIC_INTERFACE(CoeffBasedProduct) - typedef typename Base::PlainObject PlainObject; - - private: - - typedef typename internal::traits::_LhsNested _LhsNested; - typedef typename internal::traits::_RhsNested _RhsNested; - - enum { - PacketSize = internal::packet_traits::size, - InnerSize = internal::traits::InnerSize, - Unroll = CoeffReadCost != Dynamic && CoeffReadCost <= EIGEN_UNROLLING_LIMIT, - CanVectorizeInner = internal::traits::CanVectorizeInner - }; - - typedef internal::product_coeff_impl ScalarCoeffImpl; - - typedef CoeffBasedProduct LazyCoeffBasedProductType; - - public: - - EIGEN_DEVICE_FUNC - inline CoeffBasedProduct(const CoeffBasedProduct& other) - : Base(), m_lhs(other.m_lhs), m_rhs(other.m_rhs) - {} - - template - EIGEN_DEVICE_FUNC - inline CoeffBasedProduct(const Lhs& lhs, const Rhs& rhs) - : m_lhs(lhs), m_rhs(rhs) - { - // we don't allow taking products of matrices of different real types, as that wouldn't be vectorizable. - // We still allow to mix T and complex. - EIGEN_STATIC_ASSERT((internal::scalar_product_traits::Defined), - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - eigen_assert(lhs.cols() == rhs.rows() - && "invalid matrix product" - && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); - } - - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); } - - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const - { - Scalar res; - ScalarCoeffImpl::run(row, col, m_lhs, m_rhs, res); - return res; - } - - /* Allow index-based non-packet access. It is impossible though to allow index-based packed access, - * which is why we don't set the LinearAccessBit. - */ - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const Scalar coeff(Index index) const - { - Scalar res; - const Index row = RowsAtCompileTime == 1 ? 0 : index; - const Index col = RowsAtCompileTime == 1 ? index : 0; - ScalarCoeffImpl::run(row, col, m_lhs, m_rhs, res); - return res; - } - - template - EIGEN_STRONG_INLINE const PacketScalar packet(Index row, Index col) const - { - PacketScalar res; - internal::product_packet_impl - ::run(row, col, m_lhs, m_rhs, res); - return res; - } - - // Implicit conversion to the nested type (trigger the evaluation of the product) - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE operator const PlainObject& () const - { - m_result.lazyAssign(*this); - return m_result; - } - - EIGEN_DEVICE_FUNC const _LhsNested& lhs() const { return m_lhs; } - EIGEN_DEVICE_FUNC const _RhsNested& rhs() const { return m_rhs; } - - EIGEN_DEVICE_FUNC - const Diagonal diagonal() const - { return reinterpret_cast(*this); } - - template - EIGEN_DEVICE_FUNC - const Diagonal diagonal() const - { return reinterpret_cast(*this); } - - EIGEN_DEVICE_FUNC - const Diagonal diagonal(Index index) const - { return reinterpret_cast(*this).diagonal(index); } - - protected: - typename internal::add_const_on_value_type::type m_lhs; - typename internal::add_const_on_value_type::type m_rhs; - - mutable PlainObject m_result; -}; - -namespace internal { - -// here we need to overload the nested rule for products -// such that the nested type is a const reference to a plain matrix -template -struct nested, N, PlainObject> -{ - typedef PlainObject const& type; -}; - -/*************************************************************************** -* Normal product .coeff() implementation (with meta-unrolling) -***************************************************************************/ - -/************************************** -*** Scalar path - no vectorization *** -**************************************/ - -template -struct product_coeff_impl -{ - typedef typename Lhs::Index Index; - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res) - { - product_coeff_impl::run(row, col, lhs, rhs, res); - res += lhs.coeff(row, UnrollingIndex) * rhs.coeff(UnrollingIndex, col); - } -}; - -template -struct product_coeff_impl -{ - typedef typename Lhs::Index Index; - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res) - { - res = lhs.coeff(row, 0) * rhs.coeff(0, col); - } -}; - -template -struct product_coeff_impl -{ - typedef typename Lhs::Index Index; - EIGEN_DEVICE_FUNC - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar& res) - { - eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix"); - res = lhs.coeff(row, 0) * rhs.coeff(0, col); - for(Index i = 1; i < lhs.cols(); ++i) - res += lhs.coeff(row, i) * rhs.coeff(i, col); - } -}; - -/******************************************* -*** Scalar path with inner vectorization *** -*******************************************/ - -template -struct product_coeff_vectorized_unroller -{ - typedef typename Lhs::Index Index; - enum { PacketSize = packet_traits::size }; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres) - { - product_coeff_vectorized_unroller::run(row, col, lhs, rhs, pres); - pres = padd(pres, pmul( lhs.template packet(row, UnrollingIndex) , rhs.template packet(UnrollingIndex, col) )); - } -}; - -template -struct product_coeff_vectorized_unroller<0, Lhs, Rhs, Packet> -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::PacketScalar &pres) - { - pres = pmul(lhs.template packet(row, 0) , rhs.template packet(0, col)); - } -}; - -template -struct product_coeff_impl -{ - typedef typename Lhs::PacketScalar Packet; - typedef typename Lhs::Index Index; - enum { PacketSize = packet_traits::size }; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res) - { - Packet pres; - product_coeff_vectorized_unroller::run(row, col, lhs, rhs, pres); - res = predux(pres); - } -}; - -template -struct product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res) - { - res = lhs.row(row).transpose().cwiseProduct(rhs.col(col)).sum(); - } -}; - -// NOTE the 3 following specializations are because taking .col(0) on a vector is a bit slower -// NOTE maybe they are now useless since we have a specialization for Block -template -struct product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index /*row*/, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res) - { - res = lhs.transpose().cwiseProduct(rhs.col(col)).sum(); - } -}; - -template -struct product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res) - { - res = lhs.row(row).transpose().cwiseProduct(rhs).sum(); - } -}; - -template -struct product_coeff_vectorized_dyn_selector -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res) - { - res = lhs.transpose().cwiseProduct(rhs).sum(); - } -}; - -template -struct product_coeff_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar &res) - { - product_coeff_vectorized_dyn_selector::run(row, col, lhs, rhs, res); - } -}; - -/******************* -*** Packet path *** -*******************/ - -template -struct product_packet_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res) - { - product_packet_impl::run(row, col, lhs, rhs, res); - res = pmadd(pset1(lhs.coeff(row, UnrollingIndex)), rhs.template packet(UnrollingIndex, col), res); - } -}; - -template -struct product_packet_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res) - { - product_packet_impl::run(row, col, lhs, rhs, res); - res = pmadd(lhs.template packet(row, UnrollingIndex), pset1(rhs.coeff(UnrollingIndex, col)), res); - } -}; - -template -struct product_packet_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res) - { - res = pmul(pset1(lhs.coeff(row, 0)),rhs.template packet(0, col)); - } -}; - -template -struct product_packet_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res) - { - res = pmul(lhs.template packet(row, 0), pset1(rhs.coeff(0, col))); - } -}; - -template -struct product_packet_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res) - { - eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix"); - res = pmul(pset1(lhs.coeff(row, 0)),rhs.template packet(0, col)); - for(Index i = 1; i < lhs.cols(); ++i) - res = pmadd(pset1(lhs.coeff(row, i)), rhs.template packet(i, col), res); - } -}; - -template -struct product_packet_impl -{ - typedef typename Lhs::Index Index; - static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res) - { - eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix"); - res = pmul(lhs.template packet(row, 0), pset1(rhs.coeff(0, col))); - for(Index i = 1; i < lhs.cols(); ++i) - res = pmadd(lhs.template packet(row, i), pset1(rhs.coeff(i, col)), res); - } -}; - -} // end namespace internal - -#endif // EIGEN_TEST_EVALUATORS - -#endif // -class GeneralProduct - : public ProductBase, Lhs, Rhs> -{ - enum { - MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime) - }; - public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct) - - typedef typename Lhs::Scalar LhsScalar; - typedef typename Rhs::Scalar RhsScalar; - typedef Scalar ResScalar; - - GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) - { - typedef internal::scalar_product_op BinOp; - EIGEN_CHECK_BINARY_COMPATIBILIY(BinOp,LhsScalar,RhsScalar); - } - - template - inline void evalTo(Dest& dst) const - { - if((m_rhs.rows()+dst.rows()+dst.cols())<20 && m_rhs.rows()>0) - dst.noalias() = m_lhs .lazyProduct( m_rhs ); - else - { - dst.setZero(); - scaleAndAddTo(dst,Scalar(1)); - } - } - - template - inline void addTo(Dest& dst) const - { - if((m_rhs.rows()+dst.rows()+dst.cols())<20 && m_rhs.rows()>0) - dst.noalias() += m_lhs .lazyProduct( m_rhs ); - else - scaleAndAddTo(dst,Scalar(1)); - } - - template - inline void subTo(Dest& dst) const - { - if((m_rhs.rows()+dst.rows()+dst.cols())<20 && m_rhs.rows()>0) - dst.noalias() -= m_lhs .lazyProduct( m_rhs ); - else - scaleAndAddTo(dst,Scalar(-1)); - } - - template void scaleAndAddTo(Dest& dst, const Scalar& alpha) const - { - eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols()); - - typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(m_lhs); - typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(m_rhs); - - Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs) - * RhsBlasTraits::extractScalarFactor(m_rhs); - - typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar, - Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType; - - typedef internal::gemm_functor< - Scalar, Index, - internal::general_matrix_matrix_product< - Index, - LhsScalar, (_ActualLhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate), - RhsScalar, (_ActualRhsType::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate), - (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>, - _ActualLhsType, _ActualRhsType, Dest, BlockingType> GemmFunctor; - - BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), true); - - internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit); - } -}; -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template @@ -534,7 +453,6 @@ struct generic_product_impl }; } // end namespace internal -#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 06c64714a..7db3e3d38 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -261,7 +261,6 @@ struct general_product_to_triangular_selector } }; -#ifdef EIGEN_TEST_EVALUATORS template template TriangularView& TriangularViewImpl::_assignProduct(const ProductType& prod, const Scalar& alpha) @@ -272,19 +271,7 @@ TriangularView& TriangularViewImpl::_ass return derived(); } -#else -template -template -TriangularView& TriangularViewImpl::assignProduct(const ProductBase& prod, const Scalar& alpha) -{ - eigen_assert(derived().rows() == prod.rows() && derived().cols() == prod.cols()); - general_product_to_triangular_selector - ::run(derived().nestedExpression().const_cast_derived(), prod.derived(), alpha); - - return derived(); -} -#endif } // end namespace Eigen #endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index 0ab3f3a56..4e507b6cf 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -459,58 +459,6 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix -struct traits > - : traits, Lhs, Rhs> > -{}; -} - -template -struct SelfadjointProductMatrix - : public ProductBase, Lhs, Rhs > -{ - EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix) - - SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - enum { - LhsIsUpper = (LhsMode&(Upper|Lower))==Upper, - LhsIsSelfAdjoint = (LhsMode&SelfAdjoint)==SelfAdjoint, - RhsIsUpper = (RhsMode&(Upper|Lower))==Upper, - RhsIsSelfAdjoint = (RhsMode&SelfAdjoint)==SelfAdjoint - }; - - template void scaleAndAddTo(Dest& dst, const Scalar& alpha) const - { - eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols()); - - typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(m_lhs); - typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(m_rhs); - - Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs) - * RhsBlasTraits::extractScalarFactor(m_rhs); - - internal::product_selfadjoint_matrix::Flags &RowMajorBit) ? RowMajor : ColMajor, LhsIsSelfAdjoint, - NumTraits::IsComplex && EIGEN_LOGICAL_XOR(LhsIsUpper,bool(LhsBlasTraits::NeedToConjugate)), - EIGEN_LOGICAL_XOR(RhsIsUpper, - internal::traits::Flags &RowMajorBit) ? RowMajor : ColMajor, RhsIsSelfAdjoint, - NumTraits::IsComplex && EIGEN_LOGICAL_XOR(RhsIsUpper,bool(RhsBlasTraits::NeedToConjugate)), - internal::traits::Flags&RowMajorBit ? RowMajor : ColMajor> - ::run( - lhs.rows(), rhs.cols(), // sizes - &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info - &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info - &dst.coeffRef(0,0), dst.outerStride(), // result info - actualAlpha // alpha - ); - } -}; -#endif // EIGEN_TEST_EVALUATORS -#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template @@ -560,8 +508,6 @@ struct selfadjoint_product_impl } // end namespace internal -#endif // EIGEN_ENABLE_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_H diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index 020205c12..d9c041f0c 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -168,117 +168,6 @@ EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product -struct traits > - : traits, Lhs, Rhs> > -{}; -} - -template -struct SelfadjointProductMatrix - : public ProductBase, Lhs, Rhs > -{ - EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix) - - enum { - LhsUpLo = LhsMode&(Upper|Lower) - }; - - SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - template void scaleAndAddTo(Dest& dest, const Scalar& alpha) const - { - typedef typename Dest::Scalar ResScalar; - typedef typename Base::RhsScalar RhsScalar; - typedef Map, Aligned> MappedDest; - - eigen_assert(dest.rows()==m_lhs.rows() && dest.cols()==m_rhs.cols()); - - typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(m_lhs); - typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(m_rhs); - - Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs) - * RhsBlasTraits::extractScalarFactor(m_rhs); - - enum { - EvalToDest = (Dest::InnerStrideAtCompileTime==1), - UseRhs = (_ActualRhsType::InnerStrideAtCompileTime==1) - }; - - internal::gemv_static_vector_if static_dest; - internal::gemv_static_vector_if static_rhs; - - ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(), - EvalToDest ? dest.data() : static_dest.data()); - - ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,rhs.size(), - UseRhs ? const_cast(rhs.data()) : static_rhs.data()); - - if(!EvalToDest) - { - #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - Index size = dest.size(); - EIGEN_DENSE_STORAGE_CTOR_PLUGIN - #endif - MappedDest(actualDestPtr, dest.size()) = dest; - } - - if(!UseRhs) - { - #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - Index size = rhs.size(); - EIGEN_DENSE_STORAGE_CTOR_PLUGIN - #endif - Map(actualRhsPtr, rhs.size()) = rhs; - } - - - internal::selfadjoint_matrix_vector_product::Flags&RowMajorBit) ? RowMajor : ColMajor, int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run - ( - lhs.rows(), // size - &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info - actualRhsPtr, 1, // rhs info - actualDestPtr, // result info - actualAlpha // scale factor - ); - - if(!EvalToDest) - dest = MappedDest(actualDestPtr, dest.size()); - } -}; - -namespace internal { -template -struct traits > - : traits, Lhs, Rhs> > -{}; -} - -template -struct SelfadjointProductMatrix - : public ProductBase, Lhs, Rhs > -{ - EIGEN_PRODUCT_PUBLIC_INTERFACE(SelfadjointProductMatrix) - - enum { - RhsUpLo = RhsMode&(Upper|Lower) - }; - - SelfadjointProductMatrix(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - template void scaleAndAddTo(Dest& dest, const Scalar& alpha) const - { - // let's simply transpose the product - Transpose destT(dest); - SelfadjointProductMatrix, int(RhsUpLo)==Upper ? Lower : Upper, false, - Transpose, 0, true>(m_rhs.transpose(), m_lhs.transpose()).scaleAndAddTo(destT, alpha); - } -}; - -#else // EIGEN_TEST_EVALUATORS - namespace internal { template @@ -378,8 +267,6 @@ struct selfadjoint_product_impl } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index fda6e2486..c2d0817ea 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -368,59 +368,9 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix -struct traits > - : traits, Lhs, Rhs> > -{}; -#endif } // end namespace internal -#ifndef EIGEN_TEST_EVALUATORS -template -struct TriangularProduct - : public ProductBase, Lhs, Rhs > -{ - EIGEN_PRODUCT_PUBLIC_INTERFACE(TriangularProduct) - - TriangularProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - template void scaleAndAddTo(Dest& dst, const Scalar& alpha) const - { - typename internal::add_const_on_value_type::type lhs = LhsBlasTraits::extract(m_lhs); - typename internal::add_const_on_value_type::type rhs = RhsBlasTraits::extract(m_rhs); - - Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs) - * RhsBlasTraits::extractScalarFactor(m_rhs); - - typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, - Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType; - - enum { IsLower = (Mode&Lower) == Lower }; - Index stripedRows = ((!LhsIsTriangular) || (IsLower)) ? lhs.rows() : (std::min)(lhs.rows(),lhs.cols()); - Index stripedCols = ((LhsIsTriangular) || (!IsLower)) ? rhs.cols() : (std::min)(rhs.cols(),rhs.rows()); - Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows())) - : ((IsLower) ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols())); - - BlockingType blocking(stripedRows, stripedCols, stripedDepth); - - internal::product_triangular_matrix_matrix::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate, - (internal::traits<_ActualRhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate, - (internal::traits::Flags&RowMajorBit) ? RowMajor : ColMajor> - ::run( - stripedRows, stripedCols, stripedDepth, // sizes - &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info - &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info - &dst.coeffRef(0,0), dst.outerStride(), // result info - actualAlpha, blocking - ); - } -}; -#endif // EIGEN_TEST_EVALUATORS -#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template struct triangular_product_impl @@ -470,7 +420,6 @@ struct triangular_product_impl }; } // end namespace internal -#endif // EIGEN_ENABLE_EVALUATORS } // end namespace Eigen diff --git a/Eigen/src/Core/products/TriangularMatrixVector.h b/Eigen/src/Core/products/TriangularMatrixVector.h index 19167c232..92d64e384 100644 --- a/Eigen/src/Core/products/TriangularMatrixVector.h +++ b/Eigen/src/Core/products/TriangularMatrixVector.h @@ -157,61 +157,11 @@ EIGEN_DONT_INLINE void triangular_matrix_vector_product -struct traits > - : traits, Lhs, Rhs> > -{}; - -template -struct traits > - : traits, Lhs, Rhs> > -{}; -#endif - template struct trmv_selector; } // end namespace internal -#ifndef EIGEN_TEST_EVALUATORS -template -struct TriangularProduct - : public ProductBase, Lhs, Rhs > -{ - EIGEN_PRODUCT_PUBLIC_INTERFACE(TriangularProduct) - - TriangularProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - template void scaleAndAddTo(Dest& dst, const Scalar& alpha) const - { - eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols()); - - internal::trmv_selector::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(m_lhs, m_rhs, dst, alpha); - } -}; - -template -struct TriangularProduct - : public ProductBase, Lhs, Rhs > -{ - EIGEN_PRODUCT_PUBLIC_INTERFACE(TriangularProduct) - - TriangularProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} - - template void scaleAndAddTo(Dest& dst, const Scalar& alpha) const - { - eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols()); - - Transpose dstT(dst); - internal::trmv_selector<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower), - (int(internal::traits::Flags)&RowMajorBit) ? ColMajor : RowMajor> - ::run(m_rhs.transpose(),m_lhs.transpose(), dstT, alpha); - } -}; - -#else // EIGEN_TEST_EVALUATORS namespace internal { template @@ -240,7 +190,6 @@ struct triangular_product_impl }; } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS namespace internal { diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 99aa9b372..9ec57468b 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -139,10 +139,6 @@ template class ArrayWrapper; template class MatrixWrapper; namespace internal { -#ifndef EIGEN_TEST_EVALUATROS -template struct solve_retval_base; -template struct solve_retval; -#endif template struct kernel_retval_base; template struct kernel_retval; template struct image_retval_base; diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index e8ac6bee7..f9b908e22 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -374,46 +374,6 @@ namespace Eigen { * documentation in a single line. **/ -#ifndef EIGEN_TEST_EVALUATORS - -#define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \ - typedef typename Eigen::internal::traits::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex. */ \ - typedef typename Eigen::NumTraits::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex, T were corresponding to RealScalar. */ \ - typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \ - typedef typename Eigen::internal::nested::type Nested; \ - typedef typename Eigen::internal::traits::StorageKind StorageKind; \ - typedef typename Eigen::internal::traits::Index Index; \ - enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ - ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ - Flags = Eigen::internal::traits::Flags, \ - CoeffReadCost = Eigen::internal::traits::CoeffReadCost, \ - SizeAtCompileTime = Base::SizeAtCompileTime, \ - MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \ - IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; - - -#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \ - typedef typename Eigen::internal::traits::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex. */ \ - typedef typename Eigen::NumTraits::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex, T were corresponding to RealScalar. */ \ - typedef typename Base::PacketScalar PacketScalar; \ - typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \ - typedef typename Eigen::internal::nested::type Nested; \ - typedef typename Eigen::internal::traits::StorageKind StorageKind; \ - typedef typename Eigen::internal::traits::Index Index; \ - enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ - ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ - MaxRowsAtCompileTime = Eigen::internal::traits::MaxRowsAtCompileTime, \ - MaxColsAtCompileTime = Eigen::internal::traits::MaxColsAtCompileTime, \ - Flags = Eigen::internal::traits::Flags, \ - CoeffReadCost = Eigen::internal::traits::CoeffReadCost, \ - SizeAtCompileTime = Base::SizeAtCompileTime, \ - MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \ - IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \ - using Base::derived; \ - using Base::const_cast_derived; - -#else - // TODO The EIGEN_DENSE_PUBLIC_INTERFACE should not exists anymore #define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \ @@ -450,8 +410,6 @@ namespace Eigen { using Base::derived; \ using Base::const_cast_derived; -#endif // EIGEN_TEST_EVALUATORS - #define EIGEN_PLAIN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b) #define EIGEN_PLAIN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b) diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index bce009d16..f2536714e 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -125,43 +125,6 @@ template type; }; -#ifndef EIGEN_TEST_EVALUATORS -template -class compute_matrix_flags -{ - enum { - row_major_bit = Options&RowMajor ? RowMajorBit : 0, - is_dynamic_size_storage = MaxRows==Dynamic || MaxCols==Dynamic, - - aligned_bit = - ( - ((Options&DontAlign)==0) - && ( -#if EIGEN_ALIGN_STATICALLY - ((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) -#else - 0 -#endif - - || - -#if EIGEN_ALIGN - is_dynamic_size_storage -#else - 0 -#endif - - ) - ) ? AlignedBit : 0, - packet_access_bit = packet_traits::Vectorizable && aligned_bit ? PacketAccessBit : 0 - }; - - public: - enum { ret = LinearAccessBit | LvalueBit | DirectAccessBit | NestByRefBit | packet_access_bit | row_major_bit | aligned_bit }; -}; - -#else // EIGEN_TEST_EVALUATORS - template class compute_matrix_flags { @@ -172,9 +135,7 @@ class compute_matrix_flags // However, I (Gael) think that DirectAccessBit should only matter at the evaluation stage. enum { ret = DirectAccessBit | LvalueBit | NestByRefBit | row_major_bit }; }; -#endif -#ifdef EIGEN_ENABLE_EVALUATORS template class compute_matrix_evaluator_flags { @@ -209,8 +170,6 @@ class compute_matrix_evaluator_flags enum { ret = LinearAccessBit | DirectAccessBit | packet_access_bit | row_major_bit | aligned_bit }; }; -#endif // EIGEN_ENABLE_EVALUATORS - template struct size_at_compile_time { enum { ret = (_Rows==Dynamic || _Cols==Dynamic) ? Dynamic : _Rows * _Cols }; @@ -361,58 +320,6 @@ struct transfer_constness }; - -#ifndef EIGEN_TEST_EVALUATORS - -/** \internal Determines how a given expression should be nested into another one. - * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be - * nested into the bigger product expression. The choice is between nesting the expression b+c as-is, or - * evaluating that expression b+c into a temporary variable d, and nest d so that the resulting expression is - * a*d. Evaluating can be beneficial for example if every coefficient access in the resulting expression causes - * many coefficient accesses in the nested expressions -- as is the case with matrix product for example. - * - * \param T the type of the expression being nested - * \param n the number of coefficient accesses in the nested expression for each coefficient access in the bigger expression. - * - * Note that if no evaluation occur, then the constness of T is preserved. - * - * Example. Suppose that a, b, and c are of type Matrix3d. The user forms the expression a*(b+c). - * b+c is an expression "sum of matrices", which we will denote by S. In order to determine how to nest it, - * the Product expression uses: nested::type, which turns out to be Matrix3d because the internal logic of - * nested determined that in this case it was better to evaluate the expression b+c into a temporary. On the other hand, - * since a is of type Matrix3d, the Product expression nests it as nested::type, which turns out to be - * const Matrix3d&, because the internal logic of nested determined that since a was already a matrix, there was no point - * in copying it into another matrix. - */ -template::type> struct nested -{ - enum { - // for the purpose of this test, to keep it reasonably simple, we arbitrarily choose a value of Dynamic values. - // the choice of 10000 makes it larger than any practical fixed value and even most dynamic values. - // in extreme cases where these assumptions would be wrong, we would still at worst suffer performance issues - // (poor choice of temporaries). - // it's important that this value can still be squared without integer overflowing. - DynamicAsInteger = 10000, - ScalarReadCost = NumTraits::Scalar>::ReadCost, - ScalarReadCostAsInteger = ScalarReadCost == Dynamic ? int(DynamicAsInteger) : int(ScalarReadCost), - CoeffReadCost = traits::CoeffReadCost, - CoeffReadCostAsInteger = CoeffReadCost == Dynamic ? int(DynamicAsInteger) : int(CoeffReadCost), - NAsInteger = n == Dynamic ? int(DynamicAsInteger) : n, - CostEvalAsInteger = (NAsInteger+1) * ScalarReadCostAsInteger + CoeffReadCostAsInteger, - CostNoEvalAsInteger = NAsInteger * CoeffReadCostAsInteger - }; - - typedef typename conditional< - ( (int(traits::Flags) & EvalBeforeNestingBit) || - int(CostEvalAsInteger) < int(CostNoEvalAsInteger) - ), - PlainObject, - typename ref_selector::type - >::type type; -}; - -#else - // When using evaluators, we never evaluate when assembling the expression!! // TODO: get rid of this nested class since it's just an alias for ref_selector. template struct nested @@ -420,12 +327,20 @@ template struct nested typedef typename ref_selector::type type; }; -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_ENABLE_EVALUATORS // However, we still need a mechanism to detect whether an expression which is evaluated multiple time // has to be evaluated into a temporary. -// That's the purpose of this new nested_eval helper: +// That's the purpose of this new nested_eval helper: +/** \internal Determines how a given expression should be nested when evaluated multiple times. + * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be + * evaluated into the bigger product expression. The choice is between nesting the expression b+c as-is, or + * evaluating that expression b+c into a temporary variable d, and nest d so that the resulting expression is + * a*d. Evaluating can be beneficial for example if every coefficient access in the resulting expression causes + * many coefficient accesses in the nested expressions -- as is the case with matrix product for example. + * + * \param T the type of the expression being nested. + * \param n the number of coefficient accesses in the nested expression for each coefficient access in the bigger expression. + * \param PlainObject the type of the temporary if needed. + */ template::type> struct nested_eval { enum { @@ -453,7 +368,6 @@ template::type> struc typename ref_selector::type >::type type; }; -#endif template EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Geometry/AlignedBox.h b/Eigen/src/Geometry/AlignedBox.h index 1d1daaa61..d6c5c1293 100644 --- a/Eigen/src/Geometry/AlignedBox.h +++ b/Eigen/src/Geometry/AlignedBox.h @@ -71,11 +71,7 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim) template inline explicit AlignedBox(const MatrixBase& a_p) { -#ifndef EIGEN_TEST_EVALUATORS - typename internal::nested::type p(a_p.derived()); -#else typename internal::nested_eval::type p(a_p.derived()); -#endif m_min = p; m_max = p; } diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index 43314b04c..d1881d84d 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -49,9 +49,6 @@ struct traits > Flags = ColsAtCompileTime==1 ? (TmpFlags & ~RowMajorBit) : RowsAtCompileTime==1 ? (TmpFlags | RowMajorBit) : TmpFlags -#ifndef EIGEN_TEST_EVALUATORS - , CoeffReadCost = _MatrixTypeNested::CoeffReadCost -#endif // EIGEN_TEST_EVALUATORS }; }; @@ -80,39 +77,6 @@ template class Homogeneous const NestedExpression& nestedExpression() const { return m_matrix; } -#ifndef EIGEN_TEST_EVALUATORS - inline Scalar coeff(Index row, Index col) const - { - if( (int(Direction)==Vertical && row==m_matrix.rows()) - || (int(Direction)==Horizontal && col==m_matrix.cols())) - return 1; - return m_matrix.coeff(row, col); - } - - template - inline const internal::homogeneous_right_product_impl - operator* (const MatrixBase& rhs) const - { - eigen_assert(int(Direction)==Horizontal); - return internal::homogeneous_right_product_impl(m_matrix,rhs.derived()); - } - - template friend - inline const internal::homogeneous_left_product_impl - operator* (const MatrixBase& lhs, const Homogeneous& rhs) - { - eigen_assert(int(Direction)==Vertical); - return internal::homogeneous_left_product_impl(lhs.derived(),rhs.m_matrix); - } - - template friend - inline const internal::homogeneous_left_product_impl > - operator* (const Transform& lhs, const Homogeneous& rhs) - { - eigen_assert(int(Direction)==Vertical); - return internal::homogeneous_left_product_impl >(lhs,rhs.m_matrix); - } -#else template inline const Product operator* (const MatrixBase& rhs) const @@ -136,7 +100,6 @@ template class Homogeneous eigen_assert(int(Direction)==Vertical); return Product, Homogeneous>(lhs,rhs); } -#endif template EIGEN_STRONG_INLINE typename internal::result_of::type @@ -338,8 +301,6 @@ struct homogeneous_right_product_impl,Rhs> typename Rhs::Nested m_rhs; }; -#ifdef EIGEN_TEST_EVALUATORS - template struct evaluator_traits > { @@ -427,8 +388,6 @@ struct generic_product_impl, Homogeneous::cross(const MatrixBase& other) const // Note that there is no need for an expression here since the compiler // optimize such a small temporary very well (even within a complex expression) -#ifndef EIGEN_TEST_EVALUATORS - typename internal::nested::type lhs(derived()); - typename internal::nested::type rhs(other.derived()); -#else typename internal::nested_eval::type lhs(derived()); typename internal::nested_eval::type rhs(other.derived()); -#endif return typename cross_product_return_type::type( numext::conj(lhs.coeff(1) * rhs.coeff(2) - lhs.coeff(2) * rhs.coeff(1)), numext::conj(lhs.coeff(2) * rhs.coeff(0) - lhs.coeff(0) * rhs.coeff(2)), @@ -81,13 +76,8 @@ MatrixBase::cross3(const MatrixBase& other) const EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Derived,4) EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,4) -#ifndef EIGEN_TEST_EVALUATORS - typedef typename internal::nested::type DerivedNested; - typedef typename internal::nested::type OtherDerivedNested; -#else typedef typename internal::nested_eval::type DerivedNested; typedef typename internal::nested_eval::type OtherDerivedNested; -#endif DerivedNested lhs(derived()); OtherDerivedNested rhs(other.derived()); @@ -114,13 +104,8 @@ VectorwiseOp::cross(const MatrixBase& ot EIGEN_STATIC_ASSERT((internal::is_same::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) -#ifndef EIGEN_TEST_EVALUATORS - typename internal::nested::type mat(_expression()); - typename internal::nested::type vec(other.derived()); -#else typename internal::nested_eval::type mat(_expression()); typename internal::nested_eval::type vec(other.derived()); -#endif CrossReturnType res(_expression().rows(),_expression().cols()); if(Direction==Vertical) diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index 850940302..3f0067286 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -217,11 +217,7 @@ struct traits > typedef _Scalar Scalar; typedef Matrix<_Scalar,4,1,_Options> Coefficients; enum{ -#ifndef EIGEN_TEST_EVALUATORS - IsAligned = internal::traits::Flags & AlignedBit, -#else IsAligned = (internal::traits::EvaluatorFlags & AlignedBit) != 0, -#endif Flags = IsAligned ? (AlignedBit | LvalueBit) : LvalueBit }; }; diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index bcf3e2723..89e9cc1a4 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -62,7 +62,6 @@ struct transform_construct_from_matrix; template struct transform_take_affine_part; -#ifdef EIGEN_TEST_EVALUATORS template struct traits > { @@ -78,7 +77,6 @@ struct traits > Flags = 0 }; }; -#endif } // end namespace internal @@ -374,10 +372,8 @@ public: inline QTransform toQTransform(void) const; #endif -#ifdef EIGEN_TEST_EVALUATORS Index rows() const { return int(Mode)==int(Projective) ? m_matrix.cols() : (m_matrix.cols()-1); } Index cols() const { return m_matrix.cols(); } -#endif /** shortcut for m_matrix(row,col); * \sa MatrixBase::operator(Index,Index) const */ diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h index 0c6a09861..4ded2995f 100644 --- a/Eigen/src/Householder/HouseholderSequence.h +++ b/Eigen/src/Householder/HouseholderSequence.h @@ -73,8 +73,6 @@ struct traits > }; }; -#ifdef EIGEN_TEST_EVALUATORS - struct HouseholderSequenceShape {}; template @@ -83,7 +81,6 @@ struct evaluator_traits > { typedef HouseholderSequenceShape Shape; }; -#endif template struct hseq_side_dependent_impl diff --git a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h index 92af28cc8..98b169868 100644 --- a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h +++ b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h @@ -87,16 +87,6 @@ class DiagonalPreconditioner x = m_invdiag.array() * b.array() ; } -#ifndef EIGEN_TEST_EVALUATORS - template inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "DiagonalPreconditioner is not initialized."); - eigen_assert(m_invdiag.size()==b.rows() - && "DiagonalPreconditioner::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } -#else template inline const Solve solve(const MatrixBase& b) const { @@ -105,31 +95,12 @@ class DiagonalPreconditioner && "DiagonalPreconditioner::solve(): invalid number of rows of the right hand side matrix b"); return Solve(*this, b.derived()); } -#endif protected: Vector m_invdiag; bool m_isInitialized; }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef DiagonalPreconditioner<_MatrixType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} -#endif // EIGEN_TEST_EVALUATORS /** \ingroup IterativeLinearSolvers_Module * \brief A naive preconditioner which approximates any matrix as the identity matrix diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index b4d1d2a79..051940dc7 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -182,24 +182,6 @@ public: ~BiCGSTAB() {} -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A - * \a x0 as an initial solution. - * - * \sa compute() - */ - template - inline const internal::solve_retval_with_guess - solveWithGuess(const MatrixBase& b, const Guess& x0) const - { - eigen_assert(m_isInitialized && "BiCGSTAB is not initialized."); - eigen_assert(Base::rows()==b.rows() - && "BiCGSTAB::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval_with_guess - (*this, b.derived(), x0); - } -#endif - /** \internal */ template void _solve_with_guess_impl(const Rhs& b, Dest& x) const @@ -234,25 +216,6 @@ protected: }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef BiCGSTAB<_MatrixType, _Preconditioner> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} // end namespace internal -#endif - } // end namespace Eigen #endif // EIGEN_BICGSTAB_H diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index b0273aaaf..f72cf86a5 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -192,24 +192,6 @@ public: ConjugateGradient(const MatrixType& A) : Base(A) {} ~ConjugateGradient() {} - -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A - * \a x0 as an initial solution. - * - * \sa compute() - */ - template - inline const internal::solve_retval_with_guess - solveWithGuess(const MatrixBase& b, const Guess& x0) const - { - eigen_assert(m_isInitialized && "ConjugateGradient is not initialized."); - eigen_assert(Base::rows()==b.rows() - && "ConjugateGradient::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval_with_guess - (*this, b.derived(), x0); - } -#endif /** \internal */ template @@ -245,25 +227,6 @@ protected: }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} // end namespace internal -#endif - } // end namespace Eigen #endif // EIGEN_CONJUGATE_GRADIENT_H diff --git a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h index f337c5fb0..7adbbc489 100644 --- a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h +++ b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h @@ -171,17 +171,6 @@ class IncompleteLUT : public SparseSolverBase > x = m_P * x; } -#ifndef EIGEN_TEST_EVALUATORS - template inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "IncompleteLUT is not initialized."); - eigen_assert(cols()==b.rows() - && "IncompleteLUT::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } -#endif - protected: /** keeps off-diagonal entries; drops diagonal entries */ @@ -451,25 +440,6 @@ void IncompleteLUT::factorize(const _MatrixType& amat) m_info = Success; } -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef IncompleteLUT<_MatrixType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_INCOMPLETE_LUT_H diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h index 26487dbb2..fd9285087 100644 --- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h @@ -162,38 +162,6 @@ public: return m_error; } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "IterativeSolverBase is not initialized."); - eigen_assert(rows()==b.rows() - && "IterativeSolverBase::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(derived(), b.derived()); - } -#endif - -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval - solve(const SparseMatrixBase& b) const - { - eigen_assert(m_isInitialized && "IterativeSolverBase is not initialized."); - eigen_assert(rows()==b.rows() - && "IterativeSolverBase::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); - } -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_TEST_EVALUATORS /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A * and \a x0 as an initial solution. * @@ -207,7 +175,6 @@ public: eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b"); return SolveWithGuess(derived(), b.derived(), x0); } -#endif // EIGEN_TEST_EVALUATORS /** \returns Success if the iterations converged, and NoConvergence otherwise. */ ComputationInfo info() const @@ -216,25 +183,6 @@ public: return m_info; } -#ifndef EIGEN_TEST_EVALUATORS - /** \internal */ - template - void _solve_sparse(const Rhs& b, SparseMatrix &dest) const - { - eigen_assert(rows()==b.rows()); - - int rhsCols = b.cols(); - int size = b.rows(); - Eigen::Matrix tb(size); - Eigen::Matrix tx(size); - for(int k=0; k void _solve_impl(const Rhs& b, SparseMatrix &dest) const @@ -252,7 +200,6 @@ public: dest.col(k) = tx.sparseView(0); } } -#endif // EIGEN_TEST_EVALUATORS protected: void init() @@ -275,25 +222,6 @@ protected: mutable bool m_analysisIsOk, m_factorizationIsOk; }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef IterativeSolverBase Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec().derived()._solve_sparse(rhs(),dst); - } -}; - -} // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_ITERATIVE_SOLVER_BASE_H diff --git a/Eigen/src/LU/Determinant.h b/Eigen/src/LU/Determinant.h index 9726bd96a..d6a3c1e5a 100644 --- a/Eigen/src/LU/Determinant.h +++ b/Eigen/src/LU/Determinant.h @@ -92,11 +92,7 @@ template inline typename internal::traits::Scalar MatrixBase::determinant() const { eigen_assert(rows() == cols()); -#ifdef EIGEN_TEST_EVALUATORS typedef typename internal::nested_eval::type Nested; -#else - typedef typename internal::nested::type Nested; -#endif return internal::determinant_impl::type>::run(derived()); } diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index daf99e305..fdf2e0642 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -220,7 +220,6 @@ template class FullPivLU * * \sa TriangularView::solve(), kernel(), inverse() */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -228,15 +227,6 @@ template class FullPivLU eigen_assert(m_isInitialized && "LU is not initialized."); return Solve(*this, b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "LU is not initialized."); - return internal::solve_retval(*this, b.derived()); - } -#endif /** \returns the determinant of the matrix of which * *this is the LU decomposition. It has only linear complexity @@ -380,22 +370,12 @@ template class FullPivLU * * \sa MatrixBase::inverse() */ -#ifdef EIGEN_TEST_EVALUATORS inline const Inverse inverse() const { eigen_assert(m_isInitialized && "LU is not initialized."); eigen_assert(m_lu.rows() == m_lu.cols() && "You can't take the inverse of a non-square matrix!"); return Inverse(*this); } -#else - inline const internal::solve_retval inverse() const - { - eigen_assert(m_isInitialized && "LU is not initialized."); - eigen_assert(m_lu.rows() == m_lu.cols() && "You can't take the inverse of a non-square matrix!"); - return internal::solve_retval - (*this, MatrixType::Identity(m_lu.rows(), m_lu.cols())); - } -#endif MatrixType reconstructedMatrix() const; @@ -752,22 +732,8 @@ void FullPivLU<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - EIGEN_MAKE_SOLVE_HELPERS(FullPivLU<_MatrixType>,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(), dst); - } -}; -#endif /***** Implementation of inverse() *****************************************************/ -#ifdef EIGEN_TEST_EVALUATORS template struct Assignment >, internal::assign_op, Dense2Dense, Scalar> { @@ -778,7 +744,6 @@ struct Assignment >, internal::assign_ dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; -#endif } // end namespace internal /******* MatrixBase methods *****************************************************************/ diff --git a/Eigen/src/LU/InverseImpl.h b/Eigen/src/LU/InverseImpl.h index e10fee48f..e5f270d19 100644 --- a/Eigen/src/LU/InverseImpl.h +++ b/Eigen/src/LU/InverseImpl.h @@ -43,12 +43,8 @@ struct compute_inverse static inline void run(const MatrixType& matrix, ResultType& result) { typedef typename MatrixType::Scalar Scalar; -#ifdef EIGEN_TEST_EVALUATORS typename internal::evaluator::type matrixEval(matrix); result.coeffRef(0,0) = Scalar(1) / matrixEval.coeff(0,0); -#else - result.coeffRef(0,0) = Scalar(1) / matrix.coeff(0,0); -#endif } }; @@ -285,46 +281,8 @@ struct compute_inverse_and_det_with_check *** MatrixBase methods *** *************************/ -#ifndef EIGEN_TEST_EVALUATORS -template -struct traits > -{ - typedef typename MatrixType::PlainObject ReturnType; -}; - -template -struct inverse_impl : public ReturnByValue > -{ - typedef typename MatrixType::Index Index; - typedef typename internal::eval::type MatrixTypeNested; - typedef typename remove_all::type MatrixTypeNestedCleaned; - MatrixTypeNested m_matrix; - - EIGEN_DEVICE_FUNC - inverse_impl(const MatrixType& matrix) - : m_matrix(matrix) - {} - - EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } - EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } - - template - EIGEN_DEVICE_FUNC - inline void evalTo(Dest& dst) const - { - const int Size = EIGEN_PLAIN_ENUM_MIN(MatrixType::ColsAtCompileTime,Dest::ColsAtCompileTime); - EIGEN_ONLY_USED_FOR_DEBUG(Size); - eigen_assert(( (Size<=1) || (Size>4) || (extract_data(m_matrix)!=extract_data(dst))) - && "Aliasing problem detected in inverse(), you need to do inverse().eval() here."); - - compute_inverse::run(m_matrix, dst); - } -}; -#endif } // end namespace internal -#ifdef EIGEN_TEST_EVALUATORS - namespace internal { // Specialization for "dense = dense_xpr.inverse()" @@ -352,8 +310,6 @@ struct Assignment, internal::assign_op, Den } // end namespace internal -#endif - /** \lu_module * * \returns the matrix inverse of this matrix. @@ -371,7 +327,6 @@ struct Assignment, internal::assign_op, Den * * \sa computeInverseAndDetWithCheck() */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Inverse MatrixBase::inverse() const { @@ -379,15 +334,6 @@ inline const Inverse MatrixBase::inverse() const eigen_assert(rows() == cols()); return Inverse(derived()); } -#else -template -inline const internal::inverse_impl MatrixBase::inverse() const -{ - EIGEN_STATIC_ASSERT(!NumTraits::IsInteger,THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES) - eigen_assert(rows() == cols()); - return internal::inverse_impl(derived()); -} -#endif /** \lu_module * @@ -422,11 +368,7 @@ inline void MatrixBase::computeInverseAndDetWithCheck( // for larger sizes, evaluating has negligible cost and limits code size. typedef typename internal::conditional< RowsAtCompileTime == 2, -#ifndef EIGEN_TEST_EVALUATORS - typename internal::remove_all::type>::type, -#else typename internal::remove_all::type>::type, -#endif PlainObject >::type MatrixType; internal::compute_inverse_and_det_with_check::run diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index 57076f3a4..a4d22ce5f 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -142,7 +142,6 @@ template class PartialPivLU * * \sa TriangularView::solve(), inverse(), computeInverse() */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -150,15 +149,6 @@ template class PartialPivLU eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); return Solve(*this, b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); - return internal::solve_retval(*this, b.derived()); - } -#endif /** \returns the inverse of the matrix of which *this is the LU decomposition. * @@ -167,20 +157,11 @@ template class PartialPivLU * * \sa MatrixBase::inverse(), LU::inverse() */ -#ifdef EIGEN_TEST_EVALUATORS inline const Inverse inverse() const { eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); return Inverse(*this); } -#else - inline const internal::solve_retval inverse() const - { - eigen_assert(m_isInitialized && "PartialPivLU is not initialized."); - return internal::solve_retval - (*this, MatrixType::Identity(m_lu.rows(), m_lu.cols())); - } -#endif /** \returns the determinant of the matrix of which * *this is the LU decomposition. It has only linear complexity @@ -490,22 +471,7 @@ MatrixType PartialPivLU::reconstructedMatrix() const namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - EIGEN_MAKE_SOLVE_HELPERS(PartialPivLU<_MatrixType>,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(), dst); - } -}; -#endif - /***** Implementation of inverse() *****************************************************/ -#ifdef EIGEN_TEST_EVALUATORS template struct Assignment >, internal::assign_op, Dense2Dense, Scalar> { @@ -516,7 +482,6 @@ struct Assignment >, internal::assi dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; -#endif } // end namespace internal /******** MatrixBase methods *******/ diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h index 95d53c850..bb8e0d1a8 100644 --- a/Eigen/src/PaStiXSupport/PaStiXSupport.h +++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h @@ -153,22 +153,6 @@ class PastixBase : public SparseSolverBase { clean(); } - -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "Pastix solver is not initialized."); - eigen_assert(rows()==b.rows() - && "PastixBase::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } -#endif template bool _solve_impl(const MatrixBase &b, MatrixBase &x) const; @@ -227,22 +211,6 @@ class PastixBase : public SparseSolverBase return m_info; } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval - solve(const SparseMatrixBase& b) const - { - eigen_assert(m_isInitialized && "Pastix LU, LLT or LDLT is not initialized."); - eigen_assert(rows()==b.rows() - && "PastixBase::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); - } -#endif // EIGEN_TEST_EVALUATORS - protected: // Initialize the Pastix data structure, check the matrix @@ -693,38 +661,6 @@ class PastixLDLT : public PastixBase< PastixLDLT<_MatrixType, _UpLo> > } }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef PastixBase<_MatrixType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef PastixBase<_MatrixType> Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - this->defaultEvalTo(dst); - } -}; - -} // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif diff --git a/Eigen/src/PardisoSupport/PardisoSupport.h b/Eigen/src/PardisoSupport/PardisoSupport.h index e3b1c5bc0..e1b0e1818 100644 --- a/Eigen/src/PardisoSupport/PardisoSupport.h +++ b/Eigen/src/PardisoSupport/PardisoSupport.h @@ -172,36 +172,6 @@ class PardisoImpl : public SparseSolveBase Derived& factorize(const MatrixType& matrix); Derived& compute(const MatrixType& matrix); - -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "Pardiso solver is not initialized."); - eigen_assert(rows()==b.rows() - && "PardisoImpl::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } - - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval - solve(const SparseMatrixBase& b) const - { - eigen_assert(m_isInitialized && "Pardiso solver is not initialized."); - eigen_assert(rows()==b.rows() - && "PardisoImpl::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); - } -#endif template bool _solve_impl(const MatrixBase &b, MatrixBase& x) const; @@ -546,38 +516,6 @@ class PardisoLDLT : public PardisoImpl< PardisoLDLT > } }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef PardisoImpl<_Derived> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef PardisoImpl Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - this->defaultEvalTo(dst); - } -}; - -} // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_PARDISOSUPPORT_H diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index 96904c65f..adf737276 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -147,7 +147,6 @@ template class ColPivHouseholderQR * Example: \include ColPivHouseholderQR_solve.cpp * Output: \verbinclude ColPivHouseholderQR_solve.out */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -155,15 +154,6 @@ template class ColPivHouseholderQR eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); return Solve(*this, b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); - return internal::solve_retval(*this, b.derived()); - } -#endif HouseholderSequenceType householderQ() const; HouseholderSequenceType matrixQ() const @@ -304,22 +294,11 @@ template class ColPivHouseholderQR * \note If this matrix is not invertible, the returned matrix has undefined coefficients. * Use isInvertible() to first determine whether this matrix is invertible. */ -#ifdef EIGEN_TEST_EVALUATORS inline const Inverse inverse() const { eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); return Inverse(*this); } -#else - inline const - internal::solve_retval - inverse() const - { - eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); - return internal::solve_retval - (*this, MatrixType::Identity(m_qr.rows(), m_qr.cols())); - } -#endif inline Index rows() const { return m_qr.rows(); } inline Index cols() const { return m_qr.cols(); } @@ -582,21 +561,6 @@ void ColPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType & namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - EIGEN_MAKE_SOLVE_HELPERS(ColPivHouseholderQR<_MatrixType>,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(), dst); - } -}; -#endif - -#ifdef EIGEN_TEST_EVALUATORS template struct Assignment >, internal::assign_op, Dense2Dense, Scalar> { @@ -607,7 +571,6 @@ struct Assignment >, interna dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; -#endif } // end namespace internal diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index ea15da041..710c64a45 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -151,7 +151,6 @@ template class FullPivHouseholderQR * Example: \include FullPivHouseholderQR_solve.cpp * Output: \verbinclude FullPivHouseholderQR_solve.out */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -159,15 +158,6 @@ template class FullPivHouseholderQR eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); return Solve(*this, b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); - return internal::solve_retval(*this, b.derived()); - } -#endif /** \returns Expression object representing the matrix Q */ @@ -298,22 +288,11 @@ template class FullPivHouseholderQR * \note If this matrix is not invertible, the returned matrix has undefined coefficients. * Use isInvertible() to first determine whether this matrix is invertible. */ -#ifdef EIGEN_TEST_EVALUATORS inline const Inverse inverse() const { eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); return Inverse(*this); } -#else - inline const - internal::solve_retval - inverse() const - { - eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized."); - return internal::solve_retval - (*this, MatrixType::Identity(m_qr.rows(), m_qr.cols())); - } -#endif inline Index rows() const { return m_qr.rows(); } inline Index cols() const { return m_qr.cols(); } @@ -556,21 +535,6 @@ void FullPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - EIGEN_MAKE_SOLVE_HELPERS(FullPivHouseholderQR<_MatrixType>,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(), dst); - } -}; -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_TEST_EVALUATORS template struct Assignment >, internal::assign_op, Dense2Dense, Scalar> { @@ -581,7 +545,6 @@ struct Assignment >, intern dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols())); } }; -#endif /** \ingroup QR_Module * @@ -589,7 +552,6 @@ struct Assignment >, intern * * \tparam MatrixType type of underlying dense matrix */ -// #ifndef EIGEN_TEST_EVALUATORS template struct FullPivHouseholderQRMatrixQReturnType : public ReturnByValue > { @@ -645,12 +607,10 @@ protected: typename IntDiagSizeVectorType::Nested m_rowsTranspositions; }; -// #ifdef EIGEN_TEST_EVALUATORS // template // struct evaluator > // : public evaluator > > // {}; -// #endif } // end namespace internal diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index 8604fff6f..0b0c9d1bd 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -117,7 +117,6 @@ template class HouseholderQR * Example: \include HouseholderQR_solve.cpp * Output: \verbinclude HouseholderQR_solve.out */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -125,15 +124,6 @@ template class HouseholderQR eigen_assert(m_isInitialized && "HouseholderQR is not initialized."); return Solve(*this, b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "HouseholderQR is not initialized."); - return internal::solve_retval(*this, b.derived()); - } -#endif /** This method returns an expression of the unitary matrix Q as a sequence of Householder transformations. * @@ -351,24 +341,6 @@ void HouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) c } #endif -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - EIGEN_MAKE_SOLVE_HELPERS(HouseholderQR<_MatrixType>,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(), dst); - } -}; - -} // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - /** Performs the QR factorization of the given matrix \a matrix. The result of * the factorization is stored into \c *this, and a reference to \c *this * is returned. diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h index 483aaef07..bcdc981d7 100644 --- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h @@ -124,21 +124,6 @@ class SPQR : public SparseSolverBase > * Get the number of columns of the input matrix. */ inline Index cols() const { return m_cR->ncol; } - -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval solve(const MatrixBase& B) const - { - eigen_assert(m_isInitialized && " The QR factorization should be computed first, call compute()"); - eigen_assert(this->rows()==B.rows() - && "SPQR::solve(): invalid number of rows of the right hand side matrix B"); - return internal::solve_retval(*this, B.derived()); - } -#endif // EIGEN_TEST_EVALUATORS template void _solve_impl(const MatrixBase &b, MatrixBase &dest) const @@ -292,24 +277,5 @@ struct SPQRMatrixQTransposeReturnType{ const SPQRType& m_spqr; }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef SPQR<_MatrixType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} // end namespace internal -#endif - }// End namespace Eigen #endif diff --git a/Eigen/src/SVD/SVDBase.h b/Eigen/src/SVD/SVDBase.h index 60a5aabb6..27b732b80 100644 --- a/Eigen/src/SVD/SVDBase.h +++ b/Eigen/src/SVD/SVDBase.h @@ -200,7 +200,6 @@ public: * \note SVD solving is implicitly least-squares. Thus, this method serves both purposes of exact solving and least-squares solving. * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$. */ -#ifdef EIGEN_TEST_EVALUATORS template inline const Solve solve(const MatrixBase& b) const @@ -209,16 +208,6 @@ public: eigen_assert(computeU() && computeV() && "SVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); return Solve(derived(), b.derived()); } -#else - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "SVD is not initialized."); - eigen_assert(computeU() && computeV() && "SVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); - return internal::solve_retval(*this, b.derived()); - } -#endif #ifndef EIGEN_PARSED_BY_DOXYGEN template @@ -273,23 +262,6 @@ void SVDBase::_solve_impl(const RhsType &rhs, DstType &dst) const } #endif -namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef SVDBase SVDType; - EIGEN_MAKE_SOLVE_HELPERS(SVDType,Rhs) - - template void evalTo(Dest& dst) const - { - dec().derived()._solve_impl(rhs(), dst); - } -}; -#endif -} // end namespace internal - template bool SVDBase::allocate(Index rows, Index cols, unsigned int computationOptions) { diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index ac8cd29b0..3c8cef5db 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -84,36 +84,6 @@ class SimplicialCholeskyBase : public SparseSolverBase return m_info; } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "Simplicial LLT or LDLT is not initialized."); - eigen_assert(rows()==b.rows() - && "SimplicialCholeskyBase::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } - - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval - solve(const SparseMatrixBase& b) const - { - eigen_assert(m_isInitialized && "Simplicial LLT or LDLT is not initialized."); - eigen_assert(rows()==b.rows() - && "SimplicialCholesky::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); - } -#endif // EIGEN_TEST_EVALUATORS - /** \returns the permutation P * \sa permutationPinv() */ const PermutationMatrix& permutationP() const @@ -157,11 +127,7 @@ class SimplicialCholeskyBase : public SparseSolverBase /** \internal */ template -#ifndef EIGEN_TEST_EVALUATORS - void _solve(const MatrixBase &b, MatrixBase &dest) const -#else void _solve_impl(const MatrixBase &b, MatrixBase &dest) const -#endif { eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()"); eigen_assert(m_matrix.rows()==b.rows()); @@ -186,14 +152,12 @@ class SimplicialCholeskyBase : public SparseSolverBase if(m_P.size()>0) dest = m_Pinv * dest; } - -#ifdef EIGEN_TEST_EVALUATORS + template void _solve_impl(const SparseMatrixBase &b, SparseMatrixBase &dest) const { internal::solve_sparse_through_dense_panels(derived(), b, dest); } -#endif #endif // EIGEN_PARSED_BY_DOXYGEN @@ -578,11 +542,7 @@ public: /** \internal */ template -#ifndef EIGEN_TEST_EVALUATORS - void _solve(const MatrixBase &b, MatrixBase &dest) const -#else void _solve_impl(const MatrixBase &b, MatrixBase &dest) const -#endif { eigen_assert(Base::m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()"); eigen_assert(Base::m_matrix.rows()==b.rows()); @@ -618,14 +578,12 @@ public: dest = Base::m_Pinv * dest; } -#ifdef EIGEN_TEST_EVALUATORS /** \internal */ template void _solve_impl(const SparseMatrixBase &b, SparseMatrixBase &dest) const { internal::solve_sparse_through_dense_panels(*this, b, dest); } -#endif Scalar determinant() const { @@ -667,38 +625,6 @@ void SimplicialCholeskyBase::ordering(const MatrixType& a, CholMatrixTy ap.template selfadjointView() = a.template selfadjointView().twistedBy(m_P); } -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef SimplicialCholeskyBase Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec().derived()._solve(rhs(),dst); - } -}; - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef SimplicialCholeskyBase Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - this->defaultEvalTo(dst); - } -}; - -} // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SIMPLICIAL_CHOLESKY_H diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 398008597..19d9eaa42 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -39,10 +39,8 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r // Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs) Index estimated_nnz_prod = lhs.nonZeros() + rhs.nonZeros(); -#ifdef EIGEN_TEST_EVALUATORS typename evaluator::type lhsEval(lhs); typename evaluator::type rhsEval(rhs); -#endif res.setZero(); res.reserve(Index(estimated_nnz_prod)); @@ -52,19 +50,11 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r res.startVec(j); Index nnz = 0; -#ifndef EIGEN_TEST_EVALUATORS - for (typename Rhs::InnerIterator rhsIt(rhs, j); rhsIt; ++rhsIt) -#else for (typename evaluator::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt) -#endif { Scalar y = rhsIt.value(); Index k = rhsIt.index(); -#ifndef EIGEN_TEST_EVALUATORS - for (typename Lhs::InnerIterator lhsIt(lhs, k); lhsIt; ++lhsIt) -#else for (typename evaluator::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt) -#endif { Index i = lhsIt.index(); Scalar x = lhsIt.value(); diff --git a/Eigen/src/SparseCore/MappedSparseMatrix.h b/Eigen/src/SparseCore/MappedSparseMatrix.h index 9205b906f..d9aabd049 100644 --- a/Eigen/src/SparseCore/MappedSparseMatrix.h +++ b/Eigen/src/SparseCore/MappedSparseMatrix.h @@ -176,7 +176,6 @@ class MappedSparseMatrix::ReverseInnerIterator const Index m_end; }; -#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template @@ -202,7 +201,6 @@ struct evaluator > }; } -#endif } // end namespace Eigen diff --git a/Eigen/src/SparseCore/SparseAssign.h b/Eigen/src/SparseCore/SparseAssign.h index 36c9fe845..97c079d3f 100644 --- a/Eigen/src/SparseCore/SparseAssign.h +++ b/Eigen/src/SparseCore/SparseAssign.h @@ -12,117 +12,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS - -template -template -Derived& SparseMatrixBase::operator=(const EigenBase &other) -{ - other.derived().evalTo(derived()); - return derived(); -} - -template -template -Derived& SparseMatrixBase::operator=(const ReturnByValue& other) -{ - other.evalTo(derived()); - return derived(); -} - -template -template -inline Derived& SparseMatrixBase::operator=(const SparseMatrixBase& other) -{ - return assign(other.derived()); -} - -template -inline Derived& SparseMatrixBase::operator=(const Derived& other) -{ -// if (other.isRValue()) -// derived().swap(other.const_cast_derived()); -// else - return assign(other.derived()); -} - -template -template -inline Derived& SparseMatrixBase::assign(const OtherDerived& other) -{ - const bool transpose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); - const Index outerSize = (int(OtherDerived::Flags) & RowMajorBit) ? Index(other.rows()) : Index(other.cols()); - if ((!transpose) && other.isRValue()) - { - // eval without temporary - derived().resize(Index(other.rows()), Index(other.cols())); - derived().setZero(); - derived().reserve((std::max)(this->rows(),this->cols())*2); - for (Index j=0; j -template -inline void SparseMatrixBase::assignGeneric(const OtherDerived& other) -{ - //const bool transpose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); - eigen_assert(( ((internal::traits::SupportedAccessPatterns&OuterRandomAccessPattern)==OuterRandomAccessPattern) || - (!((Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit)))) && - "the transpose operation is supposed to be handled in SparseMatrix::operator="); - - enum { Flip = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit) }; - - const Index outerSize = Index(other.outerSize()); - //typedef typename internal::conditional, Derived>::type TempType; - // thanks to shallow copies, we always eval to a tempary - Derived temp(Index(other.rows()), Index(other.cols())); - - temp.reserve((std::max)(this->rows(),this->cols())*2); - for (Index j=0; j -// inline Derived& operator=(const SparseSparseProduct& product); -// -// template -// Derived& operator+=(const SparseMatrixBase& other); -// template -// Derived& operator-=(const SparseMatrixBase& other); -// -// Derived& operator*=(const Scalar& other); -// Derived& operator/=(const Scalar& other); -// -// template -// Derived& operator*=(const SparseMatrixBase& other); - -#else // EIGEN_TEST_EVALUATORS - template template Derived& SparseMatrixBase::operator=(const EigenBase &other) @@ -298,8 +187,6 @@ struct Assignment, internal::assign_opindex(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; - class ReverseInnerIterator: public XprType::ReverseInnerIterator - { - typedef typename BlockImpl::Index Index; - public: - inline ReverseInnerIterator(const BlockType& xpr, Index outer) - : XprType::ReverseInnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) - {} - inline Index row() const { return IsRowMajor ? m_outer : this->index(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; -#endif // EIGEN_TEST_EVALUATORS inline BlockImpl(const XprType& xpr, Index i) : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) @@ -66,14 +39,6 @@ public: Index nonZeros() const { -#ifndef EIGEN_TEST_EVALUATORS - Index nnz = 0; - Index end = m_outerStart + m_outerSize.value(); - for(Index j=m_outerStart; j::type EvaluatorType; EvaluatorType matEval(m_matrix); Index nnz = 0; @@ -82,7 +47,6 @@ public: for(typename EvaluatorType::InnerIterator it(matEval, j); it; ++it) ++nnz; return nnz; -#endif // EIGEN_TEST_EVALUATORS } inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } @@ -120,31 +84,6 @@ public: protected: enum { OuterSize = IsRowMajor ? BlockRows : BlockCols }; public: - -#ifndef EIGEN_TEST_EVALUATORS - class InnerIterator: public SparseMatrixType::InnerIterator - { - public: - inline InnerIterator(const BlockType& xpr, Index outer) - : SparseMatrixType::InnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) - {} - inline Index row() const { return IsRowMajor ? m_outer : this->index(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; - class ReverseInnerIterator: public SparseMatrixType::ReverseInnerIterator - { - public: - inline ReverseInnerIterator(const BlockType& xpr, Index outer) - : SparseMatrixType::ReverseInnerIterator(xpr.m_matrix, xpr.m_outerStart + outer), m_outer(outer) - {} - inline Index row() const { return IsRowMajor ? m_outer : this->index(); } - inline Index col() const { return IsRowMajor ? this->index() : m_outer; } - protected: - Index m_outer; - }; -#endif // EIGEN_TEST_EVALUATORS inline sparse_matrix_block_impl(const SparseMatrixType& xpr, Index i) : m_matrix(xpr), m_outerStart(i), m_outerSize(OuterSize) @@ -434,34 +373,6 @@ public: m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0)); } -#ifndef EIGEN_TEST_EVALUATORS - typedef internal::GenericSparseBlockInnerIteratorImpl InnerIterator; - - class ReverseInnerIterator : public _MatrixTypeNested::ReverseInnerIterator - { - typedef typename _MatrixTypeNested::ReverseInnerIterator Base; - const BlockType& m_block; - Index m_begin; - public: - - EIGEN_STRONG_INLINE ReverseInnerIterator(const BlockType& block, Index outer) - : Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())), - m_block(block), - m_begin(IsRowMajor ? block.m_startCol.value() : block.m_startRow.value()) - { - while( (Base::operator bool()) && (Base::index() >= (IsRowMajor ? m_block.m_startCol.value()+block.m_blockCols.value() : m_block.m_startRow.value()+block.m_blockRows.value())) ) - Base::operator--(); - } - - inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } - inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } - inline Index row() const { return Base::row() - m_block.m_startRow.value(); } - inline Index col() const { return Base::col() - m_block.m_startCol.value(); } - - inline operator bool() const { return Base::operator bool() && Base::index() >= m_begin; } - }; -#endif // EIGEN_TEST_EVALUATORS - inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } Index startRow() const { return m_startRow.value(); } Index startCol() const { return m_startCol.value(); } @@ -574,9 +485,6 @@ namespace internal { inline operator bool() const { return m_outerPos < m_end; } }; -#ifdef EIGEN_TEST_EVALUATORS - -// template struct unary_evaluator, IteratorBased > : public evaluator_base > @@ -690,9 +598,6 @@ public: inline operator bool() const { return m_outerPos < m_end; } }; - -#endif // EIGEN_TEST_EVALUATORS - } // end namespace internal diff --git a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h index 3a7e72cd2..5993c1caf 100644 --- a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h @@ -38,256 +38,6 @@ class sparse_cwise_binary_op_inner_iterator_selector; } // end namespace internal -#ifndef EIGEN_TEST_EVALUATORS - -template -class CwiseBinaryOpImpl - : public SparseMatrixBase > -{ - public: - class InnerIterator; - class ReverseInnerIterator; - typedef CwiseBinaryOp Derived; - EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) - CwiseBinaryOpImpl() - { - typedef typename internal::traits::StorageKind LhsStorageKind; - typedef typename internal::traits::StorageKind RhsStorageKind; - EIGEN_STATIC_ASSERT(( - (!internal::is_same::value) - || ((Lhs::Flags&RowMajorBit) == (Rhs::Flags&RowMajorBit))), - THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH); - } -}; - -template -class CwiseBinaryOpImpl::InnerIterator - : public internal::sparse_cwise_binary_op_inner_iterator_selector::InnerIterator> -{ - public: - typedef internal::sparse_cwise_binary_op_inner_iterator_selector< - BinaryOp,Lhs,Rhs, InnerIterator> Base; - - EIGEN_STRONG_INLINE InnerIterator(const CwiseBinaryOpImpl& binOp, Index outer) - : Base(binOp.derived(),outer) - {} -}; - -/*************************************************************************** -* Implementation of inner-iterators -***************************************************************************/ - -// template struct internal::func_is_conjunction { enum { ret = false }; }; -// template struct internal::func_is_conjunction > { enum { ret = true }; }; - -// TODO generalize the internal::scalar_product_op specialization to all conjunctions if any ! - -namespace internal { - -// sparse - sparse (generic) -template -class sparse_cwise_binary_op_inner_iterator_selector -{ - typedef CwiseBinaryOp CwiseBinaryXpr; - typedef typename traits::Scalar Scalar; - typedef typename traits::Index Index; - typedef typename traits::_LhsNested _LhsNested; - typedef typename traits::_RhsNested _RhsNested; - typedef typename _LhsNested::InnerIterator LhsIterator; - typedef typename _RhsNested::InnerIterator RhsIterator; - - public: - - EIGEN_STRONG_INLINE sparse_cwise_binary_op_inner_iterator_selector(const CwiseBinaryXpr& xpr, Index outer) - : m_lhsIter(xpr.lhs(),outer), m_rhsIter(xpr.rhs(),outer), m_functor(xpr.functor()) - { - this->operator++(); - } - - EIGEN_STRONG_INLINE Derived& operator++() - { - if (m_lhsIter && m_rhsIter && (m_lhsIter.index() == m_rhsIter.index())) - { - m_id = m_lhsIter.index(); - m_value = m_functor(m_lhsIter.value(), m_rhsIter.value()); - ++m_lhsIter; - ++m_rhsIter; - } - else if (m_lhsIter && (!m_rhsIter || (m_lhsIter.index() < m_rhsIter.index()))) - { - m_id = m_lhsIter.index(); - m_value = m_functor(m_lhsIter.value(), Scalar(0)); - ++m_lhsIter; - } - else if (m_rhsIter && (!m_lhsIter || (m_lhsIter.index() > m_rhsIter.index()))) - { - m_id = m_rhsIter.index(); - m_value = m_functor(Scalar(0), m_rhsIter.value()); - ++m_rhsIter; - } - else - { - m_value = 0; // this is to avoid a compilation warning - m_id = -1; - } - return *static_cast(this); - } - - EIGEN_STRONG_INLINE Scalar value() const { return m_value; } - - EIGEN_STRONG_INLINE Index index() const { return m_id; } - EIGEN_STRONG_INLINE Index row() const { return Lhs::IsRowMajor ? m_lhsIter.row() : index(); } - EIGEN_STRONG_INLINE Index col() const { return Lhs::IsRowMajor ? index() : m_lhsIter.col(); } - - EIGEN_STRONG_INLINE operator bool() const { return m_id>=0; } - - protected: - LhsIterator m_lhsIter; - RhsIterator m_rhsIter; - const BinaryOp& m_functor; - Scalar m_value; - Index m_id; -}; - -// sparse - sparse (product) -template -class sparse_cwise_binary_op_inner_iterator_selector, Lhs, Rhs, Derived, Sparse, Sparse> -{ - typedef scalar_product_op BinaryFunc; - typedef CwiseBinaryOp CwiseBinaryXpr; - typedef typename CwiseBinaryXpr::Scalar Scalar; - typedef typename CwiseBinaryXpr::Index Index; - typedef typename traits::_LhsNested _LhsNested; - typedef typename _LhsNested::InnerIterator LhsIterator; - typedef typename traits::_RhsNested _RhsNested; - typedef typename _RhsNested::InnerIterator RhsIterator; - public: - - EIGEN_STRONG_INLINE sparse_cwise_binary_op_inner_iterator_selector(const CwiseBinaryXpr& xpr, Index outer) - : m_lhsIter(xpr.lhs(),outer), m_rhsIter(xpr.rhs(),outer), m_functor(xpr.functor()) - { - while (m_lhsIter && m_rhsIter && (m_lhsIter.index() != m_rhsIter.index())) - { - if (m_lhsIter.index() < m_rhsIter.index()) - ++m_lhsIter; - else - ++m_rhsIter; - } - } - - EIGEN_STRONG_INLINE Derived& operator++() - { - ++m_lhsIter; - ++m_rhsIter; - while (m_lhsIter && m_rhsIter && (m_lhsIter.index() != m_rhsIter.index())) - { - if (m_lhsIter.index() < m_rhsIter.index()) - ++m_lhsIter; - else - ++m_rhsIter; - } - return *static_cast(this); - } - - EIGEN_STRONG_INLINE Scalar value() const { return m_functor(m_lhsIter.value(), m_rhsIter.value()); } - - EIGEN_STRONG_INLINE Index index() const { return m_lhsIter.index(); } - EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); } - EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); } - - EIGEN_STRONG_INLINE operator bool() const { return (m_lhsIter && m_rhsIter); } - - protected: - LhsIterator m_lhsIter; - RhsIterator m_rhsIter; - const BinaryFunc& m_functor; -}; - -// sparse - dense (product) -template -class sparse_cwise_binary_op_inner_iterator_selector, Lhs, Rhs, Derived, Sparse, Dense> -{ - typedef scalar_product_op BinaryFunc; - typedef CwiseBinaryOp CwiseBinaryXpr; - typedef typename CwiseBinaryXpr::Scalar Scalar; - typedef typename CwiseBinaryXpr::Index Index; - typedef typename traits::_LhsNested _LhsNested; - typedef typename traits::RhsNested RhsNested; - typedef typename _LhsNested::InnerIterator LhsIterator; - enum { IsRowMajor = (int(Lhs::Flags)&RowMajorBit)==RowMajorBit }; - public: - - EIGEN_STRONG_INLINE sparse_cwise_binary_op_inner_iterator_selector(const CwiseBinaryXpr& xpr, Index outer) - : m_rhs(xpr.rhs()), m_lhsIter(xpr.lhs(),typename _LhsNested::Index(outer)), m_functor(xpr.functor()), m_outer(outer) - {} - - EIGEN_STRONG_INLINE Derived& operator++() - { - ++m_lhsIter; - return *static_cast(this); - } - - EIGEN_STRONG_INLINE Scalar value() const - { return m_functor(m_lhsIter.value(), - m_rhs.coeff(IsRowMajor?m_outer:m_lhsIter.index(),IsRowMajor?m_lhsIter.index():m_outer)); } - - EIGEN_STRONG_INLINE Index index() const { return m_lhsIter.index(); } - EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); } - EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); } - - EIGEN_STRONG_INLINE operator bool() const { return m_lhsIter; } - - protected: - RhsNested m_rhs; - LhsIterator m_lhsIter; - const BinaryFunc m_functor; - const Index m_outer; -}; - -// sparse - dense (product) -template -class sparse_cwise_binary_op_inner_iterator_selector, Lhs, Rhs, Derived, Dense, Sparse> -{ - typedef scalar_product_op BinaryFunc; - typedef CwiseBinaryOp CwiseBinaryXpr; - typedef typename CwiseBinaryXpr::Scalar Scalar; - typedef typename CwiseBinaryXpr::Index Index; - typedef typename traits::_RhsNested _RhsNested; - typedef typename _RhsNested::InnerIterator RhsIterator; - - enum { IsRowMajor = (int(Rhs::Flags)&RowMajorBit)==RowMajorBit }; - public: - - EIGEN_STRONG_INLINE sparse_cwise_binary_op_inner_iterator_selector(const CwiseBinaryXpr& xpr, Index outer) - : m_xpr(xpr), m_rhsIter(xpr.rhs(),outer), m_functor(xpr.functor()), m_outer(outer) - {} - - EIGEN_STRONG_INLINE Derived& operator++() - { - ++m_rhsIter; - return *static_cast(this); - } - - EIGEN_STRONG_INLINE Scalar value() const - { return m_functor(m_xpr.lhs().coeff(IsRowMajor?m_outer:m_rhsIter.index(),IsRowMajor?m_rhsIter.index():m_outer), m_rhsIter.value()); } - - EIGEN_STRONG_INLINE Index index() const { return m_rhsIter.index(); } - EIGEN_STRONG_INLINE Index row() const { return m_rhsIter.row(); } - EIGEN_STRONG_INLINE Index col() const { return m_rhsIter.col(); } - - EIGEN_STRONG_INLINE operator bool() const { return m_rhsIter; } - - protected: - const CwiseBinaryXpr& m_xpr; - RhsIterator m_rhsIter; - const BinaryFunc& m_functor; - const Index m_outer; -}; - -} // end namespace internal - -#else // EIGEN_TEST_EVALUATORS - namespace internal { @@ -590,10 +340,6 @@ protected: } -#endif - - - /*************************************************************************** * Implementation of SparseMatrixBase and SparseCwise functions/operators ***************************************************************************/ diff --git a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h index 0e0f9f5f5..6036fd0a7 100644 --- a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h @@ -11,136 +11,6 @@ #define EIGEN_SPARSE_CWISE_UNARY_OP_H namespace Eigen { - -#ifndef EIGEN_TEST_EVALUATORS - -template -class CwiseUnaryOpImpl - : public SparseMatrixBase > -{ - public: - - typedef CwiseUnaryOp Derived; - EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) - - class InnerIterator; - class ReverseInnerIterator; - - protected: - typedef typename internal::traits::_XprTypeNested _MatrixTypeNested; - typedef typename _MatrixTypeNested::InnerIterator MatrixTypeIterator; - typedef typename _MatrixTypeNested::ReverseInnerIterator MatrixTypeReverseIterator; -}; - -template -class CwiseUnaryOpImpl::InnerIterator - : public CwiseUnaryOpImpl::MatrixTypeIterator -{ - typedef typename CwiseUnaryOpImpl::Scalar Scalar; - typedef typename CwiseUnaryOpImpl::MatrixTypeIterator Base; - public: - - EIGEN_STRONG_INLINE InnerIterator(const CwiseUnaryOpImpl& unaryOp, typename CwiseUnaryOpImpl::Index outer) - : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) - {} - - EIGEN_STRONG_INLINE InnerIterator& operator++() - { Base::operator++(); return *this; } - - EIGEN_STRONG_INLINE typename CwiseUnaryOpImpl::Scalar value() const { return m_functor(Base::value()); } - - protected: - const UnaryOp m_functor; - private: - typename CwiseUnaryOpImpl::Scalar& valueRef(); -}; - -template -class CwiseUnaryOpImpl::ReverseInnerIterator - : public CwiseUnaryOpImpl::MatrixTypeReverseIterator -{ - typedef typename CwiseUnaryOpImpl::Scalar Scalar; - typedef typename CwiseUnaryOpImpl::MatrixTypeReverseIterator Base; - public: - - EIGEN_STRONG_INLINE ReverseInnerIterator(const CwiseUnaryOpImpl& unaryOp, typename CwiseUnaryOpImpl::Index outer) - : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) - {} - - EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() - { Base::operator--(); return *this; } - - EIGEN_STRONG_INLINE typename CwiseUnaryOpImpl::Scalar value() const { return m_functor(Base::value()); } - - protected: - const UnaryOp m_functor; - private: - typename CwiseUnaryOpImpl::Scalar& valueRef(); -}; - -template -class CwiseUnaryViewImpl - : public SparseMatrixBase > -{ - public: - - class InnerIterator; - class ReverseInnerIterator; - - typedef CwiseUnaryView Derived; - EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) - - protected: - typedef typename internal::traits::_MatrixTypeNested _MatrixTypeNested; - typedef typename _MatrixTypeNested::InnerIterator MatrixTypeIterator; - typedef typename _MatrixTypeNested::ReverseInnerIterator MatrixTypeReverseIterator; -}; - -template -class CwiseUnaryViewImpl::InnerIterator - : public CwiseUnaryViewImpl::MatrixTypeIterator -{ - typedef typename CwiseUnaryViewImpl::Scalar Scalar; - typedef typename CwiseUnaryViewImpl::MatrixTypeIterator Base; - public: - - EIGEN_STRONG_INLINE InnerIterator(const CwiseUnaryViewImpl& unaryOp, typename CwiseUnaryViewImpl::Index outer) - : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) - {} - - EIGEN_STRONG_INLINE InnerIterator& operator++() - { Base::operator++(); return *this; } - - EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar value() const { return m_functor(Base::value()); } - EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar& valueRef() { return m_functor(Base::valueRef()); } - - protected: - const ViewOp m_functor; -}; - -template -class CwiseUnaryViewImpl::ReverseInnerIterator - : public CwiseUnaryViewImpl::MatrixTypeReverseIterator -{ - typedef typename CwiseUnaryViewImpl::Scalar Scalar; - typedef typename CwiseUnaryViewImpl::MatrixTypeReverseIterator Base; - public: - - EIGEN_STRONG_INLINE ReverseInnerIterator(const CwiseUnaryViewImpl& unaryOp, typename CwiseUnaryViewImpl::Index outer) - : Base(unaryOp.derived().nestedExpression(),outer), m_functor(unaryOp.derived().functor()) - {} - - EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() - { Base::operator--(); return *this; } - - EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar value() const { return m_functor(Base::value()); } - EIGEN_STRONG_INLINE typename CwiseUnaryViewImpl::Scalar& valueRef() { return m_functor(Base::valueRef()); } - - protected: - const ViewOp m_functor; -}; - -#else // EIGEN_TEST_EVALUATORS namespace internal { @@ -291,8 +161,6 @@ class unary_evaluator, IteratorBased>::InnerItera } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - template EIGEN_STRONG_INLINE Derived& SparseMatrixBase::operator*=(const Scalar& other) diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 803d98e2d..04c838a71 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -30,18 +30,10 @@ struct sparse_time_dense_product_impl::type Rhs; typedef typename internal::remove_all::type Res; typedef typename Lhs::Index Index; -#ifndef EIGEN_TEST_EVALUATORS - typedef typename Lhs::InnerIterator LhsInnerIterator; -#else typedef typename evaluator::InnerIterator LhsInnerIterator; -#endif static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) { -#ifndef EIGEN_TEST_EVALUATORS - const Lhs &lhsEval(lhs); -#else - typename evaluator::type lhsEval(lhs); -#endif + typename evaluator::type lhsEval(lhs); for(Index c=0; c::type Rhs; typedef typename internal::remove_all::type Res; typedef typename Lhs::Index Index; -#ifndef EIGEN_TEST_EVALUATORS - typedef typename Lhs::InnerIterator LhsInnerIterator; -#else typedef typename evaluator::InnerIterator LhsInnerIterator; -#endif static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) { -#ifndef EIGEN_TEST_EVALUATORS - const Lhs &lhsEval(lhs); -#else - typename evaluator::type lhsEval(lhs); -#endif + typename evaluator::type lhsEval(lhs); for(Index c=0; c::type Rhs; typedef typename internal::remove_all::type Res; typedef typename Lhs::Index Index; -#ifndef EIGEN_TEST_EVALUATORS - typedef typename Lhs::InnerIterator LhsInnerIterator; -#else typedef typename evaluator::InnerIterator LhsInnerIterator; -#endif static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) { -#ifndef EIGEN_TEST_EVALUATORS - const Lhs &lhsEval(lhs); -#else - typename evaluator::type lhsEval(lhs); -#endif + typename evaluator::type lhsEval(lhs); for(Index j=0; j::type Rhs; typedef typename internal::remove_all::type Res; typedef typename Lhs::Index Index; -#ifndef EIGEN_TEST_EVALUATORS - typedef typename Lhs::InnerIterator LhsInnerIterator; -#else typedef typename evaluator::InnerIterator LhsInnerIterator; -#endif static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) { -#ifndef EIGEN_TEST_EVALUATORS - const Lhs &lhsEval(lhs); -#else - typename evaluator::type lhsEval(lhs); -#endif + typename evaluator::type lhsEval(lhs); for(Index j=0; j struct SparseDenseProductReturnType -{ - typedef SparseTimeDenseProduct Type; -}; - -template struct SparseDenseProductReturnType -{ - typedef typename internal::conditional< - Lhs::IsRowMajor, - SparseDenseOuterProduct, - SparseDenseOuterProduct >::type Type; -}; - -template struct DenseSparseProductReturnType -{ - typedef DenseTimeSparseProduct Type; -}; - -template struct DenseSparseProductReturnType -{ - typedef typename internal::conditional< - Rhs::IsRowMajor, - SparseDenseOuterProduct, - SparseDenseOuterProduct >::type Type; -}; - -namespace internal { - -template -struct traits > -{ - typedef Sparse StorageKind; - typedef typename scalar_product_traits::Scalar, - typename traits::Scalar>::ReturnType Scalar; - typedef typename Lhs::Index Index; - typedef typename Lhs::Nested LhsNested; - typedef typename Rhs::Nested RhsNested; - typedef typename remove_all::type _LhsNested; - typedef typename remove_all::type _RhsNested; - - enum { - LhsCoeffReadCost = traits<_LhsNested>::CoeffReadCost, - RhsCoeffReadCost = traits<_RhsNested>::CoeffReadCost, - - RowsAtCompileTime = Tr ? int(traits::RowsAtCompileTime) : int(traits::RowsAtCompileTime), - ColsAtCompileTime = Tr ? int(traits::ColsAtCompileTime) : int(traits::ColsAtCompileTime), - MaxRowsAtCompileTime = Tr ? int(traits::MaxRowsAtCompileTime) : int(traits::MaxRowsAtCompileTime), - MaxColsAtCompileTime = Tr ? int(traits::MaxColsAtCompileTime) : int(traits::MaxColsAtCompileTime), - - Flags = Tr ? RowMajorBit : 0, - - CoeffReadCost = LhsCoeffReadCost + RhsCoeffReadCost + NumTraits::MulCost - }; -}; - -} // end namespace internal - -template -class SparseDenseOuterProduct - : public SparseMatrixBase > -{ - public: - - typedef SparseMatrixBase Base; - EIGEN_DENSE_PUBLIC_INTERFACE(SparseDenseOuterProduct) - typedef internal::traits Traits; - - private: - - typedef typename Traits::LhsNested LhsNested; - typedef typename Traits::RhsNested RhsNested; - typedef typename Traits::_LhsNested _LhsNested; - typedef typename Traits::_RhsNested _RhsNested; - - public: - - class InnerIterator; - - EIGEN_STRONG_INLINE SparseDenseOuterProduct(const Lhs& lhs, const Rhs& rhs) - : m_lhs(lhs), m_rhs(rhs) - { - EIGEN_STATIC_ASSERT(!Tr,YOU_MADE_A_PROGRAMMING_MISTAKE); - } - - EIGEN_STRONG_INLINE SparseDenseOuterProduct(const Rhs& rhs, const Lhs& lhs) - : m_lhs(lhs), m_rhs(rhs) - { - EIGEN_STATIC_ASSERT(Tr,YOU_MADE_A_PROGRAMMING_MISTAKE); - } - - EIGEN_STRONG_INLINE Index rows() const { return Tr ? Index(m_rhs.rows()) : m_lhs.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return Tr ? m_lhs.cols() : Index(m_rhs.cols()); } - - EIGEN_STRONG_INLINE const _LhsNested& lhs() const { return m_lhs; } - EIGEN_STRONG_INLINE const _RhsNested& rhs() const { return m_rhs; } - - protected: - LhsNested m_lhs; - RhsNested m_rhs; -}; - -template -class SparseDenseOuterProduct::InnerIterator : public _LhsNested::InnerIterator -{ - typedef typename _LhsNested::InnerIterator Base; - typedef typename SparseDenseOuterProduct::Index Index; - public: - EIGEN_STRONG_INLINE InnerIterator(const SparseDenseOuterProduct& prod, Index outer) - : Base(prod.lhs(), 0), m_outer(outer), m_empty(false), m_factor(get(prod.rhs(), outer, typename internal::traits::StorageKind() )) - {} - - inline Index outer() const { return m_outer; } - inline Index row() const { return Transpose ? m_outer : Base::index(); } - inline Index col() const { return Transpose ? Base::index() : m_outer; } - - inline Scalar value() const { return Base::value() * m_factor; } - inline operator bool() const { return Base::operator bool() && !m_empty; } - - protected: - Scalar get(const _RhsNested &rhs, Index outer, Dense = Dense()) const - { - return rhs.coeff(outer); - } - - Scalar get(const _RhsNested &rhs, Index outer, Sparse = Sparse()) - { - typename Traits::_RhsNested::InnerIterator it(rhs, outer); - if (it && it.index()==0 && it.value()!=Scalar(0)) - return it.value(); - m_empty = true; - return Scalar(0); - } - - Index m_outer; - bool m_empty; - Scalar m_factor; -}; - -namespace internal { -template -struct traits > - : traits, Lhs, Rhs> > -{ - typedef Dense StorageKind; - typedef MatrixXpr XprKind; -}; - -} // end namespace internal - -template -class SparseTimeDenseProduct - : public ProductBase, Lhs, Rhs> -{ - public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(SparseTimeDenseProduct) - - SparseTimeDenseProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) - {} - - template void scaleAndAddTo(Dest& dest, const Scalar& alpha) const - { - internal::sparse_time_dense_product(m_lhs, m_rhs, dest, alpha); - } - - private: - SparseTimeDenseProduct& operator=(const SparseTimeDenseProduct&); -}; - - -// dense = dense * sparse -namespace internal { -template -struct traits > - : traits, Lhs, Rhs> > -{ - typedef Dense StorageKind; -}; -} // end namespace internal - -template -class DenseTimeSparseProduct - : public ProductBase, Lhs, Rhs> -{ - public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(DenseTimeSparseProduct) - - DenseTimeSparseProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) - {} - - template void scaleAndAddTo(Dest& dest, const Scalar& alpha) const - { - Transpose lhs_t(m_lhs); - Transpose rhs_t(m_rhs); - Transpose dest_t(dest); - internal::sparse_time_dense_product(rhs_t, lhs_t, dest_t, alpha); - } - - private: - DenseTimeSparseProduct& operator=(const DenseTimeSparseProduct&); -}; - -#endif // EIGEN_TEST_EVALUATORS - -#ifdef EIGEN_TEST_EVALUATORS - namespace internal { template @@ -514,8 +276,6 @@ struct product_evaluator, OuterProduct, DenseS } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SPARSEDENSEPRODUCT_H diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index 9f465a828..0cb2bd572 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -24,171 +24,8 @@ namespace Eigen { // for that particular case // The two other cases are symmetric. -#ifndef EIGEN_TEST_EVALUATORS - -namespace internal { - -template -struct traits > -{ - typedef typename remove_all::type _Lhs; - typedef typename remove_all::type _Rhs; - typedef typename _Lhs::Scalar Scalar; - // propagate the index type of the sparse matrix - typedef typename conditional< is_diagonal<_Lhs>::ret, - typename traits::Index, - typename traits::Index>::type Index; - typedef Sparse StorageKind; - typedef MatrixXpr XprKind; - enum { - RowsAtCompileTime = _Lhs::RowsAtCompileTime, - ColsAtCompileTime = _Rhs::ColsAtCompileTime, - - MaxRowsAtCompileTime = _Lhs::MaxRowsAtCompileTime, - MaxColsAtCompileTime = _Rhs::MaxColsAtCompileTime, - - SparseFlags = is_diagonal<_Lhs>::ret ? int(_Rhs::Flags) : int(_Lhs::Flags), - Flags = (SparseFlags&RowMajorBit), - CoeffReadCost = Dynamic - }; -}; - -enum {SDP_IsDiagonal, SDP_IsSparseRowMajor, SDP_IsSparseColMajor}; -template -class sparse_diagonal_product_inner_iterator_selector; - -} // end namespace internal - -template -class SparseDiagonalProduct - : public SparseMatrixBase >, - internal::no_assignment_operator -{ - typedef typename Lhs::Nested LhsNested; - typedef typename Rhs::Nested RhsNested; - - typedef typename internal::remove_all::type _LhsNested; - typedef typename internal::remove_all::type _RhsNested; - - enum { - LhsMode = internal::is_diagonal<_LhsNested>::ret ? internal::SDP_IsDiagonal - : (_LhsNested::Flags&RowMajorBit) ? internal::SDP_IsSparseRowMajor : internal::SDP_IsSparseColMajor, - RhsMode = internal::is_diagonal<_RhsNested>::ret ? internal::SDP_IsDiagonal - : (_RhsNested::Flags&RowMajorBit) ? internal::SDP_IsSparseRowMajor : internal::SDP_IsSparseColMajor - }; - - public: - - EIGEN_SPARSE_PUBLIC_INTERFACE(SparseDiagonalProduct) - - typedef internal::sparse_diagonal_product_inner_iterator_selector - <_LhsNested,_RhsNested,SparseDiagonalProduct,LhsMode,RhsMode> InnerIterator; - - // We do not want ReverseInnerIterator for diagonal-sparse products, - // but this dummy declaration is needed to make diag * sparse * diag compile. - class ReverseInnerIterator; - - EIGEN_STRONG_INLINE SparseDiagonalProduct(const Lhs& lhs, const Rhs& rhs) - : m_lhs(lhs), m_rhs(rhs) - { - eigen_assert(lhs.cols() == rhs.rows() && "invalid sparse matrix * diagonal matrix product"); - } - - EIGEN_STRONG_INLINE Index rows() const { return Index(m_lhs.rows()); } - EIGEN_STRONG_INLINE Index cols() const { return Index(m_rhs.cols()); } - - EIGEN_STRONG_INLINE const _LhsNested& lhs() const { return m_lhs; } - EIGEN_STRONG_INLINE const _RhsNested& rhs() const { return m_rhs; } - - protected: - LhsNested m_lhs; - RhsNested m_rhs; -}; -#endif - namespace internal { -#ifndef EIGEN_TEST_EVALUATORS - - - -template -class sparse_diagonal_product_inner_iterator_selector - - : public CwiseUnaryOp,const Rhs>::InnerIterator -{ - typedef typename CwiseUnaryOp,const Rhs>::InnerIterator Base; - typedef typename Rhs::Index Index; - public: - inline sparse_diagonal_product_inner_iterator_selector( - const SparseDiagonalProductType& expr, Index outer) - : Base(expr.rhs()*(expr.lhs().diagonal().coeff(outer)), outer) - {} -}; - -template -class sparse_diagonal_product_inner_iterator_selector - - : public CwiseBinaryOp< - scalar_product_op, - const typename Rhs::ConstInnerVectorReturnType, - const typename Lhs::DiagonalVectorType>::InnerIterator -{ - typedef typename CwiseBinaryOp< - scalar_product_op, - const typename Rhs::ConstInnerVectorReturnType, - const typename Lhs::DiagonalVectorType>::InnerIterator Base; - typedef typename Rhs::Index Index; - Index m_outer; - public: - inline sparse_diagonal_product_inner_iterator_selector( - const SparseDiagonalProductType& expr, Index outer) - : Base(expr.rhs().innerVector(outer) .cwiseProduct(expr.lhs().diagonal()), 0), m_outer(outer) - {} - - inline Index outer() const { return m_outer; } - inline Index col() const { return m_outer; } -}; - -template -class sparse_diagonal_product_inner_iterator_selector - - : public CwiseUnaryOp,const Lhs>::InnerIterator -{ - typedef typename CwiseUnaryOp,const Lhs>::InnerIterator Base; - typedef typename Lhs::Index Index; - public: - inline sparse_diagonal_product_inner_iterator_selector( - const SparseDiagonalProductType& expr, Index outer) - : Base(expr.lhs()*expr.rhs().diagonal().coeff(outer), outer) - {} -}; - -template -class sparse_diagonal_product_inner_iterator_selector - - : public CwiseBinaryOp< - scalar_product_op, - const typename Lhs::ConstInnerVectorReturnType, - const Transpose >::InnerIterator -{ - typedef typename CwiseBinaryOp< - scalar_product_op, - const typename Lhs::ConstInnerVectorReturnType, - const Transpose >::InnerIterator Base; - typedef typename Lhs::Index Index; - Index m_outer; - public: - inline sparse_diagonal_product_inner_iterator_selector( - const SparseDiagonalProductType& expr, Index outer) - : Base(expr.lhs().innerVector(outer) .cwiseProduct(expr.rhs().diagonal().transpose()), 0), m_outer(outer) - {} - - inline Index outer() const { return m_outer; } - inline Index row() const { return m_outer; } -}; - -#else // EIGEN_TEST_EVALUATORS enum { SDP_AsScalarProduct, SDP_AsCwiseProduct @@ -303,23 +140,8 @@ protected: : SparseXprType::ColsAtCompileTime>::type m_diagCoeffNested; }; -#endif // EIGEN_TEST_EVALUATORS - - } // end namespace internal -#ifndef EIGEN_TEST_EVALUATORS - -// SparseMatrixBase functions -template -template -const SparseDiagonalProduct -SparseMatrixBase::operator*(const DiagonalBase &other) const -{ - return SparseDiagonalProduct(this->derived(), other.derived()); -} -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SPARSE_DIAGONAL_PRODUCT_H diff --git a/Eigen/src/SparseCore/SparseDot.h b/Eigen/src/SparseCore/SparseDot.h index a63cb003c..b10c8058f 100644 --- a/Eigen/src/SparseCore/SparseDot.h +++ b/Eigen/src/SparseCore/SparseDot.h @@ -26,12 +26,8 @@ SparseMatrixBase::dot(const MatrixBase& other) const eigen_assert(size() == other.size()); eigen_assert(other.size()>0 && "you are using a non initialized vector"); -#ifndef EIGEN_TEST_EVALUATORS - typename Derived::InnerIterator i(derived(),0); -#else typename internal::evaluator::type thisEval(derived()); typename internal::evaluator::InnerIterator i(thisEval, 0); -#endif Scalar res(0); while (i) { @@ -54,24 +50,12 @@ SparseMatrixBase::dot(const SparseMatrixBase& other) cons eigen_assert(size() == other.size()); -#ifndef EIGEN_TEST_EVALUATORS - typedef typename Derived::Nested Nested; - typedef typename OtherDerived::Nested OtherNested; - typedef typename internal::remove_all::type NestedCleaned; - typedef typename internal::remove_all::type OtherNestedCleaned; - - Nested nthis(derived()); - OtherNested nother(other.derived()); - - typename NestedCleaned::InnerIterator i(nthis,0); - typename OtherNestedCleaned::InnerIterator j(nother,0); -#else typename internal::evaluator::type thisEval(derived()); typename internal::evaluator::InnerIterator i(thisEval, 0); typename internal::evaluator::type otherEval(other.derived()); typename internal::evaluator::InnerIterator j(otherEval, 0); -#endif + Scalar res(0); while (i && j) { diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 9e2dae1b3..72368ebf3 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -52,9 +52,6 @@ struct traits > MaxRowsAtCompileTime = Dynamic, MaxColsAtCompileTime = Dynamic, Flags = _Options | NestByRefBit | LvalueBit, -#ifndef EIGEN_TEST_EVALUATORS - CoeffReadCost = NumTraits::ReadCost, -#endif SupportedAccessPatterns = InnerRandomAccessPattern }; }; @@ -77,9 +74,6 @@ struct traits, DiagIndex> MaxRowsAtCompileTime = Dynamic, MaxColsAtCompileTime = 1, Flags = 0 -#ifndef EIGEN_TEST_EVALUATORS - , CoeffReadCost = _MatrixTypeNested::CoeffReadCost*10 -#endif }; }; @@ -653,13 +647,9 @@ class SparseMatrix EIGEN_STATIC_ASSERT((internal::is_same::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) check_template_parameters(); -#ifndef EIGEN_TEST_EVALUATORS - *this = other.derived(); -#else const bool needToTranspose = (Flags & RowMajorBit) != (internal::evaluator::Flags & RowMajorBit); if (needToTranspose) *this = other.derived(); else internal::call_assignment_no_alias(*this, other.derived()); -#endif } /** Constructs a sparse matrix from the sparse selfadjoint view \a other */ @@ -668,11 +658,7 @@ class SparseMatrix : m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0) { check_template_parameters(); -#ifndef EIGEN_TEST_EVALUATORS - *this = other; -#else Base::operator=(other); -#endif } /** Copy constructor (it performs a deep copy) */ @@ -737,19 +723,6 @@ class SparseMatrix } #ifndef EIGEN_PARSED_BY_DOXYGEN -#ifndef EIGEN_TEST_EVALUATORS - template - inline SparseMatrix& operator=(const SparseSparseProduct& product) - { return Base::operator=(product); } - - template - inline SparseMatrix& operator=(const ReturnByValue& other) - { - initAssignment(other); - return Base::operator=(other.derived()); - } -#endif // EIGEN_TEST_EVALUATORS - template inline SparseMatrix& operator=(const EigenBase& other) { return Base::operator=(other.derived()); } @@ -1071,69 +1044,6 @@ void SparseMatrix::sumupDuplicates() m_data.resize(m_outerIndex[m_outerSize]); } -#ifndef EIGEN_TEST_EVALUATORS -template -template -EIGEN_DONT_INLINE SparseMatrix& SparseMatrix::operator=(const SparseMatrixBase& other) -{ - EIGEN_STATIC_ASSERT((internal::is_same::value), - YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) - - const bool needToTranspose = (Flags & RowMajorBit) != (OtherDerived::Flags & RowMajorBit); - if (needToTranspose) - { - // two passes algorithm: - // 1 - compute the number of coeffs per dest inner vector - // 2 - do the actual copy/eval - // Since each coeff of the rhs has to be evaluated twice, let's evaluate it if needed - typedef typename internal::nested::type OtherCopy; - typedef typename internal::remove_all::type _OtherCopy; - OtherCopy otherCopy(other.derived()); - - SparseMatrix dest(other.rows(),other.cols()); - Eigen::Map > (dest.m_outerIndex,dest.outerSize()).setZero(); - - // pass 1 - // FIXME the above copy could be merged with that pass - for (Index j=0; j positions(dest.outerSize()); - for (Index j=0; jswap(dest); - return *this; - } - else - { - if(other.isRValue()) - initAssignment(other.derived()); - // there is no special optimization - return Base::operator=(other.derived()); - } -} -#else template template EIGEN_DONT_INLINE SparseMatrix& SparseMatrix::operator=(const SparseMatrixBase& other) @@ -1199,7 +1109,6 @@ EIGEN_DONT_INLINE SparseMatrix& SparseMatrix EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& SparseMatrix<_Scalar,_Options,_Index>::insertUncompressed(Index row, Index col) @@ -1340,7 +1249,6 @@ EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& Sparse return (m_data.value(p) = 0); } -#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template @@ -1370,7 +1278,6 @@ struct evaluator > }; } -#endif } // end namespace Eigen diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index aff4c8b6f..b5c50d93a 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -79,13 +79,6 @@ template class SparseMatrixBase : public EigenBase * constructed from this one. See the \ref flags "list of flags". */ -#ifndef EIGEN_TEST_EVALUATORS - CoeffReadCost = internal::traits::CoeffReadCost, - /**< This is a rough measure of how expensive it is to read one coefficient from - * this expression. - */ -#endif - IsRowMajor = Flags&RowMajorBit ? 1 : 0, InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime) @@ -189,11 +182,6 @@ template class SparseMatrixBase : public EigenBase public: -#ifndef EIGEN_TEST_EVALUATORS - template - inline Derived& operator=(const SparseSparseProduct& product); -#endif - friend std::ostream & operator << (std::ostream & s, const SparseMatrixBase& m) { typedef typename Derived::Nested Nested; @@ -265,36 +253,6 @@ template class SparseMatrixBase : public EigenBase EIGEN_STRONG_INLINE const EIGEN_SPARSE_CWISE_PRODUCT_RETURN_TYPE cwiseProduct(const MatrixBase &other) const; -#ifndef EIGEN_TEST_EVALUATORS - // sparse * sparse - template - const typename SparseSparseProductReturnType::Type - operator*(const SparseMatrixBase &other) const; - - // sparse * diagonal - template - const SparseDiagonalProduct - operator*(const DiagonalBase &other) const; - - // diagonal * sparse - template friend - const SparseDiagonalProduct - operator*(const DiagonalBase &lhs, const SparseMatrixBase& rhs) - { return SparseDiagonalProduct(lhs.derived(), rhs.derived()); } - - /** dense * sparse (return a dense object unless it is an outer product) */ - template friend - const typename DenseSparseProductReturnType::Type - operator*(const MatrixBase& lhs, const Derived& rhs) - { return typename DenseSparseProductReturnType::Type(lhs.derived(),rhs); } - - /** sparse * dense (returns a dense object unless it is an outer product) */ - template - const typename SparseDenseProductReturnType::Type - operator*(const MatrixBase &other) const - { return typename SparseDenseProductReturnType::Type(derived(), other.derived()); } - -#else // EIGEN_TEST_EVALUATORS // sparse * diagonal template const Product @@ -323,7 +281,6 @@ template class SparseMatrixBase : public EigenBase const Product operator*(const MatrixBase &lhs, const SparseMatrixBase& rhs) { return Product(lhs.derived(), rhs.derived()); } -#endif // EIGEN_TEST_EVALUATORS /** \returns an expression of P H P^-1 where H is the matrix represented by \c *this */ SparseSymmetricPermutationProduct twistedBy(const PermutationMatrix& perm) const @@ -360,18 +317,6 @@ template class SparseMatrixBase : public EigenBase Block innerVectors(Index outerStart, Index outerSize); const Block innerVectors(Index outerStart, Index outerSize) const; -#ifndef EIGEN_TEST_EVALUATORS - /** \internal use operator= */ - template - void evalTo(MatrixBase& dst) const - { - dst.setZero(); - for (Index j=0; j toDense() const { return derived(); diff --git a/Eigen/src/SparseCore/SparsePermutation.h b/Eigen/src/SparseCore/SparsePermutation.h index a888ae9e1..228796bf8 100644 --- a/Eigen/src/SparseCore/SparsePermutation.h +++ b/Eigen/src/SparseCore/SparsePermutation.h @@ -103,48 +103,6 @@ struct permut_sparsematrix_product_retval } -#ifndef EIGEN_TEST_EVALUATORS - -/** \returns the matrix with the permutation applied to the columns - */ -template -inline const internal::permut_sparsematrix_product_retval, SparseDerived, OnTheRight, false> -operator*(const SparseMatrixBase& matrix, const PermutationBase& perm) -{ - return internal::permut_sparsematrix_product_retval, SparseDerived, OnTheRight, false>(perm, matrix.derived()); -} - -/** \returns the matrix with the permutation applied to the rows - */ -template -inline const internal::permut_sparsematrix_product_retval, SparseDerived, OnTheLeft, false> -operator*( const PermutationBase& perm, const SparseMatrixBase& matrix) -{ - return internal::permut_sparsematrix_product_retval, SparseDerived, OnTheLeft, false>(perm, matrix.derived()); -} - - - -/** \returns the matrix with the inverse permutation applied to the columns. - */ -template -inline const internal::permut_sparsematrix_product_retval, SparseDerived, OnTheRight, true> -operator*(const SparseMatrixBase& matrix, const Transpose >& tperm) -{ - return internal::permut_sparsematrix_product_retval, SparseDerived, OnTheRight, true>(tperm.nestedPermutation(), matrix.derived()); -} - -/** \returns the matrix with the inverse permutation applied to the rows. - */ -template -inline const internal::permut_sparsematrix_product_retval, SparseDerived, OnTheLeft, true> -operator*(const Transpose >& tperm, const SparseMatrixBase& matrix) -{ - return internal::permut_sparsematrix_product_retval, SparseDerived, OnTheLeft, true>(tperm.nestedPermutation(), matrix.derived()); -} - -#else // EIGEN_TEST_EVALUATORS - namespace internal { template struct product_promote_storage_type { typedef Sparse ret; }; @@ -274,8 +232,6 @@ operator*(const Transpose >& tperm, const SparseMat return Product >, SparseDerived>(tperm, matrix.derived()); } -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SPARSE_SELFADJOINTVIEW_H diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index 18f40b9d9..b68fe986a 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -12,182 +12,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS - -template -struct SparseSparseProductReturnType -{ - typedef typename internal::traits::Scalar Scalar; - typedef typename internal::traits::Index Index; - enum { - LhsRowMajor = internal::traits::Flags & RowMajorBit, - RhsRowMajor = internal::traits::Flags & RowMajorBit, - TransposeRhs = (!LhsRowMajor) && RhsRowMajor, - TransposeLhs = LhsRowMajor && (!RhsRowMajor) - }; - - typedef typename internal::conditional, - typename internal::nested::type>::type LhsNested; - - typedef typename internal::conditional, - typename internal::nested::type>::type RhsNested; - - typedef SparseSparseProduct Type; -}; - -namespace internal { -template -struct traits > -{ - typedef MatrixXpr XprKind; - // clean the nested types: - typedef typename remove_all::type _LhsNested; - typedef typename remove_all::type _RhsNested; - typedef typename _LhsNested::Scalar Scalar; - typedef typename promote_index_type::Index, - typename traits<_RhsNested>::Index>::type Index; - - enum { - LhsCoeffReadCost = _LhsNested::CoeffReadCost, - RhsCoeffReadCost = _RhsNested::CoeffReadCost, - LhsFlags = _LhsNested::Flags, - RhsFlags = _RhsNested::Flags, - - RowsAtCompileTime = _LhsNested::RowsAtCompileTime, - ColsAtCompileTime = _RhsNested::ColsAtCompileTime, - MaxRowsAtCompileTime = _LhsNested::MaxRowsAtCompileTime, - MaxColsAtCompileTime = _RhsNested::MaxColsAtCompileTime, - - InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(_LhsNested::ColsAtCompileTime, _RhsNested::RowsAtCompileTime), - - EvalToRowMajor = (RhsFlags & LhsFlags & RowMajorBit), - - RemovedBits = ~(EvalToRowMajor ? 0 : RowMajorBit), - - Flags = (int(LhsFlags | RhsFlags) & HereditaryBits & RemovedBits) - | EvalBeforeAssigningBit - | EvalBeforeNestingBit, - - CoeffReadCost = Dynamic - }; - - typedef Sparse StorageKind; -}; - -} // end namespace internal - -template -class SparseSparseProduct : internal::no_assignment_operator, - public SparseMatrixBase > -{ - public: - - typedef SparseMatrixBase Base; - EIGEN_DENSE_PUBLIC_INTERFACE(SparseSparseProduct) - - private: - - typedef typename internal::traits::_LhsNested _LhsNested; - typedef typename internal::traits::_RhsNested _RhsNested; - - public: - - template - EIGEN_STRONG_INLINE SparseSparseProduct(const Lhs& lhs, const Rhs& rhs) - : m_lhs(lhs), m_rhs(rhs), m_tolerance(0), m_conservative(true) - { - init(); - } - - template - EIGEN_STRONG_INLINE SparseSparseProduct(const Lhs& lhs, const Rhs& rhs, const RealScalar& tolerance) - : m_lhs(lhs), m_rhs(rhs), m_tolerance(tolerance), m_conservative(false) - { - init(); - } - - SparseSparseProduct pruned(const Scalar& reference = 0, const RealScalar& epsilon = NumTraits::dummy_precision()) const - { - using std::abs; - return SparseSparseProduct(m_lhs,m_rhs,abs(reference)*epsilon); - } - - template - void evalTo(Dest& result) const - { - if(m_conservative) - internal::conservative_sparse_sparse_product_selector<_LhsNested, _RhsNested, Dest>::run(lhs(),rhs(),result); - else - internal::sparse_sparse_product_with_pruning_selector<_LhsNested, _RhsNested, Dest>::run(lhs(),rhs(),result,m_tolerance); - } - - EIGEN_STRONG_INLINE Index rows() const { return m_lhs.rows(); } - EIGEN_STRONG_INLINE Index cols() const { return m_rhs.cols(); } - - EIGEN_STRONG_INLINE const _LhsNested& lhs() const { return m_lhs; } - EIGEN_STRONG_INLINE const _RhsNested& rhs() const { return m_rhs; } - - protected: - void init() - { - eigen_assert(m_lhs.cols() == m_rhs.rows()); - - enum { - ProductIsValid = _LhsNested::ColsAtCompileTime==Dynamic - || _RhsNested::RowsAtCompileTime==Dynamic - || int(_LhsNested::ColsAtCompileTime)==int(_RhsNested::RowsAtCompileTime), - AreVectors = _LhsNested::IsVectorAtCompileTime && _RhsNested::IsVectorAtCompileTime, - SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(_LhsNested,_RhsNested) - }; - // note to the lost user: - // * for a dot product use: v1.dot(v2) - // * for a coeff-wise product use: v1.cwise()*v2 - EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes), - INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS) - EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors), - INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION) - EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT) - } - - LhsNested m_lhs; - RhsNested m_rhs; - RealScalar m_tolerance; - bool m_conservative; -}; - -// sparse = sparse * sparse -template -template -inline Derived& SparseMatrixBase::operator=(const SparseSparseProduct& product) -{ - product.evalTo(derived()); - return derived(); -} - -/** \returns an expression of the product of two sparse matrices. - * By default a conservative product preserving the symbolic non zeros is performed. - * The automatic pruning of the small values can be achieved by calling the pruned() function - * in which case a totally different product algorithm is employed: - * \code - * C = (A*B).pruned(); // supress numerical zeros (exact) - * C = (A*B).pruned(ref); - * C = (A*B).pruned(ref,epsilon); - * \endcode - * where \c ref is a meaningful non zero reference value. - * */ -template -template -inline const typename SparseSparseProductReturnType::Type -SparseMatrixBase::operator*(const SparseMatrixBase &other) const -{ - return typename SparseSparseProductReturnType::Type(derived(), other.derived()); -} - -#else // EIGEN_TEST_EVALUATORS - - /** \returns an expression of the product of two sparse matrices. * By default a conservative product preserving the symbolic non zeros is performed. * The automatic pruning of the small values can be achieved by calling the pruned() function @@ -256,8 +80,6 @@ protected: } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SPARSEPRODUCT_H diff --git a/Eigen/src/SparseCore/SparseRedux.h b/Eigen/src/SparseCore/SparseRedux.h index cf78d0e91..763f2296b 100644 --- a/Eigen/src/SparseCore/SparseRedux.h +++ b/Eigen/src/SparseCore/SparseRedux.h @@ -18,14 +18,9 @@ SparseMatrixBase::sum() const { eigen_assert(rows()>0 && cols()>0 && "you are using a non initialized matrix"); Scalar res(0); -#ifndef EIGEN_TEST_EVALUATORS - for (Index j=0; j::type thisEval(derived()); for (Index j=0; j::InnerIterator iter(thisEval,j); iter; ++iter) -#endif res += iter.value(); return res; } diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 4235d6c4c..69ac1a398 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -11,16 +11,6 @@ #define EIGEN_SPARSE_SELFADJOINTVIEW_H namespace Eigen { - -#ifndef EIGEN_TEST_EVALUATORS - -template -class SparseSelfAdjointTimeDenseProduct; - -template -class DenseTimeSparseSelfAdjointProduct; - -#endif // EIGEN_TEST_EVALUATORS /** \ingroup SparseCore_Module * \class SparseSelfAdjointView @@ -80,76 +70,40 @@ template class SparseSelfAdjointView * Note that there is no algorithmic advantage of performing such a product compared to a general sparse-sparse matrix product. * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product. */ -#ifndef EIGEN_TEST_EVALUATORS - template - SparseSparseProduct - operator*(const SparseMatrixBase& rhs) const - { - return SparseSparseProduct(*this, rhs.derived()); - } -#else template Product operator*(const SparseMatrixBase& rhs) const { return Product(*this, rhs.derived()); } -#endif /** \returns an expression of the matrix product between a sparse matrix \a lhs and a sparse self-adjoint matrix \a rhs. * * Note that there is no algorithmic advantage of performing such a product compared to a general sparse-sparse matrix product. * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product. */ -#ifndef EIGEN_TEST_EVALUATORS - template friend - SparseSparseProduct - operator*(const SparseMatrixBase& lhs, const SparseSelfAdjointView& rhs) - { - return SparseSparseProduct(lhs.derived(), rhs); - } -#else // EIGEN_TEST_EVALUATORS template friend Product operator*(const SparseMatrixBase& lhs, const SparseSelfAdjointView& rhs) { return Product(lhs.derived(), rhs); } -#endif // EIGEN_TEST_EVALUATORS /** Efficient sparse self-adjoint matrix times dense vector/matrix product */ -#ifndef EIGEN_TEST_EVALUATORS - template - SparseSelfAdjointTimeDenseProduct - operator*(const MatrixBase& rhs) const - { - return SparseSelfAdjointTimeDenseProduct(m_matrix, rhs.derived()); - } -#else template Product operator*(const MatrixBase& rhs) const { return Product(*this, rhs.derived()); } -#endif /** Efficient dense vector/matrix times sparse self-adjoint matrix product */ -#ifndef EIGEN_TEST_EVALUATORS - template friend - DenseTimeSparseSelfAdjointProduct - operator*(const MatrixBase& lhs, const SparseSelfAdjointView& rhs) - { - return DenseTimeSparseSelfAdjointProduct(lhs.derived(), rhs.m_matrix); - } -#else template friend Product operator*(const MatrixBase& lhs, const SparseSelfAdjointView& rhs) { return Product(lhs.derived(), rhs); } -#endif /** Perform a symmetric rank K update of the selfadjoint matrix \c *this: * \f$ this = this + \alpha ( u u^* ) \f$ where \a u is a vector or matrix. @@ -177,7 +131,7 @@ template class SparseSelfAdjointView } /** \returns an expression of P H P^-1 */ -// #ifndef EIGEN_TEST_EVALUATORS + // TODO implement twists in a more evaluator friendly fashion SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode> twistedBy(const PermutationMatrix& perm) const { return SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode>(m_matrix, perm); @@ -189,7 +143,6 @@ template class SparseSelfAdjointView permutedMatrix.evalTo(*this); return *this; } -// #endif // EIGEN_TEST_EVALUATORS SparseSelfAdjointView& operator=(const SparseSelfAdjointView& src) { @@ -251,98 +204,6 @@ SparseSelfAdjointView::rankUpdate(const SparseMatrixBase -struct traits > - : traits, Lhs, Rhs> > -{ - typedef Dense StorageKind; -}; -} - -template -class SparseSelfAdjointTimeDenseProduct - : public ProductBase, Lhs, Rhs> -{ - public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(SparseSelfAdjointTimeDenseProduct) - - SparseSelfAdjointTimeDenseProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) - {} - - template void scaleAndAddTo(Dest& dest, const Scalar& alpha) const - { - EIGEN_ONLY_USED_FOR_DEBUG(alpha); - // TODO use alpha - eigen_assert(alpha==Scalar(1) && "alpha != 1 is not implemented yet, sorry"); - typedef typename internal::remove_all::type _Lhs; - typedef typename _Lhs::InnerIterator LhsInnerIterator; - enum { - LhsIsRowMajor = (_Lhs::Flags&RowMajorBit)==RowMajorBit, - ProcessFirstHalf = - ((Mode&(Upper|Lower))==(Upper|Lower)) - || ( (Mode&Upper) && !LhsIsRowMajor) - || ( (Mode&Lower) && LhsIsRowMajor), - ProcessSecondHalf = !ProcessFirstHalf - }; - for (typename _Lhs::Index j=0; j -struct traits > - : traits, Lhs, Rhs> > -{}; -} - -template -class DenseTimeSparseSelfAdjointProduct - : public ProductBase, Lhs, Rhs> -{ - public: - EIGEN_PRODUCT_PUBLIC_INTERFACE(DenseTimeSparseSelfAdjointProduct) - - DenseTimeSparseSelfAdjointProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) - {} - - template void scaleAndAddTo(Dest& /*dest*/, const Scalar& /*alpha*/) const - { - // TODO - } - - private: - DenseTimeSparseSelfAdjointProduct& operator=(const DenseTimeSparseSelfAdjointProduct&); -}; - -#else // EIGEN_TEST_EVALUATORS - namespace internal { template @@ -486,8 +347,6 @@ protected: } // namespace internal -#endif // EIGEN_TEST_EVALUATORS - /*************************************************************************** * Implementation of symmetric copies and permutations ***************************************************************************/ @@ -644,7 +503,7 @@ void permute_symm_to_symm(const MatrixType& mat, SparseMatrix(this); } const Derived& derived() const { return *static_cast(this); } -#ifdef EIGEN_TEST_EVALUATORS /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A. * * \sa compute() @@ -91,17 +90,15 @@ class SparseSolverBase : internal::noncopyable eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b"); return Solve(derived(), b.derived()); } -#endif - - -#ifndef EIGEN_PARSED_BY_DOXYGEN + + #ifndef EIGEN_PARSED_BY_DOXYGEN /** \internal default implementation of solving with a sparse rhs */ template void _solve_impl(const SparseMatrixBase &b, SparseMatrixBase &dest) const { internal::solve_sparse_through_dense_panels(derived(), b.derived(), dest.derived()); } -#endif // EIGEN_PARSED_BY_DOXYGEN + #endif // EIGEN_PARSED_BY_DOXYGEN protected: diff --git a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h index b42b33e55..f291f8cef 100644 --- a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h +++ b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h @@ -47,10 +47,8 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r else res.resize(rows, cols); - #ifdef EIGEN_TEST_EVALUATORS typename evaluator::type lhsEval(lhs); typename evaluator::type rhsEval(rhs); - #endif res.reserve(estimated_nnz_prod); double ratioColRes = double(estimated_nnz_prod)/double(lhs.rows()*rhs.cols()); @@ -61,20 +59,12 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r // let's do a more accurate determination of the nnz ratio for the current column j of res tempVector.init(ratioColRes); tempVector.setZero(); -#ifndef EIGEN_TEST_EVALUATORS - for (typename Rhs::InnerIterator rhsIt(rhs, j); rhsIt; ++rhsIt) -#else for (typename evaluator::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt) -#endif { // FIXME should be written like this: tmp += rhsIt.value() * lhs.col(rhsIt.index()) tempVector.restart(); Scalar x = rhsIt.value(); -#ifndef EIGEN_TEST_EVALUATORS - for (typename Lhs::InnerIterator lhsIt(lhs, rhsIt.index()); lhsIt; ++lhsIt) -#else for (typename evaluator::InnerIterator lhsIt(lhsEval, rhsIt.index()); lhsIt; ++lhsIt) -#endif { tempVector.coeffRef(lhsIt.index()) += lhsIt.value() * x; } @@ -153,10 +143,6 @@ struct sparse_sparse_product_with_pruning_selector struct sparse_sparse_product_with_pruning_selector { @@ -204,7 +190,6 @@ struct sparse_sparse_product_with_pruning_selector(colLhs, rhs, res, tolerance); } }; -#endif } // end namespace internal diff --git a/Eigen/src/SparseCore/SparseTranspose.h b/Eigen/src/SparseCore/SparseTranspose.h index f5eff6133..fae7cae97 100644 --- a/Eigen/src/SparseCore/SparseTranspose.h +++ b/Eigen/src/SparseCore/SparseTranspose.h @@ -12,55 +12,6 @@ namespace Eigen { -#ifndef EIGEN_TEST_EVALUATORS -template class TransposeImpl - : public SparseMatrixBase > -{ - typedef typename internal::remove_all::type _MatrixTypeNested; - public: - - EIGEN_SPARSE_PUBLIC_INTERFACE(Transpose ) - - class InnerIterator; - class ReverseInnerIterator; - - inline Index nonZeros() const { return derived().nestedExpression().nonZeros(); } -}; - -// NOTE: VC10 trigger an ICE if don't put typename TransposeImpl:: in front of Index, -// a typedef typename TransposeImpl::Index Index; -// does not fix the issue. -// An alternative is to define the nested class in the parent class itself. -template class TransposeImpl::InnerIterator - : public _MatrixTypeNested::InnerIterator -{ - typedef typename _MatrixTypeNested::InnerIterator Base; - typedef typename TransposeImpl::Index Index; - public: - - EIGEN_STRONG_INLINE InnerIterator(const TransposeImpl& trans, typename TransposeImpl::Index outer) - : Base(trans.derived().nestedExpression(), outer) - {} - Index row() const { return Base::col(); } - Index col() const { return Base::row(); } -}; - -template class TransposeImpl::ReverseInnerIterator - : public _MatrixTypeNested::ReverseInnerIterator -{ - typedef typename _MatrixTypeNested::ReverseInnerIterator Base; - typedef typename TransposeImpl::Index Index; - public: - - EIGEN_STRONG_INLINE ReverseInnerIterator(const TransposeImpl& xpr, typename TransposeImpl::Index outer) - : Base(xpr.derived().nestedExpression(), outer) - {} - Index row() const { return Base::col(); } - Index col() const { return Base::row(); } -}; - -#else // EIGEN_TEST_EVALUATORS - // Implement nonZeros() for transpose. I'm not sure that's the best approach for that. // Perhaps it should be implemented in Transpose<> itself. template class TransposeImpl @@ -119,8 +70,6 @@ struct unary_evaluator, IteratorBased> } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - } // end namespace Eigen #endif // EIGEN_SPARSETRANSPOSE_H diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index 3df7a4cd4..744c3d730 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -40,11 +40,6 @@ protected: typedef typename internal::remove_reference::type MatrixTypeNestedNonRef; typedef typename internal::remove_all::type MatrixTypeNestedCleaned; -#ifndef EIGEN_TEST_EVALUATORS - template - typename internal::plain_matrix_type_column_major::type - solve(const MatrixBase& other) const; -#else // EIGEN_TEST_EVALUATORS template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const { @@ -52,7 +47,6 @@ protected: dst = rhs; this->solveInPlace(dst); } -#endif // EIGEN_TEST_EVALUATORS template void solveInPlace(MatrixBase& other) const; template void solveInPlace(SparseMatrixBase& other) const; @@ -163,7 +157,6 @@ class TriangularViewImpl::ReverseInnerIterator : public } }; -#ifdef EIGEN_TEST_EVALUATORS namespace internal { template @@ -270,7 +263,6 @@ protected: }; } // end namespace internal -#endif template template diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 28bf89bca..8de227b88 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -43,26 +43,6 @@ EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(Derived, -=) \ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, *=) \ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) -#ifndef EIGEN_TEST_EVALUATORS - -#define _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, BaseClass) \ - typedef BaseClass Base; \ - typedef typename Eigen::internal::traits::Scalar Scalar; \ - typedef typename Eigen::NumTraits::Real RealScalar; \ - typedef typename Eigen::internal::nested::type Nested; \ - typedef typename Eigen::internal::traits::StorageKind StorageKind; \ - typedef typename Eigen::internal::traits::Index Index; \ - enum { RowsAtCompileTime = Eigen::internal::traits::RowsAtCompileTime, \ - ColsAtCompileTime = Eigen::internal::traits::ColsAtCompileTime, \ - Flags = Eigen::internal::traits::Flags, \ - CoeffReadCost = Eigen::internal::traits::CoeffReadCost, \ - SizeAtCompileTime = Base::SizeAtCompileTime, \ - IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \ - using Base::derived; \ - using Base::const_cast_derived; - -#else // EIGEN_TEST_EVALUATORS - #define _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, BaseClass) \ typedef BaseClass Base; \ typedef typename Eigen::internal::traits::Scalar Scalar; \ @@ -78,8 +58,6 @@ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) using Base::derived; \ using Base::const_cast_derived; -#endif // EIGEN_TEST_EVALUATORS - #define EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) \ _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, Eigen::SparseMatrixBase) @@ -156,13 +134,11 @@ template struct plain_matrix_type typedef SparseMatrix<_Scalar, _Options, _Index> type; }; -#ifdef EIGEN_TEST_EVALUATORS template struct solve_traits { typedef typename sparse_eval::type PlainObject; }; -#endif template struct generic_xpr_base diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index 0f9aa9dd1..c9f9d61e9 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -422,30 +422,6 @@ class SparseVector::ReverseInnerIterator namespace internal { -#ifndef EIGEN_TEST_EVALUATORS -template< typename Dest, typename Src> -struct sparse_vector_assign_selector { - static void run(Dest& dst, const Src& src) { - eigen_internal_assert(src.innerSize()==src.size()); - for(typename Src::InnerIterator it(src, 0); it; ++it) - dst.insert(it.index()) = it.value(); - } -}; - -template< typename Dest, typename Src> -struct sparse_vector_assign_selector { - static void run(Dest& dst, const Src& src) { - eigen_internal_assert(src.outerSize()==src.size()); - for(typename Dest::Index i=0; i struct evaluator > : evaluator_base > @@ -492,7 +468,6 @@ struct sparse_vector_assign_selector { } } }; -#endif // EIGEN_TEST_EVALUATORS template< typename Dest, typename Src> struct sparse_vector_assign_selector { diff --git a/Eigen/src/SparseCore/SparseView.h b/Eigen/src/SparseCore/SparseView.h index c1c50f6e9..d10cc5a35 100644 --- a/Eigen/src/SparseCore/SparseView.h +++ b/Eigen/src/SparseCore/SparseView.h @@ -40,10 +40,6 @@ public: RealScalar m_epsilon = NumTraits::dummy_precision()) : m_matrix(mat), m_reference(m_reference), m_epsilon(m_epsilon) {} -#ifndef EIGEN_TEST_EVALUATORS - class InnerIterator; -#endif // EIGEN_TEST_EVALUATORS - inline Index rows() const { return m_matrix.rows(); } inline Index cols() const { return m_matrix.cols(); } @@ -63,43 +59,6 @@ protected: RealScalar m_epsilon; }; -#ifndef EIGEN_TEST_EVALUATORS -template -class SparseView::InnerIterator : public _MatrixTypeNested::InnerIterator -{ - typedef typename SparseView::Index Index; -public: - typedef typename _MatrixTypeNested::InnerIterator IterBase; - InnerIterator(const SparseView& view, Index outer) : - IterBase(view.m_matrix, outer), m_view(view) - { - incrementToNonZero(); - } - - EIGEN_STRONG_INLINE InnerIterator& operator++() - { - IterBase::operator++(); - incrementToNonZero(); - return *this; - } - - using IterBase::value; - -protected: - const SparseView& m_view; - -private: - void incrementToNonZero() - { - while((bool(*this)) && internal::isMuchSmallerThan(value(), m_view.reference(), m_view.epsilon())) - { - IterBase::operator++(); - } - } -}; - -#else // EIGEN_TEST_EVALUATORS - namespace internal { // TODO find a way to unify the two following variants @@ -230,8 +189,6 @@ struct unary_evaluator, IndexBased> } // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - template const SparseView MatrixBase::sparseView(const Scalar& reference, const typename NumTraits::Real& epsilon) const diff --git a/Eigen/src/SparseCore/TriangularSolver.h b/Eigen/src/SparseCore/TriangularSolver.h index 012a1bb75..98062e9c6 100644 --- a/Eigen/src/SparseCore/TriangularSolver.h +++ b/Eigen/src/SparseCore/TriangularSolver.h @@ -23,149 +23,6 @@ template::Flags) & RowMajorBit> struct sparse_solve_triangular_selector; -#ifndef EIGEN_TEST_EVALUATORS -// forward substitution, row-major -template -struct sparse_solve_triangular_selector -{ - typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::Index Index; - static void run(const Lhs& lhs, Rhs& other) - { - for(Index col=0 ; col -struct sparse_solve_triangular_selector -{ - typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::Index Index; - static void run(const Lhs& lhs, Rhs& other) - { - for(Index col=0 ; col=0 ; --i) - { - Scalar tmp = other.coeff(i,col); - Scalar l_ii = 0; - typename Lhs::InnerIterator it(lhs, i); - while(it && it.index() -struct sparse_solve_triangular_selector -{ - typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::Index Index; - static void run(const Lhs& lhs, Rhs& other) - { - for(Index col=0 ; col -struct sparse_solve_triangular_selector -{ - typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::Index Index; - static void run(const Lhs& lhs, Rhs& other) - { - for(Index col=0 ; col=0; --i) - { - Scalar& tmp = other.coeffRef(i,col); - if (tmp!=Scalar(0)) // optimization when other is actually sparse - { - if(!(Mode & UnitDiag)) - { - // TODO replace this by a binary search. make sure the binary search is safe for partially sorted elements - typename Lhs::ReverseInnerIterator it(lhs, i); - while(it && it.index()!=i) - --it; - eigen_assert(it && it.index()==i); - other.coeffRef(i,col) /= it.value(); - } - typename Lhs::InnerIterator it(lhs, i); - for(; it && it.index() struct sparse_solve_triangular_selector @@ -316,7 +173,6 @@ struct sparse_solve_triangular_selector } }; -#endif // EIGEN_TEST_EVALUATORS } // end namespace internal template @@ -338,18 +194,6 @@ void TriangularViewImpl::solveInPlace(MatrixBase -template -typename internal::plain_matrix_type_column_major::type -TriangularViewImpl::solve(const MatrixBase& other) const -{ - typename internal::plain_matrix_type_column_major::type res(other); - solveInPlace(res); - return res; -} -#endif - // pure sparse path namespace internal { diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index fb01f99cd..14d7e713e 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -173,36 +173,6 @@ class SparseLU : public SparseSolverBase >, m_diagpivotthresh = thresh; } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. - * - * \warning the destination matrix X in X = this->solve(B) must be colmun-major. - * - * \sa compute() - */ - template - inline const internal::solve_retval solve(const MatrixBase& B) const - { - eigen_assert(m_factorizationIsOk && "SparseLU is not initialized."); - eigen_assert(rows()==B.rows() - && "SparseLU::solve(): invalid number of rows of the right hand side matrix B"); - return internal::solve_retval(*this, B.derived()); - } - - /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval solve(const SparseMatrixBase& B) const - { - eigen_assert(m_factorizationIsOk && "SparseLU is not initialized."); - eigen_assert(rows()==B.rows() - && "SparseLU::solve(): invalid number of rows of the right hand side matrix B"); - return internal::sparse_solve_retval(*this, B.derived()); - } -#else - #ifdef EIGEN_PARSED_BY_DOXYGEN /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. * @@ -214,8 +184,6 @@ class SparseLU : public SparseSolverBase >, inline const Solve solve(const MatrixBase& B) const; #endif // EIGEN_PARSED_BY_DOXYGEN -#endif // EIGEN_TEST_EVALUATORS - /** \brief Reports whether previous computation was successful. * * \returns \c Success if computation was succesful, @@ -749,37 +717,6 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator const MatrixUType& m_mapU; }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef SparseLU<_MatrixType,Derived> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef SparseLU<_MatrixType,Derived> Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - this->defaultEvalTo(dst); - } -}; -} // end namespace internal -#endif - } // End namespace Eigen #endif diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 8e946b045..1a6c84e00 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -202,26 +202,6 @@ class SparseQR : public SparseSolverBase > m_threshold = threshold; } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval solve(const MatrixBase& B) const - { - eigen_assert(m_isInitialized && "The factorization should be called first, use compute()"); - eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix"); - return internal::solve_retval(*this, B.derived()); - } - template - inline const internal::sparse_solve_retval solve(const SparseMatrixBase& B) const - { - eigen_assert(m_isInitialized && "The factorization should be called first, use compute()"); - eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix"); - return internal::sparse_solve_retval(*this, B.derived()); - } -#else /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A. * * \sa compute() @@ -240,7 +220,6 @@ class SparseQR : public SparseSolverBase > eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix"); return Solve(*this, B.derived()); } -#endif // EIGEN_TEST_EVALUATORS /** \brief Reports whether previous computation was successful. * @@ -577,36 +556,6 @@ void SparseQR::factorize(const MatrixType& mat) m_info = Success; } -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef SparseQR<_MatrixType,OrderingType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef SparseQR<_MatrixType, OrderingType> Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec, Rhs) - - template void evalTo(Dest& dst) const - { - this->defaultEvalTo(dst); - } -}; -} // end namespace internal -#endif // EIGEN_TEST_EVALUATORS - template struct SparseQR_QProduct : ReturnByValue > { diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h index fcecd4fcf..0137585ca 100644 --- a/Eigen/src/SuperLUSupport/SuperLUSupport.h +++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h @@ -336,34 +336,6 @@ class SuperLUBase : public SparseSolverBase derived().analyzePattern(matrix); derived().factorize(matrix); } - -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "SuperLU is not initialized."); - eigen_assert(rows()==b.rows() - && "SuperLU::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } - - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval solve(const SparseMatrixBase& b) const - { - eigen_assert(m_isInitialized && "SuperLU is not initialized."); - eigen_assert(rows()==b.rows() - && "SuperLU::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); - } -#endif // EIGEN_TEST_EVALUATORS /** Performs a symbolic decomposition on the sparcity of \a matrix. * @@ -995,37 +967,6 @@ void SuperILU::_solve_impl(const MatrixBase &b, MatrixBase -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef SuperLUBase<_MatrixType,Derived> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec().derived()._solve_impl(rhs(),dst); - } -}; - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef SuperLUBase<_MatrixType,Derived> Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - this->defaultEvalTo(dst); - } -}; - -} // end namespace internal -#endif } // end namespace Eigen #endif // EIGEN_SUPERLUSUPPORT_H diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h index 845c8076a..7fada5567 100644 --- a/Eigen/src/UmfPackSupport/UmfPackSupport.h +++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h @@ -201,34 +201,6 @@ class UmfPackLU : public SparseSolverBase > analyzePattern(matrix); factorize(matrix); } - -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::solve_retval solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "UmfPackLU is not initialized."); - eigen_assert(rows()==b.rows() - && "UmfPackLU::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } - - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A. - * - * \sa compute() - */ - template - inline const internal::sparse_solve_retval solve(const SparseMatrixBase& b) const - { - eigen_assert(m_isInitialized && "UmfPackLU is not initialized."); - eigen_assert(rows()==b.rows() - && "UmfPackLU::solve(): invalid number of rows of the right hand side matrix b"); - return internal::sparse_solve_retval(*this, b.derived()); - } -#endif /** Performs a symbolic decomposition on the sparcity of \a matrix. * @@ -401,37 +373,6 @@ bool UmfPackLU::_solve_impl(const MatrixBase &b, MatrixBas return true; } -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef UmfPackLU<_MatrixType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -template -struct sparse_solve_retval, Rhs> - : sparse_solve_retval_base, Rhs> -{ - typedef UmfPackLU<_MatrixType> Dec; - EIGEN_MAKE_SPARSE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - this->defaultEvalTo(dst); - } -}; - -} // end namespace internal -#endif } // end namespace Eigen #endif // EIGEN_UMFPACKSUPPORT_H diff --git a/Eigen/src/misc/Solve.h b/Eigen/src/misc/Solve.h deleted file mode 100644 index ebdd981d0..000000000 --- a/Eigen/src/misc/Solve.h +++ /dev/null @@ -1,78 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_MISC_SOLVE_H -#define EIGEN_MISC_SOLVE_H - -#ifndef EIGEN_TEST_EVALUATORS - -namespace Eigen { - -namespace internal { - -/** \class solve_retval_base - * - */ -template -struct traits > -{ - typedef typename DecompositionType::MatrixType MatrixType; - typedef Matrix ReturnType; -}; - -template struct solve_retval_base - : public ReturnByValue > -{ - typedef typename remove_all::type RhsNestedCleaned; - typedef _DecompositionType DecompositionType; - typedef ReturnByValue Base; - typedef typename Base::Index Index; - - solve_retval_base(const DecompositionType& dec, const Rhs& rhs) - : m_dec(dec), m_rhs(rhs) - {} - - inline Index rows() const { return m_dec.cols(); } - inline Index cols() const { return m_rhs.cols(); } - inline const DecompositionType& dec() const { return m_dec; } - inline const RhsNestedCleaned& rhs() const { return m_rhs; } - - template inline void evalTo(Dest& dst) const - { - static_cast*>(this)->evalTo(dst); - } - - protected: - const DecompositionType& m_dec; - typename Rhs::Nested m_rhs; -}; - -} // end namespace internal - -#define EIGEN_MAKE_SOLVE_HELPERS(DecompositionType,Rhs) \ - typedef typename DecompositionType::MatrixType MatrixType; \ - typedef typename MatrixType::Scalar Scalar; \ - typedef typename MatrixType::RealScalar RealScalar; \ - typedef typename MatrixType::Index Index; \ - typedef Eigen::internal::solve_retval_base Base; \ - using Base::dec; \ - using Base::rhs; \ - using Base::rows; \ - using Base::cols; \ - solve_retval(const DecompositionType& dec, const Rhs& rhs) \ - : Base(dec, rhs) {} - -} // end namespace Eigen -#endif // EIGEN_TEST_EVALUATORS -#endif // EIGEN_MISC_SOLVE_H diff --git a/Eigen/src/misc/SparseSolve.h b/Eigen/src/misc/SparseSolve.h deleted file mode 100644 index 2396dc8e8..000000000 --- a/Eigen/src/misc/SparseSolve.h +++ /dev/null @@ -1,134 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2010 Gael Guennebaud -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#ifndef EIGEN_SPARSE_SOLVE_H -#define EIGEN_SPARSE_SOLVE_H - -#ifndef EIGEN_TEST_EVALUATORS - -namespace Eigen { - -namespace internal { - -template struct sparse_solve_retval_base; -template struct sparse_solve_retval; - -template -struct traits > -{ - typedef typename DecompositionType::MatrixType MatrixType; - typedef SparseMatrix ReturnType; -}; - -template struct sparse_solve_retval_base - : public ReturnByValue > -{ - typedef typename remove_all::type RhsNestedCleaned; - typedef _DecompositionType DecompositionType; - typedef ReturnByValue Base; - typedef typename Base::Index Index; - - sparse_solve_retval_base(const DecompositionType& dec, const Rhs& rhs) - : m_dec(dec), m_rhs(rhs) - {} - - inline Index rows() const { return m_dec.cols(); } - inline Index cols() const { return m_rhs.cols(); } - inline const DecompositionType& dec() const { return m_dec; } - inline const RhsNestedCleaned& rhs() const { return m_rhs; } - - template inline void evalTo(Dest& dst) const - { - static_cast*>(this)->evalTo(dst); - } - - protected: - template - inline void defaultEvalTo(SparseMatrix& dst) const - { - // we process the sparse rhs per block of NbColsAtOnce columns temporarily stored into a dense matrix. - static const int NbColsAtOnce = 4; - int rhsCols = m_rhs.cols(); - int size = m_rhs.rows(); - // the temporary matrices do not need more columns than NbColsAtOnce: - int tmpCols = (std::min)(rhsCols, NbColsAtOnce); - Eigen::Matrix tmp(size,tmpCols); - Eigen::Matrix tmpX(size,tmpCols); - for(int k=0; k(rhsCols-k, NbColsAtOnce); - tmp.leftCols(actualCols) = m_rhs.middleCols(k,actualCols); - tmpX.leftCols(actualCols) = m_dec.solve(tmp.leftCols(actualCols)); - dst.middleCols(k,actualCols) = tmpX.leftCols(actualCols).sparseView(); - } - } - const DecompositionType& m_dec; - typename Rhs::Nested m_rhs; -}; - -#define EIGEN_MAKE_SPARSE_SOLVE_HELPERS(DecompositionType,Rhs) \ - typedef typename DecompositionType::MatrixType MatrixType; \ - typedef typename MatrixType::Scalar Scalar; \ - typedef typename MatrixType::RealScalar RealScalar; \ - typedef typename MatrixType::Index Index; \ - typedef Eigen::internal::sparse_solve_retval_base Base; \ - using Base::dec; \ - using Base::rhs; \ - using Base::rows; \ - using Base::cols; \ - sparse_solve_retval(const DecompositionType& dec, const Rhs& rhs) \ - : Base(dec, rhs) {} - - - -template struct solve_retval_with_guess; - -template -struct traits > -{ - typedef typename DecompositionType::MatrixType MatrixType; - typedef Matrix ReturnType; -}; - -template struct solve_retval_with_guess - : public ReturnByValue > -{ - typedef typename DecompositionType::Index Index; - - solve_retval_with_guess(const DecompositionType& dec, const Rhs& rhs, const Guess& guess) - : m_dec(dec), m_rhs(rhs), m_guess(guess) - {} - - inline Index rows() const { return m_dec.cols(); } - inline Index cols() const { return m_rhs.cols(); } - - template inline void evalTo(Dest& dst) const - { - dst = m_guess; - m_dec._solveWithGuess(m_rhs,dst); - } - - protected: - const DecompositionType& m_dec; - const typename Rhs::Nested m_rhs; - const typename Guess::Nested m_guess; -}; - -} // namepsace internal - -} // end namespace Eigen - -#endif // EIGEN_TEST_EVALUATORS - -#endif // EIGEN_SPARSE_SOLVE_H diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 075b8d3de..530e9e4e1 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -139,11 +139,6 @@ endif(TEST_LIB) set_property(GLOBAL PROPERTY EIGEN_CURRENT_SUBPROJECT "Official") add_custom_target(BuildOfficial) -option(EIGEN_TEST_NO_EVALUATORS "Disable evaluators in unit tests" OFF) -if(EIGEN_TEST_NO_EVALUATORS) - add_definitions("-DEIGEN_TEST_NO_EVALUATORS=1") -endif(EIGEN_TEST_NO_EVALUATORS) - ei_add_test(meta) ei_add_test(sizeof) ei_add_test(dynalloc) diff --git a/test/evaluators.cpp b/test/evaluators.cpp index 2ca453b1c..f41968da8 100644 --- a/test/evaluators.cpp +++ b/test/evaluators.cpp @@ -1,16 +1,4 @@ -#ifndef EIGEN_ENABLE_EVALUATORS -#define EIGEN_ENABLE_EVALUATORS -#endif - -#ifdef EIGEN_TEST_EVALUATORS -#undef EIGEN_TEST_EVALUATORS -#endif - -#ifdef EIGEN_TEST_NO_EVALUATORS -#undef EIGEN_TEST_NO_EVALUATORS -#endif - #include "main.h" namespace Eigen { diff --git a/test/mixingtypes.cpp b/test/mixingtypes.cpp index 976e21e37..048f7255a 100644 --- a/test/mixingtypes.cpp +++ b/test/mixingtypes.cpp @@ -53,12 +53,10 @@ template void mixingtypes(int size = SizeAtCompileType) mf+mf; VERIFY_RAISES_ASSERT(mf+md); VERIFY_RAISES_ASSERT(mf+mcf); -#ifndef EIGEN_TEST_EVALUATORS - // they do not even compile when using evaluators - VERIFY_RAISES_ASSERT(vf=vd); - VERIFY_RAISES_ASSERT(vf+=vd); - VERIFY_RAISES_ASSERT(mcd=md); -#endif + // the following do not even compile since the introduction of evaluators +// VERIFY_RAISES_ASSERT(vf=vd); +// VERIFY_RAISES_ASSERT(vf+=vd); +// VERIFY_RAISES_ASSERT(mcd=md); // check scalar products VERIFY_IS_APPROX(vcf * sf , vcf * complex(sf)); diff --git a/test/nesting_ops.cpp b/test/nesting_ops.cpp index 114dd5e41..6e772c70f 100644 --- a/test/nesting_ops.cpp +++ b/test/nesting_ops.cpp @@ -11,12 +11,7 @@ template void run_nesting_ops(const MatrixType& _m) { -#ifndef EIGEN_TEST_EVALUATORS - // TODO, with evaluator, the following is not correct anymore: - typename MatrixType::Nested m(_m); -#else typename internal::nested_eval::type m(_m); -#endif // Make really sure that we are in debug mode! VERIFY_RAISES_ASSERT(eigen_assert(false)); diff --git a/test/sparse_vector.cpp b/test/sparse_vector.cpp index 6cd5a9a8c..5eea9edfd 100644 --- a/test/sparse_vector.cpp +++ b/test/sparse_vector.cpp @@ -71,10 +71,7 @@ template void sparse_vector(int rows, int cols) VERIFY_IS_APPROX(v1.dot(v2), refV1.dot(refV2)); VERIFY_IS_APPROX(v1.dot(refV2), refV1.dot(refV2)); -#ifdef EIGEN_TEST_EVALUATORS - // the following did not compiled without evaluators VERIFY_IS_APPROX(m1*v2, refM1*refV2); -#endif VERIFY_IS_APPROX(v1.dot(m1*v2), refV1.dot(refM1*refV2)); int i = internal::random(0,rows-1); VERIFY_IS_APPROX(v1.dot(m1.col(i)), refV1.dot(refM1.col(i))); diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp index 303eb6cf0..2f839cf51 100644 --- a/test/vectorization_logic.cpp +++ b/test/vectorization_logic.cpp @@ -45,22 +45,14 @@ std::string demangle_flags(int f) template bool test_assign(const Dst&, const Src&, int traversal, int unrolling) { -#ifdef EIGEN_TEST_EVALUATORS typedef internal::copy_using_evaluator_traits,internal::evaluator, internal::assign_op > traits; -#else - typedef internal::assign_traits traits; -#endif bool res = traits::Traversal==traversal && traits::Unrolling==unrolling; if(!res) { std::cerr << "Src: " << demangle_flags(Src::Flags) << std::endl; -#ifdef EIGEN_TEST_EVALUATORS std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; -#endif std::cerr << "Dst: " << demangle_flags(Dst::Flags) << std::endl; -#ifdef EIGEN_TEST_EVALUATORS std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; -#endif traits::debug(); std::cerr << " Expected Traversal == " << demangle_traversal(traversal) << " got " << demangle_traversal(traits::Traversal) << "\n"; @@ -73,22 +65,14 @@ bool test_assign(const Dst&, const Src&, int traversal, int unrolling) template bool test_assign(int traversal, int unrolling) { -#ifdef EIGEN_TEST_EVALUATORS typedef internal::copy_using_evaluator_traits,internal::evaluator, internal::assign_op > traits; -#else - typedef internal::assign_traits traits; -#endif bool res = traits::Traversal==traversal && traits::Unrolling==unrolling; if(!res) { std::cerr << "Src: " << demangle_flags(Src::Flags) << std::endl; -#ifdef EIGEN_TEST_EVALUATORS std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; -#endif std::cerr << "Dst: " << demangle_flags(Dst::Flags) << std::endl; -#ifdef EIGEN_TEST_EVALUATORS std::cerr << " " << demangle_flags(internal::evaluator::Flags) << std::endl; -#endif traits::debug(); std::cerr << " Expected Traversal == " << demangle_traversal(traversal) << " got " << demangle_traversal(traits::Traversal) << "\n"; @@ -101,19 +85,13 @@ bool test_assign(int traversal, int unrolling) template bool test_redux(const Xpr&, int traversal, int unrolling) { -#ifdef EIGEN_TEST_EVALUATORS typedef internal::redux_traits,internal::redux_evaluator > traits; -#else - typedef internal::redux_traits,Xpr> traits; -#endif bool res = traits::Traversal==traversal && traits::Unrolling==unrolling; if(!res) { std::cerr << demangle_flags(Xpr::Flags) << std::endl; -#ifdef EIGEN_TEST_EVALUATORS std::cerr << demangle_flags(internal::evaluator::Flags) << std::endl; -#endif traits::debug(); std::cerr << " Expected Traversal == " << demangle_traversal(traversal) diff --git a/unsupported/Eigen/IterativeSolvers b/unsupported/Eigen/IterativeSolvers index aa15403db..ff0d59b6e 100644 --- a/unsupported/Eigen/IterativeSolvers +++ b/unsupported/Eigen/IterativeSolvers @@ -24,9 +24,6 @@ */ //@{ -#include "../../Eigen/src/misc/Solve.h" -#include "../../Eigen/src/misc/SparseSolve.h" - #ifndef EIGEN_MPL2_ONLY #include "src/IterativeSolvers/IterationController.h" #include "src/IterativeSolvers/ConstrainedConjGrad.h" diff --git a/unsupported/Eigen/SparseExtra b/unsupported/Eigen/SparseExtra index b5597902a..819cffa27 100644 --- a/unsupported/Eigen/SparseExtra +++ b/unsupported/Eigen/SparseExtra @@ -37,9 +37,6 @@ */ -#include "../../Eigen/src/misc/Solve.h" -#include "../../Eigen/src/misc/SparseSolve.h" - #include "src/SparseExtra/DynamicSparseMatrix.h" #include "src/SparseExtra/BlockOfDynamicSparseMatrix.h" #include "src/SparseExtra/RandomSetter.h" diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h index fe0bfd948..0e1b7d977 100644 --- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h @@ -139,24 +139,6 @@ class DGMRES : public IterativeSolverBase > ~DGMRES() {} -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A - * \a x0 as an initial solution. - * - * \sa compute() - */ - template - inline const internal::solve_retval_with_guess - solveWithGuess(const MatrixBase& b, const Guess& x0) const - { - eigen_assert(m_isInitialized && "DGMRES is not initialized."); - eigen_assert(Base::rows()==b.rows() - && "DGMRES::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval_with_guess - (*this, b.derived(), x0); - } -#endif - /** \internal */ template void _solve_with_guess_impl(const Rhs& b, Dest& x) const @@ -525,23 +507,5 @@ int DGMRES<_MatrixType, _Preconditioner>::dgmresApplyDeflation(const RhsType &x, return 0; } -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - - template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef DGMRES<_MatrixType, _Preconditioner> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; -} // end namespace internal -#endif - } // end namespace Eigen #endif diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index fd76a9d2c..cd15ce0bf 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -316,24 +316,6 @@ public: */ void set_restart(const int restart) { m_restart=restart; } -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A - * \a x0 as an initial solution. - * - * \sa compute() - */ - template - inline const internal::solve_retval_with_guess - solveWithGuess(const MatrixBase& b, const Guess& x0) const - { - eigen_assert(m_isInitialized && "GMRES is not initialized."); - eigen_assert(Base::rows()==b.rows() - && "GMRES::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval_with_guess - (*this, b.derived(), x0); - } -#endif - /** \internal */ template void _solve_with_guess_impl(const Rhs& b, Dest& x) const @@ -367,24 +349,6 @@ protected: }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - - template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef GMRES<_MatrixType, _Preconditioner> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} // end namespace internal -#endif } // end namespace Eigen #endif // EIGEN_GMRES_H diff --git a/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h b/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h index 1ee1c89b2..dd43de6b3 100644 --- a/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h +++ b/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h @@ -106,18 +106,6 @@ class IncompleteCholesky : public SparseSolverBase inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_factorizationIsOk && "IncompleteLLT did not succeed"); - eigen_assert(m_isInitialized && "IncompleteLLT is not initialized."); - eigen_assert(cols()==b.rows() - && "IncompleteLLT::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } -#endif protected: SparseMatrix m_L; // The lower part stored in CSC @@ -263,25 +251,6 @@ inline void IncompleteCholesky::updateList(const Idx } } -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef IncompleteCholesky<_Scalar, _UpLo, OrderingType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve(rhs(),dst); - } -}; - -} // end namespace internal -#endif - } // end namespace Eigen #endif diff --git a/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h b/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h index e86f65644..7d08c3515 100644 --- a/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h +++ b/unsupported/Eigen/src/IterativeSolvers/IncompleteLU.h @@ -81,40 +81,10 @@ class IncompleteLU : public SparseSolverBase > x = m_lu.template triangularView().solve(x); } -#ifndef EIGEN_TEST_EVALUATORS - template inline const internal::solve_retval - solve(const MatrixBase& b) const - { - eigen_assert(m_isInitialized && "IncompleteLU is not initialized."); - eigen_assert(cols()==b.rows() - && "IncompleteLU::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval(*this, b.derived()); - } -#endif - protected: FactorType m_lu; }; -#ifndef EIGEN_TEST_EVALUATORS -namespace internal { - -template -struct solve_retval, Rhs> - : solve_retval_base, Rhs> -{ - typedef IncompleteLU<_MatrixType> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } -}; - -} // end namespace internal -#endif - } // end namespace Eigen #endif // EIGEN_INCOMPLETE_LU_H diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index 28d5c692d..aaf42c78a 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -246,24 +246,6 @@ namespace Eigen { /** Destructor. */ ~MINRES(){} -#ifndef EIGEN_TEST_EVALUATORS - /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A - * \a x0 as an initial solution. - * - * \sa compute() - */ - template - inline const internal::solve_retval_with_guess - solveWithGuess(const MatrixBase& b, const Guess& x0) const - { - eigen_assert(m_isInitialized && "MINRES is not initialized."); - eigen_assert(Base::rows()==b.rows() - && "MINRES::solve(): invalid number of rows of the right hand side matrix b"); - return internal::solve_retval_with_guess - (*this, b.derived(), x0); - } -#endif - /** \internal */ template void _solve_with_guess_impl(const Rhs& b, Dest& x) const @@ -296,25 +278,6 @@ namespace Eigen { protected: }; - -#ifndef EIGEN_TEST_EVALUATORS - namespace internal { - - template - struct solve_retval, Rhs> - : solve_retval_base, Rhs> - { - typedef MINRES<_MatrixType,_UpLo,_Preconditioner> Dec; - EIGEN_MAKE_SOLVE_HELPERS(Dec,Rhs) - - template void evalTo(Dest& dst) const - { - dec()._solve_impl(rhs(),dst); - } - }; - - } // end namespace internal -#endif } // end namespace Eigen diff --git a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h index 4210df68a..e4dc1c1de 100644 --- a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h +++ b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h @@ -352,7 +352,6 @@ class DynamicSparseMatrix::ReverseInnerIterator : public const Index m_outer; }; -#ifdef EIGEN_ENABLE_EVALUATORS namespace internal { template @@ -382,7 +381,6 @@ struct evaluator > }; } -#endif } // end namespace Eigen diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 970a05bbd..e45efbd39 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -6,10 +6,6 @@ include_directories(../../test ../../unsupported ../../Eigen ${CMAKE_CURRENT_BINARY_DIR}/../../test) -if(EIGEN_TEST_NO_EVALUATORS) - add_definitions("-DEIGEN_TEST_NO_EVALUATORS=1") -endif(EIGEN_TEST_NO_EVALUATORS) - find_package(GoogleHash) if(GOOGLEHASH_FOUND) add_definitions("-DEIGEN_GOOGLEHASH_SUPPORT") @@ -33,22 +29,18 @@ endif(ADOLC_FOUND) ei_add_test(NonLinearOptimization) ei_add_test(NumericalDiff) -if(EIGEN_TEST_NO_EVALUATORS) -ei_add_test(autodiff_scalar) -ei_add_test(autodiff) -endif() +# TODO ei_add_test(autodiff_scalar) +# TODO ei_add_test(autodiff) if (NOT CMAKE_CXX_COMPILER MATCHES "clang\\+\\+$") ei_add_test(BVH) endif() -if(EIGEN_TEST_NO_EVALUATORS) -ei_add_test(matrix_exponential) -ei_add_test(matrix_function) -ei_add_test(matrix_power) -ei_add_test(matrix_square_root) -ei_add_test(alignedvector3) -endif() +# TODO ei_add_test(matrix_exponential) +# TODO ei_add_test(matrix_function) +# TODO ei_add_test(matrix_power) +# TODO ei_add_test(matrix_square_root) +# TODO ei_add_test(alignedvector3) ei_add_test(FFT) @@ -101,9 +93,7 @@ ei_add_test(gmres) ei_add_test(minres) ei_add_test(levenberg_marquardt) ei_add_test(bdcsvd) -if(EIGEN_TEST_NO_EVALUATORS) -ei_add_test(kronecker_product) -endif() +# TODO ei_add_test(kronecker_product) option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." OFF) if(EIGEN_TEST_CXX11) From 060e835ee9f7a3abb9ff3d9f2522027bf2f98efe Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 18 Sep 2014 17:30:21 +0200 Subject: [PATCH 0798/1103] Add evaluator for the experimental AlignedVector3 --- unsupported/Eigen/AlignedVector3 | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+) diff --git a/unsupported/Eigen/AlignedVector3 b/unsupported/Eigen/AlignedVector3 index 7b45e6cce..cd5f49edb 100644 --- a/unsupported/Eigen/AlignedVector3 +++ b/unsupported/Eigen/AlignedVector3 @@ -57,6 +57,9 @@ template class AlignedVector3 inline Index rows() const { return 3; } inline Index cols() const { return 1; } + + Scalar* data() { return m_coeffs.data(); } + const Scalar* data() const { return m_coeffs.data(); } inline const Scalar& coeff(Index row, Index col) const { return m_coeffs.coeff(row, col); } @@ -183,6 +186,29 @@ template class AlignedVector3 } }; +namespace internal { + +template +struct evaluator > + : evaluator,Aligned> >::type +{ + typedef AlignedVector3 XprType; + typedef Map,Aligned> MapType; + typedef typename evaluator::type Base; + + typedef evaluator type; + typedef evaluator nestedType; + + evaluator(const XprType &m) : Base(MapType(m.data())), m_map(m.data()) + { + ::new (static_cast(this)) Base(m_map); + } + + const MapType m_map; +}; + +} + //@} } From 62bce6e5e6da71dd8d85ae229d24b9f9f13d1681 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 18 Sep 2014 17:31:17 +0200 Subject: [PATCH 0799/1103] Make MatrixFunction use nested_eval instead of nested --- .../Eigen/src/MatrixFunctions/MatrixExponential.h | 5 +++-- .../Eigen/src/MatrixFunctions/MatrixFunction.h | 11 ++++++----- .../Eigen/src/MatrixFunctions/MatrixLogarithm.h | 11 ++++++----- .../Eigen/src/MatrixFunctions/MatrixSquareRoot.h | 8 +++++--- unsupported/test/CMakeLists.txt | 10 +++++----- 5 files changed, 25 insertions(+), 20 deletions(-) diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h index 160120d03..9e0545660 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixExponential.h @@ -392,14 +392,15 @@ template struct MatrixExponentialReturnValue template inline void evalTo(ResultType& result) const { - internal::matrix_exp_compute(m_src, result); + const typename internal::nested_eval::type tmp(m_src); + internal::matrix_exp_compute(tmp, result); } Index rows() const { return m_src.rows(); } Index cols() const { return m_src.cols(); } protected: - const typename internal::nested::type m_src; + const typename internal::nested::type m_src; }; namespace internal { diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h index a35c11be5..b68aae5e8 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixFunction.h @@ -485,7 +485,7 @@ template class MatrixFunctionReturnValue typedef typename internal::stem_function::type StemFunction; protected: - typedef typename internal::nested::type DerivedNested; + typedef typename internal::nested::type DerivedNested; public: @@ -503,18 +503,19 @@ template class MatrixFunctionReturnValue template inline void evalTo(ResultType& result) const { - typedef typename internal::remove_all::type DerivedNestedClean; - typedef internal::traits Traits; + typedef typename internal::nested_eval::type NestedEvalType; + typedef typename internal::remove_all::type NestedEvalTypeClean; + typedef internal::traits Traits; static const int RowsAtCompileTime = Traits::RowsAtCompileTime; static const int ColsAtCompileTime = Traits::ColsAtCompileTime; - static const int Options = DerivedNestedClean::Options; + static const int Options = NestedEvalTypeClean::Options; typedef std::complex::Real> ComplexScalar; typedef Matrix DynMatrixType; typedef internal::MatrixFunctionAtomic AtomicType; AtomicType atomic(m_f); - internal::matrix_function_compute::run(m_A, atomic, result); + internal::matrix_function_compute::run(m_A, atomic, result); } Index rows() const { return m_A.rows(); } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h index d46ccc145..42b60b9b1 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixLogarithm.h @@ -310,7 +310,7 @@ public: typedef typename Derived::Index Index; protected: - typedef typename internal::nested::type DerivedNested; + typedef typename internal::nested::type DerivedNested; public: @@ -327,17 +327,18 @@ public: template inline void evalTo(ResultType& result) const { - typedef typename internal::remove_all::type DerivedNestedClean; - typedef internal::traits Traits; + typedef typename internal::nested_eval::type DerivedEvalType; + typedef typename internal::remove_all::type DerivedEvalTypeClean; + typedef internal::traits Traits; static const int RowsAtCompileTime = Traits::RowsAtCompileTime; static const int ColsAtCompileTime = Traits::ColsAtCompileTime; - static const int Options = DerivedNestedClean::Options; + static const int Options = DerivedEvalTypeClean::Options; typedef std::complex::Real> ComplexScalar; typedef Matrix DynMatrixType; typedef internal::MatrixLogarithmAtomic AtomicType; AtomicType atomic; - internal::matrix_function_compute::run(m_A, atomic, result); + internal::matrix_function_compute::run(m_A, atomic, result); } Index rows() const { return m_A.rows(); } diff --git a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h index 8ca4f4864..3a4d6eb3f 100644 --- a/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h +++ b/unsupported/Eigen/src/MatrixFunctions/MatrixSquareRoot.h @@ -320,7 +320,7 @@ template class MatrixSquareRootReturnValue { protected: typedef typename Derived::Index Index; - typedef typename internal::nested::type DerivedNested; + typedef typename internal::nested::type DerivedNested; public: /** \brief Constructor. @@ -338,8 +338,10 @@ template class MatrixSquareRootReturnValue template inline void evalTo(ResultType& result) const { - typedef typename internal::remove_all::type DerivedNestedClean; - internal::matrix_sqrt_compute::run(m_src, result); + typedef typename internal::nested_eval::type DerivedEvalType; + typedef typename internal::remove_all::type DerivedEvalTypeClean; + DerivedEvalType tmp(m_src); + internal::matrix_sqrt_compute::run(tmp, result); } Index rows() const { return m_src.rows(); } diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index e45efbd39..7e2186d2d 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -36,11 +36,11 @@ if (NOT CMAKE_CXX_COMPILER MATCHES "clang\\+\\+$") ei_add_test(BVH) endif() -# TODO ei_add_test(matrix_exponential) -# TODO ei_add_test(matrix_function) -# TODO ei_add_test(matrix_power) -# TODO ei_add_test(matrix_square_root) -# TODO ei_add_test(alignedvector3) +ei_add_test(matrix_exponential) +ei_add_test(matrix_function) +ei_add_test(matrix_power) +ei_add_test(matrix_square_root) +ei_add_test(alignedvector3) ei_add_test(FFT) From 2ae20d558b33653b8ef2fe17255ed171997bcf79 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 18 Sep 2014 22:08:49 +0200 Subject: [PATCH 0800/1103] Update KroneckerProduct wrt evaluator changes --- .../KroneckerProduct/KroneckerTensorProduct.h | 36 +++++++++++++++---- unsupported/test/CMakeLists.txt | 2 +- 2 files changed, 31 insertions(+), 7 deletions(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index ca66d4d89..72e25db19 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -154,16 +154,41 @@ void KroneckerProductSparse::evalTo(Dest& dst) const dst.resize(this->rows(), this->cols()); dst.resizeNonZeros(0); + // 1 - evaluate the operands if needed: + typedef typename internal::nested_eval::type Lhs1; + typedef typename internal::remove_all::type Lhs1Cleaned; + const Lhs1 lhs1(m_A); + typedef typename internal::nested_eval::type Rhs1; + typedef typename internal::remove_all::type Rhs1Cleaned; + const Rhs1 rhs1(m_B); + + // 2 - construct a SparseView for dense operands + typedef typename internal::conditional::StorageKind,Sparse>::value, Lhs1, SparseView >::type Lhs2; + typedef typename internal::remove_all::type Lhs2Cleaned; + const Lhs2 lhs2(lhs1); + typedef typename internal::conditional::StorageKind,Sparse>::value, Rhs1, SparseView >::type Rhs2; + typedef typename internal::remove_all::type Rhs2Cleaned; + const Rhs2 rhs2(rhs1); + + // 3 - construct respective evaluators + typedef typename internal::evaluator::type LhsEval; + LhsEval lhsEval(lhs2); + typedef typename internal::evaluator::type RhsEval; + RhsEval rhsEval(rhs2); + + typedef typename LhsEval::InnerIterator LhsInnerIterator; + typedef typename RhsEval::InnerIterator RhsInnerIterator; + // compute number of non-zeros per innervectors of dst { VectorXi nnzA = VectorXi::Zero(Dest::IsRowMajor ? m_A.rows() : m_A.cols()); for (Index kA=0; kA < m_A.outerSize(); ++kA) - for (typename Lhs::InnerIterator itA(m_A,kA); itA; ++itA) + for (LhsInnerIterator itA(lhsEval,kA); itA; ++itA) nnzA(Dest::IsRowMajor ? itA.row() : itA.col())++; VectorXi nnzB = VectorXi::Zero(Dest::IsRowMajor ? m_B.rows() : m_B.cols()); for (Index kB=0; kB < m_B.outerSize(); ++kB) - for (typename Rhs::InnerIterator itB(m_B,kB); itB; ++itB) + for (RhsInnerIterator itB(rhsEval,kB); itB; ++itB) nnzB(Dest::IsRowMajor ? itB.row() : itB.col())++; Matrix nnzAB = nnzB * nnzA.transpose(); @@ -174,9 +199,9 @@ void KroneckerProductSparse::evalTo(Dest& dst) const { for (Index kB=0; kB < m_B.outerSize(); ++kB) { - for (typename Lhs::InnerIterator itA(m_A,kA); itA; ++itA) + for (LhsInnerIterator itA(lhsEval,kA); itA; ++itA) { - for (typename Rhs::InnerIterator itB(m_B,kB); itB; ++itB) + for (RhsInnerIterator itB(rhsEval,kB); itB; ++itB) { const Index i = itA.row() * Br + itB.row(), j = itA.col() * Bc + itB.col(); @@ -201,8 +226,7 @@ struct traits > Rows = size_at_compile_time::RowsAtCompileTime, traits::RowsAtCompileTime>::ret, Cols = size_at_compile_time::ColsAtCompileTime, traits::ColsAtCompileTime>::ret, MaxRows = size_at_compile_time::MaxRowsAtCompileTime, traits::MaxRowsAtCompileTime>::ret, - MaxCols = size_at_compile_time::MaxColsAtCompileTime, traits::MaxColsAtCompileTime>::ret, - CoeffReadCost = Lhs::CoeffReadCost + Rhs::CoeffReadCost + NumTraits::MulCost + MaxCols = size_at_compile_time::MaxColsAtCompileTime, traits::MaxColsAtCompileTime>::ret }; typedef Matrix ReturnType; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 7e2186d2d..04863a9ad 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -93,7 +93,7 @@ ei_add_test(gmres) ei_add_test(minres) ei_add_test(levenberg_marquardt) ei_add_test(bdcsvd) -# TODO ei_add_test(kronecker_product) +ei_add_test(kronecker_product) option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." OFF) if(EIGEN_TEST_CXX11) From e70506dd8f63241d92ebc3df2d9c01c9af5564c8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 18 Sep 2014 22:46:46 +0200 Subject: [PATCH 0801/1103] Fix inner-stride of AlignedVector3 --- unsupported/Eigen/AlignedVector3 | 16 +++++++--------- unsupported/test/CMakeLists.txt | 4 ++-- 2 files changed, 9 insertions(+), 11 deletions(-) diff --git a/unsupported/Eigen/AlignedVector3 b/unsupported/Eigen/AlignedVector3 index cd5f49edb..35493e87b 100644 --- a/unsupported/Eigen/AlignedVector3 +++ b/unsupported/Eigen/AlignedVector3 @@ -60,6 +60,7 @@ template class AlignedVector3 Scalar* data() { return m_coeffs.data(); } const Scalar* data() const { return m_coeffs.data(); } + Index innerStride() const { return 1; } inline const Scalar& coeff(Index row, Index col) const { return m_coeffs.coeff(row, col); } @@ -184,27 +185,24 @@ template class AlignedVector3 { return m_coeffs.template head<3>().isApprox(other,eps); } + + CoeffType& coeffs() { return m_coeffs; } + const CoeffType& coeffs() const { return m_coeffs; } }; namespace internal { template struct evaluator > - : evaluator,Aligned> >::type + : evaluator >::type { typedef AlignedVector3 XprType; - typedef Map,Aligned> MapType; - typedef typename evaluator::type Base; + typedef typename evaluator >::type Base; typedef evaluator type; typedef evaluator nestedType; - evaluator(const XprType &m) : Base(MapType(m.data())), m_map(m.data()) - { - ::new (static_cast(this)) Base(m_map); - } - - const MapType m_map; + evaluator(const XprType &m) : Base(m.coeffs()) {} }; } diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 04863a9ad..48b61cde0 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -29,8 +29,8 @@ endif(ADOLC_FOUND) ei_add_test(NonLinearOptimization) ei_add_test(NumericalDiff) -# TODO ei_add_test(autodiff_scalar) -# TODO ei_add_test(autodiff) +ei_add_test(autodiff_scalar) +ei_add_test(autodiff) if (NOT CMAKE_CXX_COMPILER MATCHES "clang\\+\\+$") ei_add_test(BVH) From 07c5500d709eb7914bd46a1c4b3629bf42783f1d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 19 Sep 2014 09:58:20 +0200 Subject: [PATCH 0802/1103] Introduce a compilation error when using the wrong InnerIterator type. --- Eigen/src/SparseCore/SparseMatrix.h | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 72368ebf3..9e7124ff2 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -887,6 +887,11 @@ class SparseMatrix::InnerIterator const Index m_outer; Index m_id; Index m_end; + private: + // If you get here, then you're not using the right InnerIterator type, e.g.: + // SparseMatrix A; + // SparseMatrix::InnerIterator it(A,0); + template InnerIterator(const SparseMatrixBase&,Index outer); }; template From 755e77266f7c1005688654f7e7b070894c43bd18 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 19 Sep 2014 09:58:56 +0200 Subject: [PATCH 0803/1103] Fix SparseQR for row-major inputs. --- Eigen/src/SparseQR/SparseQR.h | 38 +++++++++++++++++++++++++---------- 1 file changed, 27 insertions(+), 11 deletions(-) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 1a6c84e00..6d85ea9be 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -284,9 +284,11 @@ template void SparseQR::analyzePattern(const MatrixType& mat) { eigen_assert(mat.isCompressed() && "SparseQR requires a sparse matrix in compressed mode. Call .makeCompressed() before passing it to SparseQR"); + // Copy to a column major matrix if the input is rowmajor + typename internal::conditional::type matCpy(mat); // Compute the column fill reducing ordering OrderingType ord; - ord(mat, m_perm_c); + ord(matCpy, m_perm_c); Index n = mat.cols(); Index m = mat.rows(); Index diagSize = (std::min)(m,n); @@ -299,7 +301,7 @@ void SparseQR::analyzePattern(const MatrixType& mat) // Compute the column elimination tree of the permuted matrix m_outputPerm_c = m_perm_c.inverse(); - internal::coletree(mat, m_etree, m_firstRowElt, m_outputPerm_c.indices().data()); + internal::coletree(matCpy, m_etree, m_firstRowElt, m_outputPerm_c.indices().data()); m_isEtreeOk = true; m_R.resize(m, n); @@ -337,21 +339,35 @@ void SparseQR::factorize(const MatrixType& mat) m_R.setZero(); m_Q.setZero(); + m_pmat = mat; if(!m_isEtreeOk) { m_outputPerm_c = m_perm_c.inverse(); - internal::coletree(mat, m_etree, m_firstRowElt, m_outputPerm_c.indices().data()); + internal::coletree(m_pmat, m_etree, m_firstRowElt, m_outputPerm_c.indices().data()); m_isEtreeOk = true; } - - m_pmat = mat; + m_pmat.uncompress(); // To have the innerNonZeroPtr allocated + // Apply the fill-in reducing permutation lazily: - for (int i = 0; i < n; i++) { - Index p = m_perm_c.size() ? m_perm_c.indices()(i) : i; - m_pmat.outerIndexPtr()[p] = mat.outerIndexPtr()[i]; - m_pmat.innerNonZeroPtr()[p] = mat.outerIndexPtr()[i+1] - mat.outerIndexPtr()[i]; + // If the input is row major, copy the original column indices, + // otherwise directly use the input matrix + // + IndexVector originalOuterIndicesCpy; + const Index *originalOuterIndices = mat.outerIndexPtr(); + if(MatrixType::IsRowMajor) + { + originalOuterIndicesCpy = IndexVector::Map(m_pmat.outerIndexPtr(),n+1); + originalOuterIndices = originalOuterIndicesCpy.data(); + } + + for (int i = 0; i < n; i++) + { + Index p = m_perm_c.size() ? m_perm_c.indices()(i) : i; + m_pmat.outerIndexPtr()[p] = originalOuterIndices[i]; + m_pmat.innerNonZeroPtr()[p] = originalOuterIndices[i+1] - originalOuterIndices[i]; + } } /* Compute the default threshold as in MatLab, see: @@ -386,7 +402,7 @@ void SparseQR::factorize(const MatrixType& mat) // all the nodes (with indexes lower than rank) reachable through the column elimination tree (etree) rooted at node k. // Note: if the diagonal entry does not exist, then its contribution must be explicitly added, // thus the trick with found_diag that permits to do one more iteration on the diagonal element if this one has not been found. - for (typename MatrixType::InnerIterator itp(m_pmat, col); itp || !found_diag; ++itp) + for (typename QRMatrixType::InnerIterator itp(m_pmat, col); itp || !found_diag; ++itp) { Index curIdx = nonzeroCol; if(itp) curIdx = itp.row(); @@ -544,7 +560,7 @@ void SparseQR::factorize(const MatrixType& mat) if(nonzeroCol Date: Fri, 19 Sep 2014 10:10:29 +0200 Subject: [PATCH 0804/1103] Added tag before-evaluators for changeset 9452eb38f812194a676edc1b9eb9d08b7bc0f297 From 0a18eecab332d0dd87154b9eef7ff993a4bb625c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 19 Sep 2014 13:25:28 +0200 Subject: [PATCH 0805/1103] bug #100: add support for explicit scalar to Array conversion (as enable implicit conversion is much more tricky) --- Eigen/src/Core/Array.h | 18 ++++++++++++++-- Eigen/src/Core/ArrayBase.h | 6 ++++++ Eigen/src/Core/PlainObjectBase.h | 36 +++++++++++++++++++++++++++++++- test/array.cpp | 25 ++++++++++++++++++++++ 4 files changed, 82 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h index 28d6f1443..eaee8847b 100644 --- a/Eigen/src/Core/Array.h +++ b/Eigen/src/Core/Array.h @@ -74,6 +74,21 @@ class Array { return Base::operator=(other); } + + /** Set all the entries to \a value. + * \sa DenseBase::setConstant(), DenseBase::fill() + */ + /* This overload is needed because the usage of + * using Base::operator=; + * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped + * the usage of 'using'. This should be done only for operator=. + */ + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Array& operator=(const Scalar &value) + { + Base::setConstant(value); + return *this; + } /** Copies the value of the expression \a other into \c *this with automatic resizing. * @@ -99,7 +114,7 @@ class Array { return Base::_set(other); } - + /** Default constructor. * * For fixed-size matrices, does nothing. @@ -144,7 +159,6 @@ class Array } #endif - #ifndef EIGEN_PARSED_BY_DOXYGEN template EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index 4e80634b9..48a0006d5 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -122,6 +122,12 @@ template class ArrayBase { internal::call_assignment(derived(), other.derived()); } + + /** Set all the entries to \a value. + * \sa DenseBase::setConstant(), DenseBase::fill() */ + EIGEN_DEVICE_FUNC + Derived& operator=(const Scalar &value) + { Base::setConstant(value); return derived(); } EIGEN_DEVICE_FUNC Derived& operator+=(const Scalar& scalar); diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 11aec1552..3b0e56445 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -700,9 +700,12 @@ class PlainObjectBase : public internal::dense_xpr_base::type m_storage.data()[1] = Scalar(val1); } + // The argument is convertible to the Index type and we either have a non 1x1 Matrix, or a dynamic-sized Array, + // then the argument is meant to be the size of the object. template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if::value,T>::type* = 0) + EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if< (Base::SizeAtCompileTime!=1 || !internal::is_convertible::value) + && ((!internal::is_same::XprKind,ArrayXpr>::value || Base::SizeAtCompileTime==Dynamic)),T>::type* = 0) { // NOTE MSVC 2008 complains if we directly put bool(NumTraits::IsInteger) as the EIGEN_STATIC_ASSERT argument. const bool is_integer = NumTraits::IsInteger; @@ -710,6 +713,8 @@ class PlainObjectBase : public internal::dense_xpr_base::type FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED) resize(size); } + + // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitely converted) template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if::value,T>::type* = 0) @@ -718,6 +723,7 @@ class PlainObjectBase : public internal::dense_xpr_base::type m_storage.data()[0] = val0; } + // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type match the index type) template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init1(const Index& val0, @@ -730,18 +736,21 @@ class PlainObjectBase : public internal::dense_xpr_base::type m_storage.data()[0] = Scalar(val0); } + // Initialize a fixed size matrix from a pointer to raw data template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init1(const Scalar* data){ this->_set_noalias(ConstMapType(data)); } + // Initialize an arbitrary matrix from a dense expression template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init1(const DenseBase& other){ this->_set_noalias(other); } + // Initialize an arbitrary matrix from a generic Eigen expression template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void _init1(const EigenBase& other){ @@ -762,6 +771,31 @@ class PlainObjectBase : public internal::dense_xpr_base::type { this->derived() = r; } + + // For fixed -size arrays: + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Scalar& val0, + typename internal::enable_if< Base::SizeAtCompileTime!=Dynamic + && Base::SizeAtCompileTime!=1 + && internal::is_convertible::value + && internal::is_same::XprKind,ArrayXpr>::value,T>::type* = 0) + { + Base::setConstant(val0); + } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void _init1(const Index& val0, + typename internal::enable_if< (!internal::is_same::value) + && (internal::is_same::value) + && Base::SizeAtCompileTime!=Dynamic + && Base::SizeAtCompileTime!=1 + && internal::is_convertible::value + && internal::is_same::XprKind,ArrayXpr>::value,T*>::type* = 0) + { + Base::setConstant(val0); + } template friend struct internal::matrix_swap_impl; diff --git a/test/array.cpp b/test/array.cpp index 010fead2d..ac9be097d 100644 --- a/test/array.cpp +++ b/test/array.cpp @@ -81,6 +81,31 @@ template void array(const ArrayType& m) VERIFY_IS_APPROX(m3.rowwise() += rv1, m1.rowwise() + rv1); m3 = m1; VERIFY_IS_APPROX(m3.rowwise() -= rv1, m1.rowwise() - rv1); + + // Conversion from scalar + VERIFY_IS_APPROX((m3 = s1), ArrayType::Constant(rows,cols,s1)); + VERIFY_IS_APPROX((m3 = 1), ArrayType::Constant(rows,cols,1)); + VERIFY_IS_APPROX((m3.topLeftCorner(rows,cols) = 1), ArrayType::Constant(rows,cols,1)); + typedef Array FixedArrayType; + FixedArrayType f1(s1); + VERIFY_IS_APPROX(f1, FixedArrayType::Constant(s1)); + FixedArrayType f2(numext::real(s1)); + VERIFY_IS_APPROX(f2, FixedArrayType::Constant(numext::real(s1))); + FixedArrayType f3((int)100*numext::real(s1)); + VERIFY_IS_APPROX(f3, FixedArrayType::Constant((int)100*numext::real(s1))); + f1.setRandom(); + FixedArrayType f4(f1.data()); + VERIFY_IS_APPROX(f4, f1); + + // Check possible conflicts with 1D ctor + typedef Array OneDArrayType; + OneDArrayType o1(rows); + VERIFY(o1.size()==rows); + OneDArrayType o4((int)rows); + VERIFY(o4.size()==rows); } template void comparisons(const ArrayType& m) From 03dd4dd91a5d8963f56eebe3b9d2eb924bc06e02 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 19 Sep 2014 15:25:48 +0200 Subject: [PATCH 0806/1103] Unify unit test for BDC and Jacobi SVD. This reveals some numerical issues in BDCSVD. --- test/jacobisvd.cpp | 413 ++---------------------------- test/svd_common.h | 454 +++++++++++++++++++++++++++++++++ unsupported/test/bdcsvd.cpp | 229 +++++------------ unsupported/test/jacobisvd.cpp | 198 -------------- unsupported/test/svd_common.h | 261 ------------------- 5 files changed, 538 insertions(+), 1017 deletions(-) create mode 100644 test/svd_common.h delete mode 100644 unsupported/test/jacobisvd.cpp delete mode 100644 unsupported/test/svd_common.h diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index cd04db5be..bfcadce95 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // Copyright (C) 2009 Benoit Jacob // // This Source Code Form is subject to the terms of the Mozilla @@ -14,273 +14,47 @@ #include "main.h" #include -template -void jacobisvd_check_full(const MatrixType& m, const JacobiSVD& svd) -{ - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime - }; - - typedef typename MatrixType::Scalar Scalar; - typedef Matrix MatrixUType; - typedef Matrix MatrixVType; - - MatrixType sigma = MatrixType::Zero(rows,cols); - sigma.diagonal() = svd.singularValues().template cast(); - MatrixUType u = svd.matrixU(); - MatrixVType v = svd.matrixV(); - - VERIFY_IS_APPROX(m, u * sigma * v.adjoint()); - VERIFY_IS_UNITARY(u); - VERIFY_IS_UNITARY(v); -} - -template -void jacobisvd_compare_to_full(const MatrixType& m, - unsigned int computationOptions, - const JacobiSVD& referenceSvd) -{ - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - Index diagSize = (std::min)(rows, cols); - - JacobiSVD svd(m, computationOptions); - - VERIFY_IS_APPROX(svd.singularValues(), referenceSvd.singularValues()); - if(computationOptions & ComputeFullU) - VERIFY_IS_APPROX(svd.matrixU(), referenceSvd.matrixU()); - if(computationOptions & ComputeThinU) - VERIFY_IS_APPROX(svd.matrixU(), referenceSvd.matrixU().leftCols(diagSize)); - if(computationOptions & ComputeFullV) - VERIFY_IS_APPROX(svd.matrixV(), referenceSvd.matrixV()); - if(computationOptions & ComputeThinV) - VERIFY_IS_APPROX(svd.matrixV(), referenceSvd.matrixV().leftCols(diagSize)); -} - -template -void jacobisvd_solve(const MatrixType& m, unsigned int computationOptions) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime - }; - - typedef Matrix RhsType; - typedef Matrix SolutionType; - - RhsType rhs = RhsType::Random(rows, internal::random(1, cols)); - JacobiSVD svd(m, computationOptions); - - if(internal::is_same::value) svd.setThreshold(1e-8); - else if(internal::is_same::value) svd.setThreshold(1e-4); - - SolutionType x = svd.solve(rhs); - - RealScalar residual = (m*x-rhs).norm(); - // Check that there is no significantly better solution in the neighborhood of x - if(!test_isMuchSmallerThan(residual,rhs.norm())) - { - // If the residual is very small, then we have an exact solution, so we are already good. - for(int k=0;k::epsilon(); - RealScalar residual_y = (m*y-rhs).norm(); - VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); - - y.row(k) = x.row(k).array() - 2*NumTraits::epsilon(); - residual_y = (m*y-rhs).norm(); - VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); - } - } - - // evaluate normal equation which works also for least-squares solutions - if(internal::is_same::value) - { - // This test is not stable with single precision. - // This is probably because squaring m signicantly affects the precision. - VERIFY_IS_APPROX(m.adjoint()*m*x,m.adjoint()*rhs); - } - - // check minimal norm solutions - { - // generate a full-rank m x n problem with m MatrixType2; - typedef Matrix RhsType2; - typedef Matrix MatrixType2T; - Index rank = RankAtCompileTime2==Dynamic ? internal::random(1,cols) : Index(RankAtCompileTime2); - MatrixType2 m2(rank,cols); - int guard = 0; - do { - m2.setRandom(); - } while(m2.jacobiSvd().setThreshold(test_precision()).rank()!=rank && (++guard)<10); - VERIFY(guard<10); - RhsType2 rhs2 = RhsType2::Random(rank); - // use QR to find a reference minimal norm solution - HouseholderQR qr(m2.adjoint()); - Matrix tmp = qr.matrixQR().topLeftCorner(rank,rank).template triangularView().adjoint().solve(rhs2); - tmp.conservativeResize(cols); - tmp.tail(cols-rank).setZero(); - SolutionType x21 = qr.householderQ() * tmp; - // now check with SVD - JacobiSVD svd2(m2, computationOptions); - SolutionType x22 = svd2.solve(rhs2); - VERIFY_IS_APPROX(m2*x21, rhs2); - VERIFY_IS_APPROX(m2*x22, rhs2); - VERIFY_IS_APPROX(x21, x22); - - // Now check with a rank deficient matrix - typedef Matrix MatrixType3; - typedef Matrix RhsType3; - Index rows3 = RowsAtCompileTime3==Dynamic ? internal::random(rank+1,2*cols) : Index(RowsAtCompileTime3); - Matrix C = Matrix::Random(rows3,rank); - MatrixType3 m3 = C * m2; - RhsType3 rhs3 = C * rhs2; - JacobiSVD svd3(m3, computationOptions); - SolutionType x3 = svd3.solve(rhs3); - VERIFY_IS_APPROX(m3*x3, rhs3); - VERIFY_IS_APPROX(m3*x21, rhs3); - VERIFY_IS_APPROX(m2*x3, rhs2); - - VERIFY_IS_APPROX(x21, x3); - } -} - -template -void jacobisvd_test_all_computation_options(const MatrixType& m) -{ - if (QRPreconditioner == NoQRPreconditioner && m.rows() != m.cols()) - return; - JacobiSVD fullSvd(m, ComputeFullU|ComputeFullV); - CALL_SUBTEST(( jacobisvd_check_full(m, fullSvd) )); - CALL_SUBTEST(( jacobisvd_solve(m, ComputeFullU | ComputeFullV) )); - - #if defined __INTEL_COMPILER - // remark #111: statement is unreachable - #pragma warning disable 111 - #endif - if(QRPreconditioner == FullPivHouseholderQRPreconditioner) - return; - - CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeFullU, fullSvd) )); - CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeFullV, fullSvd) )); - CALL_SUBTEST(( jacobisvd_compare_to_full(m, 0, fullSvd) )); - - if (MatrixType::ColsAtCompileTime == Dynamic) { - // thin U/V are only available with dynamic number of columns - CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeFullU|ComputeThinV, fullSvd) )); - CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinV, fullSvd) )); - CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinU|ComputeFullV, fullSvd) )); - CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinU , fullSvd) )); - CALL_SUBTEST(( jacobisvd_compare_to_full(m, ComputeThinU|ComputeThinV, fullSvd) )); - CALL_SUBTEST(( jacobisvd_solve(m, ComputeFullU | ComputeThinV) )); - CALL_SUBTEST(( jacobisvd_solve(m, ComputeThinU | ComputeFullV) )); - CALL_SUBTEST(( jacobisvd_solve(m, ComputeThinU | ComputeThinV) )); - - // test reconstruction - typedef typename MatrixType::Index Index; - Index diagSize = (std::min)(m.rows(), m.cols()); - JacobiSVD svd(m, ComputeThinU | ComputeThinV); - VERIFY_IS_APPROX(m, svd.matrixU().leftCols(diagSize) * svd.singularValues().asDiagonal() * svd.matrixV().leftCols(diagSize).adjoint()); - } -} +#define SVD_DEFAULT(M) JacobiSVD +#define SVD_FOR_MIN_NORM(M) JacobiSVD +#include "svd_common.h" +// Check all variants of JacobiSVD template void jacobisvd(const MatrixType& a = MatrixType(), bool pickrandom = true) { MatrixType m = a; if(pickrandom) - { - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; - Index diagSize = (std::min)(a.rows(), a.cols()); - RealScalar s = std::numeric_limits::max_exponent10/4; - s = internal::random(1,s); - Matrix d = Matrix::Random(diagSize); - for(Index k=0; k(-s,s)); - m = Matrix::Random(a.rows(),diagSize) * d.asDiagonal() * Matrix::Random(diagSize,a.cols()); - // cancel some coeffs - Index n = internal::random(0,m.size()-1); - for(Index i=0; i(0,m.rows()-1), internal::random(0,m.cols()-1)) = Scalar(0); - } + svd_fill_random(m); - CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); - CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); - CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); - CALL_SUBTEST(( jacobisvd_test_all_computation_options(m) )); + CALL_SUBTEST(( svd_test_all_computation_options >(m, true) )); // check full only + CALL_SUBTEST(( svd_test_all_computation_options >(m, false) )); + CALL_SUBTEST(( svd_test_all_computation_options >(m, false) )); + if(m.rows()==m.cols()) + CALL_SUBTEST(( svd_test_all_computation_options >(m, false) )); } template void jacobisvd_verify_assert(const MatrixType& m) { - typedef typename MatrixType::Scalar Scalar; + svd_verify_assert >(m); typedef typename MatrixType::Index Index; Index rows = m.rows(); Index cols = m.cols(); enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, ColsAtCompileTime = MatrixType::ColsAtCompileTime }; - typedef Matrix RhsType; - - RhsType rhs(rows); - - JacobiSVD svd; - VERIFY_RAISES_ASSERT(svd.matrixU()) - VERIFY_RAISES_ASSERT(svd.singularValues()) - VERIFY_RAISES_ASSERT(svd.matrixV()) - VERIFY_RAISES_ASSERT(svd.solve(rhs)) MatrixType a = MatrixType::Zero(rows, cols); a.setZero(); - svd.compute(a, 0); - VERIFY_RAISES_ASSERT(svd.matrixU()) - VERIFY_RAISES_ASSERT(svd.matrixV()) - svd.singularValues(); - VERIFY_RAISES_ASSERT(svd.solve(rhs)) if (ColsAtCompileTime == Dynamic) { - svd.compute(a, ComputeThinU); - svd.matrixU(); - VERIFY_RAISES_ASSERT(svd.matrixV()) - VERIFY_RAISES_ASSERT(svd.solve(rhs)) - - svd.compute(a, ComputeThinV); - svd.matrixV(); - VERIFY_RAISES_ASSERT(svd.matrixU()) - VERIFY_RAISES_ASSERT(svd.solve(rhs)) - JacobiSVD svd_fullqr; VERIFY_RAISES_ASSERT(svd_fullqr.compute(a, ComputeFullU|ComputeThinV)) VERIFY_RAISES_ASSERT(svd_fullqr.compute(a, ComputeThinU|ComputeThinV)) VERIFY_RAISES_ASSERT(svd_fullqr.compute(a, ComputeThinU|ComputeFullV)) } - else - { - VERIFY_RAISES_ASSERT(svd.compute(a, ComputeThinU)) - VERIFY_RAISES_ASSERT(svd.compute(a, ComputeThinV)) - } } template @@ -296,165 +70,17 @@ void jacobisvd_method() VERIFY_IS_APPROX(m.jacobiSvd(ComputeFullU|ComputeFullV).solve(m), m); } -// work around stupid msvc error when constructing at compile time an expression that involves -// a division by zero, even if the numeric type has floating point -template -EIGEN_DONT_INLINE Scalar zero() { return Scalar(0); } - -// workaround aggressive optimization in ICC -template EIGEN_DONT_INLINE T sub(T a, T b) { return a - b; } - -template -void jacobisvd_inf_nan() -{ - // all this function does is verify we don't iterate infinitely on nan/inf values - - JacobiSVD svd; - typedef typename MatrixType::Scalar Scalar; - Scalar some_inf = Scalar(1) / zero(); - VERIFY(sub(some_inf, some_inf) != sub(some_inf, some_inf)); - svd.compute(MatrixType::Constant(10,10,some_inf), ComputeFullU | ComputeFullV); - - Scalar nan = std::numeric_limits::quiet_NaN(); - VERIFY(nan != nan); - svd.compute(MatrixType::Constant(10,10,nan), ComputeFullU | ComputeFullV); - - MatrixType m = MatrixType::Zero(10,10); - m(internal::random(0,9), internal::random(0,9)) = some_inf; - svd.compute(m, ComputeFullU | ComputeFullV); - - m = MatrixType::Zero(10,10); - m(internal::random(0,9), internal::random(0,9)) = nan; - svd.compute(m, ComputeFullU | ComputeFullV); - - // regression test for bug 791 - m.resize(3,3); - m << 0, 2*NumTraits::epsilon(), 0.5, - 0, -0.5, 0, - nan, 0, 0; - svd.compute(m, ComputeFullU | ComputeFullV); - - m.resize(4,4); - m << 1, 0, 0, 0, - 0, 3, 1, 2e-308, - 1, 0, 1, nan, - 0, nan, nan, 0; - svd.compute(m, ComputeFullU | ComputeFullV); -} - -// Regression test for bug 286: JacobiSVD loops indefinitely with some -// matrices containing denormal numbers. -void jacobisvd_underoverflow() -{ -#if defined __INTEL_COMPILER -// shut up warning #239: floating point underflow -#pragma warning push -#pragma warning disable 239 -#endif - Matrix2d M; - M << -7.90884e-313, -4.94e-324, - 0, 5.60844e-313; - JacobiSVD svd; - svd.compute(M,ComputeFullU|ComputeFullV); - jacobisvd_check_full(M,svd); - - VectorXd value_set(9); - value_set << 0, 1, -1, 5.60844e-313, -5.60844e-313, 4.94e-324, -4.94e-324, -4.94e-223, 4.94e-223; - Array4i id(0,0,0,0); - int k = 0; - do - { - M << value_set(id(0)), value_set(id(1)), value_set(id(2)), value_set(id(3)); - svd.compute(M,ComputeFullU|ComputeFullV); - jacobisvd_check_full(M,svd); - - id(k)++; - if(id(k)>=value_set.size()) - { - while(k<3 && id(k)>=value_set.size()) id(++k)++; - id.head(k).setZero(); - k=0; - } - - } while((id svd3; - svd3.compute(M3,ComputeFullU|ComputeFullV); // just check we don't loop indefinitely - jacobisvd_check_full(M3,svd3); -} - -void jacobisvd_preallocate() -{ - Vector3f v(3.f, 2.f, 1.f); - MatrixXf m = v.asDiagonal(); - - internal::set_is_malloc_allowed(false); - VERIFY_RAISES_ASSERT(VectorXf tmp(10);) - JacobiSVD svd; - internal::set_is_malloc_allowed(true); - svd.compute(m); - VERIFY_IS_APPROX(svd.singularValues(), v); - - JacobiSVD svd2(3,3); - internal::set_is_malloc_allowed(false); - svd2.compute(m); - internal::set_is_malloc_allowed(true); - VERIFY_IS_APPROX(svd2.singularValues(), v); - VERIFY_RAISES_ASSERT(svd2.matrixU()); - VERIFY_RAISES_ASSERT(svd2.matrixV()); - svd2.compute(m, ComputeFullU | ComputeFullV); - VERIFY_IS_APPROX(svd2.matrixU(), Matrix3f::Identity()); - VERIFY_IS_APPROX(svd2.matrixV(), Matrix3f::Identity()); - internal::set_is_malloc_allowed(false); - svd2.compute(m); - internal::set_is_malloc_allowed(true); - - JacobiSVD svd3(3,3,ComputeFullU|ComputeFullV); - internal::set_is_malloc_allowed(false); - svd2.compute(m); - internal::set_is_malloc_allowed(true); - VERIFY_IS_APPROX(svd2.singularValues(), v); - VERIFY_IS_APPROX(svd2.matrixU(), Matrix3f::Identity()); - VERIFY_IS_APPROX(svd2.matrixV(), Matrix3f::Identity()); - internal::set_is_malloc_allowed(false); - svd2.compute(m, ComputeFullU|ComputeFullV); - internal::set_is_malloc_allowed(true); -} - void test_jacobisvd() { CALL_SUBTEST_3(( jacobisvd_verify_assert(Matrix3f()) )); CALL_SUBTEST_4(( jacobisvd_verify_assert(Matrix4d()) )); CALL_SUBTEST_7(( jacobisvd_verify_assert(MatrixXf(10,12)) )); CALL_SUBTEST_8(( jacobisvd_verify_assert(MatrixXcd(7,5)) )); + + svd_all_trivial_2x2(jacobisvd); + svd_all_trivial_2x2(jacobisvd); for(int i = 0; i < g_repeat; i++) { - Matrix2cd m; - m << 0, 1, - 0, 1; - CALL_SUBTEST_1(( jacobisvd(m, false) )); - m << 1, 0, - 1, 0; - CALL_SUBTEST_1(( jacobisvd(m, false) )); - - Matrix2d n; - n << 0, 0, - 0, 0; - CALL_SUBTEST_2(( jacobisvd(n, false) )); - n << 0, 0, - 0, 1; - CALL_SUBTEST_2(( jacobisvd(n, false) )); - CALL_SUBTEST_3(( jacobisvd() )); CALL_SUBTEST_4(( jacobisvd() )); CALL_SUBTEST_5(( jacobisvd >() )); @@ -473,8 +99,8 @@ void test_jacobisvd() (void) c; // Test on inf/nan matrix - CALL_SUBTEST_7( jacobisvd_inf_nan() ); - CALL_SUBTEST_10( jacobisvd_inf_nan() ); + CALL_SUBTEST_7( (svd_inf_nan, MatrixXf>()) ); + CALL_SUBTEST_10( (svd_inf_nan, MatrixXd>()) ); } CALL_SUBTEST_7(( jacobisvd(MatrixXf(internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2), internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2))) )); @@ -488,8 +114,7 @@ void test_jacobisvd() CALL_SUBTEST_7( JacobiSVD(10,10) ); // Check that preallocation avoids subsequent mallocs - CALL_SUBTEST_9( jacobisvd_preallocate() ); + CALL_SUBTEST_9( svd_preallocate() ); - // Regression check for bug 286 - CALL_SUBTEST_2( jacobisvd_underoverflow() ); + CALL_SUBTEST_2( svd_underoverflow() ); } diff --git a/test/svd_common.h b/test/svd_common.h new file mode 100644 index 000000000..4631939e5 --- /dev/null +++ b/test/svd_common.h @@ -0,0 +1,454 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2008-2014 Gael Guennebaud +// Copyright (C) 2009 Benoit Jacob +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef SVD_DEFAULT +#error a macro SVD_DEFAULT(MatrixType) must be defined prior to including svd_common.h +#endif + +#ifndef SVD_FOR_MIN_NORM +#error a macro SVD_FOR_MIN_NORM(MatrixType) must be defined prior to including svd_common.h +#endif + +// Check that the matrix m is properly reconstructed and that the U and V factors are unitary +// The SVD must have already been computed. +template +void svd_check_full(const MatrixType& m, const SvdType& svd) +{ + typedef typename MatrixType::Index Index; + Index rows = m.rows(); + Index cols = m.cols(); + + enum { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime + }; + + typedef typename MatrixType::Scalar Scalar; + typedef Matrix MatrixUType; + typedef Matrix MatrixVType; + + MatrixType sigma = MatrixType::Zero(rows,cols); + sigma.diagonal() = svd.singularValues().template cast(); + MatrixUType u = svd.matrixU(); + MatrixVType v = svd.matrixV(); + + VERIFY_IS_APPROX(m, u * sigma * v.adjoint()); + VERIFY_IS_UNITARY(u); + VERIFY_IS_UNITARY(v); +} + +// Compare partial SVD defined by computationOptions to a full SVD referenceSvd +template +void svd_compare_to_full(const MatrixType& m, + unsigned int computationOptions, + const SvdType& referenceSvd) +{ + typedef typename MatrixType::Index Index; + Index rows = m.rows(); + Index cols = m.cols(); + Index diagSize = (std::min)(rows, cols); + + SvdType svd(m, computationOptions); + + VERIFY_IS_APPROX(svd.singularValues(), referenceSvd.singularValues()); + if(computationOptions & ComputeFullU) VERIFY_IS_APPROX(svd.matrixU(), referenceSvd.matrixU()); + if(computationOptions & ComputeThinU) VERIFY_IS_APPROX(svd.matrixU(), referenceSvd.matrixU().leftCols(diagSize)); + if(computationOptions & ComputeFullV) VERIFY_IS_APPROX(svd.matrixV(), referenceSvd.matrixV()); + if(computationOptions & ComputeThinV) VERIFY_IS_APPROX(svd.matrixV(), referenceSvd.matrixV().leftCols(diagSize)); +} + +// +template +void svd_least_square(const MatrixType& m, unsigned int computationOptions) +{ + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + typedef typename MatrixType::Index Index; + Index rows = m.rows(); + Index cols = m.cols(); + + enum { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime + }; + + typedef Matrix RhsType; + typedef Matrix SolutionType; + + RhsType rhs = RhsType::Random(rows, internal::random(1, cols)); + SvdType svd(m, computationOptions); + + if(internal::is_same::value) svd.setThreshold(1e-8); + else if(internal::is_same::value) svd.setThreshold(1e-4); + + SolutionType x = svd.solve(rhs); + + RealScalar residual = (m*x-rhs).norm(); + // Check that there is no significantly better solution in the neighborhood of x + if(!test_isMuchSmallerThan(residual,rhs.norm())) + { + // If the residual is very small, then we have an exact solution, so we are already good. + for(int k=0;k::epsilon(); + RealScalar residual_y = (m*y-rhs).norm(); + VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); + + y.row(k) = x.row(k).array() - 2*NumTraits::epsilon(); + residual_y = (m*y-rhs).norm(); + VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); + } + } + + // evaluate normal equation which works also for least-squares solutions + if(internal::is_same::value) + { + // This test is not stable with single precision. + // This is probably because squaring m signicantly affects the precision. + VERIFY_IS_APPROX(m.adjoint()*m*x,m.adjoint()*rhs); + } +} + +// check minimal norm solutions, the inoput matrix m is only used to recover problem size +template +void svd_min_norm(const MatrixType& m, unsigned int computationOptions) +{ + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::Index Index; + Index cols = m.cols(); + + enum { + ColsAtCompileTime = MatrixType::ColsAtCompileTime + }; + + typedef Matrix SolutionType; + + // generate a full-rank m x n problem with m MatrixType2; + typedef Matrix RhsType2; + typedef Matrix MatrixType2T; + Index rank = RankAtCompileTime2==Dynamic ? internal::random(1,cols) : Index(RankAtCompileTime2); + MatrixType2 m2(rank,cols); + int guard = 0; + do { + m2.setRandom(); + } while(SVD_FOR_MIN_NORM(MatrixType2)(m2).setThreshold(test_precision()).rank()!=rank && (++guard)<10); + VERIFY(guard<10); + RhsType2 rhs2 = RhsType2::Random(rank); + // use QR to find a reference minimal norm solution + HouseholderQR qr(m2.adjoint()); + Matrix tmp = qr.matrixQR().topLeftCorner(rank,rank).template triangularView().adjoint().solve(rhs2); + tmp.conservativeResize(cols); + tmp.tail(cols-rank).setZero(); + SolutionType x21 = qr.householderQ() * tmp; + // now check with SVD + SVD_FOR_MIN_NORM(MatrixType2) svd2(m2, computationOptions); + SolutionType x22 = svd2.solve(rhs2); + VERIFY_IS_APPROX(m2*x21, rhs2); + VERIFY_IS_APPROX(m2*x22, rhs2); + VERIFY_IS_APPROX(x21, x22); + + // Now check with a rank deficient matrix + typedef Matrix MatrixType3; + typedef Matrix RhsType3; + Index rows3 = RowsAtCompileTime3==Dynamic ? internal::random(rank+1,2*cols) : Index(RowsAtCompileTime3); + Matrix C = Matrix::Random(rows3,rank); + MatrixType3 m3 = C * m2; + RhsType3 rhs3 = C * rhs2; + SVD_FOR_MIN_NORM(MatrixType3) svd3(m3, computationOptions); + SolutionType x3 = svd3.solve(rhs3); + VERIFY_IS_APPROX(m3*x3, rhs3); + VERIFY_IS_APPROX(m3*x21, rhs3); + VERIFY_IS_APPROX(m2*x3, rhs2); + + VERIFY_IS_APPROX(x21, x3); +} + +// Check full, compare_to_full, least_square, and min_norm for all possible compute-options +template +void svd_test_all_computation_options(const MatrixType& m, bool full_only) +{ +// if (QRPreconditioner == NoQRPreconditioner && m.rows() != m.cols()) +// return; + SvdType fullSvd(m, ComputeFullU|ComputeFullV); + CALL_SUBTEST(( svd_check_full(m, fullSvd) )); + CALL_SUBTEST(( svd_least_square(m, ComputeFullU | ComputeFullV) )); + CALL_SUBTEST(( svd_min_norm(m, ComputeFullU | ComputeFullV) )); + + #if defined __INTEL_COMPILER + // remark #111: statement is unreachable + #pragma warning disable 111 + #endif + if(full_only) + return; + + CALL_SUBTEST(( svd_compare_to_full(m, ComputeFullU, fullSvd) )); + CALL_SUBTEST(( svd_compare_to_full(m, ComputeFullV, fullSvd) )); + CALL_SUBTEST(( svd_compare_to_full(m, 0, fullSvd) )); + + if (MatrixType::ColsAtCompileTime == Dynamic) { + // thin U/V are only available with dynamic number of columns + CALL_SUBTEST(( svd_compare_to_full(m, ComputeFullU|ComputeThinV, fullSvd) )); + CALL_SUBTEST(( svd_compare_to_full(m, ComputeThinV, fullSvd) )); + CALL_SUBTEST(( svd_compare_to_full(m, ComputeThinU|ComputeFullV, fullSvd) )); + CALL_SUBTEST(( svd_compare_to_full(m, ComputeThinU , fullSvd) )); + CALL_SUBTEST(( svd_compare_to_full(m, ComputeThinU|ComputeThinV, fullSvd) )); + + CALL_SUBTEST(( svd_least_square(m, ComputeFullU | ComputeThinV) )); + CALL_SUBTEST(( svd_least_square(m, ComputeThinU | ComputeFullV) )); + CALL_SUBTEST(( svd_least_square(m, ComputeThinU | ComputeThinV) )); + + CALL_SUBTEST(( svd_min_norm(m, ComputeFullU | ComputeThinV) )); + CALL_SUBTEST(( svd_min_norm(m, ComputeThinU | ComputeFullV) )); + CALL_SUBTEST(( svd_min_norm(m, ComputeThinU | ComputeThinV) )); + + // test reconstruction + typedef typename MatrixType::Index Index; + Index diagSize = (std::min)(m.rows(), m.cols()); + SvdType svd(m, ComputeThinU | ComputeThinV); + VERIFY_IS_APPROX(m, svd.matrixU().leftCols(diagSize) * svd.singularValues().asDiagonal() * svd.matrixV().leftCols(diagSize).adjoint()); + } +} + +template +void svd_fill_random(MatrixType &m) +{ + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + typedef typename MatrixType::Index Index; + Index diagSize = (std::min)(m.rows(), m.cols()); + RealScalar s = std::numeric_limits::max_exponent10/4; + s = internal::random(1,s); + Matrix d = Matrix::Random(diagSize); + for(Index k=0; k(-s,s)); + m = Matrix::Random(m.rows(),diagSize) * d.asDiagonal() * Matrix::Random(diagSize,m.cols()); + // cancel some coeffs + Index n = internal::random(0,m.size()-1); + for(Index i=0; i(0,m.rows()-1), internal::random(0,m.cols()-1)) = Scalar(0); +} + + +// work around stupid msvc error when constructing at compile time an expression that involves +// a division by zero, even if the numeric type has floating point +template +EIGEN_DONT_INLINE Scalar zero() { return Scalar(0); } + +// workaround aggressive optimization in ICC +template EIGEN_DONT_INLINE T sub(T a, T b) { return a - b; } + +// all this function does is verify we don't iterate infinitely on nan/inf values +template +void svd_inf_nan() +{ + SvdType svd; + typedef typename MatrixType::Scalar Scalar; + Scalar some_inf = Scalar(1) / zero(); + VERIFY(sub(some_inf, some_inf) != sub(some_inf, some_inf)); + svd.compute(MatrixType::Constant(10,10,some_inf), ComputeFullU | ComputeFullV); + + Scalar nan = std::numeric_limits::quiet_NaN(); + VERIFY(nan != nan); + svd.compute(MatrixType::Constant(10,10,nan), ComputeFullU | ComputeFullV); + + MatrixType m = MatrixType::Zero(10,10); + m(internal::random(0,9), internal::random(0,9)) = some_inf; + svd.compute(m, ComputeFullU | ComputeFullV); + + m = MatrixType::Zero(10,10); + m(internal::random(0,9), internal::random(0,9)) = nan; + svd.compute(m, ComputeFullU | ComputeFullV); + + // regression test for bug 791 + m.resize(3,3); + m << 0, 2*NumTraits::epsilon(), 0.5, + 0, -0.5, 0, + nan, 0, 0; + svd.compute(m, ComputeFullU | ComputeFullV); + + m.resize(4,4); + m << 1, 0, 0, 0, + 0, 3, 1, 2e-308, + 1, 0, 1, nan, + 0, nan, nan, 0; + svd.compute(m, ComputeFullU | ComputeFullV); +} + +// Regression test for bug 286: JacobiSVD loops indefinitely with some +// matrices containing denormal numbers. +void svd_underoverflow() +{ +#if defined __INTEL_COMPILER +// shut up warning #239: floating point underflow +#pragma warning push +#pragma warning disable 239 +#endif + Matrix2d M; + M << -7.90884e-313, -4.94e-324, + 0, 5.60844e-313; + SVD_DEFAULT(Matrix2d) svd; + svd.compute(M,ComputeFullU|ComputeFullV); + svd_check_full(M,svd); + + // Check all 2x2 matrices made with the following coefficients: + VectorXd value_set(9); + value_set << 0, 1, -1, 5.60844e-313, -5.60844e-313, 4.94e-324, -4.94e-324, -4.94e-223, 4.94e-223; + Array4i id(0,0,0,0); + int k = 0; + do + { + M << value_set(id(0)), value_set(id(1)), value_set(id(2)), value_set(id(3)); + svd.compute(M,ComputeFullU|ComputeFullV); + svd_check_full(M,svd); + + id(k)++; + if(id(k)>=value_set.size()) + { + while(k<3 && id(k)>=value_set.size()) id(++k)++; + id.head(k).setZero(); + k=0; + } + + } while((id +void svd_all_trivial_2x2( void (*cb)(const MatrixType&,bool) ) +{ + MatrixType M; + VectorXd value_set(3); + value_set << 0, 1, -1; + Array4i id(0,0,0,0); + int k = 0; + do + { + M << value_set(id(0)), value_set(id(1)), value_set(id(2)), value_set(id(3)); + + cb(M,false); + + id(k)++; + if(id(k)>=value_set.size()) + { + while(k<3 && id(k)>=value_set.size()) id(++k)++; + id.head(k).setZero(); + k=0; + } + + } while((id +void svd_verify_assert(const MatrixType& m) +{ + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::Index Index; + Index rows = m.rows(); + Index cols = m.cols(); + + enum { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime + }; + + typedef Matrix RhsType; + RhsType rhs(rows); + SvdType svd; + VERIFY_RAISES_ASSERT(svd.matrixU()) + VERIFY_RAISES_ASSERT(svd.singularValues()) + VERIFY_RAISES_ASSERT(svd.matrixV()) + VERIFY_RAISES_ASSERT(svd.solve(rhs)) + MatrixType a = MatrixType::Zero(rows, cols); + a.setZero(); + svd.compute(a, 0); + VERIFY_RAISES_ASSERT(svd.matrixU()) + VERIFY_RAISES_ASSERT(svd.matrixV()) + svd.singularValues(); + VERIFY_RAISES_ASSERT(svd.solve(rhs)) + + if (ColsAtCompileTime == Dynamic) + { + svd.compute(a, ComputeThinU); + svd.matrixU(); + VERIFY_RAISES_ASSERT(svd.matrixV()) + VERIFY_RAISES_ASSERT(svd.solve(rhs)) + svd.compute(a, ComputeThinV); + svd.matrixV(); + VERIFY_RAISES_ASSERT(svd.matrixU()) + VERIFY_RAISES_ASSERT(svd.solve(rhs)) + } + else + { + VERIFY_RAISES_ASSERT(svd.compute(a, ComputeThinU)) + VERIFY_RAISES_ASSERT(svd.compute(a, ComputeThinV)) + } +} + +#undef SVD_DEFAULT +#undef SVD_FOR_MIN_NORM diff --git a/unsupported/test/bdcsvd.cpp b/unsupported/test/bdcsvd.cpp index 115a649b0..4ad991522 100644 --- a/unsupported/test/bdcsvd.cpp +++ b/unsupported/test/bdcsvd.cpp @@ -10,204 +10,105 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/ -#include "svd_common.h" +// discard stack allocation as that too bypasses malloc +#define EIGEN_STACK_ALLOCATION_LIMIT 0 +#define EIGEN_RUNTIME_NO_MALLOC + +#include "main.h" +#include #include #include -// check if "svd" is the good image of "m" -template -void bdcsvd_check_full(const MatrixType& m, const BDCSVD& svd) -{ - svd_check_full< MatrixType, BDCSVD< MatrixType > >(m, svd); -} -// Compare to a reference value -template -void bdcsvd_compare_to_full(const MatrixType& m, - unsigned int computationOptions, - const BDCSVD& referenceSvd) -{ - svd_compare_to_full< MatrixType, BDCSVD< MatrixType > >(m, computationOptions, referenceSvd); -} // end bdcsvd_compare_to_full +#define SVD_DEFAULT(M) BDCSVD +// #define SVD_FOR_MIN_NORM(M) BDCSVD +#define SVD_FOR_MIN_NORM(M) JacobiSVD +#include "../../test/svd_common.h" - -template -void bdcsvd_solve(const MatrixType& m, unsigned int computationOptions) -{ - svd_solve< MatrixType, BDCSVD< MatrixType > >(m, computationOptions); -} // end template bdcsvd_solve - - -// test the computations options -template -void bdcsvd_test_all_computation_options(const MatrixType& m) -{ - BDCSVD fullSvd(m, ComputeFullU|ComputeFullV); - svd_test_computation_options_1< MatrixType, BDCSVD< MatrixType > >(m, fullSvd); - svd_test_computation_options_2< MatrixType, BDCSVD< MatrixType > >(m, fullSvd); -} // end bdcsvd_test_all_computation_options - - -// Call a test with all the computations options +// Check all variants of JacobiSVD template void bdcsvd(const MatrixType& a = MatrixType(), bool pickrandom = true) { - MatrixType m = pickrandom ? MatrixType::Random(a.rows(), a.cols()) : a; - bdcsvd_test_all_computation_options(m); -} // end template bdcsvd + MatrixType m = a; + if(pickrandom) + svd_fill_random(m); + CALL_SUBTEST(( svd_test_all_computation_options >(m, false) )); +} -// verify assert -template -void bdcsvd_verify_assert(const MatrixType& m) -{ - svd_verify_assert< MatrixType, BDCSVD< MatrixType > >(m); -}// end template bdcsvd_verify_assert - - -// test weird values -template -void bdcsvd_inf_nan() -{ - svd_inf_nan< MatrixType, BDCSVD< MatrixType > >(); -}// end template bdcsvd_inf_nan - - - -void bdcsvd_preallocate() -{ - svd_preallocate< BDCSVD< MatrixXf > >(); -} // end bdcsvd_preallocate - +// template +// void bdcsvd_method() +// { +// enum { Size = MatrixType::RowsAtCompileTime }; +// typedef typename MatrixType::RealScalar RealScalar; +// typedef Matrix RealVecType; +// MatrixType m = MatrixType::Identity(); +// VERIFY_IS_APPROX(m.bdcSvd().singularValues(), RealVecType::Ones()); +// VERIFY_RAISES_ASSERT(m.bdcSvd().matrixU()); +// VERIFY_RAISES_ASSERT(m.bdcSvd().matrixV()); +// VERIFY_IS_APPROX(m.bdcSvd(ComputeFullU|ComputeFullV).solve(m), m); +// } // compare the Singular values returned with Jacobi and Bdc template void compare_bdc_jacobi(const MatrixType& a = MatrixType(), unsigned int computationOptions = 0) { - std::cout << "debut compare" << std::endl; MatrixType m = MatrixType::Random(a.rows(), a.cols()); BDCSVD bdc_svd(m); JacobiSVD jacobi_svd(m); VERIFY_IS_APPROX(bdc_svd.singularValues(), jacobi_svd.singularValues()); - if(computationOptions & ComputeFullU) - VERIFY_IS_APPROX(bdc_svd.matrixU(), jacobi_svd.matrixU()); - if(computationOptions & ComputeThinU) - VERIFY_IS_APPROX(bdc_svd.matrixU(), jacobi_svd.matrixU()); - if(computationOptions & ComputeFullV) - VERIFY_IS_APPROX(bdc_svd.matrixV(), jacobi_svd.matrixV()); - if(computationOptions & ComputeThinV) - VERIFY_IS_APPROX(bdc_svd.matrixV(), jacobi_svd.matrixV()); - std::cout << "fin compare" << std::endl; -} // end template compare_bdc_jacobi + if(computationOptions & ComputeFullU) VERIFY_IS_APPROX(bdc_svd.matrixU(), jacobi_svd.matrixU()); + if(computationOptions & ComputeThinU) VERIFY_IS_APPROX(bdc_svd.matrixU(), jacobi_svd.matrixU()); + if(computationOptions & ComputeFullV) VERIFY_IS_APPROX(bdc_svd.matrixV(), jacobi_svd.matrixV()); + if(computationOptions & ComputeThinV) VERIFY_IS_APPROX(bdc_svd.matrixV(), jacobi_svd.matrixV()); +} - -// call the tests void test_bdcsvd() { - // test of Dynamic defined Matrix (42, 42) of float - CALL_SUBTEST_11(( bdcsvd_verify_assert > - (Matrix(42,42)) )); - CALL_SUBTEST_11(( compare_bdc_jacobi > - (Matrix(42,42), 0) )); - CALL_SUBTEST_11(( bdcsvd > - (Matrix(42,42)) )); - - // test of Dynamic defined Matrix (50, 50) of double - CALL_SUBTEST_13(( bdcsvd_verify_assert > - (Matrix(50,50)) )); - CALL_SUBTEST_13(( compare_bdc_jacobi > - (Matrix(50,50), 0) )); - CALL_SUBTEST_13(( bdcsvd > - (Matrix(50, 50)) )); - - // test of Dynamic defined Matrix (22, 22) of complex double - CALL_SUBTEST_14(( bdcsvd_verify_assert,Dynamic,Dynamic> > - (Matrix,Dynamic,Dynamic>(22,22)) )); - CALL_SUBTEST_14(( compare_bdc_jacobi,Dynamic,Dynamic> > - (Matrix, Dynamic, Dynamic> (22,22), 0) )); - CALL_SUBTEST_14(( bdcsvd,Dynamic,Dynamic> > - (Matrix,Dynamic,Dynamic>(22, 22)) )); - - // test of Dynamic defined Matrix (10, 10) of int - //CALL_SUBTEST_15(( bdcsvd_verify_assert > - // (Matrix(10,10)) )); - //CALL_SUBTEST_15(( compare_bdc_jacobi > - // (Matrix(10,10), 0) )); - //CALL_SUBTEST_15(( bdcsvd > - // (Matrix(10, 10)) )); + CALL_SUBTEST_3(( svd_verify_assert >(Matrix3f()) )); + CALL_SUBTEST_4(( svd_verify_assert >(Matrix4d()) )); + CALL_SUBTEST_7(( svd_verify_assert >(MatrixXf(10,12)) )); + CALL_SUBTEST_8(( svd_verify_assert >(MatrixXcd(7,5)) )); +// svd_all_trivial_2x2(bdcsvd); +// svd_all_trivial_2x2(bdcsvd); - // test of Dynamic defined Matrix (8, 6) of double - - CALL_SUBTEST_16(( bdcsvd_verify_assert > - (Matrix(8,6)) )); - CALL_SUBTEST_16(( compare_bdc_jacobi > - (Matrix(8, 6), 0) )); - CALL_SUBTEST_16(( bdcsvd > - (Matrix(8, 6)) )); - - - - // test of Dynamic defined Matrix (36, 12) of float - CALL_SUBTEST_17(( compare_bdc_jacobi > - (Matrix(36, 12), 0) )); - CALL_SUBTEST_17(( bdcsvd > - (Matrix(36, 12)) )); - - // test of Dynamic defined Matrix (5, 8) of double - CALL_SUBTEST_18(( compare_bdc_jacobi > - (Matrix(5, 8), 0) )); - CALL_SUBTEST_18(( bdcsvd > - (Matrix(5, 8)) )); - - - // non regression tests - CALL_SUBTEST_3(( bdcsvd_verify_assert(Matrix3f()) )); - CALL_SUBTEST_4(( bdcsvd_verify_assert(Matrix4d()) )); - CALL_SUBTEST_7(( bdcsvd_verify_assert(MatrixXf(10,12)) )); - CALL_SUBTEST_8(( bdcsvd_verify_assert(MatrixXcd(7,5)) )); - - // SUBTESTS 1 and 2 on specifics matrix for(int i = 0; i < g_repeat; i++) { - Matrix2cd m; - m << 0, 1, - 0, 1; - CALL_SUBTEST_1(( bdcsvd(m, false) )); - m << 1, 0, - 1, 0; - CALL_SUBTEST_1(( bdcsvd(m, false) )); +// CALL_SUBTEST_3(( bdcsvd() )); +// CALL_SUBTEST_4(( bdcsvd() )); +// CALL_SUBTEST_5(( bdcsvd >() )); - Matrix2d n; - n << 0, 0, - 0, 0; - CALL_SUBTEST_2(( bdcsvd(n, false) )); - n << 0, 0, - 0, 1; - CALL_SUBTEST_2(( bdcsvd(n, false) )); + int r = internal::random(1, EIGEN_TEST_MAX_SIZE/2), + c = internal::random(1, EIGEN_TEST_MAX_SIZE/2); - // Statics matrix don't work with BDSVD yet - // bdc algo on a random 3x3 float matrix - // CALL_SUBTEST_3(( bdcsvd() )); - // bdc algo on a random 4x4 double matrix - // CALL_SUBTEST_4(( bdcsvd() )); - // bdc algo on a random 3x5 float matrix - // CALL_SUBTEST_5(( bdcsvd >() )); - - int r = internal::random(1, 30), - c = internal::random(1, 30); - CALL_SUBTEST_7(( bdcsvd(MatrixXf(r,c)) )); - CALL_SUBTEST_8(( bdcsvd(MatrixXcd(r,c)) )); + TEST_SET_BUT_UNUSED_VARIABLE(r) + TEST_SET_BUT_UNUSED_VARIABLE(c) + + CALL_SUBTEST_6(( bdcsvd(Matrix(r,2)) )); + CALL_SUBTEST_7(( bdcsvd(MatrixXf(r,c)) )); + CALL_SUBTEST_7(( compare_bdc_jacobi(MatrixXf(r,c)) )); + CALL_SUBTEST_10(( bdcsvd(MatrixXd(r,c)) )); + CALL_SUBTEST_10(( compare_bdc_jacobi(MatrixXd(r,c)) )); + CALL_SUBTEST_8(( bdcsvd(MatrixXcd(r,c)) )); + CALL_SUBTEST_8(( compare_bdc_jacobi(MatrixXcd(r,c)) )); (void) r; (void) c; // Test on inf/nan matrix - CALL_SUBTEST_7( bdcsvd_inf_nan() ); + CALL_SUBTEST_7( (svd_inf_nan, MatrixXf>()) ); + CALL_SUBTEST_10( (svd_inf_nan, MatrixXd>()) ); } - CALL_SUBTEST_7(( bdcsvd(MatrixXf(internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2), internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2))) )); - CALL_SUBTEST_8(( bdcsvd(MatrixXcd(internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/3), internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/3))) )); + // test matrixbase method +// CALL_SUBTEST_1(( bdcsvd_method() )); +// CALL_SUBTEST_3(( bdcsvd_method() )); // Test problem size constructors CALL_SUBTEST_7( BDCSVD(10,10) ); -} // end test_bdcsvd + // Check that preallocation avoids subsequent mallocs + CALL_SUBTEST_9( svd_preallocate() ); + + CALL_SUBTEST_2( svd_underoverflow() ); +} + diff --git a/unsupported/test/jacobisvd.cpp b/unsupported/test/jacobisvd.cpp deleted file mode 100644 index b4e884eee..000000000 --- a/unsupported/test/jacobisvd.cpp +++ /dev/null @@ -1,198 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2009 Benoit Jacob -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -#include "svd_common.h" - -template -void jacobisvd_check_full(const MatrixType& m, const JacobiSVD& svd) -{ - svd_check_full >(m, svd); -} - -template -void jacobisvd_compare_to_full(const MatrixType& m, - unsigned int computationOptions, - const JacobiSVD& referenceSvd) -{ - svd_compare_to_full >(m, computationOptions, referenceSvd); -} - - -template -void jacobisvd_solve(const MatrixType& m, unsigned int computationOptions) -{ - svd_solve< MatrixType, JacobiSVD< MatrixType, QRPreconditioner > >(m, computationOptions); -} - - - -template -void jacobisvd_test_all_computation_options(const MatrixType& m) -{ - - if (QRPreconditioner == NoQRPreconditioner && m.rows() != m.cols()) - return; - - JacobiSVD< MatrixType, QRPreconditioner > fullSvd(m, ComputeFullU|ComputeFullV); - svd_test_computation_options_1< MatrixType, JacobiSVD< MatrixType, QRPreconditioner > >(m, fullSvd); - - if(QRPreconditioner == FullPivHouseholderQRPreconditioner) - return; - svd_test_computation_options_2< MatrixType, JacobiSVD< MatrixType, QRPreconditioner > >(m, fullSvd); - -} - -template -void jacobisvd(const MatrixType& a = MatrixType(), bool pickrandom = true) -{ - MatrixType m = pickrandom ? MatrixType::Random(a.rows(), a.cols()) : a; - - jacobisvd_test_all_computation_options(m); - jacobisvd_test_all_computation_options(m); - jacobisvd_test_all_computation_options(m); - jacobisvd_test_all_computation_options(m); -} - - -template -void jacobisvd_verify_assert(const MatrixType& m) -{ - - svd_verify_assert >(m); - - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime - }; - - MatrixType a = MatrixType::Zero(rows, cols); - a.setZero(); - - if (ColsAtCompileTime == Dynamic) - { - JacobiSVD svd_fullqr; - VERIFY_RAISES_ASSERT(svd_fullqr.compute(a, ComputeFullU|ComputeThinV)) - VERIFY_RAISES_ASSERT(svd_fullqr.compute(a, ComputeThinU|ComputeThinV)) - VERIFY_RAISES_ASSERT(svd_fullqr.compute(a, ComputeThinU|ComputeFullV)) - } -} - -template -void jacobisvd_method() -{ - enum { Size = MatrixType::RowsAtCompileTime }; - typedef typename MatrixType::RealScalar RealScalar; - typedef Matrix RealVecType; - MatrixType m = MatrixType::Identity(); - VERIFY_IS_APPROX(m.jacobiSvd().singularValues(), RealVecType::Ones()); - VERIFY_RAISES_ASSERT(m.jacobiSvd().matrixU()); - VERIFY_RAISES_ASSERT(m.jacobiSvd().matrixV()); - VERIFY_IS_APPROX(m.jacobiSvd(ComputeFullU|ComputeFullV).solve(m), m); -} - - - -template -void jacobisvd_inf_nan() -{ - svd_inf_nan >(); -} - - -// Regression test for bug 286: JacobiSVD loops indefinitely with some -// matrices containing denormal numbers. -void jacobisvd_bug286() -{ -#if defined __INTEL_COMPILER -// shut up warning #239: floating point underflow -#pragma warning push -#pragma warning disable 239 -#endif - Matrix2d M; - M << -7.90884e-313, -4.94e-324, - 0, 5.60844e-313; -#if defined __INTEL_COMPILER -#pragma warning pop -#endif - JacobiSVD svd; - svd.compute(M); // just check we don't loop indefinitely -} - - -void jacobisvd_preallocate() -{ - svd_preallocate< JacobiSVD >(); -} - -void test_jacobisvd() -{ - CALL_SUBTEST_11(( jacobisvd > - (Matrix(16, 6)) )); - - CALL_SUBTEST_3(( jacobisvd_verify_assert(Matrix3f()) )); - CALL_SUBTEST_4(( jacobisvd_verify_assert(Matrix4d()) )); - CALL_SUBTEST_7(( jacobisvd_verify_assert(MatrixXf(10,12)) )); - CALL_SUBTEST_8(( jacobisvd_verify_assert(MatrixXcd(7,5)) )); - - for(int i = 0; i < g_repeat; i++) { - Matrix2cd m; - m << 0, 1, - 0, 1; - CALL_SUBTEST_1(( jacobisvd(m, false) )); - m << 1, 0, - 1, 0; - CALL_SUBTEST_1(( jacobisvd(m, false) )); - - Matrix2d n; - n << 0, 0, - 0, 0; - CALL_SUBTEST_2(( jacobisvd(n, false) )); - n << 0, 0, - 0, 1; - CALL_SUBTEST_2(( jacobisvd(n, false) )); - - CALL_SUBTEST_3(( jacobisvd() )); - CALL_SUBTEST_4(( jacobisvd() )); - CALL_SUBTEST_5(( jacobisvd >() )); - CALL_SUBTEST_6(( jacobisvd >(Matrix(10,2)) )); - - int r = internal::random(1, 30), - c = internal::random(1, 30); - CALL_SUBTEST_7(( jacobisvd(MatrixXf(r,c)) )); - CALL_SUBTEST_8(( jacobisvd(MatrixXcd(r,c)) )); - (void) r; - (void) c; - - // Test on inf/nan matrix - CALL_SUBTEST_7( jacobisvd_inf_nan() ); - } - - CALL_SUBTEST_7(( jacobisvd(MatrixXf(internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2), internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/2))) )); - CALL_SUBTEST_8(( jacobisvd(MatrixXcd(internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/3), internal::random(EIGEN_TEST_MAX_SIZE/4, EIGEN_TEST_MAX_SIZE/3))) )); - - - // test matrixbase method - CALL_SUBTEST_1(( jacobisvd_method() )); - CALL_SUBTEST_3(( jacobisvd_method() )); - - - // Test problem size constructors - CALL_SUBTEST_7( JacobiSVD(10,10) ); - - // Check that preallocation avoids subsequent mallocs - CALL_SUBTEST_9( jacobisvd_preallocate() ); - - // Regression check for bug 286 - CALL_SUBTEST_2( jacobisvd_bug286() ); -} diff --git a/unsupported/test/svd_common.h b/unsupported/test/svd_common.h deleted file mode 100644 index 6a96e7c74..000000000 --- a/unsupported/test/svd_common.h +++ /dev/null @@ -1,261 +0,0 @@ -// This file is part of Eigen, a lightweight C++ template library -// for linear algebra. -// -// Copyright (C) 2008 Gael Guennebaud -// Copyright (C) 2009 Benoit Jacob -// -// Copyright (C) 2013 Gauthier Brun -// Copyright (C) 2013 Nicolas Carre -// Copyright (C) 2013 Jean Ceccato -// Copyright (C) 2013 Pierre Zoppitelli -// -// This Source Code Form is subject to the terms of the Mozilla -// Public License v. 2.0. If a copy of the MPL was not distributed -// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. - -// discard stack allocation as that too bypasses malloc -#define EIGEN_STACK_ALLOCATION_LIMIT 0 -#define EIGEN_RUNTIME_NO_MALLOC - -#include "main.h" -#include -#include - - -// check if "svd" is the good image of "m" -template -void svd_check_full(const MatrixType& m, const SVD& svd) -{ - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime - }; - - typedef typename MatrixType::Scalar Scalar; - typedef Matrix MatrixUType; - typedef Matrix MatrixVType; - - - MatrixType sigma = MatrixType::Zero(rows, cols); - sigma.diagonal() = svd.singularValues().template cast(); - MatrixUType u = svd.matrixU(); - MatrixVType v = svd.matrixV(); - VERIFY_IS_APPROX(m, u * sigma * v.adjoint()); - VERIFY_IS_UNITARY(u); - VERIFY_IS_UNITARY(v); -} // end svd_check_full - - - -// Compare to a reference value -template -void svd_compare_to_full(const MatrixType& m, - unsigned int computationOptions, - const SVD& referenceSvd) -{ - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - Index diagSize = (std::min)(rows, cols); - - SVD svd(m, computationOptions); - - VERIFY_IS_APPROX(svd.singularValues(), referenceSvd.singularValues()); - if(computationOptions & ComputeFullU) - VERIFY_IS_APPROX(svd.matrixU(), referenceSvd.matrixU()); - if(computationOptions & ComputeThinU) - VERIFY_IS_APPROX(svd.matrixU(), referenceSvd.matrixU().leftCols(diagSize)); - if(computationOptions & ComputeFullV) - VERIFY_IS_APPROX(svd.matrixV(), referenceSvd.matrixV()); - if(computationOptions & ComputeThinV) - VERIFY_IS_APPROX(svd.matrixV(), referenceSvd.matrixV().leftCols(diagSize)); -} // end svd_compare_to_full - - - -template -void svd_solve(const MatrixType& m, unsigned int computationOptions) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime - }; - - typedef Matrix RhsType; - typedef Matrix SolutionType; - - RhsType rhs = RhsType::Random(rows, internal::random(1, cols)); - SVD svd(m, computationOptions); - SolutionType x = svd.solve(rhs); - // evaluate normal equation which works also for least-squares solutions - VERIFY_IS_APPROX(m.adjoint()*m*x,m.adjoint()*rhs); -} // end svd_solve - - -// test computations options -// 2 functions because Jacobisvd can return before the second function -template -void svd_test_computation_options_1(const MatrixType& m, const SVD& fullSvd) -{ - svd_check_full< MatrixType, SVD >(m, fullSvd); - svd_solve< MatrixType, SVD >(m, ComputeFullU | ComputeFullV); -} - - -template -void svd_test_computation_options_2(const MatrixType& m, const SVD& fullSvd) -{ - svd_compare_to_full< MatrixType, SVD >(m, ComputeFullU, fullSvd); - svd_compare_to_full< MatrixType, SVD >(m, ComputeFullV, fullSvd); - svd_compare_to_full< MatrixType, SVD >(m, 0, fullSvd); - - if (MatrixType::ColsAtCompileTime == Dynamic) { - // thin U/V are only available with dynamic number of columns - - svd_compare_to_full< MatrixType, SVD >(m, ComputeFullU|ComputeThinV, fullSvd); - svd_compare_to_full< MatrixType, SVD >(m, ComputeThinV, fullSvd); - svd_compare_to_full< MatrixType, SVD >(m, ComputeThinU|ComputeFullV, fullSvd); - svd_compare_to_full< MatrixType, SVD >(m, ComputeThinU , fullSvd); - svd_compare_to_full< MatrixType, SVD >(m, ComputeThinU|ComputeThinV, fullSvd); - svd_solve(m, ComputeFullU | ComputeThinV); - svd_solve(m, ComputeThinU | ComputeFullV); - svd_solve(m, ComputeThinU | ComputeThinV); - - typedef typename MatrixType::Index Index; - Index diagSize = (std::min)(m.rows(), m.cols()); - SVD svd(m, ComputeThinU | ComputeThinV); - VERIFY_IS_APPROX(m, svd.matrixU().leftCols(diagSize) * svd.singularValues().asDiagonal() * svd.matrixV().leftCols(diagSize).adjoint()); - } -} - -template -void svd_verify_assert(const MatrixType& m) -{ - typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; - Index rows = m.rows(); - Index cols = m.cols(); - - enum { - RowsAtCompileTime = MatrixType::RowsAtCompileTime, - ColsAtCompileTime = MatrixType::ColsAtCompileTime - }; - - typedef Matrix RhsType; - RhsType rhs(rows); - SVD svd; - VERIFY_RAISES_ASSERT(svd.matrixU()) - VERIFY_RAISES_ASSERT(svd.singularValues()) - VERIFY_RAISES_ASSERT(svd.matrixV()) - VERIFY_RAISES_ASSERT(svd.solve(rhs)) - MatrixType a = MatrixType::Zero(rows, cols); - a.setZero(); - svd.compute(a, 0); - VERIFY_RAISES_ASSERT(svd.matrixU()) - VERIFY_RAISES_ASSERT(svd.matrixV()) - svd.singularValues(); - VERIFY_RAISES_ASSERT(svd.solve(rhs)) - - if (ColsAtCompileTime == Dynamic) - { - svd.compute(a, ComputeThinU); - svd.matrixU(); - VERIFY_RAISES_ASSERT(svd.matrixV()) - VERIFY_RAISES_ASSERT(svd.solve(rhs)) - svd.compute(a, ComputeThinV); - svd.matrixV(); - VERIFY_RAISES_ASSERT(svd.matrixU()) - VERIFY_RAISES_ASSERT(svd.solve(rhs)) - } - else - { - VERIFY_RAISES_ASSERT(svd.compute(a, ComputeThinU)) - VERIFY_RAISES_ASSERT(svd.compute(a, ComputeThinV)) - } -} - -// work around stupid msvc error when constructing at compile time an expression that involves -// a division by zero, even if the numeric type has floating point -template -EIGEN_DONT_INLINE Scalar zero() { return Scalar(0); } - -// workaround aggressive optimization in ICC -template EIGEN_DONT_INLINE T sub(T a, T b) { return a - b; } - - -template -void svd_inf_nan() -{ - // all this function does is verify we don't iterate infinitely on nan/inf values - - SVD svd; - typedef typename MatrixType::Scalar Scalar; - Scalar some_inf = Scalar(1) / zero(); - VERIFY(sub(some_inf, some_inf) != sub(some_inf, some_inf)); - svd.compute(MatrixType::Constant(10,10,some_inf), ComputeFullU | ComputeFullV); - - Scalar some_nan = zero () / zero (); - VERIFY(some_nan != some_nan); - svd.compute(MatrixType::Constant(10,10,some_nan), ComputeFullU | ComputeFullV); - - MatrixType m = MatrixType::Zero(10,10); - m(internal::random(0,9), internal::random(0,9)) = some_inf; - svd.compute(m, ComputeFullU | ComputeFullV); - - m = MatrixType::Zero(10,10); - m(internal::random(0,9), internal::random(0,9)) = some_nan; - svd.compute(m, ComputeFullU | ComputeFullV); -} - - -template -void svd_preallocate() -{ - Vector3f v(3.f, 2.f, 1.f); - MatrixXf m = v.asDiagonal(); - - internal::set_is_malloc_allowed(false); - VERIFY_RAISES_ASSERT(VectorXf v(10);) - SVD svd; - internal::set_is_malloc_allowed(true); - svd.compute(m); - VERIFY_IS_APPROX(svd.singularValues(), v); - - SVD svd2(3,3); - internal::set_is_malloc_allowed(false); - svd2.compute(m); - internal::set_is_malloc_allowed(true); - VERIFY_IS_APPROX(svd2.singularValues(), v); - VERIFY_RAISES_ASSERT(svd2.matrixU()); - VERIFY_RAISES_ASSERT(svd2.matrixV()); - svd2.compute(m, ComputeFullU | ComputeFullV); - VERIFY_IS_APPROX(svd2.matrixU(), Matrix3f::Identity()); - VERIFY_IS_APPROX(svd2.matrixV(), Matrix3f::Identity()); - internal::set_is_malloc_allowed(false); - svd2.compute(m); - internal::set_is_malloc_allowed(true); - - SVD svd3(3,3,ComputeFullU|ComputeFullV); - internal::set_is_malloc_allowed(false); - svd2.compute(m); - internal::set_is_malloc_allowed(true); - VERIFY_IS_APPROX(svd2.singularValues(), v); - VERIFY_IS_APPROX(svd2.matrixU(), Matrix3f::Identity()); - VERIFY_IS_APPROX(svd2.matrixV(), Matrix3f::Identity()); - internal::set_is_malloc_allowed(false); - svd2.compute(m, ComputeFullU|ComputeFullV); - internal::set_is_malloc_allowed(true); -} - - - - - From 60663a510a9d1c1dfb3518e95b495170ab4d66ba Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Fri, 19 Sep 2014 21:05:01 +0000 Subject: [PATCH 0807/1103] 32-bit floats/ints, 64-bit doubles pass packetmath tests, complex 32/64-bit remaining --- Eigen/src/Core/arch/AltiVec/Complex.h | 61 ++++--- Eigen/src/Core/arch/AltiVec/PacketMath.h | 221 ++++++++++++++--------- 2 files changed, 171 insertions(+), 111 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index d90cfe0de..94ecec2de 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -7,8 +7,8 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. -#ifndef EIGEN_COMPLEX_ALTIVEC_H -#define EIGEN_COMPLEX_ALTIVEC_H +#ifndef EIGEN_COMPLEX32_ALTIVEC_H +#define EIGEN_COMPLEX32_ALTIVEC_H namespace Eigen { @@ -21,16 +21,6 @@ static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4f_ZERO_, (Packet4ui)p4 static Packet2ul p2ul_CONJ_XOR = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui)p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; #endif -static Packet16uc p16uc_COMPLEX_RE = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; -static Packet16uc p16uc_COMPLEX_IM = vec_sld(p16uc_DUPLICATE, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; -static Packet16uc p16uc_COMPLEX_REV = vec_sld(p16uc_REVERSE, p16uc_REVERSE, 8);//{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; -static Packet16uc p16uc_COMPLEX_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);//{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; -static Packet16uc p16uc_PSET_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; -static Packet16uc p16uc_PSET_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_COMPLEX_RE, (Packet4ui)p16uc_COMPLEX_IM);//{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; -static Packet16uc p16uc_COMPLEX_MASK16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);//{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; -static Packet16uc p16uc_COMPLEX_TRANSPOSE_0 = vec_add(p16uc_PSET_HI, p16uc_COMPLEX_MASK16);//{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; -static Packet16uc p16uc_COMPLEX_TRANSPOSE_1 = vec_add(p16uc_PSET_LO, p16uc_COMPLEX_MASK16);//{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; - //---------- float ---------- struct Packet2cf { @@ -71,7 +61,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex((const float *)&from); else res.v = ploadu((const float *)&from); - res.v = vec_perm(res.v, res.v, p16uc_PSET_HI); + res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI); return res; } @@ -100,17 +90,24 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, con { Packet4f v1, v2; + std::cout << "a: " << a.v << ", b: " << b.v << std::endl; // Permute and multiply the real parts of a and b - v1 = vec_perm(a.v, a.v, p16uc_COMPLEX_RE); + v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD); + std::cout << "v1 = a(re): " << v1 << std::endl; // Get the imaginary parts of a - v2 = vec_perm(a.v, a.v, p16uc_COMPLEX_IM); + v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN); + std::cout << "v2 = a(im): " << v2 << std::endl; // multiply a_re * b v1 = vec_madd(v1, b.v, p4f_ZERO); + std::cout << "v1+ b : " << v1 << std::endl; // multiply a_im * b and get the conjugate result v2 = vec_madd(v2, b.v, p4f_ZERO); + std::cout << "v2+ b : " << v2 << std::endl; v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR); + std::cout << "~v2 : " << v2 << std::endl; // permute back to a proper order - v2 = vec_perm(v2, v2, p16uc_COMPLEX_REV); + v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV); + std::cout << "v2 : " << v2 << std::endl; return Packet2cf(vec_add(v1, v2)); } @@ -144,7 +141,7 @@ template<> EIGEN_STRONG_INLINE std::complex pfirst(const Pack template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { Packet4f rev_a; - rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2); + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2); return Packet2cf(rev_a); } @@ -185,7 +182,11 @@ struct palign_impl { if (Offset==1) { +#ifdef _BIG_ENDIAN first.v = vec_sld(first.v, second.v, 8); +#else + first.v = vec_sld(second.v, first.v, 8); +#endif } } }; @@ -228,18 +229,18 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, con // TODO optimize it for AltiVec Packet2cf res = conj_helper().pmul(a,b); Packet4f s = vec_madd(b.v, b.v, p4f_ZERO); - return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV)))); + return Packet2cf(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX32_REV)))); } template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip(const Packet2cf& x) { - return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); + return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV)); } EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) { - Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); - kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); + Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); kernel.packet[0].v = tmp; } @@ -289,7 +290,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd pset1(const std::complex((const double *)&from); else res.v = ploadu((const double *)&from); - res.v = vec_perm(res.v, res.v, p16uc_PSET_HI); + res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI); return res; } template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, DenseIndex stride) @@ -317,9 +318,9 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, con Packet2d a_re, a_im, v1, v2; // Permute and multiply the real parts of a and b - a_re = vec_perm(a.v, a.v, p16uc_PSET_HI); + a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI); // Get the imaginary parts of a - a_im = vec_perm(a.v, a.v, p16uc_PSET_HI); + a_im = vec_perm(a.v, a.v, p16uc_PSET64_HI); // multiply a_re * b v1 = vec_madd(a_re, b.v, p2d_ZERO_); // multiply a_im * b and get the conjugate result @@ -353,7 +354,7 @@ template<> EIGEN_STRONG_INLINE std::complex pfirst(const Pac template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { Packet2d rev_a; - rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX_REV2); + rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2); return Packet1cd(rev_a); } @@ -432,18 +433,18 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, con // TODO optimize it for AltiVec Packet1cd res = conj_helper().pmul(a,b); Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_); - return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX_REV)))); + return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX32_REV)))); } template<> EIGEN_STRONG_INLINE Packet1cd pcplxflip(const Packet1cd& x) { - return Packet1cd(vec_perm(x.v, x.v, p16uc_COMPLEX_REV)); + return Packet1cd(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV)); } EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) { - Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_0); - kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_COMPLEX_TRANSPOSE_1); + Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI); + kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); kernel.packet[0].v = tmp; } @@ -451,4 +452,4 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) } // end namespace Eigen -#endif // EIGEN_COMPLEX_ALTIVEC_H +#endif // EIGEN_COMPLEX32_ALTIVEC_H diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 0489ad886..aea3e79dc 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -62,21 +62,6 @@ typedef __vector unsigned char Packet16uc; #define DST_CHAN 1 #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride)) -// Handle endianness properly while loading constants -// Define global static constants: -#ifdef _BIG_ENDIAN -static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; -static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; -static Packet16uc p16uc_REVERSE = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; -static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); //{ 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15} -static Packet16uc p16uc_DUPLICATE = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; -#else -static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 }; -static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 }; -static Packet16uc p16uc_REVERSE = { 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15 }; -static Packet16uc p16uc_FORWARD = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; -static Packet16uc p16uc_DUPLICATE = { 4,5,6,7, 4,5,6,7, 0,1,2,3, 0,1,2,3 }; -#endif // _BIG_ENDIAN // These constants are endian-agnostic static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0} @@ -87,6 +72,40 @@ static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1} static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0} static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000} +static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; +static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; + +static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 }; +static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 }; + +// Handle endianness properly while loading constants +// Define global static constants: +#ifdef _BIG_ENDIAN +static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; +static Packet16uc p16uc_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; +#else +// Mask alignment +#define _EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0 +#define _EIGEN_ALIGNED_PTR(x) ((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT) + +static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; +static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; +static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; +static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; + +#endif // _BIG_ENDIAN + +static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; +static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; + +static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; + +static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16); //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; +static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16); //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; + template<> struct packet_traits : default_packet_traits { typedef Packet4f type; @@ -121,7 +140,19 @@ template<> struct packet_traits : default_packet_traits template<> struct unpacket_traits { typedef float type; enum {size=4}; typedef Packet4f half; }; template<> struct unpacket_traits { typedef int type; enum {size=4}; typedef Packet4i half; }; -/* + +inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v) +{ + union { + Packet16uc v; + unsigned char n[16]; + } vt; + vt.v = v; + for (int i=0; i< 16; i++) + s << (int)vt.n[i] << ", "; + return s; +} + inline std::ostream & operator <<(std::ostream & s, const Packet4f & v) { union { @@ -154,7 +185,7 @@ inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v) s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; return s; } - +/* inline std::ostream & operator <<(std::ostream & s, const Packetbi & v) { union { @@ -164,8 +195,8 @@ inline std::ostream & operator <<(std::ostream & s, const Packetbi & v) vt.v = v; s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; return s; -} -*/ +}*/ + // Need to define them first or we get specialization after instantiation errors template<> EIGEN_STRONG_INLINE Packet4f pload(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } @@ -370,12 +401,12 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) template<> EIGEN_STRONG_INLINE Packet4i ploadu(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD - return (Packet4i) vec_vsx_ld((long)from & 15, from); // align the data + return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from)); } template<> EIGEN_STRONG_INLINE Packet4f ploadu(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD - return (Packet4f) vec_vsx_ld((long)from & 15, from); // align the data + return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from)); } #endif @@ -384,16 +415,17 @@ template<> EIGEN_STRONG_INLINE Packet4f ploaddup(const float* from) Packet4f p; if((ptrdiff_t(from) % 16) == 0) p = pload(from); else p = ploadu(from); - return vec_perm(p, p, p16uc_DUPLICATE); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); } template<> EIGEN_STRONG_INLINE Packet4i ploaddup(const int* from) { Packet4i p; if((ptrdiff_t(from) % 16) == 0) p = pload(from); else p = ploadu(from); - return vec_perm(p, p, p16uc_DUPLICATE); + return vec_perm(p, p, p16uc_DUPLICATE32_HI); } +#ifdef _BIG_ENDIAN template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE @@ -430,6 +462,19 @@ template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& f vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part } +#else +// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX +template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to)); +} +template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) +{ + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to)); +} +#endif template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } @@ -437,8 +482,8 @@ template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { vec_dst template<> EIGEN_STRONG_INLINE float pfirst(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } template<> EIGEN_STRONG_INLINE int pfirst(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } -template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } -template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } +template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); } +template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); } template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); } template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); } @@ -485,7 +530,11 @@ template<> EIGEN_STRONG_INLINE int predux(const Packet4i& a) { Packet4i sum; sum = vec_sums(a, p4i_ZERO); +#ifdef _BIG_ENDIAN sum = vec_sld(sum, p4i_ZERO, 12); +#else + sum = vec_sld(p4i_ZERO, sum, 4); +#endif return pfirst(sum); } @@ -572,8 +621,25 @@ struct palign_impl { static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) { - if (Offset!=0) - first = vec_sld(first, second, Offset*4); +#ifdef _BIG_ENDIAN + switch (Offset % 4) { + case 1: + first = vec_sld(first, second, 4); break; + case 2: + first = vec_sld(first, second, 8); break; + case 3: + first = vec_sld(first, second, 12); break; + } +#else + switch (Offset % 4) { + case 1: + first = vec_sld(second, first, 12); break; + case 2: + first = vec_sld(second, first, 8); break; + case 3: + first = vec_sld(second, first, 4); break; + } +#endif } }; @@ -582,8 +648,25 @@ struct palign_impl { static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) { - if (Offset!=0) - first = vec_sld(first, second, Offset*4); +#ifdef _BIG_ENDIAN + switch (Offset % 4) { + case 1: + first = vec_sld(first, second, 4); break; + case 2: + first = vec_sld(first, second, 8); break; + case 3: + first = vec_sld(first, second, 12); break; + } +#else + switch (Offset % 4) { + case 1: + first = vec_sld(second, first, 12); break; + case 2: + first = vec_sld(second, first, 8); break; + case 3: + first = vec_sld(second, first, 4); break; + } +#endif } }; @@ -625,24 +708,15 @@ static Packet2d p2l_ONE = { 1, 1 }; static Packet2d p2d_ONE = { 1.0, 1.0 }; static Packet2d p2d_ZERO = { 0.0, 0.0 }; static Packet2d p2d_ZERO_ = { (double) 0x8000000000000000, (double) 0x8000000000000000 }; -static Packet2d p2d_COUNTDOWN = { 1.0, 0.0 }; - -#ifdef _BIG_ENDIAN -static Packet16uc p16uc_SPLATQ1 = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; -static Packet16uc p16uc_SPLATQ2 = { 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; -#else -static Packet16uc p16uc_SPLATQ1 = { 4,5,6,7, 0,1,2,3, 4,5,6,7, 0,1,3,4 }; -static Packet16uc p16uc_SPLATQ2 = { 12,13,14,15, 8,9,10,11, 12,13,14,15, 8,9,10,11 }; -#endif - +static Packet2d p2d_COUNTDOWN = { 0.0, 1.0 }; static EIGEN_STRONG_INLINE Packet2d vec_splat_dbl(Packet2d& a, int index) { switch (index) { case 0: - return (Packet2d) vec_perm(a, a, p16uc_SPLATQ1); + return (Packet2d) vec_perm(a, a, p16uc_PSET64_HI); case 1: - return (Packet2d) vec_perm(a, a, p16uc_SPLATQ2); + return (Packet2d) vec_perm(a, a, p16uc_PSET64_LO); } return a; } @@ -665,17 +739,17 @@ template<> struct packet_traits : default_packet_traits template<> struct unpacket_traits { typedef double type; enum {size=2}; typedef Packet2d half; }; -/* + inline std::ostream & operator <<(std::ostream & s, const Packet2d & v) { union { - Packetdf v; + Packet2d v; double n[2]; } vt; vt.v = v; s << vt.n[0] << ", " << vt.n[1]; return s; -}*/ +} // Need to define them first or we get specialization after instantiation errors template<> EIGEN_STRONG_INLINE Packet2d pload(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return (Packet2d) vec_ld(0, (const float *) from); } //FIXME @@ -698,7 +772,7 @@ pbroadcast4(const double *a, a1 = vec_splat_dbl(a1, 1); a3 = pload(a+2); a2 = vec_splat_dbl(a3, 0); - a3 = vec_splat_dbl(a1, 1); + a3 = vec_splat_dbl(a3, 1); } template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, DenseIndex stride) { @@ -755,43 +829,30 @@ template<> EIGEN_STRONG_INLINE Packet2d pxor(const Packet2d& a, const template<> EIGEN_STRONG_INLINE Packet2d pandnot(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); } - template<> EIGEN_STRONG_INLINE Packet2d ploadu(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD - return (Packet2d) vec_vsx_ld((int)((size_t)(from) & 15), (const float *)from); // align the data + return (Packet2d) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from)); } template<> EIGEN_STRONG_INLINE Packet2d ploaddup(const double* from) { Packet2d p; if((ptrdiff_t(from) % 16) == 0) p = pload(from); else p = ploadu(from); - return vec_perm(p, p, p16uc_DUPLICATE); + return vec_perm(p, p, p16uc_PSET64_HI); } template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& from) { - EIGEN_DEBUG_UNALIGNED_STORE - // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html - // Warning: not thread safe! - Packet16uc MSQ, LSQ, edges; - Packet16uc edgeAlign, align; - - MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword - LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword - edgeAlign = vec_lvsl(0, to); // permute map to extract edges - edges=vec_perm(LSQ,MSQ,edgeAlign); // extract the edges - align = vec_lvsr( 0, to ); // permute map to misalign data - MSQ = vec_perm(edges,(Packet16uc)from,align); // misalign the data (MSQ) - LSQ = vec_perm((Packet16uc)from,edges,align); // misalign the data (LSQ) - vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first - vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part + EIGEN_DEBUG_ALIGNED_STORE + vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to)); } + template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { vec_dstt((const float *) addr, DST_CTRL(2,2,32), DST_CHAN); } template<> EIGEN_STRONG_INLINE double pfirst(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; } -template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return (Packet2d)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return (Packet2d)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE64); } template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); } @@ -807,9 +868,10 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) { Packet2d v[2], sum; - v[0] = (Packet2d) vec_sld((Packet4ui) vecs[0], (Packet4ui) vecs[1], 8); - v[1] = (Packet2d) vec_sld((Packet4ui) vecs[1], (Packet4ui) vecs[0], 8); - sum = vec_add(v[0], v[1]); + v[0] = vec_add(vecs[0], (Packet2d) vec_sld((Packet4ui) vecs[0], (Packet4ui) vecs[0], 8)); + v[1] = vec_add(vecs[1], (Packet2d) vec_sld((Packet4ui) vecs[1], (Packet4ui) vecs[1], 8)); + + sum = (Packet2d) vec_sld((Packet4ui) v[1], (Packet4ui) v[0], 8); return sum; } @@ -817,26 +879,19 @@ template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) // mul template<> EIGEN_STRONG_INLINE double predux_mul(const Packet2d& a) { - Packet2d prod; - prod = pmul(a, (Packet2d)vec_sld((Packet4ui) a, (Packet4ui)a, 8)); - return pfirst(pmul(prod, (Packet2d)vec_sld((Packet4ui) prod, (Packet4ui) prod, 4))); + return pfirst(pmul(a, (Packet2d)vec_sld((Packet4ui) a, (Packet4ui) a, 8))); } // min template<> EIGEN_STRONG_INLINE double predux_min(const Packet2d& a) { - Packet2d b, res; - b = vec_min(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)); - res = vec_min(b, (Packet2d) vec_sld((Packet4ui) b, (Packet4ui) b, 4)); - return pfirst(res); + return pfirst(vec_min(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8))); } // max template<> EIGEN_STRONG_INLINE double predux_max(const Packet2d& a) { - Packet2d res; - res = vec_max(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)); - return pfirst(res); + return pfirst(vec_max(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8))); } template @@ -844,16 +899,20 @@ struct palign_impl { static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) { - if (Offset!=0) - first = (Packet2d) vec_sld((Packet4ui) first, (Packet4ui) second, Offset*8); + if (Offset == 1) +#ifdef _BIG_ENDIAN + first = (Packet2d) vec_sld((Packet4ui) first, (Packet4ui) second, 8); +#else + first = (Packet2d) vec_sld((Packet4ui) second, (Packet4ui) first, 8); +#endif } }; EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { Packet2d t0, t1; - t0 = (Packet2d) vec_sld((Packet4ui) kernel.packet[0], (Packet4ui) kernel.packet[1], 8); - t1 = (Packet2d) vec_sld((Packet4ui) kernel.packet[1], (Packet4ui) kernel.packet[0], 8); + t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI); + t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO); kernel.packet[0] = t0; kernel.packet[1] = t1; } From 80de35b6c562bca343fd5933bacc525dc1aeeb39 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Fri, 19 Sep 2014 22:05:18 +0100 Subject: [PATCH 0808/1103] Remove double return statement in PlainObjectBase::_set() --- Eigen/src/Core/PlainObjectBase.h | 1 - 1 file changed, 1 deletion(-) diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index 3b0e56445..0605e29aa 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -647,7 +647,6 @@ class PlainObjectBase : public internal::dense_xpr_base::type { internal::call_assignment(this->derived(), other.derived()); return this->derived(); - return this->derived(); } /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which From 10f8aabb611e7b727ff67141b030689c0a53ca78 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Sat, 20 Sep 2014 22:31:31 +0000 Subject: [PATCH 0809/1103] VSX port passes packetmath_[1-5] tests! --- Eigen/src/Core/arch/AltiVec/Complex.h | 80 ++++++++---------------- Eigen/src/Core/arch/AltiVec/PacketMath.h | 22 ++++--- 2 files changed, 38 insertions(+), 64 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 94ecec2de..58f5fbd27 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -14,11 +14,11 @@ namespace Eigen { namespace internal { -#ifdef _BIG_ENDIAN static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; +#ifdef _BIG_ENDIAN #else -static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4f_ZERO_, (Packet4ui)p4i_ZERO);//{ 0x80000000, 0x00000000, 0x80000000, 0x00000000 }; -static Packet2ul p2ul_CONJ_XOR = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui)p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; #endif //---------- float ---------- @@ -90,24 +90,17 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, con { Packet4f v1, v2; - std::cout << "a: " << a.v << ", b: " << b.v << std::endl; // Permute and multiply the real parts of a and b v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD); - std::cout << "v1 = a(re): " << v1 << std::endl; // Get the imaginary parts of a v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN); - std::cout << "v2 = a(im): " << v2 << std::endl; // multiply a_re * b v1 = vec_madd(v1, b.v, p4f_ZERO); - std::cout << "v1+ b : " << v1 << std::endl; // multiply a_im * b and get the conjugate result v2 = vec_madd(v2, b.v, p4f_ZERO); - std::cout << "v2+ b : " << v2 << std::endl; v2 = (Packet4f) vec_xor((Packet4ui)v2, p4ui_CONJ_XOR); - std::cout << "~v2 : " << v2 << std::endl; // permute back to a proper order v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV); - std::cout << "v2 : " << v2 << std::endl; return Packet2cf(vec_add(v1, v2)); } @@ -156,9 +149,13 @@ template<> EIGEN_STRONG_INLINE std::complex predux(const Packe template<> EIGEN_STRONG_INLINE Packet2cf preduxp(const Packet2cf* vecs) { Packet4f b1, b2; - +#ifdef _BIG_ENDIAN b1 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8); b2 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8); +#else + b1 = (Packet4f) vec_sld(vecs[1].v, vecs[0].v, 8); + b2 = (Packet4f) vec_sld(vecs[0].v, vecs[1].v, 8); +#endif b2 = (Packet4f) vec_sld(b2, b2, 8); b2 = padd(b1, b2); @@ -283,16 +280,8 @@ template<> EIGEN_STRONG_INLINE void pstore >(std::complex< template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } template<> EIGEN_STRONG_INLINE Packet1cd pset1(const std::complex& from) -{ - Packet1cd res; - /* On AltiVec we cannot load 64-bit registers, so wa have to take care of alignment */ - if((ptrdiff_t(&from) % 16) == 0) - res.v = pload((const double *)&from); - else - res.v = ploadu((const double *)&from); - res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI); - return res; -} +{ /* here we really have to use unaligned loads :( */ return ploadu(&from); } + template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, DenseIndex stride) { std::complex EIGEN_ALIGN16 af[2]; @@ -311,7 +300,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet1c template<> EIGEN_STRONG_INLINE Packet1cd padd(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_add(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd psub(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_sub(a.v,b.v)); } template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); } -template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR)); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); } template<> EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) { @@ -320,13 +309,13 @@ template<> EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, con // Permute and multiply the real parts of a and b a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI); // Get the imaginary parts of a - a_im = vec_perm(a.v, a.v, p16uc_PSET64_HI); + a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO); // multiply a_re * b - v1 = vec_madd(a_re, b.v, p2d_ZERO_); + v1 = vec_madd(a_re, b.v, p2d_ZERO); // multiply a_im * b and get the conjugate result - v2 = vec_madd(a_im, b.v, p2d_ZERO_); + v2 = vec_madd(a_im, b.v, p2d_ZERO); v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8); - v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR); + v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR1); return Packet1cd(vec_add(v1, v2)); } @@ -351,48 +340,31 @@ template<> EIGEN_STRONG_INLINE std::complex pfirst(const Pac return res[0]; } -template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) -{ - Packet2d rev_a; - rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2); - return Packet1cd(rev_a); -} +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } template<> EIGEN_STRONG_INLINE std::complex predux(const Packet1cd& a) { - Packet2d b; - b = (Packet2d) vec_sld((Packet4ui) a.v, (Packet4ui) a.v, 8); - b = padd(a.v, b); - return pfirst(Packet1cd(b)); + return pfirst(a); } template<> EIGEN_STRONG_INLINE Packet1cd preduxp(const Packet1cd* vecs) { - Packet2d b1, b2; - - b1 = (Packet2d) vec_sld((Packet4ui) vecs[0].v, (Packet4ui) vecs[1].v, 8); - b2 = (Packet2d) vec_sld((Packet4ui) vecs[1].v, (Packet4ui) vecs[0].v, 8); - b2 = (Packet2d) vec_sld((Packet4ui) b2, (Packet4ui) b2, 8); - b2 = padd(b1, b2); - - return Packet1cd(b2); + return vecs[0]; } template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet1cd& a) { - Packet2d b; - Packet1cd prod; - b = (Packet2d) vec_sld((Packet4ui) a.v, (Packet4ui) a.v, 8); - prod = pmul(a, Packet1cd(b)); - - return pfirst(prod); + return pfirst(a); } template struct palign_impl { - static EIGEN_STRONG_INLINE void run(Packet1cd& first, const Packet1cd& second) - { } + static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) + { + // FIXME is it sure we never have to align a Packet1cd? + // Even though a std::complex has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... + } }; template<> struct conj_helper @@ -436,9 +408,9 @@ template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, con return Packet1cd(pdiv(res.v, vec_add(s,vec_perm(s, s, p16uc_COMPLEX32_REV)))); } -template<> EIGEN_STRONG_INLINE Packet1cd pcplxflip(const Packet1cd& x) +EIGEN_STRONG_INLINE Packet1cd pcplxflip/**/(const Packet1cd& x) { - return Packet1cd(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV)); + return Packet1cd(preverse(Packet2d(x.v))); } EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index aea3e79dc..f319f4266 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -94,18 +94,21 @@ static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; - #endif // _BIG_ENDIAN static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 }; - -static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; -static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; - static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16); //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23}; static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16); //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31}; +static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 }; + +#ifdef _BIG_ENDIAN +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +#else +static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_PSET64_HI, p16uc_PSET64_LO, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; +#endif // _BIG_ENDIAN + template<> struct packet_traits : default_packet_traits { typedef Packet4f type; @@ -703,12 +706,11 @@ typedef __vector double Packet2d; typedef __vector unsigned long long Packet2ul; typedef __vector long long Packet2l; -static Packet2l p2l_ZERO = { 0x0, 0x0 }; -static Packet2d p2l_ONE = { 1, 1 }; +static Packet2l p2l_ZERO = (Packet2l) p4i_ZERO; static Packet2d p2d_ONE = { 1.0, 1.0 }; -static Packet2d p2d_ZERO = { 0.0, 0.0 }; -static Packet2d p2d_ZERO_ = { (double) 0x8000000000000000, (double) 0x8000000000000000 }; -static Packet2d p2d_COUNTDOWN = { 0.0, 1.0 }; +static Packet2d p2d_ZERO = (Packet2d) p4f_ZERO; +static Packet2d p2d_ZERO_ = { -0.0, -0.0 }; +static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ONE, (Packet16uc) p2d_ZERO, 8); static EIGEN_STRONG_INLINE Packet2d vec_splat_dbl(Packet2d& a, int index) { From c0205ca4af402eadcf1a21bc0949afa4c9d9712a Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Sun, 21 Sep 2014 08:12:22 +0000 Subject: [PATCH 0810/1103] VSX supports vec_div, implement where appropriate (float/doubles) --- Eigen/src/Core/arch/AltiVec/PacketMath.h | 22 ++++++---------------- 1 file changed, 6 insertions(+), 16 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index f319f4266..3555c521d 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -336,6 +336,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul(const Packet4i& a, const */ template<> EIGEN_STRONG_INLINE Packet4f pdiv(const Packet4f& a, const Packet4f& b) { +#ifndef __VSX__ // VSX actually provides a div instruction Packet4f t, y_0, y_1, res; // Altivec does not offer a divide instruction, we have to do a reciprocal approximation @@ -345,8 +346,10 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv(const Packet4f& a, const t = vec_nmsub(y_0, b, p4f_ONE); y_1 = vec_madd(y_0, t, y_0); - res = vec_madd(a, y_1, p4f_ZERO); - return res; + return vec_madd(a, y_1, p4f_ZERO); +#else + return vec_div(a, b); +#endif } template<> EIGEN_STRONG_INLINE Packet4i pdiv(const Packet4i& /*a*/, const Packet4i& /*b*/) @@ -801,20 +804,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return psub template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } template<> EIGEN_STRONG_INLINE Packet2d pmul(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); } -template<> EIGEN_STRONG_INLINE Packet2d pdiv(const Packet2d& a, const Packet2d& b) -{ - Packet2d t, y_0, y_1, res; - - // Altivec does not offer a divide instruction, we have to do a reciprocal approximation - y_0 = vec_re(b); - - // Do one Newton-Raphson iteration to get the needed accuracy - t = vec_nmsub(y_0, b, p2d_ONE); - y_1 = vec_madd(y_0, t, y_0); - - res = vec_madd(a, y_1, p2d_ZERO); - return res; -} +template<> EIGEN_STRONG_INLINE Packet2d pdiv(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); } // for some weird raisons, it has to be overloaded for packet of integers template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } From 974fe38ca38fb812f3ce4be105cf2ff69f40aef1 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Sun, 21 Sep 2014 11:24:30 +0000 Subject: [PATCH 0811/1103] prefetch are noops on VSX --- Eigen/src/Core/arch/AltiVec/PacketMath.h | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 3555c521d..1274f72dd 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -28,7 +28,7 @@ namespace internal { // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS -#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32 #endif typedef __vector float Packet4f; @@ -482,8 +482,16 @@ template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& } #endif -template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } -template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } +template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { +#ifndef __VSX__ + vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); +#endif +} +template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { +#ifndef __VSX__ + vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); +#endif +} template<> EIGEN_STRONG_INLINE float pfirst(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } template<> EIGEN_STRONG_INLINE int pfirst(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } From 56408504e4e3fa5f9c59d9edac14ca1ba1255e5a Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Sun, 21 Sep 2014 13:59:30 +0300 Subject: [PATCH 0812/1103] fix compile error on big endian altivec --- Eigen/src/Core/arch/AltiVec/Complex.h | 3 ++- Eigen/src/Core/arch/AltiVec/PacketMath.h | 5 +++-- 2 files changed, 5 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 58f5fbd27..30e2088ef 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -242,6 +242,7 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) } //---------- double ---------- +#ifdef __VSX__ struct Packet1cd { EIGEN_STRONG_INLINE Packet1cd() {} @@ -419,7 +420,7 @@ EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO); kernel.packet[0].v = tmp; } - +#endif // __VSX__ } // end namespace internal } // end namespace Eigen diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 1274f72dd..728f8b4d6 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -76,12 +76,12 @@ static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 }; static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 }; -static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 }; // Handle endianness properly while loading constants // Define global static constants: #ifdef _BIG_ENDIAN +static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; static Packet16uc p16uc_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; @@ -91,6 +91,7 @@ static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((P #define _EIGEN_ALIGNED_PTR(x) ((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT) static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; +static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; @@ -337,7 +338,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pdiv(const Packet4f& a, const Packet4f& b) { #ifndef __VSX__ // VSX actually provides a div instruction - Packet4f t, y_0, y_1, res; + Packet4f t, y_0, y_1; // Altivec does not offer a divide instruction, we have to do a reciprocal approximation y_0 = vec_re(b); From de38ff24993af77f188ab97d458d2884eeffd1ba Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Sun, 21 Sep 2014 11:56:07 +0000 Subject: [PATCH 0813/1103] prefetch are noops on VSX, actually disable the prefetch trait --- Eigen/src/Core/arch/AltiVec/PacketMath.h | 10 ++-------- 1 file changed, 2 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 728f8b4d6..1b86e1227 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -483,16 +483,10 @@ template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& } #endif -template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { #ifndef __VSX__ - vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); +template<> EIGEN_STRONG_INLINE void prefetch(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } +template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } #endif -} -template<> EIGEN_STRONG_INLINE void prefetch(const int* addr) { -#ifndef __VSX__ - vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); -#endif -} template<> EIGEN_STRONG_INLINE float pfirst(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } template<> EIGEN_STRONG_INLINE int pfirst(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } From 5fa69422a2d0f2d508c52e1b0eb39ed7e0d63726 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sun, 21 Sep 2014 14:19:23 +0100 Subject: [PATCH 0814/1103] Fix copy-and-paste typo in SolveWithGuess assignment This fixes compilation of code snippets in BiCGSTAB docs. --- Eigen/src/IterativeLinearSolvers/SolveWithGuess.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h index 46dd5babe..6cd4872dc 100644 --- a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h +++ b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h @@ -93,12 +93,12 @@ protected: PlainObject m_result; }; -// Specialization for "dst = dec.solve(rhs)" +// Specialization for "dst = dec.solveWithGuess(rhs)" // NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere template struct Assignment, internal::assign_op, Dense2Dense, Scalar> { - typedef Solve SrcXprType; + typedef SolveWithGuess SrcXprType; static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &) { // FIXME shall we resize dst here? From 333905b0c2e2e377744c7ed326ead34625510530 Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sun, 21 Sep 2014 14:20:08 +0100 Subject: [PATCH 0815/1103] Fix typos in docs for IterativeLinearSolvers module --- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 2 +- .../src/IterativeLinearSolvers/IterativeSolverBase.h | 12 ++++++------ 2 files changed, 7 insertions(+), 7 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index 051940dc7..42da7d14b 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -143,7 +143,7 @@ struct traits > * step execution example starting with a random guess and printing the evolution * of the estimated error: * \include BiCGSTAB_step_by_step.cpp - * Note that such a step by step excution is slightly slower. + * Note that such a step by step execution is slightly slower. * * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner */ diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h index fd9285087..1ca7e0561 100644 --- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h @@ -60,10 +60,10 @@ public: ~IterativeSolverBase() {} - /** Initializes the iterative solver for the sparcity pattern of the matrix \a A for further solving \c Ax=b problems. + /** Initializes the iterative solver for the sparsity pattern of the matrix \a A for further solving \c Ax=b problems. * - * Currently, this function mostly call analyzePattern on the preconditioner. In the future - * we might, for instance, implement column reodering for faster matrix vector products. + * Currently, this function mostly calls analyzePattern on the preconditioner. In the future + * we might, for instance, implement column reordering for faster matrix vector products. */ Derived& analyzePattern(const MatrixType& A) { @@ -76,7 +76,7 @@ public: /** Initializes the iterative solver with the numerical values of the matrix \a A for further solving \c Ax=b problems. * - * Currently, this function mostly call factorize on the preconditioner. + * Currently, this function mostly calls factorize on the preconditioner. * * \warning this class stores a reference to the matrix A as well as some * precomputed values that depend on it. Therefore, if \a A is changed @@ -95,8 +95,8 @@ public: /** Initializes the iterative solver with the matrix \a A for further solving \c Ax=b problems. * - * Currently, this function mostly initialized/compute the preconditioner. In the future - * we might, for instance, implement column reodering for faster matrix vector products. + * Currently, this function mostly initializes/computes the preconditioner. In the future + * we might, for instance, implement column reordering for faster matrix vector products. * * \warning this class stores a reference to the matrix A as well as some * precomputed values that depend on it. Therefore, if \a A is changed From abba11bdcf3e9f98a616e59570086edbf0d762b5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 22 Sep 2014 15:22:52 +0200 Subject: [PATCH 0816/1103] Many improvements in Divide&Conquer SVD: - Fix many numerical issues, in particular regarding deflation. - Add heavy debugging output to help track numerical issues (there are still fews) - Make use of Eiegn's apply-inplane-rotation feature. --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 536 +++++++++++++++++--------- 1 file changed, 364 insertions(+), 172 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index 0167872af..b13ee415e 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -11,6 +11,7 @@ // Copyright (C) 2013 Jean Ceccato // Copyright (C) 2013 Pierre Zoppitelli // Copyright (C) 2013 Jitse Niesen +// Copyright (C) 2014 Gael Guennebaud // // Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -18,7 +19,8 @@ #ifndef EIGEN_BDCSVD_H #define EIGEN_BDCSVD_H - +// #define EIGEN_BDCSVD_DEBUG_VERBOSE +// #define EIGEN_BDCSVD_SANITY_CHECKS namespace Eigen { template class BDCSVD; @@ -165,6 +167,7 @@ private: template void copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naivev); static void structured_update(Block A, const MatrixXr &B, Index n1); + static RealScalar secularEq(RealScalar x, const ArrayXr& col0, const ArrayXr& diag, const ArrayXr& diagShifted, RealScalar shift, Index n); protected: MatrixXr m_naiveU, m_naiveV; @@ -189,11 +192,12 @@ public: }; //end class BDCSVD -// Methode to allocate ans initialize matrix and attributs +// Methode to allocate ans initialize matrix and attributes template void BDCSVD::allocate(Index rows, Index cols, unsigned int computationOptions) { m_isTranspose = (cols > rows); + if (Base::allocate(rows, cols, computationOptions)) return; @@ -233,7 +237,7 @@ BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsign allocate(matrix.rows(), matrix.cols(), computationOptions); using std::abs; - //**** step 1 Bidiagonalization m_isTranspose = (matrix.cols()>matrix.rows()) ; + //**** step 1 Bidiagonalization MatrixType copy; if (m_isTranspose) copy = matrix.adjoint(); else copy = matrix; @@ -264,8 +268,11 @@ BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsign break; } } +// std::cout << "m_naiveU\n" << m_naiveU << "\n\n"; +// std::cout << "m_naiveV\n" << m_naiveV << "\n\n"; if(m_isTranspose) copyUV(bid.householderV(), bid.householderU(), m_naiveV, m_naiveU); else copyUV(bid.householderU(), bid.householderV(), m_naiveU, m_naiveV); +// std::cout << "DONE\n"; m_isInitialized = true; return *this; }// end compute @@ -278,18 +285,16 @@ void BDCSVD::copyUV(const HouseholderU &householderU, const Househol // Note exchange of U and V: m_matrixU is set from m_naiveV and vice versa if (computeU()) { - Index Ucols = m_computeThinU ? m_nonzeroSingularValues : householderU.cols(); + Index Ucols = m_computeThinU ? m_diagSize : householderU.cols(); m_matrixU = MatrixX::Identity(householderU.cols(), Ucols); - Index blockCols = m_computeThinU ? m_nonzeroSingularValues : m_diagSize; - m_matrixU.topLeftCorner(m_diagSize, blockCols) = naiveV.template cast().topLeftCorner(m_diagSize, blockCols); + m_matrixU.topLeftCorner(m_diagSize, m_diagSize) = naiveV.template cast().topLeftCorner(m_diagSize, m_diagSize); householderU.applyThisOnTheLeft(m_matrixU); } if (computeV()) { - Index Vcols = m_computeThinV ? m_nonzeroSingularValues : householderV.cols(); + Index Vcols = m_computeThinV ? m_diagSize : householderV.cols(); m_matrixV = MatrixX::Identity(householderV.cols(), Vcols); - Index blockCols = m_computeThinV ? m_nonzeroSingularValues : m_diagSize; - m_matrixV.topLeftCorner(m_diagSize, blockCols) = naiveU.template cast().topLeftCorner(m_diagSize, blockCols); + m_matrixV.topLeftCorner(m_diagSize, m_diagSize) = naiveU.template cast().topLeftCorner(m_diagSize, m_diagSize); householderV.applyThisOnTheLeft(m_matrixV); } } @@ -308,7 +313,7 @@ void BDCSVD::structured_update(Block A, co Index n = A.rows(); if(n>100) { - // If the matrices are large enough, let's exploit the sparse strucure of A by + // If the matrices are large enough, let's exploit the sparse structure of A by // splitting it in half (wrt n1), and packing the non-zero columns. DenseIndex n2 = n - n1; MatrixXr A1(n1,n), A2(n2,n), B1(n,n), B2(n,n); @@ -385,6 +390,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, // right submatrix before the left one. divide(k + 1 + firstCol, lastCol, k + 1 + firstRowW, k + 1 + firstColW, shift); divide(firstCol, k - 1 + firstCol, firstRowW, firstColW + 1, shift + 1); + if (m_compU) { lambda = m_naiveU(firstCol + k, firstCol + k); @@ -417,6 +423,13 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, c0 = alphaK * lambda / r0; s0 = betaK * phi / r0; } + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(m_naiveU.allFinite()); + assert(m_naiveV.allFinite()); + assert(m_computed.allFinite()); +#endif + if (m_compU) { MatrixXr q1 (m_naiveU.col(firstCol + k).segment(firstCol, k + 1)); @@ -449,6 +462,13 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, m_naiveU.row(1).segment(firstCol + 1, k).setZero(); m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1).setZero(); } + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(m_naiveU.allFinite()); + assert(m_naiveV.allFinite()); + assert(m_computed.allFinite()); +#endif + m_computed(firstCol + shift, firstCol + shift) = r0; m_computed.col(firstCol + shift).segment(firstCol + shift + 1, k) = alphaK * l.transpose().real(); m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) = betaK * f.transpose().real(); @@ -461,11 +481,22 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, VectorType singVals; computeSVDofM(firstCol + shift, n, UofSVD, singVals, VofSVD); +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(UofSVD.allFinite()); + assert(VofSVD.allFinite()); +#endif + if (m_compU) structured_update(m_naiveU.block(firstCol, firstCol, n + 1, n + 1), UofSVD, (n+2)/2); else m_naiveU.middleCols(firstCol, n + 1) *= UofSVD; // FIXME this requires a temporary, and exploit that there are 2 rows at compile time if (m_compV) structured_update(m_naiveV.block(firstRowW, firstColW, n, n), VofSVD, (n+1)/2); +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(m_naiveU.allFinite()); + assert(m_naiveV.allFinite()); + assert(m_computed.allFinite()); +#endif + m_computed.block(firstCol + shift, firstCol + shift, n, n).setZero(); m_computed.block(firstCol + shift, firstCol + shift, n, n).diagonal() = singVals; }// end divide @@ -491,18 +522,78 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec U.resize(n+1, n+1); if (m_compV) V.resize(n, n); - if (col0.hasNaN() || diag.hasNaN()) return; + if (col0.hasNaN() || diag.hasNaN()) { std::cout << "\n\nHAS NAN\n\n"; return; } - ArrayXr shifts(n), mus(n), zhat(n); - computeSingVals(col0, diag, singVals, shifts, mus); - perturbCol0(col0, diag, singVals, shifts, mus, zhat); - computeSingVecs(zhat, diag, singVals, shifts, mus, U, V); + ArrayXr shifts(n), mus(n), zhat(n); +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "computeSVDofM using:\n"; + std::cout << " z: " << col0.transpose() << "\n"; + std::cout << " d: " << diag.transpose() << "\n"; +#endif + + // Compute singVals, shifts, and mus + computeSingVals(col0, diag, singVals, shifts, mus); + +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << " sing-val: " << singVals.transpose() << "\n"; + std::cout << " mu: " << mus.transpose() << "\n"; + std::cout << " shift: " << shifts.transpose() << "\n"; +#endif + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(singVals.allFinite()); + assert(mus.allFinite()); + assert(shifts.allFinite()); +#endif + + // Compute zhat + perturbCol0(col0, diag, singVals, shifts, mus, zhat); +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << " zhat: " << zhat.transpose() << "\n"; + { + Index actual_n = n; + while(actual_n>1 && col0(actual_n-1)==0) --actual_n; + std::cout << "\n\n mus: " << mus.head(actual_n).transpose() << "\n\n"; + std::cout << " check1: " << ((singVals.array()-(shifts+mus)) / singVals.array()).head(actual_n).transpose() << "\n\n"; + std::cout << " check2: " << ((singVals.array()-diag) / singVals.array()).head(actual_n).transpose() << "\n\n"; + std::cout << " check3: " << ((diag.segment(1,actual_n-1)-singVals.head(actual_n-1).array()) / singVals.head(actual_n-1).array()).transpose() << "\n\n\n"; + } +#endif + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(zhat.allFinite()); +#endif + + computeSingVecs(zhat, diag, singVals, shifts, mus, U, V); + +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "U^T U: " << (U.transpose() * U - MatrixType(MatrixType::Identity(U.cols(),U.cols()))).norm() << "\n"; + std::cout << "V^T V: " << (V.transpose() * V - MatrixType(MatrixType::Identity(V.cols(),V.cols()))).norm() << "\n"; +#endif + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert((U.transpose() * U - MatrixType(MatrixType::Identity(U.cols(),U.cols()))).norm() < 1e-14 * n); + assert((V.transpose() * V - MatrixType(MatrixType::Identity(V.cols(),V.cols()))).norm() < 1e-14 * n); + assert(m_naiveU.allFinite()); + assert(m_naiveV.allFinite()); + assert(m_computed.allFinite()); +#endif + // Reverse order so that singular values in increased order - singVals.reverseInPlace(); - U.leftCols(n) = U.leftCols(n).rowwise().reverse().eval(); // FIXME this requires a temporary - if (m_compV) V = V.rowwise().reverse().eval(); // FIXME this requires a temporary + // Because of deflation, the zeros singular-values are already at the end + Index actual_n = n; + while(actual_n>1 && singVals(actual_n-1)==0) --actual_n; + singVals.head(actual_n).reverseInPlace(); + U.leftCols(actual_n) = U.leftCols(actual_n).rowwise().reverse().eval(); // FIXME this requires a temporary + if (m_compV) V.leftCols(actual_n) = V.leftCols(actual_n).rowwise().reverse().eval(); // FIXME this requires a temporary +} + +template +typename BDCSVD::RealScalar BDCSVD::secularEq(RealScalar mu, const ArrayXr& col0, const ArrayXr& diag, const ArrayXr& diagShifted, RealScalar shift, Index n) +{ + return 1 + (col0.square() / ((diagShifted - mu) )/( (diag + shift + mu))).head(n).sum(); } template @@ -514,6 +605,19 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia using std::max; Index n = col0.size(); + Index actual_n = n; + while(actual_n>1 && col0(actual_n-1)==0) --actual_n; + Index m = 0; + Array perm(actual_n); + { + for(Index k=0;k::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // otherwise, use secular equation to find singular value RealScalar left = diag(k); - RealScalar right = (k != n-1) ? diag(k+1) : (diag(n-1) + col0.matrix().norm()); + RealScalar right = (k != actual_n-1) ? diag(k+1) : (diag(actual_n-1) + col0.matrix().norm()); // first decide whether it's closer to the left end or the right end RealScalar mid = left + (right-left) / 2; - RealScalar fMid = 1 + (col0.square() / ((diag + mid) * (diag - mid))).sum(); + RealScalar fMid = 1 + (col0.square() / ((diag + mid) * (diag - mid))).head(actual_n).sum(); - RealScalar shift = (k == n-1 || fMid > 0) ? left : right; + RealScalar shift = (k == actual_n-1 || fMid > 0) ? left : right; // measure everything relative to shift ArrayXr diagShifted = diag - shift; @@ -543,7 +647,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia if (shift == left) { muPrev = (right - left) * 0.1; - if (k == n-1) muCur = right - left; + if (k == actual_n-1) muCur = right - left; else muCur = (right - left) * 0.5; } else @@ -552,8 +656,8 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia muCur = -(right - left) * 0.5; } - RealScalar fPrev = 1 + (col0.square() / ((diagShifted - muPrev) * (diag + shift + muPrev))).sum(); - RealScalar fCur = 1 + (col0.square() / ((diagShifted - muCur) * (diag + shift + muCur))).sum(); + RealScalar fPrev = secularEq(muPrev, col0, diag, diagShifted, shift, actual_n); + RealScalar fCur = secularEq(muCur, col0, diag, diagShifted, shift, actual_n); if (abs(fPrev) < abs(fCur)) { swap(fPrev, fCur); @@ -563,7 +667,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // rational interpolation: fit a function of the form a / mu + b through the two previous // iterates and use its zero to compute the next iterate bool useBisection = false; - while (abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (max)(abs(muCur), abs(muPrev)) && fCur != fPrev && !useBisection) + while (abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (max)(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) { ++m_numIters; @@ -573,7 +677,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia muPrev = muCur; fPrev = fCur; muCur = -a / b; - fCur = 1 + (col0.square() / ((diagShifted - muCur) * (diag + shift + muCur))).sum(); + fCur = secularEq(muCur, col0, diag, diagShifted, shift, actual_n); if (shift == left && (muCur < 0 || muCur > right - left)) useBisection = true; if (shift == right && (muCur < -(right - left) || muCur > 0)) useBisection = true; @@ -595,14 +699,19 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia rightShifted = -1e-30; } - RealScalar fLeft = 1 + (col0.square() / ((diagShifted - leftShifted) * (diag + shift + leftShifted))).sum(); - RealScalar fRight = 1 + (col0.square() / ((diagShifted - rightShifted) * (diag + shift + rightShifted))).sum(); - assert(fLeft * fRight < 0); + RealScalar fLeft = secularEq(leftShifted, col0, diag, diagShifted, shift, actual_n); + RealScalar fRight = secularEq(rightShifted, col0, diag, diagShifted, shift, actual_n); + +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + if(fLeft * fRight>0) + std::cout << fLeft << " * " << fRight << " == " << fLeft * fRight << " ; " << left << " - " << right << " -> " << leftShifted << " " << rightShifted << " shift=" << shift << "\n"; +#endif + eigen_internal_assert(fLeft * fRight < 0); while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (max)(abs(leftShifted), abs(rightShifted))) { RealScalar midShifted = (leftShifted + rightShifted) / 2; - RealScalar fMid = 1 + (col0.square() / ((diagShifted - midShifted) * (diag + shift + midShifted))).sum(); + RealScalar fMid = secularEq(midShifted, col0, diag, diagShifted, shift, actual_n); if (fLeft * fMid < 0) { rightShifted = midShifted; @@ -638,27 +747,53 @@ void BDCSVD::perturbCol0 { using std::sqrt; Index n = col0.size(); + + // Ignore trailing zeros: + Index actual_n = n; + while(actual_n>1 && col0(actual_n-1)==0) --actual_n; + // Deflated non-zero singular values are kept in-place, + // we thus compute an indirection array to properly ignore all deflated entries. + // TODO compute it once! + Index m = 0; // size of the deflated problem + Array perm(actual_n); + { + for(Index k=0;k 0) zhat(k) = tmp; - else zhat(k) = -tmp; + RealScalar dk = diag(k); + RealScalar prod = (singVals(actual_n-1) + dk) * (mus(actual_n-1) + (shifts(actual_n-1) - dk)); + + for(Index l = 0; l 0.9 ) + std::cout << " " << ((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) << " == (" << (singVals(j)+dk) << " * " << (mus(j)+(shifts(j)-dk)) << ") / (" << (diag(i)+dk) << " * " << (diag(i)-dk) << ")\n"; +#endif + } + } +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "zhat(" << k << ") = sqrt( " << prod << ") ; " << (singVals(actual_n-1) + dk) << " * " << mus(actual_n-1) + shifts(actual_n-1) << " - " << dk << "\n"; +#endif + RealScalar tmp = sqrt(prod); + zhat(k) = col0(k) > 0 ? tmp : -tmp; } } } @@ -670,6 +805,23 @@ void BDCSVD::computeSingVecs const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V) { Index n = zhat.size(); + + // Deflated non-zero singular values are kept in-place, + // we thus compute an indirection array to properly ignore all deflated entries. + // TODO compute it once! + Index actual_n = n; + while(actual_n>1 && zhat(actual_n-1)==0) --actual_n; + Index m = 0; + Array perm(actual_n); + { + for(Index k=0;k::computeSingVecs } else { - U.col(k).head(n) = zhat / (((diag - shifts(k)) - mus(k)) * (diag + singVals[k])); - U(n,k) = 0; + U.col(k).setZero(); + for(Index l=0;l::deflation43(Index firstCol, Index shift, Index i, Index using std::abs; using std::sqrt; using std::pow; - RealScalar c = m_computed(firstCol + shift, firstCol + shift); - RealScalar s = m_computed(i, firstCol + shift); - RealScalar r = sqrt(pow(abs(c), 2) + pow(abs(s), 2)); + Index start = firstCol + shift; + RealScalar c = m_computed(start, start); + RealScalar s = m_computed(start+i, start); + RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s)); if (r == 0) { - m_computed(i, i) = 0; + m_computed(start+i, start+i) = 0; return; } - c/=r; - s/=r; - m_computed(firstCol + shift, firstCol + shift) = r; - m_computed(i, firstCol + shift) = 0; - m_computed(i, i) = 0; - if (m_compU) - { - m_naiveU.col(firstCol).segment(firstCol,size) = - c * m_naiveU.col(firstCol).segment(firstCol, size) - - s * m_naiveU.col(i).segment(firstCol, size) ; - - m_naiveU.col(i).segment(firstCol, size) = - (c + s*s/c) * m_naiveU.col(i).segment(firstCol, size) + - (s/c) * m_naiveU.col(firstCol).segment(firstCol,size); - } + m_computed(start,start) = r; + m_computed(start+i, start) = 0; + m_computed(start+i, start+i) = 0; + + JacobiRotation J(c/r,-s/r); + if (m_compU) m_naiveU.middleRows(firstCol, size+1).applyOnTheRight(firstCol, firstCol+i, J); + else m_naiveU.applyOnTheRight(firstCol, firstCol+i, J); }// end deflation 43 // page 13 // i,j >= 1, i != j and |di - dj| < epsilon * norm2(M) // We apply two rotations to have zj = 0; +// TODO deflation44 is still broken and not properly tested template void BDCSVD::deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size) { @@ -740,9 +898,20 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi using std::sqrt; using std::conj; using std::pow; - RealScalar c = m_computed(firstColm, firstColm + j - 1); - RealScalar s = m_computed(firstColm, firstColm + i - 1); - RealScalar r = sqrt(pow(abs(c), 2) + pow(abs(s), 2)); + RealScalar c = m_computed(firstColm+i, firstColm); + RealScalar s = m_computed(firstColm+j, firstColm); + RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s)); +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "deflation 4.4: " << i << "," << j << " -> " << c << " " << s << " " << r << " ; " + << m_computed(firstColm + i-1, firstColm) << " " + << m_computed(firstColm + i, firstColm) << " " + << m_computed(firstColm + i+1, firstColm) << " " + << m_computed(firstColm + i+2, firstColm) << "\n"; + std::cout << m_computed(firstColm + i-1, firstColm + i-1) << " " + << m_computed(firstColm + i, firstColm+i) << " " + << m_computed(firstColm + i+1, firstColm+i+1) << " " + << m_computed(firstColm + i+2, firstColm+i+2) << "\n"; +#endif if (r==0) { m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j); @@ -753,25 +922,15 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi m_computed(firstColm + i, firstColm) = r; m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j); m_computed(firstColm + j, firstColm) = 0; + + JacobiRotation J(c,s); if (m_compU) { - m_naiveU.col(firstColu + i).segment(firstColu, size) = - c * m_naiveU.col(firstColu + i).segment(firstColu, size) - - s * m_naiveU.col(firstColu + j).segment(firstColu, size) ; - - m_naiveU.col(firstColu + j).segment(firstColu, size) = - (c + s*s/c) * m_naiveU.col(firstColu + j).segment(firstColu, size) + - (s/c) * m_naiveU.col(firstColu + i).segment(firstColu, size); + m_naiveU.middleRows(firstColu, size).applyOnTheRight(firstColu + i, firstColu + j, J); } if (m_compV) { - m_naiveV.col(firstColW + i).segment(firstRowW, size - 1) = - c * m_naiveV.col(firstColW + i).segment(firstRowW, size - 1) + - s * m_naiveV.col(firstColW + j).segment(firstRowW, size - 1) ; - - m_naiveV.col(firstColW + j).segment(firstRowW, size - 1) = - (c + s*s/c) * m_naiveV.col(firstColW + j).segment(firstRowW, size - 1) - - (s/c) * m_naiveV.col(firstColW + i).segment(firstRowW, size - 1); + m_naiveU.middleRows(firstRowW, size-1).applyOnTheRight(firstColW + i, firstColW + j, J.transpose()); } }// end deflation 44 @@ -780,104 +939,137 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi template void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift) { - //condition 4.1 using std::sqrt; using std::abs; + using std::max; const Index length = lastCol + 1 - firstCol; - RealScalar norm1 = m_computed.block(firstCol+shift, firstCol+shift, length, 1).squaredNorm(); - RealScalar norm2 = m_computed.block(firstCol+shift, firstCol+shift, length, length).diagonal().squaredNorm(); - RealScalar epsilon = 10 * NumTraits::epsilon() * sqrt(norm1 + norm2); - if (m_computed(firstCol + shift, firstCol + shift) < epsilon) - m_computed(firstCol + shift, firstCol + shift) = epsilon; + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(m_naiveU.allFinite()); + assert(m_naiveV.allFinite()); + assert(m_computed.allFinite()); +#endif + + Block col0(m_computed, firstCol+shift, firstCol+shift, length, 1); + Diagonal fulldiag(m_computed); + VectorBlock,Dynamic> diag(fulldiag, firstCol+shift, length); + + RealScalar epsilon = 8 * NumTraits::epsilon() * (max)(col0.cwiseAbs().maxCoeff(), diag.cwiseAbs().maxCoeff()); + + //condition 4.1 + if (diag(0) < epsilon) + { +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "deflation 4.1, because " << diag(0) << " < " << epsilon << "\n"; +#endif + diag(0) = epsilon; + } //condition 4.2 - for (Index i=firstCol + shift + 1;i<=lastCol + shift;i++) - if (abs(m_computed(i, firstCol + shift)) < epsilon) - m_computed(i, firstCol + shift) = 0; + for (Index i=1;i firstCol + shift + k) permutation[p] = j++; - else if (j> lastCol + shift) permutation[p] = i++; - else if (m_computed(i, i) < m_computed(j, j)) permutation[p] = j++; - else permutation[p] = i++; - } - //we do the permutation - RealScalar aux; - //we stock the current index of each col - //and the column of each index - Index *realInd = new Index[length]; // FIXME avoid repeated dynamic memory allocation - Index *realCol = new Index[length]; // FIXME avoid repeated dynamic memory allocation - for (int pos = 0; pos< length; pos++) - { - realCol[pos] = pos + firstCol + shift; - realInd[pos] = pos; - } - const Index Zero = firstCol + shift; - VectorType temp; - for (int i = 1; i < length - 1; i++) - { - const Index I = i + Zero; - const Index realI = realInd[i]; - const Index j = permutation[length - i] - Zero; - const Index J = realCol[j]; - - //diag displace - aux = m_computed(I, I); - m_computed(I, I) = m_computed(J, J); - m_computed(J, J) = aux; - - //firstrow displace - aux = m_computed(I, Zero); - m_computed(I, Zero) = m_computed(J, Zero); - m_computed(J, Zero) = aux; - - // change columns - if (m_compU) - { - temp = m_naiveU.col(I - shift).segment(firstCol, length + 1); - m_naiveU.col(I - shift).segment(firstCol, length + 1) = m_naiveU.col(J - shift).segment(firstCol, length + 1); - m_naiveU.col(J - shift).segment(firstCol, length + 1) = temp; - } - else - { - temp = m_naiveU.col(I - shift).segment(0, 2); - m_naiveU.col(I - shift).template head<2>() = m_naiveU.col(J - shift).segment(0, 2); - m_naiveU.col(J - shift).template head<2>() = temp; } - if (m_compV) - { - const Index CWI = I + firstColW - Zero; - const Index CWJ = J + firstColW - Zero; - temp = m_naiveV.col(CWI).segment(firstRowW, length); - m_naiveV.col(CWI).segment(firstRowW, length) = m_naiveV.col(CWJ).segment(firstRowW, length); - m_naiveV.col(CWJ).segment(firstRowW, length) = temp; - } - - //update real pos - realCol[realI] = J; - realCol[j] = I; - realInd[J - Zero] = realI; - realInd[I - Zero] = j; - } - for (Index i = firstCol + shift + 1; i k) permutation[p] = j++; + else if (j >= length) permutation[p] = i++; + else if (diag(i) < diag(j)) permutation[p] = j++; + else permutation[p] = i++; + } + } + + // Current index of each col, and current column of each index + Index *realInd = new Index[length]; // FIXME avoid repeated dynamic memory allocation + Index *realCol = new Index[length]; // FIXME avoid repeated dynamic memory allocation + + for(int pos = 0; pos< length; pos++) + { + realCol[pos] = pos; + realInd[pos] = pos; + } + + for(Index i = 1; i < length - 1; i++) + { + const Index pi = permutation[length - i]; + const Index J = realCol[pi]; + + using std::swap; + // swap diaognal and first column entries: + swap(diag(i), diag(J)); + swap(col0(i), col0(J)); + + // change columns + if (m_compU) m_naiveU.col(firstCol+i).segment(firstCol, length + 1).swap(m_naiveU.col(firstCol+J).segment(firstCol, length + 1)); + else m_naiveU.col(firstCol+i).segment(0, 2) .swap(m_naiveU.col(firstCol+J).segment(0, 2)); + if (m_compV) m_naiveV.col(firstColW + i).segment(firstRowW, length).swap(m_naiveV.col(firstColW + J).segment(firstRowW, length)); + + //update real pos + const Index realI = realInd[i]; + realCol[realI] = J; + realCol[pi] = i; + realInd[J] = realI; + realInd[i] = pi; + } + delete[] permutation; + delete[] realInd; + delete[] realCol; + } + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + for(int k=2;k::epsilon()*diag(i+1)) +// if ((diag(i+1) - diag(i)) < epsilon) + { +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "deflation 4.4 with i = " << i << " because " << (diag(i+1) - diag(i)) << " < " << epsilon << "\n"; +#endif + deflation44(firstCol, firstCol + shift, firstRowW, firstColW, i, i + 1, length); + } + +#ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(m_naiveU.allFinite()); + assert(m_naiveV.allFinite()); + assert(m_computed.allFinite()); +#endif }//end deflation From f9d6d3780f1709cc8e3722b55963dc2359cf414f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 22 Sep 2014 17:34:17 +0200 Subject: [PATCH 0817/1103] bug #879: fix compilation of tri1=mat*tri2 by copying tri2 into a full temporary. --- Eigen/src/Core/TriangularMatrix.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 0d315dd50..36f04a5e8 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -207,7 +207,8 @@ template class TriangularView TransposeMode = (Mode & Upper ? Lower : 0) | (Mode & Lower ? Upper : 0) | (Mode & (UnitDiag)) - | (Mode & (ZeroDiag)) + | (Mode & (ZeroDiag)), + IsVectorAtCompileTime = false }; EIGEN_DEVICE_FUNC From ae514ddfe561ef220bc9bbf8c0b7b5005b60d9c8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 22 Sep 2014 22:49:20 +0200 Subject: [PATCH 0818/1103] bug #880: manually edit the DartConfiguration.tcl file to get it working with cmake 3.0.x --- cmake/EigenConfigureTesting.cmake | 45 +++++-------------------------- 1 file changed, 7 insertions(+), 38 deletions(-) diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake index 0b5de95bb..190509958 100644 --- a/cmake/EigenConfigureTesting.cmake +++ b/cmake/EigenConfigureTesting.cmake @@ -14,49 +14,18 @@ add_dependencies(check buildtests) # check whether /bin/bash exists find_file(EIGEN_BIN_BASH_EXISTS "/bin/bash" PATHS "/" NO_DEFAULT_PATH) -# CMake/Ctest does not allow us to change the build command, -# so we have to workaround by directly editing the generated DartConfiguration.tcl file -# save CMAKE_MAKE_PROGRAM -set(CMAKE_MAKE_PROGRAM_SAVE ${CMAKE_MAKE_PROGRAM}) -# and set a fake one -set(CMAKE_MAKE_PROGRAM "@EIGEN_MAKECOMMAND_PLACEHOLDER@") - # This call activates testing and generates the DartConfiguration.tcl include(CTest) set(EIGEN_TEST_BUILD_FLAGS "" CACHE STRING "Options passed to the build command of unit tests") -# overwrite default DartConfiguration.tcl -# The worarounds are different for each version of the MSVC IDE -set(EIGEN_TEST_TARGET buildtests) -if(MSVC_IDE) - if(CMAKE_MAKE_PROGRAM_SAVE MATCHES "devenv") # devenv - set(EIGEN_BUILD_COMMAND "${CMAKE_MAKE_PROGRAM_SAVE} Eigen.sln /build Release /project ${EIGEN_TEST_TARGET}") - else() # msbuild - set(EIGEN_BUILD_COMMAND "${CMAKE_MAKE_PROGRAM_SAVE} ${EIGEN_TEST_TARGET}.vcxproj /p:Configuration=\${CTEST_CONFIGURATION_TYPE}") - endif() - - # append the build flags if provided - if(NOT "${EIGEN_TEST_BUILD_FLAGS}" MATCHES "^[ \t]*$") - set(EIGEN_BUILD_COMMAND "${EIGEN_BUILD_COMMAND} ${EIGEN_TEST_BUILD_FLAGS}") - endif() - - # apply the dartconfig hack ... - set(EIGEN_MAKECOMMAND_PLACEHOLDER "${EIGEN_BUILD_COMMAND}\n#") -else() - # for make and nmake - set(EIGEN_BUILD_COMMAND "${CMAKE_MAKE_PROGRAM_SAVE} ${EIGEN_TEST_TARGET} ${EIGEN_TEST_BUILD_FLAGS}") - set(EIGEN_MAKECOMMAND_PLACEHOLDER "${EIGEN_BUILD_COMMAND}") -endif() - -configure_file(${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl ${CMAKE_BINARY_DIR}/DartConfiguration.tcl) - -# restore default CMAKE_MAKE_PROGRAM -set(CMAKE_MAKE_PROGRAM ${CMAKE_MAKE_PROGRAM_SAVE}) - -# un-set temporary variables so that it is like they never existed -unset(CMAKE_MAKE_PROGRAM_SAVE) -unset(EIGEN_MAKECOMMAND_PLACEHOLDER) +# Overwrite default DartConfiguration.tcl such that ctest can build our unit tests. +# Recall that our unit tests are not in the "all" target, so we have to explicitely ask ctest to build our custom 'buildtests' target. +# At this stage, we can also add custom flags to the build tool through the user defined EIGEN_TEST_BUILD_FLAGS variable. +file(READ "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" EIGEN_DART_CONFIG_FILE) +string(REGEX REPLACE "MakeCommand:.*DefaultCTestConfigurationType" "MakeCommand: ${CMAKE_COMMAND} --build . --target buildtests --config '' -- -i ${EIGEN_TEST_BUILD_FLAGS}\nDefaultCTestConfigurationType" + EIGEN_DART_CONFIG_FILE2 ${EIGEN_DART_CONFIG_FILE}) +file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" ${EIGEN_DART_CONFIG_FILE2}) configure_file(${CMAKE_CURRENT_SOURCE_DIR}/CTestCustom.cmake.in ${CMAKE_BINARY_DIR}/CTestCustom.cmake) From ff46ec0f240ef84e2293b33b265c703e9b765c2e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 22 Sep 2014 23:33:28 +0200 Subject: [PATCH 0819/1103] bug #881: make SparseMatrixBase::isApprox(SparseMatrixBase) exploits sparse computations instead of converting the operands to dense matrices. --- Eigen/src/SparseCore/SparseFuzzy.h | 30 ++++++++++++++----------- Eigen/src/SparseCore/SparseMatrixBase.h | 3 +-- test/sparse_basic.cpp | 4 ++++ 3 files changed, 22 insertions(+), 15 deletions(-) diff --git a/Eigen/src/SparseCore/SparseFuzzy.h b/Eigen/src/SparseCore/SparseFuzzy.h index 45f36e9eb..3e67cbf5f 100644 --- a/Eigen/src/SparseCore/SparseFuzzy.h +++ b/Eigen/src/SparseCore/SparseFuzzy.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -10,17 +10,21 @@ #ifndef EIGEN_SPARSE_FUZZY_H #define EIGEN_SPARSE_FUZZY_H -// template -// template -// bool SparseMatrixBase::isApprox( -// const OtherDerived& other, -// typename NumTraits::Real prec -// ) const -// { -// const typename internal::nested::type nested(derived()); -// const typename internal::nested::type otherNested(other.derived()); -// return (nested - otherNested).cwise().abs2().sum() -// <= prec * prec * (std::min)(nested.cwise().abs2().sum(), otherNested.cwise().abs2().sum()); -// } +namespace Eigen { + +template +template +bool SparseMatrixBase::isApprox(const SparseMatrixBase& other, const RealScalar &prec) const +{ + using std::min; + const typename internal::nested_eval::type actualA(derived()); + typename internal::conditional::type, + const PlainObject>::type actualB(other.derived()); + + return (actualA - actualB).squaredNorm() <= prec * prec * (min)(actualA.squaredNorm(), actualB.squaredNorm()); +} + +} // end namespace Eigen #endif // EIGEN_SPARSE_FUZZY_H diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index b5c50d93a..a438c6487 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -324,8 +324,7 @@ template class SparseMatrixBase : public EigenBase template bool isApprox(const SparseMatrixBase& other, - const RealScalar& prec = NumTraits::dummy_precision()) const - { return toDense().isApprox(other.toDense(),prec); } + const RealScalar& prec = NumTraits::dummy_precision()) const; template bool isApprox(const MatrixBase& other, diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index c86534bad..e5b6d5a0a 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -304,6 +304,10 @@ template void sparse_basic(const SparseMatrixType& re VERIFY_IS_APPROX(m2.transpose(), refMat2.transpose()); VERIFY_IS_APPROX(SparseMatrixType(m2.adjoint()), refMat2.adjoint()); + + // check isApprox handles opposite storage order + typename Transpose::PlainObject m3(m2); + VERIFY(m2.isApprox(m3)); } From 3878e6f1703c22782ef9897fe46c7f1fc6a47ae5 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 23 Sep 2014 10:25:12 +0200 Subject: [PATCH 0820/1103] Add a true ctest unit test for failtests --- test/CMakeLists.txt | 3 +++ 1 file changed, 3 insertions(+) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 530e9e4e1..4132e6481 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -324,3 +324,6 @@ endif(CUDA_FOUND) endif(EIGEN_TEST_NVCC) + +file(MAKE_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/failtests) +add_test(NAME failtests WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/failtests COMMAND ${CMAKE_COMMAND} ${Eigen_SOURCE_DIR} -G "${CMAKE_GENERATOR}" -DEIGEN_FAILTEST=ON) From 72569f17ecbc37309291aa7f333c8236aba8ee3f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 23 Sep 2014 10:26:02 +0200 Subject: [PATCH 0821/1103] bug #882: add const-correctness failtests for CwiseUnaryView, TriangularView, and SelfAdjointView. --- failtest/CMakeLists.txt | 6 ++++++ ...cwiseunaryview_nonconst_ctor_on_const_xpr.cpp | 15 +++++++++++++++ ...iseunaryview_on_const_type_actually_const.cpp | 16 ++++++++++++++++ ...elfadjointview_nonconst_ctor_on_const_xpr.cpp | 15 +++++++++++++++ ...fadjointview_on_const_type_actually_const.cpp | 16 ++++++++++++++++ ...triangularview_nonconst_ctor_on_const_xpr.cpp | 15 +++++++++++++++ ...iangularview_on_const_type_actually_const.cpp | 16 ++++++++++++++++ 7 files changed, 99 insertions(+) create mode 100644 failtest/cwiseunaryview_nonconst_ctor_on_const_xpr.cpp create mode 100644 failtest/cwiseunaryview_on_const_type_actually_const.cpp create mode 100644 failtest/selfadjointview_nonconst_ctor_on_const_xpr.cpp create mode 100644 failtest/selfadjointview_on_const_type_actually_const.cpp create mode 100644 failtest/triangularview_nonconst_ctor_on_const_xpr.cpp create mode 100644 failtest/triangularview_on_const_type_actually_const.cpp diff --git a/failtest/CMakeLists.txt b/failtest/CMakeLists.txt index 5afa2ac82..83b6c81a9 100644 --- a/failtest/CMakeLists.txt +++ b/failtest/CMakeLists.txt @@ -7,6 +7,9 @@ ei_add_failtest("block_nonconst_ctor_on_const_xpr_1") ei_add_failtest("block_nonconst_ctor_on_const_xpr_2") ei_add_failtest("transpose_nonconst_ctor_on_const_xpr") ei_add_failtest("diagonal_nonconst_ctor_on_const_xpr") +ei_add_failtest("cwiseunaryview_nonconst_ctor_on_const_xpr") +ei_add_failtest("triangularview_nonconst_ctor_on_const_xpr") +ei_add_failtest("selfadjointview_nonconst_ctor_on_const_xpr") ei_add_failtest("const_qualified_block_method_retval_0") ei_add_failtest("const_qualified_block_method_retval_1") @@ -25,6 +28,9 @@ ei_add_failtest("block_on_const_type_actually_const_0") ei_add_failtest("block_on_const_type_actually_const_1") ei_add_failtest("transpose_on_const_type_actually_const") ei_add_failtest("diagonal_on_const_type_actually_const") +ei_add_failtest("cwiseunaryview_on_const_type_actually_const") +ei_add_failtest("triangularview_on_const_type_actually_const") +ei_add_failtest("selfadjointview_on_const_type_actually_const") if (EIGEN_FAILTEST_FAILURE_COUNT) message(FATAL_ERROR diff --git a/failtest/cwiseunaryview_nonconst_ctor_on_const_xpr.cpp b/failtest/cwiseunaryview_nonconst_ctor_on_const_xpr.cpp new file mode 100644 index 000000000..e23cf8fd8 --- /dev/null +++ b/failtest/cwiseunaryview_nonconst_ctor_on_const_xpr.cpp @@ -0,0 +1,15 @@ +#include "../Eigen/Core" + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +#define CV_QUALIFIER const +#else +#define CV_QUALIFIER +#endif + +using namespace Eigen; + +void foo(CV_QUALIFIER Matrix3d &m){ + CwiseUnaryView,Matrix3d> t(m); +} + +int main() {} diff --git a/failtest/cwiseunaryview_on_const_type_actually_const.cpp b/failtest/cwiseunaryview_on_const_type_actually_const.cpp new file mode 100644 index 000000000..fcd41dfdb --- /dev/null +++ b/failtest/cwiseunaryview_on_const_type_actually_const.cpp @@ -0,0 +1,16 @@ +#include "../Eigen/Core" + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +#define CV_QUALIFIER const +#else +#define CV_QUALIFIER +#endif + +using namespace Eigen; + +void foo(){ + MatrixXf m; + CwiseUnaryView,CV_QUALIFIER MatrixXf>(m).coeffRef(0, 0) = 1.0f; +} + +int main() {} diff --git a/failtest/selfadjointview_nonconst_ctor_on_const_xpr.cpp b/failtest/selfadjointview_nonconst_ctor_on_const_xpr.cpp new file mode 100644 index 000000000..a240f8184 --- /dev/null +++ b/failtest/selfadjointview_nonconst_ctor_on_const_xpr.cpp @@ -0,0 +1,15 @@ +#include "../Eigen/Core" + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +#define CV_QUALIFIER const +#else +#define CV_QUALIFIER +#endif + +using namespace Eigen; + +void foo(CV_QUALIFIER Matrix3d &m){ + SelfAdjointView t(m); +} + +int main() {} diff --git a/failtest/selfadjointview_on_const_type_actually_const.cpp b/failtest/selfadjointview_on_const_type_actually_const.cpp new file mode 100644 index 000000000..19aaad6d0 --- /dev/null +++ b/failtest/selfadjointview_on_const_type_actually_const.cpp @@ -0,0 +1,16 @@ +#include "../Eigen/Core" + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +#define CV_QUALIFIER const +#else +#define CV_QUALIFIER +#endif + +using namespace Eigen; + +void foo(){ + MatrixXf m; + SelfAdjointView(m).coeffRef(0, 0) = 1.0f; +} + +int main() {} diff --git a/failtest/triangularview_nonconst_ctor_on_const_xpr.cpp b/failtest/triangularview_nonconst_ctor_on_const_xpr.cpp new file mode 100644 index 000000000..807447e4b --- /dev/null +++ b/failtest/triangularview_nonconst_ctor_on_const_xpr.cpp @@ -0,0 +1,15 @@ +#include "../Eigen/Core" + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +#define CV_QUALIFIER const +#else +#define CV_QUALIFIER +#endif + +using namespace Eigen; + +void foo(CV_QUALIFIER Matrix3d &m){ + TriangularView t(m); +} + +int main() {} diff --git a/failtest/triangularview_on_const_type_actually_const.cpp b/failtest/triangularview_on_const_type_actually_const.cpp new file mode 100644 index 000000000..0a381a612 --- /dev/null +++ b/failtest/triangularview_on_const_type_actually_const.cpp @@ -0,0 +1,16 @@ +#include "../Eigen/Core" + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +#define CV_QUALIFIER const +#else +#define CV_QUALIFIER +#endif + +using namespace Eigen; + +void foo(){ + MatrixXf m; + TriangularView(m).coeffRef(0, 0) = 1.0f; +} + +int main() {} From de0d8a010e8cee66901786e0e2819beeaa5cb253 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 23 Sep 2014 12:58:14 +0200 Subject: [PATCH 0822/1103] Suppress stupid gcc-4.4 warning --- Eigen/src/SparseCore/SparseFuzzy.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseFuzzy.h b/Eigen/src/SparseCore/SparseFuzzy.h index 3e67cbf5f..a76c1a5e0 100644 --- a/Eigen/src/SparseCore/SparseFuzzy.h +++ b/Eigen/src/SparseCore/SparseFuzzy.h @@ -18,7 +18,7 @@ bool SparseMatrixBase::isApprox(const SparseMatrixBase& o { using std::min; const typename internal::nested_eval::type actualA(derived()); - typename internal::conditional::type, const PlainObject>::type actualB(other.derived()); From 36448c9e287935b8c408791bf88b2907292d6c80 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 23 Sep 2014 14:28:23 +0200 Subject: [PATCH 0823/1103] Make constructors explicit if they could lead to unintended implicit conversion --- Eigen/src/Cholesky/LDLT.h | 12 +-- Eigen/src/Cholesky/LLT.h | 12 +-- Eigen/src/CholmodSupport/CholmodSupport.h | 2 +- Eigen/src/Core/ArrayBase.h | 4 +- Eigen/src/Core/ArrayWrapper.h | 4 +- Eigen/src/Core/BandMatrix.h | 6 +- Eigen/src/Core/CoreEvaluators.h | 57 ++++++------ Eigen/src/Core/CwiseUnaryOp.h | 2 +- Eigen/src/Core/CwiseUnaryView.h | 2 +- Eigen/src/Core/DenseBase.h | 4 +- Eigen/src/Core/DenseStorage.h | 16 ++-- Eigen/src/Core/Diagonal.h | 16 ++-- Eigen/src/Core/DiagonalMatrix.h | 20 +++-- Eigen/src/Core/Flagged.h | 4 +- Eigen/src/Core/ForceAlignedAccess.h | 6 +- Eigen/src/Core/GlobalFunctions.h | 4 +- Eigen/src/Core/Inverse.h | 2 +- Eigen/src/Core/Map.h | 2 +- Eigen/src/Core/MapBase.h | 2 +- Eigen/src/Core/Matrix.h | 2 +- Eigen/src/Core/MatrixBase.h | 6 +- Eigen/src/Core/NestByValue.h | 2 +- Eigen/src/Core/NoAlias.h | 4 +- Eigen/src/Core/PermutationMatrix.h | 11 +-- Eigen/src/Core/PlainObjectBase.h | 2 +- Eigen/src/Core/ProductEvaluators.h | 18 ++-- Eigen/src/Core/Redux.h | 2 +- Eigen/src/Core/ReturnByValue.h | 2 +- Eigen/src/Core/Reverse.h | 6 +- Eigen/src/Core/SelfAdjointView.h | 6 +- Eigen/src/Core/Solve.h | 2 +- Eigen/src/Core/Stride.h | 4 +- Eigen/src/Core/Transpose.h | 7 +- Eigen/src/Core/Transpositions.h | 6 +- Eigen/src/Core/TriangularMatrix.h | 35 ++++---- Eigen/src/Core/VectorwiseOp.h | 88 +++++++++++-------- Eigen/src/Core/Visitor.h | 2 +- .../Core/products/SelfadjointRank2Update.h | 8 +- Eigen/src/Core/util/BlasUtil.h | 2 +- Eigen/src/Eigenvalues/ComplexEigenSolver.h | 4 +- Eigen/src/Eigenvalues/ComplexSchur.h | 4 +- Eigen/src/Eigenvalues/EigenSolver.h | 4 +- .../src/Eigenvalues/GeneralizedEigenSolver.h | 4 +- .../GeneralizedSelfAdjointEigenSolver.h | 2 +- .../src/Eigenvalues/HessenbergDecomposition.h | 4 +- Eigen/src/Eigenvalues/RealQZ.h | 4 +- Eigen/src/Eigenvalues/RealSchur.h | 4 +- .../src/Eigenvalues/SelfAdjointEigenSolver.h | 4 +- Eigen/src/Eigenvalues/Tridiagonalization.h | 4 +- Eigen/src/Geometry/Homogeneous.h | 8 +- Eigen/src/Geometry/Quaternion.h | 6 +- Eigen/src/Geometry/Rotation2D.h | 2 +- .../BasicPreconditioners.h | 4 +- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 2 +- .../ConjugateGradient.h | 2 +- .../IterativeLinearSolvers/IncompleteLUT.h | 2 +- .../IterativeSolverBase.h | 2 +- Eigen/src/LU/FullPivLU.h | 2 +- Eigen/src/LU/PartialPivLU.h | 4 +- Eigen/src/PaStiXSupport/PaStiXSupport.h | 6 +- Eigen/src/PardisoSupport/PardisoSupport.h | 6 +- Eigen/src/QR/ColPivHouseholderQR.h | 2 +- Eigen/src/QR/FullPivHouseholderQR.h | 2 +- Eigen/src/QR/HouseholderQR.h | 2 +- Eigen/src/SPQRSupport/SuiteSparseQRSupport.h | 10 +-- Eigen/src/SVD/JacobiSVD.h | 4 +- Eigen/src/SVD/UpperBidiagonalization.h | 4 +- Eigen/src/SparseCholesky/SimplicialCholesky.h | 16 ++-- Eigen/src/SparseCore/AmbiVector.h | 4 +- Eigen/src/SparseCore/CompressedStorage.h | 2 +- Eigen/src/SparseCore/MappedSparseMatrix.h | 2 +- Eigen/src/SparseCore/SparseBlock.h | 4 +- Eigen/src/SparseCore/SparseCwiseBinaryOp.h | 8 +- Eigen/src/SparseCore/SparseCwiseUnaryOp.h | 4 +- Eigen/src/SparseCore/SparseDenseProduct.h | 9 +- Eigen/src/SparseCore/SparseDiagonalProduct.h | 4 +- Eigen/src/SparseCore/SparseMatrix.h | 6 +- Eigen/src/SparseCore/SparseMatrixBase.h | 15 ++-- Eigen/src/SparseCore/SparsePermutation.h | 4 +- Eigen/src/SparseCore/SparseProduct.h | 2 +- Eigen/src/SparseCore/SparseSelfAdjointView.h | 6 +- Eigen/src/SparseCore/SparseTranspose.h | 2 +- Eigen/src/SparseCore/SparseTriangularView.h | 4 +- Eigen/src/SparseCore/SparseVector.h | 12 +-- Eigen/src/SparseCore/SparseView.h | 6 +- Eigen/src/SparseLU/SparseLU.h | 6 +- Eigen/src/SparseQR/SparseQR.h | 6 +- Eigen/src/SuperLUSupport/SuperLUSupport.h | 2 +- Eigen/src/UmfPackSupport/UmfPackSupport.h | 2 +- Eigen/src/misc/Kernel.h | 2 +- Eigen/src/plugins/ArrayCwiseUnaryOps.h | 80 ++++++++++------- Eigen/src/plugins/CommonCwiseUnaryOps.h | 37 ++++---- Eigen/src/plugins/MatrixCwiseUnaryOps.h | 26 +++--- 93 files changed, 416 insertions(+), 355 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index 32c770654..b4016cd6c 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -85,7 +85,7 @@ template class LDLT * according to the specified problem \a size. * \sa LDLT() */ - LDLT(Index size) + explicit LDLT(Index size) : m_matrix(size, size), m_transpositions(size), m_temporary(size), @@ -98,7 +98,7 @@ template class LDLT * This calculates the decomposition for the input \a matrix. * \sa LDLT(Index size) */ - LDLT(const MatrixType& matrix) + explicit LDLT(const MatrixType& matrix) : m_matrix(matrix.rows(), matrix.cols()), m_transpositions(matrix.rows()), m_temporary(matrix.rows()), @@ -406,16 +406,16 @@ template struct LDLT_Traits { typedef const TriangularView MatrixL; typedef const TriangularView MatrixU; - static inline MatrixL getL(const MatrixType& m) { return m; } - static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); } + static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); } + static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); } }; template struct LDLT_Traits { typedef const TriangularView MatrixL; typedef const TriangularView MatrixU; - static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); } - static inline MatrixU getU(const MatrixType& m) { return m; } + static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); } + static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); } }; } // end namespace internal diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index cb9e0eb7b..90194e64d 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -83,10 +83,10 @@ template class LLT * according to the specified problem \a size. * \sa LLT() */ - LLT(Index size) : m_matrix(size, size), + explicit LLT(Index size) : m_matrix(size, size), m_isInitialized(false) {} - LLT(const MatrixType& matrix) + explicit LLT(const MatrixType& matrix) : m_matrix(matrix.rows(), matrix.cols()), m_isInitialized(false) { @@ -351,8 +351,8 @@ template struct LLT_Traits { typedef const TriangularView MatrixL; typedef const TriangularView MatrixU; - static inline MatrixL getL(const MatrixType& m) { return m; } - static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); } + static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); } + static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); } static bool inplace_decomposition(MatrixType& m) { return llt_inplace::blocked(m)==-1; } }; @@ -361,8 +361,8 @@ template struct LLT_Traits { typedef const TriangularView MatrixL; typedef const TriangularView MatrixU; - static inline MatrixL getL(const MatrixType& m) { return m.adjoint(); } - static inline MatrixU getU(const MatrixType& m) { return m; } + static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); } + static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); } static bool inplace_decomposition(MatrixType& m) { return llt_inplace::blocked(m)==-1; } }; diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h index 3524ffb2d..0c3b86dd2 100644 --- a/Eigen/src/CholmodSupport/CholmodSupport.h +++ b/Eigen/src/CholmodSupport/CholmodSupport.h @@ -180,7 +180,7 @@ class CholmodBase : public SparseSolverBase cholmod_start(&m_cholmod); } - CholmodBase(const MatrixType& matrix) + explicit CholmodBase(const MatrixType& matrix) : m_cholmodFactor(0), m_info(Success) { m_shiftOffset[0] = m_shiftOffset[1] = RealScalar(0.0); diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index 48a0006d5..b37b49ac4 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -158,9 +158,9 @@ template class ArrayBase /** \returns an \link Eigen::MatrixBase Matrix \endlink expression of this array * \sa MatrixBase::array() */ EIGEN_DEVICE_FUNC - MatrixWrapper matrix() { return derived(); } + MatrixWrapper matrix() { return MatrixWrapper(derived()); } EIGEN_DEVICE_FUNC - const MatrixWrapper matrix() const { return derived(); } + const MatrixWrapper matrix() const { return MatrixWrapper(derived()); } // template // inline void evalTo(Dest& dst) const { dst = matrix(); } diff --git a/Eigen/src/Core/ArrayWrapper.h b/Eigen/src/Core/ArrayWrapper.h index ed5210272..0b89c58cb 100644 --- a/Eigen/src/Core/ArrayWrapper.h +++ b/Eigen/src/Core/ArrayWrapper.h @@ -55,7 +55,7 @@ class ArrayWrapper : public ArrayBase > typedef typename internal::nested::type NestedExpressionType; EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {} + explicit EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {} EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } @@ -198,7 +198,7 @@ class MatrixWrapper : public MatrixBase > typedef typename internal::nested::type NestedExpressionType; EIGEN_DEVICE_FUNC - inline MatrixWrapper(ExpressionType& a_matrix) : m_expression(a_matrix) {} + explicit inline MatrixWrapper(ExpressionType& a_matrix) : m_expression(a_matrix) {} EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } diff --git a/Eigen/src/Core/BandMatrix.h b/Eigen/src/Core/BandMatrix.h index b0ebe1160..e59ee3da9 100644 --- a/Eigen/src/Core/BandMatrix.h +++ b/Eigen/src/Core/BandMatrix.h @@ -204,7 +204,7 @@ class BandMatrix : public BandMatrixBase::Index Index; typedef typename internal::traits::CoefficientsType CoefficientsType; - inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs) + explicit inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs) : m_coeffs(1+supers+subs,cols), m_rows(rows), m_supers(supers), m_subs(subs) { @@ -266,7 +266,7 @@ class BandMatrixWrapper : public BandMatrixBase::CoefficientsType CoefficientsType; typedef typename internal::traits::Index Index; - inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs) + explicit inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs) : m_coeffs(coeffs), m_rows(rows), m_supers(supers), m_subs(subs) { @@ -314,7 +314,7 @@ class TridiagonalMatrix : public BandMatrix Base; typedef typename Base::Index Index; public: - TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {} + explicit TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {} inline typename Base::template DiagonalIntReturnType<1>::Type super() { return Base::template diagonal<1>(); } diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 09a83a382..ab4320c2a 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -34,6 +34,11 @@ template struct storage_kind_to_shape; template<> struct storage_kind_to_shape { typedef DenseShape Shape; }; + +// FIXME Is this necessary? And why was it not before refactoring??? +template<> struct storage_kind_to_shape { typedef PermutationShape Shape; }; + + // Evaluators have to be specialized with respect to various criteria such as: // - storage/structure/shape // - scalar type @@ -87,7 +92,7 @@ template struct evaluator : public unary_evaluator { typedef unary_evaluator Base; - evaluator(const T& xpr) : Base(xpr) {} + explicit evaluator(const T& xpr) : Base(xpr) {} }; @@ -150,7 +155,7 @@ struct evaluator > : RowsAtCompileTime) {} - evaluator(const PlainObjectType& m) + explicit evaluator(const PlainObjectType& m) : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) { } @@ -242,7 +247,7 @@ struct evaluator > evaluator() {} - evaluator(const XprType& m) + explicit evaluator(const XprType& m) : evaluator >(m) { } }; @@ -260,7 +265,7 @@ struct unary_evaluator, IndexBased> Flags = evaluator::Flags ^ RowMajorBit }; - unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} + explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -337,7 +342,7 @@ struct evaluator > | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit) // FIXME EvalBeforeNestingBit should be needed anymore }; - evaluator(const XprType& n) + explicit evaluator(const XprType& n) : m_functor(n.functor()) { } @@ -387,7 +392,7 @@ struct unary_evaluator, IndexBased > | (functor_traits::PacketAccess ? PacketAccessBit : 0)) }; - unary_evaluator(const XprType& op) + explicit unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -433,7 +438,7 @@ struct evaluator > typedef CwiseBinaryOp XprType; typedef binary_evaluator > Base; - evaluator(const XprType& xpr) : Base(xpr) {} + explicit evaluator(const XprType& xpr) : Base(xpr) {} }; template @@ -461,7 +466,7 @@ struct binary_evaluator, IndexBased, IndexBase Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit) }; - binary_evaluator(const XprType& xpr) + explicit binary_evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) @@ -515,7 +520,7 @@ struct unary_evaluator, IndexBased> Flags = (evaluator::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit)) }; - unary_evaluator(const XprType& op) + explicit unary_evaluator(const XprType& op) : m_unaryOp(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -573,7 +578,7 @@ struct mapbase_evaluator : evaluator_base CoeffReadCost = NumTraits::ReadCost }; - mapbase_evaluator(const XprType& map) + explicit mapbase_evaluator(const XprType& map) : m_data(const_cast(map.data())), m_xpr(map) { @@ -663,7 +668,7 @@ struct evaluator > Flags = KeepsPacketAccess ? int(Flags2) : (int(Flags2) & ~PacketAccessBit) }; - evaluator(const XprType& map) + explicit evaluator(const XprType& map) : mapbase_evaluator(map) { } }; @@ -680,7 +685,7 @@ struct evaluator > Flags = evaluator >::Flags }; - evaluator(const XprType& ref) + explicit evaluator(const XprType& ref) : mapbase_evaluator(ref) { } }; @@ -731,7 +736,7 @@ struct evaluator > Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit }; typedef block_evaluator block_evaluator_type; - evaluator(const XprType& block) : block_evaluator_type(block) {} + explicit evaluator(const XprType& block) : block_evaluator_type(block) {} }; // no direct-access => dispatch to a unary evaluator @@ -741,7 +746,7 @@ struct block_evaluator XprType; - block_evaluator(const XprType& block) + explicit block_evaluator(const XprType& block) : unary_evaluator(block) {} }; @@ -752,7 +757,7 @@ struct unary_evaluator, IndexBa { typedef Block XprType; - unary_evaluator(const XprType& block) + explicit unary_evaluator(const XprType& block) : m_argImpl(block.nestedExpression()), m_startRow(block.startRow()), m_startCol(block.startCol()) @@ -831,7 +836,7 @@ struct block_evaluator XprType; - block_evaluator(const XprType& block) + explicit block_evaluator(const XprType& block) : mapbase_evaluator(block) { // FIXME this should be an internal assertion @@ -857,7 +862,7 @@ struct evaluator > Flags = (unsigned int)evaluator::Flags & evaluator::Flags & HereditaryBits }; - evaluator(const XprType& select) + explicit evaluator(const XprType& select) : m_conditionImpl(select.conditionMatrix()), m_thenImpl(select.thenMatrix()), m_elseImpl(select.elseMatrix()) @@ -911,7 +916,7 @@ struct unary_evaluator > Flags = (evaluator::Flags & HereditaryBits & ~RowMajorBit) | (traits::Flags & RowMajorBit) }; - unary_evaluator(const XprType& replicate) + explicit unary_evaluator(const XprType& replicate) : m_arg(replicate.nestedExpression()), m_argImpl(m_arg), m_rows(replicate.nestedExpression().rows()), @@ -975,7 +980,7 @@ struct evaluator > Flags = (traits::Flags&RowMajorBit) | (evaluator::Flags&HereditaryBits) }; - evaluator(const XprType expr) + explicit evaluator(const XprType expr) : m_expr(expr) {} @@ -1012,7 +1017,7 @@ struct evaluator_wrapper_base Flags = evaluator::Flags }; - evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} + explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} typedef typename ArgType::Index Index; typedef typename ArgType::Scalar Scalar; @@ -1074,7 +1079,7 @@ struct unary_evaluator > { typedef MatrixWrapper XprType; - unary_evaluator(const XprType& wrapper) + explicit unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; @@ -1085,7 +1090,7 @@ struct unary_evaluator > { typedef ArrayWrapper XprType; - unary_evaluator(const XprType& wrapper) + explicit unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; @@ -1131,7 +1136,7 @@ struct unary_evaluator > }; typedef internal::reverse_packet_cond reverse_packet; - unary_evaluator(const XprType& reverse) + explicit unary_evaluator(const XprType& reverse) : m_argImpl(reverse.nestedExpression()), m_rows(ReverseRow ? reverse.nestedExpression().rows() : 0), m_cols(ReverseCol ? reverse.nestedExpression().cols() : 0) @@ -1212,7 +1217,7 @@ struct evaluator > Flags = (unsigned int)evaluator::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit) & ~RowMajorBit }; - evaluator(const XprType& diagonal) + explicit evaluator(const XprType& diagonal) : m_argImpl(diagonal.nestedExpression()), m_index(diagonal.index()) { } @@ -1275,7 +1280,7 @@ class EvalToTemp typedef typename dense_xpr_base::type Base; EIGEN_GENERIC_PUBLIC_INTERFACE(EvalToTemp) - EvalToTemp(const ArgType& arg) + explicit EvalToTemp(const ArgType& arg) : m_arg(arg) { } @@ -1309,7 +1314,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - evaluator(const XprType& xpr) + explicit evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index 79a872934..708e3e818 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -63,7 +63,7 @@ class CwiseUnaryOp : internal::no_assignment_operator, typedef typename internal::remove_all::type NestedExpression; EIGEN_DEVICE_FUNC - inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) + explicit inline CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp()) : m_xpr(xpr), m_functor(func) {} EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index 71249a39c..61fd8ee35 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -63,7 +63,7 @@ class CwiseUnaryView : public CwiseUnaryViewImpl::type NestedExpression; - inline CwiseUnaryView(const MatrixType& mat, const ViewOp& func = ViewOp()) + explicit inline CwiseUnaryView(const MatrixType& mat, const ViewOp& func = ViewOp()) : m_matrix(mat), m_functor(func) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryView) diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 6078af553..101ef6f20 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -275,6 +275,7 @@ template class DenseBase // TODO flagged is temporarly disabled. It seems useless now template + EIGEN_DEPRECATED const Derived& flagged() const { return derived(); } @@ -282,8 +283,9 @@ template class DenseBase EIGEN_DEVICE_FUNC CommaInitializer operator<< (const DenseBase& other); + typedef Transpose TransposeReturnType; EIGEN_DEVICE_FUNC - Eigen::Transpose transpose(); + TransposeReturnType transpose(); typedef typename internal::add_const >::type ConstTransposeReturnType; EIGEN_DEVICE_FUNC ConstTransposeReturnType transpose() const; diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 59f515495..852648639 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -130,7 +130,7 @@ template class DenseSt public: EIGEN_DEVICE_FUNC DenseStorage() {} EIGEN_DEVICE_FUNC - DenseStorage(internal::constructor_without_unaligned_array_assert) + explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(internal::constructor_without_unaligned_array_assert()) {} EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data) {} @@ -155,7 +155,7 @@ template class DenseStorage class DenseStorage class DenseStorage class DenseStorage class DenseStorage(size)), m_rows(nbRows), m_cols(nbCols) @@ -350,7 +350,7 @@ template class DenseStorage(size)), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } DenseStorage(const DenseStorage& other) @@ -416,7 +416,7 @@ template class DenseStorage(size)), m_rows(nbRows) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } DenseStorage(const DenseStorage& other) diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index 1ffcd97f9..ba7ddbb14 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -70,7 +70,7 @@ template class Diagonal EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal) EIGEN_DEVICE_FUNC - inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {} + explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal) @@ -189,7 +189,7 @@ template inline typename MatrixBase::DiagonalReturnType MatrixBase::diagonal() { - return derived(); + return DiagonalReturnType(derived()); } /** This is the const version of diagonal(). */ @@ -238,20 +238,20 @@ MatrixBase::diagonal(Index index) const * * \sa MatrixBase::diagonal(), class Diagonal */ template -template -inline typename MatrixBase::template DiagonalIndexReturnType::Type +template +inline typename MatrixBase::template DiagonalIndexReturnType::Type MatrixBase::diagonal() { - return derived(); + return typename DiagonalIndexReturnType::Type(derived()); } /** This is the const version of diagonal(). */ template -template -inline typename MatrixBase::template ConstDiagonalIndexReturnType::Type +template +inline typename MatrixBase::template ConstDiagonalIndexReturnType::Type MatrixBase::diagonal() const { - return derived(); + return typename ConstDiagonalIndexReturnType::Type(derived()); } } // end namespace Eigen diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index 44c249aa6..e3dc71336 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -63,24 +63,26 @@ class DiagonalBase : public EigenBase return Product(derived(),matrix.derived()); } + typedef DiagonalWrapper, const DiagonalVectorType> > InverseReturnType; EIGEN_DEVICE_FUNC - inline const DiagonalWrapper, const DiagonalVectorType> > + inline const InverseReturnType inverse() const { - return diagonal().cwiseInverse(); + return InverseReturnType(diagonal().cwiseInverse()); } + typedef DiagonalWrapper, const DiagonalVectorType> > ScalarMultipleReturnType; EIGEN_DEVICE_FUNC - inline const DiagonalWrapper, const DiagonalVectorType> > + inline const ScalarMultipleReturnType operator*(const Scalar& scalar) const { - return diagonal() * scalar; + return ScalarMultipleReturnType(diagonal() * scalar); } EIGEN_DEVICE_FUNC - friend inline const DiagonalWrapper, const DiagonalVectorType> > + friend inline const ScalarMultipleReturnType operator*(const Scalar& scalar, const DiagonalBase& other) { - return other.diagonal() * scalar; + return ScalarMultipleReturnType(other.diagonal() * scalar); } }; @@ -144,7 +146,7 @@ class DiagonalMatrix /** Constructs a diagonal matrix with given dimension */ EIGEN_DEVICE_FUNC - inline DiagonalMatrix(Index dim) : m_diagonal(dim) {} + explicit inline DiagonalMatrix(Index dim) : m_diagonal(dim) {} /** 2D constructor. */ EIGEN_DEVICE_FUNC @@ -253,7 +255,7 @@ class DiagonalWrapper /** Constructor from expression of diagonal coefficients to wrap. */ EIGEN_DEVICE_FUNC - inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {} + explicit inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {} /** \returns a const reference to the wrapped expression of diagonal coefficients. */ EIGEN_DEVICE_FUNC @@ -276,7 +278,7 @@ template inline const DiagonalWrapper MatrixBase::asDiagonal() const { - return derived(); + return DiagonalWrapper(derived()); } /** \returns true if *this is approximately equal to a diagonal matrix, diff --git a/Eigen/src/Core/Flagged.h b/Eigen/src/Core/Flagged.h index 1f2955fc1..6ce11edf3 100644 --- a/Eigen/src/Core/Flagged.h +++ b/Eigen/src/Core/Flagged.h @@ -48,7 +48,7 @@ template clas ExpressionType, const ExpressionType&>::type ExpressionTypeNested; typedef typename ExpressionType::InnerIterator InnerIterator; - inline Flagged(const ExpressionType& matrix) : m_matrix(matrix) {} + explicit inline Flagged(const ExpressionType& matrix) : m_matrix(matrix) {} inline Index rows() const { return m_matrix.rows(); } inline Index cols() const { return m_matrix.cols(); } @@ -132,7 +132,7 @@ template inline const Flagged DenseBase::flagged() const { - return derived(); + return Flagged(derived()); } } // end namespace Eigen diff --git a/Eigen/src/Core/ForceAlignedAccess.h b/Eigen/src/Core/ForceAlignedAccess.h index 807c7a293..065acfa64 100644 --- a/Eigen/src/Core/ForceAlignedAccess.h +++ b/Eigen/src/Core/ForceAlignedAccess.h @@ -39,7 +39,7 @@ template class ForceAlignedAccess typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(ForceAlignedAccess) - inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {} + explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {} inline Index rows() const { return m_expression.rows(); } inline Index cols() const { return m_expression.cols(); } @@ -127,7 +127,7 @@ template inline typename internal::add_const_on_value_type,Derived&>::type>::type MatrixBase::forceAlignedAccessIf() const { - return derived(); + return derived(); // FIXME This should not work but apparently is never used } /** \returns an expression of *this with forced aligned access if \a Enable is true. @@ -138,7 +138,7 @@ template inline typename internal::conditional,Derived&>::type MatrixBase::forceAlignedAccessIf() { - return derived(); + return derived(); // FIXME This should not work but apparently is never used } } // end namespace Eigen diff --git a/Eigen/src/Core/GlobalFunctions.h b/Eigen/src/Core/GlobalFunctions.h index 2067a2a6e..ee67b7d3c 100644 --- a/Eigen/src/Core/GlobalFunctions.h +++ b/Eigen/src/Core/GlobalFunctions.h @@ -15,7 +15,7 @@ template \ inline const Eigen::CwiseUnaryOp, const Derived> \ NAME(const Eigen::ArrayBase& x) { \ - return x.derived(); \ + return Eigen::CwiseUnaryOp, const Derived>(x.derived()); \ } #define EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(NAME,FUNCTOR) \ @@ -30,7 +30,7 @@ { \ static inline typename NAME##_retval >::type run(const Eigen::ArrayBase& x) \ { \ - return x.derived(); \ + return typename NAME##_retval >::type(x.derived()); \ } \ }; diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 5cfa7e50c..706796b78 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -51,7 +51,7 @@ public: typedef typename internal::nested::type XprTypeNested; typedef typename internal::remove_all::type XprTypeNestedCleaned; - Inverse(const XprType &xpr) + explicit Inverse(const XprType &xpr) : m_xpr(xpr) {} diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h index 87c1787bf..098f1c096 100644 --- a/Eigen/src/Core/Map.h +++ b/Eigen/src/Core/Map.h @@ -122,7 +122,7 @@ template class Ma * \param a_stride optional Stride object, passing the strides. */ EIGEN_DEVICE_FUNC - inline Map(PointerArgType dataPtr, const StrideType& a_stride = StrideType()) + explicit inline Map(PointerArgType dataPtr, const StrideType& a_stride = StrideType()) : Base(cast_to_pointer_type(dataPtr)), m_stride(a_stride) { PlainObjectType::Base::_check_template_params(); diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index 6d3b344e8..3e68b1e91 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -128,7 +128,7 @@ template class MapBase } EIGEN_DEVICE_FUNC - inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime) + explicit inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime) { EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived) checkSanity(); diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 8a5821548..bb6c75387 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -214,7 +214,7 @@ class Matrix // FIXME is it still needed EIGEN_DEVICE_FUNC - Matrix(internal::constructor_without_unaligned_array_assert) + explicit Matrix(internal::constructor_without_unaligned_array_assert) : Base(internal::constructor_without_unaligned_array_assert()) { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED } diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 9dbbd6fb5..048060e6b 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -327,7 +327,7 @@ template class MatrixBase NoAlias noalias(); - // TODO forceAlignedAccess is temporarly disabled + // TODO forceAlignedAccess is temporarily disabled // Need to find a nicer workaround. inline const Derived& forceAlignedAccess() const { return derived(); } inline Derived& forceAlignedAccess() { return derived(); } @@ -343,10 +343,10 @@ template class MatrixBase /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix * \sa ArrayBase::matrix() */ - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper array() { return derived(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper array() { return ArrayWrapper(derived()); } /** \returns a const \link Eigen::ArrayBase Array \endlink expression of this matrix * \sa ArrayBase::matrix() */ - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper array() const { return derived(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper array() const { return ArrayWrapper(derived()); } /////////// LU module /////////// diff --git a/Eigen/src/Core/NestByValue.h b/Eigen/src/Core/NestByValue.h index a893b1761..248dd8eb0 100644 --- a/Eigen/src/Core/NestByValue.h +++ b/Eigen/src/Core/NestByValue.h @@ -40,7 +40,7 @@ template class NestByValue typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(NestByValue) - inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {} + explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {} inline Index rows() const { return m_expression.rows(); } inline Index cols() const { return m_expression.cols(); } diff --git a/Eigen/src/Core/NoAlias.h b/Eigen/src/Core/NoAlias.h index 097c9c062..0ade75255 100644 --- a/Eigen/src/Core/NoAlias.h +++ b/Eigen/src/Core/NoAlias.h @@ -33,7 +33,7 @@ class NoAlias public: typedef typename ExpressionType::Scalar Scalar; - NoAlias(ExpressionType& expression) : m_expression(expression) {} + explicit NoAlias(ExpressionType& expression) : m_expression(expression) {} template EIGEN_DEVICE_FUNC @@ -100,7 +100,7 @@ class NoAlias template NoAlias MatrixBase::noalias() { - return derived(); + return NoAlias(derived()); } } // end namespace Eigen diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 200518173..98c83047a 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -73,6 +73,7 @@ class PermutationBase : public EigenBase typedef PermutationMatrix PlainPermutationType; using Base::derived; + typedef Transpose TransposeReturnType; #endif /** Copies the other permutation into *this */ @@ -198,14 +199,14 @@ class PermutationBase : public EigenBase * * \note \note_try_to_help_rvo */ - inline Transpose inverse() const - { return derived(); } + inline TransposeReturnType inverse() const + { return TransposeReturnType(derived()); } /** \returns the tranpose permutation matrix. * * \note \note_try_to_help_rvo */ - inline Transpose transpose() const - { return derived(); } + inline TransposeReturnType transpose() const + { return TransposeReturnType(derived()); } /**** multiplication helpers to hopefully get RVO ****/ @@ -301,7 +302,7 @@ class PermutationMatrix : public PermutationBase::type // FIXME is it still needed ? /** \internal */ EIGEN_DEVICE_FUNC - PlainObjectBase(internal::constructor_without_unaligned_array_assert) + explicit PlainObjectBase(internal::constructor_without_unaligned_array_assert) : m_storage(internal::constructor_without_unaligned_array_assert()) { // _check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index f880e7696..69e569908 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -35,7 +35,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - evaluator(const XprType& xpr) : Base(xpr) {} + explicit evaluator(const XprType& xpr) : Base(xpr) {} }; // Catch scalar * ( A * B ) and transform it to (A*scalar) * B @@ -50,7 +50,7 @@ struct evaluator, const Produ typedef evaluator type; typedef evaluator nestedType; - evaluator(const XprType& xpr) + explicit evaluator(const XprType& xpr) : Base(xpr.functor().m_other * xpr.nestedExpression().lhs() * xpr.nestedExpression().rhs()) {} }; @@ -66,7 +66,7 @@ struct evaluator, DiagIndex> > typedef evaluator type; typedef evaluator nestedType; - evaluator(const XprType& xpr) + explicit evaluator(const XprType& xpr) : Base(Diagonal, DiagIndex>( Product(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()), xpr.index() )) @@ -103,7 +103,7 @@ struct product_evaluator, ProductTag, LhsShape typedef typename XprType::PlainObject PlainObject; typedef typename evaluator::type Base; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); @@ -242,7 +242,7 @@ struct generic_product_impl struct sub { template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } }; struct adds { Scalar m_scale; - adds(const Scalar& s) : m_scale(s) {} + explicit adds(const Scalar& s) : m_scale(s) {} template void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += m_scale * src; } @@ -377,7 +377,7 @@ struct product_evaluator, ProductTag, DenseShape, typedef typename XprType::PacketScalar PacketScalar; typedef typename XprType::PacketReturnType PacketReturnType; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : m_lhs(xpr.lhs()), m_rhs(xpr.rhs()), m_lhsImpl(m_lhs), // FIXME the creation of the evaluator objects should result in a no-op, but check that! @@ -508,7 +508,7 @@ struct product_evaluator, LazyCoeffBasedProduc typedef Product XprType; typedef Product BaseProduct; typedef product_evaluator Base; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : Base(BaseProduct(xpr.lhs(),xpr.rhs())) {} }; @@ -739,7 +739,7 @@ struct product_evaluator, ProductTag, DiagonalSha StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) { } @@ -783,7 +783,7 @@ struct product_evaluator, ProductTag, DenseShape, enum { StorageOrder = int(Lhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal()) { } diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index c6c355d43..14a2671e9 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -340,7 +340,7 @@ class redux_evaluator { public: typedef _XprType XprType; - redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} + explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index f4e12a93b..5fcd9e3fc 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -103,7 +103,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - evaluator(const XprType& xpr) + explicit evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h index 01de90800..9ba6ea2e6 100644 --- a/Eigen/src/Core/Reverse.h +++ b/Eigen/src/Core/Reverse.h @@ -89,7 +89,7 @@ template class Reverse typedef internal::reverse_packet_cond reverse_packet; public: - inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { } + explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { } EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse) @@ -184,7 +184,7 @@ template inline typename DenseBase::ReverseReturnType DenseBase::reverse() { - return derived(); + return ReverseReturnType(derived()); } /** This is the const version of reverse(). */ @@ -192,7 +192,7 @@ template inline const typename DenseBase::ConstReverseReturnType DenseBase::reverse() const { - return derived(); + return ConstReverseReturnType(derived()); } /** This is the "in place" version of reverse: it reverses \c *this. diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 19cb232c9..f5fbd7215 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -67,7 +67,7 @@ template class SelfAdjointView typedef typename MatrixType::PlainObject PlainObject; EIGEN_DEVICE_FUNC - inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix) + explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix) {} EIGEN_DEVICE_FUNC @@ -258,7 +258,7 @@ template typename MatrixBase::template ConstSelfAdjointViewReturnType::Type MatrixBase::selfadjointView() const { - return derived(); + return typename ConstSelfAdjointViewReturnType::Type(derived()); } template @@ -266,7 +266,7 @@ template typename MatrixBase::template SelfAdjointViewReturnType::Type MatrixBase::selfadjointView() { - return derived(); + return typename SelfAdjointViewReturnType::Type(derived()); } } // end namespace Eigen diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index 7b12be1e6..641ffa218 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -121,7 +121,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - evaluator(const SolveType& solve) + explicit evaluator(const SolveType& solve) : m_result(solve.rows(), solve.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/Core/Stride.h b/Eigen/src/Core/Stride.h index 187774978..e46faad34 100644 --- a/Eigen/src/Core/Stride.h +++ b/Eigen/src/Core/Stride.h @@ -93,7 +93,7 @@ class InnerStride : public Stride<0, Value> public: typedef DenseIndex Index; EIGEN_DEVICE_FUNC InnerStride() : Base() {} - EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} + EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} // FIXME making this explicit could break valid code }; /** \brief Convenience specialization of Stride to specify only an outer stride @@ -105,7 +105,7 @@ class OuterStride : public Stride public: typedef DenseIndex Index; EIGEN_DEVICE_FUNC OuterStride() : Base() {} - EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} + EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} // FIXME making this explicit could break valid code }; } // end namespace Eigen diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 144bb2c01..57d6fd2fe 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -64,7 +64,7 @@ template class Transpose typedef typename internal::remove_all::type NestedExpression; EIGEN_DEVICE_FUNC - inline Transpose(MatrixType& a_matrix) : m_matrix(a_matrix) {} + explicit inline Transpose(MatrixType& a_matrix) : m_matrix(a_matrix) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose) @@ -169,7 +169,7 @@ template inline Transpose DenseBase::transpose() { - return derived(); + return TransposeReturnType(derived()); } /** This is the const version of transpose(). @@ -207,8 +207,7 @@ template inline const typename MatrixBase::AdjointReturnType MatrixBase::adjoint() const { - return this->transpose(); // in the complex case, the .conjugate() is be implicit here - // due to implicit conversion to return type + return AdjointReturnType(this->transpose()); } /*************************************************************************** diff --git a/Eigen/src/Core/Transpositions.h b/Eigen/src/Core/Transpositions.h index 92261118f..77e7d6f45 100644 --- a/Eigen/src/Core/Transpositions.h +++ b/Eigen/src/Core/Transpositions.h @@ -240,7 +240,7 @@ class Map > typedef typename TranspositionType::IndicesType IndicesType; public: - Transpose(const TranspositionType& t) : m_transpositions(t) {} + explicit Transpose(const TranspositionType& t) : m_transpositions(t) {} inline int size() const { return m_transpositions.size(); } diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 36f04a5e8..36bbd46e1 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -211,8 +211,9 @@ template class TriangularView IsVectorAtCompileTime = false }; + // FIXME This, combined with const_cast_derived in transpose() leads to a const-correctness loophole EIGEN_DEVICE_FUNC - inline TriangularView(const MatrixType& matrix) : m_matrix(matrix) + explicit inline TriangularView(const MatrixType& matrix) : m_matrix(matrix) {} using Base::operator=; @@ -229,32 +230,36 @@ template class TriangularView EIGEN_DEVICE_FUNC NestedExpression& nestedExpression() { return *const_cast(&m_matrix); } + typedef TriangularView ConjugateReturnType; /** \sa MatrixBase::conjugate() */ EIGEN_DEVICE_FUNC - inline TriangularView conjugate() - { return m_matrix.conjugate(); } + inline ConjugateReturnType conjugate() + { return ConjugateReturnType(m_matrix.conjugate()); } /** \sa MatrixBase::conjugate() const */ - EIGEN_DEVICE_FUNC - inline const TriangularView conjugate() const - { return m_matrix.conjugate(); } + EIGEN_DEVICE_FUNC + inline const ConjugateReturnType conjugate() const + { return ConjugateReturnType(m_matrix.conjugate()); } + + typedef TriangularView AdjointReturnType; /** \sa MatrixBase::adjoint() const */ EIGEN_DEVICE_FUNC - inline const TriangularView adjoint() const - { return m_matrix.adjoint(); } + inline const AdjointReturnType adjoint() const + { return AdjointReturnType(m_matrix.adjoint()); } - /** \sa MatrixBase::transpose() */ + typedef TriangularView,TransposeMode> TransposeReturnType; + /** \sa MatrixBase::transpose() */ EIGEN_DEVICE_FUNC - inline TriangularView,TransposeMode> transpose() + inline TransposeReturnType transpose() { EIGEN_STATIC_ASSERT_LVALUE(MatrixType) - return m_matrix.const_cast_derived().transpose(); + return TransposeReturnType(m_matrix.const_cast_derived().transpose()); } /** \sa MatrixBase::transpose() const */ EIGEN_DEVICE_FUNC - inline const TriangularView,TransposeMode> transpose() const + inline const TransposeReturnType transpose() const { - return m_matrix.transpose(); + return TransposeReturnType(m_matrix.transpose()); } template @@ -556,7 +561,7 @@ template typename MatrixBase::template TriangularViewReturnType::Type MatrixBase::triangularView() { - return derived(); + return typename TriangularViewReturnType::Type(derived()); } /** This is the const version of MatrixBase::triangularView() */ @@ -565,7 +570,7 @@ template typename MatrixBase::template ConstTriangularViewReturnType::Type MatrixBase::triangularView() const { - return derived(); + return typename ConstTriangularViewReturnType::Type(derived()); } /** \returns true if *this is approximately equal to an upper triangular matrix, diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index a8130e902..e340c1433 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -65,7 +65,7 @@ class PartialReduxExpr : internal::no_assignment_operator, typedef typename internal::traits::MatrixTypeNested MatrixTypeNested; typedef typename internal::traits::_MatrixTypeNested _MatrixTypeNested; - PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp()) + explicit PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp()) : m_matrix(mat), m_functor(func) {} Index rows() const { return (Direction==Vertical ? 1 : m_matrix.rows()); } @@ -128,7 +128,7 @@ struct member_redux { >::type result_type; template struct Cost { enum { value = (Size-1) * functor_traits::Cost }; }; - member_redux(const BinaryOp func) : m_functor(func) {} + member_redux(const BinaryOp func) : m_functor(func) {} // FIXME this should actually be explicit, but lets not exaggerate template inline result_type operator()(const DenseBase& mat) const { return mat.redux(m_functor); } @@ -249,7 +249,7 @@ template class VectorwiseOp public: - inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {} + explicit inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {} /** \internal */ inline const ExpressionType& _expression() const { return m_matrix; } @@ -266,6 +266,21 @@ template class VectorwiseOp redux(const BinaryOp& func = BinaryOp()) const { return typename ReduxReturnType::Type(_expression(), func); } + typedef typename ReturnType::Type MinCoeffReturnType; + typedef typename ReturnType::Type MaxCoeffReturnType; + typedef typename ReturnType::Type SquareNormReturnType; + typedef typename ReturnType::Type NormReturnType; + typedef typename ReturnType::Type BlueNormReturnType; + typedef typename ReturnType::Type StableNormReturnType; + typedef typename ReturnType::Type HypotNormReturnType; + typedef typename ReturnType::Type SumReturnType; + typedef typename ReturnType::Type MeanReturnType; + typedef typename ReturnType::Type AllReturnType; + typedef typename ReturnType::Type AnyReturnType; + typedef PartialReduxExpr, Direction> CountReturnType; + typedef typename ReturnType::Type ProdReturnType; + typedef Reverse ReverseReturnType; + /** \returns a row (or column) vector expression of the smallest coefficient * of each column (or row) of the referenced expression. * @@ -275,8 +290,8 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_minCoeff.out * * \sa DenseBase::minCoeff() */ - const typename ReturnType::Type minCoeff() const - { return _expression(); } + const MinCoeffReturnType minCoeff() const + { return MinCoeffReturnType(_expression()); } /** \returns a row (or column) vector expression of the largest coefficient * of each column (or row) of the referenced expression. @@ -287,8 +302,8 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_maxCoeff.out * * \sa DenseBase::maxCoeff() */ - const typename ReturnType::Type maxCoeff() const - { return _expression(); } + const MaxCoeffReturnType maxCoeff() const + { return MaxCoeffReturnType(_expression()); } /** \returns a row (or column) vector expression of the squared norm * of each column (or row) of the referenced expression. @@ -298,8 +313,8 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_squaredNorm.out * * \sa DenseBase::squaredNorm() */ - const typename ReturnType::Type squaredNorm() const - { return _expression(); } + const SquareNormReturnType squaredNorm() const + { return SquareNormReturnType(_expression()); } /** \returns a row (or column) vector expression of the norm * of each column (or row) of the referenced expression. @@ -309,8 +324,8 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_norm.out * * \sa DenseBase::norm() */ - const typename ReturnType::Type norm() const - { return _expression(); } + const NormReturnType norm() const + { return NormReturnType(_expression()); } /** \returns a row (or column) vector expression of the norm @@ -319,8 +334,8 @@ template class VectorwiseOp * This is a vector with real entries, even if the original matrix has complex entries. * * \sa DenseBase::blueNorm() */ - const typename ReturnType::Type blueNorm() const - { return _expression(); } + const BlueNormReturnType blueNorm() const + { return BlueNormReturnType(_expression()); } /** \returns a row (or column) vector expression of the norm @@ -329,8 +344,8 @@ template class VectorwiseOp * This is a vector with real entries, even if the original matrix has complex entries. * * \sa DenseBase::stableNorm() */ - const typename ReturnType::Type stableNorm() const - { return _expression(); } + const StableNormReturnType stableNorm() const + { return StableNormReturnType(_expression()); } /** \returns a row (or column) vector expression of the norm @@ -339,8 +354,8 @@ template class VectorwiseOp * This is a vector with real entries, even if the original matrix has complex entries. * * \sa DenseBase::hypotNorm() */ - const typename ReturnType::Type hypotNorm() const - { return _expression(); } + const HypotNormReturnType hypotNorm() const + { return HypotNormReturnType(_expression()); } /** \returns a row (or column) vector expression of the sum * of each column (or row) of the referenced expression. @@ -349,31 +364,31 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_sum.out * * \sa DenseBase::sum() */ - const typename ReturnType::Type sum() const - { return _expression(); } + const SumReturnType sum() const + { return SumReturnType(_expression()); } /** \returns a row (or column) vector expression of the mean * of each column (or row) of the referenced expression. * * \sa DenseBase::mean() */ - const typename ReturnType::Type mean() const - { return _expression(); } + const MeanReturnType mean() const + { return MeanReturnType(_expression()); } /** \returns a row (or column) vector expression representing * whether \b all coefficients of each respective column (or row) are \c true. * This expression can be assigned to a vector with entries of type \c bool. * * \sa DenseBase::all() */ - const typename ReturnType::Type all() const - { return _expression(); } + const AllReturnType all() const + { return AllReturnType(_expression()); } /** \returns a row (or column) vector expression representing * whether \b at \b least one coefficient of each respective column (or row) is \c true. * This expression can be assigned to a vector with entries of type \c bool. * * \sa DenseBase::any() */ - const typename ReturnType::Type any() const - { return _expression(); } + const AnyReturnType any() const + { return Any(_expression()); } /** \returns a row (or column) vector expression representing * the number of \c true coefficients of each respective column (or row). @@ -384,8 +399,8 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_count.out * * \sa DenseBase::count() */ - const PartialReduxExpr, Direction> count() const - { return _expression(); } + const CountReturnType count() const + { return CountReturnType(_expression()); } /** \returns a row (or column) vector expression of the product * of each column (or row) of the referenced expression. @@ -394,8 +409,8 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_prod.out * * \sa DenseBase::prod() */ - const typename ReturnType::Type prod() const - { return _expression(); } + const ProdReturnType prod() const + { return ProdReturnType(_expression()); } /** \returns a matrix expression @@ -405,8 +420,8 @@ template class VectorwiseOp * Output: \verbinclude Vectorwise_reverse.out * * \sa DenseBase::reverse() */ - const Reverse reverse() const - { return Reverse( _expression() ); } + const ReverseReturnType reverse() const + { return ReverseReturnType( _expression() ); } typedef Replicate ReplicateReturnType; const ReplicateReturnType replicate(Index factor) const; @@ -550,7 +565,8 @@ template class VectorwiseOp /////////// Geometry module /////////// - Homogeneous homogeneous() const; + typedef Homogeneous HomogeneousReturnType; + HomogeneousReturnType homogeneous() const; typedef typename ExpressionType::PlainObject CrossReturnType; template @@ -595,7 +611,7 @@ template inline const typename DenseBase::ConstColwiseReturnType DenseBase::colwise() const { - return derived(); + return ConstColwiseReturnType(derived()); } /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations @@ -606,7 +622,7 @@ template inline typename DenseBase::ColwiseReturnType DenseBase::colwise() { - return derived(); + return ColwiseReturnType(derived()); } /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations @@ -620,7 +636,7 @@ template inline const typename DenseBase::ConstRowwiseReturnType DenseBase::rowwise() const { - return derived(); + return ConstRowwiseReturnType(derived()); } /** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations @@ -631,7 +647,7 @@ template inline typename DenseBase::RowwiseReturnType DenseBase::rowwise() { - return derived(); + return RowwiseReturnType(derived()); } } // end namespace Eigen diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h index 810ec28e7..02bd4eff3 100644 --- a/Eigen/src/Core/Visitor.h +++ b/Eigen/src/Core/Visitor.h @@ -58,7 +58,7 @@ template class visitor_evaluator { public: - visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} + explicit visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; diff --git a/Eigen/src/Core/products/SelfadjointRank2Update.h b/Eigen/src/Core/products/SelfadjointRank2Update.h index 8594a97ce..2ae364111 100644 --- a/Eigen/src/Core/products/SelfadjointRank2Update.h +++ b/Eigen/src/Core/products/SelfadjointRank2Update.h @@ -79,11 +79,11 @@ SelfAdjointView& SelfAdjointView if (IsRowMajor) actualAlpha = numext::conj(actualAlpha); - internal::selfadjoint_rank2_update_selector::type>::type, - typename internal::remove_all::type>::type, + typedef typename internal::remove_all::type>::type UType; + typedef typename internal::remove_all::type>::type VType; + internal::selfadjoint_rank2_update_selector - ::run(_expression().const_cast_derived().data(),_expression().outerStride(),actualU,actualV,actualAlpha); + ::run(_expression().const_cast_derived().data(),_expression().outerStride(),UType(actualU),VType(actualV),actualAlpha); return *this; } diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h index 0d8e2705a..9f9115c2a 100644 --- a/Eigen/src/Core/util/BlasUtil.h +++ b/Eigen/src/Core/util/BlasUtil.h @@ -231,7 +231,7 @@ struct blas_traits > enum { IsTransposed = Base::IsTransposed ? 0 : 1 }; - static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); } + static inline ExtractType extract(const XprType& x) { return ExtractType(Base::extract(x.nestedExpression())); } static inline Scalar extractScalarFactor(const XprType& x) { return Base::extractScalarFactor(x.nestedExpression()); } }; diff --git a/Eigen/src/Eigenvalues/ComplexEigenSolver.h b/Eigen/src/Eigenvalues/ComplexEigenSolver.h index af434bc9b..25082546e 100644 --- a/Eigen/src/Eigenvalues/ComplexEigenSolver.h +++ b/Eigen/src/Eigenvalues/ComplexEigenSolver.h @@ -104,7 +104,7 @@ template class ComplexEigenSolver * according to the specified problem \a size. * \sa ComplexEigenSolver() */ - ComplexEigenSolver(Index size) + explicit ComplexEigenSolver(Index size) : m_eivec(size, size), m_eivalues(size), m_schur(size), @@ -122,7 +122,7 @@ template class ComplexEigenSolver * * This constructor calls compute() to compute the eigendecomposition. */ - ComplexEigenSolver(const MatrixType& matrix, bool computeEigenvectors = true) + explicit ComplexEigenSolver(const MatrixType& matrix, bool computeEigenvectors = true) : m_eivec(matrix.rows(),matrix.cols()), m_eivalues(matrix.cols()), m_schur(matrix.rows()), diff --git a/Eigen/src/Eigenvalues/ComplexSchur.h b/Eigen/src/Eigenvalues/ComplexSchur.h index 89e6cade3..a3a5a4649 100644 --- a/Eigen/src/Eigenvalues/ComplexSchur.h +++ b/Eigen/src/Eigenvalues/ComplexSchur.h @@ -91,7 +91,7 @@ template class ComplexSchur * * \sa compute() for an example. */ - ComplexSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) + explicit ComplexSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) : m_matT(size,size), m_matU(size,size), m_hess(size), @@ -109,7 +109,7 @@ template class ComplexSchur * * \sa matrixT() and matrixU() for examples. */ - ComplexSchur(const MatrixType& matrix, bool computeU = true) + explicit ComplexSchur(const MatrixType& matrix, bool computeU = true) : m_matT(matrix.rows(),matrix.cols()), m_matU(matrix.rows(),matrix.cols()), m_hess(matrix.rows()), diff --git a/Eigen/src/Eigenvalues/EigenSolver.h b/Eigen/src/Eigenvalues/EigenSolver.h index d2563d470..8a83b85bb 100644 --- a/Eigen/src/Eigenvalues/EigenSolver.h +++ b/Eigen/src/Eigenvalues/EigenSolver.h @@ -118,7 +118,7 @@ template class EigenSolver * according to the specified problem \a size. * \sa EigenSolver() */ - EigenSolver(Index size) + explicit EigenSolver(Index size) : m_eivec(size, size), m_eivalues(size), m_isInitialized(false), @@ -143,7 +143,7 @@ template class EigenSolver * * \sa compute() */ - EigenSolver(const MatrixType& matrix, bool computeEigenvectors = true) + explicit EigenSolver(const MatrixType& matrix, bool computeEigenvectors = true) : m_eivec(matrix.rows(), matrix.cols()), m_eivalues(matrix.cols()), m_isInitialized(false), diff --git a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h index dc240e13e..c20ea03e6 100644 --- a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h +++ b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h @@ -122,7 +122,7 @@ template class GeneralizedEigenSolver * according to the specified problem \a size. * \sa GeneralizedEigenSolver() */ - GeneralizedEigenSolver(Index size) + explicit GeneralizedEigenSolver(Index size) : m_eivec(size, size), m_alphas(size), m_betas(size), @@ -145,7 +145,7 @@ template class GeneralizedEigenSolver * * \sa compute() */ - GeneralizedEigenSolver(const MatrixType& A, const MatrixType& B, bool computeEigenvectors = true) + explicit GeneralizedEigenSolver(const MatrixType& A, const MatrixType& B, bool computeEigenvectors = true) : m_eivec(A.rows(), A.cols()), m_alphas(A.cols()), m_betas(A.cols()), diff --git a/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h index 07bf1ea09..1ce1f5f58 100644 --- a/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h @@ -74,7 +74,7 @@ class GeneralizedSelfAdjointEigenSolver : public SelfAdjointEigenSolver<_MatrixT * * \sa compute() for an example */ - GeneralizedSelfAdjointEigenSolver(Index size) + explicit GeneralizedSelfAdjointEigenSolver(Index size) : Base(size) {} diff --git a/Eigen/src/Eigenvalues/HessenbergDecomposition.h b/Eigen/src/Eigenvalues/HessenbergDecomposition.h index 3db0c0106..2615a9f23 100644 --- a/Eigen/src/Eigenvalues/HessenbergDecomposition.h +++ b/Eigen/src/Eigenvalues/HessenbergDecomposition.h @@ -97,7 +97,7 @@ template class HessenbergDecomposition * * \sa compute() for an example. */ - HessenbergDecomposition(Index size = Size==Dynamic ? 2 : Size) + explicit HessenbergDecomposition(Index size = Size==Dynamic ? 2 : Size) : m_matrix(size,size), m_temp(size), m_isInitialized(false) @@ -115,7 +115,7 @@ template class HessenbergDecomposition * * \sa matrixH() for an example. */ - HessenbergDecomposition(const MatrixType& matrix) + explicit HessenbergDecomposition(const MatrixType& matrix) : m_matrix(matrix), m_temp(matrix.rows()), m_isInitialized(false) diff --git a/Eigen/src/Eigenvalues/RealQZ.h b/Eigen/src/Eigenvalues/RealQZ.h index 5706eeebe..ae10ff91e 100644 --- a/Eigen/src/Eigenvalues/RealQZ.h +++ b/Eigen/src/Eigenvalues/RealQZ.h @@ -83,7 +83,7 @@ namespace Eigen { * * \sa compute() for an example. */ - RealQZ(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) : + explicit RealQZ(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) : m_S(size, size), m_T(size, size), m_Q(size, size), @@ -101,7 +101,7 @@ namespace Eigen { * * This constructor calls compute() to compute the QZ decomposition. */ - RealQZ(const MatrixType& A, const MatrixType& B, bool computeQZ = true) : + explicit RealQZ(const MatrixType& A, const MatrixType& B, bool computeQZ = true) : m_S(A.rows(),A.cols()), m_T(A.rows(),A.cols()), m_Q(A.rows(),A.cols()), diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h index 64d136341..10f5fb174 100644 --- a/Eigen/src/Eigenvalues/RealSchur.h +++ b/Eigen/src/Eigenvalues/RealSchur.h @@ -80,7 +80,7 @@ template class RealSchur * * \sa compute() for an example. */ - RealSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) + explicit RealSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) : m_matT(size, size), m_matU(size, size), m_workspaceVector(size), @@ -100,7 +100,7 @@ template class RealSchur * Example: \include RealSchur_RealSchur_MatrixType.cpp * Output: \verbinclude RealSchur_RealSchur_MatrixType.out */ - RealSchur(const MatrixType& matrix, bool computeU = true) + explicit RealSchur(const MatrixType& matrix, bool computeU = true) : m_matT(matrix.rows(),matrix.cols()), m_matU(matrix.rows(),matrix.cols()), m_workspaceVector(matrix.rows()), diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index a6bbdac6b..1dd2ab45b 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -133,7 +133,7 @@ template class SelfAdjointEigenSolver * \sa compute() for an example */ EIGEN_DEVICE_FUNC - SelfAdjointEigenSolver(Index size) + explicit SelfAdjointEigenSolver(Index size) : m_eivec(size, size), m_eivalues(size), m_subdiag(size > 1 ? size - 1 : 1), @@ -156,7 +156,7 @@ template class SelfAdjointEigenSolver * \sa compute(const MatrixType&, int) */ EIGEN_DEVICE_FUNC - SelfAdjointEigenSolver(const MatrixType& matrix, int options = ComputeEigenvectors) + explicit SelfAdjointEigenSolver(const MatrixType& matrix, int options = ComputeEigenvectors) : m_eivec(matrix.rows(), matrix.cols()), m_eivalues(matrix.cols()), m_subdiag(matrix.rows() > 1 ? matrix.rows() - 1 : 1), diff --git a/Eigen/src/Eigenvalues/Tridiagonalization.h b/Eigen/src/Eigenvalues/Tridiagonalization.h index e3a27f275..f20528f86 100644 --- a/Eigen/src/Eigenvalues/Tridiagonalization.h +++ b/Eigen/src/Eigenvalues/Tridiagonalization.h @@ -112,7 +112,7 @@ template class Tridiagonalization * * \sa compute() for an example. */ - Tridiagonalization(Index size = Size==Dynamic ? 2 : Size) + explicit Tridiagonalization(Index size = Size==Dynamic ? 2 : Size) : m_matrix(size,size), m_hCoeffs(size > 1 ? size-1 : 1), m_isInitialized(false) @@ -128,7 +128,7 @@ template class Tridiagonalization * Example: \include Tridiagonalization_Tridiagonalization_MatrixType.cpp * Output: \verbinclude Tridiagonalization_Tridiagonalization_MatrixType.out */ - Tridiagonalization(const MatrixType& matrix) + explicit Tridiagonalization(const MatrixType& matrix) : m_matrix(matrix), m_hCoeffs(matrix.cols() > 1 ? matrix.cols()-1 : 1), m_isInitialized(false) diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index d1881d84d..ede203ef9 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -68,7 +68,7 @@ template class Homogeneous typedef MatrixBase Base; EIGEN_DENSE_PUBLIC_INTERFACE(Homogeneous) - inline Homogeneous(const MatrixType& matrix) + explicit inline Homogeneous(const MatrixType& matrix) : m_matrix(matrix) {} @@ -128,7 +128,7 @@ inline typename MatrixBase::HomogeneousReturnType MatrixBase::homogeneous() const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived); - return derived(); + return HomogeneousReturnType(derived()); } /** \geometry_module @@ -143,7 +143,7 @@ template inline Homogeneous VectorwiseOp::homogeneous() const { - return _expression(); + return HomogeneousReturnType(_expression()); } /** \geometry_module @@ -323,7 +323,7 @@ struct unary_evaluator, IndexBased> typedef evaluator type; typedef evaluator nestedType; - unary_evaluator(const XprType& op) + explicit unary_evaluator(const XprType& op) : Base(), m_temp(op) { ::new (static_cast(this)) Base(m_temp); diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index 3f0067286..216e5b12f 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -251,7 +251,7 @@ public: inline Quaternion(const Scalar& w, const Scalar& x, const Scalar& y, const Scalar& z) : m_coeffs(x, y, z, w){} /** Constructs and initialize a quaternion from the array data */ - inline Quaternion(const Scalar* data) : m_coeffs(data) {} + explicit inline Quaternion(const Scalar* data) : m_coeffs(data) {} /** Copy constructor */ template EIGEN_STRONG_INLINE Quaternion(const QuaternionBase& other) { this->Base::operator=(other); } @@ -351,7 +351,7 @@ class Map, _Options > * \code *coeffs == {x, y, z, w} \endcode * * If the template parameter _Options is set to #Aligned, then the pointer coeffs must be aligned. */ - EIGEN_STRONG_INLINE Map(const Scalar* coeffs) : m_coeffs(coeffs) {} + explicit EIGEN_STRONG_INLINE Map(const Scalar* coeffs) : m_coeffs(coeffs) {} inline const Coefficients& coeffs() const { return m_coeffs;} @@ -388,7 +388,7 @@ class Map, _Options > * \code *coeffs == {x, y, z, w} \endcode * * If the template parameter _Options is set to #Aligned, then the pointer coeffs must be aligned. */ - EIGEN_STRONG_INLINE Map(Scalar* coeffs) : m_coeffs(coeffs) {} + explicit EIGEN_STRONG_INLINE Map(Scalar* coeffs) : m_coeffs(coeffs) {} inline Coefficients& coeffs() { return m_coeffs; } inline const Coefficients& coeffs() const { return m_coeffs; } diff --git a/Eigen/src/Geometry/Rotation2D.h b/Eigen/src/Geometry/Rotation2D.h index 1cac343a5..c82762acd 100644 --- a/Eigen/src/Geometry/Rotation2D.h +++ b/Eigen/src/Geometry/Rotation2D.h @@ -59,7 +59,7 @@ protected: public: /** Construct a 2D counter clock wise rotation from the angle \a a in radian. */ - inline Rotation2D(const Scalar& a) : m_angle(a) {} + explicit inline Rotation2D(const Scalar& a) : m_angle(a) {} /** \returns the rotation angle */ inline Scalar angle() const { return m_angle; } diff --git a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h index 98b169868..3991afa8f 100644 --- a/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h +++ b/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h @@ -43,7 +43,7 @@ class DiagonalPreconditioner DiagonalPreconditioner() : m_isInitialized(false) {} template - DiagonalPreconditioner(const MatType& mat) : m_invdiag(mat.cols()) + explicit DiagonalPreconditioner(const MatType& mat) : m_invdiag(mat.cols()) { compute(mat); } @@ -114,7 +114,7 @@ class IdentityPreconditioner IdentityPreconditioner() {} template - IdentityPreconditioner(const MatrixType& ) {} + explicit IdentityPreconditioner(const MatrixType& ) {} template IdentityPreconditioner& analyzePattern(const MatrixType& ) { return *this; } diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index 42da7d14b..224fe913f 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -178,7 +178,7 @@ public: * this class becomes invalid. Call compute() to update it with the new * matrix A, or modify a copy of A. */ - BiCGSTAB(const MatrixType& A) : Base(A) {} + explicit BiCGSTAB(const MatrixType& A) : Base(A) {} ~BiCGSTAB() {} diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index f72cf86a5..b5ef6d60f 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -189,7 +189,7 @@ public: * this class becomes invalid. Call compute() to update it with the new * matrix A, or modify a copy of A. */ - ConjugateGradient(const MatrixType& A) : Base(A) {} + explicit ConjugateGradient(const MatrixType& A) : Base(A) {} ~ConjugateGradient() {} diff --git a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h index 7adbbc489..8ed9bdecc 100644 --- a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h +++ b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h @@ -116,7 +116,7 @@ class IncompleteLUT : public SparseSolverBase > {} template - IncompleteLUT(const MatrixType& mat, const RealScalar& droptol=NumTraits::dummy_precision(), int fillfactor = 10) + explicit IncompleteLUT(const MatrixType& mat, const RealScalar& droptol=NumTraits::dummy_precision(), int fillfactor = 10) : m_droptol(droptol),m_fillfactor(fillfactor), m_analysisIsOk(false),m_factorizationIsOk(false) { diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h index 1ca7e0561..f33c868bb 100644 --- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h @@ -52,7 +52,7 @@ public: * this class becomes invalid. Call compute() to update it with the new * matrix A, or modify a copy of A. */ - IterativeSolverBase(const MatrixType& A) + explicit IterativeSolverBase(const MatrixType& A) { init(); compute(A); diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index fdf2e0642..96f2cebee 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -94,7 +94,7 @@ template class FullPivLU * \param matrix the matrix of which to compute the LU decomposition. * It is required to be nonzero. */ - FullPivLU(const MatrixType& matrix); + explicit FullPivLU(const MatrixType& matrix); /** Computes the LU decomposition of the given matrix. * diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index a4d22ce5f..d04e4191b 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -92,7 +92,7 @@ template class PartialPivLU * according to the specified problem \a size. * \sa PartialPivLU() */ - PartialPivLU(Index size); + explicit PartialPivLU(Index size); /** Constructor. * @@ -101,7 +101,7 @@ template class PartialPivLU * \warning The matrix should have full rank (e.g. if it's square, it should be invertible). * If you need to deal with non-full rank, use class FullPivLU instead. */ - PartialPivLU(const MatrixType& matrix); + explicit PartialPivLU(const MatrixType& matrix); PartialPivLU& compute(const MatrixType& matrix); diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h index bb8e0d1a8..a96c27695 100644 --- a/Eigen/src/PaStiXSupport/PaStiXSupport.h +++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h @@ -417,7 +417,7 @@ class PastixLU : public PastixBase< PastixLU<_MatrixType> > init(); } - PastixLU(const MatrixType& matrix):Base() + explicit PastixLU(const MatrixType& matrix):Base() { init(); compute(matrix); @@ -527,7 +527,7 @@ class PastixLLT : public PastixBase< PastixLLT<_MatrixType, _UpLo> > init(); } - PastixLLT(const MatrixType& matrix):Base() + explicit PastixLLT(const MatrixType& matrix):Base() { init(); compute(matrix); @@ -608,7 +608,7 @@ class PastixLDLT : public PastixBase< PastixLDLT<_MatrixType, _UpLo> > init(); } - PastixLDLT(const MatrixType& matrix):Base() + explicit PastixLDLT(const MatrixType& matrix):Base() { init(); compute(matrix); diff --git a/Eigen/src/PardisoSupport/PardisoSupport.h b/Eigen/src/PardisoSupport/PardisoSupport.h index e1b0e1818..054af6635 100644 --- a/Eigen/src/PardisoSupport/PardisoSupport.h +++ b/Eigen/src/PardisoSupport/PardisoSupport.h @@ -391,7 +391,7 @@ class PardisoLU : public PardisoImpl< PardisoLU > pardisoInit(Base::ScalarIsComplex ? 13 : 11); } - PardisoLU(const MatrixType& matrix) + explicit PardisoLU(const MatrixType& matrix) : Base() { pardisoInit(Base::ScalarIsComplex ? 13 : 11); @@ -442,7 +442,7 @@ class PardisoLLT : public PardisoImpl< PardisoLLT > pardisoInit(Base::ScalarIsComplex ? 4 : 2); } - PardisoLLT(const MatrixType& matrix) + explicit PardisoLLT(const MatrixType& matrix) : Base() { pardisoInit(Base::ScalarIsComplex ? 4 : 2); @@ -500,7 +500,7 @@ class PardisoLDLT : public PardisoImpl< PardisoLDLT > pardisoInit(Base::ScalarIsComplex ? ( bool(Options&Symmetric) ? 6 : -4 ) : -2); } - PardisoLDLT(const MatrixType& matrix) + explicit PardisoLDLT(const MatrixType& matrix) : Base() { pardisoInit(Base::ScalarIsComplex ? ( bool(Options&Symmetric) ? 6 : -4 ) : -2); diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index adf737276..de77e8411 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -117,7 +117,7 @@ template class ColPivHouseholderQR * * \sa compute() */ - ColPivHouseholderQR(const MatrixType& matrix) + explicit ColPivHouseholderQR(const MatrixType& matrix) : m_qr(matrix.rows(), matrix.cols()), m_hCoeffs((std::min)(matrix.rows(),matrix.cols())), m_colsPermutation(PermIndexType(matrix.cols())), diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 710c64a45..5712d175c 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -120,7 +120,7 @@ template class FullPivHouseholderQR * * \sa compute() */ - FullPivHouseholderQR(const MatrixType& matrix) + explicit FullPivHouseholderQR(const MatrixType& matrix) : m_qr(matrix.rows(), matrix.cols()), m_hCoeffs((std::min)(matrix.rows(), matrix.cols())), m_rows_transpositions((std::min)(matrix.rows(), matrix.cols())), diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index 0b0c9d1bd..f22008494 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -91,7 +91,7 @@ template class HouseholderQR * * \sa compute() */ - HouseholderQR(const MatrixType& matrix) + explicit HouseholderQR(const MatrixType& matrix) : m_qr(matrix.rows(), matrix.cols()), m_hCoeffs((std::min)(matrix.rows(),matrix.cols())), m_temp(matrix.cols()), diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h index bcdc981d7..44f6a1acb 100644 --- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h @@ -65,7 +65,7 @@ class SPQR : public SparseSolverBase > typedef typename _MatrixType::RealScalar RealScalar; typedef UF_long Index ; typedef SparseMatrix MatrixType; - typedef PermutationMatrix PermutationType; + typedef Map > PermutationType; public: SPQR() : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) @@ -73,7 +73,7 @@ class SPQR : public SparseSolverBase > cholmod_l_start(&m_cc); } - SPQR(const _MatrixType& matrix) + explicit SPQR(const _MatrixType& matrix) : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) { cholmod_l_start(&m_cc); @@ -164,11 +164,7 @@ class SPQR : public SparseSolverBase > PermutationType colsPermutation() const { eigen_assert(m_isInitialized && "Decomposition is not initialized."); - Index n = m_cR->ncol; - PermutationType colsPerm(n); - for(Index j = 0; j ncol); } /** * Gets the rank of the matrix. diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index f2a72faa3..59f88a15a 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -550,7 +550,7 @@ template class JacobiSVD * according to the specified problem size. * \sa JacobiSVD() */ - JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0) + explicit JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0) { allocate(rows, cols, computationOptions); } @@ -565,7 +565,7 @@ template class JacobiSVD * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not * available with the (non-default) FullPivHouseholderQR preconditioner. */ - JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0) + explicit JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0) { compute(matrix, computationOptions); } diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 40b1237a0..eaa6bb86e 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -37,7 +37,7 @@ template class UpperBidiagonalization typedef Matrix SuperDiagVectorType; typedef HouseholderSequence< const MatrixType, - CwiseUnaryOp, const Diagonal > + const typename internal::remove_all::ConjugateReturnType>::type > HouseholderUSequenceType; typedef HouseholderSequence< const typename internal::remove_all::type, @@ -53,7 +53,7 @@ template class UpperBidiagonalization */ UpperBidiagonalization() : m_householder(), m_bidiagonal(), m_isInitialized(false) {} - UpperBidiagonalization(const MatrixType& matrix) + explicit UpperBidiagonalization(const MatrixType& matrix) : m_householder(matrix.rows(), matrix.cols()), m_bidiagonal(matrix.cols(), matrix.cols()), m_isInitialized(false) diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index 3c8cef5db..0e8fa6628 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -57,7 +57,7 @@ class SimplicialCholeskyBase : public SparseSolverBase : m_info(Success), m_shiftOffset(0), m_shiftScale(1) {} - SimplicialCholeskyBase(const MatrixType& matrix) + explicit SimplicialCholeskyBase(const MatrixType& matrix) : m_info(Success), m_shiftOffset(0), m_shiftScale(1) { derived().compute(matrix); @@ -239,8 +239,8 @@ template struct traits CholMatrixType; typedef TriangularView MatrixL; typedef TriangularView MatrixU; - static inline MatrixL getL(const MatrixType& m) { return m; } - static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); } + static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); } + static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); } }; template struct traits > @@ -253,8 +253,8 @@ template struct traits CholMatrixType; typedef TriangularView MatrixL; typedef TriangularView MatrixU; - static inline MatrixL getL(const MatrixType& m) { return m; } - static inline MatrixU getU(const MatrixType& m) { return m.adjoint(); } + static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); } + static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); } }; template struct traits > @@ -303,7 +303,7 @@ public: /** Default constructor */ SimplicialLLT() : Base() {} /** Constructs and performs the LLT factorization of \a matrix */ - SimplicialLLT(const MatrixType& matrix) + explicit SimplicialLLT(const MatrixType& matrix) : Base(matrix) {} /** \returns an expression of the factor L */ @@ -393,7 +393,7 @@ public: SimplicialLDLT() : Base() {} /** Constructs and performs the LLT factorization of \a matrix */ - SimplicialLDLT(const MatrixType& matrix) + explicit SimplicialLDLT(const MatrixType& matrix) : Base(matrix) {} /** \returns a vector expression of the diagonal D */ @@ -473,7 +473,7 @@ public: public: SimplicialCholesky() : Base(), m_LDLT(true) {} - SimplicialCholesky(const MatrixType& matrix) + explicit SimplicialCholesky(const MatrixType& matrix) : Base(), m_LDLT(true) { compute(matrix); diff --git a/Eigen/src/SparseCore/AmbiVector.h b/Eigen/src/SparseCore/AmbiVector.h index 17fff96a7..5c9c3101e 100644 --- a/Eigen/src/SparseCore/AmbiVector.h +++ b/Eigen/src/SparseCore/AmbiVector.h @@ -27,7 +27,7 @@ class AmbiVector typedef _Index Index; typedef typename NumTraits::Real RealScalar; - AmbiVector(Index size) + explicit AmbiVector(Index size) : m_buffer(0), m_zero(0), m_size(0), m_allocatedSize(0), m_allocatedElements(0), m_mode(-1) { resize(size); @@ -288,7 +288,7 @@ class AmbiVector<_Scalar,_Index>::Iterator * In practice, all coefficients having a magnitude smaller than \a epsilon * are skipped. */ - Iterator(const AmbiVector& vec, const RealScalar& epsilon = 0) + explicit Iterator(const AmbiVector& vec, const RealScalar& epsilon = 0) : m_vector(vec) { using std::abs; diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index a667cb56e..a93550340 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -36,7 +36,7 @@ class CompressedStorage : m_values(0), m_indices(0), m_size(0), m_allocatedSize(0) {} - CompressedStorage(size_t size) + explicit CompressedStorage(size_t size) : m_values(0), m_indices(0), m_size(0), m_allocatedSize(0) { resize(size); diff --git a/Eigen/src/SparseCore/MappedSparseMatrix.h b/Eigen/src/SparseCore/MappedSparseMatrix.h index d9aabd049..2852c669a 100644 --- a/Eigen/src/SparseCore/MappedSparseMatrix.h +++ b/Eigen/src/SparseCore/MappedSparseMatrix.h @@ -192,7 +192,7 @@ struct evaluator > }; evaluator() : m_matrix(0) {} - evaluator(const MappedSparseMatrixType &mat) : m_matrix(&mat) {} + explicit evaluator(const MappedSparseMatrixType &mat) : m_matrix(&mat) {} operator MappedSparseMatrixType&() { return m_matrix->const_cast_derived(); } operator const MappedSparseMatrixType&() const { return *m_matrix; } diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 635d58d86..f8b9d5ad6 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -443,7 +443,7 @@ namespace internal { Index m_end; public: - EIGEN_STRONG_INLINE GenericSparseBlockInnerIteratorImpl(const BlockType& block, Index outer = 0) + explicit EIGEN_STRONG_INLINE GenericSparseBlockInnerIteratorImpl(const BlockType& block, Index outer = 0) : m_block(block), m_outerPos( (IsRowMajor ? block.m_startCol.value() : block.m_startRow.value()) - 1), // -1 so that operator++ finds the first non-zero entry @@ -512,7 +512,7 @@ struct unary_evaluator, IteratorBa typedef typename internal::conditional::type InnerIterator; - unary_evaluator(const XprType& op) + explicit unary_evaluator(const XprType& op) : m_argImpl(op.nestedExpression()), m_block(op) {} diff --git a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h index 5993c1caf..94ca9b1a4 100644 --- a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h @@ -117,7 +117,7 @@ public: Flags = XprType::Flags }; - binary_evaluator(const XprType& xpr) + explicit binary_evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) @@ -195,7 +195,7 @@ public: Flags = XprType::Flags }; - binary_evaluator(const XprType& xpr) + explicit binary_evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) @@ -260,7 +260,7 @@ public: Flags = XprType::Flags }; - binary_evaluator(const XprType& xpr) + explicit binary_evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) @@ -326,7 +326,7 @@ public: Flags = XprType::Flags }; - binary_evaluator(const XprType& xpr) + explicit binary_evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) diff --git a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h index 6036fd0a7..32b7bc949 100644 --- a/Eigen/src/SparseCore/SparseCwiseUnaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseUnaryOp.h @@ -29,7 +29,7 @@ struct unary_evaluator, IteratorBased> Flags = XprType::Flags }; - unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) {} + explicit unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) {} protected: typedef typename evaluator::InnerIterator EvalIterator; @@ -104,7 +104,7 @@ struct unary_evaluator, IteratorBased> Flags = XprType::Flags }; - unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) {} + explicit unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) {} protected: typedef typename evaluator::InnerIterator EvalIterator; diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 04c838a71..7af6a12a9 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -173,7 +173,8 @@ protected: typedef Product ProdXprType; // if the actual left-hand side is a dense vector, - // then build a sparse-view so that we can seamlessly iterator over it. + // then build a sparse-view so that we can seamlessly iterate over it. + // FIXME ActualLhs does not seem to be necessary typedef typename conditional::StorageKind,Sparse>::value, Lhs1, SparseView >::type ActualLhs; typedef typename conditional::StorageKind,Sparse>::value, @@ -228,7 +229,7 @@ public: Scalar m_factor; }; - sparse_dense_outer_product_evaluator(const ActualLhs &lhs, const ActualRhs &rhs) + sparse_dense_outer_product_evaluator(const Lhs1 &lhs, const ActualRhs &rhs) : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs) {} @@ -253,7 +254,7 @@ struct product_evaluator, OuterProduct, Sparse typedef Product XprType; typedef typename XprType::PlainObject PlainObject; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs()) {} @@ -268,7 +269,7 @@ struct product_evaluator, OuterProduct, DenseS typedef Product XprType; typedef typename XprType::PlainObject PlainObject; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs()) {} diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index 0cb2bd572..be935e9f3 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -44,7 +44,7 @@ struct product_evaluator, ProductTag, Diagonal enum { CoeffReadCost = Dynamic, Flags = Rhs::Flags&RowMajorBit }; // FIXME CoeffReadCost & Flags typedef sparse_diagonal_product_evaluator Base; - product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) {} + explicit product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) {} }; template @@ -57,7 +57,7 @@ struct product_evaluator, ProductTag, SparseSh enum { CoeffReadCost = Dynamic, Flags = Lhs::Flags&RowMajorBit }; // FIXME CoeffReadCost & Flags typedef sparse_diagonal_product_evaluator, Lhs::Flags&RowMajorBit?SDP_AsCwiseProduct:SDP_AsScalarProduct> Base; - product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal()) {} + explicit product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal().transpose()) {} }; template diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 9e7124ff2..7ccdd12a5 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -89,6 +89,8 @@ class SparseMatrix EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(SparseMatrix, -=) typedef MappedSparseMatrix Map; + typedef Diagonal DiagonalReturnType; + using Base::IsRowMajor; typedef internal::CompressedStorage Storage; enum { @@ -621,7 +623,7 @@ class SparseMatrix } /** \returns a const expression of the diagonal coefficients */ - const Diagonal diagonal() const { return *this; } + const DiagonalReturnType diagonal() const { return DiagonalReturnType(*this); } /** Default constructor yielding an empty \c 0 \c x \c 0 matrix */ inline SparseMatrix() @@ -1272,7 +1274,7 @@ struct evaluator > }; evaluator() : m_matrix(0) {} - evaluator(const SparseMatrixType &mat) : m_matrix(&mat) {} + explicit evaluator(const SparseMatrixType &mat) : m_matrix(&mat) {} operator SparseMatrixType&() { return m_matrix->const_cast_derived(); } operator const SparseMatrixType&() const { return *m_matrix; } diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index a438c6487..f96042f27 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -94,6 +94,9 @@ template class SparseMatrixBase : public EigenBase CwiseUnaryOp, Eigen::Transpose >, Transpose >::type AdjointReturnType; + typedef Transpose TransposeReturnType; + template struct SelfAdjointViewReturnType { typedef SelfAdjointView Type; }; + typedef typename internal::add_const >::type ConstTransposeReturnType; // FIXME storage order do not match evaluator storage order typedef SparseMatrix PlainObject; @@ -114,6 +117,8 @@ template class SparseMatrixBase : public EigenBase /** \internal Represents a matrix with all coefficients equal to one another*/ typedef CwiseNullaryOp,Matrix > ConstantReturnType; + /** type of the equivalent dense matrix */ + typedef Matrix DenseMatrixType; /** type of the equivalent square matrix */ typedef Matrix SquareMatrixType; @@ -303,9 +308,9 @@ template class SparseMatrixBase : public EigenBase RealScalar norm() const; RealScalar blueNorm() const; - Transpose transpose() { return derived(); } - const Transpose transpose() const { return derived(); } - const AdjointReturnType adjoint() const { return transpose(); } + TransposeReturnType transpose() { return TransposeReturnType(derived()); } + const ConstTransposeReturnType transpose() const { return ConstTransposeReturnType(derived()); } + const AdjointReturnType adjoint() const { return AdjointReturnType(transpose()); } // inner-vector typedef Block InnerVectorReturnType; @@ -317,9 +322,9 @@ template class SparseMatrixBase : public EigenBase Block innerVectors(Index outerStart, Index outerSize); const Block innerVectors(Index outerStart, Index outerSize) const; - Matrix toDense() const + DenseMatrixType toDense() const { - return derived(); + return DenseMatrixType(derived()); } template diff --git a/Eigen/src/SparseCore/SparsePermutation.h b/Eigen/src/SparseCore/SparsePermutation.h index 228796bf8..3bfa2ae1b 100644 --- a/Eigen/src/SparseCore/SparsePermutation.h +++ b/Eigen/src/SparseCore/SparsePermutation.h @@ -166,7 +166,7 @@ struct product_evaluator, ProductTag, Permutat typedef typename traits >::ReturnType PlainObject; typedef typename evaluator::type Base; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); @@ -185,7 +185,7 @@ struct product_evaluator, ProductTag, SparseSh typedef typename traits >::ReturnType PlainObject; typedef typename evaluator::type Base; - product_evaluator(const XprType& xpr) + explicit product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index b68fe986a..ae707376a 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -59,7 +59,7 @@ struct evaluator > > typedef evaluator type; typedef evaluator nestedType; - evaluator(const XprType& xpr) + explicit evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { using std::abs; diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 69ac1a398..247d4e6d3 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -53,7 +53,7 @@ template class SparseSelfAdjointView typedef typename MatrixType::Nested MatrixTypeNested; typedef typename internal::remove_all::type _MatrixTypeNested; - inline SparseSelfAdjointView(const MatrixType& matrix) : m_matrix(matrix) + explicit inline SparseSelfAdjointView(const MatrixType& matrix) : m_matrix(matrix) { eigen_assert(rows()==cols() && "SelfAdjointView is only for squared matrices"); } @@ -172,14 +172,14 @@ template template const SparseSelfAdjointView SparseMatrixBase::selfadjointView() const { - return derived(); + return SparseSelfAdjointView(derived()); } template template SparseSelfAdjointView SparseMatrixBase::selfadjointView() { - return derived(); + return SparseSelfAdjointView(derived()); } /*************************************************************************** diff --git a/Eigen/src/SparseCore/SparseTranspose.h b/Eigen/src/SparseCore/SparseTranspose.h index fae7cae97..c3d2d1a16 100644 --- a/Eigen/src/SparseCore/SparseTranspose.h +++ b/Eigen/src/SparseCore/SparseTranspose.h @@ -62,7 +62,7 @@ struct unary_evaluator, IteratorBased> Flags = XprType::Flags }; - unary_evaluator(const XprType& op) :m_argImpl(op.nestedExpression()) {} + explicit unary_evaluator(const XprType& op) :m_argImpl(op.nestedExpression()) {} protected: typename evaluator::nestedType m_argImpl; diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index 744c3d730..e200bc815 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -185,7 +185,7 @@ public: Flags = XprType::Flags }; - unary_evaluator(const XprType &xpr) : m_argImpl(xpr.nestedExpression()) {} + explicit unary_evaluator(const XprType &xpr) : m_argImpl(xpr.nestedExpression()) {} class InnerIterator : public EvalIterator { @@ -269,7 +269,7 @@ template inline const TriangularView SparseMatrixBase::triangularView() const { - return derived(); + return TriangularView(derived()); } } // end namespace Eigen diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index c9f9d61e9..dfbb2be2e 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -221,7 +221,7 @@ class SparseVector inline SparseVector() : m_size(0) { check_template_parameters(); resize(0); } - inline SparseVector(Index size) : m_size(0) { check_template_parameters(); resize(size); } + explicit inline SparseVector(Index size) : m_size(0) { check_template_parameters(); resize(size); } inline SparseVector(Index rows, Index cols) : m_size(0) { check_template_parameters(); resize(rows,cols); } @@ -360,14 +360,14 @@ template class SparseVector::InnerIterator { public: - InnerIterator(const SparseVector& vec, Index outer=0) + explicit InnerIterator(const SparseVector& vec, Index outer=0) : m_data(vec.m_data), m_id(0), m_end(static_cast(m_data.size())) { EIGEN_UNUSED_VARIABLE(outer); eigen_assert(outer==0); } - InnerIterator(const internal::CompressedStorage& data) + explicit InnerIterator(const internal::CompressedStorage& data) : m_data(data), m_id(0), m_end(static_cast(m_data.size())) {} @@ -392,14 +392,14 @@ template class SparseVector::ReverseInnerIterator { public: - ReverseInnerIterator(const SparseVector& vec, Index outer=0) + explicit ReverseInnerIterator(const SparseVector& vec, Index outer=0) : m_data(vec.m_data), m_id(static_cast(m_data.size())), m_start(0) { EIGEN_UNUSED_VARIABLE(outer); eigen_assert(outer==0); } - ReverseInnerIterator(const internal::CompressedStorage& data) + explicit ReverseInnerIterator(const internal::CompressedStorage& data) : m_data(data), m_id(static_cast(m_data.size())), m_start(0) {} @@ -435,7 +435,7 @@ struct evaluator > Flags = SparseVectorType::Flags }; - evaluator(const SparseVectorType &mat) : m_matrix(mat) {} + explicit evaluator(const SparseVectorType &mat) : m_matrix(mat) {} operator SparseVectorType&() { return m_matrix.const_cast_derived(); } operator const SparseVectorType&() const { return m_matrix; } diff --git a/Eigen/src/SparseCore/SparseView.h b/Eigen/src/SparseCore/SparseView.h index d10cc5a35..40a3019fa 100644 --- a/Eigen/src/SparseCore/SparseView.h +++ b/Eigen/src/SparseCore/SparseView.h @@ -36,7 +36,7 @@ public: EIGEN_SPARSE_PUBLIC_INTERFACE(SparseView) typedef typename internal::remove_all::type NestedExpression; - SparseView(const MatrixType& mat, const Scalar& m_reference = Scalar(0), + explicit SparseView(const MatrixType& mat, const Scalar& m_reference = Scalar(0), RealScalar m_epsilon = NumTraits::dummy_precision()) : m_matrix(mat), m_reference(m_reference), m_epsilon(m_epsilon) {} @@ -111,7 +111,7 @@ struct unary_evaluator, IteratorBased> Flags = XprType::Flags }; - unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {} + explicit unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {} protected: typename evaluator::nestedType m_argImpl; @@ -180,7 +180,7 @@ struct unary_evaluator, IndexBased> Flags = XprType::Flags }; - unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {} + explicit unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {} protected: typename evaluator::nestedType m_argImpl; diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 14d7e713e..cd4ea5b13 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -95,7 +95,7 @@ class SparseLU : public SparseSolverBase >, { initperfvalues(); } - SparseLU(const MatrixType& matrix):m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1) + explicit SparseLU(const MatrixType& matrix):m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1) { initperfvalues(); compute(matrix); @@ -650,7 +650,7 @@ struct SparseLUMatrixLReturnType : internal::no_assignment_operator { typedef typename MappedSupernodalType::Index Index; typedef typename MappedSupernodalType::Scalar Scalar; - SparseLUMatrixLReturnType(const MappedSupernodalType& mapL) : m_mapL(mapL) + explicit SparseLUMatrixLReturnType(const MappedSupernodalType& mapL) : m_mapL(mapL) { } Index rows() { return m_mapL.rows(); } Index cols() { return m_mapL.cols(); } @@ -667,7 +667,7 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator { typedef typename MatrixLType::Index Index; typedef typename MatrixLType::Scalar Scalar; - SparseLUMatrixUReturnType(const MatrixLType& mapL, const MatrixUType& mapU) + explicit SparseLUMatrixUReturnType(const MatrixLType& mapL, const MatrixUType& mapU) : m_mapL(mapL),m_mapU(mapU) { } Index rows() { return m_mapL.rows(); } diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 6d85ea9be..f6a9af672 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -88,7 +88,7 @@ class SparseQR : public SparseSolverBase > * * \sa compute() */ - SparseQR(const MatrixType& mat) : m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) + explicit SparseQR(const MatrixType& mat) : m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false) { compute(mat); } @@ -636,7 +636,7 @@ struct SparseQRMatrixQReturnType : public EigenBase DenseMatrix; - SparseQRMatrixQReturnType(const SparseQRType& qr) : m_qr(qr) {} + explicit SparseQRMatrixQReturnType(const SparseQRType& qr) : m_qr(qr) {} template SparseQR_QProduct operator*(const MatrixBase& other) { @@ -672,7 +672,7 @@ struct SparseQRMatrixQReturnType : public EigenBase struct SparseQRMatrixQTransposeReturnType { - SparseQRMatrixQTransposeReturnType(const SparseQRType& qr) : m_qr(qr) {} + explicit SparseQRMatrixQTransposeReturnType(const SparseQRType& qr) : m_qr(qr) {} template SparseQR_QProduct operator*(const MatrixBase& other) { diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h index 0137585ca..ef73587a7 100644 --- a/Eigen/src/SuperLUSupport/SuperLUSupport.h +++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h @@ -469,7 +469,7 @@ class SuperLU : public SuperLUBase<_MatrixType,SuperLU<_MatrixType> > SuperLU() : Base() { init(); } - SuperLU(const MatrixType& matrix) : Base() + explicit SuperLU(const MatrixType& matrix) : Base() { init(); Base::compute(matrix); diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h index 7fada5567..1c5800ffa 100644 --- a/Eigen/src/UmfPackSupport/UmfPackSupport.h +++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h @@ -142,7 +142,7 @@ class UmfPackLU : public SparseSolverBase > UmfPackLU() { init(); } - UmfPackLU(const MatrixType& matrix) + explicit UmfPackLU(const MatrixType& matrix) { init(); compute(matrix); diff --git a/Eigen/src/misc/Kernel.h b/Eigen/src/misc/Kernel.h index b9e1518fd..4b03e44c1 100644 --- a/Eigen/src/misc/Kernel.h +++ b/Eigen/src/misc/Kernel.h @@ -41,7 +41,7 @@ template struct kernel_retval_base typedef ReturnByValue Base; typedef typename Base::Index Index; - kernel_retval_base(const DecompositionType& dec) + explicit kernel_retval_base(const DecompositionType& dec) : m_dec(dec), m_rank(dec.rank()), m_cols(m_rank==dec.cols() ? 1 : dec.cols() - m_rank) diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index f6d7d8944..4e64826da 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -1,5 +1,24 @@ +// These are already defined in MatrixCwiseUnaryOps.h +//typedef CwiseUnaryOp, const Derived> AbsReturnType; +//typedef CwiseUnaryOp, const Derived> Abs2ReturnType; +//typedef CwiseUnaryOp, const Derived> SqrtReturnType; +//typedef CwiseUnaryOp, const Derived> CwiseInverseReturnType; +//typedef CwiseUnaryOp >, const Derived> CwiseScalarEqualReturnType; + +typedef CwiseUnaryOp, const Derived> ExpReturnType; +typedef CwiseUnaryOp, const Derived> LogReturnType; +typedef CwiseUnaryOp, const Derived> CosReturnType; +typedef CwiseUnaryOp, const Derived> SinReturnType; +typedef CwiseUnaryOp, const Derived> AcosReturnType; +typedef CwiseUnaryOp, const Derived> AsinReturnType; +typedef CwiseUnaryOp, const Derived> TanReturnType; +typedef CwiseUnaryOp, const Derived> AtanReturnType; +typedef CwiseUnaryOp, const Derived> PowReturnType; +typedef CwiseUnaryOp, const Derived> SquareReturnType; +typedef CwiseUnaryOp, const Derived> CubeReturnType; + /** \returns an expression of the coefficient-wise absolute value of \c *this * * Example: \include Cwise_abs.cpp @@ -8,10 +27,10 @@ * \sa abs2() */ EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> +EIGEN_STRONG_INLINE const AbsReturnType abs() const { - return derived(); + return AbsReturnType(derived()); } /** \returns an expression of the coefficient-wise squared absolute value of \c *this @@ -22,10 +41,10 @@ abs() const * \sa abs(), square() */ EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> +EIGEN_STRONG_INLINE const Abs2ReturnType abs2() const { - return derived(); + return Abs2ReturnType(derived()); } /** \returns an expression of the coefficient-wise exponential of *this. @@ -39,10 +58,10 @@ abs2() const * \sa pow(), log(), sin(), cos() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const ExpReturnType exp() const { - return derived(); + return ExpReturnType(derived()); } /** \returns an expression of the coefficient-wise logarithm of *this. @@ -56,10 +75,10 @@ exp() const * \sa exp() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const LogReturnType log() const { - return derived(); + return LogReturnType(derived()); } /** \returns an expression of the coefficient-wise square root of *this. @@ -73,10 +92,10 @@ log() const * \sa pow(), square() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const SqrtReturnType sqrt() const { - return derived(); + return SqrtReturnType(derived()); } /** \returns an expression of the coefficient-wise cosine of *this. @@ -90,10 +109,10 @@ sqrt() const * \sa sin(), acos() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const CosReturnType cos() const { - return derived(); + return CosReturnType(derived()); } @@ -108,10 +127,10 @@ cos() const * \sa cos(), asin() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const SinReturnType sin() const { - return derived(); + return SinReturnType(derived()); } /** \returns an expression of the coefficient-wise arc cosine of *this. @@ -122,10 +141,10 @@ sin() const * \sa cos(), asin() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const AcosReturnType acos() const { - return derived(); + return AcosReturnType(derived()); } /** \returns an expression of the coefficient-wise arc sine of *this. @@ -136,10 +155,10 @@ acos() const * \sa sin(), acos() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const AsinReturnType asin() const { - return derived(); + return AsinReturnType(derived()); } /** \returns an expression of the coefficient-wise tan of *this. @@ -150,10 +169,10 @@ asin() const * \sa cos(), sin() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, Derived> +inline const TanReturnType tan() const { - return derived(); + return TanReturnType(derived()); } /** \returns an expression of the coefficient-wise arc tan of *this. @@ -163,10 +182,10 @@ tan() const * * \sa cos(), sin(), tan() */ -inline const CwiseUnaryOp, Derived> +inline const AtanReturnType atan() const { - return derived(); + return AtanReturnType(derived()); } /** \returns an expression of the coefficient-wise power of *this to the given exponent. @@ -180,11 +199,10 @@ atan() const * \sa exp(), log() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const PowReturnType pow(const Scalar& exponent) const { - return CwiseUnaryOp, const Derived> - (derived(), internal::scalar_pow_op(exponent)); + return PowReturnType(derived(), internal::scalar_pow_op(exponent)); } @@ -196,10 +214,10 @@ pow(const Scalar& exponent) const * \sa operator/(), operator*() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const CwiseInverseReturnType inverse() const { - return derived(); + return CwiseInverseReturnType(derived()); } /** \returns an expression of the coefficient-wise square of *this. @@ -210,10 +228,10 @@ inverse() const * \sa operator/(), operator*(), abs2() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const SquareReturnType square() const { - return derived(); + return SquareReturnType(derived()); } /** \returns an expression of the coefficient-wise cube of *this. @@ -224,10 +242,10 @@ square() const * \sa square(), pow() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> +inline const CubeReturnType cube() const { - return derived(); + return CubeReturnType(derived()); } #define EIGEN_MAKE_SCALAR_CWISE_UNARY_OP(METHOD_NAME,FUNCTOR) \ diff --git a/Eigen/src/plugins/CommonCwiseUnaryOps.h b/Eigen/src/plugins/CommonCwiseUnaryOps.h index a17153e64..050bce03c 100644 --- a/Eigen/src/plugins/CommonCwiseUnaryOps.h +++ b/Eigen/src/plugins/CommonCwiseUnaryOps.h @@ -14,6 +14,8 @@ /** \internal Represents a scalar multiple of an expression */ typedef CwiseUnaryOp, const Derived> ScalarMultipleReturnType; +typedef CwiseUnaryOp >, const Derived> ScalarComplexMultipleReturnType; + /** \internal Represents a quotient of an expression by a scalar*/ typedef CwiseUnaryOp, const Derived> ScalarQuotient1ReturnType; /** \internal the return type of conjugate() */ @@ -36,13 +38,16 @@ typedef CwiseUnaryOp, const Derived> ImagReturn /** \internal the return type of imag() */ typedef CwiseUnaryView, Derived> NonConstImagReturnType; +typedef CwiseUnaryOp, const Derived> NegativeReturnType; +//typedef CwiseUnaryOp, const Derived> + #endif // not EIGEN_PARSED_BY_DOXYGEN /** \returns an expression of the opposite of \c *this */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp::Scalar>, const Derived> -operator-() const { return derived(); } +inline const NegativeReturnType +operator-() const { return NegativeReturnType(derived()); } /** \returns an expression of \c *this scaled by the scalar factor \a scalar */ @@ -50,8 +55,7 @@ EIGEN_DEVICE_FUNC inline const ScalarMultipleReturnType operator*(const Scalar& scalar) const { - return CwiseUnaryOp, const Derived> - (derived(), internal::scalar_multiple_op(scalar)); + return ScalarMultipleReturnType(derived(), internal::scalar_multiple_op(scalar)); } #ifdef EIGEN_PARSED_BY_DOXYGEN @@ -60,20 +64,18 @@ const ScalarMultipleReturnType operator*(const RealScalar& scalar) const; /** \returns an expression of \c *this divided by the scalar value \a scalar */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp::Scalar>, const Derived> +inline const ScalarQuotient1ReturnType operator/(const Scalar& scalar) const { - return CwiseUnaryOp, const Derived> - (derived(), internal::scalar_quotient1_op(scalar)); + return ScalarQuotient1ReturnType(derived(), internal::scalar_quotient1_op(scalar)); } /** Overloaded for efficient real matrix times complex scalar value */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp >, const Derived> +inline const ScalarComplexMultipleReturnType operator*(const std::complex& scalar) const { - return CwiseUnaryOp >, const Derived> - (*static_cast(this), internal::scalar_multiple2_op >(scalar)); + return ScalarComplexMultipleReturnType(derived(), internal::scalar_multiple2_op >(scalar)); } EIGEN_DEVICE_FUNC @@ -86,6 +88,9 @@ inline friend const CwiseUnaryOp& scalar, const StorageBaseType& matrix) { return matrix*scalar; } + +template struct CastXpr { typedef typename internal::cast_return_type, const Derived> >::type Type; }; + /** \returns an expression of *this with the \a Scalar type casted to * \a NewScalar. * @@ -95,10 +100,10 @@ operator*(const std::complex& scalar, const StorageBaseType& matrix) */ template EIGEN_DEVICE_FUNC -typename internal::cast_return_type::Scalar, NewType>, const Derived> >::type +typename CastXpr::Type cast() const { - return derived(); + return typename CastXpr::Type(derived()); } /** \returns an expression of the complex conjugate of \c *this. @@ -116,14 +121,14 @@ conjugate() const * \sa imag() */ EIGEN_DEVICE_FUNC inline RealReturnType -real() const { return derived(); } +real() const { return RealReturnType(derived()); } /** \returns an read-only expression of the imaginary part of \c *this. * * \sa real() */ EIGEN_DEVICE_FUNC inline const ImagReturnType -imag() const { return derived(); } +imag() const { return ImagReturnType(derived()); } /** \brief Apply a unary operator coefficient-wise * \param[in] func Functor implementing the unary operator @@ -176,11 +181,11 @@ unaryViewExpr(const CustomViewOp& func = CustomViewOp()) const * \sa imag() */ EIGEN_DEVICE_FUNC inline NonConstRealReturnType -real() { return derived(); } +real() { return NonConstRealReturnType(derived()); } /** \returns a non const expression of the imaginary part of \c *this. * * \sa real() */ EIGEN_DEVICE_FUNC inline NonConstImagReturnType -imag() { return derived(); } +imag() { return NonConstImagReturnType(derived()); } diff --git a/Eigen/src/plugins/MatrixCwiseUnaryOps.h b/Eigen/src/plugins/MatrixCwiseUnaryOps.h index 1bb15f862..045a2ebb2 100644 --- a/Eigen/src/plugins/MatrixCwiseUnaryOps.h +++ b/Eigen/src/plugins/MatrixCwiseUnaryOps.h @@ -10,6 +10,11 @@ // This file is a base class plugin containing matrix specifics coefficient wise functions. +typedef CwiseUnaryOp, const Derived> AbsReturnType; +typedef CwiseUnaryOp, const Derived> Abs2ReturnType; +typedef CwiseUnaryOp, const Derived> SqrtReturnType; +typedef CwiseUnaryOp, const Derived> CwiseInverseReturnType; +typedef CwiseUnaryOp >, const Derived> CwiseScalarEqualReturnType; /** \returns an expression of the coefficient-wise absolute value of \c *this * * Example: \include MatrixBase_cwiseAbs.cpp @@ -18,8 +23,8 @@ * \sa cwiseAbs2() */ EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> -cwiseAbs() const { return derived(); } +EIGEN_STRONG_INLINE const AbsReturnType +cwiseAbs() const { return AbsReturnType(derived()); } /** \returns an expression of the coefficient-wise squared absolute value of \c *this * @@ -29,8 +34,8 @@ cwiseAbs() const { return derived(); } * \sa cwiseAbs() */ EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE const CwiseUnaryOp, const Derived> -cwiseAbs2() const { return derived(); } +EIGEN_STRONG_INLINE const Abs2ReturnType +cwiseAbs2() const { return Abs2ReturnType(derived()); } /** \returns an expression of the coefficient-wise square root of *this. * @@ -40,8 +45,8 @@ cwiseAbs2() const { return derived(); } * \sa cwisePow(), cwiseSquare() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> -cwiseSqrt() const { return derived(); } +inline const SqrtReturnType +cwiseSqrt() const { return SqrtReturnType(derived()); } /** \returns an expression of the coefficient-wise inverse of *this. * @@ -51,8 +56,8 @@ cwiseSqrt() const { return derived(); } * \sa cwiseProduct() */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp, const Derived> -cwiseInverse() const { return derived(); } +inline const CwiseInverseReturnType +cwiseInverse() const { return CwiseInverseReturnType(derived()); } /** \returns an expression of the coefficient-wise == operator of \c *this and a scalar \a s * @@ -64,9 +69,8 @@ cwiseInverse() const { return derived(); } * \sa cwiseEqual(const MatrixBase &) const */ EIGEN_DEVICE_FUNC -inline const CwiseUnaryOp >, const Derived> +inline const CwiseScalarEqualReturnType cwiseEqual(const Scalar& s) const { - return CwiseUnaryOp >,const Derived> - (derived(), std::bind1st(std::equal_to(), s)); + return CwiseScalarEqualReturnType(derived(), std::bind1st(std::equal_to(), s)); } From eb13ada3aa4e8514bc297e99badcc87c5f71533d Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 23 Sep 2014 17:21:27 +0200 Subject: [PATCH 0824/1103] Renamed CwiseInverseReturnType to InverseReturnType for ArrayBase::inverse() --- Eigen/src/plugins/ArrayCwiseUnaryOps.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index 4e64826da..bb1f96bdc 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -4,7 +4,7 @@ //typedef CwiseUnaryOp, const Derived> AbsReturnType; //typedef CwiseUnaryOp, const Derived> Abs2ReturnType; //typedef CwiseUnaryOp, const Derived> SqrtReturnType; -//typedef CwiseUnaryOp, const Derived> CwiseInverseReturnType; +typedef CwiseUnaryOp, const Derived> InverseReturnType; //typedef CwiseUnaryOp >, const Derived> CwiseScalarEqualReturnType; typedef CwiseUnaryOp, const Derived> ExpReturnType; @@ -214,10 +214,10 @@ pow(const Scalar& exponent) const * \sa operator/(), operator*() */ EIGEN_DEVICE_FUNC -inline const CwiseInverseReturnType +inline const InverseReturnType inverse() const { - return CwiseInverseReturnType(derived()); + return InverseReturnType(derived()); } /** \returns an expression of the coefficient-wise square of *this. From 7817bc19a40caa1d7d5d56e995cc887af58b42a2 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 23 Sep 2014 17:23:34 +0200 Subject: [PATCH 0825/1103] Removed FIXME, as it is actually necessary. --- Eigen/src/SparseCore/SparseDenseProduct.h | 1 - 1 file changed, 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index 7af6a12a9..ed3cb1990 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -174,7 +174,6 @@ protected: // if the actual left-hand side is a dense vector, // then build a sparse-view so that we can seamlessly iterate over it. - // FIXME ActualLhs does not seem to be necessary typedef typename conditional::StorageKind,Sparse>::value, Lhs1, SparseView >::type ActualLhs; typedef typename conditional::StorageKind,Sparse>::value, From 421feea3b2ca7d1b4c8af567c282519f293ce38f Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 23 Sep 2014 18:55:42 +0200 Subject: [PATCH 0826/1103] member_redux constructor is explicit too. Renamed some typedefs for more consistency. --- Eigen/src/Core/VectorwiseOp.h | 16 ++++++++-------- Eigen/src/plugins/ArrayCwiseUnaryOps.h | 8 +++----- Eigen/src/plugins/MatrixCwiseUnaryOps.h | 18 +++++++++--------- 3 files changed, 20 insertions(+), 22 deletions(-) diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index e340c1433..2815a16c9 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -128,7 +128,7 @@ struct member_redux { >::type result_type; template struct Cost { enum { value = (Size-1) * functor_traits::Cost }; }; - member_redux(const BinaryOp func) : m_functor(func) {} // FIXME this should actually be explicit, but lets not exaggerate + explicit member_redux(const BinaryOp func) : m_functor(func) {} template inline result_type operator()(const DenseBase& mat) const { return mat.redux(m_functor); } @@ -165,10 +165,10 @@ template class VectorwiseOp typedef typename internal::remove_all::type ExpressionTypeNestedCleaned; template class Functor, - typename Scalar=typename internal::traits::Scalar> struct ReturnType + typename Scalar_=Scalar> struct ReturnType { typedef PartialReduxExpr, + Functor, Direction > Type; }; @@ -176,7 +176,7 @@ template class VectorwiseOp template struct ReduxReturnType { typedef PartialReduxExpr::Scalar>, + internal::member_redux, Direction > Type; }; @@ -264,11 +264,11 @@ template class VectorwiseOp template const typename ReduxReturnType::Type redux(const BinaryOp& func = BinaryOp()) const - { return typename ReduxReturnType::Type(_expression(), func); } + { return typename ReduxReturnType::Type(_expression(), internal::member_redux(func)); } typedef typename ReturnType::Type MinCoeffReturnType; typedef typename ReturnType::Type MaxCoeffReturnType; - typedef typename ReturnType::Type SquareNormReturnType; + typedef typename ReturnType::Type SquaredNormReturnType; typedef typename ReturnType::Type NormReturnType; typedef typename ReturnType::Type BlueNormReturnType; typedef typename ReturnType::Type StableNormReturnType; @@ -313,8 +313,8 @@ template class VectorwiseOp * Output: \verbinclude PartialRedux_squaredNorm.out * * \sa DenseBase::squaredNorm() */ - const SquareNormReturnType squaredNorm() const - { return SquareNormReturnType(_expression()); } + const SquaredNormReturnType squaredNorm() const + { return SquaredNormReturnType(_expression()); } /** \returns a row (or column) vector expression of the norm * of each column (or row) of the referenced expression. diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index bb1f96bdc..f6f526d2b 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -1,11 +1,9 @@ -// These are already defined in MatrixCwiseUnaryOps.h -//typedef CwiseUnaryOp, const Derived> AbsReturnType; -//typedef CwiseUnaryOp, const Derived> Abs2ReturnType; -//typedef CwiseUnaryOp, const Derived> SqrtReturnType; +typedef CwiseUnaryOp, const Derived> AbsReturnType; +typedef CwiseUnaryOp, const Derived> Abs2ReturnType; +typedef CwiseUnaryOp, const Derived> SqrtReturnType; typedef CwiseUnaryOp, const Derived> InverseReturnType; -//typedef CwiseUnaryOp >, const Derived> CwiseScalarEqualReturnType; typedef CwiseUnaryOp, const Derived> ExpReturnType; typedef CwiseUnaryOp, const Derived> LogReturnType; diff --git a/Eigen/src/plugins/MatrixCwiseUnaryOps.h b/Eigen/src/plugins/MatrixCwiseUnaryOps.h index 045a2ebb2..c99ee94ec 100644 --- a/Eigen/src/plugins/MatrixCwiseUnaryOps.h +++ b/Eigen/src/plugins/MatrixCwiseUnaryOps.h @@ -10,9 +10,9 @@ // This file is a base class plugin containing matrix specifics coefficient wise functions. -typedef CwiseUnaryOp, const Derived> AbsReturnType; -typedef CwiseUnaryOp, const Derived> Abs2ReturnType; -typedef CwiseUnaryOp, const Derived> SqrtReturnType; +typedef CwiseUnaryOp, const Derived> CwiseAbsReturnType; +typedef CwiseUnaryOp, const Derived> CwiseAbs2ReturnType; +typedef CwiseUnaryOp, const Derived> CwiseSqrtReturnType; typedef CwiseUnaryOp, const Derived> CwiseInverseReturnType; typedef CwiseUnaryOp >, const Derived> CwiseScalarEqualReturnType; /** \returns an expression of the coefficient-wise absolute value of \c *this @@ -23,8 +23,8 @@ typedef CwiseUnaryOp >, const Derived> Cwis * \sa cwiseAbs2() */ EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE const AbsReturnType -cwiseAbs() const { return AbsReturnType(derived()); } +EIGEN_STRONG_INLINE const CwiseAbsReturnType +cwiseAbs() const { return CwiseAbsReturnType(derived()); } /** \returns an expression of the coefficient-wise squared absolute value of \c *this * @@ -34,8 +34,8 @@ cwiseAbs() const { return AbsReturnType(derived()); } * \sa cwiseAbs() */ EIGEN_DEVICE_FUNC -EIGEN_STRONG_INLINE const Abs2ReturnType -cwiseAbs2() const { return Abs2ReturnType(derived()); } +EIGEN_STRONG_INLINE const CwiseAbs2ReturnType +cwiseAbs2() const { return CwiseAbs2ReturnType(derived()); } /** \returns an expression of the coefficient-wise square root of *this. * @@ -45,8 +45,8 @@ cwiseAbs2() const { return Abs2ReturnType(derived()); } * \sa cwisePow(), cwiseSquare() */ EIGEN_DEVICE_FUNC -inline const SqrtReturnType -cwiseSqrt() const { return SqrtReturnType(derived()); } +inline const CwiseSqrtReturnType +cwiseSqrt() const { return CwiseSqrtReturnType(derived()); } /** \returns an expression of the coefficient-wise inverse of *this. * From 13cbc751c92248e0dfa969bfe5f5060773ac9972 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 23 Sep 2014 22:10:32 +0200 Subject: [PATCH 0827/1103] bug #880: automatically preserves buildtool flags when modifying DartConfiguration.tcl file. --- cmake/EigenConfigureTesting.cmake | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake index 190509958..05c8c1466 100644 --- a/cmake/EigenConfigureTesting.cmake +++ b/cmake/EigenConfigureTesting.cmake @@ -23,7 +23,12 @@ set(EIGEN_TEST_BUILD_FLAGS "" CACHE STRING "Options passed to the build command # Recall that our unit tests are not in the "all" target, so we have to explicitely ask ctest to build our custom 'buildtests' target. # At this stage, we can also add custom flags to the build tool through the user defined EIGEN_TEST_BUILD_FLAGS variable. file(READ "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" EIGEN_DART_CONFIG_FILE) -string(REGEX REPLACE "MakeCommand:.*DefaultCTestConfigurationType" "MakeCommand: ${CMAKE_COMMAND} --build . --target buildtests --config '' -- -i ${EIGEN_TEST_BUILD_FLAGS}\nDefaultCTestConfigurationType" +# try to grab the default flags +string(REGEX MATCH "MakeCommand:.*-- (.*)DefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) +if(NOT CMAKE_MATCH_1) +string(REGEX MATCH "MakeCommand:.*gmake (.*)DefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) +endif() +string(REGEX REPLACE "MakeCommand:.*DefaultCTestConfigurationType" "MakeCommand: ${CMAKE_COMMAND} --build . --target buildtests --config \"\${CTEST_CONFIGURATION_TYPE}\" -- ${CMAKE_MATCH_1} ${EIGEN_TEST_BUILD_FLAGS}\nDefaultCTestConfigurationType" EIGEN_DART_CONFIG_FILE2 ${EIGEN_DART_CONFIG_FILE}) file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" ${EIGEN_DART_CONFIG_FILE2}) From 446001ef51be920649cb4f3c07848967b6788532 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 24 Sep 2014 09:39:09 +0200 Subject: [PATCH 0828/1103] Fix nested_eval > which wrongly returned a Product<> expression --- Eigen/src/Core/Inverse.h | 3 +-- Eigen/src/Core/ProductEvaluators.h | 10 +++++++--- test/geo_orthomethods.cpp | 5 ++++- 3 files changed, 12 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index 706796b78..f3b0dff87 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -25,8 +25,7 @@ struct traits > typedef typename XprType::PlainObject PlainObject; typedef traits BaseTraits; enum { - Flags = BaseTraits::Flags & RowMajorBit, - CoeffReadCost = Dynamic + Flags = BaseTraits::Flags & RowMajorBit }; }; diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 69e569908..c944ec9fc 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -97,11 +97,12 @@ struct product_evaluator, ProductTag, LhsShape : public evaluator::PlainObject>::type { typedef Product XprType; -// enum { -// CoeffReadCost = 0 // FIXME why is it needed? (this was already the case before the evaluators, see traits) -// }; typedef typename XprType::PlainObject PlainObject; typedef typename evaluator::type Base; + enum { + Flags = Base::Flags | EvalBeforeNestingBit +// CoeffReadCost = 0 // FIXME why is it needed? (this was already the case before the evaluators, see traits) + }; explicit product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) @@ -508,6 +509,9 @@ struct product_evaluator, LazyCoeffBasedProduc typedef Product XprType; typedef Product BaseProduct; typedef product_evaluator Base; + enum { + Flags = Base::Flags | EvalBeforeNestingBit + }; explicit product_evaluator(const XprType& xpr) : Base(BaseProduct(xpr.lhs(),xpr.rhs())) {} diff --git a/test/geo_orthomethods.cpp b/test/geo_orthomethods.cpp index 7f8beb205..e178df257 100644 --- a/test/geo_orthomethods.cpp +++ b/test/geo_orthomethods.cpp @@ -39,7 +39,10 @@ template void orthomethods_3() (v0.cross(v1)).normalized(), (v0.cross(v1).cross(v0)).normalized(); VERIFY(mat3.isUnitary()); - + + mat3.setRandom(); + VERIFY_IS_APPROX(v0.cross(mat3*v1), -(mat3*v1).cross(v0)); + VERIFY_IS_APPROX(v0.cross(mat3.lazyProduct(v1)), -(mat3.lazyProduct(v1)).cross(v0)); // colwise/rowwise cross product mat3.setRandom(); From 27d6b4daf90488e105d70f45e5d5684b8f1a906d Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 24 Sep 2014 14:37:13 +0200 Subject: [PATCH 0829/1103] Tridiagonalization::diagonal() and ::subDiagonal() did not work. Added unit-test --- Eigen/src/Eigenvalues/Tridiagonalization.h | 11 ++++------- test/eigensolver_selfadjoint.cpp | 11 ++++++++++- 2 files changed, 14 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Eigenvalues/Tridiagonalization.h b/Eigen/src/Eigenvalues/Tridiagonalization.h index f20528f86..bedd1cb34 100644 --- a/Eigen/src/Eigenvalues/Tridiagonalization.h +++ b/Eigen/src/Eigenvalues/Tridiagonalization.h @@ -91,10 +91,8 @@ template class Tridiagonalization >::type DiagonalReturnType; typedef typename internal::conditional::IsComplex, - typename internal::add_const_on_value_type >::RealReturnType>::type, - const Diagonal< - Block > + typename internal::add_const_on_value_type::RealReturnType>::type, + const Diagonal >::type SubDiagonalReturnType; /** \brief Return type of matrixQ() */ @@ -307,7 +305,7 @@ typename Tridiagonalization::DiagonalReturnType Tridiagonalization::diagonal() const { eigen_assert(m_isInitialized && "Tridiagonalization is not initialized."); - return m_matrix.diagonal(); + return m_matrix.diagonal().real(); } template @@ -315,8 +313,7 @@ typename Tridiagonalization::SubDiagonalReturnType Tridiagonalization::subDiagonal() const { eigen_assert(m_isInitialized && "Tridiagonalization is not initialized."); - Index n = m_matrix.rows(); - return Block(m_matrix, 1, 0, n-1,n-1).diagonal(); + return m_matrix.template diagonal<-1>().real(); } namespace internal { diff --git a/test/eigensolver_selfadjoint.cpp b/test/eigensolver_selfadjoint.cpp index 3851f9df2..41b598795 100644 --- a/test/eigensolver_selfadjoint.cpp +++ b/test/eigensolver_selfadjoint.cpp @@ -111,8 +111,17 @@ template void selfadjointeigensolver(const MatrixType& m) // test Tridiagonalization's methods Tridiagonalization tridiag(symmC); - // FIXME tridiag.matrixQ().adjoint() does not work + VERIFY_IS_APPROX(tridiag.diagonal(), tridiag.matrixT().template diagonal()); + VERIFY_IS_APPROX(tridiag.subDiagonal(), tridiag.matrixT().template diagonal<-1>()); + MatrixType T = tridiag.matrixT(); + if(rows>1 && cols>1) { + // FIXME check that upper and lower part are 0: + //VERIFY(T.topRightCorner(rows-2, cols-2).template triangularView().isZero()); + } + VERIFY_IS_APPROX(tridiag.diagonal(), T.diagonal().real()); + VERIFY_IS_APPROX(tridiag.subDiagonal(), T.template diagonal<1>().real()); VERIFY_IS_APPROX(MatrixType(symmC.template selfadjointView()), tridiag.matrixQ() * tridiag.matrixT().eval() * MatrixType(tridiag.matrixQ()).adjoint()); + VERIFY_IS_APPROX(MatrixType(symmC.template selfadjointView()), tridiag.matrixQ() * tridiag.matrixT() * tridiag.matrixQ().adjoint()); // Test computation of eigenvalues from tridiagonal matrix if(rows > 1) From 4ba8aa14822ae0b444baa9e0d41d8744cdcd6b03 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 25 Sep 2014 16:05:17 +0200 Subject: [PATCH 0830/1103] Fix bug #884: No malloc for zero-sized matrices or for Ref without temporaries --- Eigen/src/Core/util/Memory.h | 2 ++ test/nomalloc.cpp | 35 ++++++++++++++++++++++++++++++++++- 2 files changed, 36 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 30133ba67..af46c449c 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -454,6 +454,8 @@ template inline T* conditional_aligned_realloc_new(T* pt template inline T* conditional_aligned_new_auto(size_t size) { + if(size==0) + return 0; // short-cut. Also fixes Bug 884 check_size_for_overflow(size); T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); if(NumTraits::RequireInitialization) diff --git a/test/nomalloc.cpp b/test/nomalloc.cpp index cbd02dd21..5458806f5 100644 --- a/test/nomalloc.cpp +++ b/test/nomalloc.cpp @@ -165,6 +165,38 @@ void ctms_decompositions() Eigen::JacobiSVD jSVD; jSVD.compute(A, ComputeFullU | ComputeFullV); } +void test_zerosized() { + // default constructors: + Eigen::MatrixXd A; + Eigen::VectorXd v; + // explicit zero-sized: + Eigen::ArrayXXd A0(0,0); + Eigen::ArrayXd v0(0); + + // assigning empty objects to each other: + A=A0; + v=v0; +} + +template void test_reference(const MatrixType& m) { + typedef typename MatrixType::Scalar Scalar; + enum { Flag = MatrixType::IsRowMajor ? Eigen::RowMajor : Eigen::ColMajor}; + enum { TransposeFlag = !MatrixType::IsRowMajor ? Eigen::RowMajor : Eigen::ColMajor}; + typename MatrixType::Index rows = m.rows(), cols=m.cols(); + // Dynamic reference: + typedef Eigen::Ref > Ref; + typedef Eigen::Ref > RefT; + + Ref r1(m); + Ref r2(m.block(rows/3, cols/4, rows/2, cols/2)); + RefT r3(m.transpose()); + RefT r4(m.topLeftCorner(rows/2, cols/2).transpose()); + + VERIFY_RAISES_ASSERT(RefT r5(m)); + VERIFY_RAISES_ASSERT(Ref r6(m.transpose())); + VERIFY_RAISES_ASSERT(Ref r7(Scalar(2) * m)); +} + void test_nomalloc() { // check that our operator new is indeed called: @@ -175,5 +207,6 @@ void test_nomalloc() // Check decomposition modules with dynamic matrices that have a known compile-time max size (ctms) CALL_SUBTEST_4(ctms_decompositions()); - + CALL_SUBTEST_5(test_zerosized()); + CALL_SUBTEST_6(test_reference(Matrix())); } From 6d0f0b8cecd0ee944503d885fd34278044bab746 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Thu, 25 Sep 2014 16:06:16 +0000 Subject: [PATCH 0831/1103] add VSX identifier --- cmake/EigenTesting.cmake | 2 ++ 1 file changed, 2 insertions(+) diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index 6383c6eac..65bb29413 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -424,6 +424,8 @@ macro(ei_get_cxxflags VAR) ei_is_64bit_env(IS_64BIT_ENV) if(EIGEN_TEST_NEON) set(${VAR} NEON) + elseif(EIGEN_TEST_VSX) + set(${VAR} VSX) elseif(EIGEN_TEST_ALTIVEC) set(${VAR} ALVEC) elseif(EIGEN_TEST_FMA) From ce2035af868d57b85b4078c15843d5c1f51047dc Mon Sep 17 00:00:00 2001 From: Jitse Niesen Date: Sat, 27 Sep 2014 23:25:58 +0100 Subject: [PATCH 0832/1103] New doc page on implementing a new expression class. --- doc/Manual.dox | 1 + doc/NewExpressionType.dox | 137 +++++++++++++++++++++ doc/examples/make_circulant.cpp | 11 ++ doc/examples/make_circulant.cpp.entry | 5 + doc/examples/make_circulant.cpp.evaluator | 33 +++++ doc/examples/make_circulant.cpp.expression | 20 +++ doc/examples/make_circulant.cpp.main | 8 ++ doc/examples/make_circulant.cpp.preamble | 4 + doc/examples/make_circulant.cpp.traits | 19 +++ 9 files changed, 238 insertions(+) create mode 100644 doc/NewExpressionType.dox create mode 100644 doc/examples/make_circulant.cpp create mode 100644 doc/examples/make_circulant.cpp.entry create mode 100644 doc/examples/make_circulant.cpp.evaluator create mode 100644 doc/examples/make_circulant.cpp.expression create mode 100644 doc/examples/make_circulant.cpp.main create mode 100644 doc/examples/make_circulant.cpp.preamble create mode 100644 doc/examples/make_circulant.cpp.traits diff --git a/doc/Manual.dox b/doc/Manual.dox index 43af857a5..bf1a33229 100644 --- a/doc/Manual.dox +++ b/doc/Manual.dox @@ -13,6 +13,7 @@ namespace Eigen { - \subpage TopicUsingIntelMKL - \subpage TopicCUDA - \subpage TopicTemplateKeyword + - \subpage TopicNewExpressionType - \subpage UserManual_UnderstandingEigen */ diff --git a/doc/NewExpressionType.dox b/doc/NewExpressionType.dox new file mode 100644 index 000000000..ad8b7f86b --- /dev/null +++ b/doc/NewExpressionType.dox @@ -0,0 +1,137 @@ +namespace Eigen { + +/** \page TopicNewExpressionType Adding a new expression type + +This page describes with the help of an example how to implement a new +light-weight expression type in %Eigen. This consists of three parts: +the expression type itself, a traits class containing compile-time +information about the expression, and the evaluator class which is +used to evaluate the expression to a matrix. + +\b TO \b DO: Write a page explaining the design, with details on +vectorization etc., and refer to that page here. + + +\eigenAutoToc + +\section TopicSetting The setting + +A circulant matrix is a matrix where each column is the same as the +column to the left, except that it is cyclically shifted downwards. +For example, here is a 4-by-4 circulant matrix: +\f[ \begin{bmatrix} + 1 & 8 & 4 & 2 \\ + 2 & 1 & 8 & 4 \\ + 4 & 2 & 1 & 8 \\ + 8 & 4 & 2 & 1 +\end{bmatrix} \f] +A circulant matrix is uniquely determined by its first column. We wish +to write a function \c makeCirculant which, given the first column, +returns an expression representing the circulant matrix. + +For simplicity, we restrict the \c makeCirculant function to dense +matrices. It may make sense to also allow arrays, or sparse matrices, +but we will not do so here. We also do not want to support +vectorization. + + +\section TopicPreamble Getting started + +We will present the file implementing the \c makeCirculant function +part by part. We start by including the appropriate header files and +forward declaring the expression class, which we will call +\c Circulant. The \c makeCirculant function will return an object of +this type. The class \c Circulant is in fact a class template; the +template argument \c ArgType refers to the type of the vector passed +to the \c makeCirculant function. + +\include make_circulant.cpp.preamble + + +\section TopicTraits The traits class + +For every expression class \c X, there should be a traits class +\c Traits in the \c Eigen::internal namespace containing +information about \c X known as compile time. + +As explained in \ref TopicSetting, we designed the \c Circulant +expression class to refer to dense matrices. The entries of the +circulant matrix have the same type as the entries of the vector +passed to the \c makeCirculant function. The type used to index the +entries is also the same. Again for simplicity, we will only return +column-major matrices. Finally, the circulant matrix is a square +matrix (number of rows equals number of columns), and the number of +rows equals the number of rows of the column vector passed to the +\c makeCirculant function. If this is a dynamic-size vector, then the +size of the circulant matrix is not known at compile-time. + +This leads to the following code: + +\include make_circulant.cpp.traits + + +\section TopicExpression The expression class + +The next step is to define the expression class itself. In our case, +we want to inherit from \c MatrixBase in order to expose the interface +for dense matrices. In the constructor, we check that we are passed a +column vector (see \ref TopicAssertions) and we store the vector from +which we are going to build the circulant matrix in the member +variable \c m_arg. Finally, the expression class should compute the +size of the corresponding circulant matrix. As explained above, this +is a square matrix with as many columns as the vector used to +construct the matrix. + +\b TO \b DO: What about the \c Nested typedef? It seems to be +necessary; is this only temporary? + +\include make_circulant.cpp.expression + + +\section TopicEvaluator The evaluator + +The last big fragment implements the evaluator for the \c Circulant +expression. The evaluator computes the entries of the circulant +matrix; this is done in the \c .coeff() member function. The entries +are computed by finding the corresponding entry of the vector from +which the circulant matrix is constructed. Getting this entry may +actually be non-trivial when the circulant matrix is constructed from +a vector which is given by a complicated expression, so we use the +evaluator which corresponds to the vector. + +The \c CoeffReadCost constant records the cost of computing an entry +of the circulant matrix; we ignore the index computation and say that +this is the same as the cost of computing an entry of the vector from +which the circulant matrix is constructed. + +In the constructor, we save the evaluator for the column vector which +defined the circulant matrix. We also save the size of that vector; +remember that we can query an expression object to find the size but +not the evaluator. + +\include make_circulant.cpp.evaluator + + +\section TopicEntry The entry point + +After all this, the \c makeCirculant function is very simple. It +simply creates an expression object and returns it. + +\include make_circulant.cpp.entry + + +\section TopicMain A simple main function for testing + +Finally, a short \c main function that shows how the \c makeCirculant +function can be called. + +\include make_circulant.cpp.main + +If all the fragments are combined, the following output is produced, +showing that the program works as expected: + +\verbinclude make_circulant.out + +*/ +} + diff --git a/doc/examples/make_circulant.cpp b/doc/examples/make_circulant.cpp new file mode 100644 index 000000000..92e6aaa2b --- /dev/null +++ b/doc/examples/make_circulant.cpp @@ -0,0 +1,11 @@ +/* +This program is presented in several fragments in the doc page. +Every fragment is in its own file; this file simply combines them. +*/ + +#include "make_circulant.cpp.preamble" +#include "make_circulant.cpp.traits" +#include "make_circulant.cpp.expression" +#include "make_circulant.cpp.evaluator" +#include "make_circulant.cpp.entry" +#include "make_circulant.cpp.main" diff --git a/doc/examples/make_circulant.cpp.entry b/doc/examples/make_circulant.cpp.entry new file mode 100644 index 000000000..f9d2eb8a9 --- /dev/null +++ b/doc/examples/make_circulant.cpp.entry @@ -0,0 +1,5 @@ +template +Circulant makeCirculant(const Eigen::MatrixBase& arg) +{ + return Circulant(arg.derived()); +} diff --git a/doc/examples/make_circulant.cpp.evaluator b/doc/examples/make_circulant.cpp.evaluator new file mode 100644 index 000000000..98713cdc0 --- /dev/null +++ b/doc/examples/make_circulant.cpp.evaluator @@ -0,0 +1,33 @@ +namespace Eigen { + namespace internal { + template + struct evaluator > + : evaluator_base > + { + typedef Circulant XprType; + typedef typename nested_eval::type ArgTypeNested; + typedef typename remove_all::type ArgTypeNestedCleaned; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::Index Index; + + enum { + CoeffReadCost = evaluator::CoeffReadCost, + Flags = Eigen::ColMajor + }; + + evaluator(const XprType& xpr) + : m_argImpl(xpr.m_arg), m_rows(xpr.rows()) + { } + + CoeffReturnType coeff(Index row, Index col) const + { + Index index = row - col; + if (index < 0) index += m_rows; + return m_argImpl.coeff(index); + } + + typename evaluator::nestedType m_argImpl; + const Index m_rows; + }; + } +} diff --git a/doc/examples/make_circulant.cpp.expression b/doc/examples/make_circulant.cpp.expression new file mode 100644 index 000000000..a68bcb730 --- /dev/null +++ b/doc/examples/make_circulant.cpp.expression @@ -0,0 +1,20 @@ +template +class Circulant : public Eigen::MatrixBase > +{ +public: + Circulant(const ArgType& arg) + : m_arg(arg) + { + EIGEN_STATIC_ASSERT(ArgType::ColsAtCompileTime == 1, + YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX); + } + + typedef typename Eigen::internal::ref_selector::type Nested; + + typedef typename Eigen::internal::traits::Index Index; + Index rows() const { return m_arg.rows(); } + Index cols() const { return m_arg.rows(); } + + typedef typename Eigen::internal::ref_selector::type ArgTypeNested; + ArgTypeNested m_arg; +}; diff --git a/doc/examples/make_circulant.cpp.main b/doc/examples/make_circulant.cpp.main new file mode 100644 index 000000000..877f97f62 --- /dev/null +++ b/doc/examples/make_circulant.cpp.main @@ -0,0 +1,8 @@ +int main() +{ + Eigen::VectorXd vec(4); + vec << 1, 2, 4, 8; + Eigen::MatrixXd mat; + mat = makeCirculant(vec); + std::cout << mat << std::endl; +} diff --git a/doc/examples/make_circulant.cpp.preamble b/doc/examples/make_circulant.cpp.preamble new file mode 100644 index 000000000..e575cce14 --- /dev/null +++ b/doc/examples/make_circulant.cpp.preamble @@ -0,0 +1,4 @@ +#include +#include + +template class Circulant; diff --git a/doc/examples/make_circulant.cpp.traits b/doc/examples/make_circulant.cpp.traits new file mode 100644 index 000000000..f91e43717 --- /dev/null +++ b/doc/examples/make_circulant.cpp.traits @@ -0,0 +1,19 @@ +namespace Eigen { + namespace internal { + template + struct traits > + { + typedef Eigen::Dense StorageKind; + typedef Eigen::MatrixXpr XprKind; + typedef typename ArgType::Index Index; + typedef typename ArgType::Scalar Scalar; + enum { + Flags = Eigen::ColMajor, + RowsAtCompileTime = ArgType::RowsAtCompileTime, + ColsAtCompileTime = ArgType::RowsAtCompileTime, + MaxRowsAtCompileTime = ArgType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = ArgType::MaxRowsAtCompileTime + }; + }; + } +} From 74cde0c925ae91598276c3dcb1c0e1250b770dfc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 28 Sep 2014 09:16:13 +0200 Subject: [PATCH 0833/1103] Add missing return derived() in ArrayBase::operator= --- Eigen/src/Core/ArrayBase.h | 1 + 1 file changed, 1 insertion(+) diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h index b37b49ac4..d42693d4b 100644 --- a/Eigen/src/Core/ArrayBase.h +++ b/Eigen/src/Core/ArrayBase.h @@ -121,6 +121,7 @@ template class ArrayBase Derived& operator=(const ArrayBase& other) { internal::call_assignment(derived(), other.derived()); + return derived(); } /** Set all the entries to \a value. From 75e269c77b82fdac35fa8516f256bb44df1e52cc Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Sat, 20 Sep 2014 14:57:42 +0200 Subject: [PATCH 0834/1103] Fixed warning on implicit integer conversion in test case code by using type VectorXd::Index instead of int. --- unsupported/test/NonLinearOptimization.cpp | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/unsupported/test/NonLinearOptimization.cpp b/unsupported/test/NonLinearOptimization.cpp index 75974f84f..724ea7b5b 100644 --- a/unsupported/test/NonLinearOptimization.cpp +++ b/unsupported/test/NonLinearOptimization.cpp @@ -246,9 +246,9 @@ struct hybrj_functor : Functor int operator()(const VectorXd &x, VectorXd &fvec) { double temp, temp1, temp2; - const int n = x.size(); + const VectorXd::Index n = x.size(); assert(fvec.size()==n); - for (int k = 0; k < n; k++) + for (VectorXd::Index k = 0; k < n; k++) { temp = (3. - 2.*x[k])*x[k]; temp1 = 0.; @@ -261,12 +261,12 @@ struct hybrj_functor : Functor } int df(const VectorXd &x, MatrixXd &fjac) { - const int n = x.size(); + const VectorXd::Index n = x.size(); assert(fjac.rows()==n); assert(fjac.cols()==n); - for (int k = 0; k < n; k++) + for (VectorXd::Index k = 0; k < n; k++) { - for (int j = 0; j < n; j++) + for (VectorXd::Index j = 0; j < n; j++) fjac(k,j) = 0.; fjac(k,k) = 3.- 4.*x[k]; if (k) fjac(k,k-1) = -1.; @@ -351,10 +351,10 @@ struct hybrd_functor : Functor int operator()(const VectorXd &x, VectorXd &fvec) const { double temp, temp1, temp2; - const int n = x.size(); + const VectorXd::Index n = x.size(); assert(fvec.size()==n); - for (int k=0; k < n; k++) + for (VectorXd::Index k=0; k < n; k++) { temp = (3. - 2.*x[k])*x[k]; temp1 = 0.; @@ -455,7 +455,7 @@ struct lmstr_functor : Functor assert(jac_row.size()==x.size()); double tmp1, tmp2, tmp3, tmp4; - int i = rownb-2; + VectorXd::Index i = rownb-2; tmp1 = i+1; tmp2 = 16 - i - 1; tmp3 = (i>=8)? tmp2 : tmp1; From 3bd31e21b5aac1aaf1c0e217f472846450b2d5fb Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Sat, 20 Sep 2014 15:00:36 +0200 Subject: [PATCH 0835/1103] Fixed compiler warning on implicit integer conversion by separating index type for matrix and permutation matrix which may not be equal. --- unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h b/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h index f5290dee4..db3a0ef2c 100644 --- a/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h +++ b/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h @@ -19,18 +19,19 @@ namespace Eigen { namespace internal { -template +template void lmqrsolv( Matrix &s, - const PermutationMatrix &iPerm, + const PermutationMatrix &iPerm, const Matrix &diag, const Matrix &qtb, Matrix &x, Matrix &sdiag) { + typedef typename Matrix::Index Index; /* Local variables */ - Index i, j, k, l; + Index i, j, k; Scalar temp; Index n = s.cols(); Matrix wa(n); @@ -52,7 +53,7 @@ void lmqrsolv( /* prepare the row of d to be eliminated, locating the */ /* diagonal element using p from the qr factorization. */ - l = iPerm.indices()(j); + const PermIndex l = iPerm.indices()(j); if (diag[l] == 0.) break; sdiag.tail(n-j).setZero(); From edaefeb9786d529cdf17a7d396a4c1962b23cdd3 Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Sun, 21 Sep 2014 10:01:12 +0200 Subject: [PATCH 0836/1103] Using Kernel::Index type instead of int to prevent possible implicit conversion from long to int. --- Eigen/src/Core/AssignEvaluator.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 8ab71446c..5d5095a67 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -182,7 +182,7 @@ struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel &kernel, int outer) + static EIGEN_STRONG_INLINE void run(Kernel &kernel, typename Kernel::Index outer) { kernel.assignCoeffByOuterInner(outer, Index); copy_using_evaluator_DefaultTraversal_InnerUnrolling::run(kernel, outer); @@ -192,7 +192,7 @@ struct copy_using_evaluator_DefaultTraversal_InnerUnrolling template struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel&, int) { } + static EIGEN_STRONG_INLINE void run(Kernel&, typename Kernel::Index) { } }; /*********************** @@ -249,7 +249,7 @@ struct copy_using_evaluator_innervec_CompleteUnrolling template struct copy_using_evaluator_innervec_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel &kernel, int outer) + static EIGEN_STRONG_INLINE void run(Kernel &kernel, typename Kernel::Index outer) { kernel.template assignPacketByOuterInner(outer, Index); enum { NextIndex = Index + packet_traits::size }; @@ -260,7 +260,7 @@ struct copy_using_evaluator_innervec_InnerUnrolling template struct copy_using_evaluator_innervec_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel &, int) { } + static EIGEN_STRONG_INLINE void run(Kernel &, typename Kernel::Index) { } }; /*************************************************************************** From d1ef3c35465dca670002f47ad07c1618fc2d8905 Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Sun, 21 Sep 2014 10:21:20 +0200 Subject: [PATCH 0837/1103] Changed Diagonal::index() to return an Index type instead of int to prevent possible implicit conversion from long to int. Added inline keyword to member methods. --- Eigen/src/Core/Diagonal.h | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index ba7ddbb14..6ffc0c762 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -80,7 +80,6 @@ template class Diagonal EIGEN_USING_STD_MATH(min); return m_index.value()<0 ? (min)(Index(m_matrix.cols()),Index(m_matrix.rows()+m_index.value())) : (min)(Index(m_matrix.rows()),Index(m_matrix.cols()-m_index.value())); - } EIGEN_DEVICE_FUNC @@ -148,14 +147,14 @@ template class Diagonal } EIGEN_DEVICE_FUNC - const typename internal::remove_all::type& + inline const typename internal::remove_all::type& nestedExpression() const { return m_matrix; } EIGEN_DEVICE_FUNC - int index() const + inline Index index() const { return m_index.value(); } @@ -172,7 +171,7 @@ template class Diagonal EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; } - // triger a compile time error is someone try to call packet + // trigger a compile time error is someone try to call packet template typename MatrixType::PacketReturnType packet(Index) const; template typename MatrixType::PacketReturnType packet(Index,Index) const; }; From b2755edcddfc60c508e78f1e49c6cf137894195d Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Sun, 21 Sep 2014 23:15:35 +0200 Subject: [PATCH 0838/1103] Replaced hard coded int types with Index types preventing implicit integer conversions. --- Eigen/src/Core/PermutationMatrix.h | 23 ++++++++++++----------- 1 file changed, 12 insertions(+), 11 deletions(-) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 98c83047a..b67046144 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -119,7 +119,7 @@ class PermutationBase : public EigenBase void evalTo(MatrixBase& other) const { other.setZero(); - for (int i=0; i * * \returns a reference to *this. * - * \warning This is much slower than applyTranspositionOnTheRight(int,int): + * \warning This is much slower than applyTranspositionOnTheRight(Index,Index): * this has linear complexity and requires a lot of branching. * - * \sa applyTranspositionOnTheRight(int,int) + * \sa applyTranspositionOnTheRight(Index,Index) */ Derived& applyTranspositionOnTheLeft(Index i, Index j) { @@ -186,7 +186,7 @@ class PermutationBase : public EigenBase * * This is a fast operation, it only consists in swapping two indices. * - * \sa applyTranspositionOnTheLeft(int,int) + * \sa applyTranspositionOnTheLeft(Index,Index) */ Derived& applyTranspositionOnTheRight(Index i, Index j) { @@ -216,13 +216,13 @@ class PermutationBase : public EigenBase template void assignTranspose(const PermutationBase& other) { - for (int i=0; i void assignProduct(const Lhs& lhs, const Rhs& rhs) { eigen_assert(lhs.cols() == rhs.rows()); - for (int i=0; i >& other) : m_indices(other.nestedPermutation().size()) { - for (int i=0; i PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs) @@ -596,7 +596,7 @@ struct permut_matrix_product_retval } else { - for(int i = 0; i < n; ++i) + for(Index i = 0; i < n; ++i) { Block (dst, ((Side==OnTheLeft) ^ Transposed) ? m_permutation.indices().coeff(i) : i) @@ -645,19 +645,20 @@ class Transpose > MaxColsAtCompileTime = Traits::MaxColsAtCompileTime }; typedef typename Traits::Scalar Scalar; + typedef typename Traits::Index Index; #endif Transpose(const PermutationType& p) : m_permutation(p) {} - inline int rows() const { return m_permutation.rows(); } - inline int cols() const { return m_permutation.cols(); } + inline Index rows() const { return m_permutation.rows(); } + inline Index cols() const { return m_permutation.cols(); } #ifndef EIGEN_PARSED_BY_DOXYGEN template void evalTo(MatrixBase& other) const { other.setZero(); - for (int i=0; i Date: Sun, 21 Sep 2014 23:19:29 +0200 Subject: [PATCH 0839/1103] Correcting the ReturnType in traits> to include the correct Index type. Fixed mixup of types Rhs::Index and Lhs:Index in various loop variables. Added explicit type conversion for arithmetic expressions which may return a wider type. --- .../KroneckerProduct/KroneckerTensorProduct.h | 31 ++++++++++--------- 1 file changed, 17 insertions(+), 14 deletions(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index 72e25db19..608c72021 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -48,8 +48,8 @@ class KroneckerProductBase : public ReturnByValue */ Scalar coeff(Index row, Index col) const { - return m_A.coeff(row / m_B.rows(), col / m_B.cols()) * - m_B.coeff(row % m_B.rows(), col % m_B.cols()); + return m_A.coeff(typename Lhs::Index(row / m_B.rows()), typename Lhs::Index(col / m_B.cols())) * + m_B.coeff(typename Rhs::Index(row % m_B.rows()), typename Rhs::Index(col % m_B.cols())); } /*! @@ -59,7 +59,7 @@ class KroneckerProductBase : public ReturnByValue Scalar coeff(Index i) const { EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived); - return m_A.coeff(i / m_A.size()) * m_B.coeff(i % m_A.size()); + return m_A.coeff(typename Lhs::Index(i / m_A.size())) * m_B.coeff(typename Rhs::Index(i % m_A.size())); } protected: @@ -148,10 +148,12 @@ template template void KroneckerProductSparse::evalTo(Dest& dst) const { - typedef typename Base::Index Index; - const Index Br = m_B.rows(), - Bc = m_B.cols(); - dst.resize(this->rows(), this->cols()); + typedef typename Dest::Index DestIndex; + const typename Rhs::Index Br = m_B.rows(), + Bc = m_B.cols(); + eigen_assert(this->rows() <= NumTraits::highest()); + eigen_assert(this->cols() <= NumTraits::highest()); + dst.resize(DestIndex(this->rows()), DestIndex(this->cols())); dst.resizeNonZeros(0); // 1 - evaluate the operands if needed: @@ -182,12 +184,12 @@ void KroneckerProductSparse::evalTo(Dest& dst) const // compute number of non-zeros per innervectors of dst { VectorXi nnzA = VectorXi::Zero(Dest::IsRowMajor ? m_A.rows() : m_A.cols()); - for (Index kA=0; kA < m_A.outerSize(); ++kA) + for (typename Lhs::Index kA=0; kA < m_A.outerSize(); ++kA) for (LhsInnerIterator itA(lhsEval,kA); itA; ++itA) nnzA(Dest::IsRowMajor ? itA.row() : itA.col())++; VectorXi nnzB = VectorXi::Zero(Dest::IsRowMajor ? m_B.rows() : m_B.cols()); - for (Index kB=0; kB < m_B.outerSize(); ++kB) + for (typename Rhs::Index kB=0; kB < m_B.outerSize(); ++kB) for (RhsInnerIterator itB(rhsEval,kB); itB; ++itB) nnzB(Dest::IsRowMajor ? itB.row() : itB.col())++; @@ -195,16 +197,17 @@ void KroneckerProductSparse::evalTo(Dest& dst) const dst.reserve(VectorXi::Map(nnzAB.data(), nnzAB.size())); } - for (Index kA=0; kA < m_A.outerSize(); ++kA) + for (typename Lhs::Index kA=0; kA < m_A.outerSize(); ++kA) { - for (Index kB=0; kB < m_B.outerSize(); ++kB) + for (typename Rhs::Index kB=0; kB < m_B.outerSize(); ++kB) { for (LhsInnerIterator itA(lhsEval,kA); itA; ++itA) { for (RhsInnerIterator itB(rhsEval,kB); itB; ++itB) { - const Index i = itA.row() * Br + itB.row(), - j = itA.col() * Bc + itB.col(); + const DestIndex + i = DestIndex(itA.row() * Br + itB.row()), + j = DestIndex(itA.col() * Bc + itB.col()); dst.insert(i,j) = itA.value() * itB.value(); } } @@ -259,7 +262,7 @@ struct traits > CoeffReadCost = Dynamic }; - typedef SparseMatrix ReturnType; + typedef SparseMatrix ReturnType; }; } // end namespace internal From 2c4cace56ca2a937d73700c2e8348dadb711758b Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Mon, 22 Sep 2014 15:54:34 +0200 Subject: [PATCH 0840/1103] Using Index instead of hard coded int type to prevent potential implicit integer conversion --- Eigen/src/Core/GeneralProduct.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index e05ff8dce..4eb01b1b3 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -261,7 +261,7 @@ template<> struct gemv_dense_sense_selector if(!evalToDest) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = dest.size(); + Index size = dest.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif if(!alphaIsCompatible) @@ -327,7 +327,7 @@ template<> struct gemv_dense_sense_selector if(!DirectlyUseRhs) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = actualRhs.size(); + Index size = actualRhs.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif Map(actualRhsPtr, actualRhs.size()) = actualRhs; From 821ff0ecfb4efe86e5045742682c918b01e5a263 Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Mon, 22 Sep 2014 16:12:35 +0200 Subject: [PATCH 0841/1103] Using Index instead of hard coded int type to prevent potential implicit integer conversion --- Eigen/src/Core/products/SelfadjointMatrixVector.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index d9c041f0c..372a44e47 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -218,7 +218,7 @@ struct selfadjoint_product_impl if(!EvalToDest) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = dest.size(); + Index size = dest.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif MappedDest(actualDestPtr, dest.size()) = dest; @@ -227,7 +227,7 @@ struct selfadjoint_product_impl if(!UseRhs) { #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN - int size = rhs.size(); + Index size = rhs.size(); EIGEN_DENSE_STORAGE_CTOR_PLUGIN #endif Map(actualRhsPtr, rhs.size()) = rhs; From 2946992ad4bc929d231282b477a23175cc1c0090 Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Mon, 22 Sep 2014 17:59:02 +0200 Subject: [PATCH 0842/1103] Using StorageIndexType for loop assigning initial permutation. Adding assert for index overflow. --- Eigen/src/Core/PermutationMatrix.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index b67046144..d5113ce19 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -374,7 +374,9 @@ class PermutationMatrix : public PermutationBase >& other) : m_indices(other.nestedPermutation().size()) { - for (typename IndicesType::Index i=0; i::highest()); + for (StorageIndexType i=0; i PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs) From f0a62c90bcb59bbea04cbb3cb4cc818cda9251ce Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 29 Sep 2014 10:27:51 +0200 Subject: [PATCH 0843/1103] Avoid comparisons between different index types. --- Eigen/src/Core/PermutationMatrix.h | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index d5113ce19..4846f2ae1 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -303,7 +303,9 @@ class PermutationMatrix : public PermutationBase::highest()); + } /** Copy constructor. */ template @@ -375,7 +377,8 @@ class PermutationMatrix : public PermutationBase::highest()); - for (StorageIndexType i=0; i From 9a04cd307c1cbb271373d6c4f32b5921da6144ef Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Mon, 22 Sep 2014 18:47:33 +0200 Subject: [PATCH 0844/1103] Added implicit integer conversion by using explicit integer type conversion. Adding assert to catch overflow. --- Eigen/src/SparseLU/SparseLU.h | 7 +++++-- Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h | 7 +++++-- 2 files changed, 10 insertions(+), 4 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index cd4ea5b13..1789830b9 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -675,8 +675,11 @@ struct SparseLUMatrixUReturnType : internal::no_assignment_operator template void solveInPlace(MatrixBase &X) const { - Index nrhs = X.cols(); - Index n = X.rows(); + /* Explicit type conversion as the Index type of MatrixBase may be wider than Index */ + eigen_assert(X.rows() <= NumTraits::highest()); + eigen_assert(X.cols() <= NumTraits::highest()); + Index nrhs = Index(X.cols()); + Index n = Index(X.rows()); // Backward solve with U for (Index k = m_mapL.nsuper(); k >= 0; k--) { diff --git a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h index ad6f2183f..6b0b83e0b 100644 --- a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h +++ b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h @@ -233,8 +233,11 @@ template template void MappedSuperNodalMatrix::solveInPlace( MatrixBase&X) const { - Index n = X.rows(); - Index nrhs = X.cols(); + /* Explicit type conversion as the Index type of MatrixBase may be wider than Index */ + eigen_assert(X.rows() <= NumTraits::highest()); + eigen_assert(X.cols() <= NumTraits::highest()); + Index n = Index(X.rows()); + Index nrhs = Index(X.cols()); const Scalar * Lval = valuePtr(); // Nonzero values Matrix work(n, nrhs); // working vector work.setZero(); From bc34ee3365eadd742eefe59561ad11c51adecd41 Mon Sep 17 00:00:00 2001 From: Georg Drenkhahn Date: Mon, 22 Sep 2014 18:56:36 +0200 Subject: [PATCH 0845/1103] Using Index type instead of hard coded int type to prevent potential implicit integer conversion. --- Eigen/src/SparseLU/SparseLU_kernel_bmod.h | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU_kernel_bmod.h b/Eigen/src/SparseLU/SparseLU_kernel_bmod.h index 0d0283b13..cad149ded 100644 --- a/Eigen/src/SparseLU/SparseLU_kernel_bmod.h +++ b/Eigen/src/SparseLU/SparseLU_kernel_bmod.h @@ -31,13 +31,13 @@ namespace internal { template struct LU_kernel_bmod { template - static EIGEN_DONT_INLINE void run(const int segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda, + static EIGEN_DONT_INLINE void run(const Index segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros); }; template template -EIGEN_DONT_INLINE void LU_kernel_bmod::run(const int segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda, +EIGEN_DONT_INLINE void LU_kernel_bmod::run(const Index segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros) { typedef typename ScalarVector::Scalar Scalar; @@ -45,7 +45,7 @@ EIGEN_DONT_INLINE void LU_kernel_bmod::run(const int segsi // The result of triangular solve is in tempv[*]; // The result of matric-vector update is in dense[*] Index isub = lptr + no_zeros; - int i; + Index i; Index irow; for (i = 0; i < ((SegSizeAtCompileTime==Dynamic)?segsize:SegSizeAtCompileTime); i++) { @@ -92,13 +92,13 @@ EIGEN_DONT_INLINE void LU_kernel_bmod::run(const int segsi template <> struct LU_kernel_bmod<1> { template - static EIGEN_DONT_INLINE void run(const int /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr, + static EIGEN_DONT_INLINE void run(const Index /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros); }; template -EIGEN_DONT_INLINE void LU_kernel_bmod<1>::run(const int /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr, +EIGEN_DONT_INLINE void LU_kernel_bmod<1>::run(const Index /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros) { typedef typename ScalarVector::Scalar Scalar; From 76c3cf694987382c7d92c3fb7cdf65215bfdeb5e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 29 Sep 2014 10:33:16 +0200 Subject: [PATCH 0846/1103] Re-enable -Wshorten-64-to-32 compilation flag. --- CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index b3753edb0..ea42cc8db 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -143,7 +143,7 @@ if(NOT MSVC) ei_add_cxx_compiler_flag("-Wpointer-arith") ei_add_cxx_compiler_flag("-Wwrite-strings") ei_add_cxx_compiler_flag("-Wformat-security") -# ei_add_cxx_compiler_flag("-Wshorten-64-to-32") + ei_add_cxx_compiler_flag("-Wshorten-64-to-32") ei_add_cxx_compiler_flag("-Wenum-conversion") ei_add_cxx_compiler_flag("-Wc++11-extensions") From abd3502e9ea3e659c39dd5edc17d6deabd26e048 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 29 Sep 2014 13:36:57 +0200 Subject: [PATCH 0847/1103] Introduce a generic InnerIterator classes compatible with evaluators. --- Eigen/src/Core/CoreEvaluators.h | 3 - Eigen/src/Core/CoreIterators.h | 145 ++++++++++++++++------ Eigen/src/Core/DenseBase.h | 6 +- Eigen/src/Core/util/Constants.h | 10 ++ Eigen/src/Core/util/ForwardDeclarations.h | 1 + 5 files changed, 123 insertions(+), 42 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index ab4320c2a..50caf4bab 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -17,9 +17,6 @@ namespace Eigen { namespace internal { -struct IndexBased {}; -struct IteratorBased {}; - // This class returns the evaluator kind from the expression storage kind. // Default assumes index based accessors template diff --git a/Eigen/src/Core/CoreIterators.h b/Eigen/src/Core/CoreIterators.h index 6da4683d2..7feebc4e4 100644 --- a/Eigen/src/Core/CoreIterators.h +++ b/Eigen/src/Core/CoreIterators.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2010 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -15,47 +15,116 @@ namespace Eigen { /* This file contains the respective InnerIterator definition of the expressions defined in Eigen/Core */ -/** \ingroup SparseCore_Module - * \class InnerIterator - * \brief An InnerIterator allows to loop over the element of a sparse (or dense) matrix or expression - * - * todo +namespace internal { + +template +class inner_iterator_selector; + +} + +/** \class InnerIterator + * \brief An InnerIterator allows to loop over the element of any matrix expression. + * + * \warning To be used with care because an evaluator is constructed every time an InnerIterator iterator is constructed. + * + * TODO: add a usage example */ - -// generic version for dense matrix and expressions -template class DenseBase::InnerIterator +template +class InnerIterator { - protected: - typedef typename Derived::Scalar Scalar; - typedef typename Derived::Index Index; - - enum { IsRowMajor = (Derived::Flags&RowMajorBit)==RowMajorBit }; - public: - EIGEN_STRONG_INLINE InnerIterator(const Derived& expr, Index outer) - : m_expression(expr), m_inner(0), m_outer(outer), m_end(expr.innerSize()) - {} - - EIGEN_STRONG_INLINE Scalar value() const - { - return (IsRowMajor) ? m_expression.coeff(m_outer, m_inner) - : m_expression.coeff(m_inner, m_outer); - } - - EIGEN_STRONG_INLINE InnerIterator& operator++() { m_inner++; return *this; } - - EIGEN_STRONG_INLINE Index index() const { return m_inner; } - inline Index row() const { return IsRowMajor ? m_outer : index(); } - inline Index col() const { return IsRowMajor ? index() : m_outer; } - - EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; } - - protected: - const Derived& m_expression; - Index m_inner; - const Index m_outer; - const Index m_end; +protected: + typedef internal::inner_iterator_selector::Kind> IteratorType; + typedef typename internal::evaluator::type EvaluatorType; + typedef typename internal::traits::Scalar Scalar; + typedef typename internal::traits::Index Index; +public: + /** Construct an iterator over the \a outerId -th row or column of \a xpr */ + InnerIterator(const XprType &xpr, const Index &outerId) + : m_eval(xpr), m_iter(m_eval, outerId, xpr.innerSize()) + {} + + /// \returns the value of the current coefficient. + EIGEN_STRONG_INLINE Scalar value() const { return m_iter.value(); } + /** Increment the iterator \c *this to the next non-zero coefficient. + * Explicit zeros are not skipped over. To skip explicit zeros, see class SparseView + */ + EIGEN_STRONG_INLINE InnerIterator& operator++() { m_iter.operator++(); return *this; } + /// \returns the column or row index of the current coefficient. + EIGEN_STRONG_INLINE Index index() const { return m_iter.index(); } + /// \returns the row index of the current coefficient. + EIGEN_STRONG_INLINE Index row() const { return m_iter.row(); } + /// \returns the column index of the current coefficient. + EIGEN_STRONG_INLINE Index col() const { return m_iter.col(); } + /// \returns \c true if the iterator \c *this still references a valid coefficient. + EIGEN_STRONG_INLINE operator bool() const { return m_iter; } + +protected: + EvaluatorType m_eval; + IteratorType m_iter; +private: + // If you get here, then you're not using the right InnerIterator type, e.g.: + // SparseMatrix A; + // SparseMatrix::InnerIterator it(A,0); + template InnerIterator(const EigenBase&,Index outer); }; +namespace internal { + +// Generic inner iterator implementation for dense objects +template +class inner_iterator_selector +{ +protected: + typedef typename evaluator::type EvaluatorType; + typedef typename traits::Scalar Scalar; + typedef typename traits::Index Index; + enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit }; + +public: + EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &innerSize) + : m_eval(eval), m_inner(0), m_outer(outerId), m_end(innerSize) + {} + + EIGEN_STRONG_INLINE Scalar value() const + { + return (IsRowMajor) ? m_eval.coeff(m_outer, m_inner) + : m_eval.coeff(m_inner, m_outer); + } + + EIGEN_STRONG_INLINE inner_iterator_selector& operator++() { m_inner++; return *this; } + + EIGEN_STRONG_INLINE Index index() const { return m_inner; } + inline Index row() const { return IsRowMajor ? m_outer : index(); } + inline Index col() const { return IsRowMajor ? index() : m_outer; } + + EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; } + +protected: + const EvaluatorType& m_eval; + Index m_inner; + const Index m_outer; + const Index m_end; +}; + +// For iterator-based evaluator, inner-iterator is already implemented as +// evaluator<>::InnerIterator +template +class inner_iterator_selector + : public evaluator::InnerIterator +{ +protected: + typedef typename evaluator::InnerIterator Base; + typedef typename evaluator::type EvaluatorType; + typedef typename traits::Index Index; + +public: + EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &/*innerSize*/) + : Base(eval, outerId) + {} +}; + +} // end namespace internal + } // end namespace Eigen #endif // EIGEN_COREITERATORS_H diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 101ef6f20..c4128e192 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -50,7 +50,11 @@ template class DenseBase using internal::special_scalar_op_base::Scalar, typename NumTraits::Scalar>::Real>::operator*; - class InnerIterator; + + /** Inner iterator type to iterate over the coefficients of a row or column. + * \sa class InnerIterator + */ + typedef InnerIterator InnerIterator; typedef typename internal::traits::StorageKind StorageKind; diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 2c9fb443d..5c7d70af6 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -468,6 +468,16 @@ struct SelfAdjointShape { static std::string debugName() { return "SelfAdj struct PermutationShape { static std::string debugName() { return "PermutationShape"; } }; struct SparseShape { static std::string debugName() { return "SparseShape"; } }; +namespace internal { + + // random access iterators based on coeff*() accessors. +struct IndexBased {}; + +// evaluator based on iterators to access coefficients. +struct IteratorBased {}; + +} // end namespace internal + } // end namespace Eigen #endif // EIGEN_CONSTANTS_H diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 9ec57468b..be156a44a 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -137,6 +137,7 @@ template struct CommaInitializer; template class ReturnByValue; template class ArrayWrapper; template class MatrixWrapper; +template class InnerIterator; namespace internal { template struct kernel_retval_base; From 842e31cf5c8fd31f394156ada84a1aeeab89ef7e Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 29 Sep 2014 13:37:49 +0200 Subject: [PATCH 0848/1103] Let KroneckerProduct exploits the recently introduced generic InnerIterator class. --- .../KroneckerProduct/KroneckerTensorProduct.h | 33 ++++++------------- unsupported/test/kronecker_product.cpp | 12 +++++++ 2 files changed, 22 insertions(+), 23 deletions(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index 608c72021..b459360df 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -157,40 +157,27 @@ void KroneckerProductSparse::evalTo(Dest& dst) const dst.resizeNonZeros(0); // 1 - evaluate the operands if needed: - typedef typename internal::nested_eval::type Lhs1; + typedef typename internal::nested_eval::type Lhs1; typedef typename internal::remove_all::type Lhs1Cleaned; const Lhs1 lhs1(m_A); - typedef typename internal::nested_eval::type Rhs1; + typedef typename internal::nested_eval::type Rhs1; typedef typename internal::remove_all::type Rhs1Cleaned; const Rhs1 rhs1(m_B); - - // 2 - construct a SparseView for dense operands - typedef typename internal::conditional::StorageKind,Sparse>::value, Lhs1, SparseView >::type Lhs2; - typedef typename internal::remove_all::type Lhs2Cleaned; - const Lhs2 lhs2(lhs1); - typedef typename internal::conditional::StorageKind,Sparse>::value, Rhs1, SparseView >::type Rhs2; - typedef typename internal::remove_all::type Rhs2Cleaned; - const Rhs2 rhs2(rhs1); - - // 3 - construct respective evaluators - typedef typename internal::evaluator::type LhsEval; - LhsEval lhsEval(lhs2); - typedef typename internal::evaluator::type RhsEval; - RhsEval rhsEval(rhs2); - - typedef typename LhsEval::InnerIterator LhsInnerIterator; - typedef typename RhsEval::InnerIterator RhsInnerIterator; + + // 2 - construct respective iterators + typedef InnerIterator LhsInnerIterator; + typedef InnerIterator RhsInnerIterator; // compute number of non-zeros per innervectors of dst { VectorXi nnzA = VectorXi::Zero(Dest::IsRowMajor ? m_A.rows() : m_A.cols()); for (typename Lhs::Index kA=0; kA < m_A.outerSize(); ++kA) - for (LhsInnerIterator itA(lhsEval,kA); itA; ++itA) + for (LhsInnerIterator itA(lhs1,kA); itA; ++itA) nnzA(Dest::IsRowMajor ? itA.row() : itA.col())++; VectorXi nnzB = VectorXi::Zero(Dest::IsRowMajor ? m_B.rows() : m_B.cols()); for (typename Rhs::Index kB=0; kB < m_B.outerSize(); ++kB) - for (RhsInnerIterator itB(rhsEval,kB); itB; ++itB) + for (RhsInnerIterator itB(rhs1,kB); itB; ++itB) nnzB(Dest::IsRowMajor ? itB.row() : itB.col())++; Matrix nnzAB = nnzB * nnzA.transpose(); @@ -201,9 +188,9 @@ void KroneckerProductSparse::evalTo(Dest& dst) const { for (typename Rhs::Index kB=0; kB < m_B.outerSize(); ++kB) { - for (LhsInnerIterator itA(lhsEval,kA); itA; ++itA) + for (LhsInnerIterator itA(lhs1,kA); itA; ++itA) { - for (RhsInnerIterator itB(rhsEval,kB); itB; ++itB) + for (RhsInnerIterator itB(rhs1,kB); itB; ++itB) { const DestIndex i = DestIndex(itA.row() * Br + itB.row()), diff --git a/unsupported/test/kronecker_product.cpp b/unsupported/test/kronecker_product.cpp index 753a2d417..02411a262 100644 --- a/unsupported/test/kronecker_product.cpp +++ b/unsupported/test/kronecker_product.cpp @@ -216,5 +216,17 @@ void test_kronecker_product() sC2 = kroneckerProduct(sA,sB); dC = kroneckerProduct(dA,dB); VERIFY_IS_APPROX(MatrixXf(sC2),dC); + + sC2 = kroneckerProduct(dA,sB); + dC = kroneckerProduct(dA,dB); + VERIFY_IS_APPROX(MatrixXf(sC2),dC); + + sC2 = kroneckerProduct(sA,dB); + dC = kroneckerProduct(dA,dB); + VERIFY_IS_APPROX(MatrixXf(sC2),dC); + + sC2 = kroneckerProduct(2*sA,sB); + dC = kroneckerProduct(2*dA,dB); + VERIFY_IS_APPROX(MatrixXf(sC2),dC); } } From 56a0bbbbee6b882995e065504c15fadbcfb1a6ef Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 29 Sep 2014 18:28:18 +0200 Subject: [PATCH 0849/1103] Fix compilation with GCC --- Eigen/src/Core/DenseBase.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index c4128e192..14c1902e4 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -54,7 +54,7 @@ template class DenseBase /** Inner iterator type to iterate over the coefficients of a row or column. * \sa class InnerIterator */ - typedef InnerIterator InnerIterator; + typedef Eigen::InnerIterator InnerIterator; typedef typename internal::traits::StorageKind StorageKind; From 15c946338ff8c9ada9e2daa1044121115d45a0f6 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 29 Sep 2014 19:20:01 +0200 Subject: [PATCH 0850/1103] Related to bug #880: Accept make as well a gmake when searching the MakeCommand. And don't include \n in match expression --- cmake/EigenConfigureTesting.cmake | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake index 05c8c1466..1f19e81a9 100644 --- a/cmake/EigenConfigureTesting.cmake +++ b/cmake/EigenConfigureTesting.cmake @@ -24,9 +24,9 @@ set(EIGEN_TEST_BUILD_FLAGS "" CACHE STRING "Options passed to the build command # At this stage, we can also add custom flags to the build tool through the user defined EIGEN_TEST_BUILD_FLAGS variable. file(READ "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" EIGEN_DART_CONFIG_FILE) # try to grab the default flags -string(REGEX MATCH "MakeCommand:.*-- (.*)DefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) +string(REGEX MATCH "MakeCommand:.*-- (.*)\nDefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) if(NOT CMAKE_MATCH_1) -string(REGEX MATCH "MakeCommand:.*gmake (.*)DefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) +string(REGEX MATCH "MakeCommand:.*make (.*)\nDefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) endif() string(REGEX REPLACE "MakeCommand:.*DefaultCTestConfigurationType" "MakeCommand: ${CMAKE_COMMAND} --build . --target buildtests --config \"\${CTEST_CONFIGURATION_TYPE}\" -- ${CMAKE_MATCH_1} ${EIGEN_TEST_BUILD_FLAGS}\nDefaultCTestConfigurationType" EIGEN_DART_CONFIG_FILE2 ${EIGEN_DART_CONFIG_FILE}) From e404841235cf9d995f59e9b629ebf111f9271e0b Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 29 Sep 2014 19:28:10 +0000 Subject: [PATCH 0851/1103] make sure that regex does not match cmake --- cmake/EigenConfigureTesting.cmake | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/cmake/EigenConfigureTesting.cmake b/cmake/EigenConfigureTesting.cmake index 1f19e81a9..737f77232 100644 --- a/cmake/EigenConfigureTesting.cmake +++ b/cmake/EigenConfigureTesting.cmake @@ -26,7 +26,7 @@ file(READ "${CMAKE_CURRENT_BINARY_DIR}/DartConfiguration.tcl" EIGEN_DART_CONFIG # try to grab the default flags string(REGEX MATCH "MakeCommand:.*-- (.*)\nDefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) if(NOT CMAKE_MATCH_1) -string(REGEX MATCH "MakeCommand:.*make (.*)\nDefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) +string(REGEX MATCH "MakeCommand:.*[^c]make (.*)\nDefaultCTestConfigurationType" EIGEN_DUMMY ${EIGEN_DART_CONFIG_FILE}) endif() string(REGEX REPLACE "MakeCommand:.*DefaultCTestConfigurationType" "MakeCommand: ${CMAKE_COMMAND} --build . --target buildtests --config \"\${CTEST_CONFIGURATION_TYPE}\" -- ${CMAKE_MATCH_1} ${EIGEN_TEST_BUILD_FLAGS}\nDefaultCTestConfigurationType" EIGEN_DART_CONFIG_FILE2 ${EIGEN_DART_CONFIG_FILE}) From 12d59465cb78ba058192ec4ce3039aa097bf2696 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 30 Sep 2014 14:57:54 +0200 Subject: [PATCH 0852/1103] bug #884: Copy constructor of Ref shall never malloc, constructing from other RefBase shall only malloc if the memory layout is incompatible. --- Eigen/src/Core/Ref.h | 9 +++++++++ test/nomalloc.cpp | 32 +++++++++++++++++++++++++++++--- 2 files changed, 38 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index 09921c9e7..2653f2bbe 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -232,6 +232,15 @@ template class Ref< construct(expr.derived(), typename Traits::template match::type()); } + inline Ref(const Ref& other) : Base(other) { + // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy + } + + template + inline Ref(const RefBase& other) { + construct(other.derived(), typename Traits::template match::type()); + } + protected: template diff --git a/test/nomalloc.cpp b/test/nomalloc.cpp index 5458806f5..306664210 100644 --- a/test/nomalloc.cpp +++ b/test/nomalloc.cpp @@ -21,7 +21,7 @@ // discard stack allocation as that too bypasses malloc #define EIGEN_STACK_ALLOCATION_LIMIT 0 // any heap allocation will raise an assert -#define EIGEN_NO_MALLOC +#define EIGEN_RUNTIME_NO_MALLOC #include "main.h" #include @@ -183,9 +183,11 @@ template void test_reference(const MatrixType& m) { enum { Flag = MatrixType::IsRowMajor ? Eigen::RowMajor : Eigen::ColMajor}; enum { TransposeFlag = !MatrixType::IsRowMajor ? Eigen::RowMajor : Eigen::ColMajor}; typename MatrixType::Index rows = m.rows(), cols=m.cols(); + typedef Eigen::Matrix MatrixX; + typedef Eigen::Matrix MatrixXT; // Dynamic reference: - typedef Eigen::Ref > Ref; - typedef Eigen::Ref > RefT; + typedef Eigen::Ref Ref; + typedef Eigen::Ref RefT; Ref r1(m); Ref r2(m.block(rows/3, cols/4, rows/2, cols/2)); @@ -195,10 +197,30 @@ template void test_reference(const MatrixType& m) { VERIFY_RAISES_ASSERT(RefT r5(m)); VERIFY_RAISES_ASSERT(Ref r6(m.transpose())); VERIFY_RAISES_ASSERT(Ref r7(Scalar(2) * m)); + + // Copy constructors shall also never malloc + Ref r8 = r1; + RefT r9 = r3; + + // Initializing from a compatible Ref shall also never malloc + Eigen::Ref > r10=r8, r11=m; + + // Initializing from an incompatible Ref will malloc: + typedef Eigen::Ref RefAligned; + VERIFY_RAISES_ASSERT(RefAligned r12=r10); + VERIFY_RAISES_ASSERT(Ref r13=r10); // r10 has more dynamic strides + } void test_nomalloc() { + // create some dynamic objects + Eigen::MatrixXd M1 = MatrixXd::Random(3,3); + Ref R1 = 2.0*M1; // Ref requires temporary + + // from here on prohibit malloc: + Eigen::internal::set_is_malloc_allowed(false); + // check that our operator new is indeed called: VERIFY_RAISES_ASSERT(MatrixXd dummy(MatrixXd::Random(3,3))); CALL_SUBTEST_1(nomalloc(Matrix()) ); @@ -207,6 +229,10 @@ void test_nomalloc() // Check decomposition modules with dynamic matrices that have a known compile-time max size (ctms) CALL_SUBTEST_4(ctms_decompositions()); + CALL_SUBTEST_5(test_zerosized()); + CALL_SUBTEST_6(test_reference(Matrix())); + CALL_SUBTEST_7(test_reference(R1)); + CALL_SUBTEST_8(Ref R2 = M1.topRows<2>(); test_reference(R2)); } From 81517eebc1c371cedc057130af7513ddd72de35a Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 30 Sep 2014 16:42:04 +0200 Subject: [PATCH 0853/1103] Missing explicit --- Eigen/src/Core/CoreEvaluators.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 50caf4bab..2980d7a40 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -231,7 +231,7 @@ struct evaluator > evaluator() {} - evaluator(const XprType& m) + explicit evaluator(const XprType& m) : evaluator >(m) { } }; From 6d26deb894fa62a429b717efd4d4b3aa31aba88a Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 30 Sep 2014 16:43:19 +0200 Subject: [PATCH 0854/1103] Missing outerStride in AlignedVector3 resulted in infinite recursion --- unsupported/Eigen/AlignedVector3 | 1 + 1 file changed, 1 insertion(+) diff --git a/unsupported/Eigen/AlignedVector3 b/unsupported/Eigen/AlignedVector3 index 35493e87b..1fce00525 100644 --- a/unsupported/Eigen/AlignedVector3 +++ b/unsupported/Eigen/AlignedVector3 @@ -61,6 +61,7 @@ template class AlignedVector3 Scalar* data() { return m_coeffs.data(); } const Scalar* data() const { return m_coeffs.data(); } Index innerStride() const { return 1; } + Index outerStride() const { return m_coeffs.outerStride(); } inline const Scalar& coeff(Index row, Index col) const { return m_coeffs.coeff(row, col); } From 01875049125a8da1da73b3ea0ad9319d6d80f4c2 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 30 Sep 2014 16:43:52 +0200 Subject: [PATCH 0855/1103] Avoid `unneeded-internal-declaration' warning --- test/linearstructure.cpp | 2 ++ 1 file changed, 2 insertions(+) diff --git a/test/linearstructure.cpp b/test/linearstructure.cpp index 87dfa1b6b..8e3cc9a86 100644 --- a/test/linearstructure.cpp +++ b/test/linearstructure.cpp @@ -92,6 +92,8 @@ template void real_complex(DenseIndex rows = MatrixType::Ro void test_linearstructure() { + g_called = true; + VERIFY(g_called); // avoid `unneeded-internal-declaration` warning. for(int i = 0; i < g_repeat; i++) { CALL_SUBTEST_1( linearStructure(Matrix()) ); CALL_SUBTEST_2( linearStructure(Matrix2f()) ); From 5180bb5e470c8b9cf9395270fca0f068c001ba44 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 30 Sep 2014 16:59:28 +0200 Subject: [PATCH 0856/1103] Add missing default ctor in Rotation2D --- Eigen/src/Geometry/Rotation2D.h | 7 +++++-- test/geo_transformations.cpp | 9 +++++++++ 2 files changed, 14 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Geometry/Rotation2D.h b/Eigen/src/Geometry/Rotation2D.h index c82762acd..776c36144 100644 --- a/Eigen/src/Geometry/Rotation2D.h +++ b/Eigen/src/Geometry/Rotation2D.h @@ -60,6 +60,9 @@ public: /** Construct a 2D counter clock wise rotation from the angle \a a in radian. */ explicit inline Rotation2D(const Scalar& a) : m_angle(a) {} + + /** Default constructor wihtout initialization. The represented rotation is undefined. */ + Rotation2D() {} /** \returns the rotation angle */ inline Scalar angle() const { return m_angle; } @@ -81,10 +84,10 @@ public: /** Applies the rotation to a 2D vector */ Vector2 operator* (const Vector2& vec) const { return toRotationMatrix() * vec; } - + template Rotation2D& fromRotationMatrix(const MatrixBase& m); - Matrix2 toRotationMatrix(void) const; + Matrix2 toRotationMatrix() const; /** \returns the spherical interpolation between \c *this and \a other using * parameter \a t. It is in fact equivalent to a linear interpolation. diff --git a/test/geo_transformations.cpp b/test/geo_transformations.cpp index 7d9080333..1768d7b8a 100644 --- a/test/geo_transformations.cpp +++ b/test/geo_transformations.cpp @@ -397,6 +397,15 @@ template void transformations() t20 = Translation2(v20) * (Rotation2D(s0) * Eigen::Scaling(s0)); t21 = Translation2(v20) * Rotation2D(s0) * Eigen::Scaling(s0); VERIFY_IS_APPROX(t20,t21); + + // check basic features + { + Rotation2D r1; // default ctor + r1 = Rotation2D(s0); // copy assignment + VERIFY_IS_APPROX(r1.angle(),s0); + Rotation2D r2(r1); // copy ctor + VERIFY_IS_APPROX(r2.angle(),s0); + } } template void transform_alignment() From 9d3c69952bb99180c42b24911c1b10cad9e6fa00 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Wed, 1 Oct 2014 09:43:56 +0000 Subject: [PATCH 0857/1103] fixed to make big-endian VSX work as well --- Eigen/src/Core/arch/AltiVec/Complex.h | 2 ++ Eigen/src/Core/arch/AltiVec/PacketMath.h | 26 ++++++++++++++++++------ 2 files changed, 22 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index 30e2088ef..f9b93a42b 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -16,6 +16,8 @@ namespace internal { static Packet4ui p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_ZERO_);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; #ifdef _BIG_ENDIAN +static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; +static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 }; #else static Packet2ul p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO, (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 }; static Packet2ul p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 }; diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 1b86e1227..fa02f57a1 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -78,18 +78,24 @@ static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 }; static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 }; static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 }; +// Mask alignment +#ifdef __PPC64__ +#define _EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0 +#else +#define _EIGEN_MASK_ALIGNMENT 0xfffffff0 +#endif + +#define _EIGEN_ALIGNED_PTR(x) ((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT) + // Handle endianness properly while loading constants // Define global static constants: #ifdef _BIG_ENDIAN static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); +static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; static Packet16uc p16uc_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 }; static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16}; #else -// Mask alignment -#define _EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0 -#define _EIGEN_ALIGNED_PTR(x) ((ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT) - static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 }; static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 }; @@ -716,7 +722,12 @@ static Packet2l p2l_ZERO = (Packet2l) p4i_ZERO; static Packet2d p2d_ONE = { 1.0, 1.0 }; static Packet2d p2d_ZERO = (Packet2d) p4f_ZERO; static Packet2d p2d_ZERO_ = { -0.0, -0.0 }; + +#ifdef _BIG_ENDIAN +static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ZERO, (Packet16uc) p2d_ONE, 8); +#else static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ONE, (Packet16uc) p2d_ZERO, 8); +#endif static EIGEN_STRONG_INLINE Packet2d vec_splat_dbl(Packet2d& a, int index) { @@ -862,11 +873,14 @@ template<> EIGEN_STRONG_INLINE double predux(const Packet2d& a) template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) { Packet2d v[2], sum; - v[0] = vec_add(vecs[0], (Packet2d) vec_sld((Packet4ui) vecs[0], (Packet4ui) vecs[0], 8)); v[1] = vec_add(vecs[1], (Packet2d) vec_sld((Packet4ui) vecs[1], (Packet4ui) vecs[1], 8)); - + +#ifdef _BIG_ENDIAN + sum = (Packet2d) vec_sld((Packet4ui) v[0], (Packet4ui) v[1], 8); +#else sum = (Packet2d) vec_sld((Packet4ui) v[1], (Packet4ui) v[0], 8); +#endif return sum; } From 1fa6fe2abde9753807dc3f85bbf5e6eac3f31313 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 1 Oct 2014 14:33:55 +0200 Subject: [PATCH 0858/1103] template keyword not allowed before non-template function call --- test/eigensolver_selfadjoint.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/eigensolver_selfadjoint.cpp b/test/eigensolver_selfadjoint.cpp index 41b598795..8ff48462e 100644 --- a/test/eigensolver_selfadjoint.cpp +++ b/test/eigensolver_selfadjoint.cpp @@ -111,7 +111,7 @@ template void selfadjointeigensolver(const MatrixType& m) // test Tridiagonalization's methods Tridiagonalization tridiag(symmC); - VERIFY_IS_APPROX(tridiag.diagonal(), tridiag.matrixT().template diagonal()); + VERIFY_IS_APPROX(tridiag.diagonal(), tridiag.matrixT().diagonal()); VERIFY_IS_APPROX(tridiag.subDiagonal(), tridiag.matrixT().template diagonal<-1>()); MatrixType T = tridiag.matrixT(); if(rows>1 && cols>1) { From 1c236f4c9ae78cc58156eebe3b2bb43588897af4 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 1 Oct 2014 20:21:42 -0700 Subject: [PATCH 0859/1103] Added tests for tensors of const values and tensors of stringswwq:: --- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 2 +- unsupported/test/CMakeLists.txt | 2 + .../test/cxx11_tensor_of_const_values.cpp | 105 +++++++++++++ unsupported/test/cxx11_tensor_of_strings.cpp | 142 ++++++++++++++++++ 4 files changed, 250 insertions(+), 1 deletion(-) create mode 100644 unsupported/test/cxx11_tensor_of_const_values.cpp create mode 100644 unsupported/test/cxx11_tensor_of_strings.cpp diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 28ae7b3c6..13109f514 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -301,7 +301,7 @@ struct TensorEvaluator, Devi EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { m_impl.evalSubExprsIfNeeded(NULL); - if (data && m_impl.data()) { + if (internal::is_arithmetic::value && data && m_impl.data()) { Index contiguous_values = 1; for (int i = 0; i < NumDims; ++i) { contiguous_values *= dimensions()[i]; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 615ff3e6d..8d4e7db66 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -106,6 +106,8 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_convolution "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") + ei_add_test(cxx11_tensor_of_const_values "-std=c++0x") + ei_add_test(cxx11_tensor_of_strings "-std=c++0x") ei_add_test(cxx11_tensor_intdiv "-std=c++0x") ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_of_const_values.cpp b/unsupported/test/cxx11_tensor_of_const_values.cpp new file mode 100644 index 000000000..f179a0c21 --- /dev/null +++ b/unsupported/test/cxx11_tensor_of_const_values.cpp @@ -0,0 +1,105 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +static void test_assign() +{ + float data1[6]; + TensorMap> mat1(data1, 2, 3); + float data2[6]; + const TensorMap> mat2(data2, 2, 3); + + for (int i = 0; i < 6; ++i) { + data1[i] = i; + data2[i] = -i; + } + + Tensor rslt1; + rslt1 = mat1; + Tensor rslt2; + rslt2 = mat2; + + Tensor rslt3 = mat1; + Tensor rslt4 = mat2; + + Tensor rslt5(mat1); + Tensor rslt6(mat2); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_APPROX(rslt1(i,j), static_cast(i + 2*j)); + VERIFY_IS_APPROX(rslt2(i,j), static_cast(-i - 2*j)); + VERIFY_IS_APPROX(rslt3(i,j), static_cast(i + 2*j)); + VERIFY_IS_APPROX(rslt4(i,j), static_cast(-i - 2*j)); + VERIFY_IS_APPROX(rslt5(i,j), static_cast(i + 2*j)); + VERIFY_IS_APPROX(rslt6(i,j), static_cast(-i - 2*j)); + } + } +} + + +static void test_plus() +{ + float data1[6]; + TensorMap> mat1(data1, 2, 3); + float data2[6]; + TensorMap> mat2(data2, 2, 3); + + for (int i = 0; i < 6; ++i) { + data1[i] = i; + data2[i] = -i; + } + + Tensor sum1; + sum1 = mat1 + mat2; + Tensor sum2; + sum2 = mat2 + mat1; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_APPROX(sum1(i,j), 0.0f); + VERIFY_IS_APPROX(sum2(i,j), 0.0f); + } + } +} + + +static void test_plus_equal() +{ + float data1[6]; + TensorMap> mat1(data1, 2, 3); + float data2[6]; + TensorMap> mat2(data2, 2, 3); + + for (int i = 0; i < 6; ++i) { + data1[i] = i; + data2[i] = -i; + } + mat2 += mat1; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_APPROX(mat2(i,j), 0.0f); + } + } +} + + +void test_cxx11_tensor_of_const_values() +{ + CALL_SUBTEST(test_assign()); + CALL_SUBTEST(test_plus()); + CALL_SUBTEST(test_plus_equal()); +} diff --git a/unsupported/test/cxx11_tensor_of_strings.cpp b/unsupported/test/cxx11_tensor_of_strings.cpp new file mode 100644 index 000000000..0ffa341c4 --- /dev/null +++ b/unsupported/test/cxx11_tensor_of_strings.cpp @@ -0,0 +1,142 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" +#include +#include + +using std::string; +using Eigen::Tensor; +using Eigen::TensorMap; + +static void test_assign() +{ + string data1[6]; + TensorMap> mat1(data1, 2, 3); + string data2[6]; + const TensorMap> mat2(data2, 2, 3); + + for (int i = 0; i < 6; ++i) { + std::ostringstream s1; + s1 << "abc" << i*3; + data1[i] = s1.str(); + std::ostringstream s2; + s2 << "def" << i*5; + data2[i] = s2.str(); + } + + Tensor rslt1; + rslt1 = mat1; + Tensor rslt2; + rslt2 = mat2; + + Tensor rslt3 = mat1; + Tensor rslt4 = mat2; + + Tensor rslt5(mat1); + Tensor rslt6(mat2); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(rslt1(i,j), data1[i+2*j]); + VERIFY_IS_EQUAL(rslt2(i,j), data2[i+2*j]); + VERIFY_IS_EQUAL(rslt3(i,j), data1[i+2*j]); + VERIFY_IS_EQUAL(rslt4(i,j), data2[i+2*j]); + VERIFY_IS_EQUAL(rslt5(i,j), data1[i+2*j]); + VERIFY_IS_EQUAL(rslt6(i,j), data2[i+2*j]); + } + } +} + + +static void test_concat() +{ + Tensor t1(2, 3); + Tensor t2(2, 3); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + std::ostringstream s1; + s1 << "abc" << i + j*2; + t1(i, j) = s1.str(); + std::ostringstream s2; + s2 << "def" << i*5 + j*32; + t2(i, j) = s2.str(); + } + } + + Tensor result = t1.concatenate(t2, 1); + VERIFY_IS_EQUAL(result.dimension(0), 2); + VERIFY_IS_EQUAL(result.dimension(1), 6); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(result(i, j), t1(i, j)); + VERIFY_IS_EQUAL(result(i, j+3), t2(i, j)); + } + } +} + + +static void test_slices() +{ + Tensor data(2, 6); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + std::ostringstream s1; + s1 << "abc" << i + j*2; + data(i, j) = s1.str(); + } + } + + const Eigen::DSizes half_size{{2, 3}}; + const Eigen::DSizes first_half{{0, 0}}; + const Eigen::DSizes second_half{{0, 3}}; + + Tensor t1 = data.slice(first_half, half_size); + Tensor t2 = data.slice(second_half, half_size); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(data(i, j), t1(i, j)); + VERIFY_IS_EQUAL(data(i, j+3), t2(i, j)); + } + } +} + + +static void test_additions() +{ + Tensor data1(3); + Tensor data2(3); + for (int i = 0; i < 3; ++i) { + data1(i) = "abc"; + std::ostringstream s1; + s1 << i; + data2(i) = s1.str(); + } + + Tensor sum = data1 + data2; + for (int i = 0; i < 3; ++i) { + std::ostringstream concat; + concat << "abc" << i; + string expected = concat.str(); + VERIFY_IS_EQUAL(sum(i), expected); + } +} + + +void test_cxx11_tensor_of_strings() +{ + // Beware: none of this is likely to ever work on a GPU. + CALL_SUBTEST(test_assign()); + CALL_SUBTEST(test_concat()); + CALL_SUBTEST(test_slices()); + CALL_SUBTEST(test_additions()); +} From 7caaf6453b7b1f58d953729380d596b2d9b27835 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 1 Oct 2014 20:38:22 -0700 Subject: [PATCH 0860/1103] Added support for tensor reductions and concatenations --- unsupported/Eigen/CXX11/Tensor | 3 + .../Eigen/CXX11/src/Tensor/TensorBase.h | 28 +++ .../CXX11/src/Tensor/TensorConcatenation.h | 217 +++++++++++++++++ .../src/Tensor/TensorForwardDeclarations.h | 3 +- .../Eigen/CXX11/src/Tensor/TensorFunctors.h | 62 +++++ .../Eigen/CXX11/src/Tensor/TensorReduction.h | 226 ++++++++++++++++++ unsupported/test/CMakeLists.txt | 4 +- .../test/cxx11_tensor_concatenation.cpp | 110 +++++++++ unsupported/test/cxx11_tensor_reduction.cpp | 147 ++++++++++++ 9 files changed, 798 insertions(+), 2 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h create mode 100644 unsupported/test/cxx11_tensor_concatenation.cpp create mode 100644 unsupported/test/cxx11_tensor_reduction.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index ebe6419e8..11161a547 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -34,12 +34,15 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBase.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 2da8f8cc8..2f7c9ecda 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -204,12 +204,40 @@ class TensorBase return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); } + // Reductions. + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReductionOp, const Dims, const Derived> + sum(const Dims& dims) const { + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::SumReducer()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReductionOp, const Dims, const Derived> + maximum(const Dims& dims) const { + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::MaxReducer()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReductionOp, const Dims, const Derived> + minimum(const Dims& dims) const { + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::MinReducer()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReductionOp + reduce(const Dims& dims, const Reducer& reducer) const { + return TensorReductionOp(derived(), dims, reducer); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorBroadcastingOp broadcast(const Broadcast& broadcast) const { return TensorBroadcastingOp(derived(), broadcast); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorConcatenationOp + concatenate(const OtherDerived& other, Axis axis) const { + return TensorConcatenationOp(derived(), other.derived(), axis); + } + // Morphing operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorReshapingOp diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h new file mode 100644 index 000000000..b8e43f484 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h @@ -0,0 +1,217 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H +#define EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H + +namespace Eigen { + +/** \class TensorConcatenationOp + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor concatenation class. + * + * + */ +namespace internal { +template +struct traits > +{ + // Type promotion to handle the case where the types of the lhs and the rhs are different. + typedef typename promote_storage_type::ret Scalar; + typedef typename packet_traits::type Packet; + typedef typename promote_storage_type::StorageKind, + typename traits::StorageKind>::ret StorageKind; + typedef typename promote_index_type::Index, + typename traits::Index>::type Index; + typedef typename LhsXprType::Nested LhsNested; + typedef typename RhsXprType::Nested RhsNested; + typedef typename remove_reference::type _LhsNested; + typedef typename remove_reference::type _RhsNested; + enum { Flags = 0 }; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorConcatenationOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorConcatenationOp type; +}; + +} // end namespace internal + + +template +class TensorConcatenationOp : public TensorBase, WriteAccessors> +{ + public: + typedef typename internal::traits::Scalar Scalar; + typedef typename internal::traits::Packet Packet; + typedef typename internal::traits::StorageKind StorageKind; + typedef typename internal::traits::Index Index; + typedef typename internal::nested::type Nested; + typedef typename internal::promote_storage_type::ret CoeffReturnType; + typedef typename internal::promote_storage_type::ret PacketReturnType; + typedef typename NumTraits::Real RealScalar; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorConcatenationOp(const LhsXprType& lhs, const RhsXprType& rhs, Axis axis) + : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_axis(axis) {} + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + lhsExpression() const { return m_lhs_xpr; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + rhsExpression() const { return m_rhs_xpr; } + + EIGEN_DEVICE_FUNC Axis axis() const { return m_axis; } + + protected: + typename LhsXprType::Nested m_lhs_xpr; + typename RhsXprType::Nested m_rhs_xpr; + const Axis m_axis; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorConcatenationOp XprType; + typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::Dimensions>::value; + static const int RightNumDims = internal::array_size::Dimensions>::value; + typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + enum { + IsAligned = false, + PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device), m_axis(op.axis()) + { + EIGEN_STATIC_ASSERT(NumDims == RightNumDims, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(0 <= m_axis && m_axis < NumDims); + const Dimensions& lhs_dims = m_leftImpl.dimensions(); + const Dimensions& rhs_dims = m_rightImpl.dimensions(); + int i = 0; + for (; i < m_axis; ++i) { + eigen_assert(lhs_dims[i] > 0); + eigen_assert(lhs_dims[i] == rhs_dims[i]); + m_dimensions[i] = lhs_dims[i]; + } + eigen_assert(lhs_dims[i] > 0); // Now i == m_axis. + eigen_assert(rhs_dims[i] > 0); + m_dimensions[i] = lhs_dims[i] + rhs_dims[i]; + for (++i; i < NumDims; ++i) { + eigen_assert(lhs_dims[i] > 0); + eigen_assert(lhs_dims[i] == rhs_dims[i]); + m_dimensions[i] = lhs_dims[i]; + } + + m_leftStrides[0] = 1; + m_rightStrides[0] = 1; + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_leftStrides[i] = m_leftStrides[i-1] * lhs_dims[i-1]; + m_rightStrides[i] = m_rightStrides[i-1] * rhs_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + // TODO(phli): Add short-circuit memcpy evaluation if underlying data are linear? + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) + { + m_leftImpl.evalSubExprsIfNeeded(NULL); + m_rightImpl.evalSubExprsIfNeeded(NULL); + return true; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() + { + m_leftImpl.cleanup(); + m_rightImpl.cleanup(); + } + + // TODO(phli): attempt to speed this up. The integer divisions and modulo are slow. + // See CL/76180724 comments for more ideas. + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + // Collect dimension-wise indices (subs). + array subs; + for (int i = NumDims - 1; i > 0; --i) { + subs[i] = index / m_outputStrides[i]; + index -= subs[i] * m_outputStrides[i]; + } + subs[0] = index; + + const Dimensions& left_dims = m_leftImpl.dimensions(); + if (subs[m_axis] < left_dims[m_axis]) { + Index left_index = subs[0]; + for (int i = 1; i < NumDims; ++i) { + left_index += (subs[i] % left_dims[i]) * m_leftStrides[i]; + } + return m_leftImpl.coeff(left_index); + } else { + subs[m_axis] -= left_dims[m_axis]; + const Dimensions& right_dims = m_rightImpl.dimensions(); + Index right_index = subs[0]; + for (int i = 1; i < NumDims; ++i) { + right_index += (subs[i] % right_dims[i]) * m_rightStrides[i]; + } + return m_rightImpl.coeff(right_index); + } + } + + // TODO(phli): Add a real vectorization. + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + static const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index + packetSize - 1 < dimensions().TotalSize()); + + EIGEN_ALIGN_DEFAULT CoeffReturnType values[packetSize]; + for (int i = 0; i < packetSize; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + + Scalar* data() const { return NULL; } + + protected: + const Axis m_axis; + Dimensions m_dimensions; + array m_outputStrides; + array m_leftStrides; + array m_rightStrides; + TensorEvaluator m_leftImpl; + TensorEvaluator m_rightImpl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_CONCATENATION_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index afbcc9486..bc67586a4 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -21,8 +21,9 @@ template class TensorCwiseNullaryO template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; template class TensorSelectOp; -template class TensorReductionOp; template class TensorBroadcastingOp; +template class TensorReductionOp; +template class TensorConcatenationOp; template class TensorContractionOp; template class TensorConvolutionOp; template class TensorReshapingOp; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h new file mode 100644 index 000000000..92984336c --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h @@ -0,0 +1,62 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_FUNCTORS_H +#define EIGEN_CXX11_TENSOR_TENSOR_FUNCTORS_H + +namespace Eigen { +namespace internal { + +// Standard reduction functors +template struct SumReducer +{ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE SumReducer() : m_sum(0) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t) { + m_sum += t; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize() const { + return m_sum; + } + + private: + T m_sum; +}; + +template struct MaxReducer +{ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE MaxReducer() : m_max((std::numeric_limits::min)()) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t) { + if (t > m_max) { m_max = t; } + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize() const { + return m_max; + } + + private: + T m_max; +}; + +template struct MinReducer +{ + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE MinReducer() : m_min((std::numeric_limits::max)()) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t) { + if (t < m_min) { m_min = t; } + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize() const { + return m_min; + } + + private: + T m_min; +}; + +} // end namespace internal +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_FUNCTORS_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h new file mode 100644 index 000000000..eef992106 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -0,0 +1,226 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H +#define EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H + +namespace Eigen { + +/** \class TensorReduction + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor reduction class. + * + */ + +namespace internal { +template +struct traits > + : traits +{ + typedef typename traits::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorReductionOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorReductionOp type; +}; + +} // end namespace internal + + +template +class TensorReductionOp : public TensorBase, ReadOnlyAccessors> { + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorReductionOp(const XprType& expr, const Dims& dims) : m_expr(expr), m_dims(dims) + { } + TensorReductionOp(const XprType& expr, const Dims& dims, const Op& reducer) : m_expr(expr), m_dims(dims), m_reducer(reducer) + { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const XprType& expression() const { return m_expr; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const Dims& dims() const { return m_dims; } + const Op& reducer() const { return m_reducer; } + + protected: + typename XprType::Nested m_expr; + const Dims m_dims; + const Op m_reducer; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorReductionOp XprType; + typedef typename XprType::Index Index; + static const int NumInputDims = internal::array_size::Dimensions>::value; + static const int NumReducedDims = internal::array_size::value; + static const int NumDims = (NumInputDims==NumReducedDims) ? 1 : NumInputDims - NumReducedDims; + typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; + + enum { + IsAligned = false, + PacketAccess = false, // The code isn't vectorized properly yet + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_reducer(op.reducer()) + { + EIGEN_STATIC_ASSERT(NumInputDims >= NumReducedDims, YOU_MADE_A_PROGRAMMING_MISTAKE); + + array reduced; + for (int i = 0; i < NumInputDims; ++i) { + reduced[i] = false; + } + for (int i = 0; i < NumReducedDims; ++i) { + eigen_assert(op.dims()[i] >= 0); + eigen_assert(op.dims()[i] < NumInputDims); + reduced[op.dims()[i]] = true; + } + + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + int outputIndex = 0; + int reduceIndex = 0; + for (int i = 0; i < NumInputDims; ++i) { + if (reduced[i]) { + m_reducedDims[reduceIndex] = input_dims[i]; + ++reduceIndex; + } else { + m_dimensions[outputIndex] = input_dims[i]; + ++outputIndex; + } + } + + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } + + array strides; + strides[0] = 1; + for (int i = 1; i < NumInputDims; ++i) { + strides[i] = strides[i-1] * input_dims[i-1]; + } + outputIndex = 0; + reduceIndex = 0; + for (int i = 0; i < NumInputDims; ++i) { + if (reduced[i]) { + m_reducedStrides[reduceIndex] = strides[i]; + ++reduceIndex; + } else { + m_preservedStrides[outputIndex] = strides[i]; + ++outputIndex; + } + } + + // Special case for full reductions + if (NumInputDims == NumReducedDims) { + m_dimensions[0] = 1; + } + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + Op reducer(m_reducer); + reduce(firstInput(index), 0, reducer); + return reducer.finalize(); + } + + // TODO(bsteiner): provide a more efficient implementation. + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index + packetSize - 1 < dimensions().TotalSize()); + + EIGEN_ALIGN_DEFAULT CoeffReturnType values[packetSize]; + for (int i = 0; i < packetSize; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + + Scalar* data() const { return NULL; } + + private: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const { + Index startInput = 0; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + startInput += idx * m_preservedStrides[i]; + index -= idx * m_outputStrides[i]; + } + startInput += index * m_preservedStrides[0]; + return startInput; + } + + EIGEN_DEVICE_FUNC void reduce(Index firstIndex, int DimIndex, Op& reducer) const { + for (int j = 0; j < m_reducedDims[DimIndex]; ++j) { + const Index input = firstIndex + j * m_reducedStrides[DimIndex]; + if (DimIndex < NumReducedDims-1) { + reduce(input, DimIndex+1, reducer); + } else { + reducer.reduce(m_impl.coeff(input)); + } + } + } + + Dimensions m_dimensions; + array m_outputStrides; + array m_preservedStrides; + array m_reducedStrides; + array m_reducedDims; + Op m_reducer; + TensorEvaluator m_impl; +}; + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_REDUCTION_H diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 8d4e7db66..e83d8b54e 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -106,14 +106,16 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_convolution "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") - ei_add_test(cxx11_tensor_of_const_values "-std=c++0x") +# ei_add_test(cxx11_tensor_of_const_values "-std=c++0x") ei_add_test(cxx11_tensor_of_strings "-std=c++0x") ei_add_test(cxx11_tensor_intdiv "-std=c++0x") ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") ei_add_test(cxx11_tensor_broadcasting "-std=c++0x") + ei_add_test(cxx11_tensor_concatenation "-std=c++0x") ei_add_test(cxx11_tensor_morphing "-std=c++0x") ei_add_test(cxx11_tensor_padding "-std=c++0x") + ei_add_test(cxx11_tensor_reduction "-std=c++0x") # ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_concatenation.cpp b/unsupported/test/cxx11_tensor_concatenation.cpp new file mode 100644 index 000000000..8fd4f5f80 --- /dev/null +++ b/unsupported/test/cxx11_tensor_concatenation.cpp @@ -0,0 +1,110 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +static void test_dimension_failures() +{ + Tensor left(2, 3, 1); + Tensor right(3, 3, 1); + left.setRandom(); + right.setRandom(); + + // Okay; other dimensions are equal. + Tensor concatenation = left.concatenate(right, 0); + + // Dimension mismatches. + VERIFY_RAISES_ASSERT(concatenation = left.concatenate(right, 1)); + VERIFY_RAISES_ASSERT(concatenation = left.concatenate(right, 2)); + + // Axis > NumDims or < 0. + VERIFY_RAISES_ASSERT(concatenation = left.concatenate(right, 3)); + VERIFY_RAISES_ASSERT(concatenation = left.concatenate(right, -1)); +} + +static void test_static_dimension_failure() +{ + Tensor left(2, 3); + Tensor right(2, 3, 1); + +#ifdef CXX11_TENSOR_CONCATENATION_STATIC_DIMENSION_FAILURE + // Technically compatible, but we static assert that the inputs have same + // NumDims. + Tensor concatenation = left.concatenate(right, 0); +#endif + + // This can be worked around in this case. + Tensor concatenation = left + .reshape(Tensor::Dimensions{{2, 3, 1}}) + .concatenate(right, 0); + Tensor alternative = left + .concatenate(right.reshape(Tensor::Dimensions{{2, 3}}), 0); +} + +static void test_simple_concatenation() +{ + Tensor left(2, 3, 1); + Tensor right(2, 3, 1); + left.setRandom(); + right.setRandom(); + + Tensor concatenation = left.concatenate(right, 0); + VERIFY_IS_EQUAL(concatenation.dimension(0), 4); + VERIFY_IS_EQUAL(concatenation.dimension(1), 3); + VERIFY_IS_EQUAL(concatenation.dimension(2), 1); + for (int j = 0; j < 3; ++j) { + for (int i = 0; i < 2; ++i) { + VERIFY_IS_EQUAL(concatenation(i, j, 0), left(i, j, 0)); + } + for (int i = 2; i < 4; ++i) { + VERIFY_IS_EQUAL(concatenation(i, j, 0), right(i - 2, j, 0)); + } + } + + concatenation = left.concatenate(right, 1); + VERIFY_IS_EQUAL(concatenation.dimension(0), 2); + VERIFY_IS_EQUAL(concatenation.dimension(1), 6); + VERIFY_IS_EQUAL(concatenation.dimension(2), 1); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(concatenation(i, j, 0), left(i, j, 0)); + } + for (int j = 3; j < 6; ++j) { + VERIFY_IS_EQUAL(concatenation(i, j, 0), right(i, j - 3, 0)); + } + } + + concatenation = left.concatenate(right, 2); + VERIFY_IS_EQUAL(concatenation.dimension(0), 2); + VERIFY_IS_EQUAL(concatenation.dimension(1), 3); + VERIFY_IS_EQUAL(concatenation.dimension(2), 2); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(concatenation(i, j, 0), left(i, j, 0)); + VERIFY_IS_EQUAL(concatenation(i, j, 1), right(i, j, 0)); + } + } +} + + +// TODO(phli): Add test once we have a real vectorized implementation. +// static void test_vectorized_concatenation() {} + + +void test_cxx11_tensor_concatenation() +{ + CALL_SUBTEST(test_dimension_failures()); + CALL_SUBTEST(test_static_dimension_failure()); + CALL_SUBTEST(test_simple_concatenation()); + // CALL_SUBTEST(test_vectorized_concatenation()); +} diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp new file mode 100644 index 000000000..27135b982 --- /dev/null +++ b/unsupported/test/cxx11_tensor_reduction.cpp @@ -0,0 +1,147 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" +#include +#include + +using Eigen::Tensor; + +static void test_simple_reductions() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + array reduction_axis; + reduction_axis[0] = 1; + reduction_axis[1] = 3; + + Tensor result = tensor.sum(reduction_axis); + VERIFY_IS_EQUAL(result.dimension(0), 2); + VERIFY_IS_EQUAL(result.dimension(1), 5); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 5; ++j) { + float sum = 0.0f; + for (int k = 0; k < 3; ++k) { + for (int l = 0; l < 7; ++l) { + sum += tensor(i, k, j, l); + } + } + VERIFY_IS_APPROX(result(i, j), sum); + } + } + + reduction_axis[0] = 0; + reduction_axis[1] = 2; + result = tensor.maximum(reduction_axis); + VERIFY_IS_EQUAL(result.dimension(0), 3); + VERIFY_IS_EQUAL(result.dimension(1), 7); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 7; ++j) { + float max_val = std::numeric_limits::lowest(); + for (int k = 0; k < 2; ++k) { + for (int l = 0; l < 5; ++l) { + max_val = (std::max)(max_val, tensor(k, i, l, j)); + } + } + VERIFY_IS_APPROX(result(i, j), max_val); + } + } + + reduction_axis[0] = 0; + reduction_axis[1] = 1; + result = tensor.minimum(reduction_axis); + VERIFY_IS_EQUAL(result.dimension(0), 5); + VERIFY_IS_EQUAL(result.dimension(1), 7); + for (int i = 0; i < 5; ++i) { + for (int j = 0; j < 7; ++j) { + float min_val = (std::numeric_limits::max)(); + for (int k = 0; k < 2; ++k) { + for (int l = 0; l < 3; ++l) { + min_val = (std::min)(min_val, tensor(k, l, i, j)); + } + } + VERIFY_IS_APPROX(result(i, j), min_val); + } + } +} + + +static void test_full_reductions() +{ + Tensor tensor(2,3); + tensor.setRandom(); + array reduction_axis; + reduction_axis[0] = 0; + reduction_axis[1] = 1; + + Tensor result = tensor.sum(reduction_axis); + VERIFY_IS_EQUAL(result.dimension(0), 1); + + float sum = 0.0f; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + sum += tensor(i, j); + } + } + VERIFY_IS_APPROX(result(0), sum); + + result = tensor.square().sum(reduction_axis).sqrt(); + VERIFY_IS_EQUAL(result.dimension(0), 1); + + sum = 0.0f; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + sum += tensor(i, j) * tensor(i, j); + } + } + VERIFY_IS_APPROX(result(0), sqrtf(sum)); +} + + +struct UserReducer { + UserReducer(float offset) : offset_(offset), sum_(0.0f) {} + void reduce(const float val) { + sum_ += val * val; + } + float finalize() const { + return 1.0f / (sum_ + offset_); + } + + private: + float offset_; + float sum_; +}; + +static void test_user_defined_reductions() +{ + Tensor tensor(5,7); + tensor.setRandom(); + array reduction_axis; + reduction_axis[0] = 1; + + UserReducer reducer(10.0f); + Tensor result = tensor.reduce(reduction_axis, reducer); + VERIFY_IS_EQUAL(result.dimension(0), 5); + for (int i = 0; i < 5; ++i) { + float expected = 10.0f; + for (int j = 0; j < 7; ++j) { + expected += tensor(i, j) * tensor(i, j); + } + expected = 1.0f / expected; + VERIFY_IS_APPROX(result(i), expected); + } +} + + +void test_cxx11_tensor_reduction() +{ + CALL_SUBTEST(test_simple_reductions()); + CALL_SUBTEST(test_full_reductions()); + CALL_SUBTEST(test_user_defined_reductions()); +} From 5cc23199be743d0d1be85d709eb366e67e87a262 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 2 Oct 2014 10:30:44 -0700 Subject: [PATCH 0861/1103] More tests to validate the const-correctness of the tensor code. --- Eigen/src/Core/GenericPacketMath.h | 2 ++ Eigen/src/Core/util/XprHelper.h | 8 +++++ unsupported/test/CMakeLists.txt | 3 +- unsupported/test/cxx11_tensor_const.cpp | 39 +++++++++++++++++++++++++ 4 files changed, 51 insertions(+), 1 deletion(-) create mode 100644 unsupported/test/cxx11_tensor_const.cpp diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 6ec29d0fd..e6fea5bba 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -95,6 +95,8 @@ template struct packet_traits : default_packet_traits }; }; +template struct packet_traits : packet_traits { }; + /** \internal \returns a + b (coeff-wise) */ template EIGEN_DEVICE_FUNC inline Packet padd(const Packet& a, diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index 7c77b2263..67ca49754 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -415,6 +415,14 @@ template struct promote_storage_type { typedef A ret; }; +template struct promote_storage_type +{ + typedef A ret; +}; +template struct promote_storage_type +{ + typedef A ret; +}; /** \internal gives the plain matrix or array type to store a row/column/diagonal of a matrix type. * \param Scalar optional parameter allowing to pass a different scalar type than the one of the MatrixType. diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index e83d8b54e..a47c7bc74 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -106,7 +106,8 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_convolution "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") -# ei_add_test(cxx11_tensor_of_const_values "-std=c++0x") + ei_add_test(cxx11_tensor_const "-std=c++0x") + ei_add_test(cxx11_tensor_of_const_values "-std=c++0x") ei_add_test(cxx11_tensor_of_strings "-std=c++0x") ei_add_test(cxx11_tensor_intdiv "-std=c++0x") ei_add_test(cxx11_tensor_lvalue "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_const.cpp b/unsupported/test/cxx11_tensor_const.cpp new file mode 100644 index 000000000..0ffb02afd --- /dev/null +++ b/unsupported/test/cxx11_tensor_const.cpp @@ -0,0 +1,39 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include +using Eigen::Tensor; + + + + +static void test_simple_assign() +{ + Tensor random(2,3,7); + random.setRandom(); + + TensorMap > constant(random.data(), 2, 3, 7); + Tensor result(2,3,7); + result = constant; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL((result(i,j,k)), random(i,j,k)); + } + } + } +} + +void test_cxx11_tensor_const() +{ + CALL_SUBTEST(test_simple_assign()); +} From 8b2afe33a165ff0cc5a7afd14fcfb06cdf703235 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 2 Oct 2014 10:39:36 -0700 Subject: [PATCH 0862/1103] Fixes for the forced evaluation of tensor expressions More tests --- .../Eigen/CXX11/src/Tensor/TensorForcedEval.h | 13 +++-- unsupported/test/CMakeLists.txt | 3 +- unsupported/test/cxx11_tensor_dimension.cpp | 51 +++++++++++++++++++ unsupported/test/cxx11_tensor_forced_eval.cpp | 51 +++++++++++++++++++ 4 files changed, 110 insertions(+), 8 deletions(-) create mode 100644 unsupported/test/cxx11_tensor_dimension.cpp create mode 100644 unsupported/test/cxx11_tensor_forced_eval.cpp diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h index 6f6641de6..cb14cc7f7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h @@ -87,31 +87,28 @@ struct TensorEvaluator, Device> enum { IsAligned = true, - PacketAccess = true, + PacketAccess = (internal::packet_traits::size > 1), }; EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) : m_impl(op.expression(), device), m_op(op.expression()), m_device(device), m_buffer(NULL) { } - EIGEN_DEVICE_FUNC ~TensorEvaluator() { - eigen_assert(!m_buffer); - } - typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); } - EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() { - m_impl.evalSubExprsIfNeeded(); + EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + m_impl.evalSubExprsIfNeeded(NULL); m_buffer = (Scalar*)m_device.allocate(m_impl.dimensions().TotalSize() * sizeof(Scalar)); typedef TensorEvalToOp EvalTo; EvalTo evalToTmp(m_buffer, m_op); internal::TensorExecutor::PacketAccess>::run(evalToTmp, m_device); m_impl.cleanup(); + return true; } EIGEN_STRONG_INLINE void cleanup() { m_device.deallocate(m_buffer); @@ -129,6 +126,8 @@ struct TensorEvaluator, Device> return internal::ploadt(m_buffer + index); } + Scalar* data() const { return m_buffer; } + private: TensorEvaluator m_impl; const ArgType m_op; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index a47c7bc74..5d8913dd8 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -101,10 +101,12 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_simple "-std=c++0x") ei_add_test(cxx11_tensor_symmetry "-std=c++0x") # ei_add_test(cxx11_tensor_assign "-std=c++0x") +# ei_add_test(cxx11_tensor_dimension "-std=c++0x") ei_add_test(cxx11_tensor_comparisons "-std=c++0x") ei_add_test(cxx11_tensor_contraction "-std=c++0x") ei_add_test(cxx11_tensor_convolution "-std=c++0x") ei_add_test(cxx11_tensor_expr "-std=c++0x") + ei_add_test(cxx11_tensor_forced_eval "-std=c++0x") # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") ei_add_test(cxx11_tensor_const "-std=c++0x") ei_add_test(cxx11_tensor_of_const_values "-std=c++0x") @@ -120,6 +122,5 @@ if(EIGEN_TEST_CXX11) # ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") -# ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") # ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_dimension.cpp b/unsupported/test/cxx11_tensor_dimension.cpp new file mode 100644 index 000000000..fc0d29c50 --- /dev/null +++ b/unsupported/test/cxx11_tensor_dimension.cpp @@ -0,0 +1,51 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + + +static void test_dynamic_size() +{ + Eigen::DSizes dimensions(Eigen::array(2,3,7)); + + VERIFY_IS_EQUAL((int)Eigen::internal::array_get<0>(dimensions), 2); + VERIFY_IS_EQUAL((int)Eigen::internal::array_get<1>(dimensions), 3); + VERIFY_IS_EQUAL((int)Eigen::internal::array_get<2>(dimensions), 7); + VERIFY_IS_EQUAL(dimensions.TotalSize(), (size_t)2*3*7); +} + +static void test_fixed_size() +{ + Eigen::Sizes<2,3,7> dimensions; + + VERIFY_IS_EQUAL((int)Eigen::internal::array_get<0>(dimensions), 2); + VERIFY_IS_EQUAL((int)Eigen::internal::array_get<1>(dimensions), 3); + VERIFY_IS_EQUAL((int)Eigen::internal::array_get<2>(dimensions), 7); + VERIFY_IS_EQUAL(dimensions.TotalSize(), (size_t)2*3*7); +} + + +static void test_match() +{ + Eigen::DSizes dyn(Eigen::array(2,3,7)); + Eigen::Sizes<2,3,7> stat; + VERIFY_IS_EQUAL(Eigen::internal::dimensions_match(dyn, stat), true); +} + + +void test_cxx11_tensor_dimension() +{ + CALL_SUBTEST(test_dynamic_size()); + CALL_SUBTEST(test_fixed_size()); + CALL_SUBTEST(test_match()); +} diff --git a/unsupported/test/cxx11_tensor_forced_eval.cpp b/unsupported/test/cxx11_tensor_forced_eval.cpp new file mode 100644 index 000000000..529584a7b --- /dev/null +++ b/unsupported/test/cxx11_tensor_forced_eval.cpp @@ -0,0 +1,51 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include +#include + +using Eigen::MatrixXf; +using Eigen::Tensor; + +static void test_simple() +{ + MatrixXf m1(3,3); + MatrixXf m2(3,3); + m1.setRandom(); + m2.setRandom(); + + TensorMap> mat1(m1.data(), 3,3); + TensorMap> mat2(m2.data(), 3,3); + + Tensor mat3(3,3); + mat3 = mat1; + + typedef Tensor::DimensionPair DimPair; + Eigen::array dims({{DimPair(1, 0)}}); + + mat3 = mat3.contract(mat2, dims).eval(); + + VERIFY_IS_APPROX(mat3(0, 0), (m1*m2).eval()(0,0)); + VERIFY_IS_APPROX(mat3(0, 1), (m1*m2).eval()(0,1)); + VERIFY_IS_APPROX(mat3(0, 2), (m1*m2).eval()(0,2)); + VERIFY_IS_APPROX(mat3(1, 0), (m1*m2).eval()(1,0)); + VERIFY_IS_APPROX(mat3(1, 1), (m1*m2).eval()(1,1)); + VERIFY_IS_APPROX(mat3(1, 2), (m1*m2).eval()(1,2)); + VERIFY_IS_APPROX(mat3(2, 0), (m1*m2).eval()(2,0)); + VERIFY_IS_APPROX(mat3(2, 1), (m1*m2).eval()(2,1)); + VERIFY_IS_APPROX(mat3(2, 2), (m1*m2).eval()(2,2)); +} + + +void test_cxx11_tensor_forced_eval() +{ + CALL_SUBTEST(test_simple()); +} From b7271dffb5b1ceeee4c8bd99402ff89dcce58d74 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 2 Oct 2014 16:51:57 -0700 Subject: [PATCH 0863/1103] Generalized the gebp apis --- .../Core/products/GeneralBlockPanelKernel.h | 423 +++++++++--------- Eigen/src/Core/products/GeneralMatrixMatrix.h | 80 ++-- .../products/GeneralMatrixMatrixTriangular.h | 54 ++- .../Core/products/SelfadjointMatrixMatrix.h | 51 ++- .../Core/products/TriangularMatrixMatrix.h | 65 +-- .../Core/products/TriangularSolverMatrix.h | 49 +- Eigen/src/Core/util/BlasUtil.h | 108 ++++- unsupported/test/CMakeLists.txt | 2 +- 8 files changed, 474 insertions(+), 358 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 7da52c2e8..090c8f4e6 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -667,7 +667,7 @@ protected: * |real |cplx | no vectorization yet, would require to pack A with duplication * |cplx |real | easy vectorization */ -template +template struct gebp_kernel { typedef gebp_traits Traits; @@ -676,14 +676,15 @@ struct gebp_kernel typedef typename Traits::RhsPacket RhsPacket; typedef typename Traits::ResPacket ResPacket; typedef typename Traits::AccPacket AccPacket; - + typedef gebp_traits SwappedTraits; typedef typename SwappedTraits::ResScalar SResScalar; typedef typename SwappedTraits::LhsPacket SLhsPacket; typedef typename SwappedTraits::RhsPacket SRhsPacket; typedef typename SwappedTraits::ResPacket SResPacket; typedef typename SwappedTraits::AccPacket SAccPacket; - + + typedef typename DataMapper::LinearMapper LinearMapper; enum { Vectorizable = Traits::Vectorizable, @@ -693,14 +694,16 @@ struct gebp_kernel }; EIGEN_DONT_INLINE - void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha, + void operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB, + Index rows, Index depth, Index cols, ResScalar alpha, Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0); }; -template +template EIGEN_DONT_INLINE -void gebp_kernel - ::operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha, +void gebp_kernel + ::operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB, + Index rows, Index depth, Index cols, ResScalar alpha, Index strideA, Index strideB, Index offsetA, Index offsetB) { Traits traits; @@ -743,15 +746,15 @@ void gebp_kernel traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7); traits.initAcc(C8); traits.initAcc(C9); traits.initAcc(C10); traits.initAcc(C11); - ResScalar* r0 = &res[(j2+0)*resStride + i]; - ResScalar* r1 = &res[(j2+1)*resStride + i]; - ResScalar* r2 = &res[(j2+2)*resStride + i]; - ResScalar* r3 = &res[(j2+3)*resStride + i]; - - internal::prefetch(r0); - internal::prefetch(r1); - internal::prefetch(r2); - internal::prefetch(r3); + LinearMapper r0 = res.getLinearMapper(i, j2 + 0); + LinearMapper r1 = res.getLinearMapper(i, j2 + 1); + LinearMapper r2 = res.getLinearMapper(i, j2 + 2); + LinearMapper r3 = res.getLinearMapper(i, j2 + 3); + + r0.prefetch(0); + r1.prefetch(0); + r2.prefetch(0); + r3.prefetch(0); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; @@ -813,48 +816,48 @@ void gebp_kernel ResPacket R0, R1, R2; ResPacket alphav = pset1(alpha); - - R0 = ploadu(r0+0*Traits::ResPacketSize); - R1 = ploadu(r0+1*Traits::ResPacketSize); - R2 = ploadu(r0+2*Traits::ResPacketSize); + + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); + R2 = r0.loadPacket(2 * Traits::ResPacketSize); traits.acc(C0, alphav, R0); traits.acc(C4, alphav, R1); traits.acc(C8, alphav, R2); - pstoreu(r0+0*Traits::ResPacketSize, R0); - pstoreu(r0+1*Traits::ResPacketSize, R1); - pstoreu(r0+2*Traits::ResPacketSize, R2); - - R0 = ploadu(r1+0*Traits::ResPacketSize); - R1 = ploadu(r1+1*Traits::ResPacketSize); - R2 = ploadu(r1+2*Traits::ResPacketSize); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); + r0.storePacket(2 * Traits::ResPacketSize, R2); + + R0 = r1.loadPacket(0 * Traits::ResPacketSize); + R1 = r1.loadPacket(1 * Traits::ResPacketSize); + R2 = r1.loadPacket(2 * Traits::ResPacketSize); traits.acc(C1, alphav, R0); traits.acc(C5, alphav, R1); traits.acc(C9, alphav, R2); - pstoreu(r1+0*Traits::ResPacketSize, R0); - pstoreu(r1+1*Traits::ResPacketSize, R1); - pstoreu(r1+2*Traits::ResPacketSize, R2); - - R0 = ploadu(r2+0*Traits::ResPacketSize); - R1 = ploadu(r2+1*Traits::ResPacketSize); - R2 = ploadu(r2+2*Traits::ResPacketSize); + r1.storePacket(0 * Traits::ResPacketSize, R0); + r1.storePacket(1 * Traits::ResPacketSize, R1); + r1.storePacket(2 * Traits::ResPacketSize, R2); + + R0 = r2.loadPacket(0 * Traits::ResPacketSize); + R1 = r2.loadPacket(1 * Traits::ResPacketSize); + R2 = r2.loadPacket(2 * Traits::ResPacketSize); traits.acc(C2, alphav, R0); traits.acc(C6, alphav, R1); traits.acc(C10, alphav, R2); - pstoreu(r2+0*Traits::ResPacketSize, R0); - pstoreu(r2+1*Traits::ResPacketSize, R1); - pstoreu(r2+2*Traits::ResPacketSize, R2); - - R0 = ploadu(r3+0*Traits::ResPacketSize); - R1 = ploadu(r3+1*Traits::ResPacketSize); - R2 = ploadu(r3+2*Traits::ResPacketSize); + r2.storePacket(0 * Traits::ResPacketSize, R0); + r2.storePacket(1 * Traits::ResPacketSize, R1); + r2.storePacket(2 * Traits::ResPacketSize, R2); + + R0 = r3.loadPacket(0 * Traits::ResPacketSize); + R1 = r3.loadPacket(1 * Traits::ResPacketSize); + R2 = r3.loadPacket(2 * Traits::ResPacketSize); traits.acc(C3, alphav, R0); traits.acc(C7, alphav, R1); traits.acc(C11, alphav, R2); - pstoreu(r3+0*Traits::ResPacketSize, R0); - pstoreu(r3+1*Traits::ResPacketSize, R1); - pstoreu(r3+2*Traits::ResPacketSize, R2); + r3.storePacket(0 * Traits::ResPacketSize, R0); + r3.storePacket(1 * Traits::ResPacketSize, R1); + r3.storePacket(2 * Traits::ResPacketSize, R2); } - + // Deal with remaining columns of the rhs for(Index j2=packet_cols4; j2 traits.initAcc(C4); traits.initAcc(C8); - ResScalar* r0 = &res[(j2+0)*resStride + i]; + LinearMapper r0 = res.getLinearMapper(i, j2); + r0.prefetch(0); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB]; @@ -912,19 +916,19 @@ void gebp_kernel ResPacket R0, R1, R2; ResPacket alphav = pset1(alpha); - R0 = ploadu(r0+0*Traits::ResPacketSize); - R1 = ploadu(r0+1*Traits::ResPacketSize); - R2 = ploadu(r0+2*Traits::ResPacketSize); + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); + R2 = r0.loadPacket(2 * Traits::ResPacketSize); traits.acc(C0, alphav, R0); traits.acc(C4, alphav, R1); - traits.acc(C8 , alphav, R2); - pstoreu(r0+0*Traits::ResPacketSize, R0); - pstoreu(r0+1*Traits::ResPacketSize, R1); - pstoreu(r0+2*Traits::ResPacketSize, R2); + traits.acc(C8, alphav, R2); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); + r0.storePacket(2 * Traits::ResPacketSize, R2); } } } - + //---------- Process 2 * LhsProgress rows at once ---------- if(mr>=2*Traits::LhsProgress) { @@ -946,15 +950,15 @@ void gebp_kernel traits.initAcc(C0); traits.initAcc(C1); traits.initAcc(C2); traits.initAcc(C3); traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7); - ResScalar* r0 = &res[(j2+0)*resStride + i]; - ResScalar* r1 = &res[(j2+1)*resStride + i]; - ResScalar* r2 = &res[(j2+2)*resStride + i]; - ResScalar* r3 = &res[(j2+3)*resStride + i]; - - internal::prefetch(r0+prefetch_res_offset); - internal::prefetch(r1+prefetch_res_offset); - internal::prefetch(r2+prefetch_res_offset); - internal::prefetch(r3+prefetch_res_offset); + LinearMapper r0 = res.getLinearMapper(i, j2 + 0); + LinearMapper r1 = res.getLinearMapper(i, j2 + 1); + LinearMapper r2 = res.getLinearMapper(i, j2 + 2); + LinearMapper r3 = res.getLinearMapper(i, j2 + 3); + + r0.prefetch(prefetch_res_offset); + r1.prefetch(prefetch_res_offset); + r2.prefetch(prefetch_res_offset); + r3.prefetch(prefetch_res_offset); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; @@ -978,7 +982,7 @@ void gebp_kernel traits.madd(A1, B2, C6, B2); \ traits.madd(A0, B3, C3, T0); \ traits.madd(A1, B3, C7, B3) - + internal::prefetch(blB+(48+0)); EIGEN_GEBGP_ONESTEP(0); EIGEN_GEBGP_ONESTEP(1); @@ -1002,37 +1006,37 @@ void gebp_kernel blA += 2*Traits::LhsProgress; } #undef EIGEN_GEBGP_ONESTEP - + ResPacket R0, R1, R2, R3; ResPacket alphav = pset1(alpha); - - R0 = ploadu(r0+0*Traits::ResPacketSize); - R1 = ploadu(r0+1*Traits::ResPacketSize); - R2 = ploadu(r1+0*Traits::ResPacketSize); - R3 = ploadu(r1+1*Traits::ResPacketSize); + + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); + R2 = r1.loadPacket(0 * Traits::ResPacketSize); + R3 = r1.loadPacket(1 * Traits::ResPacketSize); traits.acc(C0, alphav, R0); traits.acc(C4, alphav, R1); traits.acc(C1, alphav, R2); traits.acc(C5, alphav, R3); - pstoreu(r0+0*Traits::ResPacketSize, R0); - pstoreu(r0+1*Traits::ResPacketSize, R1); - pstoreu(r1+0*Traits::ResPacketSize, R2); - pstoreu(r1+1*Traits::ResPacketSize, R3); - - R0 = ploadu(r2+0*Traits::ResPacketSize); - R1 = ploadu(r2+1*Traits::ResPacketSize); - R2 = ploadu(r3+0*Traits::ResPacketSize); - R3 = ploadu(r3+1*Traits::ResPacketSize); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); + r1.storePacket(0 * Traits::ResPacketSize, R2); + r1.storePacket(1 * Traits::ResPacketSize, R3); + + R0 = r2.loadPacket(0 * Traits::ResPacketSize); + R1 = r2.loadPacket(1 * Traits::ResPacketSize); + R2 = r3.loadPacket(0 * Traits::ResPacketSize); + R3 = r3.loadPacket(1 * Traits::ResPacketSize); traits.acc(C2, alphav, R0); traits.acc(C6, alphav, R1); traits.acc(C3, alphav, R2); traits.acc(C7, alphav, R3); - pstoreu(r2+0*Traits::ResPacketSize, R0); - pstoreu(r2+1*Traits::ResPacketSize, R1); - pstoreu(r3+0*Traits::ResPacketSize, R2); - pstoreu(r3+1*Traits::ResPacketSize, R3); + r2.storePacket(0 * Traits::ResPacketSize, R0); + r2.storePacket(1 * Traits::ResPacketSize, R1); + r3.storePacket(0 * Traits::ResPacketSize, R2); + r3.storePacket(1 * Traits::ResPacketSize, R3); } - + // Deal with remaining columns of the rhs for(Index j2=packet_cols4; j2 traits.initAcc(C0); traits.initAcc(C4); - ResScalar* r0 = &res[(j2+0)*resStride + i]; - internal::prefetch(r0+prefetch_res_offset); + LinearMapper r0 = res.getLinearMapper(i, j2); + r0.prefetch(prefetch_res_offset); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB]; @@ -1089,12 +1093,12 @@ void gebp_kernel ResPacket R0, R1; ResPacket alphav = pset1(alpha); - R0 = ploadu(r0+0*Traits::ResPacketSize); - R1 = ploadu(r0+1*Traits::ResPacketSize); + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r0.loadPacket(1 * Traits::ResPacketSize); traits.acc(C0, alphav, R0); traits.acc(C4, alphav, R1); - pstoreu(r0+0*Traits::ResPacketSize, R0); - pstoreu(r0+1*Traits::ResPacketSize, R1); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r0.storePacket(1 * Traits::ResPacketSize, R1); } } } @@ -1120,15 +1124,15 @@ void gebp_kernel traits.initAcc(C2); traits.initAcc(C3); - ResScalar* r0 = &res[(j2+0)*resStride + i]; - ResScalar* r1 = &res[(j2+1)*resStride + i]; - ResScalar* r2 = &res[(j2+2)*resStride + i]; - ResScalar* r3 = &res[(j2+3)*resStride + i]; - - internal::prefetch(r0+prefetch_res_offset); - internal::prefetch(r1+prefetch_res_offset); - internal::prefetch(r2+prefetch_res_offset); - internal::prefetch(r3+prefetch_res_offset); + LinearMapper r0 = res.getLinearMapper(i, j2 + 0); + LinearMapper r1 = res.getLinearMapper(i, j2 + 1); + LinearMapper r2 = res.getLinearMapper(i, j2 + 2); + LinearMapper r3 = res.getLinearMapper(i, j2 + 3); + + r0.prefetch(prefetch_res_offset); + r1.prefetch(prefetch_res_offset); + r2.prefetch(prefetch_res_offset); + r3.prefetch(prefetch_res_offset); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; @@ -1171,25 +1175,25 @@ void gebp_kernel blA += 1*LhsProgress; } #undef EIGEN_GEBGP_ONESTEP - + ResPacket R0, R1; ResPacket alphav = pset1(alpha); - - R0 = ploadu(r0+0*Traits::ResPacketSize); - R1 = ploadu(r1+0*Traits::ResPacketSize); + + R0 = r0.loadPacket(0 * Traits::ResPacketSize); + R1 = r1.loadPacket(0 * Traits::ResPacketSize); traits.acc(C0, alphav, R0); traits.acc(C1, alphav, R1); - pstoreu(r0+0*Traits::ResPacketSize, R0); - pstoreu(r1+0*Traits::ResPacketSize, R1); - - R0 = ploadu(r2+0*Traits::ResPacketSize); - R1 = ploadu(r3+0*Traits::ResPacketSize); + r0.storePacket(0 * Traits::ResPacketSize, R0); + r1.storePacket(0 * Traits::ResPacketSize, R1); + + R0 = r2.loadPacket(0 * Traits::ResPacketSize); + R1 = r3.loadPacket(0 * Traits::ResPacketSize); traits.acc(C2, alphav, R0); traits.acc(C3, alphav, R1); - pstoreu(r2+0*Traits::ResPacketSize, R0); - pstoreu(r3+0*Traits::ResPacketSize, R1); + r2.storePacket(0 * Traits::ResPacketSize, R0); + r3.storePacket(0 * Traits::ResPacketSize, R1); } - + // Deal with remaining columns of the rhs for(Index j2=packet_cols4; j2 AccPacket C0; traits.initAcc(C0); - ResScalar* r0 = &res[(j2+0)*resStride + i]; + LinearMapper r0 = res.getLinearMapper(i, j2); // performs "inner" products const RhsScalar* blB = &blockB[j2*strideB+offsetB]; @@ -1241,9 +1245,9 @@ void gebp_kernel #undef EIGEN_GEBGP_ONESTEP ResPacket R0; ResPacket alphav = pset1(alpha); - R0 = ploadu(r0+0*Traits::ResPacketSize); + R0 = r0.loadPacket(0 * Traits::ResPacketSize); traits.acc(C0, alphav, R0); - pstoreu(r0+0*Traits::ResPacketSize, R0); + r0.storePacket(0 * Traits::ResPacketSize, R0); } } } @@ -1259,7 +1263,7 @@ void gebp_kernel const LhsScalar* blA = &blockA[i*strideA+offsetA]; prefetch(&blA[0]); const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr]; - + if( (SwappedTraits::LhsProgress % 4)==0 ) { // NOTE The following piece of code wont work for 512 bit registers @@ -1268,32 +1272,32 @@ void gebp_kernel straits.initAcc(C1); straits.initAcc(C2); straits.initAcc(C3); - + const Index spk = (std::max)(1,SwappedTraits::LhsProgress/4); const Index endk = (depth/spk)*spk; const Index endk4 = (depth/(spk*4))*(spk*4); - + Index k=0; for(; k { SLhsPacket A0; SRhsPacket B_0; - + straits.loadLhsUnaligned(blB, A0); straits.loadRhsQuad(blA, B_0); straits.madd(A0,B_0,C0,B_0); - + blB += SwappedTraits::LhsProgress; blA += spk; } @@ -1317,10 +1321,10 @@ void gebp_kernel typedef typename conditional::half,SLhsPacket>::type SLhsPacketHalf; typedef typename conditional::half,SRhsPacket>::type SRhsPacketHalf; typedef typename conditional::half,SAccPacket>::type SAccPacketHalf; - - SResPacketHalf R = pgather(&res[j2*resStride + i], resStride); + + SResPacketHalf R = res.template gatherPacket(i, j2); SResPacketHalf alphav = pset1(alpha); - + if(depth-endk>0) { // We have to handle the last row of the rhs which corresponds to a half-packet @@ -1336,14 +1340,14 @@ void gebp_kernel { straits.acc(predux4(C0), alphav, R); } - pscatter(&res[j2*resStride + i], R, resStride); + res.scatterPacket(i, j2, R); } else { - SResPacket R = pgather(&res[j2*resStride + i], resStride); + SResPacket R = res.template gatherPacket(i, j2); SResPacket alphav = pset1(alpha); straits.acc(C0, alphav, R); - pscatter(&res[j2*resStride + i], R, resStride); + res.scatterPacket(i, j2, R); } } else // scalar path @@ -1355,25 +1359,25 @@ void gebp_kernel { LhsScalar A0; RhsScalar B_0, B_1; - + A0 = blA[k]; - + B_0 = blB[0]; B_1 = blB[1]; MADD(cj,A0,B_0,C0, B_0); MADD(cj,A0,B_1,C1, B_1); - + B_0 = blB[2]; B_1 = blB[3]; MADD(cj,A0,B_0,C2, B_0); MADD(cj,A0,B_1,C3, B_1); - + blB += 4; } - res[(j2+0)*resStride + i] += alpha*C0; - res[(j2+1)*resStride + i] += alpha*C1; - res[(j2+2)*resStride + i] += alpha*C2; - res[(j2+3)*resStride + i] += alpha*C3; + res(i, j2 + 0) += alpha * C0; + res(i, j2 + 1) += alpha * C1; + res(i, j2 + 2) += alpha * C2; + res(i, j2 + 3) += alpha * C3; } } } @@ -1394,7 +1398,7 @@ void gebp_kernel RhsScalar B_0 = blB[k]; MADD(cj, A0, B_0, C0, B_0); } - res[(j2+0)*resStride + i] += alpha*C0; + res(i, j2) += alpha * C0; } } } @@ -1417,15 +1421,16 @@ void gebp_kernel // // 32 33 34 35 ... // 36 36 38 39 ... -template -struct gemm_pack_lhs +template +struct gemm_pack_lhs { - EIGEN_DONT_INLINE void operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride=0, Index offset=0); + typedef typename DataMapper::LinearMapper LinearMapper; + EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0); }; -template -EIGEN_DONT_INLINE void gemm_pack_lhs - ::operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride, Index offset) +template +EIGEN_DONT_INLINE void gemm_pack_lhs + ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset) { typedef typename packet_traits::type Packet; enum { PacketSize = packet_traits::size }; @@ -1436,30 +1441,29 @@ EIGEN_DONT_INLINE void gemm_pack_lhs=depth && offset<=stride)); eigen_assert( ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) || (Pack1<=4) ); conj_if::IsComplex && Conjugate> cj; - const_blas_data_mapper lhs(_lhs,lhsStride); Index count = 0; - + const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; const Index peeled_mc0 = Pack2>=1*PacketSize ? peeled_mc1 : Pack2>1 ? (rows/Pack2)*Pack2 : 0; - + Index i=0; - + // Pack 3 packets if(Pack1>=3*PacketSize) { for(; i(&lhs(i+0*PacketSize, k)); - B = ploadu(&lhs(i+1*PacketSize, k)); - C = ploadu(&lhs(i+2*PacketSize, k)); + A = lhs.loadPacket(i+0*PacketSize, k); + B = lhs.loadPacket(i+1*PacketSize, k); + C = lhs.loadPacket(i+2*PacketSize, k); pstore(blockA+count, cj.pconj(A)); count+=PacketSize; pstore(blockA+count, cj.pconj(B)); count+=PacketSize; pstore(blockA+count, cj.pconj(C)); count+=PacketSize; @@ -1473,12 +1477,12 @@ EIGEN_DONT_INLINE void gemm_pack_lhs(&lhs(i+0*PacketSize, k)); - B = ploadu(&lhs(i+1*PacketSize, k)); + A = lhs.loadPacket(i+0*PacketSize, k); + B = lhs.loadPacket(i+1*PacketSize, k); pstore(blockA+count, cj.pconj(A)); count+=PacketSize; pstore(blockA+count, cj.pconj(B)); count+=PacketSize; } @@ -1491,11 +1495,11 @@ EIGEN_DONT_INLINE void gemm_pack_lhs(&lhs(i+0*PacketSize, k)); + A = lhs.loadPacket(i+0*PacketSize, k); pstore(blockA+count, cj.pconj(A)); count+=PacketSize; } @@ -1508,11 +1512,11 @@ EIGEN_DONT_INLINE void gemm_pack_lhs -struct gemm_pack_lhs +template +struct gemm_pack_lhs { - EIGEN_DONT_INLINE void operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride=0, Index offset=0); + typedef typename DataMapper::LinearMapper LinearMapper; + EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0); }; -template -EIGEN_DONT_INLINE void gemm_pack_lhs - ::operator()(Scalar* blockA, const Scalar* EIGEN_RESTRICT _lhs, Index lhsStride, Index depth, Index rows, Index stride, Index offset) +template +EIGEN_DONT_INLINE void gemm_pack_lhs + ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset) { typedef typename packet_traits::type Packet; enum { PacketSize = packet_traits::size }; @@ -1543,13 +1548,12 @@ EIGEN_DONT_INLINE void gemm_pack_lhs=depth && offset<=stride)); conj_if::IsComplex && Conjugate> cj; - const_blas_data_mapper lhs(_lhs,lhsStride); Index count = 0; - + // const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0; // const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0; // const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0; - + int pack = Pack1; Index i = 0; while(pack>0) @@ -1569,7 +1573,7 @@ EIGEN_DONT_INLINE void gemm_pack_lhs kernel; - for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = ploadu(&lhs(i+p+m, k)); + for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.loadPacket(i+p+m, k); ptranspose(kernel); for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p])); } @@ -1594,15 +1598,15 @@ EIGEN_DONT_INLINE void gemm_pack_lhs -struct gemm_pack_rhs +template +struct gemm_pack_rhs { typedef typename packet_traits::type Packet; + typedef typename DataMapper::LinearMapper LinearMapper; enum { PacketSize = packet_traits::size }; - EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0); + EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0); }; -template -EIGEN_DONT_INLINE void gemm_pack_rhs - ::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset) +template +EIGEN_DONT_INLINE void gemm_pack_rhs + ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset) { EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS COLMAJOR"); EIGEN_UNUSED_VARIABLE(stride); @@ -1685,27 +1690,27 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=4) { for(Index j2=packet_cols8; j2 kernel; - kernel.packet[0] = ploadu(&b0[k]); - kernel.packet[1] = ploadu(&b1[k]); - kernel.packet[2] = ploadu(&b2[k]); - kernel.packet[3] = ploadu(&b3[k]); + kernel.packet[0] = dm0.loadPacket(k); + kernel.packet[1] = dm1.loadPacket(k); + kernel.packet[2] = dm2.loadPacket(k); + kernel.packet[3] = dm3.loadPacket(k); ptranspose(kernel); pstoreu(blockB+count+0*PacketSize, cj.pconj(kernel.packet[0])); pstoreu(blockB+count+1*PacketSize, cj.pconj(kernel.packet[1])); @@ -1716,10 +1721,10 @@ EIGEN_DONT_INLINE void gemm_pack_rhs -struct gemm_pack_rhs +template +struct gemm_pack_rhs { typedef typename packet_traits::type Packet; + typedef typename DataMapper::LinearMapper LinearMapper; enum { PacketSize = packet_traits::size }; - EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0); + EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0); }; -template -EIGEN_DONT_INLINE void gemm_pack_rhs - ::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset) +template +EIGEN_DONT_INLINE void gemm_pack_rhs + ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset) { EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS ROWMAJOR"); EIGEN_UNUSED_VARIABLE(stride); @@ -1762,7 +1768,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs=8 ? (cols/8) * 8 : 0; Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0; Index count = 0; - + // if(nr>=8) // { // for(Index j2=0; j2(&rhs[k*rhsStride + j2]); + Packet A = rhs.loadPacket(k, j2); pstoreu(blockB+count, cj.pconj(A)); count += PacketSize; } else { - const Scalar* b0 = &rhs[k*rhsStride + j2]; - blockB[count+0] = cj(b0[0]); - blockB[count+1] = cj(b0[1]); - blockB[count+2] = cj(b0[2]); - blockB[count+3] = cj(b0[3]); + const LinearMapper dm0 = rhs.getLinearMapper(k, j2); + blockB[count+0] = cj(dm0(0)); + blockB[count+1] = cj(dm0(1)); + blockB[count+2] = cj(dm0(2)); + blockB[count+3] = cj(dm0(3)); count += 4; } } @@ -1825,10 +1831,9 @@ EIGEN_DONT_INLINE void gemm_pack_rhs::ReturnType ResScal static void run(Index rows, Index cols, Index depth, const LhsScalar* _lhs, Index lhsStride, const RhsScalar* _rhs, Index rhsStride, - ResScalar* res, Index resStride, + ResScalar* _res, Index resStride, ResScalar alpha, level3_blocking& blocking, GemmParallelInfo* info = 0) { - const_blas_data_mapper lhs(_lhs,lhsStride); - const_blas_data_mapper rhs(_rhs,rhsStride); + typedef const_blas_data_mapper LhsMapper; + typedef const_blas_data_mapper RhsMapper; + typedef blas_data_mapper ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); Index kc = blocking.kc(); // cache block size along the K direction Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction Index nc = (std::min)(cols,blocking.nc()); // cache block size along the N direction - gemm_pack_lhs pack_lhs; - gemm_pack_rhs pack_rhs; - gebp_kernel gebp; + gemm_pack_lhs pack_lhs; + gemm_pack_rhs pack_rhs; + gebp_kernel gebp; #ifdef EIGEN_HAS_OPENMP if(info) @@ -95,7 +99,7 @@ static void run(Index rows, Index cols, Index depth, // In order to reduce the chance that a thread has to wait for the other, // let's start by packing B'. - pack_rhs(blockB, &rhs(k,0), rhsStride, actual_kc, nc); + pack_rhs(blockB, rhs.getSubMapper(k,0), actual_kc, nc); // Pack A_k to A' in a parallel fashion: // each thread packs the sub block A_k,i to A'_i where i is the thread id. @@ -105,8 +109,8 @@ static void run(Index rows, Index cols, Index depth, // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it. while(info[tid].users!=0) {} info[tid].users += threads; - - pack_lhs(blockA+info[tid].lhs_start*actual_kc, &lhs(info[tid].lhs_start,k), lhsStride, actual_kc, info[tid].lhs_length); + + pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length); // Notify the other threads that the part A'_i is ready to go. info[tid].sync = k; @@ -119,9 +123,12 @@ static void run(Index rows, Index cols, Index depth, // At this point we have to make sure that A'_i has been updated by the thread i, // we use testAndSetOrdered to mimic a volatile access. // However, no need to wait for the B' part which has been updated by the current thread! - if(shift>0) - while(info[i].sync!=k) {} - gebp(res+info[i].lhs_start, resStride, blockA+info[i].lhs_start*actual_kc, blockB, info[i].lhs_length, actual_kc, nc, alpha); + if (shift>0) { + while(info[i].sync!=k) { + } + } + + gebp(res.getSubMapper(info[i].lhs_start, 0), blockA+info[i].lhs_start*actual_kc, blockB, info[i].lhs_length, actual_kc, nc, alpha); } // Then keep going as usual with the remaining B' @@ -130,10 +137,10 @@ static void run(Index rows, Index cols, Index depth, const Index actual_nc = (std::min)(j+nc,cols)-j; // pack B_k,j to B' - pack_rhs(blockB, &rhs(k,j), rhsStride, actual_kc, actual_nc); + pack_rhs(blockB, rhs.getSubMapper(k,j), actual_kc, actual_nc); // C_j += A' * B' - gebp(res+j*resStride, resStride, blockA, blockB, rows, actual_kc, actual_nc, alpha); + gebp(res.getSubMapper(0, j), blockA, blockB, rows, actual_kc, actual_nc, alpha); } // Release all the sub blocks A'_i of A' for the current thread, @@ -159,28 +166,33 @@ static void run(Index rows, Index cols, Index depth, ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB()); // For each horizontal panel of the rhs, and corresponding panel of the lhs... - for(Index k2=0; k2 Pack lhs's panel into a sequential chunk of memory (L2/L3 caching) - // Note that this panel will be read as many times as the number of blocks in the rhs's - // horizontal panel which is, in practice, a very low number. - pack_lhs(blockA, &lhs(0,k2), lhsStride, actual_kc, rows); - - // For each kc x nc block of the rhs's horizontal panel... - for(Index j2=0; j2 Pack lhs's panel into a sequential chunk of memory (L2/L3 caching) + // Note that this panel will be read as many times as the number of blocks in the rhs's + // horizontal panel which is, in practice, a very low number. + pack_lhs(blockA, lhs.getSubMapper(i2,k2), actual_kc, actual_mc); + + // For each kc x nc block of the rhs's horizontal panel... + for(Index j2=0; j2m_nc; computeProductBlockingSizes(this->m_kc, this->m_mc, n); } - + m_sizeA = this->m_mc * this->m_kc; m_sizeB = this->m_kc * this->m_nc; } diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index 225b994d1..daa8a1d8a 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -58,13 +58,17 @@ struct general_matrix_matrix_triangular_product::ReturnType ResScalar; static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* _lhs, Index lhsStride, - const RhsScalar* _rhs, Index rhsStride, ResScalar* res, Index resStride, const ResScalar& alpha) + const RhsScalar* _rhs, Index rhsStride, ResScalar* _res, Index resStride, const ResScalar& alpha) { - const_blas_data_mapper lhs(_lhs,lhsStride); - const_blas_data_mapper rhs(_rhs,rhsStride); - typedef gebp_traits Traits; + typedef const_blas_data_mapper LhsMapper; + typedef const_blas_data_mapper RhsMapper; + typedef blas_data_mapper ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + Index kc = depth; // cache block size along the K direction Index mc = size; // cache block size along the M direction Index nc = size; // cache block size along the N direction @@ -75,10 +79,10 @@ struct general_matrix_matrix_triangular_product pack_lhs; - gemm_pack_rhs pack_rhs; - gebp_kernel gebp; + + gemm_pack_lhs pack_lhs; + gemm_pack_rhs pack_rhs; + gebp_kernel gebp; tribb_kernel sybb; for(Index k2=0; k2 processed with gebp or skipped // 2 - the actual_mc x actual_mc symmetric block => processed with a special kernel // 3 - after the diagonal => processed with gebp or skipped if (UpLo==Lower) - gebp(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, (std::min)(size,i2), alpha, - -1, -1, 0, 0); + gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, + (std::min)(size,i2), alpha, -1, -1, 0, 0); - sybb(res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha); + + sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha); if (UpLo==Upper) { Index j2 = i2+actual_mc; - gebp(res+resStride*j2+i2, resStride, blockA, blockB+actual_kc*j2, actual_mc, actual_kc, (std::max)(Index(0), size-j2), alpha, - -1, -1, 0, 0); + gebp(res.getSubMapper(i2, j2), blockA, blockB+actual_kc*j2, actual_mc, + actual_kc, (std::max)(Index(0), size-j2), alpha, -1, -1, 0, 0); } } } @@ -129,13 +134,16 @@ struct tribb_kernel { typedef gebp_traits Traits; typedef typename Traits::ResScalar ResScalar; - + enum { BlockSize = EIGEN_PLAIN_ENUM_MAX(mr,nr) }; - void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha) + void operator()(ResScalar* _res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha) { - gebp_kernel gebp_kernel; + typedef blas_data_mapper ResMapper; + ResMapper res(_res, resStride); + gebp_kernel gebp_kernel; + Matrix buffer; // let's process the block per panel of actual_mc x BlockSize, @@ -146,7 +154,7 @@ struct tribb_kernel const RhsScalar* actual_b = blockB+j*depth; if(UpLo==Upper) - gebp_kernel(res+j*resStride, resStride, blockA, actual_b, j, depth, actualBlockSize, alpha, + gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha, -1, -1, 0, 0); // selfadjoint micro block @@ -154,12 +162,12 @@ struct tribb_kernel Index i = j; buffer.setZero(); // 1 - apply the kernel on the temporary buffer - gebp_kernel(buffer.data(), BlockSize, blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha, + gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha, -1, -1, 0, 0); // 2 - triangular accumulation for(Index j1=0; j1 lhs(_lhs,lhsStride); - const_blas_data_mapper rhs(_rhs,rhsStride); - typedef gebp_traits Traits; + typedef const_blas_data_mapper LhsMapper; + typedef const_blas_data_mapper LhsTransposeMapper; + typedef const_blas_data_mapper RhsMapper; + typedef blas_data_mapper ResMapper; + LhsMapper lhs(_lhs,lhsStride); + LhsTransposeMapper lhs_transpose(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); + Index kc = size; // cache block size along the K direction Index mc = rows; // cache block size along the M direction Index nc = cols; // cache block size along the N direction @@ -346,10 +352,10 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix gebp_kernel; + gebp_kernel gebp_kernel; symm_pack_lhs pack_lhs; - gemm_pack_rhs pack_rhs; - gemm_pack_lhs pack_lhs_transposed; + gemm_pack_rhs pack_rhs; + gemm_pack_lhs pack_lhs_transposed; for(Index k2=0; k2 transposed packed copy @@ -368,9 +374,9 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix() - (blockA, &lhs(i2, k2), lhsStride, actual_kc, actual_mc); + gemm_pack_lhs() + (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc); - gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha); + gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha); } } } @@ -414,15 +420,18 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix lhs(_lhs,lhsStride); - typedef gebp_traits Traits; + typedef const_blas_data_mapper LhsMapper; + typedef blas_data_mapper ResMapper; + LhsMapper lhs(_lhs,lhsStride); + ResMapper res(_res,resStride); + Index kc = size; // cache block size along the K direction Index mc = rows; // cache block size along the M direction Index nc = cols; // cache block size along the N direction @@ -432,8 +441,8 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix gebp_kernel; - gemm_pack_lhs pack_lhs; + gebp_kernel gebp_kernel; + gemm_pack_lhs pack_lhs; symm_pack_rhs pack_rhs; for(Index k2=0; k2& blocking) { // strip zeros @@ -117,8 +117,12 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix lhs(_lhs,lhsStride); - const_blas_data_mapper rhs(_rhs,rhsStride); + typedef const_blas_data_mapper LhsMapper; + typedef const_blas_data_mapper RhsMapper; + typedef blas_data_mapper ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); Index kc = blocking.kc(); // cache block size along the K direction Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction @@ -136,9 +140,9 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix gebp_kernel; - gemm_pack_lhs pack_lhs; - gemm_pack_rhs pack_rhs; + gebp_kernel gebp_kernel; + gemm_pack_lhs pack_lhs; + gemm_pack_rhs pack_rhs; for(Index k2=IsLower ? depth : 0; IsLower ? k2>0 : k2 skip it @@ -182,9 +186,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix() - (blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc); + gemm_pack_lhs() + (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc); - gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0); + gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, + actual_kc, cols, alpha, -1, -1, 0, 0); } } } @@ -247,7 +254,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix& blocking) { // strip zeros @@ -256,8 +263,12 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix lhs(_lhs,lhsStride); - const_blas_data_mapper rhs(_rhs,rhsStride); + typedef const_blas_data_mapper LhsMapper; + typedef const_blas_data_mapper RhsMapper; + typedef blas_data_mapper ResMapper; + LhsMapper lhs(_lhs,lhsStride); + RhsMapper rhs(_rhs,rhsStride); + ResMapper res(_res, resStride); Index kc = blocking.kc(); // cache block size along the K direction Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction @@ -275,10 +286,10 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix gebp_kernel; - gemm_pack_lhs pack_lhs; - gemm_pack_rhs pack_rhs; - gemm_pack_rhs pack_rhs_panel; + gebp_kernel gebp_kernel; + gemm_pack_lhs pack_lhs; + gemm_pack_rhs pack_rhs; + gemm_pack_rhs pack_rhs_panel; for(Index k2=IsLower ? 0 : depth; IsLower ? k20; @@ -302,7 +313,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix0) @@ -315,7 +326,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix0) @@ -349,7 +360,7 @@ EIGEN_DONT_INLINE void product_triangular_matrix_matrix& blocking) { Index cols = otherSize; - const_blas_data_mapper tri(_tri,triStride); - blas_data_mapper other(_other,otherStride); + + typedef const_blas_data_mapper TriMapper; + typedef blas_data_mapper OtherMapper; + TriMapper tri(_tri, triStride); + OtherMapper other(_other, otherStride); typedef gebp_traits Traits; + enum { SmallPanelWidth = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr), IsLower = (Mode&Lower) == Lower @@ -71,9 +75,9 @@ EIGEN_DONT_INLINE void triangular_solve_matrix conj; - gebp_kernel gebp_kernel; - gemm_pack_lhs pack_lhs; - gemm_pack_rhs pack_rhs; + gebp_kernel gebp_kernel; + gemm_pack_lhs pack_lhs; + gemm_pack_rhs pack_rhs; // the goal here is to subdivise the Rhs panels such that we keep some cache // coherence when accessing the rhs elements @@ -146,16 +150,16 @@ EIGEN_DONT_INLINE void triangular_solve_matrix0) { Index startTarget = IsLower ? k2+k1+actualPanelWidth : k2-actual_kc; - pack_lhs(blockA, &tri(startTarget,startBlock), triStride, actualPanelWidth, lengthTarget); + pack_lhs(blockA, tri.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget); - gebp_kernel(&other(startTarget,j2), otherStride, blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1), + gebp_kernel(other.getSubMapper(startTarget,j2), blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1), actualPanelWidth, actual_kc, 0, blockBOffset); } } @@ -170,9 +174,9 @@ EIGEN_DONT_INLINE void triangular_solve_matrix0) { - pack_lhs(blockA, &tri(i2, IsLower ? k2 : k2-kc), triStride, actual_kc, actual_mc); + pack_lhs(blockA, tri.getSubMapper(i2, IsLower ? k2 : k2-kc), actual_kc, actual_mc); - gebp_kernel(_other+i2, otherStride, blockA, blockB, actual_mc, actual_kc, cols, Scalar(-1), -1, -1, 0, 0); + gebp_kernel(other.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, Scalar(-1), -1, -1, 0, 0); } } } @@ -198,8 +202,11 @@ EIGEN_DONT_INLINE void triangular_solve_matrix& blocking) { Index rows = otherSize; - const_blas_data_mapper rhs(_tri,triStride); - blas_data_mapper lhs(_other,otherStride); + + typedef blas_data_mapper LhsMapper; + typedef const_blas_data_mapper RhsMapper; + LhsMapper lhs(_other, otherStride); + RhsMapper rhs(_tri, triStride); typedef gebp_traits Traits; enum { @@ -218,10 +225,10 @@ EIGEN_DONT_INLINE void triangular_solve_matrix conj; - gebp_kernel gebp_kernel; - gemm_pack_rhs pack_rhs; - gemm_pack_rhs pack_rhs_panel; - gemm_pack_lhs pack_lhs_panel; + gebp_kernel gebp_kernel; + gemm_pack_rhs pack_rhs; + gemm_pack_rhs pack_rhs_panel; + gemm_pack_lhs pack_lhs_panel; for(Index k2=IsLower ? size : 0; IsLower ? k2>0 : k20) pack_rhs(geb, &rhs(actual_k2,startPanel), triStride, actual_kc, rs); + if (rs>0) pack_rhs(geb, rhs.getSubMapper(actual_k2,startPanel), actual_kc, rs); // triangular packing (we only pack the panels off the diagonal, // neglecting the blocks overlapping the diagonal @@ -248,7 +255,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix0) pack_rhs_panel(blockB+j2*actual_kc, - &rhs(actual_k2+panelOffset, actual_j2), triStride, + rhs.getSubMapper(actual_k2+panelOffset, actual_j2), panelLength, actualPanelWidth, actual_kc, panelOffset); } @@ -276,7 +283,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix0) { - gebp_kernel(&lhs(i2,absolute_j2), otherStride, + gebp_kernel(lhs.getSubMapper(i2,absolute_j2), blockA, blockB+j2*actual_kc, actual_mc, panelLength, actualPanelWidth, Scalar(-1), @@ -303,14 +310,14 @@ EIGEN_DONT_INLINE void triangular_solve_matrix0) - gebp_kernel(_other+i2+startPanel*otherStride, otherStride, blockA, geb, + gebp_kernel(lhs.getSubMapper(i2, startPanel), blockA, geb, actual_mc, actual_kc, rs, Scalar(-1), -1, -1, 0, 0); } diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h index 0d8e2705a..25a62d528 100644 --- a/Eigen/src/Core/util/BlasUtil.h +++ b/Eigen/src/Core/util/BlasUtil.h @@ -18,13 +18,13 @@ namespace Eigen { namespace internal { // forward declarations -template +template struct gebp_kernel; -template +template struct gemm_pack_rhs; -template +template struct gemm_pack_lhs; template< @@ -117,32 +117,96 @@ template struct get_factor::R static EIGEN_STRONG_INLINE typename NumTraits::Real run(const Scalar& x) { return numext::real(x); } }; -// Lightweight helper class to access matrix coefficients. -// Yes, this is somehow redundant with Map<>, but this version is much much lighter, -// and so I hope better compilation performance (time and code quality). -template -class blas_data_mapper -{ + +template +class MatrixLinearMapper { public: - blas_data_mapper(Scalar* data, Index stride) : m_data(data), m_stride(stride) {} - EIGEN_STRONG_INLINE Scalar& operator()(Index i, Index j) - { return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride]; } + typedef typename packet_traits::type Packet; + typedef typename packet_traits::half HalfPacket; + + EIGEN_ALWAYS_INLINE MatrixLinearMapper(Scalar *data) : m_data(data) {} + + EIGEN_ALWAYS_INLINE void prefetch(int i) const { + internal::prefetch(&operator()(i)); + } + + EIGEN_ALWAYS_INLINE Scalar& operator()(Index i) const { + return m_data[i]; + } + + EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const { + return ploadt(m_data + i); + } + + EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const { + return ploadt(m_data + i); + } + + EIGEN_ALWAYS_INLINE void storePacket(Index i, Packet p) const { + pstoret(m_data + i, p); + } + protected: - Scalar* EIGEN_RESTRICT m_data; - Index m_stride; + Scalar *m_data; +}; + +// Lightweight helper class to access matrix coefficients. +template +class blas_data_mapper { + public: + typedef typename packet_traits::type Packet; + typedef typename packet_traits::half HalfPacket; + + typedef MatrixLinearMapper LinearMapper; + + EIGEN_ALWAYS_INLINE blas_data_mapper(Scalar* data, Index stride) : m_data(data), m_stride(stride) {} + + EIGEN_ALWAYS_INLINE blas_data_mapper + getSubMapper(Index i, Index j) const { + return blas_data_mapper(&operator()(i, j), m_stride); + } + + EIGEN_ALWAYS_INLINE LinearMapper getLinearMapper(Index i, Index j) const { + return LinearMapper(&operator()(i, j)); + } + + EIGEN_DEVICE_FUNC + EIGEN_ALWAYS_INLINE Scalar& operator()(Index i, Index j) const { + return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride]; + } + + EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const { + return ploadt(&operator()(i, j)); + } + + EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i, Index j) const { + return ploadt(&operator()(i, j)); + } + + template + EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, SubPacket p) const { + pscatter(&operator()(i, j), p, m_stride); + } + + template + EIGEN_ALWAYS_INLINE SubPacket gatherPacket(Index i, Index j) const { + return pgather(&operator()(i, j), m_stride); + } + + protected: + Scalar* EIGEN_RESTRICT m_data; + const Index m_stride; }; // lightweight helper class to access matrix coefficients (const version) template -class const_blas_data_mapper -{ +class const_blas_data_mapper : public blas_data_mapper { public: - const_blas_data_mapper(const Scalar* data, Index stride) : m_data(data), m_stride(stride) {} - EIGEN_STRONG_INLINE const Scalar& operator()(Index i, Index j) const - { return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride]; } - protected: - const Scalar* EIGEN_RESTRICT m_data; - Index m_stride; + EIGEN_ALWAYS_INLINE const_blas_data_mapper(const Scalar *data, Index stride) : blas_data_mapper(data, stride) {} + + EIGEN_ALWAYS_INLINE const_blas_data_mapper getSubMapper(Index i, Index j) const { + return const_blas_data_mapper(&(this->operator()(i, j)), this->m_stride); + } }; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 5d8913dd8..75423f516 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -48,7 +48,7 @@ if(MPFR_FOUND) include_directories(${MPFR_INCLUDES} ./mpreal) ei_add_property(EIGEN_TESTED_BACKENDS "MPFR C++, ") set(EIGEN_MPFR_TEST_LIBRARIES ${MPFR_LIBRARIES} ${GMP_LIBRARIES}) - ei_add_test(mpreal_support "" "${EIGEN_MPFR_TEST_LIBRARIES}" ) +# ei_add_test(mpreal_support "" "${EIGEN_MPFR_TEST_LIBRARIES}" ) else() ei_add_property(EIGEN_MISSING_BACKENDS "MPFR C++, ") endif() From 12693928228922ecf8fa3fcf14341d195e376a11 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 3 Oct 2014 10:16:59 -0700 Subject: [PATCH 0864/1103] Created the IndexPair type to store pair of tensor indices. CUDA doesn't support std::pair so we can't use them when targeting GPUs. Improved the performance on tensor contractions --- .../CXX11/src/Core/util/CXX11Workarounds.h | 4 +- .../CXX11/src/Tensor/TensorContraction.h | 741 +++++++++++++++--- .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 7 + 3 files changed, 662 insertions(+), 90 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index 3812ecd1f..227522ecb 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -69,11 +69,13 @@ template constexpr inline T const& array_ #undef STD_GET_ARR_HACK template struct array_size; +template struct array_size > { + static const size_t value = N; +}; template struct array_size > { static const size_t value = N; }; - /* Suppose you have a template of the form * template struct X; * And you want to specialize it in such a way: diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index 46624724c..1e6f276e0 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -20,6 +20,319 @@ namespace Eigen { * */ namespace internal { + +enum { + Rhs = 0, + Lhs = 1, +}; + +/* + * Implementation of the Eigen blas_data_mapper class for tensors. + */ +template +class BaseTensorContractionMapper { + public: + EIGEN_DEVICE_FUNC + BaseTensorContractionMapper(const Tensor& tensor, + const nocontract_t& nocontract_strides, + const nocontract_t& ij_strides, + const contract_t& contract_strides, + const contract_t& k_strides) : + m_tensor(tensor), + m_nocontract_strides(nocontract_strides), + m_ij_strides(ij_strides), + m_contract_strides(contract_strides), + m_k_strides(k_strides) { } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE void prefetch(int i) { } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar operator()(Index row) const { + // column major assumption + return operator()(row, 0); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar operator()(Index row, Index col) const { + return m_tensor.coeff(computeIndex(row, col)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index computeIndex(Index row, Index col) const { + const bool left = (side == Lhs); + Index nocontract_val = left ? row : col; + Index linidx = 0; + for (int i = array_size::value - 1; i > 0; i--) { + const Index idx = nocontract_val / m_ij_strides[i]; + linidx += idx * m_nocontract_strides[i]; + nocontract_val -= idx * m_ij_strides[i]; + } + if (array_size::value > array_size::value) { + if (side == Lhs && inner_dim_contiguous) { + eigen_assert(m_nocontract_strides[0] == 1); + linidx += nocontract_val; + } else { + linidx += nocontract_val * m_nocontract_strides[0]; + } + } + + Index contract_val = left ? col : row; + for (int i = array_size::value - 1; i > 0; i--) { + const Index idx = contract_val / m_k_strides[i]; + linidx += idx * m_contract_strides[i]; + contract_val -= idx * m_k_strides[i]; + } + EIGEN_STATIC_ASSERT(array_size::value > 0, YOU_MADE_A_PROGRAMMING_MISTAKE); + if (side == Rhs && inner_dim_contiguous) { + eigen_assert(m_contract_strides[0] == 1); + linidx += contract_val; + } else { + linidx += contract_val * m_contract_strides[0]; + } + + return linidx; + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE IndexPair computeIndexPair(Index row, Index col, const Index distance) const { + const bool left = (side == Lhs); + Index nocontract_val[2] = {left ? row : col, left ? row + distance : col}; + Index linidx[2] = {0, 0}; + for (int i = array_size::value - 1; i > 0; i--) { + const Index idx0 = nocontract_val[0] / m_ij_strides[i]; + const Index idx1 = nocontract_val[1] / m_ij_strides[i]; + linidx[0] += idx0 * m_nocontract_strides[i]; + linidx[1] += idx1 * m_nocontract_strides[i]; + nocontract_val[0] -= idx0 * m_ij_strides[i]; + nocontract_val[1] -= idx1 * m_ij_strides[i]; + } + if (array_size::value > array_size::value) { + if (side == Lhs && inner_dim_contiguous) { + eigen_assert(m_nocontract_strides[0] == 1); + linidx[0] += nocontract_val[0]; + linidx[1] += nocontract_val[1]; + } else { + linidx[0] += nocontract_val[0] * m_nocontract_strides[0]; + linidx[1] += nocontract_val[1] * m_nocontract_strides[0]; + } + } + + Index contract_val[2] = {left ? col : row, left ? col : row + distance}; + for (int i = array_size::value - 1; i > 0; i--) { + const Index idx0 = contract_val[0] / m_k_strides[i]; + const Index idx1 = contract_val[1] / m_k_strides[i]; + linidx[0] += idx0 * m_contract_strides[i]; + linidx[1] += idx1 * m_contract_strides[i]; + contract_val[0] -= idx0 * m_k_strides[i]; + contract_val[1] -= idx1 * m_k_strides[i]; + } + EIGEN_STATIC_ASSERT(array_size::value > 0, YOU_MADE_A_PROGRAMMING_MISTAKE); + if (side == Rhs && inner_dim_contiguous) { + eigen_assert(m_contract_strides[0] == 1); + linidx[0] += contract_val[0]; + linidx[1] += contract_val[1]; + } else { + linidx[0] += contract_val[0] * m_contract_strides[0]; + linidx[1] += contract_val[1] * m_contract_strides[0]; + } + return IndexPair(linidx[0], linidx[1]); + } + + protected: + const Tensor m_tensor; + const nocontract_t m_nocontract_strides; + const nocontract_t m_ij_strides; + const contract_t m_contract_strides; + const contract_t m_k_strides; +}; + + + +template +class TensorContractionInputMapper; + +template +class TensorContractionSubMapper { + public: + typedef typename packet_traits::type Packet; + typedef typename packet_traits::half HalfPacket; + + typedef TensorContractionInputMapper ParentMapper; + typedef TensorContractionSubMapper Self; + typedef Self LinearMapper; + + EIGEN_DEVICE_FUNC TensorContractionSubMapper(const ParentMapper& base_mapper, Index vert_offset, Index horiz_offset) + : m_base_mapper(base_mapper), m_vert_offset(vert_offset), m_horiz_offset(horiz_offset) { } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar operator()(Index i) const { + return m_base_mapper(i + m_vert_offset, m_horiz_offset); + } + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar operator()(Index i, Index j) const { + return m_base_mapper(i + m_vert_offset, j + m_horiz_offset); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const { + return m_base_mapper.loadPacket(i + m_vert_offset, m_horiz_offset); + } + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const { + return m_base_mapper.loadPacket(i + m_vert_offset, j + m_horiz_offset); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const { + return m_base_mapper.loadHalfPacket(i + m_vert_offset, m_horiz_offset); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, Packet p) const { + m_base_mapper.storePacket(i + m_vert_offset, m_horiz_offset, p); + } + + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE LinearMapper getLinearMapper(Index i, Index j) const { + return LinearMapper(m_base_mapper, i + m_vert_offset, j + m_horiz_offset); + } + + private: + const ParentMapper& m_base_mapper; + const Index m_vert_offset; + const Index m_horiz_offset; +}; + + +template::size : 1), + bool inner_dim_contiguous = false, bool inner_dim_reordered = (side != Lhs), int Alignment=Unaligned> +class TensorContractionInputMapper + : public BaseTensorContractionMapper { + + public: + typedef BaseTensorContractionMapper Base; + typedef TensorContractionSubMapper SubMapper; + + TensorContractionInputMapper(const Tensor& tensor, + const nocontract_t& nocontract_strides, + const nocontract_t& ij_strides, + const contract_t& contract_strides, + const contract_t& k_strides) + : Base(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE SubMapper getSubMapper(Index i, Index j) const { + return SubMapper(*this, i, j); + } + + typedef typename packet_traits::type Packet; + typedef typename packet_traits::half HalfPacket; + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Packet loadPacket(Index i, Index j) const { + // whole method makes column major assumption + + // don't need to add offsets for now (because operator handles that) + // current code assumes packet size must be a multiple of 2 + EIGEN_STATIC_ASSERT(packet_size % 2 == 0, YOU_MADE_A_PROGRAMMING_MISTAKE); + + if (Tensor::PacketAccess && inner_dim_contiguous && !inner_dim_reordered) { + const Index index = this->computeIndex(i, j); + eigen_assert(this->computeIndex(i+packet_size-1, j) == index + packet_size-1); + return this->m_tensor.template packet(index); + } + + const IndexPair indexPair = this->computeIndexPair(i, j, packet_size - 1); + const Index first = indexPair.first; + const Index last = indexPair.second; + + // We can always do optimized packet reads from left hand side right now, because + // the vertical matrix dimension on the left hand side is never contracting. + // On the right hand side we need to check if the contracting dimensions may have + // been shuffled first. + if (Tensor::PacketAccess && + (side == Lhs || internal::array_size::value <= 1 || !inner_dim_reordered) && + (last - first) == (packet_size - 1)) { + + return this->m_tensor.template packet(first); + } + + EIGEN_ALIGN_DEFAULT Scalar data[packet_size]; + + data[0] = this->m_tensor.coeff(first); + for (Index k = 1; k < packet_size - 1; k += 2) { + const IndexPair internal_pair = this->computeIndexPair(i + k, j, 1); + data[k] = this->m_tensor.coeff(internal_pair.first); + data[k + 1] = this->m_tensor.coeff(internal_pair.second); + } + data[packet_size - 1] = this->m_tensor.coeff(last); + + return pload(data); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE HalfPacket loadHalfPacket(Index i, Index j) const { + // whole method makes column major assumption + + // don't need to add offsets for now (because operator handles that) + const Index half_packet_size = unpacket_traits::size; + if (half_packet_size == packet_size) { + return loadPacket(i, j); + } + EIGEN_ALIGN_DEFAULT Scalar data[half_packet_size]; + for (Index k = 0; k < half_packet_size; k++) { + data[k] = operator()(i + k, j); + } + return pload(data); + } +}; + + +template +class TensorContractionInputMapper + : public BaseTensorContractionMapper { + + public: + typedef BaseTensorContractionMapper Base; + typedef TensorContractionSubMapper SubMapper; + + TensorContractionInputMapper(const Tensor& tensor, + const nocontract_t& nocontract_strides, + const nocontract_t& ij_strides, + const contract_t& contract_strides, + const contract_t& k_strides) + : Base(tensor, nocontract_strides, ij_strides, contract_strides, k_strides) { } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE SubMapper getSubMapper(Index i, Index j) const { + return SubMapper(*this, i, j); + } + + typedef typename packet_traits::type Packet; + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Packet loadPacket(Index i, Index j) const { + EIGEN_ALIGN_DEFAULT Scalar data[1]; + data[0] = this->m_tensor.coeff(this->computeIndex(i, j)); + return pload::type>(data); + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Packet loadHalfPacket(Index i, Index j) const { + return loadPacket(i, j); + } +}; + + template struct traits > { @@ -53,6 +366,14 @@ struct nested, 1, typena typedef TensorContractionOp type; }; +template +struct traits, Device_> > { + typedef Indices_ Indices; + typedef LeftArgType_ LeftArgType; + typedef RightArgType_ RightArgType; + typedef Device_ Device; +}; + } // end namespace internal @@ -102,143 +423,385 @@ template <> struct max_n_1<0> { }; -template -struct TensorEvaluator, Device> +template +struct TensorContractionEvaluatorBase { + typedef typename internal::traits::Indices Indices; + typedef typename internal::traits::LeftArgType LeftArgType; + typedef typename internal::traits::RightArgType RightArgType; + typedef typename internal::traits::Device Device; + typedef TensorContractionOp XprType; + typedef typename internal::remove_const::type Scalar; + typedef typename XprType::Packet Packet; + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + typedef array::Dimensions::count> left_dim_mapper_t; + typedef array::Dimensions::count> right_dim_mapper_t; + + typedef array::value> contract_t; + typedef array::Dimensions::count - internal::array_size::value>::size> left_nocontract_t; + typedef array::Dimensions::count - internal::array_size::value>::size> right_nocontract_t; static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * internal::array_size::value>::size; - typedef typename XprType::Index Index; + typedef DSizes Dimensions; enum { - IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, - PacketAccess = /*TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess */ - false, + IsAligned = true, + PacketAccess = (internal::packet_traits::size > 1), }; - TensorEvaluator(const XprType& op, const Device& device) - : m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionEvaluatorBase(const XprType& op, const Device& device) + : m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device), m_device(device), m_result(NULL) { - Index index = 0; - Index stride = 1; - m_shiftright = 1; + eigen_assert((internal::array_size::value > 0) && "Must contract on some indices"); - int skipped = 0; + array::Dimensions::count> lhs_strides; + lhs_strides[0] = 1; + for (int i = 0; i < TensorEvaluator::Dimensions::count-1; ++i) { + lhs_strides[i+1] = lhs_strides[i] * m_leftImpl.dimensions()[i]; + } + + array::Dimensions::count> rhs_strides; + rhs_strides[0] = 1; + for (int i = 0; i < TensorEvaluator::Dimensions::count-1; ++i) { + rhs_strides[i+1] = rhs_strides[i] * m_rightImpl.dimensions()[i]; + } + + m_i_strides[0] = 1; + m_j_strides[0] = 1; + m_k_strides[0] = 1; + + m_i_size = 1; + m_j_size = 1; + m_k_size = 1; + + // To compute the dimension, we simply concatenate the non-contracting + // dimensions of the left and then the right tensor. Additionally, we also + // compute the strides corresponding to the left non-contracting + // dimensions and right non-contracting dimensions. + m_lhs_inner_dim_contiguous = true; + int dim_idx = 0; + int nocontract_idx = 0; const typename TensorEvaluator::Dimensions& left_dims = m_leftImpl.dimensions(); - for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { - bool skip = false; - for (int j = 0; j < internal::array_size::value; ++j) { + for (int i = 0; i < TensorEvaluator::Dimensions::count; i++) { + // find if we are contracting on index i of left tensor + bool contracting = false; + for (int j = 0; j < internal::array_size::value; j++) { if (op.indices()[j].first == i) { - skip = true; - m_leftOffsets[2*skipped] = stride; - m_leftOffsets[2*skipped+1] = stride * left_dims[i]; - m_stitchsize[skipped] = left_dims[i]; + contracting = true; break; } } - if (!skip) { - m_dimensions[index++] = left_dims[i]; - m_shiftright *= left_dims[i]; - } else { - ++skipped; + if (!contracting) { + // add dimension size to output dimensions + m_dimensions[dim_idx] = left_dims[i]; + m_left_nocontract_strides[nocontract_idx] = lhs_strides[i]; + if (dim_idx != i) { + m_lhs_inner_dim_contiguous = false; + } + if (nocontract_idx+1 < internal::array_size::value) { + m_i_strides[nocontract_idx+1] = m_i_strides[nocontract_idx] * left_dims[i]; + } else { + m_i_size = m_i_strides[nocontract_idx] * left_dims[i]; + } + dim_idx++; + nocontract_idx++; } - stride *= left_dims[i]; } - stride = 1; - skipped = 0; + nocontract_idx = 0; const typename TensorEvaluator::Dimensions& right_dims = m_rightImpl.dimensions(); - for (int i = 0; i < TensorEvaluator::Dimensions::count; ++i) { - bool skip = false; - for (int j = 0; j < internal::array_size::value; ++j) { + for (int i = 0; i < TensorEvaluator::Dimensions::count; i++) { + bool contracting = false; + // find if we are contracting on index i of right tensor + for (int j = 0; j < internal::array_size::value; j++) { if (op.indices()[j].second == i) { - skip = true; - m_rightOffsets[2*skipped] = stride; - m_rightOffsets[2*skipped+1] = stride * right_dims[i]; + contracting = true; break; } } - if (!skip) { - m_dimensions[index++] = right_dims[i]; - } else { - ++skipped; + if (!contracting) { + m_dimensions[dim_idx] = right_dims[i]; + if (nocontract_idx+1 < internal::array_size::value) { + m_j_strides[nocontract_idx+1] = m_j_strides[nocontract_idx] * right_dims[i]; + } else { + m_j_size = m_j_strides[nocontract_idx] * right_dims[i]; + } + m_right_nocontract_strides[nocontract_idx] = rhs_strides[i]; + dim_idx++; + nocontract_idx++; } - stride *= right_dims[i]; } - // Scalar case + // Now compute the strides corresponding to the contracting dimensions. We + // assumed above that non-contracting axes are represented in the same order + // in the matrix as they are in the tensor. This is not the case for + // contracting axes. As the contracting axes must be of the same size in + // each tensor, we'll only look at the first tensor here. + m_rhs_inner_dim_contiguous = true; + m_rhs_inner_dim_reordered = false; + for (int i = 0; i < internal::array_size::value; i++) { + Index left = op.indices()[i].first; + Index right = op.indices()[i].second; + + Index size = left_dims[left]; + eigen_assert(size == right_dims[right] && "Contraction axes must be same size"); + + if (i+1 < internal::array_size::value) { + m_k_strides[i+1] = m_k_strides[i] * size; + } else { + m_k_size = m_k_strides[i] * size; + } + m_left_contracting_strides[i] = lhs_strides[left]; + m_right_contracting_strides[i] = rhs_strides[right]; + + if (i > 0 && right < op.indices()[i-1].second) { + m_rhs_inner_dim_reordered = true; + } + if (right != i) { + m_rhs_inner_dim_contiguous = false; + } + } + + // Scalar case. We represent the result as a 1d tensor of size 1. if (TensorEvaluator::Dimensions::count + TensorEvaluator::Dimensions::count == 2 * internal::array_size::value) { m_dimensions[0] = 1; } } - typedef typename XprType::Scalar Scalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - const Dimensions& dimensions() const { return m_dimensions; } - - void evalTo(typename XprType::Scalar* buffer) const { - for (int i = 0; i < dimensions().TotalSize(); ++i) { - buffer[i] += coeff(i); - } - } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { m_leftImpl.evalSubExprsIfNeeded(NULL); m_rightImpl.evalSubExprsIfNeeded(NULL); - return true; - } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { - m_leftImpl.cleanup(); - m_rightImpl.cleanup(); + if (data) { + evalTo(data); + return false; + } else { + m_result = static_cast(m_device.allocate(dimensions().TotalSize() * sizeof(Scalar))); + evalTo(m_result); + return true; + } } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const - { - const Index startLeft = index % m_shiftright; - const Index startRight = index / m_shiftright; - CoeffReturnType result = CoeffReturnType(0); - partialStitch(startLeft, startRight, 0, result); - return result; - } - - /* TODO: vectorization - template - EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const - { - assert(false); - }*/ - - private: - EIGEN_DEVICE_FUNC void partialStitch(Index startLeft, Index startRight, int StitchIndex, CoeffReturnType& accum) const { - Index firstLeft = (startLeft / m_leftOffsets[2*StitchIndex]) * m_leftOffsets[2*StitchIndex+1] + (startLeft % m_leftOffsets[2*StitchIndex]); - Index firstRight = (startRight / m_rightOffsets[2*StitchIndex]) * m_rightOffsets[2*StitchIndex+1] + (startRight % m_rightOffsets[2*StitchIndex]); - - for (int j = 0; j < m_stitchsize[StitchIndex]; ++j) { - const Index left = firstLeft+j*m_leftOffsets[2*StitchIndex]; - const Index right = firstRight+j*m_rightOffsets[2*StitchIndex]; - if (StitchIndex < internal::array_size::value-1) { - partialStitch(left, right, StitchIndex+1, accum); - } else { - accum += m_leftImpl.coeff(left) * m_rightImpl.coeff(right); + EIGEN_DEVICE_FUNC void evalTo(Scalar* buffer) const { + if (this->m_lhs_inner_dim_contiguous) { + if (this->m_rhs_inner_dim_contiguous) { + if (this->m_rhs_inner_dim_reordered) { + static_cast(this)->template evalTyped(buffer); + } + else { + static_cast(this)->template evalTyped(buffer); + } + } + else { + if (this->m_rhs_inner_dim_reordered) { + static_cast(this)->template evalTyped(buffer); + } + else { + static_cast(this)->template evalTyped(buffer); + } + } + } + else { + if (this->m_rhs_inner_dim_contiguous) { + if (this->m_rhs_inner_dim_reordered) { + static_cast(this)->template evalTyped(buffer); + } + else { + static_cast(this)->template evalTyped(buffer); + } + } + else { + if (this->m_rhs_inner_dim_reordered) { + static_cast(this)->template evalTyped(buffer); + } + else { + static_cast(this)->template evalTyped(buffer); + } } } } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_leftImpl.cleanup(); + m_rightImpl.cleanup(); + + if (m_result != NULL) { + m_device.deallocate(m_result); + m_result = NULL; + } + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { + return m_result[index]; + } + + template + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const { + return internal::ploadt(m_result + index); + } + + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } + + protected: + // Prevent assignment + TensorContractionEvaluatorBase& operator = (const TensorContractionEvaluatorBase&); - private: - array::value> m_leftOffsets; - array::value> m_rightOffsets; - array::value> m_stitchsize; - Index m_shiftright; Dimensions m_dimensions; + + contract_t m_k_strides; + contract_t m_left_contracting_strides; + contract_t m_right_contracting_strides; + + bool m_lhs_inner_dim_contiguous; + bool m_rhs_inner_dim_contiguous; + bool m_rhs_inner_dim_reordered; + + left_nocontract_t m_i_strides; + right_nocontract_t m_j_strides; + left_nocontract_t m_left_nocontract_strides; + right_nocontract_t m_right_nocontract_strides; + + Index m_i_size; + Index m_j_size; + Index m_k_size; + + const Device& m_device; + Scalar* m_result; TensorEvaluator m_leftImpl; TensorEvaluator m_rightImpl; }; +template +struct TensorEvaluator, Device> : + public TensorContractionEvaluatorBase, Device> > { + typedef TensorEvaluator, Device> Self; + typedef TensorContractionEvaluatorBase Base; + + typedef TensorContractionOp XprType; + typedef typename internal::remove_const::type Scalar; + typedef typename XprType::Packet Packet; + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + typedef array::Dimensions::count> left_dim_mapper_t; + typedef array::Dimensions::count> right_dim_mapper_t; + + typedef array::value> contract_t; + typedef array::Dimensions::count - internal::array_size::value>::size> left_nocontract_t; + typedef array::Dimensions::count - internal::array_size::value>::size> right_nocontract_t; + + static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * internal::array_size::value>::size; + + typedef DSizes Dimensions; + + + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) : + Base(op, device) { } + + template + EIGEN_DEVICE_FUNC void evalTyped(Scalar* buffer) const { + // columns in left side, rows in right side + const Index k = this->m_k_size; + + // rows in left side + const Index m = this->m_i_size; + + // columns in right side + const Index n = this->m_j_size; + + // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar) + this->m_device.memset(buffer, 0, m * n * sizeof(Scalar)); + + // define mr, nr, and all of my data mapper types + typedef typename internal::remove_const::type LhsScalar; + typedef typename internal::remove_const::type RhsScalar; + typedef typename internal::gebp_traits Traits; + + const Index nr = Traits::nr; + const Index mr = Traits::mr; + + typedef TensorEvaluator LeftEvaluator; + typedef TensorEvaluator RightEvaluator; + + const int lhs_packet_size = internal::packet_traits::size; + const int rhs_packet_size = internal::packet_traits::size; + + typedef internal::TensorContractionInputMapper LhsMapper; + + typedef internal::TensorContractionInputMapper RhsMapper; + + typedef internal::blas_data_mapper OutputMapper; + + + // Declare GEBP packing and kernel structs + internal::gemm_pack_lhs pack_lhs; + internal::gemm_pack_rhs pack_rhs; + internal::gebp_kernel gebp; + + // initialize data mappers + LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides, + this->m_left_contracting_strides, this->m_k_strides); + + RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides, + this->m_right_contracting_strides, this->m_k_strides); + + OutputMapper output(buffer, m); + + typedef typename internal::gemm_blocking_space BlockingType; + + // Sizes of the blocks to load in cache. See the Goto paper for details. + BlockingType blocking(m, n, k, true); + const Index kc = blocking.kc(); + const Index mc = (std::min)(m, blocking.mc()); + const Index nc = (std::min)(n, blocking.nc()); + int sizeA = mc * kc; + int sizeB = kc * nc; + + LhsScalar* blockA = static_cast(this->m_device.allocate(sizeA * sizeof(LhsScalar))); + RhsScalar* blockB = static_cast(this->m_device.allocate(sizeB * sizeof(RhsScalar))); + + for(Index i2=0; i2m_device.deallocate(blockA); + this->m_device.deallocate(blockB); + } +}; + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 5a113dc19..11590b474 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -29,6 +29,13 @@ namespace Eigen { * \sa Tensor */ +// Can't use std::pairs on cuda devices +template struct IndexPair { + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair() : first(0), second(0) { } + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair(Index f, Index s) : first(f), second(s) { } + Index first; + Index second; +}; // Boiler plate code From af2e5995e2ba48384024bbc8432bd6dbbebf71d2 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 3 Oct 2014 19:18:07 -0700 Subject: [PATCH 0865/1103] Improved support for CUDA devices. Improved contractions on GPU --- unsupported/Eigen/CXX11/Tensor | 1 + .../CXX11/src/Tensor/TensorContractionCuda.h | 1206 +++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 38 +- 3 files changed, 1237 insertions(+), 8 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 11161a547..b1bd2f676 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -44,6 +44,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h new file mode 100644 index 000000000..babe33fff --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h @@ -0,0 +1,1206 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H +#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H + +#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) + +namespace Eigen { + +template +__device__ EIGEN_STRONG_INLINE void +EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs, + const OutputMapper output, volatile Scalar* lhs_shmem, volatile Scalar* rhs_shmem, + const Index m_size, const Index n_size, const Index k_size) { + + const Index m_block_idx = blockIdx.x; + const Index n_block_idx = blockIdx.y; + + const Index base_m = 64 * m_block_idx; + const Index base_n = 64 * n_block_idx; + + // declare and initialize 64 registers for output 8x8 block + + // prefetch registers + Scalar lhs_pf0; + Scalar lhs_pf1; + Scalar lhs_pf2; + Scalar lhs_pf3; + Scalar lhs_pf4; + Scalar lhs_pf5; + Scalar lhs_pf6; + Scalar lhs_pf7; + + Scalar rhs_pf0; + Scalar rhs_pf1; + Scalar rhs_pf2; + Scalar rhs_pf3; + Scalar rhs_pf4; + Scalar rhs_pf5; + Scalar rhs_pf6; + Scalar rhs_pf7; + + // shared memory is formatted + // (contract idx in block, nocontract idx in block, block idx) + // where block idx is column major. This transposition limits the number of + // bank conflicts when reading the LHS. The core idea is that since the contracting + // index is shared by both sides, then the contracting index should be in threadIdx.x. + + // On the LHS, we pad each row inside of each block with an extra element. This makes + // each block 8 rows of 9 elements, which is 72 elements. This gives no bank conflicts + // on writes and very few 2-way conflicts on reads. There is an 8x8 grid of these blocks. + + // On the RHS we just add 8 padding elements to the end of each block. This gives no bank + // conflicts on writes and also none on reads. + + // storage indices + const Index lhs_store_idx_base = threadIdx.y * 72 + threadIdx.x * 9 + threadIdx.z; + const Index rhs_store_idx_base = threadIdx.y * 72 + threadIdx.z * 8 + threadIdx.x; + + const Index lhs_store_idx_0 = lhs_store_idx_base + 576 * 0; + const Index lhs_store_idx_1 = lhs_store_idx_base + 576 * 1; + const Index lhs_store_idx_2 = lhs_store_idx_base + 576 * 2; + const Index lhs_store_idx_3 = lhs_store_idx_base + 576 * 3; + const Index lhs_store_idx_4 = lhs_store_idx_base + 576 * 4; + const Index lhs_store_idx_5 = lhs_store_idx_base + 576 * 5; + const Index lhs_store_idx_6 = lhs_store_idx_base + 576 * 6; + const Index lhs_store_idx_7 = lhs_store_idx_base + 576 * 7; + + const Index rhs_store_idx_0 = rhs_store_idx_base + 576 * 0; + const Index rhs_store_idx_1 = rhs_store_idx_base + 576 * 1; + const Index rhs_store_idx_2 = rhs_store_idx_base + 576 * 2; + const Index rhs_store_idx_3 = rhs_store_idx_base + 576 * 3; + const Index rhs_store_idx_4 = rhs_store_idx_base + 576 * 4; + const Index rhs_store_idx_5 = rhs_store_idx_base + 576 * 5; + const Index rhs_store_idx_6 = rhs_store_idx_base + 576 * 6; + const Index rhs_store_idx_7 = rhs_store_idx_base + 576 * 7; + + // in the loading code, the following variables are important: + // threadIdx.x: the vertical position in an 8x8 block + // threadIdx.y: the vertical index of the 8x8 block in the grid + // threadIdx.z: the horizontal position in an 8x8 block + // k: the horizontal index of the 8x8 block in the grid + // + // The k parameter is implicit (it was the loop counter for a loop that went + // from 0 to <8, but now that loop is unrolled in the below code. + + const Index load_idx_vert = threadIdx.x + 8 * threadIdx.y; + const Index lhs_vert = base_m + load_idx_vert; + +#define prefetchIntoRegisters(base_k) \ + { \ + lhs_pf0 = Scalar(0); \ + lhs_pf1 = Scalar(0); \ + lhs_pf2 = Scalar(0); \ + lhs_pf3 = Scalar(0); \ + lhs_pf4 = Scalar(0); \ + lhs_pf5 = Scalar(0); \ + lhs_pf6 = Scalar(0); \ + lhs_pf7 = Scalar(0); \ + \ + rhs_pf0 = Scalar(0); \ + rhs_pf1 = Scalar(0); \ + rhs_pf2 = Scalar(0); \ + rhs_pf3 = Scalar(0); \ + rhs_pf4 = Scalar(0); \ + rhs_pf5 = Scalar(0); \ + rhs_pf6 = Scalar(0); \ + rhs_pf7 = Scalar(0); \ + \ + if (!needs_edge_check || lhs_vert < m_size) { \ + const Index lhs_horiz_0 = base_k + threadIdx.z + 0 * 8; \ + const Index lhs_horiz_1 = base_k + threadIdx.z + 1 * 8; \ + const Index lhs_horiz_2 = base_k + threadIdx.z + 2 * 8; \ + const Index lhs_horiz_3 = base_k + threadIdx.z + 3 * 8; \ + const Index lhs_horiz_4 = base_k + threadIdx.z + 4 * 8; \ + const Index lhs_horiz_5 = base_k + threadIdx.z + 5 * 8; \ + const Index lhs_horiz_6 = base_k + threadIdx.z + 6 * 8; \ + const Index lhs_horiz_7 = base_k + threadIdx.z + 7 * 8; \ + \ + if (!needs_edge_check || lhs_horiz_7 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \ + lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \ + lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \ + lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \ + lhs_pf5 = lhs(lhs_vert, lhs_horiz_5); \ + lhs_pf6 = lhs(lhs_vert, lhs_horiz_6); \ + lhs_pf7 = lhs(lhs_vert, lhs_horiz_7); \ + } else if (lhs_horiz_6 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \ + lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \ + lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \ + lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \ + lhs_pf5 = lhs(lhs_vert, lhs_horiz_5); \ + lhs_pf6 = lhs(lhs_vert, lhs_horiz_6); \ + } else if (lhs_horiz_5 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \ + lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \ + lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \ + lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \ + lhs_pf5 = lhs(lhs_vert, lhs_horiz_5); \ + } else if (lhs_horiz_4 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \ + lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \ + lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \ + lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \ + } else if (lhs_horiz_3 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \ + lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \ + lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \ + } else if (lhs_horiz_2 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \ + lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \ + } else if (lhs_horiz_1 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \ + } else if (lhs_horiz_0 < k_size) { \ + lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \ + } \ + } \ + \ + const Index rhs_vert = base_k + load_idx_vert; \ + if (!needs_edge_check || rhs_vert < k_size) { \ + const Index rhs_horiz_0 = base_n + threadIdx.z + 0 * 8; \ + const Index rhs_horiz_1 = base_n + threadIdx.z + 1 * 8; \ + const Index rhs_horiz_2 = base_n + threadIdx.z + 2 * 8; \ + const Index rhs_horiz_3 = base_n + threadIdx.z + 3 * 8; \ + const Index rhs_horiz_4 = base_n + threadIdx.z + 4 * 8; \ + const Index rhs_horiz_5 = base_n + threadIdx.z + 5 * 8; \ + const Index rhs_horiz_6 = base_n + threadIdx.z + 6 * 8; \ + const Index rhs_horiz_7 = base_n + threadIdx.z + 7 * 8; \ + \ + if (rhs_horiz_7 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \ + rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \ + rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \ + rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \ + rhs_pf5 = rhs(rhs_vert, rhs_horiz_5); \ + rhs_pf6 = rhs(rhs_vert, rhs_horiz_6); \ + rhs_pf7 = rhs(rhs_vert, rhs_horiz_7); \ + } else if (rhs_horiz_6 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \ + rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \ + rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \ + rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \ + rhs_pf5 = rhs(rhs_vert, rhs_horiz_5); \ + rhs_pf6 = rhs(rhs_vert, rhs_horiz_6); \ + } else if (rhs_horiz_5 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \ + rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \ + rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \ + rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \ + rhs_pf5 = rhs(rhs_vert, rhs_horiz_5); \ + } else if (rhs_horiz_4 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \ + rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \ + rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \ + rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \ + } else if (rhs_horiz_3 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \ + rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \ + rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \ + } else if (rhs_horiz_2 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \ + rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \ + } else if (rhs_horiz_1 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \ + } else if (rhs_horiz_0 < n_size) { \ + rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \ + } \ + } \ + } \ + +#define writeRegToShmem(_) \ + lhs_shmem[lhs_store_idx_0] = lhs_pf0; \ + rhs_shmem[rhs_store_idx_0] = rhs_pf0; \ + \ + lhs_shmem[lhs_store_idx_1] = lhs_pf1; \ + rhs_shmem[rhs_store_idx_1] = rhs_pf1; \ + \ + lhs_shmem[lhs_store_idx_2] = lhs_pf2; \ + rhs_shmem[rhs_store_idx_2] = rhs_pf2; \ + \ + lhs_shmem[lhs_store_idx_3] = lhs_pf3; \ + rhs_shmem[rhs_store_idx_3] = rhs_pf3; \ + \ + lhs_shmem[lhs_store_idx_4] = lhs_pf4; \ + rhs_shmem[rhs_store_idx_4] = rhs_pf4; \ + \ + lhs_shmem[lhs_store_idx_5] = lhs_pf5; \ + rhs_shmem[rhs_store_idx_5] = rhs_pf5; \ + \ + lhs_shmem[lhs_store_idx_6] = lhs_pf6; \ + rhs_shmem[rhs_store_idx_6] = rhs_pf6; \ + \ + lhs_shmem[lhs_store_idx_7] = lhs_pf7; \ + rhs_shmem[rhs_store_idx_7] = rhs_pf7; \ + + // declare and initialize result array +#define res(i, j) _res_##i##j +#define initResultRow(i) \ + Scalar res(i, 0) = Scalar(0); \ + Scalar res(i, 1) = Scalar(0); \ + Scalar res(i, 2) = Scalar(0); \ + Scalar res(i, 3) = Scalar(0); \ + Scalar res(i, 4) = Scalar(0); \ + Scalar res(i, 5) = Scalar(0); \ + Scalar res(i, 6) = Scalar(0); \ + Scalar res(i, 7) = Scalar(0); \ + + initResultRow(0); + initResultRow(1); + initResultRow(2); + initResultRow(3); + initResultRow(4); + initResultRow(5); + initResultRow(6); + initResultRow(7); +#undef initResultRow + + for (Index base_k = 0; base_k < k_size; base_k += 64) { + // wait for previous iteration to finish with shmem. Despite common sense, + // the code is a bit faster with this here then at bottom of loop + __syncthreads(); + + prefetchIntoRegisters(base_k); + writeRegToShmem(); + + #undef prefetchIntoRegisters + #undef writeRegToShmem + + // wait for shared mem packing to be done before starting computation + __syncthreads(); + + // compute 8x8 matrix product by outer product. This involves packing one column + // of LHS and one row of RHS into registers (takes 16 registers). + +#define lcol(i) _lcol##i + Scalar lcol(0); + Scalar lcol(1); + Scalar lcol(2); + Scalar lcol(3); + Scalar lcol(4); + Scalar lcol(5); + Scalar lcol(6); + Scalar lcol(7); + +#define rrow(j) _rrow##j + Scalar rrow(0); + Scalar rrow(1); + Scalar rrow(2); + Scalar rrow(3); + Scalar rrow(4); + Scalar rrow(5); + Scalar rrow(6); + Scalar rrow(7); + + // Now x corresponds to k, y to m, and z to n + const volatile Scalar* lhs_block = &lhs_shmem[threadIdx.x + 9 * threadIdx.y]; + const volatile Scalar* rhs_block = &rhs_shmem[threadIdx.x + 8 * threadIdx.z]; + +#define lhs_element(i, j) lhs_block[72 * ((i) + 8 * (j))] +#define rhs_element(i, j) rhs_block[72 * ((i) + 8 * (j))] + +#define loadData(i, j) \ + lcol(0) = lhs_element(0, j); \ + rrow(0) = rhs_element(i, 0); \ + lcol(1) = lhs_element(1, j); \ + rrow(1) = rhs_element(i, 1); \ + lcol(2) = lhs_element(2, j); \ + rrow(2) = rhs_element(i, 2); \ + lcol(3) = lhs_element(3, j); \ + rrow(3) = rhs_element(i, 3); \ + lcol(4) = lhs_element(4, j); \ + rrow(4) = rhs_element(i, 4); \ + lcol(5) = lhs_element(5, j); \ + rrow(5) = rhs_element(i, 5); \ + lcol(6) = lhs_element(6, j); \ + rrow(6) = rhs_element(i, 6); \ + lcol(7) = lhs_element(7, j); \ + rrow(7) = rhs_element(i, 7); \ + +#define computeCol(j) \ + res(0, j) += lcol(0) * rrow(j); \ + res(1, j) += lcol(1) * rrow(j); \ + res(2, j) += lcol(2) * rrow(j); \ + res(3, j) += lcol(3) * rrow(j); \ + res(4, j) += lcol(4) * rrow(j); \ + res(5, j) += lcol(5) * rrow(j); \ + res(6, j) += lcol(6) * rrow(j); \ + res(7, j) += lcol(7) * rrow(j); \ + +#define computePass(i) \ + loadData(i, i); \ + \ + computeCol(0); \ + computeCol(1); \ + computeCol(2); \ + computeCol(3); \ + computeCol(4); \ + computeCol(5); \ + computeCol(6); \ + computeCol(7); \ + + computePass(0); + computePass(1); + computePass(2); + computePass(3); + computePass(4); + computePass(5); + computePass(6); + computePass(7); + +#undef lcol +#undef rrow +#undef lhs_element +#undef rhs_element +#undef loadData +#undef computeCol +#undef computePass + } // end loop over k + + // we've now iterated over all of the large (ie width 64) k blocks and + // accumulated results in registers. At this point thread (x, y, z) contains + // the sum across all big k blocks of the product of little k block of index (x, y) + // with block of index (y, z). To compute the final output, we need to reduce + // the 8 threads over y by summation. +#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask) + +#define reduceRow(i, mask) \ + shuffleInc(i, 0, mask); \ + shuffleInc(i, 1, mask); \ + shuffleInc(i, 2, mask); \ + shuffleInc(i, 3, mask); \ + shuffleInc(i, 4, mask); \ + shuffleInc(i, 5, mask); \ + shuffleInc(i, 6, mask); \ + shuffleInc(i, 7, mask); \ + +#define reduceMatrix(mask) \ + reduceRow(0, mask); \ + reduceRow(1, mask); \ + reduceRow(2, mask); \ + reduceRow(3, mask); \ + reduceRow(4, mask); \ + reduceRow(5, mask); \ + reduceRow(6, mask); \ + reduceRow(7, mask); \ + + // actually perform the reduction, now each thread of index (_, y, z) + // contains the correct values in its registers that belong in the output + // block + reduceMatrix(1); + reduceMatrix(2); + reduceMatrix(4); + +#undef shuffleInc +#undef reduceRow +#undef reduceMatrix + + // now we need to copy the 64 values into main memory. We can't split work + // among threads because all variables are in registers. There's 2 ways + // to do this: + // (1) have 1 thread do 64 writes from registers into global memory + // (2) have 1 thread do 64 writes into shared memory, and then 8 threads + // each do 8 writes into global memory. We can just overwrite the shared + // memory from the problem we just solved. + // (2) is slightly faster than (1) due to less branching and more ILP + + // TODO: won't yield much gain, but could just use currently unused shared mem + // and then we won't have to sync + // wait for shared mem to be out of use + __syncthreads(); + +#define writeResultShmem(i, j) \ + lhs_shmem[i + 8 * threadIdx.y + 64 * threadIdx.z + 512 * j] = res(i, j); \ + +#define writeRow(i) \ + writeResultShmem(i, 0); \ + writeResultShmem(i, 1); \ + writeResultShmem(i, 2); \ + writeResultShmem(i, 3); \ + writeResultShmem(i, 4); \ + writeResultShmem(i, 5); \ + writeResultShmem(i, 6); \ + writeResultShmem(i, 7); \ + + if (threadIdx.x == 0) { + writeRow(0); + writeRow(1); + writeRow(2); + writeRow(3); + writeRow(4); + writeRow(5); + writeRow(6); + writeRow(7); + } +#undef writeResultShmem +#undef writeRow + + const int max_i_write = (min)((int)((m_size - base_m - threadIdx.y + 7) / 8), 8); + const int max_j_write = (min)((int)((n_size - base_n - threadIdx.z + 7) / 8), 8); + + if (threadIdx.x < max_i_write) { + if (max_j_write == 8) { + Scalar val0 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 0]; + Scalar val1 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 1]; + Scalar val2 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 2]; + Scalar val3 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 3]; + Scalar val4 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 4]; + Scalar val5 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 5]; + Scalar val6 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 6]; + Scalar val7 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 7]; + + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 0) = val0; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 1) = val1; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 2) = val2; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 3) = val3; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 4) = val4; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 5) = val5; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 6) = val6; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 7) = val7; + } else { +#pragma unroll 7 + for (int j = 0; j < max_j_write; j++) { + Scalar val = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * j]; + output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * j) = val; + } + } + } +#undef res + } + + + template +__global__ void +__launch_bounds__(512) + EigenContractionKernel(const LhsMapper lhs, const RhsMapper rhs, + const OutputMapper output, + const Index m_size, const Index n_size, const Index k_size) { + __shared__ volatile Scalar lhs_shmem[72 * 64]; + __shared__ volatile Scalar rhs_shmem[72 * 64]; + + const Index m_block_idx = blockIdx.x; + const Index n_block_idx = blockIdx.y; + + const Index base_m = 64 * m_block_idx; + const Index base_n = 64 * n_block_idx; + + if (base_m + 63 < m_size && base_n + 63 < n_size) { + EigenContractionKernelInternal(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size); + } else { + EigenContractionKernelInternal(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size); + } + } + + + + template +__device__ EIGEN_STRONG_INLINE void + EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs, + const OutputMapper output, float4* lhs_shmem4, float2* rhs_shmem2, + const Index m_size, const Index n_size, const Index k_size) { + typedef float Scalar; + + const Index m_block_idx = blockIdx.x; + const Index n_block_idx = blockIdx.y; + + const Index base_m = 64 * m_block_idx; + const Index base_n = 64 * n_block_idx; + + const Index lane = threadIdx.x + 8 * (threadIdx.y % 4); + + // prefetch registers + float4 lhs_pf0; + float4 lhs_pf1; + + float4 rhs_pf0; + float4 rhs_pf1; + + // shared memory is formatted + // (contract idx in block, nocontract idx in block, block idx) + // where block idx is column major. This transposition limits the number of + // bank conflicts when reading the LHS. The core idea is that since the contracting + // index is shared by both sides, then the contracting index should be in threadIdx.x. + + // all of these indices assume float4 loading + // this thread loads the float4 starting at this index, and then also loads + // another float4 starting 32 columns to to the right + const Index horiz_block_idx = threadIdx.z / 2; + const Index vert_block_idx = threadIdx.x / 2 + 4 * (threadIdx.y % 2); + const Index horiz_idx_in_block = threadIdx.y / 2 + 4 * (threadIdx.z % 2); + const Index vert_idx_in_block = threadIdx.x % 2; + + // there's padding in both the LHS and RHS shared memory layouts. This padding + // allows for 0 bank conflicts on all shmem stores and loads. + // LHS padding: 1 float4 on each 8x8 block of floats + // RHS padding: 1 float2 on each block, and 12 additional float2s between vertical blocks + // 3 and 4 + + // storage indices + // lhs index with respect to float4s + const Index lhs_store_idx_base = + 136 * horiz_block_idx + + 17 * vert_block_idx + + 8 * vert_idx_in_block + + horiz_idx_in_block; + + // rhs index with respect to floats + const Index rhs_store_idx_base = + 552 * horiz_block_idx + + 66 * vert_block_idx + + 32 * (horiz_idx_in_block / 4) + (horiz_idx_in_block % 4) + + 16 * vert_idx_in_block + + ((vert_block_idx < 4) ? 0 : 24); + + const Index lhs_store_idx_0 = lhs_store_idx_base + 544 * 0; + const Index lhs_store_idx_1 = lhs_store_idx_base + 544 * 1; + + const Index rhs_store_idx_0 = (rhs_store_idx_base / 2) + ((lane < 16) ? 0 : 4); + const Index rhs_store_idx_1 = rhs_store_idx_0 + 2; + const Index rhs_store_idx_2 = rhs_store_idx_0 + 1104; + const Index rhs_store_idx_3 = rhs_store_idx_1 + 1104; + + // The below diagrams show which shmem index (with respect to floats) each element + // in an 8x8 input block gets packed into: + // LHS: + // 0 4 8 12 16 20 24 28 + // 1 5 9 13 17 21 25 29 + // 2 6 10 14 18 22 26 30 + // 3 7 11 15 19 23 27 31 + // 32 36 40 44 48 52 56 60 + // ... (pack as 2 rows of float4 indexed row major, each float4 is vertical) + // + // RHS: + // 0 1 2 3 32 33 34 35 + // 4 5 6 7 36 37 38 39 + // ... (pack as 2 cols of float4 indexed col major, each float4 is horizontal) + + // Each thread in a warp loads 2 float4s. This happens in 2 instructions. On each of these + // instruction, the warp loads 2 columns (2 cols * 64 elements / col = 128 elements = 32 threads + // * 4 elements/thread). For the LHS, we're able to store the loaded float4 directly into + // shmem (using a 128 bit store instruction). For the RHS, we need to transpose the data. + // This is done with warp shuffles. Furthermore, we only use 64 bit stores for the RHS, because + // 64 bits is only 2 columns (which is all we load in a warp), and the padding for the RHS + // doesn't meet 64 bit alignment requirements (namely, the 4 consecutive floats that we want + // to load on the RHS are 8 byte aligned, not 16 byte aligned, which is required for float4). + + const Index load_idx_vert = 4 * (threadIdx.x + 8 * (threadIdx.y % 2)); + const Index load_idx_horiz = (threadIdx.y / 2) + 4 * threadIdx.z; + + const Index lhs_vert = base_m + load_idx_vert; + const Index rhs_horiz_0 = base_n + load_idx_horiz; + const Index rhs_horiz_1 = base_n + load_idx_horiz + 32; + +#define prefetchIntoRegisters(base_k) \ + { \ + lhs_pf0 = internal::pset1(0); \ + lhs_pf1 = internal::pset1(0); \ + \ + rhs_pf0 = internal::pset1(0); \ + rhs_pf1 = internal::pset1(0); \ + \ + const Index lhs_horiz_0 = base_k + load_idx_horiz; \ + const Index lhs_horiz_1 = base_k + load_idx_horiz + 32; \ + if (!needs_edge_check || lhs_vert + 3 < m_size) { \ + if (lhs_horiz_1 < k_size) { \ + lhs_pf0 = lhs.loadPacket(lhs_vert, lhs_horiz_0); \ + lhs_pf1 = lhs.loadPacket(lhs_vert, lhs_horiz_1); \ + } else if (lhs_horiz_0 < k_size) { \ + lhs_pf0 = lhs.loadPacket(lhs_vert, lhs_horiz_0); \ + } \ + } else if (lhs_vert + 2 < m_size) { \ + if (lhs_horiz_1 < k_size) { \ + lhs_pf0.x = lhs(lhs_vert + 0, lhs_horiz_0); \ + lhs_pf0.y = lhs(lhs_vert + 1, lhs_horiz_0); \ + lhs_pf0.z = lhs(lhs_vert + 2, lhs_horiz_0); \ + \ + lhs_pf1.x = lhs(lhs_vert + 0, lhs_horiz_1); \ + lhs_pf1.y = lhs(lhs_vert + 1, lhs_horiz_1); \ + lhs_pf1.z = lhs(lhs_vert + 2, lhs_horiz_1); \ + } else if (lhs_horiz_0 < k_size) { \ + lhs_pf0.x = lhs(lhs_vert + 0, lhs_horiz_0); \ + lhs_pf0.y = lhs(lhs_vert + 1, lhs_horiz_0); \ + lhs_pf0.z = lhs(lhs_vert + 2, lhs_horiz_0); \ + } \ + } else if (lhs_vert + 1 < m_size) { \ + if (lhs_horiz_1 < k_size) { \ + lhs_pf0.x = lhs(lhs_vert + 0, lhs_horiz_0); \ + lhs_pf0.y = lhs(lhs_vert + 1, lhs_horiz_0); \ + \ + lhs_pf1.x = lhs(lhs_vert + 0, lhs_horiz_1); \ + lhs_pf1.y = lhs(lhs_vert + 1, lhs_horiz_1); \ + } else if (lhs_horiz_0 < k_size) { \ + lhs_pf0.x = lhs(lhs_vert + 0, lhs_horiz_0); \ + lhs_pf0.y = lhs(lhs_vert + 1, lhs_horiz_0); \ + } \ + } else if (lhs_vert < m_size) { \ + if (lhs_horiz_1 < k_size) { \ + lhs_pf0.x = lhs(lhs_vert + 0, lhs_horiz_0); \ + lhs_pf1.x = lhs(lhs_vert + 0, lhs_horiz_1); \ + } else if (lhs_horiz_0 < k_size) { \ + lhs_pf0.x = lhs(lhs_vert + 0, lhs_horiz_0); \ + } \ +} \ + \ + const Index rhs_vert = base_k + load_idx_vert; \ + if (rhs_vert + 3 < k_size) { \ + if (!needs_edge_check || rhs_horiz_1 < n_size) { \ + rhs_pf0 = rhs.loadPacket(rhs_vert, rhs_horiz_0); \ + rhs_pf1 = rhs.loadPacket(rhs_vert, rhs_horiz_1); \ + } else if (rhs_horiz_0 < n_size) { \ + rhs_pf0 = rhs.loadPacket(rhs_vert, rhs_horiz_0); \ + } \ + } else if (rhs_vert + 2 < k_size) { \ + if (!needs_edge_check || rhs_horiz_1 < n_size) { \ + rhs_pf0.x = rhs(rhs_vert + 0, rhs_horiz_0); \ + rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz_0); \ + rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz_0); \ + \ + rhs_pf1.x = rhs(rhs_vert + 0, rhs_horiz_1); \ + rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz_1); \ + rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz_1); \ + } else if (rhs_horiz_0 < n_size) { \ + rhs_pf0.x = rhs(rhs_vert + 0, rhs_horiz_0); \ + rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz_0); \ + rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz_0); \ + } \ + } else if (rhs_vert + 1 < k_size) { \ + if (!needs_edge_check || rhs_horiz_1 < n_size) { \ + rhs_pf0.x = rhs(rhs_vert + 0, rhs_horiz_0); \ + rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz_0); \ + \ + rhs_pf1.x = rhs(rhs_vert + 0, rhs_horiz_1); \ + rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz_1); \ + } else if (rhs_horiz_0 < n_size) { \ + rhs_pf0.x = rhs(rhs_vert + 0, rhs_horiz_0); \ + rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz_0); \ + } \ + } else if (rhs_vert < k_size) { \ + if (!needs_edge_check || rhs_horiz_1 < n_size) { \ + rhs_pf0.x = rhs(rhs_vert + 0, rhs_horiz_0); \ + rhs_pf1.x = rhs(rhs_vert + 0, rhs_horiz_1); \ + } else if (rhs_horiz_0 < n_size) { \ + rhs_pf0.x = rhs(rhs_vert + 0, rhs_horiz_0); \ + } \ +} \ + \ + float swap_val0 = (lane < 16) ? rhs_pf0.z : rhs_pf0.x; \ + float swap_val1 = (lane < 16) ? rhs_pf0.w : rhs_pf0.y; \ + float swap_val2 = (lane < 16) ? rhs_pf1.z : rhs_pf1.x; \ + float swap_val3 = (lane < 16) ? rhs_pf1.w : rhs_pf1.y; \ + \ + swap_val0 = __shfl_xor(swap_val0, 16); \ + swap_val1 = __shfl_xor(swap_val1, 16); \ + swap_val2 = __shfl_xor(swap_val2, 16); \ + swap_val3 = __shfl_xor(swap_val3, 16); \ + \ + if (lane < 16) { \ + rhs_pf0.z = swap_val0; \ + rhs_pf0.w = swap_val1; \ + rhs_pf1.z = swap_val2; \ + rhs_pf1.w = swap_val3; \ + } else { \ + rhs_pf0.x = swap_val0; \ + rhs_pf0.y = swap_val1; \ + rhs_pf1.x = swap_val2; \ + rhs_pf1.y = swap_val3; \ + } \ +} \ + + +#define writeRegToShmem(_) \ + lhs_shmem4[lhs_store_idx_0] = lhs_pf0; \ + \ + rhs_shmem2[rhs_store_idx_0] = make_float2(rhs_pf0.x, rhs_pf0.z); \ + rhs_shmem2[rhs_store_idx_1] = make_float2(rhs_pf0.y, rhs_pf0.w); \ + \ + lhs_shmem4[lhs_store_idx_1] = lhs_pf1; \ + \ + rhs_shmem2[rhs_store_idx_2] = make_float2(rhs_pf1.x, rhs_pf1.z); \ + rhs_shmem2[rhs_store_idx_3] = make_float2(rhs_pf1.y, rhs_pf1.w); \ + + // declare and initialize result array +#define res(i, j) _res_##i##j +#define initResultRow(i) \ + Scalar res(i, 0) = Scalar(0); \ + Scalar res(i, 1) = Scalar(0); \ + Scalar res(i, 2) = Scalar(0); \ + Scalar res(i, 3) = Scalar(0); \ + Scalar res(i, 4) = Scalar(0); \ + Scalar res(i, 5) = Scalar(0); \ + Scalar res(i, 6) = Scalar(0); \ + Scalar res(i, 7) = Scalar(0); \ + + initResultRow(0); + initResultRow(1); + initResultRow(2); + initResultRow(3); + initResultRow(4); + initResultRow(5); + initResultRow(6); + initResultRow(7); +#undef initResultRow + + for (Index base_k = 0; base_k < k_size; base_k += 64) { + // wait for previous iteration to finish with shmem. Despite common sense, + // the code is a bit faster with this here then at bottom of loop + __syncthreads(); + + prefetchIntoRegisters(base_k); + writeRegToShmem(); + +#undef prefetchIntoRegisters +#undef writeRegoToShmem + + // wait for shared mem packing to be done before starting computation + __syncthreads(); + + // compute 8x8 matrix product by outer product. This involves packing one column + // of LHS and one row of RHS into registers (takes 16 registers). + + float4 _lcol0; + float4 _lcol1; + float2 _rrow0; + float2 _rrow1; + float2 _rrow2; + float2 _rrow3; + +#define lcol0 _lcol0.x +#define lcol1 _lcol0.y +#define lcol2 _lcol0.z +#define lcol3 _lcol0.w +#define lcol4 _lcol1.x +#define lcol5 _lcol1.y +#define lcol6 _lcol1.z +#define lcol7 _lcol1.w +#define rrow0 _rrow0.x +#define rrow1 _rrow0.y +#define rrow2 _rrow1.x +#define rrow3 _rrow1.y +#define rrow4 _rrow2.x +#define rrow5 _rrow2.y +#define rrow6 _rrow3.x +#define rrow7 _rrow3.y + + // Now x corresponds to k, y to m, and z to n + const float4* lhs_block = &lhs_shmem4[threadIdx.x + 8 * (threadIdx.y % 2) + 17 * (threadIdx.y / 2)]; + const float2* rhs_block = &rhs_shmem2[2 * threadIdx.x + 16 * (threadIdx.z % 2) + 276 * (threadIdx.z / 2)]; + +#define lhs_element(i, k) lhs_block[68 * i + 136 * k] +#define rhs_element(k, j) rhs_block[33 * k + 1104 * j + ((k < 4) ? 0 : 12)] + +#define loadData(i) \ + _lcol0 = lhs_element(0, i); \ + _rrow0 = rhs_element(i, 0); \ + _rrow1 = *(&(rhs_element(i, 0)) + 1); \ + _lcol1 = lhs_element(1, i); \ + _rrow2 = rhs_element(i, 1); \ + _rrow3 = *(&(rhs_element(i, 1)) + 1); \ + +#define computeCol(j) \ + res(0, j) += lcol0 * rrow##j; \ + res(1, j) += lcol1 * rrow##j; \ + res(2, j) += lcol2 * rrow##j; \ + res(3, j) += lcol3 * rrow##j; \ + res(4, j) += lcol4 * rrow##j; \ + res(5, j) += lcol5 * rrow##j; \ + res(6, j) += lcol6 * rrow##j; \ + res(7, j) += lcol7 * rrow##j; \ + +#define computePass(i) \ + loadData(i); \ + \ + computeCol(0); \ + computeCol(1); \ + computeCol(2); \ + computeCol(3); \ + computeCol(4); \ + computeCol(5); \ + computeCol(6); \ + computeCol(7); \ + + computePass(0); + computePass(1); + computePass(2); + computePass(3); + computePass(4); + computePass(5); + computePass(6); + computePass(7); + +#undef lcol0 +#undef lcol1 +#undef lcol2 +#undef lcol3 +#undef lcol4 +#undef lcol5 +#undef lcol6 +#undef lcol7 +#undef rrow0 +#undef rrow1 +#undef rrow2 +#undef rrow3 +#undef rrow4 +#undef rrow5 +#undef rrow6 +#undef rrow7 + +#undef computePass +#undef computeCol +#undef loadData +#undef lhs_element +#undef rhs_element + + } // end loop over k + + // we've now iterated over all of the large (ie width 64) k blocks and + // accumulated results in registers. At this point thread (x, y, z) contains + // the sum across all big k blocks of the product of little k block of index (x, y) + // with block of index (y, z). To compute the final output, we need to reduce + // the 8 threads over y by summation. +#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask) + +#define reduceRow(i, mask) \ + shuffleInc(i, 0, mask); \ + shuffleInc(i, 1, mask); \ + shuffleInc(i, 2, mask); \ + shuffleInc(i, 3, mask); \ + shuffleInc(i, 4, mask); \ + shuffleInc(i, 5, mask); \ + shuffleInc(i, 6, mask); \ + shuffleInc(i, 7, mask); \ + +#define reduceMatrix(mask) \ + reduceRow(0, mask); \ + reduceRow(1, mask); \ + reduceRow(2, mask); \ + reduceRow(3, mask); \ + reduceRow(4, mask); \ + reduceRow(5, mask); \ + reduceRow(6, mask); \ + reduceRow(7, mask); \ + + // actually perform the reduction, now each thread of index (_, y, z) + // contains the correct values in its registers that belong in the output + // block + reduceMatrix(1); + reduceMatrix(2); + reduceMatrix(4); + +#undef shuffleInc +#undef reduceRow +#undef reduceMatrix + + // now we need to copy the 64 values into main memory. We can't split work + // among threads because all variables are in registers. There's 2 ways + // to do this: + // (1) have 1 thread do 64 writes from registers into global memory + // (2) have 1 thread do 64 writes into shared memory, and then 8 threads + // each do 8 writes into global memory. We can just overwrite the shared + // memory from the problem we just solved. + // (3) Copies the values into new registers using conditional logic. + +#define makeAssignments(i) \ + val0 = res(i, 0); \ + val1 = res(i, 1); \ + val2 = res(i, 2); \ + val3 = res(i, 3); \ + val4 = res(i, 4); \ + val5 = res(i, 5); \ + val6 = res(i, 6); \ + val7 = res(i, 7); \ + + Scalar val0; + Scalar val1; + Scalar val2; + Scalar val3; + Scalar val4; + Scalar val5; + Scalar val6; + Scalar val7; + + switch (threadIdx.x) { + case 0: + makeAssignments(0); + break; + case 1: + makeAssignments(1); + break; + case 2: + makeAssignments(2); + break; + case 3: + makeAssignments(3); + break; + case 4: + makeAssignments(4); + break; + case 5: + makeAssignments(5); + break; + case 6: + makeAssignments(6); + break; + case 7: + makeAssignments(7); + break; + } + +#undef res + + const Index vert_base = base_m + 4 * threadIdx.y + (threadIdx.x % 4) + 32 * (threadIdx.x / 4); + const Index horiz_base = base_n + 4 * threadIdx.z; + + if (!needs_edge_check || vert_base < m_size) { + if (!needs_edge_check || horiz_base + 35 < n_size) { + output(vert_base, horiz_base + 0) = val0; + output(vert_base, horiz_base + 1) = val1; + output(vert_base, horiz_base + 2) = val2; + output(vert_base, horiz_base + 3) = val3; + output(vert_base, horiz_base + 32) = val4; + output(vert_base, horiz_base + 33) = val5; + output(vert_base, horiz_base + 34) = val6; + output(vert_base, horiz_base + 35) = val7; + } else if (horiz_base + 34 < n_size) { + output(vert_base, horiz_base + 0) = val0; + output(vert_base, horiz_base + 1) = val1; + output(vert_base, horiz_base + 2) = val2; + output(vert_base, horiz_base + 3) = val3; + output(vert_base, horiz_base + 32) = val4; + output(vert_base, horiz_base + 33) = val5; + output(vert_base, horiz_base + 34) = val6; + } else if (horiz_base + 33 < n_size) { + output(vert_base, horiz_base + 0) = val0; + output(vert_base, horiz_base + 1) = val1; + output(vert_base, horiz_base + 2) = val2; + output(vert_base, horiz_base + 3) = val3; + output(vert_base, horiz_base + 32) = val4; + output(vert_base, horiz_base + 33) = val5; + } else if (horiz_base + 32 < n_size) { + output(vert_base, horiz_base + 0) = val0; + output(vert_base, horiz_base + 1) = val1; + output(vert_base, horiz_base + 2) = val2; + output(vert_base, horiz_base + 3) = val3; + output(vert_base, horiz_base + 32) = val4; + } else if (horiz_base + 3 < n_size) { + output(vert_base, horiz_base + 0) = val0; + output(vert_base, horiz_base + 1) = val1; + output(vert_base, horiz_base + 2) = val2; + output(vert_base, horiz_base + 3) = val3; + } else if (horiz_base + 2 < n_size) { + output(vert_base, horiz_base + 0) = val0; + output(vert_base, horiz_base + 1) = val1; + output(vert_base, horiz_base + 2) = val2; + } else if (horiz_base + 1 < n_size) { + output(vert_base, horiz_base + 0) = val0; + output(vert_base, horiz_base + 1) = val1; + } else if (horiz_base < n_size) { + output(vert_base, horiz_base + 0) = val0; + } + } + } + + + template +__global__ void + __launch_bounds__(512) + EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs, + const OutputMapper output, + const Index m_size, const Index n_size, const Index k_size) { + __shared__ float4 lhs_shmem[(68 * 64) / 4]; + __shared__ float2 rhs_shmem[((66 * 8 + 24) * 8) / 2]; + + const Index m_block_idx = blockIdx.x; + const Index n_block_idx = blockIdx.y; + + const Index base_m = 64 * m_block_idx; + const Index base_n = 64 * n_block_idx; + + if (base_m + 63 < m_size && base_n + 63 < n_size) { + EigenFloatContractionKernelInternal(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size); + } else { + EigenFloatContractionKernelInternal(lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size); + } + } + + + template + struct TensorEvaluator, GpuDevice> : + public TensorContractionEvaluatorBase, GpuDevice> > { + + typedef GpuDevice Device; + + typedef TensorEvaluator, Device> Self; + typedef TensorContractionEvaluatorBase Base; + + typedef TensorContractionOp XprType; + typedef typename internal::remove_const::type Scalar; + typedef typename XprType::Packet Packet; + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + typedef array::Dimensions::count> left_dim_mapper_t; + typedef array::Dimensions::count> right_dim_mapper_t; + + typedef array::value> contract_t; + typedef array::Dimensions::count - internal::array_size::value> left_nocontract_t; + typedef array::Dimensions::count - internal::array_size::value> right_nocontract_t; + + static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * internal::array_size::value>::size; + + typedef DSizes Dimensions; + + // typedefs needed in evalTo + typedef typename internal::remove_const::type LhsScalar; + typedef typename internal::remove_const::type RhsScalar; + + typedef TensorEvaluator LeftEvaluator; + typedef TensorEvaluator RightEvaluator; + + typedef typename LeftEvaluator::Dimensions LeftDimensions; + typedef typename RightEvaluator::Dimensions RightDimensions; + + EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) : + Base(op, device) {} + + // We need to redefine this method to make nvcc happy + EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + this->m_leftImpl.evalSubExprsIfNeeded(NULL); + this->m_rightImpl.evalSubExprsIfNeeded(NULL); + if (data) { + evalTo(data); + return false; + } else { + this->m_result = static_cast(this->m_device.allocate(this->dimensions().TotalSize() * sizeof(Scalar))); + evalTo(this->m_result); + return true; + } + } + + void evalTo(Scalar* buffer) const { + if (this->m_lhs_inner_dim_contiguous) { + if (this->m_rhs_inner_dim_contiguous) { + if (this->m_rhs_inner_dim_reordered) { + evalTyped(buffer); + } + else { + evalTyped(buffer); + } + } + else { + if (this->m_rhs_inner_dim_reordered) { + evalTyped(buffer); + } + else { + evalTyped(buffer); + } + } + } + else { + if (this->m_rhs_inner_dim_contiguous) { + if (this->m_rhs_inner_dim_reordered) { + evalTyped(buffer); + } + else { + evalTyped(buffer); + } + } + else { + if (this->m_rhs_inner_dim_reordered) { + evalTyped(buffer); + } + else { + evalTyped(buffer); + } + } + } + } + + template + void evalTyped(Scalar* buffer) const { + // columns in left side, rows in right side + const Index k = this->m_k_size; + + // rows in left side + const Index m = this->m_i_size; + + // columns in right side + const Index n = this->m_j_size; + + // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar) + this->m_device.memset(buffer, 0, m * n * sizeof(Scalar)); + + typedef internal::TensorContractionInputMapper LhsMapper; + + typedef internal::TensorContractionInputMapper RhsMapper; + + typedef internal::blas_data_mapper OutputMapper; + + + // initialize data mappers + LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides, + this->m_left_contracting_strides, this->m_k_strides); + + RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides, + this->m_right_contracting_strides, this->m_k_strides); + + OutputMapper output(buffer, m); + + const Index m_blocks = (m + 63) / 64; + const Index n_blocks = (n + 63) / 64; + const dim3 num_blocks(m_blocks, n_blocks, 1); + const dim3 block_size(8, 8, 8); + + cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeEightByte); + if (internal::is_same::value && + internal::is_same::value) { + EigenFloatContractionKernel + <<m_device.stream()>>>(lhs, rhs, output, m, n, k); + } else { + EigenContractionKernel + <<m_device.stream()>>>(lhs, rhs, output, m, n, k); + } + + assert(cudaGetLastError() == cudaSuccess); + } + }; + +} // end namespace Eigen + +#endif // EIGEN_USE_GPU and __CUDACC__ + +#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_CUDA_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h index ef5e11537..fad342eab 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -104,19 +104,41 @@ struct GpuDevice { EIGEN_STRONG_INLINE const cudaStream_t& stream() const { return *stream_; } - /*EIGEN_DEVICE_FUNC*/ EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { +#ifndef __CUDA_ARCH__ void* result; - cudaMalloc(&result, num_bytes); + assert(cudaMalloc(&result, num_bytes) == cudaSuccess); + assert(result != NULL); return result; +#else + assert(false && "The default device should be used instead to generate kernel code"); + return NULL; +#endif } - /*EIGEN_DEVICE_FUNC */EIGEN_STRONG_INLINE void deallocate(void* buffer) const { - cudaFree(buffer); + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const { +#ifndef __CUDA_ARCH__ + assert(buffer != NULL); + assert(cudaFree(buffer) == cudaSuccess); +#else + assert(false && "The default device should be used instead to generate kernel code"); +#endif } - EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const { - cudaMemcpyAsync(dst, src, n, cudaMemcpyDeviceToDevice, *stream_); + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const { +#ifndef __CUDA_ARCH__ + assert(cudaMemcpyAsync(dst, src, n, cudaMemcpyDeviceToDevice, *stream_) == cudaSuccess); +#else + assert(false && "The default device should be used instead to generate kernel code"); +#endif } - EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { - cudaMemsetAsync(buffer, c, n, *stream_); + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { +#ifndef __CUDA_ARCH__ + assert(cudaMemsetAsync(buffer, c, n, *stream_) == cudaSuccess); +#else + assert(false && "The default device should be used instead to generate kernel code"); +#endif } EIGEN_STRONG_INLINE size_t numThreads() const { From 152f3218ac9b6941cf6dbc960c2d4a6d1099eb06 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 3 Oct 2014 19:33:44 -0700 Subject: [PATCH 0866/1103] Improved contraction test --- unsupported/test/cxx11_tensor_contraction.cpp | 32 +++++++++++++++++++ 1 file changed, 32 insertions(+) diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp index a37fcd967..2b599d30d 100644 --- a/unsupported/test/cxx11_tensor_contraction.cpp +++ b/unsupported/test/cxx11_tensor_contraction.cpp @@ -201,6 +201,37 @@ static void test_full_redux() } +static void test_contraction_of_contraction() +{ + Tensor t1(2, 2); + Tensor t2(2, 2); + Tensor t3(2, 2); + Tensor t4(2, 2); + t1.setRandom(); + t2.setRandom(); + t3.setRandom(); + t4.setRandom(); + + Eigen::array dims({{DimPair(1, 0)}}); + auto contract1 = t1.contract(t2, dims); + auto diff = t3 - contract1; + auto contract2 = t1.contract(t4, dims); + Tensor result = contract2.contract(diff, dims); + VERIFY_IS_EQUAL(result.dimension(0), 2); + VERIFY_IS_EQUAL(result.dimension(1), 2); + + Eigen::Map m1(t1.data(), 2, 2); + Eigen::Map m2(t2.data(), 2, 2); + Eigen::Map m3(t3.data(), 2, 2); + Eigen::Map m4(t4.data(), 2, 2); + Eigen::MatrixXf expected = (m1 * m4) * (m3 - m1 * m2); + VERIFY_IS_APPROX(result(0, 0), expected(0, 0)); + VERIFY_IS_APPROX(result(0, 1), expected(0, 1)); + VERIFY_IS_APPROX(result(1, 0), expected(1, 0)); + VERIFY_IS_APPROX(result(1, 1), expected(1, 1)); +} + + static void test_expr() { Tensor mat1(2, 3); @@ -328,6 +359,7 @@ void test_cxx11_tensor_contraction() CALL_SUBTEST(test_multidims()); CALL_SUBTEST(test_holes()); CALL_SUBTEST(test_full_redux()); + CALL_SUBTEST(test_contraction_of_contraction()); CALL_SUBTEST(test_expr()); CALL_SUBTEST(test_out_of_order_contraction()); CALL_SUBTEST(test_consistency()); From 95a430a2ca8489a85d0a12ffa66d260011c11745 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 3 Oct 2014 19:45:19 -0700 Subject: [PATCH 0867/1103] Vector primitives for CUDA --- Eigen/Core | 5 + Eigen/src/Core/arch/CUDA/MathFunctions.h | 75 +++++++ Eigen/src/Core/arch/CUDA/PacketMath.h | 260 +++++++++++++++++++++++ 3 files changed, 340 insertions(+) create mode 100644 Eigen/src/Core/arch/CUDA/MathFunctions.h create mode 100644 Eigen/src/Core/arch/CUDA/PacketMath.h diff --git a/Eigen/Core b/Eigen/Core index 776b7faf3..537ac16b2 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -294,6 +294,11 @@ using std::ptrdiff_t; #include "src/Core/arch/NEON/Complex.h" #endif +#if defined EIGEN_VECTORIZE_CUDA + #include "src/Core/arch/CUDA/PacketMath.h" + #include "src/Core/arch/CUDA/MathFunctions.h" +#endif + #include "src/Core/arch/Default/Settings.h" #include "src/Core/functors/BinaryFunctors.h" diff --git a/Eigen/src/Core/arch/CUDA/MathFunctions.h b/Eigen/src/Core/arch/CUDA/MathFunctions.h new file mode 100644 index 000000000..e7305c01e --- /dev/null +++ b/Eigen/src/Core/arch/CUDA/MathFunctions.h @@ -0,0 +1,75 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_MATH_FUNCTIONS_CUDA_H +#define EIGEN_MATH_FUNCTIONS_CUDA_H + +namespace Eigen { + +namespace internal { + +// Make sure this is only available when targeting a GPU: we don't want to +// introduce conflicts between these packet_traits definitions and the ones +// we'll use on the host side (SSE, AVX, ...) +#if defined(__CUDACC__) && defined(EIGEN_USE_GPU) +template<> EIGEN_STRONG_INLINE +float4 plog(const float4& a) +{ + return make_float4(logf(a.x), logf(a.y), logf(a.z), logf(a.w)); +} + +template<> EIGEN_STRONG_INLINE +double2 plog(const double2& a) +{ + return make_double2(log(a.x), log(a.y)); +} + +template<> EIGEN_STRONG_INLINE +float4 pexp(const float4& a) +{ + return make_float4(expf(a.x), expf(a.y), expf(a.z), expf(a.w)); +} + +template<> EIGEN_STRONG_INLINE +double2 pexp(const double2& a) +{ + return make_double2(exp(a.x), exp(a.y)); +} + +template<> EIGEN_STRONG_INLINE +float4 psqrt(const float4& a) +{ + return make_float4(sqrtf(a.x), sqrtf(a.y), sqrtf(a.z), sqrtf(a.w)); +} + +template<> EIGEN_STRONG_INLINE +double2 psqrt(const double2& a) +{ + return make_double2(sqrt(a.x), sqrt(a.y)); +} + +template<> EIGEN_STRONG_INLINE +float4 prsqrt(const float4& a) +{ + return make_float4(rsqrtf(a.x), rsqrtf(a.y), rsqrtf(a.z), rsqrtf(a.w)); +} + +template<> EIGEN_STRONG_INLINE +double2 prsqrt(const double2& a) +{ + return make_double2(rsqrt(a.x), rsqrt(a.y)); +} + +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_CUDA_H diff --git a/Eigen/src/Core/arch/CUDA/PacketMath.h b/Eigen/src/Core/arch/CUDA/PacketMath.h new file mode 100644 index 000000000..5b0abe2e6 --- /dev/null +++ b/Eigen/src/Core/arch/CUDA/PacketMath.h @@ -0,0 +1,260 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_PACKET_MATH_CUDA_H +#define EIGEN_PACKET_MATH_CUDA_H + +namespace Eigen { + +namespace internal { + +// Make sure this is only available when targeting a GPU: we don't want to +// introduce conflicts between these packet_traits definitions and the ones +// we'll use on the host side (SSE, AVX, ...) +#if defined(__CUDACC__) && defined(EIGEN_USE_GPU) +template<> struct is_arithmetic { enum { value = true }; }; +template<> struct is_arithmetic { enum { value = true }; }; + + +template<> struct packet_traits : default_packet_traits +{ + typedef float4 type; + typedef float4 half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=4, + HasHalfPacket = 0, + + HasDiv = 1, + HasSin = 0, + HasCos = 0, + HasLog = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + + HasBlend = 0, + }; +}; + +template<> struct packet_traits : default_packet_traits +{ + typedef double2 type; + typedef double2 half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size=2, + HasHalfPacket = 0, + + HasDiv = 1, + HasLog = 1, + HasExp = 1, + HasSqrt = 1, + HasRsqrt = 1, + + HasBlend = 0, + }; +}; + + +template<> struct unpacket_traits { typedef float type; enum {size=4}; typedef float4 half; }; +template<> struct unpacket_traits { typedef double type; enum {size=2}; typedef double2 half; }; + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pset1(const float& from) { + return make_float4(from, from, from, from); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pset1(const double& from) { + return make_double2(from, from); +} + + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 plset(const float& a) { + return make_float4(a, a+1, a+2, a+3); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 plset(const double& a) { + return make_double2(a, a+1); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 padd(const float4& a, const float4& b) { + return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 padd(const double2& a, const double2& b) { + return make_double2(a.x+b.x, a.y+b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 psub(const float4& a, const float4& b) { + return make_float4(a.x-b.x, a.y-b.y, a.z-b.z, a.w-b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 psub(const double2& a, const double2& b) { + return make_double2(a.x-b.x, a.y-b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pnegate(const float4& a) { + return make_float4(-a.x, -a.y, -a.z, -a.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pnegate(const double2& a) { + return make_double2(-a.x, -a.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pconj(const float4& a) { return a; } +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pconj(const double2& a) { return a; } + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmul(const float4& a, const float4& b) { + return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmul(const double2& a, const double2& b) { + return make_double2(a.x*b.x, a.y*b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pdiv(const float4& a, const float4& b) { + return make_float4(a.x/b.x, a.y/b.y, a.z/b.z, a.w/b.w); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pdiv(const double2& a, const double2& b) { + return make_double2(a.x/b.x, a.y/b.y); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmin(const float4& a, const float4& b) { + return make_float4(fminf(a.x, b.x), fminf(a.y, b.y), fminf(a.z, b.z), fminf(a.w, b.w)); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmin(const double2& a, const double2& b) { + return make_double2(fmin(a.x, b.x), fmin(a.y, b.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmax(const float4& a, const float4& b) { + return make_float4(fmaxf(a.x, b.x), fmaxf(a.y, b.y), fmaxf(a.z, b.z), fmaxf(a.w, b.w)); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmax(const double2& a, const double2& b) { + return make_double2(fmax(a.x, b.x), fmax(a.y, b.y)); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pload(const float* from) { + return *reinterpret_cast(from); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pload(const double* from) { + return *reinterpret_cast(from); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 ploadu(const float* from) { + return make_float4(from[0], from[1], from[2], from[3]); +} +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploadu(const double* from) { + return make_double2(from[0], from[1]); +} + +template<> EIGEN_STRONG_INLINE float4 ploaddup(const float* from) { + return make_float4(from[0], from[0], from[1], from[1]); +} +template<> EIGEN_STRONG_INLINE double2 ploaddup(const double* from) { + return make_double2(from[0], from[0]); +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore(float* to, const float4& from) { + *reinterpret_cast(to) = from; +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore(double* to, const double2& from) { + *reinterpret_cast(to) = from; +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu(float* to, const float4& from) { + to[0] = from.x; + to[1] = from.y; + to[2] = from.z; + to[3] = from.w; +} + +template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu(double* to, const double2& from) { + to[0] = from.x; + to[1] = from.y; +} + +#ifdef __CUDA_ARCH__ +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro(const float* from) { + return __ldg((const float4*)from); +} +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro(const double* from) { + return __ldg((const double2*)from); +} + +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro(const float* from) { + return make_float4(__ldg(from+0), __ldg(from+1), __ldg(from+2), __ldg(from+3)); +} +template<> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro(const double* from) { + return make_double2(__ldg(from+0), __ldg(from+1)); +} +#endif + +template<> EIGEN_DEVICE_FUNC inline float4 pgather(const float* from, int stride) { + return make_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]); +} + +template<> EIGEN_DEVICE_FUNC inline double2 pgather(const double* from, int stride) { + return make_double2(from[0*stride], from[1*stride]); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const float4& from, int stride) { + to[stride*0] = from.x; + to[stride*1] = from.y; + to[stride*2] = from.z; + to[stride*3] = from.w; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const double2& from, int stride) { + to[stride*0] = from.x; + to[stride*1] = from.y; +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { + double tmp = kernel.packet[0].y; + kernel.packet[0].y = kernel.packet[1].x; + kernel.packet[1].x = tmp; + + tmp = kernel.packet[0].z; + kernel.packet[0].z = kernel.packet[2].x; + kernel.packet[2].x = tmp; + + tmp = kernel.packet[0].w; + kernel.packet[0].w = kernel.packet[3].x; + kernel.packet[3].x = tmp; + + tmp = kernel.packet[1].z; + kernel.packet[1].z = kernel.packet[2].y; + kernel.packet[2].y = tmp; + + tmp = kernel.packet[1].w; + kernel.packet[1].w = kernel.packet[3].y; + kernel.packet[3].y = tmp; + + tmp = kernel.packet[2].w; + kernel.packet[2].w = kernel.packet[3].z; + kernel.packet[3].z = tmp; +} + +template<> EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { + double tmp = kernel.packet[0].y; + kernel.packet[0].y = kernel.packet[1].x; + kernel.packet[1].x = tmp; +} + +#endif + +} // end namespace internal + +} // end namespace Eigen + + +#endif // EIGEN_PACKET_MATH_CUDA_H From bbce6fa65d8a196f05e0428d014e0e3865e202f3 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 3 Oct 2014 19:55:35 -0700 Subject: [PATCH 0868/1103] define EIGEN_VECTORIZE_CUDA when compiling with nvcc --- Eigen/Core | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/Eigen/Core b/Eigen/Core index 537ac16b2..acdeca5f4 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -178,6 +178,11 @@ #endif #endif +#if defined __CUDACC__ + #define EIGEN_VECTORIZE_CUDA + #include +#endif + #if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE) #define EIGEN_HAS_OPENMP #endif From 7a1763995396e6119092ce5fc4ca2d536b6acb73 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 6 Oct 2014 11:41:50 +0200 Subject: [PATCH 0869/1103] Extend unit tests to check uncompressed sparse inputs in sparse solvers --- test/sparse_solver.h | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/test/sparse_solver.h b/test/sparse_solver.h index d84aff070..8c8d7f939 100644 --- a/test/sparse_solver.h +++ b/test/sparse_solver.h @@ -67,6 +67,15 @@ void check_sparse_solving(Solver& solver, const typename Solver::MatrixType& A, VERIFY(oldb.isApprox(db) && "sparse solver testing: the rhs should not be modified!"); VERIFY(x.isApprox(refX,test_precision())); } + + // test uncompressed inputs + { + Mat A2 = A; + A2.reserve((ArrayXf::Random(A.outerSize())+2).template cast().eval()); + solver.compute(A2); + Rhs x = solver.solve(b); + VERIFY(x.isApprox(refX,test_precision())); + } } template From fb53ff1edaa96bc1c6090e5ab4982751dbdfc910 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 6 Oct 2014 11:42:31 +0200 Subject: [PATCH 0870/1103] Fix SparseLU regarding uncompressed inputs and avoid manual new/delete calls. --- Eigen/src/SparseLU/SparseLU.h | 30 ++++++++++++++++-------------- 1 file changed, 16 insertions(+), 14 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 1789830b9..b02796293 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -364,30 +364,32 @@ void SparseLU::analyzePattern(const MatrixType& mat) //TODO It is possible as in SuperLU to compute row and columns scaling vectors to equilibrate the matrix mat. + // Firstly, copy the whole input matrix. + m_mat = mat; + + // Compute fill-in ordering OrderingType ord; - ord(mat,m_perm_c); + ord(m_mat,m_perm_c); // Apply the permutation to the column of the input matrix - //First copy the whole input matrix. - m_mat = mat; - if (m_perm_c.size()) { + if (m_perm_c.size()) + { m_mat.uncompress(); //NOTE: The effect of this command is only to create the InnerNonzeros pointers. FIXME : This vector is filled but not subsequently used. - //Then, permute only the column pointers - const Index * outerIndexPtr; - if (mat.isCompressed()) outerIndexPtr = mat.outerIndexPtr(); - else - { - Index *outerIndexPtr_t = new Index[mat.cols()+1]; - for(Index i = 0; i <= mat.cols(); i++) outerIndexPtr_t[i] = m_mat.outerIndexPtr()[i]; - outerIndexPtr = outerIndexPtr_t; - } + // Then, permute only the column pointers + ei_declare_aligned_stack_constructed_variable(Index,outerIndexPtr,mat.cols()+1,mat.isCompressed()?const_cast(mat.outerIndexPtr()):0); + + // If the input matrix 'mat' is uncompressed, then the outer-indices do not match the ones of m_mat, and a copy is thus needed. + if(!mat.isCompressed()) + IndexVector::Map(outerIndexPtr, mat.cols()+1) = IndexVector::Map(m_mat.outerIndexPtr(),mat.cols()+1); + + // Apply the permutation and compute the nnz per column. for (Index i = 0; i < mat.cols(); i++) { m_mat.outerIndexPtr()[m_perm_c.indices()(i)] = outerIndexPtr[i]; m_mat.innerNonZeroPtr()[m_perm_c.indices()(i)] = outerIndexPtr[i+1] - outerIndexPtr[i]; } - if(!mat.isCompressed()) delete[] outerIndexPtr; } + // Compute the column elimination tree of the permuted matrix IndexVector firstRowElt; internal::coletree(m_mat, m_etree,firstRowElt); From 893bfcf95fd7aaa49747f029bac708a02a1526cd Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 6 Oct 2014 11:54:30 +0200 Subject: [PATCH 0871/1103] bug #887: use ei_declare_aligned_stack_constructed_variable instead of manual new[]/delete[] pairs in AMD and Paralellizer --- Eigen/src/Core/products/Parallelizer.h | 4 +--- Eigen/src/OrderingMethods/Amd.h | 5 ++--- 2 files changed, 3 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/products/Parallelizer.h b/Eigen/src/Core/products/Parallelizer.h index 4079063eb..126cfbbff 100644 --- a/Eigen/src/Core/products/Parallelizer.h +++ b/Eigen/src/Core/products/Parallelizer.h @@ -129,7 +129,7 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos Index blockRows = (rows / threads); blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr; - GemmParallelInfo* info = new GemmParallelInfo[threads]; + ei_declare_aligned_stack_constructed_variable(GemmParallelInfo,info,threads,0); #pragma omp parallel num_threads(threads) { @@ -146,8 +146,6 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos if(transpose) func(c0, actualBlockCols, 0, rows, info); else func(0, rows, c0, actualBlockCols, info); } - - delete[] info; #endif } diff --git a/Eigen/src/OrderingMethods/Amd.h b/Eigen/src/OrderingMethods/Amd.h index 41b4fd7e3..ce7c0bbf3 100644 --- a/Eigen/src/OrderingMethods/Amd.h +++ b/Eigen/src/OrderingMethods/Amd.h @@ -106,7 +106,8 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation t = cnz + cnz/5 + 2*n; /* add elbow room to C */ C.resizeNonZeros(t); - Index* W = new Index[8*(n+1)]; /* get workspace */ + // get workspace + ei_declare_aligned_stack_constructed_variable(Index,W,8*(n+1),0); Index* len = W; Index* nv = W + (n+1); Index* next = W + 2*(n+1); @@ -424,8 +425,6 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation } perm.indices().conservativeResize(n); - - delete[] W; } } // namespace internal From d44d432baa142fdbe17f9d3abeab2c7629e199b8 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 6 Oct 2014 16:11:26 +0200 Subject: [PATCH 0872/1103] Re-enable products with triangular views of sparse matrices: we simply have to treat them as a sparse matrix. --- Eigen/src/Core/TriangularMatrix.h | 8 ++++--- .../ConservativeSparseSparseProduct.h | 2 ++ Eigen/src/SparseCore/SparseDenseProduct.h | 12 +++++++++- Eigen/src/SparseCore/SparseProduct.h | 13 +++++++++++ test/sparse_product.cpp | 22 +++++++++++++++---- 5 files changed, 49 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 36bbd46e1..263d54485 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -247,7 +247,7 @@ template class TriangularView inline const AdjointReturnType adjoint() const { return AdjointReturnType(m_matrix.adjoint()); } - typedef TriangularView,TransposeMode> TransposeReturnType; + typedef TriangularView TransposeReturnType; /** \sa MatrixBase::transpose() */ EIGEN_DEVICE_FUNC inline TransposeReturnType transpose() @@ -255,11 +255,13 @@ template class TriangularView EIGEN_STATIC_ASSERT_LVALUE(MatrixType) return TransposeReturnType(m_matrix.const_cast_derived().transpose()); } + + typedef TriangularView ConstTransposeReturnType; /** \sa MatrixBase::transpose() const */ EIGEN_DEVICE_FUNC - inline const TransposeReturnType transpose() const + inline const ConstTransposeReturnType transpose() const { - return TransposeReturnType(m_matrix.transpose()); + return ConstTransposeReturnType(m_matrix.transpose()); } template diff --git a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h index 19d9eaa42..a30522ff7 100644 --- a/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h +++ b/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h @@ -141,6 +141,8 @@ struct conservative_sparse_sparse_product_selector ColMajorMatrixAux; typedef typename sparse_eval::type ColMajorMatrix; + // If the result is tall and thin (in the extreme case a column vector) + // then it is faster to sort the coefficients inplace instead of transposing twice. // FIXME, the following heuristic is probably not very good. if(lhs.rows()>=rhs.cols()) { diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index ed3cb1990..5aea11425 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -146,6 +146,11 @@ struct generic_product_impl } }; +template +struct generic_product_impl + : generic_product_impl +{}; + template struct generic_product_impl { @@ -158,12 +163,17 @@ struct generic_product_impl RhsNested rhsNested(rhs); dst.setZero(); - // transpoe everything + // transpose everything Transpose dstT(dst); internal::sparse_time_dense_product(rhsNested.transpose(), lhsNested.transpose(), dstT, typename Dest::Scalar(1)); } }; +template +struct generic_product_impl + : generic_product_impl +{}; + template struct sparse_dense_outer_product_evaluator { diff --git a/Eigen/src/SparseCore/SparseProduct.h b/Eigen/src/SparseCore/SparseProduct.h index ae707376a..c62386ed1 100644 --- a/Eigen/src/SparseCore/SparseProduct.h +++ b/Eigen/src/SparseCore/SparseProduct.h @@ -33,6 +33,7 @@ SparseMatrixBase::operator*(const SparseMatrixBase &other namespace internal { +// sparse * sparse template struct generic_product_impl { @@ -48,6 +49,18 @@ struct generic_product_impl } }; +// sparse * sparse-triangular +template +struct generic_product_impl + : public generic_product_impl +{}; + +// sparse-triangular * sparse +template +struct generic_product_impl + : public generic_product_impl +{}; + template struct evaluator > > : public evaluator::PlainObject>::type diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index fa9be5440..366e27274 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -194,7 +194,7 @@ template void sparse_product() VERIFY_IS_APPROX(d3=d1*m2.transpose(), refM3=d1*refM2.transpose()); } - // test self adjoint products + // test self-adjoint and traingular-view products { DenseMatrix b = DenseMatrix::Random(rows, rows); DenseMatrix x = DenseMatrix::Random(rows, rows); @@ -202,9 +202,12 @@ template void sparse_product() DenseMatrix refUp = DenseMatrix::Zero(rows, rows); DenseMatrix refLo = DenseMatrix::Zero(rows, rows); DenseMatrix refS = DenseMatrix::Zero(rows, rows); + DenseMatrix refA = DenseMatrix::Zero(rows, rows); SparseMatrixType mUp(rows, rows); SparseMatrixType mLo(rows, rows); SparseMatrixType mS(rows, rows); + SparseMatrixType mA(rows, rows); + initSparse(density, refA, mA); do { initSparse(density, refUp, mUp, ForceRealDiag|/*ForceNonZeroDiag|*/MakeUpperTriangular); } while (refUp.isZero()); @@ -224,19 +227,30 @@ template void sparse_product() VERIFY_IS_APPROX(mS, refS); VERIFY_IS_APPROX(x=mS*b, refX=refS*b); + // sparse selfadjointView with dense matrices VERIFY_IS_APPROX(x=mUp.template selfadjointView()*b, refX=refS*b); VERIFY_IS_APPROX(x=mLo.template selfadjointView()*b, refX=refS*b); VERIFY_IS_APPROX(x=mS.template selfadjointView()*b, refX=refS*b); - // sparse selfadjointView * sparse + // sparse selfadjointView with sparse matrices SparseMatrixType mSres(rows,rows); VERIFY_IS_APPROX(mSres = mLo.template selfadjointView()*mS, refX = refLo.template selfadjointView()*refS); - // sparse * sparse selfadjointview VERIFY_IS_APPROX(mSres = mS * mLo.template selfadjointView(), refX = refS * refLo.template selfadjointView()); + + // sparse triangularView with dense matrices + VERIFY_IS_APPROX(x=mA.template triangularView()*b, refX=refA.template triangularView()*b); + VERIFY_IS_APPROX(x=mA.template triangularView()*b, refX=refA.template triangularView()*b); + VERIFY_IS_APPROX(x=b*mA.template triangularView(), refX=b*refA.template triangularView()); + VERIFY_IS_APPROX(x=b*mA.template triangularView(), refX=b*refA.template triangularView()); + + // sparse triangularView with sparse matrices + VERIFY_IS_APPROX(mSres = mA.template triangularView()*mS, refX = refA.template triangularView()*refS); + VERIFY_IS_APPROX(mSres = mS * mA.template triangularView(), refX = refS * refA.template triangularView()); + VERIFY_IS_APPROX(mSres = mA.template triangularView()*mS, refX = refA.template triangularView()*refS); + VERIFY_IS_APPROX(mSres = mS * mA.template triangularView(), refX = refS * refA.template triangularView()); } - } // New test for Bug in SparseTimeDenseProduct From dbdd8b0883ebd1ebafeb21640d6e4bead1999565 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 6 Oct 2014 19:35:57 +0200 Subject: [PATCH 0873/1103] D&C SVD: add scaling to avoid overflow, fix handling of fixed size matrices --- test/svd_common.h | 6 ++--- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 35 +++++++++++++++------------ unsupported/test/bdcsvd.cpp | 10 ++++---- 3 files changed, 28 insertions(+), 23 deletions(-) diff --git a/test/svd_common.h b/test/svd_common.h index 4631939e5..b880efce5 100644 --- a/test/svd_common.h +++ b/test/svd_common.h @@ -301,7 +301,7 @@ void svd_underoverflow() 0, 5.60844e-313; SVD_DEFAULT(Matrix2d) svd; svd.compute(M,ComputeFullU|ComputeFullV); - svd_check_full(M,svd); + CALL_SUBTEST( svd_check_full(M,svd) ); // Check all 2x2 matrices made with the following coefficients: VectorXd value_set(9); @@ -312,7 +312,7 @@ void svd_underoverflow() { M << value_set(id(0)), value_set(id(1)), value_set(id(2)), value_set(id(3)); svd.compute(M,ComputeFullU|ComputeFullV); - svd_check_full(M,svd); + CALL_SUBTEST( svd_check_full(M,svd) ); id(k)++; if(id(k)>=value_set.size()) @@ -336,7 +336,7 @@ void svd_underoverflow() SVD_DEFAULT(Matrix3d) svd3; svd3.compute(M3,ComputeFullU|ComputeFullV); // just check we don't loop indefinitely - svd_check_full(M3,svd3); + CALL_SUBTEST( svd_check_full(M3,svd3) ); } // void jacobisvd(const MatrixType& a = MatrixType(), bool pickrandom = true) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index b13ee415e..5bf9b0ae2 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -236,26 +236,29 @@ BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsign { allocate(matrix.rows(), matrix.cols(), computationOptions); using std::abs; - - //**** step 1 Bidiagonalization - MatrixType copy; - if (m_isTranspose) copy = matrix.adjoint(); - else copy = matrix; + //**** step 0 - Copy the input matrix and apply scaling to reduce over/under-flows + RealScalar scale = matrix.cwiseAbs().maxCoeff(); + if(scale==RealScalar(0)) scale = RealScalar(1); + MatrixX copy; + if (m_isTranspose) copy = matrix.adjoint()/scale; + else copy = matrix/scale; + + //**** step 1 - Bidiagonalization internal::UpperBidiagonalization bid(copy); - //**** step 2 Divide + //**** step 2 - Divide & Conquer m_naiveU.setZero(); m_naiveV.setZero(); m_computed.topRows(m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose(); m_computed.template bottomRows<1>().setZero(); divide(0, m_diagSize - 1, 0, 0, 0); - //**** step 3 copy + //**** step 3 - Copy singular values and vectors for (int i=0; i& BDCSVD::compute(const MatrixType& matrix, unsign break; } } -// std::cout << "m_naiveU\n" << m_naiveU << "\n\n"; -// std::cout << "m_naiveV\n" << m_naiveV << "\n\n"; +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "m_naiveU\n" << m_naiveU << "\n\n"; + std::cout << "m_naiveV\n" << m_naiveV << "\n\n"; +#endif if(m_isTranspose) copyUV(bid.householderV(), bid.householderU(), m_naiveV, m_naiveU); else copyUV(bid.householderU(), bid.householderV(), m_naiveU, m_naiveV); -// std::cout << "DONE\n"; + m_isInitialized = true; return *this; }// end compute @@ -569,13 +574,13 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec computeSingVecs(zhat, diag, singVals, shifts, mus, U, V); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE - std::cout << "U^T U: " << (U.transpose() * U - MatrixType(MatrixType::Identity(U.cols(),U.cols()))).norm() << "\n"; - std::cout << "V^T V: " << (V.transpose() * V - MatrixType(MatrixType::Identity(V.cols(),V.cols()))).norm() << "\n"; + std::cout << "U^T U: " << (U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() << "\n"; + std::cout << "V^T V: " << (V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() << "\n"; #endif #ifdef EIGEN_BDCSVD_SANITY_CHECKS - assert((U.transpose() * U - MatrixType(MatrixType::Identity(U.cols(),U.cols()))).norm() < 1e-14 * n); - assert((V.transpose() * V - MatrixType(MatrixType::Identity(V.cols(),V.cols()))).norm() < 1e-14 * n); + assert((U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() < 1e-14 * n); + assert((V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() < 1e-14 * n); assert(m_naiveU.allFinite()); assert(m_naiveV.allFinite()); assert(m_computed.allFinite()); diff --git a/unsupported/test/bdcsvd.cpp b/unsupported/test/bdcsvd.cpp index 4ad991522..95cdb6a2e 100644 --- a/unsupported/test/bdcsvd.cpp +++ b/unsupported/test/bdcsvd.cpp @@ -70,13 +70,13 @@ void test_bdcsvd() CALL_SUBTEST_7(( svd_verify_assert >(MatrixXf(10,12)) )); CALL_SUBTEST_8(( svd_verify_assert >(MatrixXcd(7,5)) )); -// svd_all_trivial_2x2(bdcsvd); -// svd_all_trivial_2x2(bdcsvd); + CALL_SUBTEST_1(( svd_all_trivial_2x2(bdcsvd) )); + CALL_SUBTEST_1(( svd_all_trivial_2x2(bdcsvd) )); for(int i = 0; i < g_repeat; i++) { -// CALL_SUBTEST_3(( bdcsvd() )); -// CALL_SUBTEST_4(( bdcsvd() )); -// CALL_SUBTEST_5(( bdcsvd >() )); + CALL_SUBTEST_3(( bdcsvd() )); + CALL_SUBTEST_4(( bdcsvd() )); + CALL_SUBTEST_5(( bdcsvd >() )); int r = internal::random(1, EIGEN_TEST_MAX_SIZE/2), c = internal::random(1, EIGEN_TEST_MAX_SIZE/2); From 503c176d8edc6e2877b2ade7b3ed3a5ae17cf4e7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 7 Oct 2014 09:53:27 +0200 Subject: [PATCH 0874/1103] Fix missing outer() member in DynamicSparseMatrix --- unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h index e4dc1c1de..976f9f270 100644 --- a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h +++ b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h @@ -331,6 +331,7 @@ class DynamicSparseMatrix::InnerIterator : public Sparse inline Index row() const { return IsRowMajor ? m_outer : Base::index(); } inline Index col() const { return IsRowMajor ? Base::index() : m_outer; } + inline Index outer() const { return m_outer; } protected: const Index m_outer; @@ -347,6 +348,7 @@ class DynamicSparseMatrix::ReverseInnerIterator : public inline Index row() const { return IsRowMajor ? m_outer : Base::index(); } inline Index col() const { return IsRowMajor ? Base::index() : m_outer; } + inline Index outer() const { return m_outer; } protected: const Index m_outer; From 118b1113d9a4fa1a263597cdd699e237e5f2b4ac Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 7 Oct 2014 09:53:39 +0200 Subject: [PATCH 0875/1103] Workaround MSVC issue. --- .../Eigen/src/KroneckerProduct/KroneckerTensorProduct.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index b459360df..446fcac16 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -165,8 +165,8 @@ void KroneckerProductSparse::evalTo(Dest& dst) const const Rhs1 rhs1(m_B); // 2 - construct respective iterators - typedef InnerIterator LhsInnerIterator; - typedef InnerIterator RhsInnerIterator; + typedef Eigen::InnerIterator LhsInnerIterator; + typedef Eigen::InnerIterator RhsInnerIterator; // compute number of non-zeros per innervectors of dst { From 57413492948d4c9dd6572f5bde6720b80122054a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 7 Oct 2014 18:29:28 +0200 Subject: [PATCH 0876/1103] bug #882: fix various const-correctness issues with *View classes. --- Eigen/src/CholmodSupport/CholmodSupport.h | 2 +- Eigen/src/Core/CwiseUnaryView.h | 5 +++-- Eigen/src/Core/SelfAdjointView.h | 4 +++- Eigen/src/Core/TriangularMatrix.h | 18 ++++++++---------- Eigen/src/SparseCholesky/SimplicialCholesky.h | 8 ++++---- Eigen/src/SparseCore/SparseMatrixBase.h | 4 ++-- Eigen/src/SparseCore/SparseSelfAdjointView.h | 4 ++-- Eigen/src/SparseCore/SparseTriangularView.h | 4 ++-- 8 files changed, 25 insertions(+), 24 deletions(-) diff --git a/Eigen/src/CholmodSupport/CholmodSupport.h b/Eigen/src/CholmodSupport/CholmodSupport.h index 0c3b86dd2..3eadb83a0 100644 --- a/Eigen/src/CholmodSupport/CholmodSupport.h +++ b/Eigen/src/CholmodSupport/CholmodSupport.h @@ -105,7 +105,7 @@ const cholmod_sparse viewAsCholmod(const SparseMatrix<_Scalar,_Options,_Index>& /** Returns a view of the Eigen sparse matrix \a mat as Cholmod sparse matrix. * The data are not copied but shared. */ template -cholmod_sparse viewAsCholmod(const SparseSelfAdjointView, UpLo>& mat) +cholmod_sparse viewAsCholmod(const SparseSelfAdjointView, UpLo>& mat) { cholmod_sparse res = viewAsCholmod(mat.matrix().const_cast_derived()); diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index 61fd8ee35..6384dfdc3 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -37,7 +37,8 @@ struct traits > typedef typename MatrixType::Nested MatrixTypeNested; typedef typename remove_all::type _MatrixTypeNested; enum { - Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | LvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions + FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, + Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions MatrixTypeInnerStride = inner_stride_at_compile_time::ret, // need to cast the sizeof's from size_t to int explicitly, otherwise: // "error: no integral type can represent all of the enumerator values @@ -63,7 +64,7 @@ class CwiseUnaryView : public CwiseUnaryViewImpl::type NestedExpression; - explicit inline CwiseUnaryView(const MatrixType& mat, const ViewOp& func = ViewOp()) + explicit inline CwiseUnaryView(MatrixType& mat, const ViewOp& func = ViewOp()) : m_matrix(mat), m_functor(func) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryView) diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index f5fbd7215..1c44d9c9a 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -38,7 +38,8 @@ struct traits > : traits typedef typename MatrixType::PlainObject FullMatrixType; enum { Mode = UpLo | SelfAdjoint, - Flags = MatrixTypeNestedCleaned::Flags & (HereditaryBits) + FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, + Flags = MatrixTypeNestedCleaned::Flags & (HereditaryBits|FlagsLvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved }; }; @@ -95,6 +96,7 @@ template class SelfAdjointView EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { + EIGEN_STATIC_ASSERT_LVALUE(SelfAdjointView); Base::check_coordinates_internal(row, col); return m_matrix.const_cast_derived().coeffRef(row, col); } diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 263d54485..055ed7514 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -173,7 +173,8 @@ struct traits > : traits typedef MatrixType ExpressionType; enum { Mode = _Mode, - Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | LvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) + FlagsLvalueBit = is_lvalue::value ? LvalueBit : 0, + Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | FlagsLvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit))) }; }; } @@ -213,7 +214,7 @@ template class TriangularView // FIXME This, combined with const_cast_derived in transpose() leads to a const-correctness loophole EIGEN_DEVICE_FUNC - explicit inline TriangularView(const MatrixType& matrix) : m_matrix(matrix) + explicit inline TriangularView(MatrixType& matrix) : m_matrix(matrix) {} using Base::operator=; @@ -229,14 +230,9 @@ template class TriangularView const NestedExpression& nestedExpression() const { return m_matrix; } EIGEN_DEVICE_FUNC NestedExpression& nestedExpression() { return *const_cast(&m_matrix); } - - typedef TriangularView ConjugateReturnType; - /** \sa MatrixBase::conjugate() */ - EIGEN_DEVICE_FUNC - inline ConjugateReturnType conjugate() - { return ConjugateReturnType(m_matrix.conjugate()); } + /** \sa MatrixBase::conjugate() const */ - + typedef TriangularView ConjugateReturnType; EIGEN_DEVICE_FUNC inline const ConjugateReturnType conjugate() const { return ConjugateReturnType(m_matrix.conjugate()); } @@ -253,7 +249,8 @@ template class TriangularView inline TransposeReturnType transpose() { EIGEN_STATIC_ASSERT_LVALUE(MatrixType) - return TransposeReturnType(m_matrix.const_cast_derived().transpose()); + typename MatrixType::TransposeReturnType tmp(m_matrix.const_cast_derived()); + return TransposeReturnType(tmp); } typedef TriangularView ConstTransposeReturnType; @@ -392,6 +389,7 @@ template class TriangularViewImpl<_Mat EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { + EIGEN_STATIC_ASSERT_LVALUE(TriangularViewType); Base::check_coordinates_internal(row, col); return derived().nestedExpression().const_cast_derived().coeffRef(row, col); } diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index 0e8fa6628..918a34e13 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -237,8 +237,8 @@ template struct traits CholMatrixType; - typedef TriangularView MatrixL; - typedef TriangularView MatrixU; + typedef TriangularView MatrixL; + typedef TriangularView MatrixU; static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); } static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); } }; @@ -251,8 +251,8 @@ template struct traits CholMatrixType; - typedef TriangularView MatrixL; - typedef TriangularView MatrixU; + typedef TriangularView MatrixL; + typedef TriangularView MatrixU; static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); } static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); } }; diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index f96042f27..04baabe4f 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -297,9 +297,9 @@ template class SparseMatrixBase : public EigenBase Derived& operator*=(const SparseMatrixBase& other); template - inline const TriangularView triangularView() const; + inline const TriangularView triangularView() const; - template inline const SparseSelfAdjointView selfadjointView() const; + template inline const SparseSelfAdjointView selfadjointView() const; template inline SparseSelfAdjointView selfadjointView(); template Scalar dot(const MatrixBase& other) const; diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index 247d4e6d3..5da7d2bef 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -170,9 +170,9 @@ template class SparseSelfAdjointView template template -const SparseSelfAdjointView SparseMatrixBase::selfadjointView() const +const SparseSelfAdjointView SparseMatrixBase::selfadjointView() const { - return SparseSelfAdjointView(derived()); + return SparseSelfAdjointView(derived()); } template diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index e200bc815..1f5e53155 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -266,10 +266,10 @@ protected: template template -inline const TriangularView +inline const TriangularView SparseMatrixBase::triangularView() const { - return TriangularView(derived()); + return TriangularView(derived()); } } // end namespace Eigen From 4b886e6b39e4fd82365830b1dfa1bdc75a36f302 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 8 Oct 2014 07:48:30 +0200 Subject: [PATCH 0877/1103] bug #889: fix protected typedef --- Eigen/src/IterativeLinearSolvers/SolveWithGuess.h | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h index 6cd4872dc..251c6fa1a 100644 --- a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h +++ b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h @@ -42,6 +42,7 @@ class SolveWithGuess : public internal::generic_xpr_base::Scalar Scalar; typedef typename internal::traits::PlainObject PlainObject; typedef typename internal::generic_xpr_base, MatrixXpr, typename internal::traits::StorageKind>::type Base; @@ -61,8 +62,6 @@ protected: const RhsType &m_rhs; const GuessType &m_guess; - typedef typename internal::traits::Scalar Scalar; - private: Scalar coeff(Index row, Index col) const; Scalar coeff(Index i) const; From 6c047d398daba5784da35d3b502360a5a7a83f33 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 8 Oct 2014 13:29:36 -0700 Subject: [PATCH 0878/1103] Fixed a comment --- unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h index d42167da9..4d7f9e1fd 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h @@ -18,7 +18,7 @@ namespace Eigen { * \brief The fixed sized version of the tensor class. * * The fixes sized equivalent of - * Eigen::Tensor t(3, 5, 7); + * Eigen::Tensor t(3, 5, 7); * is * Eigen::TensorFixedSize> t; */ From 0a07ac574ead83d314d518127d8d69595f6212b2 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 8 Oct 2014 13:32:41 -0700 Subject: [PATCH 0879/1103] Added support for the *= and /* operators to TensorBase --- unsupported/Eigen/CXX11/src/Tensor/TensorBase.h | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 2f7c9ecda..90a9cc2c4 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -307,11 +307,18 @@ class TensorBase : public TensorBase, const Derived, const OtherDerived>(derived(), other.derived()); } - template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator-=(const OtherDerived& other) { return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator*=(const OtherDerived& other) { + return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Derived& operator/=(const OtherDerived& other) { + return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReshapingOp From 44beee9d68e13dc299c6e2ea321aedc74c23d039 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 8 Oct 2014 14:14:20 -0700 Subject: [PATCH 0880/1103] Removed dead code --- unsupported/Eigen/CXX11/src/Tensor/TensorMap.h | 3 --- 1 file changed, 3 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index 417717b90..04849dd9f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -12,9 +12,6 @@ namespace Eigen { -template class Stride; - - /** \class TensorMap * \ingroup CXX11_Tensor_Module * From ccd70ba123562b77f132c562f76d1286af7d5fe3 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 9 Oct 2014 23:29:01 +0200 Subject: [PATCH 0881/1103] Various numerical fixes in D&C SVD: I cannot make it fail with double, but still need to tune for single precision, and carefully test with duplicated singular values --- test/svd_common.h | 6 +- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 73 +++++++++++-------- unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt | 34 +++------ unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt | 15 +--- unsupported/test/bdcsvd.cpp | 3 +- 5 files changed, 58 insertions(+), 73 deletions(-) diff --git a/test/svd_common.h b/test/svd_common.h index b880efce5..e902d2320 100644 --- a/test/svd_common.h +++ b/test/svd_common.h @@ -146,6 +146,7 @@ void svd_min_norm(const MatrixType& m, unsigned int computationOptions) m2.setRandom(); } while(SVD_FOR_MIN_NORM(MatrixType2)(m2).setThreshold(test_precision()).rank()!=rank && (++guard)<10); VERIFY(guard<10); + RhsType2 rhs2 = RhsType2::Random(rank); // use QR to find a reference minimal norm solution HouseholderQR qr(m2.adjoint()); @@ -159,7 +160,7 @@ void svd_min_norm(const MatrixType& m, unsigned int computationOptions) VERIFY_IS_APPROX(m2*x21, rhs2); VERIFY_IS_APPROX(m2*x22, rhs2); VERIFY_IS_APPROX(x21, x22); - + // Now check with a rank deficient matrix typedef Matrix MatrixType3; typedef Matrix RhsType3; @@ -172,7 +173,6 @@ void svd_min_norm(const MatrixType& m, unsigned int computationOptions) VERIFY_IS_APPROX(m3*x3, rhs3); VERIFY_IS_APPROX(m3*x21, rhs3); VERIFY_IS_APPROX(m2*x3, rhs2); - VERIFY_IS_APPROX(x21, x3); } @@ -209,7 +209,7 @@ void svd_test_all_computation_options(const MatrixType& m, bool full_only) CALL_SUBTEST(( svd_least_square(m, ComputeFullU | ComputeThinV) )); CALL_SUBTEST(( svd_least_square(m, ComputeThinU | ComputeFullV) )); CALL_SUBTEST(( svd_least_square(m, ComputeThinU | ComputeThinV) )); - + CALL_SUBTEST(( svd_min_norm(m, ComputeFullU | ComputeThinV) )); CALL_SUBTEST(( svd_min_norm(m, ComputeThinU | ComputeFullV) )); CALL_SUBTEST(( svd_min_norm(m, ComputeThinU | ComputeThinV) )); diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index 5bf9b0ae2..d5e8140a4 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -272,8 +272,8 @@ BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsign } } #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE - std::cout << "m_naiveU\n" << m_naiveU << "\n\n"; - std::cout << "m_naiveV\n" << m_naiveV << "\n\n"; +// std::cout << "m_naiveU\n" << m_naiveU << "\n\n"; +// std::cout << "m_naiveV\n" << m_naiveV << "\n\n"; #endif if(m_isTranspose) copyUV(bid.householderV(), bid.householderU(), m_naiveV, m_naiveU); else copyUV(bid.householderU(), bid.householderV(), m_naiveU, m_naiveV); @@ -612,22 +612,23 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia Index n = col0.size(); Index actual_n = n; while(actual_n>1 && col0(actual_n-1)==0) --actual_n; - Index m = 0; - Array perm(actual_n); - { - for(Index k=0;k perm(actual_n); +// { +// for(Index k=0;k::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // otherwise, use secular equation to find singular value RealScalar left = diag(k); - RealScalar right = (k != actual_n-1) ? diag(k+1) : (diag(actual_n-1) + col0.matrix().norm()); + RealScalar right; // was: = (k != actual_n-1) ? diag(k+1) : (diag(actual_n-1) + col0.matrix().norm()); + if(k==actual_n-1) + right = (diag(actual_n-1) + col0.matrix().norm()); + else + { + // Skip deflated singular values + Index l = k+1; + while(col0(l)==0) { ++l; eigen_internal_assert(l::computeSingVals(const ArrayXr& col0, const ArrayXr& dia { muPrev = (right - left) * 0.1; if (k == actual_n-1) muCur = right - left; - else muCur = (right - left) * 0.5; + else muCur = (right - left) * 0.5; } else { @@ -671,7 +681,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // rational interpolation: fit a function of the form a / mu + b through the two previous // iterates and use its zero to compute the next iterate - bool useBisection = false; + bool useBisection = fPrev*fCur>0; while (abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (max)(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) { ++m_numIters; @@ -684,32 +694,36 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia muCur = -a / b; fCur = secularEq(muCur, col0, diag, diagShifted, shift, actual_n); - if (shift == left && (muCur < 0 || muCur > right - left)) useBisection = true; + if (shift == left && (muCur < 0 || muCur > right - left)) useBisection = true; if (shift == right && (muCur < -(right - left) || muCur > 0)) useBisection = true; } // fall back on bisection method if rational interpolation did not work if (useBisection) { +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "useBisection for k = " << k << ", actual_n = " << actual_n << "\n"; +#endif RealScalar leftShifted, rightShifted; if (shift == left) { - leftShifted = 1e-30; - if (k == 0) rightShifted = right - left; - else rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe + leftShifted = RealScalar(1)/NumTraits::highest(); + // I don't understand why the case k==0 would be special there: + // if (k == 0) rightShifted = right - left; else + rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe } else { leftShifted = -(right - left) * 0.6; - rightShifted = -1e-30; + rightShifted = -RealScalar(1)/NumTraits::highest(); } RealScalar fLeft = secularEq(leftShifted, col0, diag, diagShifted, shift, actual_n); RealScalar fRight = secularEq(rightShifted, col0, diag, diagShifted, shift, actual_n); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE - if(fLeft * fRight>0) - std::cout << fLeft << " * " << fRight << " == " << fLeft * fRight << " ; " << left << " - " << right << " -> " << leftShifted << " " << rightShifted << " shift=" << shift << "\n"; + if(fLeft * fRight>=0) + std::cout << k << " : " << fLeft << " * " << fRight << " == " << fLeft * fRight << " ; " << left << " - " << right << " -> " << leftShifted << " " << rightShifted << " shift=" << shift << "\n"; #endif eigen_internal_assert(fLeft * fRight < 0); @@ -738,8 +752,9 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // perturb singular value slightly if it equals diagonal entry to avoid division by zero later // (deflation is supposed to avoid this from happening) - if (singVals[k] == left) singVals[k] *= 1 + NumTraits::epsilon(); - if (singVals[k] == right) singVals[k] *= 1 - NumTraits::epsilon(); + // - this does no seem to be necessary anymore - +// if (singVals[k] == left) singVals[k] *= 1 + NumTraits::epsilon(); +// if (singVals[k] == right) singVals[k] *= 1 - NumTraits::epsilon(); } } @@ -989,7 +1004,7 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index #endif deflation43(firstCol, shift, i, length); } - + #ifdef EIGEN_BDCSVD_SANITY_CHECKS assert(m_naiveU.allFinite()); assert(m_naiveV.allFinite()); @@ -1027,7 +1042,7 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index realInd[pos] = pos; } - for(Index i = 1; i < length - 1; i++) + for(Index i = 1; i < length; i++) { const Index pi = permutation[length - i]; const Index J = realCol[pi]; @@ -1056,7 +1071,7 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index #ifdef EIGEN_BDCSVD_SANITY_CHECKS for(int k=2;k, ...) +- To solve the secular equation using FMM: +http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf -(optional optimization) - do all the allocations in the allocate part - - support static matrices - - return a error at compilation time when using integer matrices (int, long, std::complex, ...) - -to finish the algorithm : - -implement the last part of the algorithm as described on the reference paper. - You may find more information on that part on this paper - - -to replace the call to JacobiSVD at the end of the divide algorithm, just after the call to - deflation. - -(suggested step by step resolution) - 0) comment the call to Jacobi in the last part of the divide method and everything right after - until the end of the method. What is commented can be a guideline to steps 3) 4) and 6) - 1) solve the secular equation (Characteristic equation) on the values that are not null (zi!=0 and di!=0), after the deflation - wich should be uncommented in the divide method - 2) remember the values of the singular values that are already computed (zi=0) - 3) assign the singular values found in m_computed at the right places (with the ones found in step 2) ) - in decreasing order - 4) set the firstcol to zero (except the first element) in m_computed - 5) compute all the singular vectors when CompV is set to true and only the left vectors when - CompV is set to false - 6) multiply naiveU and naiveV to the right by the matrices found, only naiveU when CompV is set to - false, /!\ if CompU is false NaiveU has only 2 rows - 7) delete everything commented in step 0) diff --git a/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt b/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt index 8563ddab8..29ab9cd40 100644 --- a/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt +++ b/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt @@ -3,19 +3,6 @@ This unsupported package is about a divide and conquer algorithm to compute SVD. The implementation follows as closely as possible the following reference paper : http://www.cs.yale.edu/publications/techreports/tr933.pdf -The code documentation uses the same names for variables as the reference paper. The code, deflation included, is -working but there are a few things that could be optimised as explained in the TODOBdsvd. - -In the code comments were put at the line where would be the third step of the algorithm so one could simply add the call -of a function doing the last part of the algorithm and that would not require any knowledge of the part we implemented. - -In the TODOBdcsvd we explain what is the main difficulty of the last part and suggest a reference paper to help solve it. - -The implemented has trouble with fixed size matrices. - -In the actual implementation, it returns matrices of zero when ask to do a svd on an int matrix. - - -Paper for the third part: +To solve the secular equation using FMM: http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf diff --git a/unsupported/test/bdcsvd.cpp b/unsupported/test/bdcsvd.cpp index 95cdb6a2e..9e5c29a8c 100644 --- a/unsupported/test/bdcsvd.cpp +++ b/unsupported/test/bdcsvd.cpp @@ -21,8 +21,7 @@ #define SVD_DEFAULT(M) BDCSVD -// #define SVD_FOR_MIN_NORM(M) BDCSVD -#define SVD_FOR_MIN_NORM(M) JacobiSVD +#define SVD_FOR_MIN_NORM(M) BDCSVD #include "../../test/svd_common.h" // Check all variants of JacobiSVD From a48b82eece435e24418beb95f16ab32be7970183 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 9 Oct 2014 23:34:05 +0200 Subject: [PATCH 0882/1103] Add a scoped_array helper class to handle locally allocated/used arrays --- Eigen/src/Core/util/Memory.h | 26 +++++++++++++++++++++++++- 1 file changed, 25 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index af46c449c..6ed158972 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -618,7 +618,7 @@ template struct smart_memmove_helper { // This helper class construct the allocated memory, and takes care of destructing and freeing the handled data // at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions. -template class aligned_stack_memory_handler +template class aligned_stack_memory_handler : noncopyable { public: /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size. @@ -646,6 +646,30 @@ template class aligned_stack_memory_handler bool m_deallocate; }; +template class scoped_array : noncopyable +{ + T* m_ptr; +public: + explicit scoped_array(std::ptrdiff_t size) + { + m_ptr = new T[size]; + } + ~scoped_array() + { + delete[] m_ptr; + } + T& operator[](std::ptrdiff_t i) { return m_ptr[i]; } + const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; } + T* &ptr() { return m_ptr; } + const T* ptr() const { return m_ptr; } + operator const T*() const { return m_ptr; } +}; + +template void swap(scoped_array &a,scoped_array &b) +{ + std::swap(a.ptr(),b.ptr()); +} + } // end namespace internal /** \internal From 538c059aa4d02a9f1a282e8fc4b181c8eb0b4f47 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 9 Oct 2014 23:35:05 +0200 Subject: [PATCH 0883/1103] bug #887: fix CompressedStorage::reallocate wrt memory leaks --- Eigen/src/SparseCore/CompressedStorage.h | 17 +++++++---------- 1 file changed, 7 insertions(+), 10 deletions(-) diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index a93550340..4f3695773 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -204,17 +204,14 @@ class CompressedStorage inline void reallocate(size_t size) { - Scalar* newValues = new Scalar[size]; - Index* newIndices = new Index[size]; + eigen_internal_assert(size!=m_allocatedSize); + internal::scoped_array newValues(size); + internal::scoped_array newIndices(size); size_t copySize = (std::min)(size, m_size); - // copy - internal::smart_copy(m_values, m_values+copySize, newValues); - internal::smart_copy(m_indices, m_indices+copySize, newIndices); - // delete old stuff - delete[] m_values; - delete[] m_indices; - m_values = newValues; - m_indices = newIndices; + internal::smart_copy(m_values, m_values+copySize, newValues.ptr()); + internal::smart_copy(m_indices, m_indices+copySize, newIndices.ptr()); + std::swap(m_values,newValues.ptr()); + std::swap(m_indices,newIndices.ptr()); m_allocatedSize = size; } From 48d537f59fd3c7efd954bb99581dac83fd1c0f16 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 9 Oct 2014 23:35:26 +0200 Subject: [PATCH 0884/1103] Fix indentation --- Eigen/src/Core/util/Memory.h | 28 ++++++++++++++-------------- 1 file changed, 14 insertions(+), 14 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 6ed158972..7b39285ea 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -578,27 +578,27 @@ template struct smart_memmove_helper; template void smart_memmove(const T* start, const T* end, T* target) { - smart_memmove_helper::RequireInitialization>::run(start, end, target); + smart_memmove_helper::RequireInitialization>::run(start, end, target); } template struct smart_memmove_helper { - static inline void run(const T* start, const T* end, T* target) - { std::memmove(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); } + static inline void run(const T* start, const T* end, T* target) + { std::memmove(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); } }; template struct smart_memmove_helper { - static inline void run(const T* start, const T* end, T* target) - { - if (uintptr_t(target) < uintptr_t(start)) - { - std::copy(start, end, target); - } - else - { - std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T); - std::copy_backward(start, end, target + count); - } + static inline void run(const T* start, const T* end, T* target) + { + if (uintptr_t(target) < uintptr_t(start)) + { + std::copy(start, end, target); } + else + { + std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T); + std::copy_backward(start, end, target + count); + } + } }; From 349c2c9235ce93b50bd8e38be3e876ee73442ccc Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 9 Oct 2014 23:35:49 +0200 Subject: [PATCH 0885/1103] bug #367: fix double copies in atWithInsertion, and add respective unit-test --- Eigen/src/SparseCore/CompressedStorage.h | 27 ++++++++++++++++++++---- test/sparse_vector.cpp | 18 ++++++++++++++-- 2 files changed, 39 insertions(+), 6 deletions(-) diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index 4f3695773..f408dd3a6 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -172,12 +172,31 @@ class CompressedStorage size_t id = searchLowerIndex(0,m_size,key); if (id>=m_size || m_indices[id]!=key) { - resize(m_size+1,1); - for (size_t j=m_size-1; j>id; --j) + if (m_allocatedSize newValues(m_allocatedSize); + internal::scoped_array newIndices(m_allocatedSize); + + // copy first chunk + internal::smart_copy(m_values, m_values +id, newValues.ptr()); + internal::smart_copy(m_indices, m_indices+id, newIndices.ptr()); + + // copy the rest + if(m_size>id) + { + internal::smart_copy(m_values +id, m_values +m_size, newValues.ptr() +id+1); + internal::smart_copy(m_indices+id, m_indices+m_size, newIndices.ptr()+id+1); + } + std::swap(m_values,newValues.ptr()); + std::swap(m_indices,newIndices.ptr()); } + else if(m_size>id) + { + internal::smart_memmove(m_values +id, m_values +m_size, m_values +id+1); + internal::smart_memmove(m_indices+id, m_indices+m_size, m_indices+id+1); + } + m_size++; m_indices[id] = key; m_values[id] = defaultValue; } diff --git a/test/sparse_vector.cpp b/test/sparse_vector.cpp index 5eea9edfd..5dc421976 100644 --- a/test/sparse_vector.cpp +++ b/test/sparse_vector.cpp @@ -23,8 +23,8 @@ template void sparse_vector(int rows, int cols) SparseVectorType v1(rows), v2(rows), v3(rows); DenseMatrix refM1 = DenseMatrix::Zero(rows, rows); DenseVector refV1 = DenseVector::Random(rows), - refV2 = DenseVector::Random(rows), - refV3 = DenseVector::Random(rows); + refV2 = DenseVector::Random(rows), + refV3 = DenseVector::Random(rows); std::vector zerocoords, nonzerocoords; initSparse(densityVec, refV1, v1, &zerocoords, &nonzerocoords); @@ -52,6 +52,20 @@ template void sparse_vector(int rows, int cols) } } VERIFY_IS_APPROX(v1, refV1); + + // test coeffRef with reallocation + { + SparseVectorType v1(rows); + DenseVector v2 = DenseVector::Zero(rows); + for(int k=0; k(0,rows-1); + Scalar v = internal::random(); + v1.coeffRef(i) += v; + v2.coeffRef(i) += v; + } + VERIFY_IS_APPROX(v1,v2); + } v1.coeffRef(nonzerocoords[0]) = Scalar(5); refV1.coeffRef(nonzerocoords[0]) = Scalar(5); From a80e17cfe86c5803965dd002c8bd9c5b80591135 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 9 Oct 2014 23:42:33 +0200 Subject: [PATCH 0886/1103] Remove unused and dangerous CompressedStorage::Map function --- Eigen/src/SparseCore/CompressedStorage.h | 9 --------- 1 file changed, 9 deletions(-) diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index f408dd3a6..2845e44e7 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -108,15 +108,6 @@ class CompressedStorage inline Index& index(size_t i) { return m_indices[i]; } inline const Index& index(size_t i) const { return m_indices[i]; } - static CompressedStorage Map(Index* indices, Scalar* values, size_t size) - { - CompressedStorage res; - res.m_indices = indices; - res.m_values = values; - res.m_allocatedSize = res.m_size = size; - return res; - } - /** \returns the largest \c k such that for all \c j in [0,k) index[\c j]\<\a key */ inline Index searchLowerIndex(Index key) const { From 767424af18a55604496f38dd4593542db97240a1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 9 Oct 2014 15:36:23 -0700 Subject: [PATCH 0887/1103] Improved the functors defined for standard reductions Added a functor to encapsulate the generation of random numbers on cpu and gpu. --- .../Eigen/CXX11/src/Tensor/TensorFunctors.h | 72 +++++++++++++++++-- unsupported/test/cxx11_tensor_reduction.cpp | 33 +++++++++ 2 files changed, 101 insertions(+), 4 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h index 92984336c..e9aa22183 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h @@ -25,12 +25,12 @@ template struct SumReducer } private: - T m_sum; + typename internal::remove_all::type m_sum; }; template struct MaxReducer { - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE MaxReducer() : m_max((std::numeric_limits::min)()) { } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE MaxReducer() : m_max(-(std::numeric_limits::max)()) { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t) { if (t > m_max) { m_max = t; } } @@ -39,7 +39,7 @@ template struct MaxReducer } private: - T m_max; + typename internal::remove_all::type m_max; }; template struct MinReducer @@ -53,9 +53,73 @@ template struct MinReducer } private: - T m_min; + typename internal::remove_all::type m_min; }; + +#if !defined (EIGEN_USE_GPU) || !defined(__CUDACC__) || !defined(__CUDA_ARCH__) +// We're not compiling a cuda kernel +template struct UniformRandomGenerator { + template + T operator()(Index, Index = 0) const { + return random(); + } + template + typename internal::packet_traits::type packetOp(Index, Index = 0) const { + const int packetSize = internal::packet_traits::size; + EIGEN_ALIGN_DEFAULT T values[packetSize]; + for (int i = 0; i < packetSize; ++i) { + values[i] = random(); + } + return internal::pload::type>(values); + } +}; + +#else + +// We're compiling a cuda kernel +template struct UniformRandomGenerator; + +template <> struct UniformRandomGenerator { + UniformRandomGenerator() { + const int tid = blockIdx.x * blockDim.x + threadIdx.x; + curand_init(0, tid, 0, &m_state); + } + + template + float operator()(Index, Index = 0) const { + return curand_uniform(&m_state); + } + template + float4 packetOp(Index, Index = 0) const { + return curand_uniform4(&m_state); + } + + private: + mutable curandStatePhilox4_32_10_t m_state; +}; + +template <> struct UniformRandomGenerator { + UniformRandomGenerator() { + const int tid = blockIdx.x * blockDim.x + threadIdx.x; + curand_init(0, tid, 0, &m_state); + } + template + double operator()(Index, Index = 0) const { + return curand_uniform_double(&m_state); + } + template + double2 packetOp(Index, Index = 0) const { + return curand_uniform2_double(&m_state); + } + + private: + mutable curandStatePhilox4_32_10_t m_state; +}; + +#endif + + } // end namespace internal } // end namespace Eigen diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp index 27135b982..da9885166 100644 --- a/unsupported/test/cxx11_tensor_reduction.cpp +++ b/unsupported/test/cxx11_tensor_reduction.cpp @@ -139,9 +139,42 @@ static void test_user_defined_reductions() } +static void test_tensor_maps() +{ + int inputs[2*3*5*7]; + TensorMap > tensor_map(inputs, 2,3,5,7); + TensorMap > tensor_map_const(inputs, 2,3,5,7); + const TensorMap > tensor_map_const_const(inputs, 2,3,5,7); + + tensor_map.setRandom(); + array reduction_axis; + reduction_axis[0] = 1; + reduction_axis[1] = 3; + + Tensor result = tensor_map.sum(reduction_axis); + Tensor result2 = tensor_map_const.sum(reduction_axis); + Tensor result3 = tensor_map_const_const.sum(reduction_axis); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 5; ++j) { + int sum = 0; + for (int k = 0; k < 3; ++k) { + for (int l = 0; l < 7; ++l) { + sum += tensor_map(i, k, j, l); + } + } + VERIFY_IS_EQUAL(result(i, j), sum); + VERIFY_IS_EQUAL(result2(i, j), sum); + VERIFY_IS_EQUAL(result3(i, j), sum); + } + } +} + + void test_cxx11_tensor_reduction() { CALL_SUBTEST(test_simple_reductions()); CALL_SUBTEST(test_full_reductions()); CALL_SUBTEST(test_user_defined_reductions()); + CALL_SUBTEST(test_tensor_maps()); } From 498b7eed25bdb3b90f2fc45dd822c96aa08db2f8 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 9 Oct 2014 15:39:13 -0700 Subject: [PATCH 0888/1103] Rewrote the TensorBase::random method to support the generation of random number on gpu. --- unsupported/Eigen/CXX11/src/Tensor/TensorBase.h | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 90a9cc2c4..d4b7846a0 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -39,9 +39,14 @@ class TensorBase } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> + EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> random() const { - return TensorCwiseNullaryOp, const Derived>(derived()); + return TensorCwiseNullaryOp, const Derived>(derived()); + } + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseNullaryOp + random() const { + return TensorCwiseNullaryOp(derived()); } // Coefficient-wise unary operators From a991f94c0e5c51555875564ce58681a82d07cd69 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 10 Oct 2014 15:20:37 -0700 Subject: [PATCH 0889/1103] Fixed the thread pool test --- test/main.h | 4 ++-- unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 2 +- unsupported/test/CMakeLists.txt | 2 +- unsupported/test/cxx11_tensor_thread_pool.cpp | 8 ++++---- 4 files changed, 8 insertions(+), 8 deletions(-) diff --git a/test/main.h b/test/main.h index b504970f3..9cb41c828 100644 --- a/test/main.h +++ b/test/main.h @@ -47,8 +47,8 @@ // protected by parenthesis against macro expansion, the min()/max() macros // are defined here and any not-parenthesized min/max call will cause a // compiler error. -#define min(A,B) please_protect_your_min_with_parentheses -#define max(A,B) please_protect_your_max_with_parentheses +//#define min(A,B) please_protect_your_min_with_parentheses +//#define max(A,B) please_protect_your_max_with_parentheses #define FORBIDDEN_IDENTIFIER (this_identifier_is_forbidden_to_avoid_clashes) this_identifier_is_forbidden_to_avoid_clashes // B0 is defined in POSIX header termios.h diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index faf965df8..84768ca09 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -131,7 +131,7 @@ class TensorExecutor const Index numblocks = size / blocksize; Index i = 0; - vector > results; + std::vector > results; results.reserve(numblocks); for (int i = 0; i < numblocks; ++i) { results.push_back(std::async(std::launch::async, &EvalRange::run, &evaluator, i*blocksize, (i+1)*blocksize)); diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 75423f516..1c4d0838a 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -122,5 +122,5 @@ if(EIGEN_TEST_CXX11) # ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") -# ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") + ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp index 2e67b2064..e02d8e4be 100644 --- a/unsupported/test/cxx11_tensor_thread_pool.cpp +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -17,9 +17,9 @@ using Eigen::Tensor; void test_cxx11_tensor_thread_pool() { - Eigen::Tensor in1(Eigen::array(2,3,7)); - Eigen::Tensor in2(Eigen::array(2,3,7)); - Eigen::Tensor out(Eigen::array(2,3,7)); + Eigen::Tensor in1(2,3,7); + Eigen::Tensor in2(2,3,7); + Eigen::Tensor out(2,3,7); in1.setRandom(); in2.setRandom(); @@ -30,7 +30,7 @@ void test_cxx11_tensor_thread_pool() for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f); + VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f); } } } From 4b36c3591f247d4be38e5a12dbed7ac0d1ad2bff Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 10 Oct 2014 15:43:21 -0700 Subject: [PATCH 0890/1103] Fixed the tensor shuffling test --- .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 133 +++++++++++++++++- .../Eigen/CXX11/src/Tensor/TensorExecutor.h | 8 +- unsupported/test/CMakeLists.txt | 2 +- unsupported/test/cxx11_tensor_fixed_size.cpp | 2 +- unsupported/test/cxx11_tensor_shuffling.cpp | 9 +- 5 files changed, 141 insertions(+), 13 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 11590b474..732c6b344 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -37,8 +37,7 @@ template struct IndexPair { Index second; }; - -// Boiler plate code +// Boilerplate code namespace internal { template struct dget { @@ -110,6 +109,11 @@ struct Sizes : internal::numeric_list { } }; +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_prod(const Sizes&) { + return Sizes::total_size; +} + #else template @@ -136,9 +140,21 @@ template Sizes(DenseIndex... indices) { } explicit Sizes(std::initializer_list l) { // todo: add assertion } +#else + EIGEN_DEVICE_FUNC explicit Sizes(const DenseIndex i0) { + } + EIGEN_DEVICE_FUNC explicit Sizes(const DenseIndex i0, const DenseIndex i1) { + } + EIGEN_DEVICE_FUNC explicit Sizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2) { + } + EIGEN_DEVICE_FUNC explicit Sizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3) { + } + EIGEN_DEVICE_FUNC explicit Sizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3, const DenseIndex i4) { + } #endif template Sizes& operator = (const T& other) { @@ -156,9 +172,14 @@ template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_prod(const Sizes&) { + return Sizes::total_size; +}; + #endif -// Boiler plate +// Boilerplate namespace internal { template struct tensor_index_linearization_helper @@ -243,6 +264,112 @@ struct DSizes : array { }; + + +// Boilerplate +namespace internal { +template +struct tensor_vsize_index_linearization_helper +{ + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index run(array const& indices, std::vector const& dimensions) + { + return array_get(indices) + + array_get(dimensions) * + tensor_vsize_index_linearization_helper::run(indices, dimensions); + } +}; + +template +struct tensor_vsize_index_linearization_helper +{ + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + Index run(array const& indices, std::vector const&) + { + return array_get(indices); + } +}; +} // end namespace internal + +template +struct VSizes : std::vector { + typedef std::vector Base; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() const { + return internal::array_prod(*static_cast(this)); + } + + EIGEN_DEVICE_FUNC VSizes() { } + EIGEN_DEVICE_FUNC explicit VSizes(const std::vector& a) : Base(a) { } + + template + EIGEN_DEVICE_FUNC explicit VSizes(const array& a) { + this->resize(NumDims); + for (int i = 0; i < NumDims; ++i) { + (*this)[i] = a[i]; + } + } + + EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0) { + this->resize(1); + (*this)[0] = i0; + } + EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1) { + this->resize(2); + (*this)[0] = i0; + (*this)[1] = i1; + } + EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2) { + this->resize(3); + (*this)[0] = i0; + (*this)[1] = i1; + (*this)[2] = i2; + } + EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3) { + this->resize(4); + (*this)[0] = i0; + (*this)[1] = i1; + (*this)[2] = i2; + (*this)[3] = i3; + } + EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3, const DenseIndex i4) { + this->resize(5); + (*this)[0] = i0; + (*this)[1] = i1; + (*this)[2] = i2; + (*this)[3] = i3; + (*this)[4] = i4; + } + + VSizes& operator = (const std::vector& other) { + *static_cast(this) = other; + return *this; + } + + // A constexpr would be so much better here + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfColMajor(const array& indices) const { + return internal::tensor_vsize_index_linearization_helper::run(indices, *static_cast(this)); + } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfRowMajor(const array& indices) const { + return internal::tensor_vsize_index_linearization_helper::run(indices, *static_cast(this)); + } +}; + + +// Boilerplate +namespace internal { +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex array_prod(const VSizes& sizes) { + DenseIndex total_size = 1; + for (int i = 0; i < sizes.size(); ++i) { + total_size *= sizes[i]; + } + return total_size; +} +} + namespace internal { template struct array_size > { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index 84768ca09..10f5a5ee7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -39,7 +39,7 @@ class TensorExecutor const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); if (needs_assign) { - const Index size = evaluator.dimensions().TotalSize(); + const Index size = array_prod(evaluator.dimensions()); for (Index i = 0; i < size; ++i) { evaluator.evalScalar(i); } @@ -60,7 +60,7 @@ class TensorExecutor const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); if (needs_assign) { - const Index size = evaluator.dimensions().TotalSize(); + const Index size = array_prod(evaluator.dimensions()); static const int PacketSize = unpacket_traits::PacketReturnType>::size; const int VectorizedSize = (size / PacketSize) * PacketSize; @@ -122,7 +122,7 @@ class TensorExecutor const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); if (needs_assign) { - const Index size = evaluator.dimensions().TotalSize(); + const Index size = array_prod(evaluator.dimensions()); static const int PacketSize = Vectorizable ? unpacket_traits::size : 1; @@ -176,7 +176,7 @@ class TensorExecutor const int num_blocks = getNumCudaMultiProcessors() * maxCudaThreadsPerMultiProcessor() / maxCudaThreadsPerBlock(); const int block_size = maxCudaThreadsPerBlock(); - const Index size = evaluator.dimensions().TotalSize(); + const Index size = array_prod(evaluator.dimensions()); EigenMetaKernel > <<>>(evaluator, size); assert(cudaGetLastError() == cudaSuccess); } diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 1c4d0838a..ac2ccaf27 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -119,7 +119,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_morphing "-std=c++0x") ei_add_test(cxx11_tensor_padding "-std=c++0x") ei_add_test(cxx11_tensor_reduction "-std=c++0x") -# ei_add_test(cxx11_tensor_shuffling "-std=c++0x") + ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_fixed_size.cpp b/unsupported/test/cxx11_tensor_fixed_size.cpp index d270486f2..b0501aaa3 100644 --- a/unsupported/test/cxx11_tensor_fixed_size.cpp +++ b/unsupported/test/cxx11_tensor_fixed_size.cpp @@ -179,7 +179,7 @@ static void test_array() for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { - VERIFY_IS_APPROX(mat3(array(i,j,k)), powf(val, 3.5f)); + VERIFY_IS_APPROX(mat3(i,j,k), powf(val, 3.5f)); val += 1.0; } } diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp index 5ab8b6821..39c623499 100644 --- a/unsupported/test/cxx11_tensor_shuffling.cpp +++ b/unsupported/test/cxx11_tensor_shuffling.cpp @@ -12,6 +12,7 @@ #include using Eigen::Tensor; +using Eigen::array; static void test_simple_shuffling() { @@ -80,10 +81,10 @@ static void test_expr_shuffling() Tensor result(5,7,3,2); - array src_slice_dim(Eigen::array(2,3,1,7)); - array src_slice_start(Eigen::array(0,0,0,0)); - array dst_slice_dim(Eigen::array(1,7,3,2)); - array dst_slice_start(Eigen::array(0,0,0,0)); + array src_slice_dim{{2,3,1,7}}; + array src_slice_start{{0,0,0,0}}; + array dst_slice_dim{{1,7,3,2}}; + array dst_slice_start{{0,0,0,0}}; for (int i = 0; i < 5; ++i) { result.slice(dst_slice_start, dst_slice_dim) = From 2ed1838aeb6d3c70c35dbd8d545fba1e7e1c68dc Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 10 Oct 2014 16:11:27 -0700 Subject: [PATCH 0891/1103] Added support for tensor chips --- unsupported/Eigen/CXX11/Tensor | 1 + .../Eigen/CXX11/src/Tensor/TensorBase.h | 12 +- .../Eigen/CXX11/src/Tensor/TensorChipping.h | 232 +++++++++++++++++ .../src/Tensor/TensorForwardDeclarations.h | 3 +- unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_chipping.cpp | 244 ++++++++++++++++++ 6 files changed, 491 insertions(+), 2 deletions(-) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h create mode 100644 unsupported/test/cxx11_tensor_chipping.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index b1bd2f676..5a6246a03 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -47,6 +47,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index d4b7846a0..cadeb3b19 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -254,6 +254,11 @@ class TensorBase slice(const StartIndices& startIndices, const Sizes& sizes) const { return TensorSlicingOp(derived(), startIndices, sizes); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorChippingOp + chip(const Index offset) const { + return TensorChippingOp(derived(), offset); + } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorPaddingOp pad(const PaddingDimensions& padding) const { @@ -327,7 +332,7 @@ class TensorBase : public TensorBase EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReshapingOp - reshape(const NewDimensions& newDimensions) { + reshape(const NewDimensions& newDimensions) const { return TensorReshapingOp(derived(), newDimensions); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE @@ -335,6 +340,11 @@ class TensorBase : public TensorBase(derived(), startIndices, sizes); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorChippingOp + chip(const Index offset) const { + return TensorChippingOp(derived(), offset); + } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp shuffle(const Shuffle& shuffle) const { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h new file mode 100644 index 000000000..9ecea9108 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h @@ -0,0 +1,232 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H +#define EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H + +namespace Eigen { + +/** \class TensorKChippingReshaping + * \ingroup CXX11_Tensor_Module + * + * \brief A chip is a thin slice, corresponding to a column or a row in a 2-d tensor. + * + * + */ + +namespace internal { +template +struct traits > : public traits +{ + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorChippingOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorChippingOp type; +}; + +} // end namespace internal + + + +template +class TensorChippingOp : public TensorBase > +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp(const XprType& expr, const Index offset) + : m_xpr(expr), m_offset(offset) {} + + EIGEN_DEVICE_FUNC + const Index offset() const { return m_offset; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorChippingOp& operator = (const OtherDerived& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + + protected: + typename XprType::Nested m_xpr; + const Index m_offset; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorChippingOp XprType; + static const int NumInputDims = internal::array_size::Dimensions>::value; + static const int NumDims = NumInputDims-1; + typedef typename XprType::Index Index; + typedef DSizes Dimensions; + + enum { + // Alignment can't be guaranteed at compile time since it depends on the + // slice offsets. + IsAligned = false, + PacketAccess = false, // not yet implemented + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_device(device) + { + // We could also support the case where NumInputDims==1 if needed. + EIGEN_STATIC_ASSERT(NumInputDims >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(NumInputDims > DimId, YOU_MADE_A_PROGRAMMING_MISTAKE); + + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + int j = 0; + for (int i = 0; i < NumInputDims; ++i) { + if (i != DimId) { + m_dimensions[j] = input_dims[i]; + ++j; + } + } + + m_stride = 1; + m_inputStride = 1; + for (int i = 0; i < DimId; ++i) { + m_stride *= input_dims[i]; + m_inputStride *= input_dims[i]; + } + m_inputStride *= input_dims[DimId]; + m_inputOffset = m_stride * op.offset(); + } + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + return m_impl.coeff(srcCoeff(index)); + } + + /* to be done + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + + }*/ + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { return NULL; } + + protected: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const + { + Index inputIndex; + if (DimId == 0) { + // m_stride is equal to 1, so let's avoid the integer division. + eigen_assert(m_stride == 1); + inputIndex = index * m_inputStride + m_inputOffset; + } else if (DimId == NumInputDims-1) { + // m_stride is aways greater than index, so let's avoid the integer division. + eigen_assert(m_stride > index); + inputIndex = index + m_inputOffset; + } else { + const Index idx = index / m_stride; + inputIndex = idx * m_inputStride + m_inputOffset; + index -= idx * m_stride; + inputIndex += index; + } + return inputIndex; + } + + Dimensions m_dimensions; + Index m_stride; + Index m_inputOffset; + Index m_inputStride; + TensorEvaluator m_impl; + const Device& m_device; +}; + + +// Eval as lvalue +template +struct TensorEvaluator, Device> + : public TensorEvaluator, Device> +{ + typedef TensorEvaluator, Device> Base; + typedef TensorChippingOp XprType; + static const int NumInputDims = internal::array_size::Dimensions>::value; + static const int NumDims = NumInputDims-1; + typedef typename XprType::Index Index; + typedef DSizes Dimensions; + + enum { + IsAligned = false, + PacketAccess = false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : Base(op, device) + { } + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) + { + return this->m_impl.coeffRef(this->srcCoeff(index)); + } + + /* to be done + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketReturnType& x) + { + } */ +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index bc67586a4..86ddd1ae8 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -21,11 +21,12 @@ template class TensorCwiseNullaryO template class TensorCwiseUnaryOp; template class TensorCwiseBinaryOp; template class TensorSelectOp; -template class TensorBroadcastingOp; template class TensorReductionOp; template class TensorConcatenationOp; template class TensorContractionOp; template class TensorConvolutionOp; +template class TensorBroadcastingOp; +template class TensorChippingOp; template class TensorReshapingOp; template class TensorSlicingOp; template class TensorPaddingOp; diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index ac2ccaf27..48435eb9c 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -115,6 +115,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") ei_add_test(cxx11_tensor_broadcasting "-std=c++0x") + ei_add_test(cxx11_tensor_chipping "-std=c++0x") ei_add_test(cxx11_tensor_concatenation "-std=c++0x") ei_add_test(cxx11_tensor_morphing "-std=c++0x") ei_add_test(cxx11_tensor_padding "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_chipping.cpp b/unsupported/test/cxx11_tensor_chipping.cpp new file mode 100644 index 000000000..8c8a0cec2 --- /dev/null +++ b/unsupported/test/cxx11_tensor_chipping.cpp @@ -0,0 +1,244 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + + +static void test_simple_chip() +{ + Tensor tensor(2,3,5,7,11); + tensor.setRandom(); + + Tensor chip1; + chip1 = tensor.chip<0>(1); + VERIFY_IS_EQUAL(chip1.dimension(0), 3); + VERIFY_IS_EQUAL(chip1.dimension(1), 5); + VERIFY_IS_EQUAL(chip1.dimension(2), 7); + VERIFY_IS_EQUAL(chip1.dimension(3), 11); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip1(i,j,k,l), tensor(1,i,j,k,l)); + } + } + } + } + + Tensor chip2 = tensor.chip<1>(1); + VERIFY_IS_EQUAL(chip2.dimension(0), 2); + VERIFY_IS_EQUAL(chip2.dimension(1), 5); + VERIFY_IS_EQUAL(chip2.dimension(2), 7); + VERIFY_IS_EQUAL(chip2.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip2(i,j,k,l), tensor(i,1,j,k,l)); + } + } + } + } + + Tensor chip3 = tensor.chip<2>(2); + VERIFY_IS_EQUAL(chip3.dimension(0), 2); + VERIFY_IS_EQUAL(chip3.dimension(1), 3); + VERIFY_IS_EQUAL(chip3.dimension(2), 7); + VERIFY_IS_EQUAL(chip3.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip3(i,j,k,l), tensor(i,j,2,k,l)); + } + } + } + } + + Tensor chip4(tensor.chip<3>(5)); + VERIFY_IS_EQUAL(chip4.dimension(0), 2); + VERIFY_IS_EQUAL(chip4.dimension(1), 3); + VERIFY_IS_EQUAL(chip4.dimension(2), 5); + VERIFY_IS_EQUAL(chip4.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(chip4(i,j,k,l), tensor(i,j,k,5,l)); + } + } + } + } + + Tensor chip5(tensor.chip<4>(7)); + VERIFY_IS_EQUAL(chip5.dimension(0), 2); + VERIFY_IS_EQUAL(chip5.dimension(1), 3); + VERIFY_IS_EQUAL(chip5.dimension(2), 5); + VERIFY_IS_EQUAL(chip5.dimension(3), 7); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(chip5(i,j,k,l), tensor(i,j,k,l,7)); + } + } + } + } +} + + +static void test_chip_in_expr() { + Tensor input1(2,3,5,7,11); + input1.setRandom(); + Tensor input2(3,5,7,11); + input2.setRandom(); + + Tensor result = input1.chip<0>(0) + input2; + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + float expected = input1(0,i,j,k,l) + input2(i,j,k,l); + VERIFY_IS_EQUAL(result(i,j,k,l), expected); + } + } + } + } + + Tensor input3(3,7,11); + input3.setRandom(); + Tensor result2 = input1.chip<0>(0).chip<1>(2) + input3; + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 7; ++j) { + for (int k = 0; k < 11; ++k) { + float expected = input1(0,i,2,j,k) + input3(i,j,k); + VERIFY_IS_EQUAL(result2(i,j,k), expected); + } + } + } +} + + +static void test_chip_as_lvalue() +{ + Tensor input1(2,3,5,7,11); + input1.setRandom(); + + Tensor input2(3,5,7,11); + input2.setRandom(); + Tensor tensor = input1; + tensor.chip<0>(1) = input2; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (i != 1) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input2(j,k,l,m)); + } + } + } + } + } + } + + Tensor input3(2,5,7,11); + input3.setRandom(); + tensor = input1; + tensor.chip<1>(1) = input3; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (j != 1) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input3(i,k,l,m)); + } + } + } + } + } + } + + Tensor input4(2,3,7,11); + input4.setRandom(); + tensor = input1; + tensor.chip<2>(3) = input4; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (k != 3) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input4(i,j,l,m)); + } + } + } + } + } + } + + Tensor input5(2,3,5,11); + input5.setRandom(); + tensor = input1; + tensor.chip<3>(4) = input5; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (l != 4) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input5(i,j,k,m)); + } + } + } + } + } + } + + Tensor input6(2,3,5,7); + input6.setRandom(); + tensor = input1; + tensor.chip<4>(5) = input6; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (m != 5) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input6(i,j,k,l)); + } + } + } + } + } + } +} + + +void test_cxx11_tensor_chipping() +{ + CALL_SUBTEST(test_simple_chip()); + CALL_SUBTEST(test_chip_in_expr()); + CALL_SUBTEST(test_chip_as_lvalue()); +} From 0219f8aed44279858330b1c07402c066f5b75459 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 10 Oct 2014 16:17:26 -0700 Subject: [PATCH 0892/1103] Added ability to print a tensor using an iostream. --- unsupported/Eigen/CXX11/Tensor | 2 + unsupported/Eigen/CXX11/src/Tensor/TensorIO.h | 44 ++++++++++++ unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_io.cpp | 70 +++++++++++++++++++ 4 files changed, 117 insertions(+) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorIO.h create mode 100644 unsupported/test/cxx11_tensor_io.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 5a6246a03..79510fd96 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -64,6 +64,8 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMap.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorIO.h" + #include "Eigen/src/Core/util/ReenableStupidWarnings.h" #endif // EIGEN_CXX11_TENSOR_MODULE diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIO.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIO.h new file mode 100644 index 000000000..959b5db73 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIO.h @@ -0,0 +1,44 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_IO_H +#define EIGEN_CXX11_TENSOR_TENSOR_IO_H + +namespace Eigen { + +template +std::ostream& operator << (std::ostream& os, const TensorBase& expr) { + // Evaluate the expression if needed + TensorForcedEvalOp eval = expr.eval(); + TensorEvaluator, DefaultDevice> tensor(eval, DefaultDevice()); + tensor.evalSubExprsIfNeeded(NULL); + + typedef typename T::Scalar Scalar; + typedef typename T::Index Index; + typedef typename TensorEvaluator, DefaultDevice>::Dimensions Dimensions; + const Index total_size = internal::array_prod(tensor.dimensions()); + + // Print the tensor as a 1d vector or a 2d matrix. + if (internal::array_size::value == 1) { + Map > array(tensor.data(), total_size); + os << array; + } else { + const Index first_dim = tensor.dimensions()[0]; + Map > matrix(tensor.data(), first_dim, total_size/first_dim); + os << matrix; + } + + // Cleanup. + tensor.cleanup(); + return os; +} + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_IO_H diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 48435eb9c..99593b562 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -124,4 +124,5 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") + ei_add_test(cxx11_tensor_io "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_io.cpp b/unsupported/test/cxx11_tensor_io.cpp new file mode 100644 index 000000000..b73c024f5 --- /dev/null +++ b/unsupported/test/cxx11_tensor_io.cpp @@ -0,0 +1,70 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" +#include +#include +#include + + +static void test_output_1d() +{ + Tensor tensor(5); + for (int i = 0; i < 5; ++i) { + tensor(i) = i; + } + + std::stringstream os; + os << tensor; + + std::string expected("0\n1\n2\n3\n4"); + VERIFY_IS_EQUAL(std::string(os.str()), expected); +} + + +static void test_output_2d() +{ + Tensor tensor(5, 3); + for (int i = 0; i < 5; ++i) { + for (int j = 0; j < 3; ++j) { + tensor(i, j) = i*j; + } + } + + std::stringstream os; + os << tensor; + + std::string expected("0 0 0\n0 1 2\n0 2 4\n0 3 6\n0 4 8"); + VERIFY_IS_EQUAL(std::string(os.str()), expected); +} + + +static void test_output_expr() +{ + Tensor tensor1(5); + Tensor tensor2(5); + for (int i = 0; i < 5; ++i) { + tensor1(i) = i; + tensor2(i) = 7; + } + + std::stringstream os; + os << tensor1 + tensor2; + + std::string expected(" 7\n 8\n 9\n10\n11"); + VERIFY_IS_EQUAL(std::string(os.str()), expected); +} + + +void test_cxx11_tensor_io() +{ + CALL_SUBTEST(test_output_1d()); + CALL_SUBTEST(test_output_2d()); + CALL_SUBTEST(test_output_expr()); +} From d3f52debc6c45d1f26bb9406207ebc5a8638a429 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 13 Oct 2014 17:18:26 +0200 Subject: [PATCH 0893/1103] Make cuda_basic test compile again by adding lots of EIGEN_DEVICE_FUNC. Although the test passes now, there might still be some missing. --- Eigen/src/Core/AssignEvaluator.h | 108 +++++++------- Eigen/src/Core/Block.h | 2 +- Eigen/src/Core/CoreEvaluators.h | 146 +++++++++---------- Eigen/src/Core/CwiseUnaryView.h | 8 +- Eigen/src/Core/Flagged.h | 26 ++-- Eigen/src/Core/ForceAlignedAccess.h | 20 +-- Eigen/src/Core/MapBase.h | 2 +- Eigen/src/Core/MatrixBase.h | 2 +- Eigen/src/Core/NestByValue.h | 20 +-- Eigen/src/Core/PlainObjectBase.h | 4 +- Eigen/src/Core/Product.h | 16 +- Eigen/src/Core/ProductEvaluators.h | 33 +++-- Eigen/src/Core/Redux.h | 19 ++- Eigen/src/Core/Ref.h | 22 +-- Eigen/src/Core/ReturnByValue.h | 2 +- Eigen/src/Core/Reverse.h | 22 +-- Eigen/src/Core/SelfAdjointView.h | 8 +- Eigen/src/Core/Solve.h | 2 +- Eigen/src/Core/Swap.h | 2 +- Eigen/src/Core/Transpose.h | 4 +- Eigen/src/Core/TriangularMatrix.h | 18 +-- Eigen/src/Core/functors/AssignmentFunctors.h | 7 +- test/cuda_basic.cu | 6 +- 23 files changed, 255 insertions(+), 244 deletions(-) diff --git a/Eigen/src/Core/AssignEvaluator.h b/Eigen/src/Core/AssignEvaluator.h index 5d5095a67..f481a7a36 100644 --- a/Eigen/src/Core/AssignEvaluator.h +++ b/Eigen/src/Core/AssignEvaluator.h @@ -166,7 +166,7 @@ struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling inner = Index % DstXprType::InnerSizeAtCompileTime }; - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { kernel.assignCoeffByOuterInner(outer, inner); copy_using_evaluator_DefaultTraversal_CompleteUnrolling::run(kernel); @@ -176,13 +176,13 @@ struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling template struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel&) { } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { } }; template struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel &kernel, typename Kernel::Index outer) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, typename Kernel::Index outer) { kernel.assignCoeffByOuterInner(outer, Index); copy_using_evaluator_DefaultTraversal_InnerUnrolling::run(kernel, outer); @@ -192,7 +192,7 @@ struct copy_using_evaluator_DefaultTraversal_InnerUnrolling template struct copy_using_evaluator_DefaultTraversal_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel&, typename Kernel::Index) { } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, typename Kernel::Index) { } }; /*********************** @@ -202,7 +202,7 @@ struct copy_using_evaluator_DefaultTraversal_InnerUnrolling template struct copy_using_evaluator_LinearTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel& kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel& kernel) { kernel.assignCoeff(Index); copy_using_evaluator_LinearTraversal_CompleteUnrolling::run(kernel); @@ -212,7 +212,7 @@ struct copy_using_evaluator_LinearTraversal_CompleteUnrolling template struct copy_using_evaluator_LinearTraversal_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel&) { } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { } }; /************************** @@ -232,7 +232,7 @@ struct copy_using_evaluator_innervec_CompleteUnrolling JointAlignment = Kernel::AssignmentTraits::JointAlignment }; - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { kernel.template assignPacketByOuterInner(outer, inner); enum { NextIndex = Index + packet_traits::size }; @@ -243,13 +243,13 @@ struct copy_using_evaluator_innervec_CompleteUnrolling template struct copy_using_evaluator_innervec_CompleteUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel&) { } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { } }; template struct copy_using_evaluator_innervec_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel &kernel, typename Kernel::Index outer) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, typename Kernel::Index outer) { kernel.template assignPacketByOuterInner(outer, Index); enum { NextIndex = Index + packet_traits::size }; @@ -260,7 +260,7 @@ struct copy_using_evaluator_innervec_InnerUnrolling template struct copy_using_evaluator_innervec_InnerUnrolling { - static EIGEN_STRONG_INLINE void run(Kernel &, typename Kernel::Index) { } + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &, typename Kernel::Index) { } }; /*************************************************************************** @@ -281,7 +281,7 @@ struct dense_assignment_loop; template struct dense_assignment_loop { - static void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static void run(Kernel &kernel) { typedef typename Kernel::Index Index; @@ -296,7 +296,7 @@ struct dense_assignment_loop template struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { typedef typename Kernel::DstEvaluatorType::XprType DstXprType; copy_using_evaluator_DefaultTraversal_CompleteUnrolling::run(kernel); @@ -307,7 +307,7 @@ template struct dense_assignment_loop { typedef typename Kernel::Index Index; - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { typedef typename Kernel::DstEvaluatorType::XprType DstXprType; @@ -330,7 +330,7 @@ struct unaligned_dense_assignment_loop { // if IsAligned = true, then do nothing template - static EIGEN_STRONG_INLINE void run(Kernel&, typename Kernel::Index, typename Kernel::Index) {} + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, typename Kernel::Index, typename Kernel::Index) {} }; template <> @@ -346,7 +346,7 @@ struct unaligned_dense_assignment_loop typename Kernel::Index end) #else template - static EIGEN_STRONG_INLINE void run(Kernel &kernel, + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, typename Kernel::Index start, typename Kernel::Index end) #endif @@ -359,7 +359,7 @@ struct unaligned_dense_assignment_loop template struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { typedef typename Kernel::Index Index; @@ -387,7 +387,7 @@ template struct dense_assignment_loop { typedef typename Kernel::Index Index; - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { typedef typename Kernel::DstEvaluatorType::XprType DstXprType; @@ -407,7 +407,7 @@ struct dense_assignment_loop struct dense_assignment_loop { - static inline void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel) { typedef typename Kernel::Index Index; @@ -423,7 +423,7 @@ struct dense_assignment_loop template struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { typedef typename Kernel::DstEvaluatorType::XprType DstXprType; copy_using_evaluator_innervec_CompleteUnrolling::run(kernel); @@ -434,7 +434,7 @@ template struct dense_assignment_loop { typedef typename Kernel::Index Index; - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { typedef typename Kernel::DstEvaluatorType::XprType DstXprType; const Index outerSize = kernel.outerSize(); @@ -450,7 +450,7 @@ struct dense_assignment_loop template struct dense_assignment_loop { - static inline void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel) { typedef typename Kernel::Index Index; const Index size = kernel.size(); @@ -462,7 +462,7 @@ struct dense_assignment_loop template struct dense_assignment_loop { - static EIGEN_STRONG_INLINE void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel) { typedef typename Kernel::DstEvaluatorType::XprType DstXprType; copy_using_evaluator_LinearTraversal_CompleteUnrolling::run(kernel); @@ -476,7 +476,7 @@ struct dense_assignment_loop template struct dense_assignment_loop { - static inline void run(Kernel &kernel) + EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel) { typedef typename Kernel::Index Index; typedef packet_traits PacketTraits; @@ -537,7 +537,7 @@ public: typedef copy_using_evaluator_traits AssignmentTraits; - generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) : m_dst(dst), m_src(src), m_functor(func), m_dstExpr(dstExpr) { #ifdef EIGEN_DEBUG_ASSIGN @@ -545,33 +545,33 @@ public: #endif } - Index size() const { return m_dstExpr.size(); } - Index innerSize() const { return m_dstExpr.innerSize(); } - Index outerSize() const { return m_dstExpr.outerSize(); } - Index rows() const { return m_dstExpr.rows(); } - Index cols() const { return m_dstExpr.cols(); } - Index outerStride() const { return m_dstExpr.outerStride(); } + EIGEN_DEVICE_FUNC Index size() const { return m_dstExpr.size(); } + EIGEN_DEVICE_FUNC Index innerSize() const { return m_dstExpr.innerSize(); } + EIGEN_DEVICE_FUNC Index outerSize() const { return m_dstExpr.outerSize(); } + EIGEN_DEVICE_FUNC Index rows() const { return m_dstExpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_dstExpr.cols(); } + EIGEN_DEVICE_FUNC Index outerStride() const { return m_dstExpr.outerStride(); } // TODO get rid of this one: - DstXprType& dstExpression() const { return m_dstExpr; } + EIGEN_DEVICE_FUNC DstXprType& dstExpression() const { return m_dstExpr; } - DstEvaluatorType& dstEvaluator() { return m_dst; } - const SrcEvaluatorType& srcEvaluator() const { return m_src; } + EIGEN_DEVICE_FUNC DstEvaluatorType& dstEvaluator() { return m_dst; } + EIGEN_DEVICE_FUNC const SrcEvaluatorType& srcEvaluator() const { return m_src; } /// Assign src(row,col) to dst(row,col) through the assignment functor. - void assignCoeff(Index row, Index col) + EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col) { m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col)); } /// \sa assignCoeff(Index,Index) - void assignCoeff(Index index) + EIGEN_DEVICE_FUNC void assignCoeff(Index index) { m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index)); } /// \sa assignCoeff(Index,Index) - void assignCoeffByOuterInner(Index outer, Index inner) + EIGEN_DEVICE_FUNC void assignCoeffByOuterInner(Index outer, Index inner) { Index row = rowIndexByOuterInner(outer, inner); Index col = colIndexByOuterInner(outer, inner); @@ -580,26 +580,26 @@ public: template - void assignPacket(Index row, Index col) + EIGEN_DEVICE_FUNC void assignPacket(Index row, Index col) { m_functor.template assignPacket(&m_dst.coeffRef(row,col), m_src.template packet(row,col)); } template - void assignPacket(Index index) + EIGEN_DEVICE_FUNC void assignPacket(Index index) { m_functor.template assignPacket(&m_dst.coeffRef(index), m_src.template packet(index)); } template - void assignPacketByOuterInner(Index outer, Index inner) + EIGEN_DEVICE_FUNC void assignPacketByOuterInner(Index outer, Index inner) { Index row = rowIndexByOuterInner(outer, inner); Index col = colIndexByOuterInner(outer, inner); assignPacket(row, col); } - static Index rowIndexByOuterInner(Index outer, Index inner) + EIGEN_DEVICE_FUNC static Index rowIndexByOuterInner(Index outer, Index inner) { typedef typename DstEvaluatorType::ExpressionTraits Traits; return int(Traits::RowsAtCompileTime) == 1 ? 0 @@ -608,7 +608,7 @@ public: : inner; } - static Index colIndexByOuterInner(Index outer, Index inner) + EIGEN_DEVICE_FUNC static Index colIndexByOuterInner(Index outer, Index inner) { typedef typename DstEvaluatorType::ExpressionTraits Traits; return int(Traits::ColsAtCompileTime) == 1 ? 0 @@ -630,7 +630,7 @@ protected: ***************************************************************************/ template -void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) +EIGEN_DEVICE_FUNC void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) { eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); @@ -647,7 +647,7 @@ void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src, co } template -void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src) +EIGEN_DEVICE_FUNC void call_dense_assignment_loop(const DstXprType& dst, const SrcXprType& src) { call_dense_assignment_loop(dst, src, internal::assign_op()); } @@ -681,26 +681,26 @@ struct Assignment; // does not has to bother about these annoying details. template -void call_assignment(Dst& dst, const Src& src) +EIGEN_DEVICE_FUNC void call_assignment(Dst& dst, const Src& src) { call_assignment(dst, src, internal::assign_op()); } template -void call_assignment(const Dst& dst, const Src& src) +EIGEN_DEVICE_FUNC void call_assignment(const Dst& dst, const Src& src) { call_assignment(dst, src, internal::assign_op()); } // Deal with AssumeAliasing template -void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) +EIGEN_DEVICE_FUNC void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==1, void*>::type = 0) { typename plain_matrix_type::type tmp(src); call_assignment_no_alias(dst, tmp, func); } template -void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==0, void*>::type = 0) +EIGEN_DEVICE_FUNC void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if::AssumeAliasing==0, void*>::type = 0) { call_assignment_no_alias(dst, src, func); } @@ -709,19 +709,19 @@ void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable // FIXME the const version should probably not be needed // When there is no aliasing, we require that 'dst' has been properly resized template class StorageBase, typename Src, typename Func> -void call_assignment(const NoAlias& dst, const Src& src, const Func& func) +EIGEN_DEVICE_FUNC void call_assignment(const NoAlias& dst, const Src& src, const Func& func) { call_assignment_no_alias(dst.expression(), src, func); } template class StorageBase, typename Src, typename Func> -void call_assignment(NoAlias& dst, const Src& src, const Func& func) +EIGEN_DEVICE_FUNC void call_assignment(NoAlias& dst, const Src& src, const Func& func) { call_assignment_no_alias(dst.expression(), src, func); } template -void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) +EIGEN_DEVICE_FUNC void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) { enum { NeedToTranspose = ( (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1) @@ -752,19 +752,19 @@ void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func) Assignment::run(actualDst, src, func); } template -void call_assignment_no_alias(Dst& dst, const Src& src) +EIGEN_DEVICE_FUNC void call_assignment_no_alias(Dst& dst, const Src& src) { call_assignment_no_alias(dst, src, internal::assign_op()); } -// forxard declaration +// forward declaration template void check_for_aliasing(const Dst &dst, const Src &src); // Generic Dense to Dense assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> struct Assignment { - static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) { eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); @@ -781,7 +781,7 @@ struct Assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> struct Assignment { - static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op &/*func*/) { eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 737e5dc24..9cf9d5432 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -178,7 +178,7 @@ template struct evaluator : public unary_evaluator { typedef unary_evaluator Base; - explicit evaluator(const T& xpr) : Base(xpr) {} + EIGEN_DEVICE_FUNC explicit evaluator(const T& xpr) : Base(xpr) {} }; @@ -145,18 +145,18 @@ struct evaluator > Derived::Options,Derived::MaxRowsAtCompileTime,Derived::MaxColsAtCompileTime>::ret }; - evaluator() + EIGEN_DEVICE_FUNC evaluator() : m_data(0), m_outerStride(IsVectorAtCompileTime ? 0 : int(IsRowMajor) ? ColsAtCompileTime : RowsAtCompileTime) {} - explicit evaluator(const PlainObjectType& m) + EIGEN_DEVICE_FUNC explicit evaluator(const PlainObjectType& m) : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) { } - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { if (IsRowMajor) return m_data[row * m_outerStride.value() + col]; @@ -164,12 +164,12 @@ struct evaluator > return m_data[row + col * m_outerStride.value()]; } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_data[index]; } - Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index col) { if (IsRowMajor) return const_cast(m_data)[row * m_outerStride.value() + col]; @@ -177,7 +177,7 @@ struct evaluator > return const_cast(m_data)[row + col * m_outerStride.value()]; } - Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return const_cast(m_data)[index]; } @@ -231,7 +231,7 @@ struct evaluator > evaluator() {} - explicit evaluator(const XprType& m) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m) : evaluator >(m) { } }; @@ -244,7 +244,7 @@ struct evaluator > evaluator() {} - explicit evaluator(const XprType& m) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m) : evaluator >(m) { } }; @@ -262,7 +262,7 @@ struct unary_evaluator, IndexBased> Flags = evaluator::Flags ^ RowMajorBit }; - explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {} typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -270,22 +270,22 @@ struct unary_evaluator, IndexBased> typedef typename XprType::PacketScalar PacketScalar; typedef typename XprType::PacketReturnType PacketReturnType; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_argImpl.coeff(col, row); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_argImpl.coeff(index); } - Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index col) { return m_argImpl.coeffRef(col, row); } - typename XprType::Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC typename XprType::Scalar& coeffRef(Index index) { return m_argImpl.coeffRef(index); } @@ -339,7 +339,7 @@ struct evaluator > | (functor_traits::IsRepeatable ? 0 : EvalBeforeNestingBit) // FIXME EvalBeforeNestingBit should be needed anymore }; - explicit evaluator(const XprType& n) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& n) : m_functor(n.functor()) { } @@ -347,12 +347,12 @@ struct evaluator > typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketScalar PacketScalar; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_functor(row, col); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_functor(index); } @@ -389,7 +389,7 @@ struct unary_evaluator, IndexBased > | (functor_traits::PacketAccess ? PacketAccessBit : 0)) }; - explicit unary_evaluator(const XprType& op) + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -398,12 +398,12 @@ struct unary_evaluator, IndexBased > typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketScalar PacketScalar; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_functor(m_argImpl.coeff(row, col)); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_functor(m_argImpl.coeff(index)); } @@ -435,7 +435,7 @@ struct evaluator > typedef CwiseBinaryOp XprType; typedef binary_evaluator > Base; - explicit evaluator(const XprType& xpr) : Base(xpr) {} + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {} }; template @@ -463,7 +463,7 @@ struct binary_evaluator, IndexBased, IndexBase Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit) }; - explicit binary_evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit binary_evaluator(const XprType& xpr) : m_functor(xpr.functor()), m_lhsImpl(xpr.lhs()), m_rhsImpl(xpr.rhs()) @@ -473,12 +473,12 @@ struct binary_evaluator, IndexBased, IndexBase typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketScalar PacketScalar; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_functor(m_lhsImpl.coeff(row, col), m_rhsImpl.coeff(row, col)); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_functor(m_lhsImpl.coeff(index), m_rhsImpl.coeff(index)); } @@ -517,7 +517,7 @@ struct unary_evaluator, IndexBased> Flags = (evaluator::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit)) }; - explicit unary_evaluator(const XprType& op) + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op) : m_unaryOp(op.functor()), m_argImpl(op.nestedExpression()) { } @@ -526,22 +526,22 @@ struct unary_evaluator, IndexBased> typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_unaryOp(m_argImpl.coeff(row, col)); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_unaryOp(m_argImpl.coeff(index)); } - Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index col) { return m_unaryOp(m_argImpl.coeffRef(row, col)); } - Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return m_unaryOp(m_argImpl.coeffRef(index)); } @@ -575,7 +575,7 @@ struct mapbase_evaluator : evaluator_base CoeffReadCost = NumTraits::ReadCost }; - explicit mapbase_evaluator(const XprType& map) + EIGEN_DEVICE_FUNC explicit mapbase_evaluator(const XprType& map) : m_data(const_cast(map.data())), m_xpr(map) { @@ -583,22 +583,22 @@ struct mapbase_evaluator : evaluator_base PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1); } - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_data[col * m_xpr.colStride() + row * m_xpr.rowStride()]; } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_data[index * m_xpr.innerStride()]; } - Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index col) { return m_data[col * m_xpr.colStride() + row * m_xpr.rowStride()]; } - Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return m_data[index * m_xpr.innerStride()]; } @@ -665,7 +665,7 @@ struct evaluator > Flags = KeepsPacketAccess ? int(Flags2) : (int(Flags2) & ~PacketAccessBit) }; - explicit evaluator(const XprType& map) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& map) : mapbase_evaluator(map) { } }; @@ -682,7 +682,7 @@ struct evaluator > Flags = evaluator >::Flags }; - explicit evaluator(const XprType& ref) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& ref) : mapbase_evaluator(ref) { } }; @@ -733,7 +733,7 @@ struct evaluator > Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit }; typedef block_evaluator block_evaluator_type; - explicit evaluator(const XprType& block) : block_evaluator_type(block) {} + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& block) : block_evaluator_type(block) {} }; // no direct-access => dispatch to a unary evaluator @@ -743,7 +743,7 @@ struct block_evaluator XprType; - explicit block_evaluator(const XprType& block) + EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block) : unary_evaluator(block) {} }; @@ -754,7 +754,7 @@ struct unary_evaluator, IndexBa { typedef Block XprType; - explicit unary_evaluator(const XprType& block) + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& block) : m_argImpl(block.nestedExpression()), m_startRow(block.startRow()), m_startCol(block.startCol()) @@ -770,22 +770,22 @@ struct unary_evaluator, IndexBa RowsAtCompileTime = XprType::RowsAtCompileTime }; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_argImpl.coeff(m_startRow.value() + row, m_startCol.value() + col); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); } - Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index col) { return m_argImpl.coeffRef(m_startRow.value() + row, m_startCol.value() + col); } - Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0); } @@ -833,7 +833,7 @@ struct block_evaluator XprType; - explicit block_evaluator(const XprType& block) + EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block) : mapbase_evaluator(block) { // FIXME this should be an internal assertion @@ -859,7 +859,7 @@ struct evaluator > Flags = (unsigned int)evaluator::Flags & evaluator::Flags & HereditaryBits }; - explicit evaluator(const XprType& select) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& select) : m_conditionImpl(select.conditionMatrix()), m_thenImpl(select.thenMatrix()), m_elseImpl(select.elseMatrix()) @@ -868,7 +868,7 @@ struct evaluator > typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { if (m_conditionImpl.coeff(row, col)) return m_thenImpl.coeff(row, col); @@ -876,7 +876,7 @@ struct evaluator > return m_elseImpl.coeff(row, col); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { if (m_conditionImpl.coeff(index)) return m_thenImpl.coeff(index); @@ -913,14 +913,14 @@ struct unary_evaluator > Flags = (evaluator::Flags & HereditaryBits & ~RowMajorBit) | (traits::Flags & RowMajorBit) }; - explicit unary_evaluator(const XprType& replicate) + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& replicate) : m_arg(replicate.nestedExpression()), m_argImpl(m_arg), m_rows(replicate.nestedExpression().rows()), m_cols(replicate.nestedExpression().cols()) {} - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { // try to avoid using modulo; this is a pure optimization strategy const Index actual_row = internal::traits::RowsAtCompileTime==1 ? 0 @@ -977,19 +977,19 @@ struct evaluator > Flags = (traits::Flags&RowMajorBit) | (evaluator::Flags&HereditaryBits) }; - explicit evaluator(const XprType expr) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType expr) : m_expr(expr) {} typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_expr.coeff(row, col); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_expr.coeff(index); } @@ -1014,7 +1014,7 @@ struct evaluator_wrapper_base Flags = evaluator::Flags }; - explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} + EIGEN_DEVICE_FUNC explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {} typedef typename ArgType::Index Index; typedef typename ArgType::Scalar Scalar; @@ -1022,22 +1022,22 @@ struct evaluator_wrapper_base typedef typename ArgType::PacketScalar PacketScalar; typedef typename ArgType::PacketReturnType PacketReturnType; - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_argImpl.coeff(row, col); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_argImpl.coeff(index); } - Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index col) { return m_argImpl.coeffRef(row, col); } - Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return m_argImpl.coeffRef(index); } @@ -1076,7 +1076,7 @@ struct unary_evaluator > { typedef MatrixWrapper XprType; - explicit unary_evaluator(const XprType& wrapper) + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; @@ -1087,7 +1087,7 @@ struct unary_evaluator > { typedef ArrayWrapper XprType; - explicit unary_evaluator(const XprType& wrapper) + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper) : evaluator_wrapper_base >(wrapper.nestedExpression()) { } }; @@ -1133,30 +1133,30 @@ struct unary_evaluator > }; typedef internal::reverse_packet_cond reverse_packet; - explicit unary_evaluator(const XprType& reverse) + EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& reverse) : m_argImpl(reverse.nestedExpression()), m_rows(ReverseRow ? reverse.nestedExpression().rows() : 0), m_cols(ReverseCol ? reverse.nestedExpression().cols() : 0) { } - CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_argImpl.coeff(ReverseRow ? m_rows.value() - row - 1 : row, ReverseCol ? m_cols.value() - col - 1 : col); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_argImpl.coeff(m_rows.value() * m_cols.value() - index - 1); } - Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index col) { return m_argImpl.coeffRef(ReverseRow ? m_rows.value() - row - 1 : row, ReverseCol ? m_cols.value() - col - 1 : col); } - Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return m_argImpl.coeffRef(m_rows.value() * m_cols.value() - index - 1); } @@ -1214,7 +1214,7 @@ struct evaluator > Flags = (unsigned int)evaluator::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit) & ~RowMajorBit }; - explicit evaluator(const XprType& diagonal) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& diagonal) : m_argImpl(diagonal.nestedExpression()), m_index(diagonal.index()) { } @@ -1223,22 +1223,22 @@ struct evaluator > typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; - CoeffReturnType coeff(Index row, Index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index) const { return m_argImpl.coeff(row + rowOffset(), row + colOffset()); } - CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_argImpl.coeff(index + rowOffset(), index + colOffset()); } - Scalar& coeffRef(Index row, Index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index row, Index) { return m_argImpl.coeffRef(row + rowOffset(), row + colOffset()); } - Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC Scalar& coeffRef(Index index) { return m_argImpl.coeffRef(index + rowOffset(), index + colOffset()); } @@ -1248,8 +1248,8 @@ protected: const internal::variable_if_dynamicindex m_index; private: - EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value() > 0 ? 0 : -m_index.value(); } - EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value() > 0 ? 0 : -m_index.value(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; } }; @@ -1311,7 +1311,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - explicit evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); @@ -1320,7 +1320,7 @@ struct evaluator > } // This constructor is used when nesting an EvalTo evaluator in another evaluator - evaluator(const ArgType& arg) + EIGEN_DEVICE_FUNC evaluator(const ArgType& arg) : m_result(arg.rows(), arg.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h index 6384dfdc3..6680f32dd 100644 --- a/Eigen/src/Core/CwiseUnaryView.h +++ b/Eigen/src/Core/CwiseUnaryView.h @@ -110,15 +110,15 @@ class CwiseUnaryViewImpl EIGEN_DENSE_PUBLIC_INTERFACE(Derived) EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl) - inline Scalar* data() { return &(this->coeffRef(0)); } - inline const Scalar* data() const { return &(this->coeff(0)); } + EIGEN_DEVICE_FUNC inline Scalar* data() { return &(this->coeffRef(0)); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return &(this->coeff(0)); } - inline Index innerStride() const + EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride() * sizeof(typename internal::traits::Scalar) / sizeof(Scalar); } - inline Index outerStride() const + EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().nestedExpression().outerStride() * sizeof(typename internal::traits::Scalar) / sizeof(Scalar); } diff --git a/Eigen/src/Core/Flagged.h b/Eigen/src/Core/Flagged.h index 6ce11edf3..2e2a50be5 100644 --- a/Eigen/src/Core/Flagged.h +++ b/Eigen/src/Core/Flagged.h @@ -50,37 +50,37 @@ template clas explicit inline Flagged(const ExpressionType& matrix) : m_matrix(matrix) {} - inline Index rows() const { return m_matrix.rows(); } - inline Index cols() const { return m_matrix.cols(); } - inline Index outerStride() const { return m_matrix.outerStride(); } - inline Index innerStride() const { return m_matrix.innerStride(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_matrix.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_matrix.innerStride(); } - inline CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index row, Index col) const { return m_matrix.coeff(row, col); } - inline CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index index) const { return m_matrix.coeff(index); } - inline const Scalar& coeffRef(Index row, Index col) const + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index row, Index col) const { return m_matrix.const_cast_derived().coeffRef(row, col); } - inline const Scalar& coeffRef(Index index) const + EIGEN_DEVICE_FUNC inline const Scalar& coeffRef(Index index) const { return m_matrix.const_cast_derived().coeffRef(index); } - inline Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { return m_matrix.const_cast_derived().coeffRef(row, col); } - inline Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { return m_matrix.const_cast_derived().coeffRef(index); } @@ -109,13 +109,13 @@ template clas m_matrix.const_cast_derived().template writePacket(index, x); } - const ExpressionType& _expression() const { return m_matrix; } + EIGEN_DEVICE_FUNC const ExpressionType& _expression() const { return m_matrix; } template - typename ExpressionType::PlainObject solveTriangular(const MatrixBase& other) const; + EIGEN_DEVICE_FUNC typename ExpressionType::PlainObject solveTriangular(const MatrixBase& other) const; template - void solveTriangularInPlace(const MatrixBase& other) const; + EIGEN_DEVICE_FUNC void solveTriangularInPlace(const MatrixBase& other) const; protected: ExpressionTypeNested m_matrix; diff --git a/Eigen/src/Core/ForceAlignedAccess.h b/Eigen/src/Core/ForceAlignedAccess.h index 065acfa64..7b08b45e6 100644 --- a/Eigen/src/Core/ForceAlignedAccess.h +++ b/Eigen/src/Core/ForceAlignedAccess.h @@ -39,29 +39,29 @@ template class ForceAlignedAccess typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(ForceAlignedAccess) - explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {} + EIGEN_DEVICE_FUNC explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {} - inline Index rows() const { return m_expression.rows(); } - inline Index cols() const { return m_expression.cols(); } - inline Index outerStride() const { return m_expression.outerStride(); } - inline Index innerStride() const { return m_expression.innerStride(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); } - inline const CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const { return m_expression.coeff(row, col); } - inline Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { return m_expression.const_cast_derived().coeffRef(row, col); } - inline const CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const { return m_expression.coeff(index); } - inline Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { return m_expression.const_cast_derived().coeffRef(index); } @@ -90,7 +90,7 @@ template class ForceAlignedAccess m_expression.const_cast_derived().template writePacket(index, x); } - operator const ExpressionType&() const { return m_expression; } + EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; } protected: const ExpressionType& m_expression; diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index 3e68b1e91..3c67edae5 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -85,7 +85,7 @@ template class MapBase * * \sa innerStride(), outerStride() */ - inline const Scalar* data() const { return m_data; } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_data; } EIGEN_DEVICE_FUNC inline const Scalar& coeff(Index rowId, Index colId) const diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 048060e6b..001513187 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -180,7 +180,7 @@ template class MatrixBase #ifdef __CUDACC__ template EIGEN_DEVICE_FUNC - const typename LazyProductReturnType::Type + const Product operator*(const MatrixBase &other) const { return this->lazyProduct(other); } #else diff --git a/Eigen/src/Core/NestByValue.h b/Eigen/src/Core/NestByValue.h index 248dd8eb0..9aeaf8d18 100644 --- a/Eigen/src/Core/NestByValue.h +++ b/Eigen/src/Core/NestByValue.h @@ -40,29 +40,29 @@ template class NestByValue typedef typename internal::dense_xpr_base::type Base; EIGEN_DENSE_PUBLIC_INTERFACE(NestByValue) - explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {} + EIGEN_DEVICE_FUNC explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {} - inline Index rows() const { return m_expression.rows(); } - inline Index cols() const { return m_expression.cols(); } - inline Index outerStride() const { return m_expression.outerStride(); } - inline Index innerStride() const { return m_expression.innerStride(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); } + EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); } + EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); } - inline const CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const { return m_expression.coeff(row, col); } - inline Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { return m_expression.const_cast_derived().coeffRef(row, col); } - inline const CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const { return m_expression.coeff(index); } - inline Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { return m_expression.const_cast_derived().coeffRef(index); } @@ -91,7 +91,7 @@ template class NestByValue m_expression.const_cast_derived().template writePacket(index, x); } - operator const ExpressionType&() const { return m_expression; } + EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; } protected: const ExpressionType m_expression; diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h index ec7621d09..58fe0f6e0 100644 --- a/Eigen/src/Core/PlainObjectBase.h +++ b/Eigen/src/Core/PlainObjectBase.h @@ -221,11 +221,11 @@ class PlainObjectBase : public internal::dense_xpr_base::type } /** \returns a const pointer to the data array of this matrix */ - EIGEN_STRONG_INLINE const Scalar *data() const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar *data() const { return m_storage.data(); } /** \returns a pointer to the data array of this matrix */ - EIGEN_STRONG_INLINE Scalar *data() + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar *data() { return m_storage.data(); } /** Resizes \c *this to a \a rows x \a cols matrix. diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index ae64d5200..cb79543ef 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -68,7 +68,7 @@ struct traits > typename RhsTraits::StorageKind, internal::product_type::ret>::ret StorageKind; typedef typename promote_index_type::type Index; + typename RhsTraits::Index>::type Index; enum { RowsAtCompileTime = LhsTraits::RowsAtCompileTime, @@ -113,18 +113,18 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, typedef typename internal::remove_all::type LhsNestedCleaned; typedef typename internal::remove_all::type RhsNestedCleaned; - Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs) + EIGEN_DEVICE_FUNC Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs) { eigen_assert(lhs.cols() == rhs.rows() && "invalid matrix product" && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); } - inline Index rows() const { return m_lhs.rows(); } - inline Index cols() const { return m_rhs.cols(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); } - const LhsNestedCleaned& lhs() const { return m_lhs; } - const RhsNestedCleaned& rhs() const { return m_rhs; } + EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; } + EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; } protected: @@ -186,7 +186,7 @@ class ProductImpl public: - Scalar coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const { EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); @@ -194,7 +194,7 @@ class ProductImpl return typename internal::evaluator::type(derived()).coeff(row,col); } - Scalar coeff(Index i) const + EIGEN_DEVICE_FUNC Scalar coeff(Index i) const { EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS); eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) ); diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index c944ec9fc..3cebbbd12 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -35,7 +35,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - explicit evaluator(const XprType& xpr) : Base(xpr) {} + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {} }; // Catch scalar * ( A * B ) and transform it to (A*scalar) * B @@ -50,7 +50,7 @@ struct evaluator, const Produ typedef evaluator type; typedef evaluator nestedType; - explicit evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr.functor().m_other * xpr.nestedExpression().lhs() * xpr.nestedExpression().rhs()) {} }; @@ -66,7 +66,7 @@ struct evaluator, DiagIndex> > typedef evaluator type; typedef evaluator nestedType; - explicit evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(Diagonal, DiagIndex>( Product(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()), xpr.index() )) @@ -104,7 +104,7 @@ struct product_evaluator, ProductTag, LhsShape // CoeffReadCost = 0 // FIXME why is it needed? (this was already the case before the evaluators, see traits) }; - explicit product_evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); @@ -378,7 +378,7 @@ struct product_evaluator, ProductTag, DenseShape, typedef typename XprType::PacketScalar PacketScalar; typedef typename XprType::PacketReturnType PacketReturnType; - explicit product_evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) : m_lhs(xpr.lhs()), m_rhs(xpr.rhs()), m_lhsImpl(m_lhs), // FIXME the creation of the evaluator objects should result in a no-op, but check that! @@ -461,7 +461,7 @@ struct product_evaluator, ProductTag, DenseShape, && (InnerSize % packet_traits::size == 0) }; - const CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index row, Index col) const { // TODO check performance regression wrt to Eigen 3.2 which has special handling of this function return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum(); @@ -471,7 +471,7 @@ struct product_evaluator, ProductTag, DenseShape, * which is why we don't set the LinearAccessBit. * TODO: this seems possible when the result is a vector */ - const CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index index) const { const Index row = RowsAtCompileTime == 1 ? 0 : index; const Index col = RowsAtCompileTime == 1 ? index : 0; @@ -512,7 +512,7 @@ struct product_evaluator, LazyCoeffBasedProduc enum { Flags = Base::Flags | EvalBeforeNestingBit }; - explicit product_evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) : Base(BaseProduct(xpr.lhs(),xpr.rhs())) {} }; @@ -694,7 +694,7 @@ public: { } - EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const { return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx); } @@ -743,19 +743,21 @@ struct product_evaluator, ProductTag, DiagonalSha StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; - explicit product_evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) { } - EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const { return m_diagImpl.coeff(row) * m_matImpl.coeff(row, col); } +#ifndef __CUDACC__ template EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const { + // NVCC complains about template keyword, so we disable this function in CUDA mode return this->template packet_impl(row,col, row, typename internal::conditional::type()); } @@ -765,7 +767,7 @@ struct product_evaluator, ProductTag, DiagonalSha { return packet(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); } - +#endif }; // dense * diagonal @@ -787,16 +789,17 @@ struct product_evaluator, ProductTag, DenseShape, enum { StorageOrder = int(Lhs::Flags) & RowMajorBit ? RowMajor : ColMajor }; - explicit product_evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal()) { } - EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const { return m_matImpl.coeff(row, col) * m_diagImpl.coeff(col); } +#ifndef __CUDACC__ template EIGEN_STRONG_INLINE PacketScalar packet(Index row, Index col) const { @@ -809,7 +812,7 @@ struct product_evaluator, ProductTag, DenseShape, { return packet(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx); } - +#endif }; /*************************************************************************** diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index 14a2671e9..f6546917e 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -277,7 +277,7 @@ struct redux_impl typedef typename packet_traits::type PacketScalar; typedef typename Derived::Index Index; - static Scalar run(const Derived &mat, const Func& func) + EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func) { eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); const Index innerSize = mat.innerSize(); @@ -319,7 +319,7 @@ struct redux_impl Size = Derived::SizeAtCompileTime, VectorizedSize = (Size / PacketSize) * PacketSize }; - static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) + EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) { eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix"); if (VectorizedSize > 0) { @@ -340,7 +340,7 @@ class redux_evaluator { public: typedef _XprType XprType; - explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} + EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -359,15 +359,17 @@ public: CoeffReadCost = evaluator::CoeffReadCost }; - Index rows() const { return m_xpr.rows(); } - Index cols() const { return m_xpr.cols(); } - Index size() const { return m_xpr.size(); } - Index innerSize() const { return m_xpr.innerSize(); } - Index outerSize() const { return m_xpr.outerSize(); } + EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); } + EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); } + EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); } + EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); } + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { return m_evaluator.coeff(row, col); } + EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { return m_evaluator.coeff(index); } @@ -379,6 +381,7 @@ public: PacketReturnType packet(Index index) const { return m_evaluator.template packet(index); } + EIGEN_DEVICE_FUNC CoeffReturnType coeffByOuterInner(Index outer, Index inner) const { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); } diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h index 2653f2bbe..27fa178cd 100644 --- a/Eigen/src/Core/Ref.h +++ b/Eigen/src/Core/Ref.h @@ -127,12 +127,12 @@ public: typedef MapBase Base; EIGEN_DENSE_PUBLIC_INTERFACE(RefBase) - inline Index innerStride() const + EIGEN_DEVICE_FUNC inline Index innerStride() const { return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1; } - inline Index outerStride() const + EIGEN_DEVICE_FUNC inline Index outerStride() const { return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer() : IsVectorAtCompileTime ? this->size() @@ -140,7 +140,7 @@ public: : this->rows(); } - RefBase() + EIGEN_DEVICE_FUNC RefBase() : Base(0,RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime), // Stride<> does not allow default ctor for Dynamic strides, so let' initialize it with dummy values: m_stride(StrideType::OuterStrideAtCompileTime==Dynamic?0:StrideType::OuterStrideAtCompileTime, @@ -154,7 +154,7 @@ protected: typedef Stride StrideBase; template - void construct(Expression& expr) + EIGEN_DEVICE_FUNC void construct(Expression& expr) { if(PlainObjectType::RowsAtCompileTime==1) { @@ -192,13 +192,13 @@ template class Ref #ifndef EIGEN_PARSED_BY_DOXYGEN template - inline Ref(PlainObjectBase& expr, + EIGEN_DEVICE_FUNC inline Ref(PlainObjectBase& expr, typename internal::enable_if::MatchAtCompileTime),Derived>::type* = 0) { Base::construct(expr); } template - inline Ref(const DenseBase& expr, + EIGEN_DEVICE_FUNC inline Ref(const DenseBase& expr, typename internal::enable_if::value&&bool(Traits::template match::MatchAtCompileTime)),Derived>::type* = 0, int = Derived::ThisConstantIsPrivateInPlainObjectBase) #else @@ -224,7 +224,7 @@ template class Ref< EIGEN_DENSE_PUBLIC_INTERFACE(Ref) template - inline Ref(const DenseBase& expr) + EIGEN_DEVICE_FUNC inline Ref(const DenseBase& expr) { // std::cout << match_helper::HasDirectAccess << "," << match_helper::OuterStrideMatch << "," << match_helper::InnerStrideMatch << "\n"; // std::cout << int(StrideType::OuterStrideAtCompileTime) << " - " << int(Derived::OuterStrideAtCompileTime) << "\n"; @@ -232,25 +232,25 @@ template class Ref< construct(expr.derived(), typename Traits::template match::type()); } - inline Ref(const Ref& other) : Base(other) { + EIGEN_DEVICE_FUNC inline Ref(const Ref& other) : Base(other) { // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy } template - inline Ref(const RefBase& other) { + EIGEN_DEVICE_FUNC inline Ref(const RefBase& other) { construct(other.derived(), typename Traits::template match::type()); } protected: template - void construct(const Expression& expr,internal::true_type) + EIGEN_DEVICE_FUNC void construct(const Expression& expr,internal::true_type) { Base::construct(expr); } template - void construct(const Expression& expr, internal::false_type) + EIGEN_DEVICE_FUNC void construct(const Expression& expr, internal::false_type) { internal::call_assignment_no_alias(m_object,expr,internal::assign_op()); Base::construct(m_object); diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index 5fcd9e3fc..4e2a81b56 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -103,7 +103,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - explicit evaluator(const XprType& xpr) + EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : m_result(xpr.rows(), xpr.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h index 9ba6ea2e6..291300a4a 100644 --- a/Eigen/src/Core/Reverse.h +++ b/Eigen/src/Core/Reverse.h @@ -89,47 +89,47 @@ template class Reverse typedef internal::reverse_packet_cond reverse_packet; public: - explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { } + EIGEN_DEVICE_FUNC explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { } EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse) - inline Index rows() const { return m_matrix.rows(); } - inline Index cols() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); } - inline Index innerStride() const + EIGEN_DEVICE_FUNC inline Index innerStride() const { return -m_matrix.innerStride(); } - inline Scalar& operator()(Index row, Index col) + EIGEN_DEVICE_FUNC inline Scalar& operator()(Index row, Index col) { eigen_assert(row >= 0 && row < rows() && col >= 0 && col < cols()); return coeffRef(row, col); } - inline Scalar& coeffRef(Index row, Index col) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col) { return m_matrix.const_cast_derived().coeffRef(ReverseRow ? m_matrix.rows() - row - 1 : row, ReverseCol ? m_matrix.cols() - col - 1 : col); } - inline CoeffReturnType coeff(Index row, Index col) const + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index row, Index col) const { return m_matrix.coeff(ReverseRow ? m_matrix.rows() - row - 1 : row, ReverseCol ? m_matrix.cols() - col - 1 : col); } - inline CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC inline CoeffReturnType coeff(Index index) const { return m_matrix.coeff(m_matrix.size() - index - 1); } - inline Scalar& coeffRef(Index index) + EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index) { return m_matrix.const_cast_derived().coeffRef(m_matrix.size() - index - 1); } - inline Scalar& operator()(Index index) + EIGEN_DEVICE_FUNC inline Scalar& operator()(Index index) { eigen_assert(index >= 0 && index < m_matrix.size()); return coeffRef(index); @@ -164,7 +164,7 @@ template class Reverse m_matrix.const_cast_derived().template writePacket(m_matrix.size() - index - PacketSize, internal::preverse(x)); } - const typename internal::remove_all::type& + EIGEN_DEVICE_FUNC const typename internal::remove_all::type& nestedExpression() const { return m_matrix; diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 1c44d9c9a..b785e8e1e 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -228,11 +228,11 @@ public: typedef typename Base::AssignmentTraits AssignmentTraits; - triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) : Base(dst, src, func, dstExpr) {} - void assignCoeff(Index row, Index col) + EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col) { eigen_internal_assert(row!=col); Scalar tmp = m_src.coeff(row,col); @@ -240,12 +240,12 @@ public: m_functor.assignCoeff(m_dst.coeffRef(col,row), numext::conj(tmp)); } - void assignDiagonalCoeff(Index id) + EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id) { Base::assignCoeff(id,id); } - void assignOppositeCoeff(Index, Index) + EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index, Index) { eigen_internal_assert(false && "should never be called"); } }; diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index 641ffa218..3905cd616 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -121,7 +121,7 @@ struct evaluator > typedef evaluator type; typedef evaluator nestedType; - explicit evaluator(const SolveType& solve) + EIGEN_DEVICE_FUNC explicit evaluator(const SolveType& solve) : m_result(solve.rows(), solve.cols()) { ::new (static_cast(this)) Base(m_result); diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index 3277cb5ba..55319320a 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -32,7 +32,7 @@ public: typedef typename Base::DstXprType DstXprType; typedef swap_assign_op Functor; - generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr) + EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr) : Base(dst, src, func, dstExpr) {} diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 57d6fd2fe..a3b95256f 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -129,8 +129,8 @@ template class TransposeImpl const Scalar >::type ScalarWithConstIfNotLvalue; - inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } - inline const Scalar* data() const { return derived().nestedExpression().data(); } + EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); } + EIGEN_DEVICE_FUNC inline const Scalar* data() const { return derived().nestedExpression().data(); } // FIXME: shall we keep the const version of coeffRef? EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index 055ed7514..defe29cd4 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -692,12 +692,12 @@ public: typedef typename Base::AssignmentTraits AssignmentTraits; - triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) + EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr) : Base(dst, src, func, dstExpr) {} #ifdef EIGEN_INTERNAL_DEBUGGING - void assignCoeff(Index row, Index col) + EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col) { eigen_internal_assert(row!=col); Base::assignCoeff(row,col); @@ -706,14 +706,14 @@ public: using Base::assignCoeff; #endif - void assignDiagonalCoeff(Index id) + EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id) { if(Mode==UnitDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(1)); else if(Mode==ZeroDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(0)); else if(Mode==0) Base::assignCoeff(id,id); } - void assignOppositeCoeff(Index row, Index col) + EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index row, Index col) { eigen_internal_assert(row!=col); if(SetOpposite) @@ -722,7 +722,7 @@ public: }; template -void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) +EIGEN_DEVICE_FUNC void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src, const Functor &func) { eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols()); @@ -746,7 +746,7 @@ void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& sr } template -void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src) +EIGEN_DEVICE_FUNC void call_triangular_assignment_loop(const DstXprType& dst, const SrcXprType& src) { call_triangular_assignment_loop(dst, src, internal::assign_op()); } @@ -759,7 +759,7 @@ template<> struct AssignmentKind { typedef Dens template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> struct Assignment { - static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) { eigen_assert(int(DstXprType::Mode) == int(SrcXprType::Mode)); @@ -770,7 +770,7 @@ struct Assignment struct Assignment { - static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) { call_triangular_assignment_loop(dst, src, func); } @@ -779,7 +779,7 @@ struct Assignment template< typename DstXprType, typename SrcXprType, typename Functor, typename Scalar> struct Assignment { - static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) + EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func) { call_triangular_assignment_loop(dst, src, func); } diff --git a/Eigen/src/Core/functors/AssignmentFunctors.h b/Eigen/src/Core/functors/AssignmentFunctors.h index d4d85a1ca..161b0aa93 100644 --- a/Eigen/src/Core/functors/AssignmentFunctors.h +++ b/Eigen/src/Core/functors/AssignmentFunctors.h @@ -123,7 +123,7 @@ struct functor_traits > { /** \internal - * \brief Template functor for scalar/packet assignment with swaping + * \brief Template functor for scalar/packet assignment with swapping * * It works as follow. For a non-vectorized evaluation loop, we have: * for(i) func(A.coeffRef(i), B.coeff(i)); @@ -142,8 +142,13 @@ template struct swap_assign_op { EIGEN_EMPTY_STRUCT_CTOR(swap_assign_op) EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const { +#ifdef __CUDACC__ + // FIXME is there some kind of cuda::swap? + Scalar t=b; const_cast(b)=a; a=t; +#else using std::swap; swap(a,const_cast(b)); +#endif } template diff --git a/test/cuda_basic.cu b/test/cuda_basic.cu index 4c7e96c10..300bced02 100644 --- a/test/cuda_basic.cu +++ b/test/cuda_basic.cu @@ -65,7 +65,7 @@ struct redux { }; template -struct prod { +struct prod_test { EIGEN_DEVICE_FUNC void operator()(int i, const typename T1::Scalar* in, typename T1::Scalar* out) const { @@ -125,8 +125,8 @@ void test_cuda_basic() CALL_SUBTEST( run_and_compare_to_cuda(redux(), nthreads, in, out) ); CALL_SUBTEST( run_and_compare_to_cuda(redux(), nthreads, in, out) ); - CALL_SUBTEST( run_and_compare_to_cuda(prod(), nthreads, in, out) ); - CALL_SUBTEST( run_and_compare_to_cuda(prod(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(prod_test(), nthreads, in, out) ); + CALL_SUBTEST( run_and_compare_to_cuda(prod_test(), nthreads, in, out) ); CALL_SUBTEST( run_and_compare_to_cuda(diagonal(), nthreads, in, out) ); CALL_SUBTEST( run_and_compare_to_cuda(diagonal(), nthreads, in, out) ); From 4c70b0a7627d45286ecbb3c73d2d774412168205 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 13 Oct 2014 10:04:04 -0700 Subject: [PATCH 0894/1103] Added support for patch extraction --- unsupported/Eigen/CXX11/Tensor | 7 + .../Eigen/CXX11/src/Tensor/TensorBase.h | 6 + .../src/Tensor/TensorForwardDeclarations.h | 1 + .../Eigen/CXX11/src/Tensor/TensorPatch.h | 212 ++++++++++++++++++ unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_patch.cpp | 103 +++++++++ 6 files changed, 330 insertions(+) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h create mode 100644 unsupported/test/cxx11_tensor_patch.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 79510fd96..0dac95e45 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -1,6 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // +// Copyright (C) 2014 Benoit Steiner // Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla @@ -27,6 +28,11 @@ #include #include +#include + +#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) +#include +#endif #include "Eigen/Core" @@ -46,6 +52,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index cadeb3b19..27c10f64f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -243,6 +243,12 @@ class TensorBase return TensorConcatenationOp(derived(), other.derived(), axis); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorPatchOp + extract_patches(const PatchDims& patch_dims) const { + return TensorPatchOp(derived(), patch_dims); + } + // Morphing operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorReshapingOp diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 86ddd1ae8..67f478822 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -25,6 +25,7 @@ template class TensorReductionOp; template class TensorConcatenationOp; template class TensorContractionOp; template class TensorConvolutionOp; +template class TensorPatchOp; template class TensorBroadcastingOp; template class TensorChippingOp; template class TensorReshapingOp; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h new file mode 100644 index 000000000..01f2daf52 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h @@ -0,0 +1,212 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_PATCH_H +#define EIGEN_CXX11_TENSOR_TENSOR_PATCH_H + +namespace Eigen { + +/** \class TensorPatch + * \ingroup CXX11_Tensor_Module + * + * \brief Tensor patch class. + * + * + */ +namespace internal { +template +struct traits > : public traits +{ + typedef typename XprType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename traits::StorageKind StorageKind; + typedef typename traits::Index Index; + typedef typename XprType::Nested Nested; + typedef typename remove_reference::type _Nested; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorPatchOp& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef TensorPatchOp type; +}; + +} // end namespace internal + + + +template +class TensorPatchOp : public TensorBase, ReadOnlyAccessors> +{ + public: + typedef typename Eigen::internal::traits::Scalar Scalar; + typedef typename Eigen::internal::traits::Packet Packet; + typedef typename Eigen::NumTraits::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested::type Nested; + typedef typename Eigen::internal::traits::StorageKind StorageKind; + typedef typename Eigen::internal::traits::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorPatchOp(const XprType& expr, const PatchDim& patch_dims) + : m_xpr(expr), m_patch_dims(patch_dims) {} + + EIGEN_DEVICE_FUNC + const PatchDim& patch_dims() const { return m_patch_dims; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } + + protected: + typename XprType::Nested m_xpr; + const PatchDim m_patch_dims; +}; + + +// Eval as rvalue +template +struct TensorEvaluator, Device> +{ + typedef TensorPatchOp XprType; + typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::Dimensions>::value + 1; + typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; + + enum { + IsAligned = false, + PacketAccess = TensorEvaluator::PacketAccess, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device) + { + Index num_patches = 1; + const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + const PatchDim& patch_dims = op.patch_dims(); + for (int i = 0; i < NumDims-1; ++i) { + m_dimensions[i] = patch_dims[i]; + num_patches *= (input_dims[i] - patch_dims[i] + 1); + } + m_dimensions[NumDims-1] = num_patches; + + m_inputStrides[0] = 1; + m_patchStrides[0] = 1; + for (int i = 1; i < NumDims-1; ++i) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_patchStrides[i] = m_patchStrides[i-1] * (input_dims[i-1] - patch_dims[i-1] + 1); + } + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } + } + + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + m_impl.evalSubExprsIfNeeded(NULL); + return true; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { + m_impl.cleanup(); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + { + // Find the location of the first element of the patch. + Index patchIndex = index / m_outputStrides[NumDims - 1]; + // Find the offset of the element wrt the location of the first element. + Index patchOffset = index - patchIndex * m_outputStrides[NumDims - 1]; + + Index inputIndex = 0; + for (int i = NumDims - 2; i > 0; --i) { + const Index patchIdx = patchIndex / m_patchStrides[i]; + patchIndex -= patchIdx * m_patchStrides[i]; + const Index offsetIdx = patchOffset / m_outputStrides[i]; + patchOffset -= offsetIdx * m_outputStrides[i]; + inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i]; + } + inputIndex += (patchIndex + patchOffset); + return m_impl.coeff(inputIndex); + } + + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); + + Index indices[2] = {index, index + packetSize - 1}; + Index patchIndices[2] = {indices[0] / m_outputStrides[NumDims - 1], + indices[1] / m_outputStrides[NumDims - 1]}; + Index patchOffsets[2] = {indices[0] - patchIndices[0] * m_outputStrides[NumDims - 1], + indices[1] - patchIndices[1] * m_outputStrides[NumDims - 1]}; + + Index inputIndices[2] = {0, 0}; + for (int i = NumDims - 2; i > 0; --i) { + const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i], + patchIndices[1] / m_patchStrides[i]}; + patchIndices[0] -= patchIdx[0] * m_patchStrides[i]; + patchIndices[1] -= patchIdx[1] * m_patchStrides[i]; + + const Index offsetIdx[2] = {patchOffsets[0] / m_outputStrides[i], + patchOffsets[1] / m_outputStrides[i]}; + patchOffsets[0] -= offsetIdx[0] * m_outputStrides[i]; + patchOffsets[1] -= offsetIdx[1] * m_outputStrides[i]; + + inputIndices[0] += (patchIdx[0] + offsetIdx[0]) * m_inputStrides[i]; + inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i]; + } + inputIndices[0] += (patchIndices[0] + patchOffsets[0]); + inputIndices[1] += (patchIndices[1] + patchOffsets[1]); + + if (inputIndices[1] - inputIndices[0] == packetSize - 1) { + PacketReturnType rslt = m_impl.template packet(inputIndices[0]); + return rslt; + } + else { + EIGEN_ALIGN_DEFAULT CoeffReturnType values[packetSize]; + values[0] = m_impl.coeff(inputIndices[0]); + values[packetSize-1] = m_impl.coeff(inputIndices[1]); + for (int i = 1; i < packetSize-1; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + } + + Scalar* data() const { return NULL; } + + protected: + Dimensions m_dimensions; + array m_outputStrides; + array m_inputStrides; + array m_patchStrides; + + TensorEvaluator m_impl; +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_PATCH_H diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 99593b562..d6c435947 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -119,6 +119,7 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_concatenation "-std=c++0x") ei_add_test(cxx11_tensor_morphing "-std=c++0x") ei_add_test(cxx11_tensor_padding "-std=c++0x") + ei_add_test(cxx11_tensor_patch "-std=c++0x") ei_add_test(cxx11_tensor_reduction "-std=c++0x") ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_patch.cpp b/unsupported/test/cxx11_tensor_patch.cpp new file mode 100644 index 000000000..e2ba5bfd8 --- /dev/null +++ b/unsupported/test/cxx11_tensor_patch.cpp @@ -0,0 +1,103 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; + +static void test_simple_patch() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + array patch_dims; + patch_dims[0] = 1; + patch_dims[1] = 1; + patch_dims[2] = 1; + patch_dims[3] = 1; + + Tensor no_patch; + no_patch = tensor.extract_patches(patch_dims); + + VERIFY_IS_EQUAL(no_patch.dimension(0), 1); + VERIFY_IS_EQUAL(no_patch.dimension(1), 1); + VERIFY_IS_EQUAL(no_patch.dimension(2), 1); + VERIFY_IS_EQUAL(no_patch.dimension(3), 1); + VERIFY_IS_EQUAL(no_patch.dimension(4), tensor.size()); + + for (int i = 0; i < tensor.size(); ++i) { + VERIFY_IS_EQUAL(tensor.data()[i], no_patch.data()[i]); + } + + patch_dims[0] = 1; + patch_dims[1] = 2; + patch_dims[2] = 2; + patch_dims[3] = 1; + Tensor twod_patch; + twod_patch = tensor.extract_patches(patch_dims); + + VERIFY_IS_EQUAL(twod_patch.dimension(0), 1); + VERIFY_IS_EQUAL(twod_patch.dimension(1), 2); + VERIFY_IS_EQUAL(twod_patch.dimension(2), 2); + VERIFY_IS_EQUAL(twod_patch.dimension(3), 1); + VERIFY_IS_EQUAL(twod_patch.dimension(4), 2*2*4*7); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 2; ++j) { + for (int k = 0; k < 4; ++k) { + for (int l = 0; l < 7; ++l) { + int patch_loc = i + 2 * (j + 2 * (k + 4 * l)); + for (int x = 0; x < 2; ++x) { + for (int y = 0; y < 2; ++y) { + VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l), twod_patch(0,x,y,0,patch_loc)); + } + } + } + } + } + } + + patch_dims[0] = 1; + patch_dims[1] = 2; + patch_dims[2] = 3; + patch_dims[3] = 5; + Tensor threed_patch; + threed_patch = tensor.extract_patches(patch_dims); + + VERIFY_IS_EQUAL(threed_patch.dimension(0), 1); + VERIFY_IS_EQUAL(threed_patch.dimension(1), 2); + VERIFY_IS_EQUAL(threed_patch.dimension(2), 3); + VERIFY_IS_EQUAL(threed_patch.dimension(3), 5); + VERIFY_IS_EQUAL(threed_patch.dimension(4), 2*2*3*3); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 2; ++j) { + for (int k = 0; k < 3; ++k) { + for (int l = 0; l < 3; ++l) { + int patch_loc = i + 2 * (j + 2 * (k + 3 * l)); + for (int x = 0; x < 2; ++x) { + for (int y = 0; y < 3; ++y) { + for (int z = 0; z < 5; ++z) { + VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l+z), threed_patch(0,x,y,z,patch_loc)); + } + } + } + } + } + } + } +} + + +void test_cxx11_tensor_patch() +{ + CALL_SUBTEST(test_simple_patch()); + // CALL_SUBTEST(test_expr_shuffling()); +} From 99d75235a9567865d2c070a2840d54c8a5ad0f43 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 13 Oct 2014 17:02:09 -0700 Subject: [PATCH 0895/1103] Misc improvements and cleanups --- Eigen/src/Core/GenericPacketMath.h | 15 +- unsupported/Eigen/CXX11/Tensor | 4 + .../CXX11/src/Core/util/CXX11Workarounds.h | 5 + .../CXX11/src/Core/util/EmulateCXX11Meta.h | 101 +++++++- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorBase.h | 8 +- .../CXX11/src/Tensor/TensorBroadcasting.h | 8 +- .../CXX11/src/Tensor/TensorConvolution.h | 12 +- .../Eigen/CXX11/src/Tensor/TensorDevice.h | 35 +++ .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 73 +++--- .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorEvalTo.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 20 +- .../Eigen/CXX11/src/Tensor/TensorExecutor.h | 36 ++- .../Eigen/CXX11/src/Tensor/TensorFixedSize.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorIntDiv.h | 24 +- .../Eigen/CXX11/src/Tensor/TensorMap.h | 22 +- .../Eigen/CXX11/src/Tensor/TensorPadding.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorShuffling.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorStorage.h | 9 +- .../Eigen/CXX11/src/Tensor/TensorStriding.h | 61 +++-- .../Eigen/CXX11/src/Tensor/TensorTraits.h | 32 +-- unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_assign.cpp | 35 ++- unsupported/test/cxx11_tensor_convolution.cpp | 70 ++++++ unsupported/test/cxx11_tensor_device.cpp | 27 ++ unsupported/test/cxx11_tensor_morphing.cpp | 5 +- unsupported/test/cxx11_tensor_of_complex.cpp | 64 +++++ unsupported/test/cxx11_tensor_thread_pool.cpp | 232 +++++++++++++++++- 29 files changed, 779 insertions(+), 140 deletions(-) create mode 100644 unsupported/test/cxx11_tensor_of_complex.cpp diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index e6fea5bba..3ef3475c7 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -359,7 +359,7 @@ pmadd(const Packet& a, /** \internal \returns a packet version of \a *from. * If LoadMode equals #Aligned, \a from must be 16 bytes aligned */ template -inline Packet ploadt(const typename unpacket_traits::type* from) +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt(const typename unpacket_traits::type* from) { if(LoadMode == Aligned) return pload(from); @@ -370,7 +370,7 @@ inline Packet ploadt(const typename unpacket_traits::type* from) /** \internal copy the packet \a from to \a *to. * If StoreMode equals #Aligned, \a to must be 16 bytes aligned */ template -inline void pstoret(Scalar* to, const Packet& from) +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& from) { if(LoadMode == Aligned) pstore(to, from); @@ -378,6 +378,17 @@ inline void pstoret(Scalar* to, const Packet& from) pstoreu(to, from); } +/** \internal \returns a packet version of \a *from. + * Unlike ploadt, ploadt_ro takes advantage of the read-only memory path on the + * hardware if available to speedup the loading of data that won't be modified + * by the current computation. + */ +template +inline Packet ploadt_ro(const typename unpacket_traits::type* from) +{ + return ploadt(from); +} + /** \internal default implementation of palign() allowing partial specialization */ template struct palign_impl diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 0dac95e45..2137f4276 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -30,6 +30,10 @@ #include #include +#ifdef EIGEN_USE_THREADS +#include +#endif + #if defined(EIGEN_USE_GPU) && defined(__CUDACC__) #include #endif diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index 227522ecb..e30eb6ad8 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -66,6 +66,11 @@ template constexpr inline T& array_ template constexpr inline T&& array_get(std::array&& a) { return (T&&) STD_GET_ARR_HACK; } template constexpr inline T const& array_get(std::array const& a) { return (T const&) STD_GET_ARR_HACK; } +template constexpr inline T& array_get(std::vector& a) { return a[I]; } +template constexpr inline T&& array_get(std::vector&& a) { return a[I]; } +template constexpr inline T const& array_get(std::vector const& a) { return a[I]; } + + #undef STD_GET_ARR_HACK template struct array_size; diff --git a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h index 4c6b95773..e45d0a3b1 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h @@ -48,7 +48,8 @@ template class array { values[2] = v3; } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4) { + EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, + const T& v4) { EIGEN_STATIC_ASSERT(n==4, YOU_MADE_A_PROGRAMMING_MISTAKE) values[0] = v1; values[1] = v2; @@ -56,7 +57,8 @@ template class array { values[3] = v4; } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4, const T& v5) { + EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4, + const T& v5) { EIGEN_STATIC_ASSERT(n==5, YOU_MADE_A_PROGRAMMING_MISTAKE) values[0] = v1; values[1] = v2; @@ -64,6 +66,43 @@ template class array { values[3] = v4; values[4] = v5; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4, + const T& v5, const T& v6) { + EIGEN_STATIC_ASSERT(n==6, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v1; + values[1] = v2; + values[2] = v3; + values[3] = v4; + values[4] = v5; + values[5] = v6; + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array(const T& v1, const T& v2, const T& v3, const T& v4, + const T& v5, const T& v6, const T& v7) { + EIGEN_STATIC_ASSERT(n==7, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v1; + values[1] = v2; + values[2] = v3; + values[3] = v4; + values[4] = v5; + values[5] = v6; + values[6] = v7; + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE array( + const T& v1, const T& v2, const T& v3, const T& v4, + const T& v5, const T& v6, const T& v7, const T& v8) { + EIGEN_STATIC_ASSERT(n==8, YOU_MADE_A_PROGRAMMING_MISTAKE) + values[0] = v1; + values[1] = v2; + values[2] = v3; + values[3] = v4; + values[4] = v5; + values[5] = v6; + values[6] = v7; + values[7] = v8; + } #ifdef EIGEN_HAS_VARIADIC_TEMPLATES array(std::initializer_list l) { @@ -93,9 +132,11 @@ template struct type_list { struct null_type { }; -template +template struct make_type_list { - typedef typename make_type_list::type tailresult; + typedef typename make_type_list::type tailresult; typedef type_list type; }; @@ -150,6 +191,23 @@ template struct gen_numeric_list_repeated { typedef typename make_type_list, type2val, type2val, type2val, type2val >::type type; }; +template struct gen_numeric_list_repeated { + typedef typename make_type_list, type2val, type2val, + type2val, type2val, type2val >::type type; +}; + +template struct gen_numeric_list_repeated { + typedef typename make_type_list, type2val, type2val, + type2val, type2val, type2val, + type2val >::type type; +}; + +template struct gen_numeric_list_repeated { + typedef typename make_type_list, type2val, type2val, + type2val, type2val, type2val, + type2val, type2val >::type type; +}; + template struct get; @@ -174,6 +232,7 @@ template <> struct arg_prod { static const int value = 1; }; + template array repeat(t v) { array array; @@ -190,6 +249,11 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Head::type array_get(const type_l return get >::value; } +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename NList::HeadType::type array_prod(const NList& l) { + return arg_prod::value; +}; + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const array& a) { t prod = 1; @@ -201,6 +265,14 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const array& /*a*/) { return 0; } +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const std::vector& a) { + eigen_assert(a.size() > 0); + t prod = 1; + for (size_t i = 0; i < a.size(); ++i) { prod *= a[i]; } + return prod; +} + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T& array_get(array& a) { return a[I]; @@ -210,12 +282,31 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T& array_get(const array& a) { return a[I]; } +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T& array_get(std::vector& a) { + return a[I]; +} +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T& array_get(const std::vector& a) { + return a[I]; +} +template struct array_size; +template struct array_size > { + static const size_t value = N; +}; +template struct array_size; +template struct array_size& > { + static const size_t value = N; +}; template struct array_size; template struct array_size > { static const size_t value = N; }; - +template struct array_size; +template struct array_size& > { + static const size_t value = N; +}; struct sum_op { template static inline bool run(A a, B b) { return a + b; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index 3bfe80c9e..e973c00d3 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -131,8 +131,8 @@ struct TensorEvaluator, Device> m_leftImpl.coeffRef(i) = m_rightImpl.coeff(i); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalPacket(Index i) { - static const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; - static const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + const int LhsStoreMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; + const int RhsLoadMode = TensorEvaluator::IsAligned ? Aligned : Unaligned; m_leftImpl.template writePacket(i, m_rightImpl.template packet(i)); } EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 27c10f64f..6018ecc66 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -30,6 +30,12 @@ class TensorBase typedef Scalar CoeffReturnType; typedef typename internal::packet_traits::type PacketReturnType; + // Dimensions + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return derived().dimensions()[n]; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index size() const { return internal::array_prod(derived().dimensions()); } + // Nullary operators EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> @@ -187,7 +193,7 @@ class TensorBase } // Contractions. - typedef std::pair DimensionPair; + typedef Eigen::IndexPair DimensionPair; template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorContractionOp diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h index 3b2a9c8b9..0e55d4de1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h @@ -48,7 +48,7 @@ struct nested, 1, typename eval -class TensorBroadcastingOp : public TensorBase, WriteAccessors> +class TensorBroadcastingOp : public TensorBase, ReadOnlyAccessors> { public: typedef typename Eigen::internal::traits::Scalar Scalar; @@ -91,7 +91,7 @@ struct TensorEvaluator, Device> PacketAccess = TensorEvaluator::PacketAccess, }; -EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_impl(op.expression(), device) { const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); @@ -141,7 +141,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const D template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - static const int packetSize = internal::unpacket_traits::size; + const int packetSize = internal::unpacket_traits::size; EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) eigen_assert(index+packetSize-1 < dimensions().TotalSize()); @@ -161,7 +161,7 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const D if (innermostLoc + packetSize <= m_impl.dimensions()[0]) { return m_impl.template packet(inputIndex); } else { - EIGEN_ALIGN_DEFAULT CoeffReturnType values[packetSize]; + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; values[0] = m_impl.coeff(inputIndex); for (int i = 1; i < packetSize; ++i) { values[i] = coeff(originalIndex+i); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 4a5fd9c79..34bdd5309 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -872,11 +872,19 @@ struct TensorEvaluator + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(const Index index) const + { + eigen_assert(m_buf); + eigen_assert(index < m_dimensions.TotalSize()); + return internal::ploadt(m_buf+index); + } + private: // No assignment (copies are needed by the kernels) TensorEvaluator& operator = (const TensorEvaluator&); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h index 75519c9f5..649bdb308 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h @@ -38,6 +38,18 @@ template class TensorDevice { return *this; } + template + EIGEN_STRONG_INLINE TensorDevice& operator+=(const OtherDerived& other) { + typedef typename OtherDerived::Scalar Scalar; + typedef TensorCwiseBinaryOp, const ExpressionType, const OtherDerived> Sum; + Sum sum(m_expression, other); + typedef TensorAssignOp Assign; + Assign assign(m_expression, sum); + static const bool Vectorize = TensorEvaluator::PacketAccess; + internal::TensorExecutor::run(assign, m_device); + return *this; + } + protected: const DeviceType& m_device; ExpressionType& m_expression; @@ -58,6 +70,18 @@ template class TensorDevice + EIGEN_STRONG_INLINE TensorDevice& operator+=(const OtherDerived& other) { + typedef typename OtherDerived::Scalar Scalar; + typedef TensorCwiseBinaryOp, const ExpressionType, const OtherDerived> Sum; + Sum sum(m_expression, other); + typedef TensorAssignOp Assign; + Assign assign(m_expression, sum); + static const bool Vectorize = TensorEvaluator::PacketAccess; + internal::TensorExecutor::run(assign, m_device); + return *this; + } + protected: const ThreadPoolDevice& m_device; ExpressionType& m_expression; @@ -79,6 +103,17 @@ template class TensorDevice return *this; } + template + EIGEN_STRONG_INLINE TensorDevice& operator+=(const OtherDerived& other) { + typedef typename OtherDerived::Scalar Scalar; + typedef TensorCwiseBinaryOp, const ExpressionType, const OtherDerived> Sum; + Sum sum(m_expression, other); + typedef TensorAssignOp Assign; + Assign assign(m_expression, sum); + internal::TensorExecutor::run(assign, m_device); + return *this; + } + protected: const GpuDevice& m_device; ExpressionType m_expression; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h index fad342eab..5a6ff70e9 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -37,23 +37,41 @@ struct DefaultDevice { // Multiple cpu cores // We should really use a thread pool here but first we need to find a portable thread pool library. #ifdef EIGEN_USE_THREADS + +typedef std::future Future; + struct ThreadPoolDevice { - ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : /*pool_(pool), */num_threads_(num_cores) { } - size_t numThreads() const { return num_threads_; } + ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : num_threads_(num_cores) { } EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { return internal::aligned_malloc(num_bytes); } + EIGEN_STRONG_INLINE void deallocate(void* buffer) const { internal::aligned_free(buffer); } + EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const { ::memcpy(dst, src, n); } + EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { ::memset(buffer, c, n); } + EIGEN_STRONG_INLINE size_t numThreads() const { + return num_threads_; + } + + template + EIGEN_STRONG_INLINE Future enqueue(Function&& f, Args&&... args) const { + return std::async(std::launch::async, f, args...); + } + template + EIGEN_STRONG_INLINE void enqueueNoFuture(Function&& f, Args&&... args) const { + std::async(std::launch::async, f, args...); + } + private: // todo: NUMA, ... size_t num_threads_; @@ -63,40 +81,33 @@ struct ThreadPoolDevice { // GPU offloading #ifdef EIGEN_USE_GPU -static int m_numMultiProcessors = 0; -static int m_maxThreadsPerBlock = 0; -static int m_maxThreadsPerMultiProcessor = 0; +static cudaDeviceProp m_deviceProperties; +static bool m_devicePropInitialized = false; + +static void initializeDeviceProp() { + if (!m_devicePropInitialized) { + assert(cudaGetDeviceProperties(&m_deviceProperties, 0) == cudaSuccess); + m_devicePropInitialized = true; + } +} static inline int getNumCudaMultiProcessors() { - if (m_numMultiProcessors == 0) { - cudaDeviceProp deviceProp; - cudaGetDeviceProperties(&deviceProp, 0); - m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock; - m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor; - m_numMultiProcessors = deviceProp.multiProcessorCount; - } - return m_numMultiProcessors; + initializeDeviceProp(); + return m_deviceProperties.multiProcessorCount; } static inline int maxCudaThreadsPerBlock() { - if (m_maxThreadsPerBlock == 0) { - cudaDeviceProp deviceProp; - cudaGetDeviceProperties(&deviceProp, 0); - m_numMultiProcessors = deviceProp.multiProcessorCount; - m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor; - m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock; - } - return m_maxThreadsPerBlock; + initializeDeviceProp(); + return m_deviceProperties.maxThreadsPerBlock; } static inline int maxCudaThreadsPerMultiProcessor() { - if (m_maxThreadsPerBlock == 0) { - cudaDeviceProp deviceProp; - cudaGetDeviceProperties(&deviceProp, 0); - m_numMultiProcessors = deviceProp.multiProcessorCount; - m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock; - m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor; - } - return m_maxThreadsPerMultiProcessor; + initializeDeviceProp(); + return m_deviceProperties.maxThreadsPerMultiProcessor; } +static inline int sharedMemPerBlock() { + initializeDeviceProp(); + return m_deviceProperties.sharedMemPerBlock; +} + struct GpuDevice { // The cudastream is not owned: the caller is responsible for its initialization and eventual destruction. @@ -141,8 +152,8 @@ struct GpuDevice { #endif } - EIGEN_STRONG_INLINE size_t numThreads() const { - // Fixme: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t numThreads() const { + // FIXME return 32; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 732c6b344..2dd8e274b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -29,7 +29,7 @@ namespace Eigen { * \sa Tensor */ -// Can't use std::pairs on cuda devices +// Can't use std::pair on cuda devices template struct IndexPair { EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair() : first(0), second(0) { } EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair(Index f, Index s) : first(f), second(s) { } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h index 587cbd5ca..ce9d73578 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h @@ -116,7 +116,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalScalar(Index i) { m_buffer[i] = m_impl.coeff(i); } - EIGEN_STRONG_INLINE void evalPacket(Index i) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalPacket(Index i) { internal::pstoret(m_buffer + i, m_impl.template packet::IsAligned ? Aligned : Unaligned>(i)); } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index 0f969036c..e324ba8d2 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -65,13 +65,13 @@ struct TensorEvaluator return m_data[index]; } - template EIGEN_STRONG_INLINE + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return internal::ploadt(m_data + index); } - template EIGEN_STRONG_INLINE + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writePacket(Index index, const Packet& x) { return internal::pstoret(m_data + index, x); @@ -113,13 +113,17 @@ struct TensorEvaluator EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { eigen_assert(m_data); +#ifdef __CUDA_ARCH__ + return __ldg(m_data+index); +#else return m_data[index]; +#endif } - template EIGEN_STRONG_INLINE + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - return internal::ploadt(m_data + index); + return internal::ploadt_ro(m_data + index); } const Scalar* data() const { return m_data; } @@ -166,7 +170,7 @@ struct TensorEvaluator, Device> } template - EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return m_functor.packetOp(index); } @@ -219,7 +223,7 @@ struct TensorEvaluator, Device> } template - EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return m_functor.packetOp(m_argImpl.template packet(index)); } @@ -278,7 +282,7 @@ struct TensorEvaluator - EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { return m_functor.packetOp(m_leftImpl.template packet(index), m_rightImpl.template packet(index)); } @@ -340,7 +344,7 @@ struct TensorEvaluator return m_condImpl.coeff(index) ? m_thenImpl.coeff(index) : m_elseImpl.coeff(index); } template - PacketReturnType packet(Index index) const + EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const { static const int PacketSize = internal::unpacket_traits::size; internal::Selector select; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index 10f5a5ee7..01fa04c64 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -10,10 +10,6 @@ #ifndef EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H #define EIGEN_CXX11_TENSOR_TENSOR_EXECUTOR_H -#ifdef EIGEN_USE_THREADS -#include -#endif - namespace Eigen { /** \class TensorExecutor @@ -62,7 +58,7 @@ class TensorExecutor { const Index size = array_prod(evaluator.dimensions()); static const int PacketSize = unpacket_traits::PacketReturnType>::size; - const int VectorizedSize = (size / PacketSize) * PacketSize; + const Index VectorizedSize = (size / PacketSize) * PacketSize; for (Index i = 0; i < VectorizedSize; i += PacketSize) { evaluator.evalPacket(i); @@ -131,10 +127,10 @@ class TensorExecutor const Index numblocks = size / blocksize; Index i = 0; - std::vector > results; + std::vector results; results.reserve(numblocks); for (int i = 0; i < numblocks; ++i) { - results.push_back(std::async(std::launch::async, &EvalRange::run, &evaluator, i*blocksize, (i+1)*blocksize)); + results.push_back(device.enqueue(&EvalRange::run, &evaluator, i*blocksize, (i+1)*blocksize)); } for (int i = 0; i < numblocks; ++i) { @@ -154,11 +150,31 @@ class TensorExecutor // GPU: the evaluation of the expression is offloaded to a GPU. #if defined(EIGEN_USE_GPU) && defined(__CUDACC__) template -__global__ void EigenMetaKernel(Evaluator eval, unsigned int size) { +__global__ void +__launch_bounds__(1024) +EigenMetaKernel(Evaluator eval, unsigned int size) { + const int first_index = blockIdx.x * blockDim.x + threadIdx.x; const int step_size = blockDim.x * gridDim.x; - for (int i = first_index; i < size; i += step_size) { - eval.evalScalar(i); + + if (!Evaluator::PacketAccess || !Evaluator::IsAligned) { + // Use the scalar path + for (int i = first_index; i < size; i += step_size) { + eval.evalScalar(i); + } + } + else { + // Use the vector path + const int PacketSize = unpacket_traits::size; + const int vectorized_step_size = step_size * PacketSize; + const int vectorized_size = (size / PacketSize) * PacketSize; + int i = first_index * PacketSize; + for ( ; i < vectorized_size; i += vectorized_step_size) { + eval.evalPacket(i); + } + for ( ; i < size; i += step_size) { + eval.evalScalar(i); + } } } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h index 4d7f9e1fd..a753c5a48 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h @@ -17,7 +17,7 @@ namespace Eigen { * * \brief The fixed sized version of the tensor class. * - * The fixes sized equivalent of + * The fixed sized equivalent of * Eigen::Tensor t(3, 5, 7); * is * Eigen::TensorFixedSize> t; @@ -41,7 +41,7 @@ class TensorFixedSize : public TensorBase::size > 1), }; typedef Dimensions_ Dimensions; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h index cf97031be..2714117ab 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h @@ -31,30 +31,34 @@ namespace internal { template struct TensorIntDivisor { public: - TensorIntDivisor() { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() { multiplier = 0; shift1 = 0; shift2 = 0; } // Must have 1 <= divider <= 2^31-1 - TensorIntDivisor(const T divider) { - static const int N = 32; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor(const T divider) { + const int N = 32; eigen_assert(divider > 0); eigen_assert(divider <= (1<<(N-1)) - 1); // fast ln2 +#ifndef __CUDA_ARCH__ const int leading_zeros = __builtin_clz(divider); - const int l = N - (leading_zeros+1); +#else + const int leading_zeros = __clz(divider); +#endif + const int log_div = N - (leading_zeros+1); - multiplier = (static_cast(1) << (N+l)) / divider - (static_cast(1) << N) + 1; - shift1 = (std::min)(1, l); - shift2 = (std::max)(0, l-1); + multiplier = (static_cast(1) << (N+log_div)) / divider - (static_cast(1) << N) + 1; + shift1 = log_div > 1 ? 1 : log_div; + shift2 = log_div > 1 ? log_div-1 : 0; } // Must have 0 <= numerator <= 2^32-1 - T divide(const T numerator) const { - static const int N = 32; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T divide(const T numerator) const { + const int N = 32; eigen_assert(numerator >= 0); eigen_assert(numerator <= (1ull< -static T operator / (const T& numerator, const TensorIntDivisor& divisor) { +static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator / (const T& numerator, const TensorIntDivisor& divisor) { return divisor.divide(numerator); } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index 04849dd9f..2c0d2cd0f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -42,26 +42,25 @@ template class TensorMap : public Tensor static const int Options = Options_; - static const std::size_t NumIndices = PlainObjectType::NumIndices; + static const Index NumIndices = PlainObjectType::NumIndices; typedef typename PlainObjectType::Dimensions Dimensions; - enum { - IsAligned = bool(EIGEN_ALIGN) && ((int(Options_)&Aligned)==Aligned), - PacketAccess = true, + IsAligned = ((int(Options_)&Aligned)==Aligned), + PacketAccess = (internal::packet_traits::size > 1), }; #ifdef EIGEN_HAS_VARIADIC_TEMPLATES template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions(array({{firstDimension, otherDimensions...}})) { + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension, IndexTypes... otherDimensions) : m_data(dataPtr), m_dimensions(firstDimension, otherDimensions...) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. - EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + EIGEN_STATIC_ASSERT((sizeof...(otherDimensions) + 1 == NumIndices || NumIndices == Dynamic), YOU_MADE_A_PROGRAMMING_MISTAKE) } #else EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions(array(firstDimension)) { + EIGEN_STRONG_INLINE TensorMap(PointerArgType dataPtr, Index firstDimension) : m_data(dataPtr), m_dimensions(firstDimension) { // The number of dimensions used to construct a tensor must be equal to the rank of the tensor. - EIGEN_STATIC_ASSERT(1 == NumIndices, YOU_MADE_A_PROGRAMMING_MISTAKE) + EIGEN_STATIC_ASSERT((1 == NumIndices || NumIndices == Dynamic), YOU_MADE_A_PROGRAMMING_MISTAKE) } #endif @@ -176,12 +175,13 @@ template class TensorMap : public Tensor template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& operator()(Index firstIndex, IndexTypes... otherIndices) { - static_assert(sizeof...(otherIndices) + 1 == NumIndices, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + static_assert(sizeof...(otherIndices) + 1 == NumIndices || NumIndices == Dynamic, "Number of indices used to access a tensor coefficient must be equal to the rank of the tensor."); + const std::size_t NumDims = sizeof...(otherIndices) + 1; if (PlainObjectType::Options&RowMajor) { - const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); + const Index index = m_dimensions.IndexOfRowMajor(array{{firstIndex, otherIndices...}}); return m_data[index]; } else { - const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); + const Index index = m_dimensions.IndexOfColMajor(array{{firstIndex, otherIndices...}}); return m_data[index]; } } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h index 7da89458f..8da6e0f26 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h @@ -144,7 +144,7 @@ struct TensorEvaluator, Device template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - static const int packetSize = internal::unpacket_traits::size; + const int packetSize = internal::unpacket_traits::size; EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) eigen_assert(index+packetSize-1 < dimensions().TotalSize()); @@ -206,7 +206,7 @@ struct TensorEvaluator, Device EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const { - static const int packetSize = internal::unpacket_traits::size; + const int packetSize = internal::unpacket_traits::size; EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; for (int i = 0; i < packetSize; ++i) { values[i] = coeff(index+i); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h index f7e7fc107..7e0063626 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h @@ -97,7 +97,7 @@ struct TensorEvaluator, Device> typedef typename XprType::Scalar Scalar; enum { - IsAligned = true, + IsAligned = false, PacketAccess = (internal::packet_traits::size > 1), }; @@ -194,7 +194,7 @@ struct TensorEvaluator, Device> typedef typename XprType::Scalar Scalar; enum { - IsAligned = true, + IsAligned = false, PacketAccess = (internal::packet_traits::size > 1), }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index 0c4f8a3d6..aaec39756 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -30,11 +30,11 @@ namespace Eigen { * * \sa Tensor */ -template class TensorStorage; +template class TensorStorage; // Pure fixed-size storage -template +template class TensorStorage { private: @@ -62,7 +62,7 @@ class TensorStorage // pure-dynamic, but without specification of all dimensions explicitly -template +template class TensorStorage : public TensorStorage::type> { @@ -79,7 +79,7 @@ class TensorStorage }; // pure dynamic -template +template class TensorStorage::type> { T *m_data; @@ -140,6 +140,7 @@ class TensorStorage, 1, typename eval -class TensorStridingOp : public TensorBase, WriteAccessors> +class TensorStridingOp : public TensorBase > { public: typedef typename Eigen::internal::traits::Scalar Scalar; @@ -97,7 +97,7 @@ struct TensorEvaluator, Device> enum { IsAligned = /*TensorEvaluator::IsAligned*/false, - PacketAccess = /*TensorEvaluator::PacketAccess*/false, + PacketAccess = TensorEvaluator::PacketAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -109,28 +109,23 @@ struct TensorEvaluator, Device> } const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; - m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; - } else { - m_inputStrides[0] = 1; - m_outputStrides[0] = 1; - } - } - for (int i = 0; i < NumDims; ++i) { - m_inputStrides[i] *= op.strides()[i]; + m_outputStrides[0] = 1; + m_inputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_inputStrides[i-1] *= op.strides()[i-1]; } + m_inputStrides[NumDims-1] *= op.strides()[NumDims-1]; } - // typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) { m_impl.evalSubExprsIfNeeded(NULL); return true; } @@ -150,16 +145,44 @@ struct TensorEvaluator, Device> return m_impl.coeff(inputIndex); } - /* template + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - return m_impl.template packet(index); - }*/ + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); + + Index inputIndices[] = {0, 0}; + Index indices[] = {index, index + packetSize - 1}; + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / m_outputStrides[i]; + const Index idx1 = indices[1] / m_outputStrides[i]; + inputIndices[0] += idx0 * m_inputStrides[i]; + inputIndices[1] += idx1 * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += indices[0] * m_inputStrides[0]; + inputIndices[1] += indices[1] * m_inputStrides[0]; + if (inputIndices[1] - inputIndices[0] == packetSize - 1) { + PacketReturnType rslt = m_impl.template packet(inputIndices[0]); + return rslt; + } + else { + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + values[0] = m_impl.coeff(inputIndices[0]); + values[packetSize-1] = m_impl.coeff(inputIndices[1]); + for (int i = 1; i < packetSize-1; ++i) { + values[i] = coeff(index+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + } Scalar* data() const { return NULL; } protected: - // Strides m_strides; Dimensions m_dimensions; array m_outputStrides; array m_inputStrides; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h index 40f805741..5940a8cf1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h @@ -70,14 +70,18 @@ struct traits > }; -template -struct traits > +template +struct traits > : public traits { typedef traits BaseTraits; typedef typename BaseTraits::Scalar Scalar; typedef typename BaseTraits::StorageKind StorageKind; typedef typename BaseTraits::Index Index; + enum { + Options = Options_, + Flags = ((BaseTraits::Flags | LvalueBit) & ~AlignedBit) | (Options&Aligned ? AlignedBit : 0), + }; }; @@ -105,16 +109,16 @@ struct eval, Eigen::Dense> typedef const TensorFixedSize& type; }; -template -struct eval, Eigen::Dense> +template +struct eval, Eigen::Dense> { - typedef const TensorMap& type; + typedef const TensorMap& type; }; -template -struct eval, Eigen::Dense> +template +struct eval, Eigen::Dense> { - typedef const TensorMap& type; + typedef const TensorMap& type; }; template @@ -141,16 +145,16 @@ struct nested, 1, typename e typedef const TensorFixedSize& type; }; -template -struct nested, 1, typename eval >::type> +template +struct nested, 1, typename eval >::type> { - typedef const TensorMap& type; + typedef const TensorMap& type; }; -template -struct nested, 1, typename eval >::type> +template +struct nested, 1, typename eval >::type> { - typedef const TensorMap& type; + typedef const TensorMap& type; }; } // end namespace internal diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index d6c435947..a7ef2b402 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -110,6 +110,7 @@ if(EIGEN_TEST_CXX11) # ei_add_test(cxx11_tensor_fixed_size "-std=c++0x") ei_add_test(cxx11_tensor_const "-std=c++0x") ei_add_test(cxx11_tensor_of_const_values "-std=c++0x") + ei_add_test(cxx11_tensor_of_complex "-std=c++0x") ei_add_test(cxx11_tensor_of_strings "-std=c++0x") ei_add_test(cxx11_tensor_intdiv "-std=c++0x") ei_add_test(cxx11_tensor_lvalue "-std=c++0x") diff --git a/unsupported/test/cxx11_tensor_assign.cpp b/unsupported/test/cxx11_tensor_assign.cpp index f2b126413..0ac3f9bf9 100644 --- a/unsupported/test/cxx11_tensor_assign.cpp +++ b/unsupported/test/cxx11_tensor_assign.cpp @@ -253,6 +253,39 @@ static void test_auto_resize() } +static void test_compound_assign() +{ + Tensor start_tensor(10); + Tensor offset_tensor(10); + start_tensor.setRandom(); + offset_tensor.setRandom(); + + Tensor tensor = start_tensor; + tensor += offset_tensor; + for (int i = 0; i < 10; ++i) { + VERIFY_IS_EQUAL(tensor(i), start_tensor(i) + offset_tensor(i)); + } + + tensor = start_tensor; + tensor -= offset_tensor; + for (int i = 0; i < 10; ++i) { + VERIFY_IS_EQUAL(tensor(i), start_tensor(i) - offset_tensor(i)); + } + + tensor = start_tensor; + tensor *= offset_tensor; + for (int i = 0; i < 10; ++i) { + VERIFY_IS_EQUAL(tensor(i), start_tensor(i) * offset_tensor(i)); + } + + tensor = start_tensor; + tensor /= offset_tensor; + for (int i = 0; i < 10; ++i) { + VERIFY_IS_EQUAL(tensor(i), start_tensor(i) / offset_tensor(i)); + } +} + + void test_cxx11_tensor_assign() { CALL_SUBTEST(test_1d()); @@ -260,5 +293,5 @@ void test_cxx11_tensor_assign() CALL_SUBTEST(test_3d()); CALL_SUBTEST(test_same_type()); CALL_SUBTEST(test_auto_resize()); - + CALL_SUBTEST(test_compound_assign()); } diff --git a/unsupported/test/cxx11_tensor_convolution.cpp b/unsupported/test/cxx11_tensor_convolution.cpp index bafe73edd..4672db463 100644 --- a/unsupported/test/cxx11_tensor_convolution.cpp +++ b/unsupported/test/cxx11_tensor_convolution.cpp @@ -64,8 +64,78 @@ static void test_expr() } +static void test_modes() { + Tensor input(3); + Tensor kernel(3); + input(0) = 1.0f; + input(1) = 2.0f; + input(2) = 3.0f; + kernel(0) = 0.5f; + kernel(1) = 1.0f; + kernel(2) = 0.0f; + + const Eigen::array dims{{0}}; + Eigen::array, 1> padding; + + // Emulate VALID mode (as defined in + // http://docs.scipy.org/doc/numpy/reference/generated/numpy.convolve.html). + padding[0] = std::make_pair(0, 0); + Tensor valid(1); + valid = input.pad(padding).convolve(kernel, dims); + VERIFY_IS_EQUAL(valid.dimension(0), 1); + VERIFY_IS_APPROX(valid(0), 2.5f); + + // Emulate SAME mode (as defined in + // http://docs.scipy.org/doc/numpy/reference/generated/numpy.convolve.html). + padding[0] = std::make_pair(1, 1); + Tensor same(3); + same = input.pad(padding).convolve(kernel, dims); + VERIFY_IS_EQUAL(same.dimension(0), 3); + VERIFY_IS_APPROX(same(0), 1.0f); + VERIFY_IS_APPROX(same(1), 2.5f); + VERIFY_IS_APPROX(same(2), 4.0f); + + // Emulate FULL mode (as defined in + // http://docs.scipy.org/doc/numpy/reference/generated/numpy.convolve.html). + padding[0] = std::make_pair(2, 2); + Tensor full(5); + full = input.pad(padding).convolve(kernel, dims); + VERIFY_IS_EQUAL(full.dimension(0), 5); + VERIFY_IS_APPROX(full(0), 0.0f); + VERIFY_IS_APPROX(full(1), 1.0f); + VERIFY_IS_APPROX(full(2), 2.5f); + VERIFY_IS_APPROX(full(3), 4.0f); + VERIFY_IS_APPROX(full(4), 1.5f); +} + + +static void test_strides() { + Tensor input(13); + Tensor kernel(3); + input.setRandom(); + kernel.setRandom(); + + const Eigen::array dims{{0}}; + const Eigen::array stride_of_3{{3}}; + const Eigen::array stride_of_2{{2}}; + + Tensor result; + result = input.stride(stride_of_3).convolve(kernel, dims).stride(stride_of_2); + + VERIFY_IS_EQUAL(result.dimension(0), 2); + VERIFY_IS_APPROX(result(0), (input(0)*kernel(0) + input(3)*kernel(1) + + input(6)*kernel(2))); + VERIFY_IS_APPROX(result(1), (input(6)*kernel(0) + input(9)*kernel(1) + + input(12)*kernel(2))); +} + + + + void test_cxx11_tensor_convolution() { CALL_SUBTEST(test_evals()); CALL_SUBTEST(test_expr()); + CALL_SUBTEST(test_modes()); + CALL_SUBTEST(test_strides()); } diff --git a/unsupported/test/cxx11_tensor_device.cpp b/unsupported/test/cxx11_tensor_device.cpp index f331cb481..26465ee11 100644 --- a/unsupported/test/cxx11_tensor_device.cpp +++ b/unsupported/test/cxx11_tensor_device.cpp @@ -123,6 +123,14 @@ static void test_forced_contextual_eval(Context* context) context->out().device(context->device()) = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f); } +template +static void test_compound_assignment(Context* context) +{ + context->out().device(context->device()) = context->in1().constant(2.718f); + context->out().device(context->device()) += context->in1() + context->in2() * 3.14f; +} + + template static void test_contraction(Context* context) { @@ -197,6 +205,15 @@ static void test_cpu() { } } + test_compound_assignment(&context); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 50; ++j) { + for (int k = 0; k < 70; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); + } + } + } + test_contraction(&context); for (int i = 0; i < 40; ++i) { for (int j = 0; j < 40; ++j) { @@ -299,6 +316,16 @@ static void test_gpu() { } } + test_compound_assignment(&context); + assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess); + for (int i = 0; i < 40; ++i) { + for (int j = 0; j < 50; ++j) { + for (int k = 0; k < 70; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); + } + } + } + test_contraction(&context); assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess); for (int i = 0; i < 40; ++i) { diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp index 2a6a97856..fd1b1fa32 100644 --- a/unsupported/test/cxx11_tensor_morphing.cpp +++ b/unsupported/test/cxx11_tensor_morphing.cpp @@ -12,6 +12,7 @@ #include using Eigen::Tensor; +using Eigen::IndexPair; static void test_simple_reshape() { @@ -52,7 +53,7 @@ static void test_reshape_in_expr() { TensorMap> tensor2(m2.data(), 3,5,7,11,13); Tensor::Dimensions newDims1{{2,3*5*7*11}}; Tensor::Dimensions newDims2{{3*5*7*11,13}}; - array::DimensionPair, 1> contract_along{{std::make_pair(1, 0)}}; + Eigen::array, 1> contract_along{{IndexPair(1, 0)}}; Tensor tensor3(2,13); tensor3 = tensor1.reshape(newDims1).contract(tensor2.reshape(newDims2), contract_along); @@ -125,7 +126,7 @@ static void test_slice_in_expr() { TensorMap> tensor1(m1.data(), 7, 7); TensorMap> tensor2(m2.data(), 3, 3); Tensor tensor3(3,1); - array::DimensionPair, 1> contract_along{{std::make_pair(1, 0)}}; + array, 1> contract_along{{IndexPair(1, 0)}}; Eigen::DSizes indices1(1,2); Eigen::DSizes sizes1(3,3); diff --git a/unsupported/test/cxx11_tensor_of_complex.cpp b/unsupported/test/cxx11_tensor_of_complex.cpp new file mode 100644 index 000000000..b5044b962 --- /dev/null +++ b/unsupported/test/cxx11_tensor_of_complex.cpp @@ -0,0 +1,64 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::TensorMap; + + + +static void test_additions() +{ + Tensor, 1> data1(3); + Tensor, 1> data2(3); + for (int i = 0; i < 3; ++i) { + data1(i) = std::complex(i, -i); + data2(i) = std::complex(i, 7 * i); + } + + Tensor, 1> sum = data1 + data2; + for (int i = 0; i < 3; ++i) { + VERIFY_IS_EQUAL(sum(i), std::complex(2*i, 6*i)); + } +} + + +static void test_contractions() +{ + Tensor, 4> t_left(30, 50, 8, 31); + Tensor, 5> t_right(8, 31, 7, 20, 10); + Tensor, 5> t_result(30, 50, 7, 20, 10); + + t_left.setRandom(); + t_right.setRandom(); + + typedef Map, Dynamic, Dynamic>> MapXcf; + MapXcf m_left(t_left.data(), 1500, 248); + MapXcf m_right(t_right.data(), 248, 1400); + Matrix, Dynamic, Dynamic> m_result(1500, 1400); + + // This contraction should be equivalent to a regular matrix multiplication + typedef Tensor::DimensionPair DimPair; + Eigen::array dims({{DimPair(2, 0), DimPair(3, 1)}}); + t_result = t_left.contract(t_right, dims); + m_result = m_left * m_right; + for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { + VERIFY_IS_APPROX(t_result.data()[i], m_result.data()[i]); + } +} + + +void test_cxx11_tensor_of_complex() +{ + CALL_SUBTEST(test_additions()); + CALL_SUBTEST(test_contractions()); +} diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp index e02d8e4be..f0de61f8b 100644 --- a/unsupported/test/cxx11_tensor_thread_pool.cpp +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -9,22 +9,23 @@ #define EIGEN_USE_THREADS - +#include #include "main.h" #include + using Eigen::Tensor; -void test_cxx11_tensor_thread_pool() +static void test_multithread_elementwise() { - Eigen::Tensor in1(2,3,7); - Eigen::Tensor in2(2,3,7); - Eigen::Tensor out(2,3,7); + Tensor in1(2,3,7); + Tensor in2(2,3,7); + Tensor out(2,3,7); in1.setRandom(); in2.setRandom(); - Eigen::ThreadPoolDevice thread_pool_device(3); + Eigen::ThreadPoolDevice thread_pool_device(internal::random(3, 11)); out.device(thread_pool_device) = in1 + in2 * 3.14f; for (int i = 0; i < 2; ++i) { @@ -35,3 +36,222 @@ void test_cxx11_tensor_thread_pool() } } } + + +static void test_multithread_compound_assignment() +{ + Tensor in1(2,3,7); + Tensor in2(2,3,7); + Tensor out(2,3,7); + + in1.setRandom(); + in2.setRandom(); + + Eigen::ThreadPoolDevice thread_pool_device(internal::random(3, 11)); + out.device(thread_pool_device) = in1; + out.device(thread_pool_device) += in2 * 3.14f; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f); + } + } + } +} + + +static void test_multithread_contraction() +{ + Tensor t_left(30, 50, 37, 31); + Tensor t_right(37, 31, 70, 2, 10); + Tensor t_result(30, 50, 70, 2, 10); + + t_left.setRandom(); + t_right.setRandom(); + + // this contraction should be equivalent to a single matrix multiplication + typedef Tensor::DimensionPair DimPair; + Eigen::array dims({{DimPair(2, 0), DimPair(3, 1)}}); + + + typedef Map MapXf; + MapXf m_left(t_left.data(), 1500, 1147); + MapXf m_right(t_right.data(), 1147, 1400); + MatrixXf m_result(1500, 1400); + + Eigen::ThreadPoolDevice thread_pool_device(4); + + // compute results by separate methods + t_result.device(thread_pool_device) = t_left.contract(t_right, dims); + m_result = m_left * m_right; + + for (ptrdiff_t i = 0; i < t_result.size(); i++) { + VERIFY(&t_result.data()[i] != &m_result.data()[i]); + if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { + std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; + assert(false); + } + } +} + + +static void test_contraction_corner_cases() +{ + Tensor t_left(32, 500); + Tensor t_right(32, 28*28); + Tensor t_result(500, 28*28); + + t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f; + t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f; + t_result = t_result.constant(NAN); + + // this contraction should be equivalent to a single matrix multiplication + typedef Tensor::DimensionPair DimPair; + Eigen::array dims{{DimPair(0, 0)}}; + + typedef Map MapXf; + MapXf m_left(t_left.data(), 32, 500); + MapXf m_right(t_right.data(), 32, 28*28); + MatrixXf m_result(500, 28*28); + + Eigen::ThreadPoolDevice thread_pool_device(12); + + // compute results by separate methods + t_result.device(thread_pool_device) = t_left.contract(t_right, dims); + m_result = m_left.transpose() * m_right; + + for (ptrdiff_t i = 0; i < t_result.size(); i++) { + assert(!isnan(t_result.data()[i])); + if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { + std::cout << "mismatch detected at index " << i << " : " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; + assert(false); + } + } + + t_left.resize(32, 1); + t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f; + t_result.resize (1, 28*28); + t_result = t_result.constant(NAN); + t_result.device(thread_pool_device) = t_left.contract(t_right, dims); + new(&m_left) MapXf(t_left.data(), 32, 1); + m_result = m_left.transpose() * m_right; + for (ptrdiff_t i = 0; i < t_result.size(); i++) { + assert(!isnan(t_result.data()[i])); + if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { + std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; + assert(false); + } + } + + t_left.resize(32, 500); + t_right.resize(32, 4); + t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f; + t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f; + t_result.resize (500, 4); + t_result = t_result.constant(NAN); + t_result.device(thread_pool_device) = t_left.contract(t_right, dims); + new(&m_left) MapXf(t_left.data(), 32, 500); + new(&m_right) MapXf(t_right.data(), 32, 4); + m_result = m_left.transpose() * m_right; + for (ptrdiff_t i = 0; i < t_result.size(); i++) { + assert(!isnan(t_result.data()[i])); + if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { + std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; + assert(false); + } + } + + t_left.resize(32, 1); + t_right.resize(32, 4); + t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f; + t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f; + t_result.resize (1, 4); + t_result = t_result.constant(NAN); + t_result.device(thread_pool_device) = t_left.contract(t_right, dims); + new(&m_left) MapXf(t_left.data(), 32, 1); + new(&m_right) MapXf(t_right.data(), 32, 4); + m_result = m_left.transpose() * m_right; + for (ptrdiff_t i = 0; i < t_result.size(); i++) { + assert(!isnan(t_result.data()[i])); + if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { + std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; + assert(false); + } + } +} + + +static void test_multithread_contraction_agrees_with_singlethread() { + int contract_size = internal::random(1, 5000); + + Tensor left(internal::random(1, 80), + contract_size, + internal::random(1, 100)); + + Tensor right(internal::random(1, 25), + internal::random(1, 37), + contract_size, + internal::random(1, 51)); + + left.setRandom(); + right.setRandom(); + + // add constants to shift values away from 0 for more precision + left += left.constant(1.5f); + right += right.constant(1.5f); + + typedef Tensor::DimensionPair DimPair; + Eigen::array dims({{DimPair(1, 2)}}); + + Eigen::ThreadPoolDevice thread_pool_device(internal::random(2, 11)); + + Tensor st_result; + st_result = left.contract(right, dims); + + Tensor tp_result(st_result.dimensions()); + tp_result.device(thread_pool_device) = left.contract(right, dims); + + VERIFY(internal::dimensions_match(st_result.dimensions(), tp_result.dimensions())); + for (ptrdiff_t i = 0; i < st_result.size(); i++) { + // if both of the values are very small, then do nothing (because the test will fail + // due to numerical precision issues when values are small) + if (fabs(st_result.data()[i] - tp_result.data()[i]) >= 1e-4) { + VERIFY_IS_APPROX(st_result.data()[i], tp_result.data()[i]); + } + } +} + + +static void test_memcpy() { + + for (int i = 0; i < 5; ++i) { + const int num_threads = internal::random(3, 11); + Eigen::ThreadPoolDevice thread_pool_device(num_threads); + + const int size = internal::random(13, 7632); + Tensor t1(size); + t1.setRandom(); + std::vector result(size); + thread_pool_device.memcpy(&result[0], t1.data(), size*sizeof(float)); + for (int i = 0; i < size; i++) { + VERIFY_IS_EQUAL(t1(i), result[i]); + } + } +} + + +void test_cxx11_tensor_thread_pool() +{ + CALL_SUBTEST(test_multithread_elementwise()); + CALL_SUBTEST(test_multithread_compound_assignment()); + + CALL_SUBTEST(test_multithread_contraction()); + + CALL_SUBTEST(test_multithread_contraction_agrees_with_singlethread()); + + // Exercise various cases that have been problematic in the past. + CALL_SUBTEST(test_contraction_corner_cases()); + + CALL_SUBTEST(test_memcpy()); +} From 0ec1fc9e114a2fb8fd0cbeee38669547f46804c8 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 14 Oct 2014 14:14:25 +0200 Subject: [PATCH 0896/1103] bug #891: Determine sizeof(void*) via CMAKE variable instead of test program --- cmake/EigenTesting.cmake | 18 +++++------------- 1 file changed, 5 insertions(+), 13 deletions(-) diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index 65bb29413..3ed002a87 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -489,20 +489,12 @@ macro(ei_set_build_string) endmacro(ei_set_build_string) macro(ei_is_64bit_env VAR) - - file(WRITE "${CMAKE_CURRENT_BINARY_DIR}/is64.cpp" - "int main() { return (sizeof(int*) == 8 ? 1 : 0); } - ") - try_run(run_res compile_res - ${CMAKE_CURRENT_BINARY_DIR} "${CMAKE_CURRENT_BINARY_DIR}/is64.cpp" - RUN_OUTPUT_VARIABLE run_output) - - if(compile_res AND run_res) - set(${VAR} ${run_res}) - elseif(CMAKE_CL_64) - set(${VAR} 1) - elseif("$ENV{Platform}" STREQUAL "X64") # nmake 64 bit + if(CMAKE_SIZEOF_VOID_P EQUAL 8) set(${VAR} 1) + elseif(CMAKE_SIZEOF_VOID_P EQUAL 4) + set(${VAR} 0) + else() + message(WARNING "Unsupported pointer size. Please contact the authors.") endif() endmacro(ei_is_64bit_env) From c26e8a1af32a907eb6c78f0ae2165af5e1f79a76 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 15 Oct 2014 11:59:21 +0200 Subject: [PATCH 0897/1103] D&C SVD: fix deflation of repeated singular values, fix sorting of singular values, fix case of complete deflation --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 229 ++++++++++++++++++-------- 1 file changed, 162 insertions(+), 67 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index d5e8140a4..d8d75624d 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -23,6 +23,10 @@ // #define EIGEN_BDCSVD_SANITY_CHECKS namespace Eigen { +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE +IOFormat bdcsvdfmt(8, 0, ", ", "\n", " [", "]"); +#endif + template class BDCSVD; namespace internal { @@ -234,6 +238,9 @@ BDCSVD >& BDCSVD >:: template BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsigned int computationOptions) { +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "\n\n\n======================================================================================================================\n\n\n"; +#endif allocate(matrix.rows(), matrix.cols(), computationOptions); using std::abs; @@ -478,9 +485,23 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, m_computed.col(firstCol + shift).segment(firstCol + shift + 1, k) = alphaK * l.transpose().real(); m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) = betaK * f.transpose().real(); +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + ArrayXr tmp1 = (m_computed.block(firstCol+shift, firstCol+shift, n, n)).jacobiSvd().singularValues(); +#endif // Second part: try to deflate singular values in combined matrix deflation(firstCol, lastCol, k, firstRowW, firstColW, shift); - +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + ArrayXr tmp2 = (m_computed.block(firstCol+shift, firstCol+shift, n, n)).jacobiSvd().singularValues(); + std::cout << "\n\nj1 = " << tmp1.transpose().format(bdcsvdfmt) << "\n"; + std::cout << "j2 = " << tmp2.transpose().format(bdcsvdfmt) << "\n\n"; + std::cout << "err: " << ((tmp1-tmp2).abs()>1e-12*tmp2.abs()).transpose() << "\n"; + static int count = 0; + std::cout << "# " << ++count << "\n\n"; + assert((tmp1-tmp2).matrix().norm() < 1e-14*tmp2.matrix().norm()); +// assert(count<681); +// assert(((tmp1-tmp2).abs()<1e-13*tmp2.abs()).all()); +#endif + // Third part: compute SVD of combined matrix MatrixXr UofSVD, VofSVD; VectorType singVals; @@ -542,9 +563,20 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec computeSingVals(col0, diag, singVals, shifts, mus); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << " j: " << (m_computed.block(firstCol, firstCol, n, n)).jacobiSvd().singularValues().transpose().reverse() << "\n\n"; std::cout << " sing-val: " << singVals.transpose() << "\n"; std::cout << " mu: " << mus.transpose() << "\n"; std::cout << " shift: " << shifts.transpose() << "\n"; + + { + Index actual_n = n; + while(actual_n>1 && col0(actual_n-1)==0) --actual_n; + std::cout << "\n\n mus: " << mus.head(actual_n).transpose() << "\n\n"; + std::cout << " check1 (expect0) : " << ((singVals.array()-(shifts+mus)) / singVals.array()).head(actual_n).transpose() << "\n\n"; + std::cout << " check2 (>0) : " << ((singVals.array()-diag) / singVals.array()).head(actual_n).transpose() << "\n\n"; + std::cout << " check3 (>0) : " << ((diag.segment(1,actual_n-1)-singVals.head(actual_n-1).array()) / singVals.head(actual_n-1).array()).transpose() << "\n\n\n"; + std::cout << " check4 (>0) : " << ((singVals.segment(1,actual_n-1)-singVals.head(actual_n-1))).transpose() << "\n\n\n"; + } #endif #ifdef EIGEN_BDCSVD_SANITY_CHECKS @@ -557,16 +589,8 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec perturbCol0(col0, diag, singVals, shifts, mus, zhat); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE std::cout << " zhat: " << zhat.transpose() << "\n"; - { - Index actual_n = n; - while(actual_n>1 && col0(actual_n-1)==0) --actual_n; - std::cout << "\n\n mus: " << mus.head(actual_n).transpose() << "\n\n"; - std::cout << " check1: " << ((singVals.array()-(shifts+mus)) / singVals.array()).head(actual_n).transpose() << "\n\n"; - std::cout << " check2: " << ((singVals.array()-diag) / singVals.array()).head(actual_n).transpose() << "\n\n"; - std::cout << " check3: " << ((diag.segment(1,actual_n-1)-singVals.head(actual_n-1).array()) / singVals.head(actual_n-1).array()).transpose() << "\n\n\n"; - } #endif - + #ifdef EIGEN_BDCSVD_SANITY_CHECKS assert(zhat.allFinite()); #endif @@ -586,19 +610,40 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec assert(m_computed.allFinite()); #endif - // Reverse order so that singular values in increased order - // Because of deflation, the zeros singular-values are already at the end Index actual_n = n; while(actual_n>1 && singVals(actual_n-1)==0) --actual_n; + + // Because of deflation, the singular values might not be completely sorted. + // Fortunately, reordering them is a O(n) problem + for(Index i=0; isingVals(i+1)) + { + using std::swap; + swap(singVals(i),singVals(i+1)); + U.col(i).swap(U.col(i+1)); + if(m_compV) V.col(i).swap(V.col(i+1)); + } + } + + // Reverse order so that singular values in increased order + // Because of deflation, the zeros singular-values are already at the end singVals.head(actual_n).reverseInPlace(); U.leftCols(actual_n) = U.leftCols(actual_n).rowwise().reverse().eval(); // FIXME this requires a temporary if (m_compV) V.leftCols(actual_n) = V.leftCols(actual_n).rowwise().reverse().eval(); // FIXME this requires a temporary + +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + JacobiSVD jsvd(m_computed.block(firstCol, firstCol, n, n) ); + std::cout << " * j: " << jsvd.singularValues().transpose() << "\n\n"; + std::cout << " * sing-val: " << singVals.transpose() << "\n"; +// std::cout << " * err: " << ((jsvd.singularValues()-singVals)>1e-13*singVals.norm()).transpose() << "\n"; +#endif } template typename BDCSVD::RealScalar BDCSVD::secularEq(RealScalar mu, const ArrayXr& col0, const ArrayXr& diag, const ArrayXr& diagShifted, RealScalar shift, Index n) { - return 1 + (col0.square() / ((diagShifted - mu) )/( (diag + shift + mu))).head(n).sum(); + return 1 + (col0.square() / ((diagShifted - mu) * (diag + shift + mu))).head(n).sum(); } template @@ -622,16 +667,16 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // } // } // perm.conservativeResize(m); - + for (Index k = 0; k < n; ++k) { if (col0(k) == 0 || actual_n==1) { // if col0(k) == 0, then entry is deflated, so singular value is on diagonal // if actual_n==1, then the deflated problem is already diagonalized - singVals(k) = diag(k); + singVals(k) = k==0 ? col0(0) : diag(k); mus(k) = 0; - shifts(k) = diag(k); + shifts(k) = k==0 ? col0(0) : diag(k); continue; } @@ -646,12 +691,23 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia Index l = k+1; while(col0(l)==0) { ++l; eigen_internal_assert(l 0) ? left : right; // measure everything relative to shift @@ -682,20 +738,26 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // rational interpolation: fit a function of the form a / mu + b through the two previous // iterates and use its zero to compute the next iterate bool useBisection = fPrev*fCur>0; - while (abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (max)(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) + while (fCur!=0 && abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (max)(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) { ++m_numIters; + // Find a and b such that the function f(mu) = a / mu + b matches the current and previous samples. RealScalar a = (fCur - fPrev) / (1/muCur - 1/muPrev); RealScalar b = fCur - a / muCur; - + // And find mu such that f(mu)==0: + RealScalar muZero = -a/b; + RealScalar fZero = secularEq(muZero, col0, diag, diagShifted, shift, actual_n); + muPrev = muCur; fPrev = fCur; - muCur = -a / b; - fCur = secularEq(muCur, col0, diag, diagShifted, shift, actual_n); + muCur = muZero; + fCur = fZero; + if (shift == left && (muCur < 0 || muCur > right - left)) useBisection = true; if (shift == right && (muCur < -(right - left) || muCur > 0)) useBisection = true; + if (abs(fCur)>abs(fPrev)) useBisection = true; } // fall back on bisection method if rational interpolation did not work @@ -710,7 +772,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia leftShifted = RealScalar(1)/NumTraits::highest(); // I don't understand why the case k==0 would be special there: // if (k == 0) rightShifted = right - left; else - rightShifted = (right - left) * 0.6; // theoretically we can take 0.5, but let's be safe + rightShifted = (k==actual_n-1) ? right : ((right - left) * 0.6); // theoretically we can take 0.5, but let's be safe } else { @@ -722,7 +784,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia RealScalar fRight = secularEq(rightShifted, col0, diag, diagShifted, shift, actual_n); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE - if(fLeft * fRight>=0) + if(!(fLeft * fRight<0)) std::cout << k << " : " << fLeft << " * " << fRight << " == " << fLeft * fRight << " ; " << left << " - " << right << " -> " << leftShifted << " " << rightShifted << " shift=" << shift << "\n"; #endif eigen_internal_assert(fLeft * fRight < 0); @@ -805,7 +867,8 @@ void BDCSVD::perturbCol0 prod *= ((singVals(j)+dk) / ((diag(i)+dk))) * ((mus(j)+(shifts(j)-dk)) / ((diag(i)-dk))); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE if(i!=k && std::abs(((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) - 1) > 0.9 ) - std::cout << " " << ((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) << " == (" << (singVals(j)+dk) << " * " << (mus(j)+(shifts(j)-dk)) << ") / (" << (diag(i)+dk) << " * " << (diag(i)-dk) << ")\n"; + std::cout << " " << ((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) << " == (" << (singVals(j)+dk) << " * " << (mus(j)+(shifts(j)-dk)) + << ") / (" << (diag(i)+dk) << " * " << (diag(i)-dk) << ")\n"; #endif } } @@ -863,8 +926,6 @@ void BDCSVD::computeSingVecs if (m_compV) { V.col(k).setZero(); -// for(Index i=1;i::deflation43(Index firstCol, Index shift, Index i, Index // page 13 -// i,j >= 1, i != j and |di - dj| < epsilon * norm2(M) +// i,j >= 1, i!=j and |di - dj| < epsilon * norm2(M) // We apply two rotations to have zj = 0; // TODO deflation44 is still broken and not properly tested template @@ -943,15 +1004,10 @@ void BDCSVD::deflation44(Index firstColu , Index firstColm, Index fi m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j); m_computed(firstColm + j, firstColm) = 0; - JacobiRotation J(c,s); - if (m_compU) - { - m_naiveU.middleRows(firstColu, size).applyOnTheRight(firstColu + i, firstColu + j, J); - } - if (m_compV) - { - m_naiveU.middleRows(firstRowW, size-1).applyOnTheRight(firstColW + i, firstColW + j, J.transpose()); - } + JacobiRotation J(c,-s); + if (m_compU) m_naiveU.middleRows(firstColu, size+1).applyOnTheRight(firstColu + i, firstColu + j, J); + else m_naiveU.applyOnTheRight(firstColu+i, firstColu+j, J); + if (m_compV) m_naiveV.middleRows(firstRowW, size).applyOnTheRight(firstColW + i, firstColW + j, J); }// end deflation 44 @@ -964,43 +1020,49 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index using std::max; const Index length = lastCol + 1 - firstCol; + Block col0(m_computed, firstCol+shift, firstCol+shift, length, 1); + Diagonal fulldiag(m_computed); + VectorBlock,Dynamic> diag(fulldiag, firstCol+shift, length); + + RealScalar maxDiag = diag.tail((std::max)(Index(1),length-1)).cwiseAbs().maxCoeff(); + RealScalar epsilon_strict = NumTraits::epsilon() * maxDiag; + RealScalar epsilon_coarse = 8 * NumTraits::epsilon() * (max)(col0.cwiseAbs().maxCoeff(), maxDiag); + #ifdef EIGEN_BDCSVD_SANITY_CHECKS assert(m_naiveU.allFinite()); assert(m_naiveV.allFinite()); assert(m_computed.allFinite()); #endif - - Block col0(m_computed, firstCol+shift, firstCol+shift, length, 1); - Diagonal fulldiag(m_computed); - VectorBlock,Dynamic> diag(fulldiag, firstCol+shift, length); - - RealScalar epsilon = 8 * NumTraits::epsilon() * (max)(col0.cwiseAbs().maxCoeff(), diag.cwiseAbs().maxCoeff()); + +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "\ndeflate:" << diag.head(k+1).transpose() << " | " << diag.segment(k+1,length-k-1).transpose() << "\n"; +#endif //condition 4.1 - if (diag(0) < epsilon) + if (diag(0) < epsilon_coarse) { #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE - std::cout << "deflation 4.1, because " << diag(0) << " < " << epsilon << "\n"; + std::cout << "deflation 4.1, because " << diag(0) << " < " << epsilon_coarse << "\n"; #endif - diag(0) = epsilon; + diag(0) = epsilon_coarse; } //condition 4.2 for (Index i=1;i::deflation(Index firstCol, Index lastCol, Index k, Index assert(m_naiveV.allFinite()); assert(m_computed.allFinite()); #endif - +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "to be sorted: " << diag.transpose() << "\n\n"; +#endif { + // Check for total deflation + // If we have a total deflation, then we have to consider col0(0)==diag(0) as a singular value during sorting + bool total_deflation = (col0.tail(length-1).array()==RealScalar(0)).all(); + // Sort the diagonal entries, since diag(1:k-1) and diag(k:length) are already sorted, let's do a sorted merge. // First, compute the respective permutation. Index *permutation = new Index[length]; // FIXME avoid repeated dynamic memory allocation { permutation[0] = 0; Index p = 1; + + // Move deflated diagonal entries at the end. for(Index i=1; i::deflation(Index firstCol, Index lastCol, Index k, Index } } + // If we have a total deflation, then we have to insert diag(0) at the right place + if(total_deflation) + { + for(Index i=1; i::Map(permutation, length) << "\n\n"; + // Current index of each col, and current column of each index Index *realInd = new Index[length]; // FIXME avoid repeated dynamic memory allocation Index *realCol = new Index[length]; // FIXME avoid repeated dynamic memory allocation @@ -1042,15 +1129,15 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index realInd[pos] = pos; } - for(Index i = 1; i < length; i++) + for(Index i = total_deflation?0:1; i < length; i++) { - const Index pi = permutation[length - i]; + const Index pi = permutation[length - (total_deflation ? i+1 : i)]; const Index J = realCol[pi]; using std::swap; - // swap diaognal and first column entries: + // swap diagonal and first column entries: swap(diag(i), diag(J)); - swap(col0(i), col0(J)); + if(i!=0 && J!=0) swap(col0(i), col0(J)); // change columns if (m_compU) m_naiveU.col(firstCol+i).segment(firstCol, length + 1).swap(m_naiveU.col(firstCol+J).segment(firstCol, length + 1)); @@ -1068,23 +1155,31 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index delete[] realInd; delete[] realCol; } +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "sorted: " << diag.transpose().format(bdcsvdfmt) << "\n"; + std::cout << " : " << col0.transpose() << "\n\n"; +#endif + + //condition 4.4 + { + Index i = length-1; + while(i>0 && (diag(i)==0 || col0(i)==0)) --i; + for(; i>1;--i) + if( (diag(i) - diag(i-1)) < NumTraits::epsilon()*diag(i) ) + { +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + std::cout << "deflation 4.4 with i = " << i << " because " << (diag(i) - diag(i-1)) << " < " << NumTraits::epsilon()*diag(i) << "\n"; +#endif + eigen_internal_assert(abs(diag(i) - diag(i-1))::epsilon()*diag(i+1)) -// if ((diag(i+1) - diag(i)) < epsilon) - { -#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE - std::cout << "deflation 4.4 with i = " << i << " because " << (diag(i+1) - diag(i)) << " < " << epsilon << "\n"; -#endif - deflation44(firstCol, firstCol + shift, firstRowW, firstColW, i, i + 1, length); - } - #ifdef EIGEN_BDCSVD_SANITY_CHECKS assert(m_naiveU.allFinite()); assert(m_naiveV.allFinite()); From 2cc41dbe8355eee2e646c5bffa2d4b6cfdad4029 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 15 Oct 2014 15:21:12 +0200 Subject: [PATCH 0898/1103] D&C SVD: fix some numerical issues by truly skipping deflated singular values when computing them --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 143 +++++++++++--------------- 1 file changed, 59 insertions(+), 84 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index d8d75624d..e8e795589 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -84,6 +84,7 @@ public: typedef Matrix MatrixXr; typedef Matrix VectorType; typedef Array ArrayXr; + typedef Array ArrayXi; /** \brief Default Constructor. * @@ -159,19 +160,16 @@ private: void allocate(Index rows, Index cols, unsigned int computationOptions); void divide(Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift); void computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V); - void computeSingVals(const ArrayXr& col0, const ArrayXr& diag, VectorType& singVals, - ArrayXr& shifts, ArrayXr& mus); - void perturbCol0(const ArrayXr& col0, const ArrayXr& diag, const VectorType& singVals, - const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat); - void computeSingVecs(const ArrayXr& zhat, const ArrayXr& diag, const VectorType& singVals, - const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V); + void computeSingVals(const ArrayXr& col0, const ArrayXr& diag, const ArrayXi& perm, VectorType& singVals, ArrayXr& shifts, ArrayXr& mus); + void perturbCol0(const ArrayXr& col0, const ArrayXr& diag, const ArrayXi& perm, const VectorType& singVals, const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat); + void computeSingVecs(const ArrayXr& zhat, const ArrayXr& diag, const ArrayXi& perm, const VectorType& singVals, const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V); void deflation43(Index firstCol, Index shift, Index i, Index size); void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size); void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift); template void copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naivev); static void structured_update(Block A, const MatrixXr &B, Index n1); - static RealScalar secularEq(RealScalar x, const ArrayXr& col0, const ArrayXr& diag, const ArrayXr& diagShifted, RealScalar shift, Index n); + static RealScalar secularEq(RealScalar x, const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm, const ArrayXr& diagShifted, RealScalar shift); protected: MatrixXr m_naiveU, m_naiveV; @@ -543,13 +541,24 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec ArrayXr diag = m_computed.block(firstCol, firstCol, n, n).diagonal(); diag(0) = 0; - // compute singular values and vectors (in decreasing order) + // Allocate space for singular values and vectors singVals.resize(n); U.resize(n+1, n+1); if (m_compV) V.resize(n, n); if (col0.hasNaN() || diag.hasNaN()) { std::cout << "\n\nHAS NAN\n\n"; return; } - + + // Many singular values might have been deflated, the zero ones have been moved to the end, + // but others are interleaved and we must ignore them at this stage. + // To this end, let's compute a permutation skipping them: + Index actual_n = n; + while(actual_n>1 && diag(actual_n-1)==0) --actual_n; + Index m = 0; // size of the deflated problem + ArrayXi perm(actual_n); + for(Index k=0;k::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec #endif // Compute singVals, shifts, and mus - computeSingVals(col0, diag, singVals, shifts, mus); + computeSingVals(col0, diag, perm, singVals, shifts, mus); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE std::cout << " j: " << (m_computed.block(firstCol, firstCol, n, n)).jacobiSvd().singularValues().transpose().reverse() << "\n\n"; @@ -586,7 +595,7 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec #endif // Compute zhat - perturbCol0(col0, diag, singVals, shifts, mus, zhat); + perturbCol0(col0, diag, perm, singVals, shifts, mus, zhat); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE std::cout << " zhat: " << zhat.transpose() << "\n"; #endif @@ -595,7 +604,7 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec assert(zhat.allFinite()); #endif - computeSingVecs(zhat, diag, singVals, shifts, mus, U, V); + computeSingVecs(zhat, diag, perm, singVals, shifts, mus, U, V); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE std::cout << "U^T U: " << (U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() << "\n"; @@ -610,9 +619,6 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec assert(m_computed.allFinite()); #endif - Index actual_n = n; - while(actual_n>1 && singVals(actual_n-1)==0) --actual_n; - // Because of deflation, the singular values might not be completely sorted. // Fortunately, reordering them is a O(n) problem for(Index i=0; i::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec } template -typename BDCSVD::RealScalar BDCSVD::secularEq(RealScalar mu, const ArrayXr& col0, const ArrayXr& diag, const ArrayXr& diagShifted, RealScalar shift, Index n) +typename BDCSVD::RealScalar BDCSVD::secularEq(RealScalar mu, const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm, const ArrayXr& diagShifted, RealScalar shift) { - return 1 + (col0.square() / ((diagShifted - mu) * (diag + shift + mu))).head(n).sum(); + Index m = perm.size(); + RealScalar res = 1; + for(Index i=0; i -void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& diag, +void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm, VectorType& singVals, ArrayXr& shifts, ArrayXr& mus) { using std::abs; @@ -657,16 +670,6 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia Index n = col0.size(); Index actual_n = n; while(actual_n>1 && col0(actual_n-1)==0) --actual_n; -// Index m = 0; -// Array perm(actual_n); -// { -// for(Index k=0;k::computeSingVals(const ArrayXr& col0, const ArrayXr& dia Index l = k+1; while(col0(l)==0) { ++l; eigen_internal_assert(l 0) ? left : right; @@ -727,8 +733,8 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia muCur = -(right - left) * 0.5; } - RealScalar fPrev = secularEq(muPrev, col0, diag, diagShifted, shift, actual_n); - RealScalar fCur = secularEq(muCur, col0, diag, diagShifted, shift, actual_n); + RealScalar fPrev = secularEq(muPrev, col0, diag, perm, diagShifted, shift); + RealScalar fCur = secularEq(muCur, col0, diag, perm, diagShifted, shift); if (abs(fPrev) < abs(fCur)) { swap(fPrev, fCur); @@ -747,7 +753,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia RealScalar b = fCur - a / muCur; // And find mu such that f(mu)==0: RealScalar muZero = -a/b; - RealScalar fZero = secularEq(muZero, col0, diag, diagShifted, shift, actual_n); + RealScalar fZero = secularEq(muZero, col0, diag, perm, diagShifted, shift); muPrev = muCur; fPrev = fCur; @@ -780,8 +786,8 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia rightShifted = -RealScalar(1)/NumTraits::highest(); } - RealScalar fLeft = secularEq(leftShifted, col0, diag, diagShifted, shift, actual_n); - RealScalar fRight = secularEq(rightShifted, col0, diag, diagShifted, shift, actual_n); + RealScalar fLeft = secularEq(leftShifted, col0, diag, perm, diagShifted, shift); + RealScalar fRight = secularEq(rightShifted, col0, diag, perm, diagShifted, shift); #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE if(!(fLeft * fRight<0)) @@ -792,7 +798,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (max)(abs(leftShifted), abs(rightShifted))) { RealScalar midShifted = (leftShifted + rightShifted) / 2; - RealScalar fMid = secularEq(midShifted, col0, diag, diagShifted, shift, actual_n); + RealScalar fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift); if (fLeft * fMid < 0) { rightShifted = midShifted; @@ -824,28 +830,13 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // zhat is perturbation of col0 for which singular vectors can be computed stably (see Section 3.1) template void BDCSVD::perturbCol0 - (const ArrayXr& col0, const ArrayXr& diag, const VectorType& singVals, + (const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm, const VectorType& singVals, const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat) { using std::sqrt; Index n = col0.size(); - - // Ignore trailing zeros: - Index actual_n = n; - while(actual_n>1 && col0(actual_n-1)==0) --actual_n; - // Deflated non-zero singular values are kept in-place, - // we thus compute an indirection array to properly ignore all deflated entries. - // TODO compute it once! - Index m = 0; // size of the deflated problem - Array perm(actual_n); - { - for(Index k=0;k::perturbCol0 { // see equation (3.6) RealScalar dk = diag(k); - RealScalar prod = (singVals(actual_n-1) + dk) * (mus(actual_n-1) + (shifts(actual_n-1) - dk)); + RealScalar prod = (singVals(last) + dk) * (mus(last) + (shifts(last) - dk)); for(Index l = 0; l::perturbCol0 } } #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE - std::cout << "zhat(" << k << ") = sqrt( " << prod << ") ; " << (singVals(actual_n-1) + dk) << " * " << mus(actual_n-1) + shifts(actual_n-1) << " - " << dk << "\n"; + std::cout << "zhat(" << k << ") = sqrt( " << prod << ") ; " << (singVals(last) + dk) << " * " << mus(last) + shifts(last) << " - " << dk << "\n"; #endif RealScalar tmp = sqrt(prod); zhat(k) = col0(k) > 0 ? tmp : -tmp; @@ -884,26 +875,11 @@ void BDCSVD::perturbCol0 // compute singular vectors template void BDCSVD::computeSingVecs - (const ArrayXr& zhat, const ArrayXr& diag, const VectorType& singVals, + (const ArrayXr& zhat, const ArrayXr& diag, const ArrayXi &perm, const VectorType& singVals, const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V) { Index n = zhat.size(); - - // Deflated non-zero singular values are kept in-place, - // we thus compute an indirection array to properly ignore all deflated entries. - // TODO compute it once! - Index actual_n = n; - while(actual_n>1 && zhat(actual_n-1)==0) --actual_n; - Index m = 0; - Array perm(actual_n); - { - for(Index k=0;k::deflation44(Index firstColu , Index firstColm, Index fi c/=r; s/=r; m_computed(firstColm + i, firstColm) = r; - m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j); + m_computed(firstColm + j, firstColm + j) = m_computed(firstColm + i, firstColm + i); m_computed(firstColm + j, firstColm) = 0; JacobiRotation J(c,-s); @@ -1117,7 +1093,6 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index } } } -// std::cout << "perm: " << Matrix::Map(permutation, length) << "\n\n"; // Current index of each col, and current column of each index Index *realInd = new Index[length]; // FIXME avoid repeated dynamic memory allocation @@ -1165,7 +1140,7 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index Index i = length-1; while(i>0 && (diag(i)==0 || col0(i)==0)) --i; for(; i>1;--i) - if( (diag(i) - diag(i-1)) < NumTraits::epsilon()*diag(i) ) + if( (diag(i) - diag(i-1)) < NumTraits::epsilon()*maxDiag ) { #ifdef EIGEN_BDCSVD_DEBUG_VERBOSE std::cout << "deflation 4.4 with i = " << i << " because " << (diag(i) - diag(i-1)) << " < " << NumTraits::epsilon()*diag(i) << "\n"; From c8060094535882e4fc46e5d54a13358d6c23b7a9 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 15 Oct 2014 16:32:16 +0200 Subject: [PATCH 0899/1103] Extend svd unit tests to stress problems with duplicated singular values. --- test/svd_common.h | 71 +++++++++++++++++++++++++++++++++++------------ 1 file changed, 54 insertions(+), 17 deletions(-) diff --git a/test/svd_common.h b/test/svd_common.h index e902d2320..347ea8046 100644 --- a/test/svd_common.h +++ b/test/svd_common.h @@ -38,7 +38,6 @@ void svd_check_full(const MatrixType& m, const SvdType& svd) sigma.diagonal() = svd.singularValues().template cast(); MatrixUType u = svd.matrixU(); MatrixVType v = svd.matrixV(); - VERIFY_IS_APPROX(m, u * sigma * v.adjoint()); VERIFY_IS_UNITARY(u); VERIFY_IS_UNITARY(v); @@ -90,31 +89,31 @@ void svd_least_square(const MatrixType& m, unsigned int computationOptions) SolutionType x = svd.solve(rhs); + // evaluate normal equation which works also for least-squares solutions + if(internal::is_same::value || svd.rank()==m.diagonal().size()) + { + // This test is not stable with single precision. + // This is probably because squaring m signicantly affects the precision. + VERIFY_IS_APPROX(m.adjoint()*(m*x),m.adjoint()*rhs); + } + RealScalar residual = (m*x-rhs).norm(); // Check that there is no significantly better solution in the neighborhood of x if(!test_isMuchSmallerThan(residual,rhs.norm())) { - // If the residual is very small, then we have an exact solution, so we are already good. - for(int k=0;k::epsilon(); + y.row(k) = (1.+2*NumTraits::epsilon())*x.row(k); RealScalar residual_y = (m*y-rhs).norm(); VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); - y.row(k) = x.row(k).array() - 2*NumTraits::epsilon(); + y.row(k) = (1.-2*NumTraits::epsilon())*x.row(k); residual_y = (m*y-rhs).norm(); VERIFY( test_isApprox(residual_y,residual) || residual < residual_y ); } } - - // evaluate normal equation which works also for least-squares solutions - if(internal::is_same::value) - { - // This test is not stable with single precision. - // This is probably because squaring m signicantly affects the precision. - VERIFY_IS_APPROX(m.adjoint()*m*x,m.adjoint()*rhs); - } } // check minimal norm solutions, the inoput matrix m is only used to recover problem size @@ -234,11 +233,49 @@ void svd_fill_random(MatrixType &m) Matrix d = Matrix::Random(diagSize); for(Index k=0; k(-s,s)); - m = Matrix::Random(m.rows(),diagSize) * d.asDiagonal() * Matrix::Random(diagSize,m.cols()); + + bool dup = internal::random(0,10) < 3; + bool unit_uv = internal::random(0,10) < (dup?7:3); // if we duplicate some diagonal entries, then increase the chance to preserve them using unitary U and V factors + + // duplicate some singular values + if(dup) + { + Index n = internal::random(0,d.size()-1); + for(Index i=0; i(0,d.size()-1)) = d(internal::random(0,d.size()-1)); + } + + Matrix U(m.rows(),diagSize); + Matrix VT(diagSize,m.cols()); + if(unit_uv) + { + // in very rare cases let's try with a pure diagonal matrix + if(internal::random(0,10) < 1) + { + U.setIdentity(); + VT.setIdentity(); + } + else + { + createRandomPIMatrixOfRank(diagSize,U.rows(), U.cols(), U); + createRandomPIMatrixOfRank(diagSize,VT.rows(), VT.cols(), VT); + } + } + else + { + U.setRandom(); + VT.setRandom(); + } + + m = U * d.asDiagonal() * VT; + // cancel some coeffs - Index n = internal::random(0,m.size()-1); - for(Index i=0; i(0,m.rows()-1), internal::random(0,m.cols()-1)) = Scalar(0); + if(!(dup && unit_uv)) + { + Index n = internal::random(0,m.size()-1); + for(Index i=0; i(0,m.rows()-1), internal::random(0,m.cols()-1)) = Scalar(0); + } } From dba55041ab62961e549ea58778dffa3eaa0cbdb5 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 15 Oct 2014 11:20:36 -0700 Subject: [PATCH 0900/1103] Added support for promises Started to improve multithreaded contractions --- unsupported/Eigen/CXX11/Tensor | 1 + .../src/Tensor/TensorContractionThreadPool.h | 351 ++++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 11 + 3 files changed, 363 insertions(+) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 2137f4276..7ec60044e 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -55,6 +55,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" +//#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h new file mode 100644 index 000000000..dc0513305 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h @@ -0,0 +1,351 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_THREAD_POOL_H +#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_THREAD_POOL_H + +// evaluator for thread pool device +#ifdef EIGEN_USE_THREADS + +namespace Eigen { +namespace internal { + +template +struct packLhsArg { + LhsScalar* blockA; + const LhsMapper& lhs; + const Index m_start; + const Index k_start; + const Index mc; + const Index kc; +}; + +template +struct packRhsAndKernelArg { + const std::vector* blockAs; + RhsScalar* blockB; + const RhsMapper& rhs; + OutputMapper& output; + const Index m; + const Index k; + const Index n; + const Index mc; + const Index kc; + const Index nc; + const Index num_threads; + const Index num_blockAs; + const Index max_m; + const Index k_block_idx; + const Index m_block_idx; + const Index n_block_idx; + const Index m_blocks; + const Index n_blocks; + std::vector* kernel_promises; + const std::vector* lhs_futures; + const bool need_to_pack; +}; + +} // end namespace internal + + +template +struct TensorEvaluator, ThreadPoolDevice> : + public TensorContractionEvaluatorBase, ThreadPoolDevice> > { + + typedef ThreadPoolDevice Device; + + typedef TensorEvaluator, Device> Self; + typedef TensorContractionEvaluatorBase Base; + + typedef TensorContractionOp XprType; + typedef typename internal::remove_const::type Scalar; + typedef typename XprType::Packet Packet; + typedef typename XprType::Index Index; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + typedef array::Dimensions::count> left_dim_mapper_t; + typedef array::Dimensions::count> right_dim_mapper_t; + + typedef array::value> contract_t; + typedef array::Dimensions::count - internal::array_size::value>::size> left_nocontract_t; + typedef array::Dimensions::count - internal::array_size::value>::size> right_nocontract_t; + + static const int NumDims = max_n_1::Dimensions::count + TensorEvaluator::Dimensions::count - 2 * internal::array_size::value>::size; + + typedef DSizes Dimensions; + + // typedefs needed in evalTo + typedef typename internal::remove_const::type LhsScalar; + typedef typename internal::remove_const::type RhsScalar; + typedef typename internal::gebp_traits Traits; + + typedef TensorEvaluator LeftEvaluator; + typedef TensorEvaluator RightEvaluator; + + TensorEvaluator(const XprType& op, const Device& device) : + Base(op, device) {} + + template + void evalTyped(Scalar* buffer) const { + // columns in left side, rows in right side + const Index k = this->m_k_size; + + // rows in left side + const Index m = this->m_i_size; + + // columns in right side + const Index n = this->m_j_size; + + // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar) + this->m_device.memset(buffer, 0, m * n * sizeof(Scalar)); + + + const int lhs_packet_size = internal::packet_traits::size; + const int rhs_packet_size = internal::packet_traits::size; + + typedef internal::TensorContractionInputMapper LhsMapper; + + typedef internal::TensorContractionInputMapper RhsMapper; + + typedef internal::blas_data_mapper OutputMapper; + + // TODO: packing could be faster sometimes if we supported row major tensor mappers + typedef internal::gemm_pack_lhs LhsPacker; + typedef internal::gemm_pack_rhs RhsPacker; + + // TODO: replace false, false with conjugate values? + typedef internal::gebp_kernel GebpKernel; + + typedef internal::packLhsArg packLArg; + typedef internal::packRhsAndKernelArg packRKArg; + + // initialize data mappers + LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides, + this->m_left_contracting_strides, this->m_k_strides); + + RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides, + this->m_right_contracting_strides, this->m_k_strides); + + OutputMapper output(buffer, m); + + LhsPacker pack_lhs; + + // compute block sizes (which depend on number of threads) + const Index num_threads = this->m_device.numThreads(); + Index mc = m; + Index nc = n; + Index kc = k; + internal::computeProductBlockingSizes(kc, mc, nc/*, num_threads*/); + eigen_assert(mc <= m); + eigen_assert(nc <= n); + eigen_assert(kc <= k); + +#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b)) + const Index k_blocks = CEIL_DIV(k, kc); + const Index n_blocks = CEIL_DIV(n, nc); + const Index m_blocks = CEIL_DIV(m, mc); + const int sizeA = mc * kc; + const int sizeB = kc * nc; + + /* cout << "m: " << m << " n: " << n << " k: " << k << endl; + cout << "mc: " << mc << " nc: " << nc << " kc: " << kc << endl; + cout << "m_blocks: " << m_blocks << " n_blocks: " << n_blocks << " k_blocks: " << k_blocks << endl; + cout << "num threads: " << num_threads << endl; + */ + + // note: m_device.allocate should return 16 byte aligned pointers, but if blockA and blockB + // aren't 16 byte aligned segfaults will happen due to SIMD instructions + // note: You can get away with allocating just a single blockA and offsets and meet the + // the alignment requirements with the assumption that + // (Traits::mr * sizeof(ResScalar)) % 16 == 0 + const Index numBlockAs = (std::min)(num_threads, m_blocks); + std::vector blockAs; + blockAs.reserve(num_threads); + for (int i = 0; i < num_threads; i++) { + blockAs.push_back(static_cast(this->m_device.allocate(sizeA * sizeof(LhsScalar)))); + } + + // To circumvent alignment issues, I'm just going to separately allocate the memory for each thread + // TODO: is this too much memory to allocate? This simplifies coding a lot, but is wasteful. + // Other options: (1) reuse memory when a thread finishes. con: tricky + // (2) allocate block B memory in each thread. con: overhead + std::vector blockBs; + blockBs.reserve(n_blocks); + for (int i = 0; i < n_blocks; i++) { + blockBs.push_back(static_cast(this->m_device.allocate(sizeB * sizeof(RhsScalar)))); + } + + // lhs_futures starts with all null futures + std::vector lhs_futures(num_threads); + + // this should really be numBlockAs * n_blocks; + const Index num_kernel_promises = num_threads * n_blocks; + Promise p; + p.set_value(); + std::vector kernel_promises(num_kernel_promises, p); + + for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) { + const Index k_start = k_block_idx * kc; + // make sure we don't overshoot right edge of left matrix + const Index actual_kc = (std::min)(k_start + kc, k) - k_start; + + for (Index m_block_idx = 0; m_block_idx < m_blocks; m_block_idx += numBlockAs) { + const int num_blocks = (std::min)(m_blocks-m_block_idx, numBlockAs); + + for (Index mt_block_idx = m_block_idx; mt_block_idx < m_block_idx+num_blocks; mt_block_idx++) { + const Index m_start = mt_block_idx * mc; + const Index actual_mc = (std::min)(m_start + mc, m) - m_start; + eigen_assert(actual_mc > 0); + + int blockAId = (k_block_idx * m_blocks + mt_block_idx) % num_threads; + for (int i = 0; i < n_blocks; ++i) { + int future_id = (blockAId * n_blocks + i); + wait_until_ready(&kernel_promises[future_id]); + kernel_promises[future_id] = Promise(); + } + const packLArg arg = { + blockAs[blockAId], // blockA + lhs, // lhs + m_start, // m + k_start, // k + actual_mc, // mc + actual_kc, // kc + }; + + lhs_futures[blockAId] = + this->m_device.enqueue(&Self::packLhs, arg); + } + + // now start kernels. + const Index m_base_start = m_block_idx * mc; + const bool need_to_pack = m_block_idx == 0; + + for (Index n_block_idx = 0; n_block_idx < n_blocks; n_block_idx++) { + const Index n_start = n_block_idx * nc; + const Index actual_nc = (std::min)(n_start + nc, n) - n_start; + + // first make sure the previous kernels are all done before overwriting rhs. Also wait if + // we're going to start new k. In both cases need_to_pack is true. + if (need_to_pack) { + for (int i = num_blocks; i < num_threads; ++i) { + int blockAId = (k_block_idx * m_blocks + i + m_block_idx) % num_threads; + int future_id = (blockAId * n_blocks + n_block_idx); + wait_until_ready(&kernel_promises[future_id]); + } + } + + packRKArg arg = { + &blockAs, // blockA + blockBs[n_block_idx], // blockB + rhs, // rhs + output, // output + m_base_start, // m + k_start, // k + n_start, // n + mc, // mc + actual_kc, // kc + actual_nc, // nc + num_threads, + numBlockAs, + m, + k_block_idx, + m_block_idx, + n_block_idx, // n_block_idx + m_blocks, // m_blocks + n_blocks, // n_blocks + &kernel_promises, // kernel_promises + &lhs_futures, // lhs_futures + need_to_pack, // need_to_pack + }; + + typedef decltype(Self::packRhsAndKernel) Func; + this->m_device.enqueueNoFuture(&Self::packRhsAndKernel, arg); + } + } + } + + // collect the last frame of kernel futures + for (int i = 0; i < kernel_promises.size(); ++i) { + wait_until_ready(&kernel_promises[i]); + } + + // deallocate all of the memory for both A and B's + for (int i = 0; i < blockAs.size(); i++) { + this->m_device.deallocate(blockAs[i]); + } + for (int i = 0; i < blockBs.size(); i++) { + this->m_device.deallocate(blockBs[i]); + } + +#undef CEIL_DIV + } + + /* + * Packs a LHS block of size (mt, kc) starting at lhs(m, k). Before packing + * the LHS block, check that all of the kernels that worked on the same + * mt_block_idx in the previous m_block are done. + */ + template + static void packLhs(const packLArg arg) { + // perform actual packing + LhsPacker pack_lhs; + pack_lhs(arg.blockA, arg.lhs.getSubMapper(arg.m_start, arg.k_start), arg.kc, arg.mc); + } + + /* + * Packs a RHS block of size (kc, nc) starting at (k, n) after checking that + * all kernels in the previous block are done. + * Then for each LHS future, we wait on the future and then call GEBP + * on the area packed by the future (which starts at + * blockA + future_idx * mt * kc) on the LHS and with the full packed + * RHS block. + * The output of this GEBP is written to output(m + i * mt, n). + */ + template + static void packRhsAndKernel(packRKArg arg) { + if (arg.need_to_pack) { + RhsPacker pack_rhs; + pack_rhs(arg.blockB, arg.rhs.getSubMapper(arg.k, arg.n), arg.kc, arg.nc); + } + + GebpKernel gebp; + for (Index mt_block_idx = 0; mt_block_idx < arg.num_blockAs; mt_block_idx++) { + const Index m_base_start = arg.m + arg.mc*mt_block_idx; + if (m_base_start < arg.max_m) { + int blockAId = (arg.k_block_idx * arg.m_blocks + mt_block_idx + arg.m_block_idx) % arg.num_threads; + + wait_until_ready(&(*arg.lhs_futures)[blockAId]); + const Index actual_mc = (std::min)(m_base_start + arg.mc, arg.max_m) - m_base_start; + gebp(arg.output.getSubMapper(m_base_start, arg.n), + (*arg.blockAs)[blockAId], arg.blockB, + actual_mc, arg.kc, arg.nc, 1.0, -1, -1, 0, 0); + + const Index set_idx = blockAId * arg.n_blocks + arg.n_block_idx; + eigen_assert(!(*arg.kernel_promises)[set_idx].ready()); + (*arg.kernel_promises)[set_idx].set_value(); + } + } + } +}; + +} // end namespace Eigen + +#endif // EIGEN_USE_THREADS +#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_THREAD_POOL_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h index 5a6ff70e9..3748879cc 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h @@ -39,6 +39,17 @@ struct DefaultDevice { #ifdef EIGEN_USE_THREADS typedef std::future Future; +typedef std::promise Promise; + +static EIGEN_STRONG_INLINE void wait_until_ready(const Future* f) { + f->wait(); + // eigen_assert(f->ready()); +} + +static EIGEN_STRONG_INLINE void wait_until_ready(Promise* p) { + p->get_future().wait(); + // eigen_assert(p->get_future().ready()); +} struct ThreadPoolDevice { ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : num_threads_(num_cores) { } From c566cfe2ba0aad4ef054a55b402209980a90d994 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 15 Oct 2014 23:37:47 +0200 Subject: [PATCH 0901/1103] Make SVD unit test even more tough --- test/svd_common.h | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/test/svd_common.h b/test/svd_common.h index 347ea8046..4c172cf9d 100644 --- a/test/svd_common.h +++ b/test/svd_common.h @@ -269,12 +269,14 @@ void svd_fill_random(MatrixType &m) m = U * d.asDiagonal() * VT; - // cancel some coeffs + // (partly) cancel some coeffs if(!(dup && unit_uv)) { - Index n = internal::random(0,m.size()-1); + Matrix samples(7); + samples << 0, 5.60844e-313, -5.60844e-313, 4.94e-324, -4.94e-324, -1./NumTraits::highest(), 1./NumTraits::highest(); + Index n = internal::random(0,m.size()-1); for(Index i=0; i(0,m.rows()-1), internal::random(0,m.cols()-1)) = Scalar(0); + m(internal::random(0,m.rows()-1), internal::random(0,m.cols()-1)) = samples(internal::random(0,6)); } } From bfdd9f3ac95d9a2b41e6f2ec1f7434331125b9e1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 15 Oct 2014 15:32:59 -0700 Subject: [PATCH 0902/1103] Made the blocking computation aware of the l3 cache Also optimized the blocking parameters to take into account the number of threads used for a computation --- Eigen/src/Core/SolveTriangular.h | 2 +- .../Core/products/GeneralBlockPanelKernel.h | 122 ++++++++++++------ Eigen/src/Core/products/GeneralMatrixMatrix.h | 18 +-- .../products/GeneralMatrixMatrixTriangular.h | 2 +- Eigen/src/Core/products/Parallelizer.h | 4 +- .../Core/products/SelfadjointMatrixMatrix.h | 6 +- .../Core/products/TriangularMatrixMatrix.h | 2 +- .../Core/products/TriangularSolverMatrix.h | 4 +- blas/level3_impl.h | 12 +- test/product_large.cpp | 7 +- unsupported/Eigen/CXX11/Tensor | 2 +- .../CXX11/src/Tensor/TensorContraction.h | 2 +- .../src/Tensor/TensorContractionThreadPool.h | 13 +- 13 files changed, 117 insertions(+), 79 deletions(-) diff --git a/Eigen/src/Core/SolveTriangular.h b/Eigen/src/Core/SolveTriangular.h index ef17f288e..e158e3162 100644 --- a/Eigen/src/Core/SolveTriangular.h +++ b/Eigen/src/Core/SolveTriangular.h @@ -96,7 +96,7 @@ struct triangular_solver_selector typedef internal::gemm_blocking_space<(Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, Rhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxRowsAtCompileTime,4> BlockingType; - BlockingType blocking(rhs.rows(), rhs.cols(), size); + BlockingType blocking(rhs.rows(), rhs.cols(), size, 1, false); triangular_solve_matrix diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 090c8f4e6..b91786037 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -26,28 +26,37 @@ inline std::ptrdiff_t manage_caching_sizes_helper(std::ptrdiff_t a, std::ptrdiff } /** \internal */ -inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1=0, std::ptrdiff_t* l2=0) +inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff_t* l2, std::ptrdiff_t* l3) { - static std::ptrdiff_t m_l1CacheSize = 0; - static std::ptrdiff_t m_l2CacheSize = 0; - if(m_l2CacheSize==0) + static bool m_cache_sizes_initialized = false; + static std::ptrdiff_t m_l1CacheSize = 32*1024; + static std::ptrdiff_t m_l2CacheSize = 256*1024; + static std::ptrdiff_t m_l3CacheSize = 2*1024*1024; + + if(!m_cache_sizes_initialized) { - m_l1CacheSize = manage_caching_sizes_helper(queryL1CacheSize(),8 * 1024); - m_l2CacheSize = manage_caching_sizes_helper(queryTopLevelCacheSize(),1*1024*1024); + int l1CacheSize, l2CacheSize, l3CacheSize; + queryCacheSizes(l1CacheSize, l2CacheSize, l3CacheSize); + m_l1CacheSize = manage_caching_sizes_helper(l1CacheSize, 8*1024); + m_l2CacheSize = manage_caching_sizes_helper(l2CacheSize, 256*1024); + m_l3CacheSize = manage_caching_sizes_helper(l3CacheSize, 8*1024*1024); + m_cache_sizes_initialized = true; } - + if(action==SetAction) { // set the cpu cache size and cache all block sizes from a global cache size in byte eigen_internal_assert(l1!=0 && l2!=0); m_l1CacheSize = *l1; m_l2CacheSize = *l2; + m_l3CacheSize = *l3; } else if(action==GetAction) { eigen_internal_assert(l1!=0 && l2!=0); *l1 = m_l1CacheSize; *l2 = m_l2CacheSize; + *l3 = m_l3CacheSize; } else { @@ -70,10 +79,11 @@ inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1=0, std::ptrdi * - the number of scalars that fit into a packet (when vectorization is enabled). * * \sa setCpuCacheSizes */ +#define CEIL(a, b) ((a)+(b)-1)/(b) + template -void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) +void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n, int num_threads) { - EIGEN_UNUSED_VARIABLE(n); // Explanations: // Let's recall the product algorithms form kc x nc horizontal panels B' on the rhs and // mc x kc blocks A' on the lhs. A' has to fit into L2 cache. Moreover, B' is processed @@ -81,43 +91,71 @@ void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) // at the register level. For vectorization purpose, these small vertical panels are unpacked, // e.g., each coefficient is replicated to fit a packet. This small vertical panel has to // stay in L1 cache. - std::ptrdiff_t l1, l2; + std::ptrdiff_t l1, l2, l3; + manage_caching_sizes(GetAction, &l1, &l2, &l3); - typedef gebp_traits Traits; - enum { - kdiv = KcFactor * 2 * Traits::nr - * Traits::RhsProgress * sizeof(RhsScalar), - mr = gebp_traits::mr, - mr_mask = (0xffffffff/mr)*mr - }; + if (num_threads > 1) { + typedef gebp_traits Traits; + typedef typename Traits::ResScalar ResScalar; + enum { + kdiv = KcFactor * (Traits::mr * sizeof(LhsScalar) + Traits::nr * sizeof(RhsScalar)), + ksub = Traits::mr * Traits::nr * sizeof(ResScalar), + k_mask = (0xffffffff/8)*8, - manage_caching_sizes(GetAction, &l1, &l2); + mr = Traits::mr, + mr_mask = (0xffffffff/mr)*mr, -// k = std::min(k, l1/kdiv); -// SizeType _m = k>0 ? l2/(4 * sizeof(LhsScalar) * k) : 0; -// if(_m 0); + } + + SizeType n_cache = (l2-l1) / (nr * sizeof(RhsScalar) * k); + SizeType n_per_thread = CEIL(n, num_threads); + if (n_cache <= n_per_thread) { + // Don't exceed the capacity of the l2 cache. + eigen_assert(n_cache >= static_cast(nr)); + n = n_cache & nr_mask; + eigen_assert(n > 0); + } else { + n = (std::min)(n, (n_per_thread + nr - 1) & nr_mask); + } + + if (l3 > l2) { + // l3 is shared between all cores, so we'll give each thread its own chunk of l3. + SizeType m_cache = (l3-l2) / (sizeof(LhsScalar) * k * num_threads); + SizeType m_per_thread = CEIL(m, num_threads); + if(m_cache < m_per_thread && m_cache >= static_cast(mr)) { + m = m_cache & mr_mask; + eigen_assert(m > 0); + } else { + m = (std::min)(m, (m_per_thread + mr - 1) & mr_mask); + } + } + } + else { + // In unit tests we do not want to use extra large matrices, + // so we reduce the block size to check the blocking strategy is not flawed #ifndef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS -// k = std::min(k,240); -// n = std::min(n,3840/sizeof(RhsScalar)); -// m = std::min(m,3840/sizeof(RhsScalar)); - - k = std::min(k,sizeof(LhsScalar)<=4 ? 360 : 240); - n = std::min(n,3840/sizeof(RhsScalar)); - m = std::min(m,3840/sizeof(RhsScalar)); + k = std::min(k,sizeof(LhsScalar)<=4 ? 360 : 240); + n = std::min(n,3840/sizeof(RhsScalar)); + m = std::min(m,3840/sizeof(RhsScalar)); #else - k = std::min(k,24); - n = std::min(n,384/sizeof(RhsScalar)); - m = std::min(m,384/sizeof(RhsScalar)); + k = std::min(k,24); + n = std::min(n,384/sizeof(RhsScalar)); + m = std::min(m,384/sizeof(RhsScalar)); #endif + } } template -inline void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) +inline void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n, int num_threads) { - computeProductBlockingSizes(k, m, n); + computeProductBlockingSizes(k, m, n, num_threads); } #ifdef EIGEN_HAS_FUSE_CJMADD @@ -1846,8 +1884,8 @@ EIGEN_DONT_INLINE void gemm_pack_rhsm_mc = ActualRows; this->m_nc = ActualCols; @@ -331,21 +331,21 @@ class gemm_blocking_spacem_mc = Transpose ? cols : rows; this->m_nc = Transpose ? rows : cols; this->m_kc = depth; - if(full_rows) + if(l3_blocking) + { + computeProductBlockingSizes(this->m_kc, this->m_mc, this->m_nc, num_threads); + } + else // no l3 blocking { DenseIndex m = this->m_mc; - computeProductBlockingSizes(this->m_kc, m, this->m_nc); - } - else // full columns - { DenseIndex n = this->m_nc; - computeProductBlockingSizes(this->m_kc, this->m_mc, n); + computeProductBlockingSizes(this->m_kc, m, n, num_threads); } m_sizeA = this->m_mc * this->m_kc; @@ -451,7 +451,7 @@ class GeneralProduct (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>, _ActualLhsType, _ActualRhsType, Dest, BlockingType> GemmFunctor; - BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), true); + BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1, true); internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), this->rows(), this->cols(), Dest::Flags&RowMajorBit); } diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index daa8a1d8a..8de39f76f 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -72,7 +72,7 @@ struct general_matrix_matrix_triangular_product(kc, mc, nc); + computeProductBlockingSizes(kc, mc, nc, 1); // !!! mc must be a multiple of nr: if(mc > Traits::nr) mc = (mc/Traits::nr)*Traits::nr; diff --git a/Eigen/src/Core/products/Parallelizer.h b/Eigen/src/Core/products/Parallelizer.h index 4079063eb..837e69415 100644 --- a/Eigen/src/Core/products/Parallelizer.h +++ b/Eigen/src/Core/products/Parallelizer.h @@ -49,8 +49,8 @@ inline void initParallel() { int nbt; internal::manage_multi_threading(GetAction, &nbt); - std::ptrdiff_t l1, l2; - internal::manage_caching_sizes(GetAction, &l1, &l2); + std::ptrdiff_t l1, l2, l3; + internal::manage_caching_sizes(GetAction, &l1, &l2, &l3); } /** \returns the max number of threads reserved for Eigen diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index d9e6084c3..21f8175d2 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -343,7 +343,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix(kc, mc, nc); + computeProductBlockingSizes(kc, mc, nc, 1); // kc must smaller than mc kc = (std::min)(kc,mc); @@ -432,10 +432,10 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix(kc, mc, nc); + computeProductBlockingSizes(kc, mc, nc, 1); std::size_t sizeB = kc*cols; ei_declare_aligned_stack_constructed_variable(Scalar, blockA, kc*mc, 0); ei_declare_aligned_stack_constructed_variable(Scalar, allocatedBlockB, sizeB, 0); diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index 77aa3e5ee..4cbb79da0 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -412,7 +412,7 @@ struct TriangularProduct Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows())) : ((IsLower) ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols())); - BlockingType blocking(stripedRows, stripedCols, stripedDepth); + BlockingType blocking(stripedRows, stripedCols, stripedDepth, 1, false); internal::product_triangular_matrix_matrix0 ? l2/(4 * sizeof(Scalar) * otherStride) : 0; subcols = std::max((subcols/Traits::nr)*Traits::nr, Traits::nr); diff --git a/blas/level3_impl.h b/blas/level3_impl.h index a05872666..37a803ced 100644 --- a/blas/level3_impl.h +++ b/blas/level3_impl.h @@ -56,7 +56,7 @@ int EIGEN_BLAS_FUNC(gemm)(char *opa, char *opb, int *m, int *n, int *k, RealScal else matrix(c, *m, *n, *ldc) *= beta; } - internal::gemm_blocking_space blocking(*m,*n,*k,true); + internal::gemm_blocking_space blocking(*m,*n,*k,1,true); int code = OP(*opa) | (OP(*opb) << 2); func[code](*m, *n, *k, a, *lda, b, *ldb, c, *ldc, alpha, blocking, 0); @@ -131,12 +131,12 @@ int EIGEN_BLAS_FUNC(trsm)(char *side, char *uplo, char *opa, char *diag, int *m, if(SIDE(*side)==LEFT) { - internal::gemm_blocking_space blocking(*m,*n,*m); + internal::gemm_blocking_space blocking(*m,*n,*m,1,false); func[code](*m, *n, a, *lda, b, *ldb, blocking); } else { - internal::gemm_blocking_space blocking(*m,*n,*n); + internal::gemm_blocking_space blocking(*m,*n,*n,1,false); func[code](*n, *m, a, *lda, b, *ldb, blocking); } @@ -222,12 +222,12 @@ int EIGEN_BLAS_FUNC(trmm)(char *side, char *uplo, char *opa, char *diag, int *m, if(SIDE(*side)==LEFT) { - internal::gemm_blocking_space blocking(*m,*n,*m); + internal::gemm_blocking_space blocking(*m,*n,*m,1,false); func[code](*m, *n, *m, a, *lda, tmp.data(), tmp.outerStride(), b, *ldb, alpha, blocking); } else { - internal::gemm_blocking_space blocking(*m,*n,*n); + internal::gemm_blocking_space blocking(*m,*n,*n,1,false); func[code](*m, *n, *n, tmp.data(), tmp.outerStride(), a, *lda, b, *ldb, alpha, blocking); } return 1; @@ -577,7 +577,7 @@ int EIGEN_BLAS_FUNC(her2k)(char *uplo, char *op, int *n, int *k, RealScalar *pal else if(*n<0) info = 3; else if(*k<0) info = 4; else if(*lda(10000,20000); - std::ptrdiff_t l2 = internal::random(1000000,2000000); - setCpuCacheSizes(l1,l2); + std::ptrdiff_t l2 = internal::random(100000,200000); + std::ptrdiff_t l3 = internal::random(1000000,2000000); + setCpuCacheSizes(l1,l2,l3); VERIFY(l1==l1CacheSize()); VERIFY(l2==l2CacheSize()); std::ptrdiff_t k1 = internal::random(10,100)*16; std::ptrdiff_t m1 = internal::random(10,100)*16; std::ptrdiff_t n1 = internal::random(10,100)*16; // only makes sure it compiles fine - internal::computeProductBlockingSizes(k1,m1,n1); + internal::computeProductBlockingSizes(k1,m1,n1,1); } { diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 7ec60044e..47447f446 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -55,7 +55,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" -//#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index 1e6f276e0..cd992daab 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -766,7 +766,7 @@ struct TensorEvaluator BlockingType; // Sizes of the blocks to load in cache. See the Goto paper for details. - BlockingType blocking(m, n, k, true); + BlockingType blocking(m, n, k, 1, true); const Index kc = blocking.kc(); const Index mc = (std::min)(m, blocking.mc()); const Index nc = (std::min)(n, blocking.nc()); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h index dc0513305..8e4c7c11d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h @@ -152,7 +152,7 @@ struct TensorEvaluator(kc, mc, nc/*, num_threads*/); + internal::computeProductBlockingSizes(kc, mc, nc, num_threads); eigen_assert(mc <= m); eigen_assert(nc <= n); eigen_assert(kc <= k); @@ -197,9 +197,10 @@ struct TensorEvaluator kernel_promises(num_kernel_promises, p); + std::vector kernel_promises(num_kernel_promises); + for (int i = 0; i < kernel_promises.size(); ++i) { + kernel_promises[i].set_value(); + } for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) { const Index k_start = k_block_idx * kc; @@ -275,8 +276,7 @@ struct TensorEvaluator) Func; - this->m_device.enqueueNoFuture(&Self::packRhsAndKernel, arg); + this->m_device.enqueueNoFuture(&Self::packRhsAndKernel, arg); } } } @@ -338,7 +338,6 @@ struct TensorEvaluator Date: Thu, 16 Oct 2014 09:19:32 -0400 Subject: [PATCH 0903/1103] quieted more g++ warnings of the form: warning: typedef XXX locally defined but not used [-Wunused-local-typedefs] --- Eigen/src/Eigenvalues/ComplexSchur_MKL.h | 1 - Eigen/src/Eigenvalues/RealSchur_MKL.h | 4 ---- Eigen/src/QR/ColPivHouseholderQR_MKL.h | 1 - Eigen/src/SVD/JacobiSVD_MKL.h | 4 ++-- 4 files changed, 2 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Eigenvalues/ComplexSchur_MKL.h b/Eigen/src/Eigenvalues/ComplexSchur_MKL.h index 91496ae5b..27aed923c 100644 --- a/Eigen/src/Eigenvalues/ComplexSchur_MKL.h +++ b/Eigen/src/Eigenvalues/ComplexSchur_MKL.h @@ -45,7 +45,6 @@ ComplexSchur >& \ ComplexSchur >::compute(const Matrix& matrix, bool computeU) \ { \ typedef Matrix MatrixType; \ - typedef MatrixType::Scalar Scalar; \ typedef MatrixType::RealScalar RealScalar; \ typedef std::complex ComplexScalar; \ \ diff --git a/Eigen/src/Eigenvalues/RealSchur_MKL.h b/Eigen/src/Eigenvalues/RealSchur_MKL.h index ad9736460..c3089b468 100644 --- a/Eigen/src/Eigenvalues/RealSchur_MKL.h +++ b/Eigen/src/Eigenvalues/RealSchur_MKL.h @@ -44,10 +44,6 @@ template<> inline \ RealSchur >& \ RealSchur >::compute(const Matrix& matrix, bool computeU) \ { \ - typedef Matrix MatrixType; \ - typedef MatrixType::Scalar Scalar; \ - typedef MatrixType::RealScalar RealScalar; \ -\ eigen_assert(matrix.cols() == matrix.rows()); \ \ lapack_int n = matrix.cols(), sdim, info; \ diff --git a/Eigen/src/QR/ColPivHouseholderQR_MKL.h b/Eigen/src/QR/ColPivHouseholderQR_MKL.h index b5b198326..7b6ba0a5e 100644 --- a/Eigen/src/QR/ColPivHouseholderQR_MKL.h +++ b/Eigen/src/QR/ColPivHouseholderQR_MKL.h @@ -49,7 +49,6 @@ ColPivHouseholderQR MatrixType; \ - typedef MatrixType::Scalar Scalar; \ typedef MatrixType::RealScalar RealScalar; \ Index rows = matrix.rows();\ Index cols = matrix.cols();\ diff --git a/Eigen/src/SVD/JacobiSVD_MKL.h b/Eigen/src/SVD/JacobiSVD_MKL.h index decda7540..14e461c4e 100644 --- a/Eigen/src/SVD/JacobiSVD_MKL.h +++ b/Eigen/src/SVD/JacobiSVD_MKL.h @@ -45,8 +45,8 @@ JacobiSVD, ColPiv JacobiSVD, ColPivHouseholderQRPreconditioner>::compute(const Matrix& matrix, unsigned int computationOptions) \ { \ typedef Matrix MatrixType; \ - typedef MatrixType::Scalar Scalar; \ - typedef MatrixType::RealScalar RealScalar; \ + /*typedef MatrixType::Scalar Scalar;*/ \ + /*typedef MatrixType::RealScalar RealScalar;*/ \ allocate(matrix.rows(), matrix.cols(), computationOptions); \ \ /*const RealScalar precision = RealScalar(2) * NumTraits::epsilon();*/ \ From d853adffdba52da6e1dd6b137724e4f4e783dcca Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 16 Oct 2014 10:10:04 -0700 Subject: [PATCH 0904/1103] Avoid calling get_future() more than once on a given promise. --- .../CXX11/src/Tensor/TensorContractionThreadPool.h | 13 ++++++++----- .../Eigen/CXX11/src/Tensor/TensorDeviceType.h | 5 ----- 2 files changed, 8 insertions(+), 10 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h index 8e4c7c11d..cf1352a31 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h @@ -198,8 +198,10 @@ struct TensorEvaluator kernel_promises(num_kernel_promises); + std::vector kernel_futures(num_kernel_promises); for (int i = 0; i < kernel_promises.size(); ++i) { kernel_promises[i].set_value(); + kernel_futures[i] = kernel_promises[i].get_future(); } for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) { @@ -218,8 +220,9 @@ struct TensorEvaluatorready()); } -static EIGEN_STRONG_INLINE void wait_until_ready(Promise* p) { - p->get_future().wait(); - // eigen_assert(p->get_future().ready()); -} - struct ThreadPoolDevice { ThreadPoolDevice(/*ThreadPool* pool, */size_t num_cores) : num_threads_(num_cores) { } From 94e47798f4e462b857a00b4ca60c954c71d16605 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 16 Oct 2014 10:41:07 -0700 Subject: [PATCH 0905/1103] Fixed the return types of unary and binary expressions to properly handle the case where it is different from the input type (e.g. abs(complex)) --- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 16 ++++++++-------- unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h | 7 ++----- unsupported/test/cxx11_tensor_of_complex.cpp | 17 +++++++++++++++++ 3 files changed, 27 insertions(+), 13 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index e324ba8d2..131326615 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -155,8 +155,8 @@ struct TensorEvaluator, Device> typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename internal::traits::Scalar CoeffReturnType; + typedef typename internal::traits::Packet PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } @@ -203,8 +203,8 @@ struct TensorEvaluator, Device> typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename internal::traits::Scalar CoeffReturnType; + typedef typename internal::traits::Packet PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_argImpl.dimensions(); } @@ -257,8 +257,8 @@ struct TensorEvaluator::Scalar CoeffReturnType; + typedef typename internal::traits::Packet PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const @@ -317,8 +317,8 @@ struct TensorEvaluator typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename internal::traits::Scalar CoeffReturnType; + typedef typename internal::traits::Packet PacketReturnType; typedef typename TensorEvaluator::Dimensions Dimensions; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h index de66da13f..6e5503de1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h @@ -84,9 +84,7 @@ struct traits > typedef typename result_of< UnaryOp(typename XprType::Scalar) >::type Scalar; - typedef typename result_of< - UnaryOp(typename XprType::Packet) - >::type Packet; + typedef typename internal::packet_traits::type Packet; typedef typename XprType::Nested XprTypeNested; typedef typename remove_reference::type _XprTypeNested; }; @@ -188,8 +186,7 @@ class TensorCwiseBinaryOp : public TensorBase::Real RealScalar; typedef typename internal::promote_storage_type::ret CoeffReturnType; - typedef typename internal::promote_storage_type::ret PacketReturnType; + typedef typename internal::packet_traits::type PacketReturnType; typedef typename Eigen::internal::nested::type Nested; typedef typename Eigen::internal::traits::StorageKind StorageKind; typedef typename Eigen::internal::traits::Index Index; diff --git a/unsupported/test/cxx11_tensor_of_complex.cpp b/unsupported/test/cxx11_tensor_of_complex.cpp index b5044b962..24b2bcb58 100644 --- a/unsupported/test/cxx11_tensor_of_complex.cpp +++ b/unsupported/test/cxx11_tensor_of_complex.cpp @@ -32,6 +32,22 @@ static void test_additions() } +static void test_abs() +{ + Tensor, 1> data1(3); + Tensor, 1> data2(3); + data1.setRandom(); + data2.setRandom(); + + Tensor abs1 = data1.abs(); + Tensor abs2 = data2.abs(); + for (int i = 0; i < 3; ++i) { + VERIFY_IS_APPROX(abs1(i), std::abs(data1(i))); + VERIFY_IS_APPROX(abs2(i), std::abs(data2(i))); + } +} + + static void test_contractions() { Tensor, 4> t_left(30, 50, 8, 31); @@ -60,5 +76,6 @@ static void test_contractions() void test_cxx11_tensor_of_complex() { CALL_SUBTEST(test_additions()); + CALL_SUBTEST(test_abs()); CALL_SUBTEST(test_contractions()); } From ae697b471c0d3961ebdb633e30046e5fe31fbe24 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 16 Oct 2014 14:52:50 -0700 Subject: [PATCH 0906/1103] Silenced a few compilation warnings Generalized a TensorMap constructor --- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 3 ++- .../Eigen/CXX11/src/Tensor/TensorBroadcasting.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorConcatenation.h | 4 ++-- unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 4 ++-- unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorMap.h | 3 ++- unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h | 6 +++--- unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 4 ++-- unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h | 2 +- unsupported/test/cxx11_tensor_fixed_size.cpp | 10 +++++----- 13 files changed, 24 insertions(+), 22 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index 879057f38..ceed09505 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -1,6 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // +// Copyright (C) 2014 Benoit Steiner // Copyright (C) 2013 Christian Seiler // // This Source Code Form is subject to the terms of the Mozilla @@ -82,7 +83,7 @@ class Tensor : public TensorBase > static const std::size_t NumIndices = NumIndices_; - typedef DSizes Dimensions; + typedef DSizes Dimensions; protected: TensorStorage m_storage; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h index 0e55d4de1..2bd158dac 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h @@ -114,7 +114,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) { m_impl.evalSubExprsIfNeeded(NULL); return true; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h index 9ecea9108..3aa3eba24 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h @@ -136,7 +136,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) { m_impl.evalSubExprsIfNeeded(NULL); return true; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h index b8e43f484..74485b15b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h @@ -140,7 +140,7 @@ struct TensorEvaluator m_outputStrides; array m_leftStrides; array m_rightStrides; TensorEvaluator m_leftImpl; TensorEvaluator m_rightImpl; + const Axis m_axis; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index cd992daab..0db34adb1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -671,10 +671,10 @@ struct TensorContractionEvaluatorBase Index m_j_size; Index m_k_size; - const Device& m_device; - Scalar* m_result; TensorEvaluator m_leftImpl; TensorEvaluator m_rightImpl; + const Device& m_device; + Scalar* m_result; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 34bdd5309..50cb10a33 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -230,7 +230,7 @@ struct TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h index 2c0d2cd0f..0a8c10ac7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMap.h @@ -64,7 +64,8 @@ template class TensorMap : public Tensor } #endif - inline TensorMap(PointerArgType dataPtr, const array& dimensions) + template + inline TensorMap(PointerArgType dataPtr, const Dimensions& dimensions) : m_data(dataPtr), m_dimensions(dimensions) { } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 13109f514..686bf5c24 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -130,8 +130,8 @@ struct TensorEvaluator, Device> Scalar* data() const { return m_impl.data(); } protected: - NewDimensions m_dimensions; TensorEvaluator m_impl; + NewDimensions m_dimensions; }; @@ -381,13 +381,13 @@ struct TensorEvaluator, Devi return inputIndex; } - Dimensions m_dimensions; array m_outputStrides; array, NumDims> m_fastOutputStrides; array m_inputStrides; - const StartIndices m_offsets; TensorEvaluator m_impl; const Device& m_device; + Dimensions m_dimensions; + const StartIndices m_offsets; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h index 8da6e0f26..89c0cff05 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h @@ -215,11 +215,11 @@ struct TensorEvaluator, Device return rslt; } - PaddingDimensions m_padding; Dimensions m_dimensions; array m_outputStrides; array m_inputStrides; TensorEvaluator m_impl; + PaddingDimensions m_padding; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h index 01f2daf52..e2fe32d67 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h @@ -120,7 +120,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) { m_impl.evalSubExprsIfNeeded(NULL); return true; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index eef992106..cbe87394b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -152,7 +152,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) { m_impl.evalSubExprsIfNeeded(NULL); return true; } @@ -217,8 +217,8 @@ struct TensorEvaluator, Device> array m_preservedStrides; array m_reducedStrides; array m_reducedDims; - Op m_reducer; TensorEvaluator m_impl; + Op m_reducer; }; } // end namespace Eigen diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h index 7e0063626..831a9f005 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h @@ -131,7 +131,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) { m_impl.evalSubExprsIfNeeded(NULL); return true; } diff --git a/unsupported/test/cxx11_tensor_fixed_size.cpp b/unsupported/test/cxx11_tensor_fixed_size.cpp index b0501aaa3..99ffc7f07 100644 --- a/unsupported/test/cxx11_tensor_fixed_size.cpp +++ b/unsupported/test/cxx11_tensor_fixed_size.cpp @@ -32,10 +32,10 @@ static void test_1d() vec1(5) = 42.0; vec2(5) = 5.0; float data3[6]; - TensorMap > > vec3(data3, 6); + TensorMap > > vec3(data3, Sizes<6>()); vec3 = vec1.sqrt(); float data4[6]; - TensorMap, RowMajor> > vec4(data4, 6); + TensorMap, RowMajor> > vec4(data4, Sizes<6>()); vec4 = vec2.sqrt(); VERIFY_IS_EQUAL((vec3.size()), 6); @@ -68,9 +68,9 @@ static void test_1d() static void test_2d() { float data1[6]; - TensorMap >> mat1(data1,2,3); + TensorMap >> mat1(data1, Sizes<2, 3>()); float data2[6]; - TensorMap, RowMajor>> mat2(data2,2,3); + TensorMap, RowMajor>> mat2(data2, Sizes<2, 3>()); VERIFY_IS_EQUAL((mat1.size()), 2*3); // VERIFY_IS_EQUAL((mat1.dimension(0)), 2); @@ -166,7 +166,7 @@ static void test_array() for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 7; ++k) { - mat1(array(i,j,k)) = val; + mat1(array{{i,j,k}}) = val; val += 1.0; } } From 65af852b54afca3c76c978c1bfd27d8a1451cab6 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 16 Oct 2014 15:02:30 -0700 Subject: [PATCH 0907/1103] Silenced one last warning --- unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index 0db34adb1..c530b27a7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -48,7 +48,7 @@ class BaseTensorContractionMapper { m_k_strides(k_strides) { } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE void prefetch(int i) { } + EIGEN_STRONG_INLINE void prefetch(int /*i*/) { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(Index row) const { From 8472e697ca606e2ea1a67cc2dfa175757666cc02 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 17 Oct 2014 15:31:11 +0200 Subject: [PATCH 0908/1103] Add lapack interface to JacobiSVD and BDCSVD --- Eigen/src/Core/util/Memory.h | 2 +- Eigen/src/SparseCore/CompressedStorage.h | 2 +- lapack/complex_double.cpp | 3 +- lapack/complex_single.cpp | 3 +- lapack/double.cpp | 3 +- lapack/eigenvalues.cpp | 23 +--- lapack/lapack_common.h | 7 +- lapack/single.cpp | 3 +- lapack/svd.cpp | 138 +++++++++++++++++++++++ 9 files changed, 157 insertions(+), 27 deletions(-) create mode 100644 lapack/svd.cpp diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index 7b39285ea..b49a18cc7 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2010 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // Copyright (C) 2008-2009 Benoit Jacob // Copyright (C) 2009 Kenneth Riddile // Copyright (C) 2010 Hauke Heibel diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index 2845e44e7..e09ad97f0 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed diff --git a/lapack/complex_double.cpp b/lapack/complex_double.cpp index 424d2b8ca..c9c575273 100644 --- a/lapack/complex_double.cpp +++ b/lapack/complex_double.cpp @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009-2011 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -15,3 +15,4 @@ #include "cholesky.cpp" #include "lu.cpp" +#include "svd.cpp" diff --git a/lapack/complex_single.cpp b/lapack/complex_single.cpp index c0b2d29ae..6d11b26cd 100644 --- a/lapack/complex_single.cpp +++ b/lapack/complex_single.cpp @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009-2011 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -15,3 +15,4 @@ #include "cholesky.cpp" #include "lu.cpp" +#include "svd.cpp" diff --git a/lapack/double.cpp b/lapack/double.cpp index d86549e19..ea78bb662 100644 --- a/lapack/double.cpp +++ b/lapack/double.cpp @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009-2011 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -15,3 +15,4 @@ #include "cholesky.cpp" #include "lu.cpp" #include "eigenvalues.cpp" +#include "svd.cpp" diff --git a/lapack/eigenvalues.cpp b/lapack/eigenvalues.cpp index 6141032ab..921c51569 100644 --- a/lapack/eigenvalues.cpp +++ b/lapack/eigenvalues.cpp @@ -7,10 +7,10 @@ // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. -#include "common.h" +#include "lapack_common.h" #include -// computes an LU factorization of a general M-by-N matrix A using partial pivoting with row interchanges +// computes eigen values and vectors of a general N-by-N matrix A EIGEN_LAPACK_FUNC(syev,(char *jobz, char *uplo, int* n, Scalar* a, int *lda, Scalar* w, Scalar* /*work*/, int* lwork, int *info)) { // TODO exploit the work buffer @@ -22,24 +22,7 @@ EIGEN_LAPACK_FUNC(syev,(char *jobz, char *uplo, int* n, Scalar* a, int *lda, Sca else if(*n<0) *info = -3; else if(*lda +// Copyright (C) 2010-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -18,6 +18,11 @@ typedef Eigen::Map > PivotsType; +#if ISCOMPLEX +#define EIGEN_LAPACK_ARG_IF_COMPLEX(X) X, +#else +#define EIGEN_LAPACK_ARG_IF_COMPLEX(X) +#endif #endif // EIGEN_LAPACK_COMMON_H diff --git a/lapack/single.cpp b/lapack/single.cpp index a64ed44e1..c7da3effa 100644 --- a/lapack/single.cpp +++ b/lapack/single.cpp @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009-2011 Gael Guennebaud +// Copyright (C) 2009-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -15,3 +15,4 @@ #include "cholesky.cpp" #include "lu.cpp" #include "eigenvalues.cpp" +#include "svd.cpp" diff --git a/lapack/svd.cpp b/lapack/svd.cpp new file mode 100644 index 000000000..ecac3bab1 --- /dev/null +++ b/lapack/svd.cpp @@ -0,0 +1,138 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "lapack_common.h" +#include +#include + +// computes the singular values/vectors a general M-by-N matrix A using divide-and-conquer +EIGEN_LAPACK_FUNC(gesdd,(char *jobz, int *m, int* n, Scalar* a, int *lda, RealScalar *s, Scalar *u, int *ldu, Scalar *vt, int *ldvt, Scalar* /*work*/, int* lwork, + EIGEN_LAPACK_ARG_IF_COMPLEX(RealScalar */*rwork*/) int * /*iwork*/, int *info)) +{ + // TODO exploit the work buffer + bool query_size = *lwork==-1; + int diag_size = (std::min)(*m,*n); + + *info = 0; + if(*jobz!='A' && *jobz!='S' && *jobz!='O' && *jobz!='N') *info = -1; + else if(*m<0) *info = -2; + else if(*n<0) *info = -3; + else if(*lda=*n && *ldvt<*n)) *info = -10; + + if(*info!=0) + { + int e = -*info; + return xerbla_(SCALAR_SUFFIX_UP"GESDD ", &e, 6); + } + + if(query_size) + { + *lwork = 0; + return 0; + } + + if(*n==0 || *m==0) + return 0; + + PlainMatrixType mat(*m,*n); + mat = matrix(a,*m,*n,*lda); + + int option = *jobz=='A' ? ComputeFullU|ComputeFullV + : *jobz=='S' ? ComputeThinU|ComputeThinV + : *jobz=='O' ? ComputeThinU|ComputeThinV + : 0; + + BDCSVD svd(mat,option); + + make_vector(s,diag_size) = svd.singularValues().head(diag_size); + + if(*jobz=='A') + { + matrix(u,*m,*m,*ldu) = svd.matrixU(); + matrix(vt,*n,*n,*ldvt) = svd.matrixV().adjoint(); + } + else if(*jobz=='S') + { + matrix(u,*m,diag_size,*ldu) = svd.matrixU(); + matrix(vt,diag_size,*n,*ldvt) = svd.matrixV().adjoint(); + } + else if(*jobz=='O' && *m>=*n) + { + matrix(a,*m,*n,*lda) = svd.matrixU(); + matrix(vt,*n,*n,*ldvt) = svd.matrixV().adjoint(); + } + else if(*jobz=='O') + { + matrix(u,*m,*m,*ldu) = svd.matrixU(); + matrix(a,diag_size,*n,*lda) = svd.matrixV().adjoint(); + } + + return 0; +} + +// computes the singular values/vectors a general M-by-N matrix A using two sided jacobi algorithm +EIGEN_LAPACK_FUNC(gesvd,(char *jobu, char *jobv, int *m, int* n, Scalar* a, int *lda, RealScalar *s, Scalar *u, int *ldu, Scalar *vt, int *ldvt, Scalar* /*work*/, int* lwork, + EIGEN_LAPACK_ARG_IF_COMPLEX(RealScalar */*rwork*/) int *info)) +{ + // TODO exploit the work buffer + bool query_size = *lwork==-1; + int diag_size = (std::min)(*m,*n); + + *info = 0; + if( *jobu!='A' && *jobu!='S' && *jobu!='O' && *jobu!='N') *info = -1; + else if((*jobv!='A' && *jobv!='S' && *jobv!='O' && *jobv!='N') + || (*jobu=='O' && *jobv=='O')) *info = -2; + else if(*m<0) *info = -3; + else if(*n<0) *info = -4; + else if(*lda svd(mat,option); + + make_vector(s,diag_size) = svd.singularValues().head(diag_size); + + if(*jobu=='A') matrix(u,*m,*m,*ldu) = svd.matrixU(); + else if(*jobu=='S') matrix(u,*m,diag_size,*ldu) = svd.matrixU(); + else if(*jobu=='O') matrix(a,*m,diag_size,*lda) = svd.matrixU(); + + if(*jobv=='A') matrix(vt,*n,*n,*ldvt) = svd.matrixV().adjoint(); + else if(*jobv=='S') matrix(vt,diag_size,*n,*ldvt) = svd.matrixV().adjoint(); + else if(*jobv=='O') matrix(a,diag_size,*n,*lda) = svd.matrixV().adjoint(); + + return 0; +} From feacfa5f83085a7289b4b2aefc3a64b576fea048 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 17 Oct 2014 15:32:06 +0200 Subject: [PATCH 0909/1103] Fix JacobiSVD wrt undeR/overflow by doing scaling prior to QR preconditioning --- Eigen/src/SVD/JacobiSVD.h | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-) diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 59f88a15a..7a8aa8d3f 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -692,21 +692,20 @@ JacobiSVD::compute(const MatrixType& matrix, unsig // limit for very small denormal numbers to be considered zero in order to avoid infinite loops (see bug 286) const RealScalar considerAsZero = RealScalar(2) * std::numeric_limits::denorm_min(); + // Scaling factor to reduce over/under-flows + RealScalar scale = matrix.cwiseAbs().maxCoeff(); + if(scale==RealScalar(0)) scale = RealScalar(1); + /*** step 1. The R-SVD step: we use a QR decomposition to reduce to the case of a square matrix */ - if(!m_qr_precond_morecols.run(*this, matrix) && !m_qr_precond_morerows.run(*this, matrix)) + if(!m_qr_precond_morecols.run(*this, matrix/scale) && !m_qr_precond_morerows.run(*this, matrix/scale)) { - m_workMatrix = matrix.block(0,0,m_diagSize,m_diagSize); + m_workMatrix = matrix.block(0,0,m_diagSize,m_diagSize) / scale; if(m_computeFullU) m_matrixU.setIdentity(m_rows,m_rows); if(m_computeThinU) m_matrixU.setIdentity(m_rows,m_diagSize); if(m_computeFullV) m_matrixV.setIdentity(m_cols,m_cols); if(m_computeThinV) m_matrixV.setIdentity(m_cols, m_diagSize); } - - // Scaling factor to reduce over/under-flows - RealScalar scale = m_workMatrix.cwiseAbs().maxCoeff(); - if(scale==RealScalar(0)) scale = RealScalar(1); - m_workMatrix /= scale; /*** step 2. The main Jacobi SVD iteration. ***/ From 4b7c3abbea97c604d56f732cd5ba5f8ff8356814 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 17 Oct 2014 15:32:55 +0200 Subject: [PATCH 0910/1103] Fix D&C SVD wrt zero matrices --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 30 ++++++++------------------- 1 file changed, 9 insertions(+), 21 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index e8e795589..e2551cf88 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -194,7 +194,7 @@ public: }; //end class BDCSVD -// Methode to allocate ans initialize matrix and attributes +// Method to allocate and initialize matrix and attributes template void BDCSVD::allocate(Index rows, Index cols, unsigned int computationOptions) { @@ -215,24 +215,6 @@ void BDCSVD::allocate(Index rows, Index cols, unsigned int computati if (m_compV) m_naiveV = MatrixXr::Zero(m_diagSize, m_diagSize); }// end allocate -// Methode which compute the BDCSVD for the int -template<> -BDCSVD >& BDCSVD >::compute(const MatrixType& matrix, unsigned int computationOptions) -{ - allocate(matrix.rows(), matrix.cols(), computationOptions); - m_nonzeroSingularValues = 0; - m_computed = Matrix::Zero(rows(), cols()); - - m_singularValues.head(m_diagSize).setZero(); - - if (m_computeFullU) m_matrixU.setZero(rows(), rows()); - if (m_computeFullV) m_matrixV.setZero(cols(), cols()); - m_isInitialized = true; - return *this; -} - - -// Methode which compute the BDCSVD template BDCSVD& BDCSVD::compute(const MatrixType& matrix, unsigned int computationOptions) { @@ -612,6 +594,8 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec #endif #ifdef EIGEN_BDCSVD_SANITY_CHECKS + assert(U.allFinite()); + assert(V.allFinite()); assert((U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() < 1e-14 * n); assert((V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() < 1e-14 * n); assert(m_naiveU.allFinite()); @@ -836,8 +820,12 @@ void BDCSVD::perturbCol0 using std::sqrt; Index n = col0.size(); Index m = perm.size(); - Index last = perm(m-1); - + if(m==0) + { + zhat.setZero(); + return; + } + Index last = last = perm(m-1); // The offset permits to skip deflated entries while computing zhat for (Index k = 0; k < n; ++k) { From a13bc2220457224152e8e301c928f28013691d15 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 17 Oct 2014 15:34:39 +0200 Subject: [PATCH 0911/1103] Ignore automalically imported lapack source files --- .hgignore | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.hgignore b/.hgignore index e33ba2e9d..769a47f1f 100644 --- a/.hgignore +++ b/.hgignore @@ -30,3 +30,5 @@ log patch a a.* +lapack/testing +lapack/reference From a370b1f2e20804f0b99bcda1c3940f00300e4f90 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 17 Oct 2014 16:52:56 +0200 Subject: [PATCH 0912/1103] Fix SparseLU::absDeterminant and add respective unit test --- Eigen/src/SparseLU/SparseLU.h | 5 ++- .../src/SparseLU/SparseLU_SupernodalMatrix.h | 4 +-- test/sparse_solver.h | 33 +++++++++++++++++++ test/sparselu.cpp | 3 ++ 4 files changed, 40 insertions(+), 5 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index b02796293..d72d7f150 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -249,14 +249,13 @@ class SparseLU : public SparseSolverBase >, eigen_assert(m_factorizationIsOk && "The matrix should be factorized first."); // Initialize with the determinant of the row matrix Scalar det = Scalar(1.); - //Note that the diagonal blocks of U are stored in supernodes, + // Note that the diagonal blocks of U are stored in supernodes, // which are available in the L part :) for (Index j = 0; j < this->cols(); ++j) { for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it) { - if(it.row() < j) continue; - if(it.row() == j) + if(it.index() == j) { det *= abs(it.value()); break; diff --git a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h index 6b0b83e0b..e8ee35a94 100644 --- a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h +++ b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h @@ -189,8 +189,8 @@ class MappedSuperNodalMatrix::InnerIterator m_idval(mat.colIndexPtr()[outer]), m_startidval(m_idval), m_endidval(mat.colIndexPtr()[outer+1]), - m_idrow(mat.rowIndexPtr()[outer]), - m_endidrow(mat.rowIndexPtr()[outer+1]) + m_idrow(mat.rowIndexPtr()[mat.supToCol()[mat.colToSup()[outer]]]), + m_endidrow(mat.rowIndexPtr()[mat.supToCol()[mat.colToSup()[outer]]+1]) {} inline InnerIterator& operator++() { diff --git a/test/sparse_solver.h b/test/sparse_solver.h index 8c8d7f939..b10eaebcc 100644 --- a/test/sparse_solver.h +++ b/test/sparse_solver.h @@ -133,7 +133,23 @@ void check_sparse_determinant(Solver& solver, const typename Solver::MatrixType& Scalar refDet = dA.determinant(); VERIFY_IS_APPROX(refDet,solver.determinant()); } +template +void check_sparse_abs_determinant(Solver& solver, const typename Solver::MatrixType& A, const DenseMat& dA) +{ + using std::abs; + typedef typename Solver::MatrixType Mat; + typedef typename Mat::Scalar Scalar; + + solver.compute(A); + if (solver.info() != Success) + { + std::cerr << "sparse solver testing: factorization failed (check_sparse_abs_determinant)\n"; + return; + } + Scalar refDet = abs(dA.determinant()); + VERIFY_IS_APPROX(refDet,solver.absDeterminant()); +} template int generate_sparse_spd_problem(Solver& , typename Solver::MatrixType& A, typename Solver::MatrixType& halfA, DenseMat& dA, int maxSize = 300) @@ -333,3 +349,20 @@ template void check_sparse_square_determinant(Solver& solver) check_sparse_determinant(solver, A, dA); } } + +template void check_sparse_square_abs_determinant(Solver& solver) +{ + typedef typename Solver::MatrixType Mat; + typedef typename Mat::Scalar Scalar; + typedef Matrix DenseMatrix; + + // generate the problem + Mat A; + DenseMatrix dA; + generate_sparse_square_problem(solver, A, dA, 30); + A.makeCompressed(); + for (int i = 0; i < g_repeat; i++) { + check_sparse_abs_determinant(solver, A, dA); + } +} + diff --git a/test/sparselu.cpp b/test/sparselu.cpp index 37980defc..52371cb12 100644 --- a/test/sparselu.cpp +++ b/test/sparselu.cpp @@ -44,6 +44,9 @@ template void test_sparselu_T() check_sparse_square_solving(sparselu_colamd); check_sparse_square_solving(sparselu_amd); check_sparse_square_solving(sparselu_natural); + + check_sparse_square_abs_determinant(sparselu_colamd); + check_sparse_square_abs_determinant(sparselu_amd); } void test_sparselu() From 7acd38d19e2f9559825c78b4be8644f3b10496fb Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 17 Oct 2014 09:49:03 -0700 Subject: [PATCH 0913/1103] Created some benchmarks for the tensor code --- bench/btl/CMakeLists.txt | 1 + bench/btl/libs/tensors/CMakeLists.txt | 44 +++++++++ bench/btl/libs/tensors/main_linear.cpp | 23 +++++ bench/btl/libs/tensors/main_matmat.cpp | 21 +++++ bench/btl/libs/tensors/main_vecmat.cpp | 21 +++++ bench/btl/libs/tensors/tensor_interface.hh | 105 +++++++++++++++++++++ unsupported/Eigen/CXX11/Core | 2 + 7 files changed, 217 insertions(+) create mode 100644 bench/btl/libs/tensors/CMakeLists.txt create mode 100644 bench/btl/libs/tensors/main_linear.cpp create mode 100644 bench/btl/libs/tensors/main_matmat.cpp create mode 100644 bench/btl/libs/tensors/main_vecmat.cpp create mode 100644 bench/btl/libs/tensors/tensor_interface.hh diff --git a/bench/btl/CMakeLists.txt b/bench/btl/CMakeLists.txt index b299d9899..9444b450c 100644 --- a/bench/btl/CMakeLists.txt +++ b/bench/btl/CMakeLists.txt @@ -97,6 +97,7 @@ ENABLE_TESTING() add_subdirectory(libs/eigen3) add_subdirectory(libs/eigen2) +add_subdirectory(libs/tensors) add_subdirectory(libs/BLAS) add_subdirectory(libs/ublas) add_subdirectory(libs/gmm) diff --git a/bench/btl/libs/tensors/CMakeLists.txt b/bench/btl/libs/tensors/CMakeLists.txt new file mode 100644 index 000000000..09d6d8e43 --- /dev/null +++ b/bench/btl/libs/tensors/CMakeLists.txt @@ -0,0 +1,44 @@ + + +if((NOT TENSOR_INCLUDE_DIR) AND Eigen_SOURCE_DIR) + # unless TENSOR_INCLUDE_DIR is defined, let's use current Eigen version + set(TENSOR_INCLUDE_DIR ${Eigen_SOURCE_DIR}) + set(TENSOR_FOUND TRUE) +else() + find_package(Tensor) +endif() + +if (TENSOR_FOUND) + + include_directories(${TENSOR_INCLUDE_DIR}) + btl_add_bench(btl_tensor_linear main_linear.cpp) + btl_add_bench(btl_tensor_vecmat main_vecmat.cpp) + btl_add_bench(btl_tensor_matmat main_matmat.cpp) + + btl_add_target_property(btl_tensor_linear COMPILE_FLAGS "-fno-exceptions -DBTL_PREFIX=tensor") + btl_add_target_property(btl_tensor_vecmat COMPILE_FLAGS "-fno-exceptions -DBTL_PREFIX=tensor") + btl_add_target_property(btl_tensor_matmat COMPILE_FLAGS "-fno-exceptions -DBTL_PREFIX=tensor") + + option(BTL_BENCH_NOGCCVEC "also bench Eigen explicit vec without GCC's auto vec" OFF) + if(CMAKE_COMPILER_IS_GNUCXX AND BTL_BENCH_NOGCCVEC) + btl_add_bench(btl_tensor_nogccvec_linear main_linear.cpp) + btl_add_bench(btl_tensor_nogccvec_vecmat main_vecmat.cpp) + btl_add_bench(btl_tensor_nogccvec_matmat main_matmat.cpp) + + btl_add_target_property(btl_tensor_nogccvec_linear COMPILE_FLAGS "-fno-exceptions -fno-tree-vectorize -DBTL_PREFIX=tensor_nogccvec") + btl_add_target_property(btl_tensor_nogccvec_vecmat COMPILE_FLAGS "-fno-exceptions -fno-tree-vectorize -DBTL_PREFIX=tensor_nogccvec") + btl_add_target_property(btl_tensor_nogccvec_matmat COMPILE_FLAGS "-fno-exceptions -fno-tree-vectorize -DBTL_PREFIX=tensor_nogccvec") + endif() + + + if(NOT BTL_NOVEC) + btl_add_bench(btl_tensor_novec_linear main_linear.cpp OFF) + btl_add_bench(btl_tensor_novec_vecmat main_vecmat.cpp OFF) + btl_add_bench(btl_tensor_novec_matmat main_matmat.cpp OFF) + btl_add_target_property(btl_tensor_novec_linear COMPILE_FLAGS "-fno-exceptions -DEIGEN_DONT_VECTORIZE -DBTL_PREFIX=tensor_novec") + btl_add_target_property(btl_tensor_novec_vecmat COMPILE_FLAGS "-fno-exceptions -DEIGEN_DONT_VECTORIZE -DBTL_PREFIX=tensor_novec") + btl_add_target_property(btl_tensor_novec_matmat COMPILE_FLAGS "-fno-exceptions -DEIGEN_DONT_VECTORIZE -DBTL_PREFIX=tensor_novec") + + endif(NOT BTL_NOVEC) + +endif (TENSOR_FOUND) diff --git a/bench/btl/libs/tensors/main_linear.cpp b/bench/btl/libs/tensors/main_linear.cpp new file mode 100644 index 000000000..e257f1e72 --- /dev/null +++ b/bench/btl/libs/tensors/main_linear.cpp @@ -0,0 +1,23 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "utilities.h" +#include "tensor_interface.hh" +#include "bench.hh" +#include "basic_actions.hh" + +BTL_MAIN; + +int main() +{ + bench > >(MIN_AXPY,MAX_AXPY,NB_POINT); + bench > >(MIN_AXPY,MAX_AXPY,NB_POINT); + + return 0; +} diff --git a/bench/btl/libs/tensors/main_matmat.cpp b/bench/btl/libs/tensors/main_matmat.cpp new file mode 100644 index 000000000..675fcfc6d --- /dev/null +++ b/bench/btl/libs/tensors/main_matmat.cpp @@ -0,0 +1,21 @@ +//===================================================== +// Copyright (C) 2014 Benoit Steiner +//===================================================== +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +// +#include "utilities.h" +#include "tensor_interface.hh" +#include "bench.hh" +#include "basic_actions.hh" + +BTL_MAIN; + +int main() +{ + bench > >(MIN_MM,MAX_MM,NB_POINT); + + return 0; +} diff --git a/bench/btl/libs/tensors/main_vecmat.cpp b/bench/btl/libs/tensors/main_vecmat.cpp new file mode 100644 index 000000000..1af00c81b --- /dev/null +++ b/bench/btl/libs/tensors/main_vecmat.cpp @@ -0,0 +1,21 @@ +//===================================================== +// Copyright (C) 2014 Benoit Steiner +//===================================================== +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +// +#include "utilities.h" +#include "tensor_interface.hh" +#include "bench.hh" +#include "basic_actions.hh" + +BTL_MAIN; + +int main() +{ + bench > >(MIN_MV,MAX_MV,NB_POINT); + + return 0; +} diff --git a/bench/btl/libs/tensors/tensor_interface.hh b/bench/btl/libs/tensors/tensor_interface.hh new file mode 100644 index 000000000..97b8e0f0b --- /dev/null +++ b/bench/btl/libs/tensors/tensor_interface.hh @@ -0,0 +1,105 @@ +//===================================================== +// Copyright (C) 2014 Benoit Steiner +//===================================================== +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. +// +#ifndef TENSOR_INTERFACE_HH +#define TENSOR_INTERFACE_HH + +#include +#include +#include "btl.hh" + +using namespace Eigen; + +template +class tensor_interface +{ +public : + typedef real real_type; + typedef typename Eigen::Tensor::Index Index; + + typedef std::vector stl_vector; + typedef std::vector stl_matrix; + + typedef Eigen::Tensor gene_matrix; + typedef Eigen::Tensor gene_vector; + + + static inline std::string name( void ) + { + return EIGEN_MAKESTRING(BTL_PREFIX); + } + + static void free_matrix(gene_matrix & /*A*/, int /*N*/) {} + + static void free_vector(gene_vector & /*B*/) {} + + static BTL_DONT_INLINE void matrix_from_stl(gene_matrix & A, stl_matrix & A_stl){ + A.resize(Eigen::array(A_stl[0].size(), A_stl.size())); + + for (unsigned int j=0; j(i,j)) = A_stl[j][i]; + } + } + } + + static BTL_DONT_INLINE void vector_from_stl(gene_vector & B, stl_vector & B_stl){ + B.resize(B_stl.size()); + + for (unsigned int i=0; i(i,j)); + } + } + } + + static inline void matrix_matrix_product(const gene_matrix & A, const gene_matrix & B, gene_matrix & X, int /*N*/){ + typedef typename Eigen::Tensor::DimensionPair DimPair; + const Eigen::array dims(DimPair(1, 0)); + X/*.noalias()*/ = A.contract(B, dims); + } + + static inline void matrix_vector_product(const gene_matrix & A, const gene_vector & B, gene_vector & X, int /*N*/){ + typedef typename Eigen::Tensor::DimensionPair DimPair; + const Eigen::array dims(DimPair(1, 0)); + X/*.noalias()*/ = A.contract(B, dims); + } + + static inline void axpy(real coef, const gene_vector & X, gene_vector & Y, int /*N*/){ + Y += X.constant(coef) * X; + } + + static inline void axpby(real a, const gene_vector & X, real b, gene_vector & Y, int /*N*/){ + Y = X.constant(a)*X + Y.constant(b)*Y; + } + + static EIGEN_DONT_INLINE void copy_matrix(const gene_matrix & source, gene_matrix & cible, int /*N*/){ + cible = source; + } + + static EIGEN_DONT_INLINE void copy_vector(const gene_vector & source, gene_vector & cible, int /*N*/){ + cible = source; + } +}; + +#endif diff --git a/unsupported/Eigen/CXX11/Core b/unsupported/Eigen/CXX11/Core index f6c3b49bb..292f09564 100644 --- a/unsupported/Eigen/CXX11/Core +++ b/unsupported/Eigen/CXX11/Core @@ -30,6 +30,8 @@ * \endcode */ +#include + // Emulate the cxx11 functionality that we need if the compiler doesn't support it. #if __cplusplus <= 199711L #include "src/Core/util/EmulateCXX11Meta.h" From f786897e4b96737767effc85bedb78f06dc46dc5 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 17 Oct 2014 15:33:27 -0700 Subject: [PATCH 0914/1103] Added access to the unerlying raw data of a tnsor slice/chip whenever possible --- .../Eigen/CXX11/src/Tensor/TensorChipping.h | 9 ++- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 21 +++++- unsupported/test/cxx11_tensor_chipping.cpp | 37 +++++++++++ unsupported/test/cxx11_tensor_morphing.cpp | 64 ++++++++++++++++++- 4 files changed, 126 insertions(+), 5 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h index 3aa3eba24..b862a8fd3 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h @@ -157,7 +157,14 @@ struct TensorEvaluator, Device> }*/ - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { + Scalar* result = m_impl.data(); + if (DimId == NumDims && result) { + return result + m_inputOffset; + } else { + return NULL; + } + } protected: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 686bf5c24..3447592eb 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -366,7 +366,26 @@ struct TensorEvaluator, Devi } } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { + Scalar* result = m_impl.data(); + if (result) { + Index offset = 0; + for (int i = 0; i < NumDims; ++i) { + if (m_dimensions[i] != m_impl.dimensions()[i]) { + offset += m_offsets[i] * m_inputStrides[i]; + for (int j = i+1; j < NumDims; ++j) { + if (m_dimensions[j] > 1) { + return NULL; + } + offset += m_offsets[j] * m_inputStrides[j]; + } + break; + } + } + return result + offset; + } + return NULL; + } protected: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const diff --git a/unsupported/test/cxx11_tensor_chipping.cpp b/unsupported/test/cxx11_tensor_chipping.cpp index 8c8a0cec2..0027b2888 100644 --- a/unsupported/test/cxx11_tensor_chipping.cpp +++ b/unsupported/test/cxx11_tensor_chipping.cpp @@ -236,9 +236,46 @@ static void test_chip_as_lvalue() } +static void test_chip_raw_data() +{ + Tensor tensor(2,3,5,7,11); + tensor.setRandom(); + + typedef TensorEvaluator(3)), DefaultDevice> Evaluator4; + auto chip = Evaluator4(tensor.chip<4>(3), DefaultDevice()); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + int chip_index = i + 2 * (j + 3 * (k + 5 * l)); + VERIFY_IS_EQUAL(chip.data()[chip_index], tensor(i,j,k,l,3)); + } + } + } + } + + typedef TensorEvaluator(0)), DefaultDevice> Evaluator0; + auto chip0 = Evaluator0(tensor.chip<0>(0), DefaultDevice()); + VERIFY_IS_EQUAL(chip0.data(), static_cast(0)); + + typedef TensorEvaluator(0)), DefaultDevice> Evaluator1; + auto chip1 = Evaluator1(tensor.chip<1>(0), DefaultDevice()); + VERIFY_IS_EQUAL(chip1.data(), static_cast(0)); + + typedef TensorEvaluator(0)), DefaultDevice> Evaluator2; + auto chip2 = Evaluator2(tensor.chip<2>(0), DefaultDevice()); + VERIFY_IS_EQUAL(chip2.data(), static_cast(0)); + + typedef TensorEvaluator(0)), DefaultDevice> Evaluator3; + auto chip3 = Evaluator3(tensor.chip<3>(0), DefaultDevice()); + VERIFY_IS_EQUAL(chip3.data(), static_cast(0)); +} + + void test_cxx11_tensor_chipping() { CALL_SUBTEST(test_simple_chip()); CALL_SUBTEST(test_chip_in_expr()); CALL_SUBTEST(test_chip_as_lvalue()); + CALL_SUBTEST(test_chip_raw_data()); } diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp index fd1b1fa32..78b0dade0 100644 --- a/unsupported/test/cxx11_tensor_morphing.cpp +++ b/unsupported/test/cxx11_tensor_morphing.cpp @@ -12,7 +12,6 @@ #include using Eigen::Tensor; -using Eigen::IndexPair; static void test_simple_reshape() { @@ -53,7 +52,8 @@ static void test_reshape_in_expr() { TensorMap> tensor2(m2.data(), 3,5,7,11,13); Tensor::Dimensions newDims1{{2,3*5*7*11}}; Tensor::Dimensions newDims2{{3*5*7*11,13}}; - Eigen::array, 1> contract_along{{IndexPair(1, 0)}}; + typedef Tensor::DimensionPair DimPair; + array contract_along{{DimPair(1, 0)}}; Tensor tensor3(2,13); tensor3 = tensor1.reshape(newDims1).contract(tensor2.reshape(newDims2), contract_along); @@ -126,7 +126,8 @@ static void test_slice_in_expr() { TensorMap> tensor1(m1.data(), 7, 7); TensorMap> tensor2(m2.data(), 3, 3); Tensor tensor3(3,1); - array, 1> contract_along{{IndexPair(1, 0)}}; + typedef Tensor::DimensionPair DimPair; + array contract_along{{DimPair(1, 0)}}; Eigen::DSizes indices1(1,2); Eigen::DSizes sizes1(3,3); @@ -190,6 +191,62 @@ static void test_slice_as_lvalue() } +static void test_slice_raw_data() +{ + Tensor tensor(3,5,7,11); + tensor.setRandom(); + + Eigen::DSizes offsets(1,2,3,4); + Eigen::DSizes extents(1,1,1,1); + typedef TensorEvaluator SliceEvaluator; + auto slice1 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice1.dimensions().TotalSize(), 1ul); + VERIFY_IS_EQUAL(slice1.data()[0], tensor(1,2,3,4)); + + extents = Eigen::DSizes(2,1,1,1); + auto slice2 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice2.dimensions().TotalSize(), 2ul); + VERIFY_IS_EQUAL(slice2.data()[0], tensor(1,2,3,4)); + VERIFY_IS_EQUAL(slice2.data()[1], tensor(2,2,3,4)); + + extents = Eigen::DSizes(1,2,1,1); + auto slice3 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice3.dimensions().TotalSize(), 2ul); + VERIFY_IS_EQUAL(slice3.data(), static_cast(0)); + + offsets = Eigen::DSizes(0,2,3,4); + extents = Eigen::DSizes(3,2,1,1); + auto slice4 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice4.dimensions().TotalSize(), 6ul); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 2; ++j) { + VERIFY_IS_EQUAL(slice4.data()[i+3*j], tensor(i,2+j,3,4)); + } + } + + offsets = Eigen::DSizes(0,0,0,4); + extents = Eigen::DSizes(3,5,7,2); + auto slice5 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice5.dimensions().TotalSize(), 210ul); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 2; ++l) { + int slice_index = i + 3 * (j + 5 * (k + 7 * l)); + VERIFY_IS_EQUAL(slice5.data()[slice_index], tensor(i,j,k,l+4)); + } + } + } + } + + offsets = Eigen::DSizes(0,0,0,0); + extents = Eigen::DSizes(3,5,7,11); + auto slice6 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice6.dimensions().TotalSize(), 3ul*5*7*11); + VERIFY_IS_EQUAL(slice6.data(), tensor.data()); +} + + void test_cxx11_tensor_morphing() { CALL_SUBTEST(test_simple_reshape()); @@ -199,4 +256,5 @@ void test_cxx11_tensor_morphing() CALL_SUBTEST(test_simple_slice()); CALL_SUBTEST(test_slice_in_expr()); CALL_SUBTEST(test_slice_as_lvalue()); + CALL_SUBTEST(test_slice_raw_data()); } From 8838b0a1fff2cb01fab3a55312a16ae20daead13 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sun, 19 Oct 2014 22:42:20 +0200 Subject: [PATCH 0915/1103] Fix SparseQR::rank for a completely empty matrix. --- Eigen/src/SparseQR/SparseQR.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index f6a9af672..879ac553d 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -378,6 +378,8 @@ void SparseQR::factorize(const MatrixType& mat) { RealScalar max2Norm = 0.0; for (int j = 0; j < n; j++) max2Norm = (max)(max2Norm, m_pmat.col(j).norm()); + if(max2Norm==RealScalar(0)) + max2Norm = RealScalar(1); pivotThreshold = 20 * (m + n) * max2Norm * NumTraits::epsilon(); } @@ -386,7 +388,7 @@ void SparseQR::factorize(const MatrixType& mat) Index nonzeroCol = 0; // Record the number of valid pivots m_Q.startVec(0); - + // Left looking rank-revealing QR factorization: compute a column of R and Q at a time for (Index col = 0; col < n; ++col) { @@ -554,7 +556,7 @@ void SparseQR::factorize(const MatrixType& mat) m_R.finalize(); m_R.makeCompressed(); m_isQSorted = false; - + m_nonzeropivots = nonzeroCol; if(nonzeroCol Date: Mon, 20 Oct 2014 10:48:40 +0200 Subject: [PATCH 0916/1103] Fix bug #894: the sign of LDLT was not re-initialized at each call of compute() --- Eigen/src/Cholesky/LDLT.h | 1 + test/cholesky.cpp | 12 +++++++----- 2 files changed, 8 insertions(+), 5 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index b4016cd6c..dfc473df1 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -433,6 +433,7 @@ LDLT& LDLT::compute(const MatrixType& a) m_transpositions.resize(size); m_isInitialized = false; m_temporary.resize(size); + m_sign = internal::ZeroSign; internal::ldlt_inplace::unblocked(m_matrix, m_transpositions, m_temporary, m_sign); diff --git a/test/cholesky.cpp b/test/cholesky.cpp index a883192ab..33e32a322 100644 --- a/test/cholesky.cpp +++ b/test/cholesky.cpp @@ -316,33 +316,35 @@ template void cholesky_definiteness(const MatrixType& m) { eigen_assert(m.rows() == 2 && m.cols() == 2); MatrixType mat; + LDLT ldlt(2); + { mat << 1, 0, 0, -1; - LDLT ldlt(mat); + ldlt.compute(mat); VERIFY(!ldlt.isNegative()); VERIFY(!ldlt.isPositive()); } { mat << 1, 2, 2, 1; - LDLT ldlt(mat); + ldlt.compute(mat); VERIFY(!ldlt.isNegative()); VERIFY(!ldlt.isPositive()); } { mat << 0, 0, 0, 0; - LDLT ldlt(mat); + ldlt.compute(mat); VERIFY(ldlt.isNegative()); VERIFY(ldlt.isPositive()); } { mat << 0, 0, 0, 1; - LDLT ldlt(mat); + ldlt.compute(mat); VERIFY(!ldlt.isNegative()); VERIFY(ldlt.isPositive()); } { mat << -1, 0, 0, 0; - LDLT ldlt(mat); + ldlt.compute(mat); VERIFY(ldlt.isNegative()); VERIFY(!ldlt.isPositive()); } From b4a9b3f4960657bf2fc156ba5c25a3a32417b861 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 20 Oct 2014 11:04:32 +0200 Subject: [PATCH 0917/1103] Add unit tests for Rotation2D's inverse(), operator*, slerp, and fix regression wrt explicit ctor change --- Eigen/src/Geometry/Rotation2D.h | 6 +++--- test/geo_transformations.cpp | 17 ++++++++++++++--- 2 files changed, 17 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Geometry/Rotation2D.h b/Eigen/src/Geometry/Rotation2D.h index 776c36144..4feb3d4d2 100644 --- a/Eigen/src/Geometry/Rotation2D.h +++ b/Eigen/src/Geometry/Rotation2D.h @@ -71,11 +71,11 @@ public: inline Scalar& angle() { return m_angle; } /** \returns the inverse rotation */ - inline Rotation2D inverse() const { return -m_angle; } + inline Rotation2D inverse() const { return Rotation2D(-m_angle); } /** Concatenates two rotations */ inline Rotation2D operator*(const Rotation2D& other) const - { return m_angle + other.m_angle; } + { return Rotation2D(m_angle + other.m_angle); } /** Concatenates two rotations */ inline Rotation2D& operator*=(const Rotation2D& other) @@ -93,7 +93,7 @@ public: * parameter \a t. It is in fact equivalent to a linear interpolation. */ inline Rotation2D slerp(const Scalar& t, const Rotation2D& other) const - { return m_angle * (1-t) + other.angle() * t; } + { return Rotation2D(m_angle * (1-t) + other.angle() * t); } /** \returns \c *this with scalar type casted to \a NewScalarType * diff --git a/test/geo_transformations.cpp b/test/geo_transformations.cpp index 1768d7b8a..e189217eb 100644 --- a/test/geo_transformations.cpp +++ b/test/geo_transformations.cpp @@ -98,7 +98,7 @@ template void transformations() Matrix3 matrot1, m; Scalar a = internal::random(-Scalar(M_PI), Scalar(M_PI)); - Scalar s0 = internal::random(); + Scalar s0 = internal::random(), s1 = internal::random(); VERIFY_IS_APPROX(v0, AngleAxisx(a, v0.normalized()) * v0); VERIFY_IS_APPROX(-v0, AngleAxisx(Scalar(M_PI), v0.unitOrthogonal()) * v0); @@ -394,10 +394,21 @@ template void transformations() Rotation2D r2d1d = r2d1.template cast(); VERIFY_IS_APPROX(r2d1d.template cast(),r2d1); - t20 = Translation2(v20) * (Rotation2D(s0) * Eigen::Scaling(s0)); - t21 = Translation2(v20) * Rotation2D(s0) * Eigen::Scaling(s0); + Rotation2D R0(s0), R1(s1); + + t20 = Translation2(v20) * (R0 * Eigen::Scaling(s0)); + t21 = Translation2(v20) * R0 * Eigen::Scaling(s0); VERIFY_IS_APPROX(t20,t21); + t20 = Translation2(v20) * (R0 * R0.inverse() * Eigen::Scaling(s0)); + t21 = Translation2(v20) * Eigen::Scaling(s0); + VERIFY_IS_APPROX(t20,t21); + + VERIFY_IS_APPROX(s0, (R0.slerp(0, R1)).angle()); + VERIFY_IS_APPROX(s1, (R0.slerp(1, R1)).angle()); + VERIFY_IS_APPROX(s0, (R0.slerp(0.5, R0)).angle()); + VERIFY_IS_APPROX(Scalar(0), (R0.slerp(0.5, R0.inverse())).angle()); + // check basic features { Rotation2D r1; // default ctor From aa5f79206fb632d141c3555338f89f59d1bb4633 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 20 Oct 2014 11:38:51 +0200 Subject: [PATCH 0918/1103] Fix bug #859: pexp(NaN) returned Inf instead of NaN --- Eigen/src/Core/arch/SSE/MathFunctions.h | 2 +- test/main.h | 20 ++++++++++++++++++++ test/packetmath.cpp | 6 ++++++ test/stable_norm.cpp | 20 -------------------- 4 files changed, 27 insertions(+), 21 deletions(-) diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index 8f78b3a6c..b549e4870 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -167,7 +167,7 @@ Packet4f pexp(const Packet4f& _x) emm0 = _mm_cvttps_epi32(fx); emm0 = _mm_add_epi32(emm0, p4i_0x7f); emm0 = _mm_slli_epi32(emm0, 23); - return pmul(y, Packet4f(_mm_castsi128_ps(emm0))); + return pmax(pmul(y, Packet4f(_mm_castsi128_ps(emm0))), _x); } template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d pexp(const Packet2d& _x) diff --git a/test/main.h b/test/main.h index 371c7e602..adf4adb31 100644 --- a/test/main.h +++ b/test/main.h @@ -432,6 +432,26 @@ void randomPermutationVector(PermutationVectorType& v, typename PermutationVecto } } +template bool isNotNaN(const T& x) +{ + return x==x; +} + +template bool isNaN(const T& x) +{ + return x!=x; +} + +template bool isInf(const T& x) +{ + return x > NumTraits::highest(); +} + +template bool isMinusInf(const T& x) +{ + return x < NumTraits::lowest(); +} + } // end namespace Eigen template struct GetDifferentType; diff --git a/test/packetmath.cpp b/test/packetmath.cpp index e716d6d9a..a4166d868 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -297,6 +297,12 @@ template void packetmath_real() data2[i] = internal::random(-87,88); } CHECK_CWISE1_IF(internal::packet_traits::HasExp, std::exp, internal::pexp); + { + data1[0] = std::numeric_limits::quiet_NaN(); + packet_helper::HasExp,Packet> h; + h.store(data2, internal::pexp(h.load(data1))); + VERIFY(isNaN(data2[0])); + } for (int i=0; i bool isNotNaN(const T& x) -{ - return x==x; -} - -template bool isNaN(const T& x) -{ - return x!=x; -} - -template bool isInf(const T& x) -{ - return x > NumTraits::highest(); -} - -template bool isMinusInf(const T& x) -{ - return x < NumTraits::lowest(); -} - // workaround aggressive optimization in ICC template EIGEN_DONT_INLINE T sub(T a, T b) { return a - b; } From 84aaa03182f6e31fa251d93730d628975c44207c Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 20 Oct 2014 13:13:43 +0200 Subject: [PATCH 0919/1103] Addendum to bug #859: pexp(NaN) for double did not return NaN, also, plog(NaN) did not return NaN. psqrt(NaN) and psqrt(-1) shall return NaN if EIGEN_FAST_MATH==0 --- Eigen/src/Core/arch/SSE/MathFunctions.h | 4 ++-- test/packetmath.cpp | 16 +++++++++++++++- 2 files changed, 17 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index b549e4870..9ffba5b41 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -52,7 +52,7 @@ Packet4f plog(const Packet4f& _x) Packet4i emm0; - Packet4f invalid_mask = _mm_cmplt_ps(x, _mm_setzero_ps()); + Packet4f invalid_mask = _mm_cmpnge_ps(x, _mm_setzero_ps()); // not greater equal is true if x is NaN Packet4f iszero_mask = _mm_cmpeq_ps(x, _mm_setzero_ps()); x = pmax(x, p4f_min_norm_pos); /* cut off denormalized stuff */ @@ -241,7 +241,7 @@ Packet2d pexp(const Packet2d& _x) emm0 = _mm_add_epi32(emm0, p4i_1023_0); emm0 = _mm_slli_epi32(emm0, 20); emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3)); - return pmul(x, Packet2d(_mm_castsi128_pd(emm0))); + return pmax(pmul(x, Packet2d(_mm_castsi128_pd(emm0))), _x); } /* evaluation of 4 sines at onces, using SSE2 intrinsics. diff --git a/test/packetmath.cpp b/test/packetmath.cpp index a4166d868..ee0502f69 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -311,8 +311,22 @@ template void packetmath_real() } if(internal::random(0,1)<0.1) data1[internal::random(0, PacketSize)] = 0; - CHECK_CWISE1_IF(internal::packet_traits::HasLog, std::log, internal::plog); CHECK_CWISE1_IF(internal::packet_traits::HasSqrt, std::sqrt, internal::psqrt); + CHECK_CWISE1_IF(internal::packet_traits::HasLog, std::log, internal::plog); + { + data1[0] = std::numeric_limits::quiet_NaN(); + packet_helper::HasLog,Packet> h; + h.store(data2, internal::plog(h.load(data1))); + VERIFY(isNaN(data2[0])); + data1[0] = -1.0f; + h.store(data2, internal::plog(h.load(data1))); + VERIFY(isNaN(data2[0])); +#if !EIGEN_FAST_MATH + h.store(data2, internal::psqrt(h.load(data1))); + VERIFY(isNaN(data2[0])); + VERIFY(isNaN(data2[1])); +#endif + } } template void packetmath_notcomplex() From 973e6a035f0207e47bf8face584f002ff169ced1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 20 Oct 2014 14:07:08 +0200 Subject: [PATCH 0920/1103] bug #718: Introduce a compilation error when using the wrong InnerIterator type with a SparseVector --- Eigen/src/SparseCore/SparseVector.h | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index dfbb2be2e..8b696a476 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -386,6 +386,11 @@ class SparseVector::InnerIterator const internal::CompressedStorage& m_data; Index m_id; const Index m_end; + private: + // If you get here, then you're not using the right InnerIterator type, e.g.: + // SparseMatrix A; + // SparseMatrix::InnerIterator it(A,0); + template InnerIterator(const SparseMatrixBase&,Index outer=0); }; template From c12b7896d0ddf95e8f060c93a5cf2fcfa1e6865c Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Mon, 20 Oct 2014 14:23:11 +0200 Subject: [PATCH 0921/1103] bug #766: Check minimum CUDA version --- test/CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 4132e6481..63ed2c7a4 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -306,7 +306,7 @@ endif() option(EIGEN_TEST_NVCC "Enable NVCC support in unit tests" OFF) if(EIGEN_TEST_NVCC) -find_package(CUDA) +find_package(CUDA 5.0) if(CUDA_FOUND) set(CUDA_PROPAGATE_HOST_FLAGS OFF) From fe57b2f963da832d14f4d7b6d4a9554ceef26e36 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 20 Oct 2014 15:55:32 +0200 Subject: [PATCH 0922/1103] bug #701: workaround (min) and (max) blocking ADL by introducing numext::mini and numext::maxi internal functions and a EIGEN_NOT_A_MACRO macro. --- Eigen/src/Cholesky/LDLT.h | 3 +-- Eigen/src/Core/Diagonal.h | 5 ++--- Eigen/src/Core/Fuzzy.h | 3 +-- Eigen/src/Core/GenericPacketMath.h | 4 ++-- Eigen/src/Core/MathFunctions.h | 16 ++++++++++++++++ Eigen/src/Core/StableNorm.h | 10 +++------- Eigen/src/Core/functors/BinaryFunctors.h | 6 ++---- Eigen/src/Core/util/Macros.h | 5 +++++ Eigen/src/Eigenvalues/EigenSolver.h | 6 ++---- Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h | 3 +-- Eigen/src/Geometry/Quaternion.h | 3 +-- Eigen/src/SVD/JacobiSVD.h | 3 +-- Eigen/src/SparseCore/SparseFuzzy.h | 3 +-- Eigen/src/SparseQR/SparseQR.h | 3 +-- unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h | 1 - unsupported/Eigen/src/BDCSVD/BDCSVD.h | 8 +++----- .../src/IterativeSolvers/IncompleteCholesky.h | 3 +-- 17 files changed, 43 insertions(+), 42 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index dfc473df1..5acbf4651 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -488,11 +488,10 @@ void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) cons // dst = D^-1 (L^-1 P b) // more precisely, use pseudo-inverse of D (see bug 241) using std::abs; - EIGEN_USING_STD_MATH(max); const typename Diagonal::RealReturnType vecD(vectorD()); // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon // as motivated by LAPACK's xGELSS: - // RealScalar tolerance = (max)(vectorD.array().abs().maxCoeff() *NumTraits::epsilon(),RealScalar(1) / NumTraits::highest()); + // RealScalar tolerance = numext::maxi(vectorD.array().abs().maxCoeff() *NumTraits::epsilon(),RealScalar(1) / NumTraits::highest()); // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest // diagonal element is not well justified and to numerical issues in some cases. // Moreover, Lapack's xSYTRS routines use 0 for the tolerance. diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index 6ffc0c762..26a58d664 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -77,9 +77,8 @@ template class Diagonal EIGEN_DEVICE_FUNC inline Index rows() const { - EIGEN_USING_STD_MATH(min); - return m_index.value()<0 ? (min)(Index(m_matrix.cols()),Index(m_matrix.rows()+m_index.value())) - : (min)(Index(m_matrix.rows()),Index(m_matrix.cols()-m_index.value())); + return m_index.value()<0 ? numext::mini(Index(m_matrix.cols()),Index(m_matrix.rows()+m_index.value())) + : numext::mini(Index(m_matrix.rows()),Index(m_matrix.cols()-m_index.value())); } EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/Fuzzy.h b/Eigen/src/Core/Fuzzy.h index 8cd069a0d..3e403a09d 100644 --- a/Eigen/src/Core/Fuzzy.h +++ b/Eigen/src/Core/Fuzzy.h @@ -22,10 +22,9 @@ struct isApprox_selector EIGEN_DEVICE_FUNC static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec) { - EIGEN_USING_STD_MATH(min); typename internal::nested_eval::type nested(x); typename internal::nested_eval::type otherNested(y); - return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * (min)(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum()); + return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * numext::mini(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum()); } }; diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index a1fcb82fb..065ccf7ac 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -126,12 +126,12 @@ pdiv(const Packet& a, /** \internal \returns the min of \a a and \a b (coeff-wise) */ template EIGEN_DEVICE_FUNC inline Packet pmin(const Packet& a, - const Packet& b) { EIGEN_USING_STD_MATH(min); return (min)(a, b); } + const Packet& b) { return numext::mini(a, b); } /** \internal \returns the max of \a a and \a b (coeff-wise) */ template EIGEN_DEVICE_FUNC inline Packet pmax(const Packet& a, - const Packet& b) { EIGEN_USING_STD_MATH(max); return (max)(a, b); } + const Packet& b) { return numext::maxi(a, b); } /** \internal \returns the absolute value of \a a */ template EIGEN_DEVICE_FUNC inline Packet diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 73859b0ee..071c234c2 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -591,6 +591,22 @@ inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random() ****************************************************************************/ namespace numext { + +template +EIGEN_DEVICE_FUNC +inline T mini(const T& x, const T& y) +{ + using std::min; + return min EIGEN_NOT_A_MACRO (x,y); +} + +template +EIGEN_DEVICE_FUNC +inline T maxi(const T& x, const T& y) +{ + using std::max; + return max EIGEN_NOT_A_MACRO (x,y); +} template EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/StableNorm.h b/Eigen/src/Core/StableNorm.h index 64d43e1b1..0b7e39827 100644 --- a/Eigen/src/Core/StableNorm.h +++ b/Eigen/src/Core/StableNorm.h @@ -17,7 +17,6 @@ namespace internal { template inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale) { - using std::max; Scalar maxCoeff = bl.cwiseAbs().maxCoeff(); if(maxCoeff>scale) @@ -58,8 +57,6 @@ blueNorm_impl(const EigenBase& _vec) typedef typename Derived::RealScalar RealScalar; typedef typename Derived::Index Index; using std::pow; - EIGEN_USING_STD_MATH(min); - EIGEN_USING_STD_MATH(max); using std::sqrt; using std::abs; const Derived& vec(_vec.derived()); @@ -136,8 +133,8 @@ blueNorm_impl(const EigenBase& _vec) } else return sqrt(amed); - asml = (min)(abig, amed); - abig = (max)(abig, amed); + asml = numext::mini(abig, amed); + abig = numext::maxi(abig, amed); if(asml <= abig*relerr) return abig; else @@ -160,7 +157,6 @@ template inline typename NumTraits::Scalar>::Real MatrixBase::stableNorm() const { - EIGEN_USING_STD_MATH(min); using std::sqrt; const Index blockSize = 4096; RealScalar scale(0); @@ -174,7 +170,7 @@ MatrixBase::stableNorm() const if (bi>0) internal::stable_norm_kernel(this->head(bi), ssq, scale, invScale); for (; bisegment(bi,(min)(blockSize, n - bi)).template forceAlignedAccessIf(), ssq, scale, invScale); + internal::stable_norm_kernel(this->segment(bi,numext::mini(blockSize, n - bi)).template forceAlignedAccessIf(), ssq, scale, invScale); return scale * sqrt(ssq); } diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h index 157d075a7..9c96181c7 100644 --- a/Eigen/src/Core/functors/BinaryFunctors.h +++ b/Eigen/src/Core/functors/BinaryFunctors.h @@ -115,7 +115,7 @@ struct functor_traits > { */ template struct scalar_min_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(min); return (min)(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return numext::mini(a, b); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pmin(a,b); } @@ -138,7 +138,7 @@ struct functor_traits > { */ template struct scalar_max_op { EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { EIGEN_USING_STD_MATH(max); return (max)(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a, const Scalar& b) const { return numext::maxi(a, b); } template EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const { return internal::pmax(a,b); } @@ -164,8 +164,6 @@ template struct scalar_hypot_op { // typedef typename NumTraits::Real result_type; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const { - EIGEN_USING_STD_MATH(max); - EIGEN_USING_STD_MATH(min); using std::sqrt; Scalar p, qp; if(_x>_y) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index f9b908e22..c09854951 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -86,6 +86,11 @@ #define EIGEN_ALIGN 0 #endif + +// This macro can be used to prevent from macro expansion, e.g.: +// std::max EIGNE_NOT_A_MACRO(a,b) +#define EIGEN_NOT_A_MACRO + // EIGEN_ALIGN_STATICALLY is the true test whether we want to align arrays on the stack or not. It takes into account both the user choice to explicitly disable // alignment (EIGEN_DONT_ALIGN_STATICALLY) and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT). Henceforth, only EIGEN_ALIGN_STATICALLY should be used. #if EIGEN_ARCH_WANTS_STACK_ALIGNMENT && !defined(EIGEN_DONT_ALIGN_STATICALLY) diff --git a/Eigen/src/Eigenvalues/EigenSolver.h b/Eigen/src/Eigenvalues/EigenSolver.h index 8a83b85bb..9372021ff 100644 --- a/Eigen/src/Eigenvalues/EigenSolver.h +++ b/Eigen/src/Eigenvalues/EigenSolver.h @@ -368,7 +368,6 @@ EigenSolver::compute(const MatrixType& matrix, bool computeEigenvect { using std::sqrt; using std::abs; - using std::max; using numext::isfinite; eigen_assert(matrix.cols() == matrix.rows()); @@ -409,7 +408,7 @@ EigenSolver::compute(const MatrixType& matrix, bool computeEigenvect { Scalar t0 = m_matT.coeff(i+1, i); Scalar t1 = m_matT.coeff(i, i+1); - Scalar maxval = (max)(abs(p),(max)(abs(t0),abs(t1))); + Scalar maxval = numext::maxi(abs(p),numext::maxi(abs(t0),abs(t1))); t0 /= maxval; t1 /= maxval; Scalar p0 = p/maxval; @@ -600,8 +599,7 @@ void EigenSolver::doComputeEigenvectors() } // Overflow control - EIGEN_USING_STD_MATH(max); - Scalar t = (max)(abs(m_matT.coeff(i,n-1)),abs(m_matT.coeff(i,n))); + Scalar t = numext::maxi(abs(m_matT.coeff(i,n-1)),abs(m_matT.coeff(i,n))); if ((eps * t) * t > Scalar(1)) m_matT.block(i, n-1, size-i, 2) /= t; diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index 1dd2ab45b..54f60b197 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -732,7 +732,6 @@ struct direct_selfadjoint_eigenvalues EIGEN_DEVICE_FUNC static inline void run(SolverType& solver, const MatrixType& mat, int options) { - EIGEN_USING_STD_MATH(max) EIGEN_USING_STD_MATH(sqrt); eigen_assert(mat.cols() == 2 && mat.cols() == mat.rows()); @@ -746,7 +745,7 @@ struct direct_selfadjoint_eigenvalues // map the matrix coefficients to [-1:1] to avoid over- and underflow. Scalar scale = mat.cwiseAbs().maxCoeff(); - scale = (max)(scale,Scalar(1)); + scale = numext::maxi(scale,Scalar(1)); MatrixType scaledMat = mat / scale; // Compute the eigenvalues diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index 216e5b12f..508eba767 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -571,7 +571,6 @@ template template inline Derived& QuaternionBase::setFromTwoVectors(const MatrixBase& a, const MatrixBase& b) { - EIGEN_USING_STD_MATH(max); using std::sqrt; Vector3 v0 = a.normalized(); Vector3 v1 = b.normalized(); @@ -587,7 +586,7 @@ inline Derived& QuaternionBase::setFromTwoVectors(const MatrixBase::dummy_precision()) { - c = (max)(c,Scalar(-1)); + c = numext::maxi(c,Scalar(-1)); Matrix m; m << v0.transpose(), v1.transpose(); JacobiSVD > svd(m, ComputeFullV); Vector3 axis = svd.matrixV().col(2); diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 7a8aa8d3f..0f7e5b8fe 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -723,8 +723,7 @@ JacobiSVD::compute(const MatrixType& matrix, unsig // if this 2x2 sub-matrix is not diagonal already... // notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't // keep us iterating forever. Similarly, small denormal numbers are considered zero. - EIGEN_USING_STD_MATH(max); - RealScalar threshold = (max)(considerAsZero, precision * (max)(abs(m_workMatrix.coeff(p,p)), + RealScalar threshold = numext::maxi(considerAsZero, precision * numext::maxi(abs(m_workMatrix.coeff(p,p)), abs(m_workMatrix.coeff(q,q)))); // We compare both values to threshold instead of calling max to be robust to NaN (See bug 791) if(abs(m_workMatrix.coeff(p,q))>threshold || abs(m_workMatrix.coeff(q,p)) > threshold) diff --git a/Eigen/src/SparseCore/SparseFuzzy.h b/Eigen/src/SparseCore/SparseFuzzy.h index a76c1a5e0..7d47eb94d 100644 --- a/Eigen/src/SparseCore/SparseFuzzy.h +++ b/Eigen/src/SparseCore/SparseFuzzy.h @@ -16,13 +16,12 @@ template template bool SparseMatrixBase::isApprox(const SparseMatrixBase& other, const RealScalar &prec) const { - using std::min; const typename internal::nested_eval::type actualA(derived()); typename internal::conditional::type, const PlainObject>::type actualB(other.derived()); - return (actualA - actualB).squaredNorm() <= prec * prec * (min)(actualA.squaredNorm(), actualB.squaredNorm()); + return (actualA - actualB).squaredNorm() <= prec * prec * numext::mini(actualA.squaredNorm(), actualB.squaredNorm()); } } // end namespace Eigen diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 879ac553d..133211488 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -325,7 +325,6 @@ template void SparseQR::factorize(const MatrixType& mat) { using std::abs; - using std::max; eigen_assert(m_analysisIsok && "analyzePattern() should be called before this step"); Index m = mat.rows(); @@ -377,7 +376,7 @@ void SparseQR::factorize(const MatrixType& mat) if(m_useDefaultThreshold) { RealScalar max2Norm = 0.0; - for (int j = 0; j < n; j++) max2Norm = (max)(max2Norm, m_pmat.col(j).norm()); + for (int j = 0; j < n; j++) max2Norm = numext::maxi(max2Norm, m_pmat.col(j).norm()); if(max2Norm==RealScalar(0)) max2Norm = RealScalar(1); pivotThreshold = 20 * (m + n) * max2Norm * NumTraits::epsilon(); diff --git a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h index 590797973..8336c2644 100644 --- a/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h +++ b/unsupported/Eigen/src/AutoDiff/AutoDiffScalar.h @@ -593,7 +593,6 @@ inline const AutoDiffScalar::Scalar,D atan2(const AutoDiffScalar& a, const AutoDiffScalar& b) { using std::atan2; - using std::max; typedef typename internal::traits::Scalar Scalar; typedef AutoDiffScalar > PlainADS; PlainADS ret; diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index e2551cf88..ac71b0aa8 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -649,7 +649,6 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia { using std::abs; using std::swap; - using std::max; Index n = col0.size(); Index actual_n = n; @@ -728,7 +727,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia // rational interpolation: fit a function of the form a / mu + b through the two previous // iterates and use its zero to compute the next iterate bool useBisection = fPrev*fCur>0; - while (fCur!=0 && abs(muCur - muPrev) > 8 * NumTraits::epsilon() * (max)(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) + while (fCur!=0 && abs(muCur - muPrev) > 8 * NumTraits::epsilon() * numext::maxi(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits::epsilon() && !useBisection) { ++m_numIters; @@ -779,7 +778,7 @@ void BDCSVD::computeSingVals(const ArrayXr& col0, const ArrayXr& dia #endif eigen_internal_assert(fLeft * fRight < 0); - while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * (max)(abs(leftShifted), abs(rightShifted))) + while (rightShifted - leftShifted > 2 * NumTraits::epsilon() * numext::maxi(abs(leftShifted), abs(rightShifted))) { RealScalar midShifted = (leftShifted + rightShifted) / 2; RealScalar fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift); @@ -981,7 +980,6 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index { using std::sqrt; using std::abs; - using std::max; const Index length = lastCol + 1 - firstCol; Block col0(m_computed, firstCol+shift, firstCol+shift, length, 1); @@ -990,7 +988,7 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index RealScalar maxDiag = diag.tail((std::max)(Index(1),length-1)).cwiseAbs().maxCoeff(); RealScalar epsilon_strict = NumTraits::epsilon() * maxDiag; - RealScalar epsilon_coarse = 8 * NumTraits::epsilon() * (max)(col0.cwiseAbs().maxCoeff(), maxDiag); + RealScalar epsilon_coarse = 8 * NumTraits::epsilon() * numext::maxi(col0.cwiseAbs().maxCoeff(), maxDiag); #ifdef EIGEN_BDCSVD_SANITY_CHECKS assert(m_naiveU.allFinite()); diff --git a/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h b/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h index dd43de6b3..35cfa315d 100644 --- a/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h +++ b/unsupported/Eigen/src/IterativeSolvers/IncompleteCholesky.h @@ -126,7 +126,6 @@ template void IncompleteCholesky::factorize(const _MatrixType& mat) { using std::sqrt; - using std::min; eigen_assert(m_analysisIsOk && "analyzePattern() should be called first"); // Dropping strategies : Keep only the p largest elements per column, where p is the number of elements in the column of the original matrix. Other strategies will be added @@ -160,7 +159,7 @@ void IncompleteCholesky::factorize(const _MatrixType for (int j = 0; j < n; j++){ for (int k = colPtr[j]; k < colPtr[j+1]; k++) vals[k] /= (m_scal(j) * m_scal(rowIdx[k])); - mindiag = (min)(vals[colPtr[j]], mindiag); + mindiag = numext::mini(vals[colPtr[j]], mindiag); } if(mindiag < Scalar(0.)) m_shift = m_shift - mindiag; From a303b6a733509db8c533e94e7a903ad7d28fef8d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 20 Oct 2014 16:46:47 +0200 Subject: [PATCH 0923/1103] bug #670: add unit test for mapped input in sparse solver. --- test/sparse_solver.h | 28 ++++++++++++++++++++++++++-- 1 file changed, 26 insertions(+), 2 deletions(-) diff --git a/test/sparse_solver.h b/test/sparse_solver.h index b10eaebcc..ee350d561 100644 --- a/test/sparse_solver.h +++ b/test/sparse_solver.h @@ -15,6 +15,7 @@ void check_sparse_solving(Solver& solver, const typename Solver::MatrixType& A, { typedef typename Solver::MatrixType Mat; typedef typename Mat::Scalar Scalar; + typedef typename Mat::Index Index; DenseRhs refX = dA.lu().solve(db); { @@ -35,8 +36,8 @@ void check_sparse_solving(Solver& solver, const typename Solver::MatrixType& A, return; } VERIFY(oldb.isApprox(b) && "sparse solver testing: the rhs should not be modified!"); - VERIFY(x.isApprox(refX,test_precision())); + x.setZero(); // test the analyze/factorize API solver.analyzePattern(A); @@ -54,8 +55,31 @@ void check_sparse_solving(Solver& solver, const typename Solver::MatrixType& A, return; } VERIFY(oldb.isApprox(b) && "sparse solver testing: the rhs should not be modified!"); - VERIFY(x.isApprox(refX,test_precision())); + + + x.setZero(); + // test with Map + MappedSparseMatrix Am(A.rows(), A.cols(), A.nonZeros(), const_cast(A.outerIndexPtr()), const_cast(A.innerIndexPtr()), const_cast(A.valuePtr())); + solver.compute(Am); + if (solver.info() != Success) + { + std::cerr << "sparse solver testing: factorization failed (check_sparse_solving)\n"; + exit(0); + return; + } + DenseRhs dx(refX); + dx.setZero(); + Map xm(dx.data(), dx.rows(), dx.cols()); + Map bm(db.data(), db.rows(), db.cols()); + xm = solver.solve(bm); + if (solver.info() != Success) + { + std::cerr << "sparse solver testing: solving failed\n"; + return; + } + VERIFY(oldb.isApprox(bm) && "sparse solver testing: the rhs should not be modified!"); + VERIFY(xm.isApprox(refX,test_precision())); } // test dense Block as the result and rhs: From 87524922dcb1e23eccccdf8ff6df2a1b9bee6308 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Tue, 21 Oct 2014 18:08:50 +0000 Subject: [PATCH 0924/1103] check for __ARM_NEON instead as it's defined in arm64 as well --- Eigen/Core | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/Core b/Eigen/Core index adab50b4a..2174c6bf8 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -180,7 +180,7 @@ #undef bool #undef vector #undef pixel - #elif defined __ARM_NEON__ + #elif defined __ARM_NEON #define EIGEN_VECTORIZE #define EIGEN_VECTORIZE_NEON #include From 0f65f2762de6f2da4d88d19175560a1a2788ebd7 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Tue, 21 Oct 2014 18:10:01 +0000 Subject: [PATCH 0925/1103] add EIGEN_TEST_NEON64, but it's a dummy, AArch64 implies NEON support so extra CXXFLAGS are needed --- CMakeLists.txt | 8 ++++++++ cmake/EigenTesting.cmake | 8 ++++++++ 2 files changed, 16 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index ea42cc8db..05d92babe 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -231,6 +231,14 @@ if(NOT MSVC) message(STATUS "Enabling NEON in tests/examples") endif() + option(EIGEN_TEST_NEON64 "Enable/Disable Neon in tests/examples" OFF) + if(EIGEN_TEST_NEON64) + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}") + message(STATUS "Enabling NEON in tests/examples") + endif() + + + check_cxx_compiler_flag("-fopenmp" COMPILER_SUPPORT_OPENMP) if(COMPILER_SUPPORT_OPENMP) option(EIGEN_TEST_OPENMP "Enable/Disable OpenMP in tests/examples" OFF) diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index 3ed002a87..3eb4e4c07 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -294,6 +294,12 @@ macro(ei_testing_print_summary) message(STATUS "ARM NEON: Using architecture defaults") endif() + if(EIGEN_TEST_NEON64) + message(STATUS "ARMv8 NEON: ON") + else() + message(STATUS "ARMv8 NEON: Using architecture defaults") + endif() + endif() # vectorization / alignment options message(STATUS "\n${EIGEN_TESTING_SUMMARY}") @@ -424,6 +430,8 @@ macro(ei_get_cxxflags VAR) ei_is_64bit_env(IS_64BIT_ENV) if(EIGEN_TEST_NEON) set(${VAR} NEON) + elseif(EIGEN_TEST_NEON64) + set(${VAR} NEON) elseif(EIGEN_TEST_VSX) set(${VAR} VSX) elseif(EIGEN_TEST_ALTIVEC) From b50861939217d2c6c92362e2eea59866b5386b33 Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Tue, 21 Oct 2014 18:10:33 +0000 Subject: [PATCH 0926/1103] working 64-bit support in PacketMath.h, Complex.h needed --- Eigen/src/Core/arch/NEON/PacketMath.h | 174 +++++++++++++++++++++++++- 1 file changed, 173 insertions(+), 1 deletion(-) diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 0504c095c..e3a13d93a 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -20,10 +20,18 @@ namespace internal { #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 #endif +#ifndef EIGEN_HAS_FUSED_MADD +#define EIGEN_HAS_FUSED_MADD 1 +#endif + +#ifndef EIGEN_HAS_FUSE_CJMADD +#define EIGEN_HAS_FUSE_CJMADD 1 +#endif + // FIXME NEON has 16 quad registers, but since the current register allocator // is so bad, it is much better to reduce it to 8 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS -#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8 +#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 #endif typedef float32x4_t Packet4f; @@ -71,6 +79,7 @@ template<> struct packet_traits : default_packet_traits Vectorizable = 1, AlignedOnScalar = 1, size = 4, + HasHalfPacket=0, HasDiv = 1, // FIXME check the Has* @@ -136,6 +145,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul(const Packet4i& a, const template<> EIGEN_STRONG_INLINE Packet4f pdiv(const Packet4f& a, const Packet4f& b) { +#ifndef __aarch64__ Packet4f inv, restep, div; // NEON does not offer a divide instruction, we have to do a reciprocal approximation @@ -154,7 +164,11 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv(const Packet4f& a, const div = vmulq_f32(a, inv); return div; +#else + return vdivq_f32(a,b); +#endif } + template<> EIGEN_STRONG_INLINE Packet4i pdiv(const Packet4i& /*a*/, const Packet4i& /*b*/) { eigen_assert(false && "packet integer division are not supported by NEON"); return pset1(0); @@ -472,6 +486,164 @@ ptranspose(PacketBlock& kernel) { kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1])); } +//---------- double ---------- +#ifdef __aarch64__ + +typedef float64x2_t Packet2d; + +template<> struct packet_traits : default_packet_traits +{ + typedef Packet2d type; + typedef Packet2d half; + enum { + Vectorizable = 1, + AlignedOnScalar = 1, + size = 2, + HasHalfPacket=0, + + HasDiv = 1, + // FIXME check the Has* + HasSin = 0, + HasCos = 0, + HasLog = 0, + HasExp = 0, + HasSqrt = 0 + }; +}; + +template<> struct unpacket_traits { typedef double type; enum {size=2}; typedef Packet2d half; }; + +template<> EIGEN_STRONG_INLINE Packet2d pset1(const double& from) { return vdupq_n_f64(from); } + +template<> EIGEN_STRONG_INLINE Packet2d plset(const double& a) +{ + Packet2d countdown = EIGEN_INIT_NEON_PACKET2(0, 1); + return vaddq_f64(pset1(a), countdown); +} +template<> EIGEN_STRONG_INLINE Packet2d padd(const Packet2d& a, const Packet2d& b) { return vaddq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d psub(const Packet2d& a, const Packet2d& b) { return vsubq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return vnegq_f64(a); } + +template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } + +template<> EIGEN_STRONG_INLINE Packet2d pmul(const Packet2d& a, const Packet2d& b) { return vmulq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pdiv(const Packet2d& a, const Packet2d& b) { return vdivq_f64(a,b); } + +// for some weird raisons, it has to be overloaded for packet of integers +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vmlaq_f64(c,a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pmin(const Packet2d& a, const Packet2d& b) { return vminq_f64(a,b); } + +template<> EIGEN_STRONG_INLINE Packet2d pmax(const Packet2d& a, const Packet2d& b) { return vmaxq_f64(a,b); } + +// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics +template<> EIGEN_STRONG_INLINE Packet2d pand(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d por(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pxor(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pandnot(const Packet2d& a, const Packet2d& b) +{ + return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b))); +} + +template<> EIGEN_STRONG_INLINE Packet2d pload(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f64(from); } + +template<> EIGEN_STRONG_INLINE Packet2d ploadu(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f64(from); } + +template<> EIGEN_STRONG_INLINE Packet2d ploaddup(const double* from) +{ + return vld1q_dup_f64(from); +} +template<> EIGEN_STRONG_INLINE void pstore(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f64(to, from); } + +template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to, from); } + +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, DenseIndex stride) +{ + Packet2d res; + res = vsetq_lane_f64(from[0*stride], res, 0); + res = vsetq_lane_f64(from[1*stride], res, 1); + return res; +} +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, DenseIndex stride) +{ + to[stride*0] = vgetq_lane_f64(from, 0); + to[stride*1] = vgetq_lane_f64(from, 1); +} +template<> EIGEN_STRONG_INLINE void prefetch(const double* addr) { EIGEN_ARM_PREFETCH(addr); } + +// FIXME only store the 2 first elements ? +template<> EIGEN_STRONG_INLINE double pfirst(const Packet2d& a) { return vgetq_lane_f64(a, 0); } + +template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return vcombine_f64(vget_high_f64(a), vget_low_f64(a)); } + +template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vabsq_f64(a); } + +template<> EIGEN_STRONG_INLINE double predux(const Packet2d& a) { return vget_low_f64(a) + vget_high_f64(a); } + +template<> EIGEN_STRONG_INLINE Packet2d preduxp(const Packet2d* vecs) +{ + float64x2_t trn1, trn2; + + // NEON zip performs interleaving of the supplied vectors. + // We perform two interleaves in a row to acquire the transposed vector + trn1 = vzip1q_f64(vecs[0], vecs[1]); + trn2 = vzip2q_f64(vecs[0], vecs[1]); + + // Do the addition of the resulting vectors + return vaddq_f64(trn1, trn2); +} +// Other reduction functions: +// mul +template<> EIGEN_STRONG_INLINE double predux_mul(const Packet2d& a) { return vget_low_f64(a) * vget_high_f64(a); } + +// min +template<> EIGEN_STRONG_INLINE double predux_min(const Packet2d& a) { return vgetq_lane_f64(vpminq_f64(a, a), 0); } + +// max +template<> EIGEN_STRONG_INLINE double predux_max(const Packet2d& a) { return vgetq_lane_f64(vpmaxq_f64(a, a), 0); } + +// this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors, +// see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074 +#define PALIGN_NEON(Offset,Type,Command) \ +template<>\ +struct palign_impl\ +{\ + EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\ + {\ + if (Offset!=0)\ + first = Command(first, second, Offset);\ + }\ +};\ + +PALIGN_NEON(0,Packet2d,vextq_f64) +PALIGN_NEON(1,Packet2d,vextq_f64) +#undef PALIGN_NEON + +EIGEN_DEVICE_FUNC inline void +ptranspose(PacketBlock& kernel) { + float64x2_t trn1 = vzip1q_f64(kernel.packet[0], kernel.packet[1]); + float64x2_t trn2 = vzip2q_f64(kernel.packet[0], kernel.packet[1]); + + kernel.packet[0] = trn1; + kernel.packet[1] = trn2; +} +#endif // __aarch64__ + } // end namespace internal } // end namespace Eigen From cf09c5f687c171d2e8b760358a4d48da1c010ce0 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 21 Oct 2014 20:40:09 +0200 Subject: [PATCH 0927/1103] Prevent CUDA `calling a __host__ function from a __host__ __device__ function is not allowed` error. --- Eigen/src/Core/MathFunctions.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h index 071c234c2..4c5fc1cae 100644 --- a/Eigen/src/Core/MathFunctions.h +++ b/Eigen/src/Core/MathFunctions.h @@ -596,7 +596,7 @@ template EIGEN_DEVICE_FUNC inline T mini(const T& x, const T& y) { - using std::min; + EIGEN_USING_STD_MATH(min); return min EIGEN_NOT_A_MACRO (x,y); } @@ -604,7 +604,7 @@ template EIGEN_DEVICE_FUNC inline T maxi(const T& x, const T& y) { - using std::max; + EIGEN_USING_STD_MATH(max); return max EIGEN_NOT_A_MACRO (x,y); } From fae4fd7a26ee31fbf1eaa5d3581916ccee3c004f Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Wed, 22 Oct 2014 07:39:49 +0000 Subject: [PATCH 0928/1103] Added ARMv8 support --- Eigen/src/Core/arch/NEON/Complex.h | 194 +++++++++++++++++++++++++- Eigen/src/Core/arch/NEON/PacketMath.h | 16 ++- 2 files changed, 201 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 42e7733d7..57de400e5 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -33,6 +33,7 @@ template<> struct packet_traits > : default_packet_traits Vectorizable = 1, AlignedOnScalar = 1, size = 2, + HasHalfPacket = 0, HasAdd = 1, HasSub = 1, @@ -88,7 +89,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, con template<> EIGEN_STRONG_INLINE Packet2cf pand (const Packet2cf& a, const Packet2cf& b) { - return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); + return Packet2cf(vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); } template<> EIGEN_STRONG_INLINE Packet2cf por (const Packet2cf& a, const Packet2cf& b) { @@ -252,7 +253,7 @@ template<> struct conj_helper template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, const Packet2cf& b) { - // TODO optimize it for AltiVec + // TODO optimize it for NEON Packet2cf res = conj_helper().pmul(a,b); Packet4f s, rev_s; @@ -265,11 +266,198 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv(const Packet2cf& a, con EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock& kernel) { - float32x4_t tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v)); + Packet4f tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v)); kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v)); kernel.packet[1].v = tmp; } +//---------- double ---------- +#ifdef __aarch64__ + +static uint64x2_t p2ul_CONJ_XOR = EIGEN_INIT_NEON_PACKET2(0x0, 0x8000000000000000); + +struct Packet1cd +{ + EIGEN_STRONG_INLINE Packet1cd() {} + EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} + Packet2d v; +}; + +template<> struct packet_traits > : default_packet_traits +{ + typedef Packet1cd type; + typedef Packet1cd half; + enum { + Vectorizable = 1, + AlignedOnScalar = 0, + size = 1, + HasHalfPacket = 0, + + HasAdd = 1, + HasSub = 1, + HasMul = 1, + HasDiv = 1, + HasNegate = 1, + HasAbs = 0, + HasAbs2 = 0, + HasMin = 0, + HasMax = 0, + HasSetLinear = 0 + }; +}; + +template<> struct unpacket_traits { typedef std::complex type; enum {size=1}; typedef Packet1cd half; }; + +template<> EIGEN_STRONG_INLINE Packet1cd pload(const std::complex* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload((const double*)from)); } +template<> EIGEN_STRONG_INLINE Packet1cd ploadu(const std::complex* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu((const double*)from)); } + +template<> EIGEN_STRONG_INLINE Packet1cd pset1(const std::complex& from) +{ /* here we really have to use unaligned loads :( */ return ploadu(&from); } + +template<> EIGEN_STRONG_INLINE Packet1cd padd(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(padd(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd psub(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(psub(a.v,b.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(a.v)); } +template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v), p2ul_CONJ_XOR))); } + +template<> EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) +{ + Packet2d v1, v2; + + // Get the real values of a + v1 = vdupq_lane_f64(vget_low_f64(a.v), 0); + // Get the real values of a + v2 = vdupq_lane_f64(vget_high_f64(a.v), 1); + // Multiply the real a with b + v1 = vmulq_f64(v1, b.v); + // Multiply the imag a with b + v2 = vmulq_f64(v2, b.v); + // Conjugate v2 + v2 = vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(v2), p2ul_CONJ_XOR)); + // Swap real/imag elements in v2. + v2 = preverse(v2); + // Add and return the result + return Packet1cd(vaddq_f64(v1, v2)); +} + +template<> EIGEN_STRONG_INLINE Packet1cd pand (const Packet1cd& a, const Packet1cd& b) +{ + return Packet1cd(vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet1cd por (const Packet1cd& a, const Packet1cd& b) +{ + return Packet1cd(vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet1cd pxor (const Packet1cd& a, const Packet1cd& b) +{ + return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); +} +template<> EIGEN_STRONG_INLINE Packet1cd pandnot(const Packet1cd& a, const Packet1cd& b) +{ + return Packet1cd(vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); +} + +template<> EIGEN_STRONG_INLINE Packet1cd ploaddup(const std::complex* from) { return pset1(*from); } + +template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } +template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } + +template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((double *)addr); } + +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, DenseIndex stride) +{ + Packet2d res; + res = vsetq_lane_f64(std::real(from[0*stride]), res, 0); + res = vsetq_lane_f64(std::imag(from[0*stride]), res, 1); + return Packet1cd(res); +} + +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet1cd>(std::complex* to, const Packet1cd& from, DenseIndex stride) +{ + to[stride*0] = std::complex(vgetq_lane_f64(from.v, 0), vgetq_lane_f64(from.v, 1)); +} + + +template<> EIGEN_STRONG_INLINE std::complex pfirst(const Packet1cd& a) +{ + std::complex EIGEN_ALIGN16 res; + pstore >(&res, a); + + return res; +} + +template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } + +template<> EIGEN_STRONG_INLINE std::complex predux(const Packet1cd& a) { return pfirst(a); } + +template<> EIGEN_STRONG_INLINE Packet1cd preduxp(const Packet1cd* vecs) { return vecs[0]; } + +template<> EIGEN_STRONG_INLINE std::complex predux_mul(const Packet1cd& a) { return pfirst(a); } + +template +struct palign_impl +{ + static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) + { + // FIXME is it sure we never have to align a Packet1cd? + // Even though a std::complex has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(a, pconj(b)); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return internal::pmul(pconj(a), b); + } +}; + +template<> struct conj_helper +{ + EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const + { return padd(pmul(x,y),c); } + + EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const + { + return pconj(internal::pmul(a, b)); + } +}; + +template<> EIGEN_STRONG_INLINE Packet1cd pdiv(const Packet1cd& a, const Packet1cd& b) +{ + // TODO optimize it for NEON + Packet1cd res = conj_helper().pmul(a,b); + Packet2d s = pmul(b.v, b.v); + Packet2d rev_s = preverse(s); + + return Packet1cd(pdiv(res.v, padd(s,rev_s))); +} + +EIGEN_STRONG_INLINE Packet1cd pcplxflip/**/(const Packet1cd& x) +{ + return Packet1cd(preverse(Packet2d(x.v))); +} + +EIGEN_STRONG_INLINE void ptranspose(PacketBlock& kernel) +{ + Packet2d tmp = vcombine_f64(vget_high_f64(kernel.packet[0].v), vget_high_f64(kernel.packet[1].v)); + kernel.packet[0].v = vcombine_f64(vget_low_f64(kernel.packet[0].v), vget_low_f64(kernel.packet[1].v)); + kernel.packet[1].v = tmp; +} +#endif // __aarch64__ } // end namespace internal diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index e3a13d93a..472f7c0fe 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -34,8 +34,10 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 #endif +typedef float32x2_t Packet2f; typedef float32x4_t Packet4f; typedef int32x4_t Packet4i; +typedef int32x2_t Packet2i; typedef uint32x4_t Packet4ui; #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ @@ -74,12 +76,12 @@ typedef uint32x4_t Packet4ui; template<> struct packet_traits : default_packet_traits { typedef Packet4f type; - typedef Packet4f half; + typedef Packet2f half; enum { Vectorizable = 1, AlignedOnScalar = 1, size = 4, - HasHalfPacket=0, + HasHalfPacket=1, HasDiv = 1, // FIXME check the Has* @@ -93,11 +95,12 @@ template<> struct packet_traits : default_packet_traits template<> struct packet_traits : default_packet_traits { typedef Packet4i type; - typedef Packet4i half; + typedef Packet2i half; enum { Vectorizable = 1, AlignedOnScalar = 1, - size=4 + size=4, + HasHalfPacket=1 // FIXME check the Has* }; }; @@ -490,16 +493,17 @@ ptranspose(PacketBlock& kernel) { #ifdef __aarch64__ typedef float64x2_t Packet2d; +typedef float64x1_t Packet1d; template<> struct packet_traits : default_packet_traits { typedef Packet2d type; - typedef Packet2d half; + typedef Packet1d half; enum { Vectorizable = 1, AlignedOnScalar = 1, size = 2, - HasHalfPacket=0, + HasHalfPacket=1, HasDiv = 1, // FIXME check the Has* From 94ed7c81e6c20ea069cfc37adce7b45dad91691c Mon Sep 17 00:00:00 2001 From: Konstantinos Margaritis Date: Wed, 22 Oct 2014 06:15:18 -0400 Subject: [PATCH 0929/1103] Bug #896: Swap order of checking __VSX__/__ALTIVEC__ --- Eigen/Core | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index adab50b4a..748e20f8e 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -162,18 +162,18 @@ #endif #endif } // end extern "C" - #elif defined __ALTIVEC__ + #elif defined __VSX__ #define EIGEN_VECTORIZE - #define EIGEN_VECTORIZE_ALTIVEC + #define EIGEN_VECTORIZE_VSX #include // We need to #undef all these ugly tokens defined in // => use __vector instead of vector #undef bool #undef vector #undef pixel - #elif defined __VSX__ + #elif defined __ALTIVEC__ #define EIGEN_VECTORIZE - #define EIGEN_VECTORIZE_VSX + #define EIGEN_VECTORIZE_ALTIVEC #include // We need to #undef all these ugly tokens defined in // => use __vector instead of vector From 04ffb9956eb75fbab8f4926235ba011ae7a79d39 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 24 Oct 2014 13:18:23 +0200 Subject: [PATCH 0930/1103] Replace TEST_SET_BUT_UNUSED_VARIABLE by already defined EIGEN_UNUSED_VARIABLE --- test/main.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/main.h b/test/main.h index adf4adb31..b3fa68476 100644 --- a/test/main.h +++ b/test/main.h @@ -61,7 +61,7 @@ #endif // shuts down ICC's remark #593: variable "XXX" was set but never used -#define TEST_SET_BUT_UNUSED_VARIABLE(X) X = X + 0; +#define TEST_SET_BUT_UNUSED_VARIABLE(X) EIGEN_UNUSED_VARIABLE(X) // the following file is automatically generated by cmake #include "split_test_helper.h" From 1fa793cb978adace700417a16d2ee54c28f42862 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Fri, 24 Oct 2014 13:19:19 +0200 Subject: [PATCH 0931/1103] Removed weird self assignment. --- unsupported/Eigen/src/BDCSVD/BDCSVD.h | 2 +- unsupported/test/bdcsvd.cpp | 2 -- 2 files changed, 1 insertion(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/unsupported/Eigen/src/BDCSVD/BDCSVD.h index ac71b0aa8..94bc03aac 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/unsupported/Eigen/src/BDCSVD/BDCSVD.h @@ -824,7 +824,7 @@ void BDCSVD::perturbCol0 zhat.setZero(); return; } - Index last = last = perm(m-1); + Index last = perm(m-1); // The offset permits to skip deflated entries while computing zhat for (Index k = 0; k < n; ++k) { diff --git a/unsupported/test/bdcsvd.cpp b/unsupported/test/bdcsvd.cpp index 9e5c29a8c..97d7880bb 100644 --- a/unsupported/test/bdcsvd.cpp +++ b/unsupported/test/bdcsvd.cpp @@ -90,8 +90,6 @@ void test_bdcsvd() CALL_SUBTEST_10(( compare_bdc_jacobi(MatrixXd(r,c)) )); CALL_SUBTEST_8(( bdcsvd(MatrixXcd(r,c)) )); CALL_SUBTEST_8(( compare_bdc_jacobi(MatrixXcd(r,c)) )); - (void) r; - (void) c; // Test on inf/nan matrix CALL_SUBTEST_7( (svd_inf_nan, MatrixXf>()) ); From c42605476755652b84643472d48e2e978c5d6f77 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Benjamin=20Chr=C3=A9tien?= Date: Fri, 24 Oct 2014 15:10:56 +0200 Subject: [PATCH 0932/1103] BDCSVD: fix CMake install (missing separator). --- unsupported/Eigen/src/BDCSVD/CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/BDCSVD/CMakeLists.txt b/unsupported/Eigen/src/BDCSVD/CMakeLists.txt index 73b89ea18..1045512f9 100644 --- a/unsupported/Eigen/src/BDCSVD/CMakeLists.txt +++ b/unsupported/Eigen/src/BDCSVD/CMakeLists.txt @@ -2,5 +2,5 @@ FILE(GLOB Eigen_BDCSVD_SRCS "*.h") INSTALL(FILES ${Eigen_BDCSVD_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}unsupported/Eigen/src/BDCSVD COMPONENT Devel + DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/BDCSVD COMPONENT Devel ) From bd2d330b25f4f57b64ddc125c70dadd42e1e3349 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 28 Oct 2014 13:31:00 +0100 Subject: [PATCH 0933/1103] Temporary workaround for bug #875: Let TriangularView::nonZeros() return nonZeros() of the nested expression --- Eigen/src/SparseCore/SparseTriangularView.h | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index 1f5e53155..e051f4486 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -50,6 +50,13 @@ protected: template void solveInPlace(MatrixBase& other) const; template void solveInPlace(SparseMatrixBase& other) const; + + inline Index nonZeros() const { + // FIXME HACK number of nonZeros is required for product logic + // this returns only an upper bound (but should be OK for most purposes) + return derived().nestedExpression().nonZeros(); + } + }; From e2e7ba9f8511e9394b1823e14b40a727352c95be Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Tue, 28 Oct 2014 14:49:44 +0100 Subject: [PATCH 0934/1103] bug #898: add inline hint to const_cast_ptr --- Eigen/src/Core/util/XprHelper.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h index f2536714e..463ef85b4 100644 --- a/Eigen/src/Core/util/XprHelper.h +++ b/Eigen/src/Core/util/XprHelper.h @@ -371,7 +371,7 @@ template::type> struc template EIGEN_DEVICE_FUNC -T* const_cast_ptr(const T* ptr) +inline T* const_cast_ptr(const T* ptr) { return const_cast(ptr); } From debc97821c775518afd54e05e19dec9eb0c3bde1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 28 Oct 2014 23:10:13 -0700 Subject: [PATCH 0935/1103] Added support for tensor references --- unsupported/Eigen/CXX11/Tensor | 2 + .../src/Tensor/TensorForwardDeclarations.h | 1 + .../Eigen/CXX11/src/Tensor/TensorRef.h | 360 ++++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorTraits.h | 40 ++ unsupported/test/CMakeLists.txt | 1 + unsupported/test/cxx11_tensor_ref.cpp | 192 ++++++++++ 6 files changed, 596 insertions(+) create mode 100644 unsupported/Eigen/CXX11/src/Tensor/TensorRef.h create mode 100644 unsupported/test/cxx11_tensor_ref.cpp diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index 47447f446..c36db96ec 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -76,6 +76,8 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMap.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorRef.h" + #include "unsupported/Eigen/CXX11/src/Tensor/TensorIO.h" #include "Eigen/src/Core/util/ReenableStupidWarnings.h" diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 67f478822..a72e11215 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -15,6 +15,7 @@ namespace Eigen { template class Tensor; template class TensorFixedSize; template class TensorMap; +template class TensorRef; template::value> class TensorBase; template class TensorCwiseNullaryOp; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h new file mode 100644 index 000000000..db2027a5f --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorRef.h @@ -0,0 +1,360 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_CXX11_TENSOR_TENSOR_REF_H +#define EIGEN_CXX11_TENSOR_TENSOR_REF_H + +namespace Eigen { + +namespace internal { + +template +class TensorLazyBaseEvaluator { + public: + TensorLazyBaseEvaluator() : m_refcount(0) { } + virtual ~TensorLazyBaseEvaluator() { } + + virtual const Dimensions& dimensions() const = 0; + virtual const Scalar* data() const = 0; + + virtual const Scalar coeff(DenseIndex index) const = 0; + virtual Scalar& coeffRef(DenseIndex index) = 0; + + void incrRefCount() { ++m_refcount; } + void decrRefCount() { --m_refcount; } + int refCount() const { return m_refcount; } + + private: + // No copy, no assigment; + TensorLazyBaseEvaluator(const TensorLazyBaseEvaluator& other); + TensorLazyBaseEvaluator& operator = (const TensorLazyBaseEvaluator& other); + + int m_refcount; +}; + +static char dummy[8]; + +template +class TensorLazyEvaluatorReadOnly : public TensorLazyBaseEvaluator::Scalar> { + public: + // typedef typename TensorEvaluator::Dimensions Dimensions; + typedef typename TensorEvaluator::Scalar Scalar; + + TensorLazyEvaluatorReadOnly(const Expr& expr, const Device& device) : m_impl(expr, device) { + m_dims = m_impl.dimensions(); + m_impl.evalSubExprsIfNeeded(NULL); + } + virtual ~TensorLazyEvaluatorReadOnly() { + m_impl.cleanup(); + } + + virtual const Dimensions& dimensions() const { + return m_dims; + } + virtual const Scalar* data() const { + return m_impl.data(); + } + + virtual const Scalar coeff(DenseIndex index) const { + return m_impl.coeff(index); + } + virtual Scalar& coeffRef(DenseIndex index) { + eigen_assert(false && "can't reference the coefficient of a rvalue"); + return *reinterpret_cast(dummy); + }; + + protected: + TensorEvaluator m_impl; + Dimensions m_dims; +}; + +template +class TensorLazyEvaluatorWritable : public TensorLazyEvaluatorReadOnly { + public: + typedef TensorLazyEvaluatorReadOnly Base; + typedef typename Base::Scalar Scalar; + + TensorLazyEvaluatorWritable(const Expr& expr, const Device& device) : Base(expr, device) { + } + virtual ~TensorLazyEvaluatorWritable() { + } + + virtual Scalar& coeffRef(DenseIndex index) { + return this->m_impl.coeffRef(index); + } +}; + +template +class TensorLazyEvaluator : public internal::conditional::value), + TensorLazyEvaluatorWritable, + TensorLazyEvaluatorReadOnly >::type { + public: + typedef typename internal::conditional::value), + TensorLazyEvaluatorWritable, + TensorLazyEvaluatorReadOnly >::type Base; + typedef typename Base::Scalar Scalar; + + TensorLazyEvaluator(const Expr& expr, const Device& device) : Base(expr, device) { + } + virtual ~TensorLazyEvaluator() { + } +}; + +} // namespace internal + + +/** \class TensorRef + * \ingroup CXX11_Tensor_Module + * + * \brief A reference to a tensor expression + * The expression will be evaluated lazily (as much as possible). + * + */ +template class TensorRef : public TensorBase > +{ + public: + typedef TensorRef Self; + typedef typename PlainObjectType::Base Base; + typedef typename Eigen::internal::nested::type Nested; + typedef typename internal::traits::StorageKind StorageKind; + typedef typename internal::traits::Index Index; + typedef typename internal::traits::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename NumTraits::Real RealScalar; + typedef typename Base::CoeffReturnType CoeffReturnType; + typedef Scalar* PointerType; + typedef PointerType PointerArgType; + + static const Index NumIndices = PlainObjectType::NumIndices; + typedef typename PlainObjectType::Dimensions Dimensions; + + enum { + IsAligned = false, + PacketAccess = false, + }; + + EIGEN_STRONG_INLINE TensorRef() : m_evaluator(NULL) { + } + + template + EIGEN_STRONG_INLINE TensorRef(const Expression& expr) : m_evaluator(new internal::TensorLazyEvaluator(expr, DefaultDevice())) { + m_evaluator->incrRefCount(); + } + + template + EIGEN_STRONG_INLINE TensorRef& operator = (const Expression& expr) { + unrefEvaluator(); + m_evaluator = new internal::TensorLazyEvaluator(expr, DefaultDevice()); + m_evaluator->incrRefCount(); + return *this; + } + + ~TensorRef() { + unrefEvaluator(); + } + + TensorRef(const TensorRef& other) : m_evaluator(other.m_evaluator) { + eigen_assert(m_evaluator->refCount() > 0); + m_evaluator->incrRefCount(); + } + + TensorRef& operator = (const TensorRef& other) { + if (this != &other) { + unrefEvaluator(); + m_evaluator = other.m_evaluator; + eigen_assert(m_evaluator->refCount() > 0); + m_evaluator->incrRefCount(); + } + return *this; + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index dimension(Index n) const { return m_evaluator->dimensions()[n]; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_evaluator->dimensions(); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Index size() const { return m_evaluator->dimensions().TotalSize(); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar* data() const { return m_evaluator->data(); } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator()(Index index) const + { + return m_evaluator->coeff(index); + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator()(Index firstIndex, IndexTypes... otherIndices) const + { + const std::size_t NumIndices = (sizeof...(otherIndices) + 1); + const array indices{{firstIndex, otherIndices...}}; + return coeff(indices); + } +#else + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1) const + { + array indices; + indices[0] = i0; + indices[1] = i1; + return coeff(indices); + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1, Index i2) const + { + array indices; + indices[0] = i0; + indices[1] = i1; + indices[2] = i2; + return coeff(indices); + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1, Index i2, Index i3) const + { + array indices; + indices[0] = i0; + indices[1] = i1; + indices[2] = i2; + indices[3] = i3; + return coeff(indices); + } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar operator()(Index i0, Index i1, Index i2, Index i3, Index i4) const + { + array indices; + indices[0] = i0; + indices[1] = i1; + indices[2] = i2; + indices[3] = i3; + indices[4] = i4; + return coeff(indices); + } +#endif + + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar coeff(const array& indices) const + { + const Dimensions& dims = this->dimensions(); + Index index = 0; + if (PlainObjectType::Options&RowMajor) { + index += indices[0]; + for (int i = 1; i < NumIndices; ++i) { + index = index * dims[i] + indices[i]; + } + } else { + index += indices[NumIndices-1]; + for (int i = NumIndices-2; i >= 0; --i) { + index = index * dims[i] + indices[i]; + } + } + return m_evaluator->coeff(index); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const Scalar coeff(Index index) const + { + return m_evaluator->coeff(index); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) + { + return m_evaluator->coeffRef(index); + } + + private: + EIGEN_STRONG_INLINE void unrefEvaluator() { + if (m_evaluator) { + m_evaluator->decrRefCount(); + if (m_evaluator->refCount() == 0) { + delete m_evaluator; + } + } + } + + internal::TensorLazyBaseEvaluator* m_evaluator; +}; + + +// evaluator for rvalues +template +struct TensorEvaluator, Device> +{ + typedef typename Derived::Index Index; + typedef typename Derived::Scalar Scalar; + typedef typename Derived::Packet Packet; + typedef typename Derived::Scalar CoeffReturnType; + typedef typename Derived::Packet PacketReturnType; + typedef typename Derived::Dimensions Dimensions; + + enum { + IsAligned = false, + PacketAccess = false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const TensorRef& m, const Device&) + : m_ref(m) + { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_ref.dimensions(); } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + return true; + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { + return m_ref.coeff(index); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) { + return m_ref.coeffRef(index); + } + + Scalar* data() const { return m_ref.data(); } + + protected: + TensorRef m_ref; +}; + + +// evaluator for lvalues +template +struct TensorEvaluator, Device> : public TensorEvaluator, Device> +{ + typedef typename Derived::Index Index; + typedef typename Derived::Scalar Scalar; + typedef typename Derived::Packet Packet; + typedef typename Derived::Scalar CoeffReturnType; + typedef typename Derived::Packet PacketReturnType; + typedef typename Derived::Dimensions Dimensions; + + typedef TensorEvaluator, Device> Base; + + enum { + IsAligned = false, + PacketAccess = false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(TensorRef& m, const Device& d) : Base(m, d) + { } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) { + return this->m_ref.coeffRef(index); + } +}; + + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_REF_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h index 5940a8cf1..5c0f78489 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h @@ -84,6 +84,20 @@ struct traits > }; }; +template +struct traits > + : public traits +{ + typedef traits BaseTraits; + typedef typename BaseTraits::Scalar Scalar; + typedef typename BaseTraits::StorageKind StorageKind; + typedef typename BaseTraits::Index Index; + enum { + Options = BaseTraits::Options, + Flags = ((BaseTraits::Flags | LvalueBit) & ~AlignedBit) | (Options&Aligned ? AlignedBit : 0), + }; +}; + template struct eval, Eigen::Dense> @@ -121,6 +135,19 @@ struct eval, Eigen::Dense> typedef const TensorMap& type; }; +template +struct eval, Eigen::Dense> +{ + typedef const TensorRef& type; +}; + +template +struct eval, Eigen::Dense> +{ + typedef const TensorRef& type; +}; + + template struct nested, 1, typename eval >::type> { @@ -145,6 +172,7 @@ struct nested, 1, typename e typedef const TensorFixedSize& type; }; + template struct nested, 1, typename eval >::type> { @@ -157,6 +185,18 @@ struct nested, 1, typename eval& type; }; +template +struct nested, 1, typename eval >::type> +{ + typedef const TensorRef& type; +}; + +template +struct nested, 1, typename eval >::type> +{ + typedef const TensorRef& type; +}; + } // end namespace internal } // end namespace Eigen diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index a7ef2b402..2b5395013 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -126,5 +126,6 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_striding "-std=c++0x") # ei_add_test(cxx11_tensor_device "-std=c++0x") ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") + ei_add_test(cxx11_tensor_ref "-std=c++0x") ei_add_test(cxx11_tensor_io "-std=c++0x") endif() diff --git a/unsupported/test/cxx11_tensor_ref.cpp b/unsupported/test/cxx11_tensor_ref.cpp new file mode 100644 index 000000000..4ff94a059 --- /dev/null +++ b/unsupported/test/cxx11_tensor_ref.cpp @@ -0,0 +1,192 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#include "main.h" + +#include + +using Eigen::Tensor; +using Eigen::RowMajor; + +static void test_simple_lvalue_ref() +{ + Tensor input(6); + input.setRandom(); + + TensorRef> ref3(input); + TensorRef> ref4 = input; + + VERIFY_IS_EQUAL(ref3.data(), input.data()); + VERIFY_IS_EQUAL(ref4.data(), input.data()); + + for (int i = 0; i < 6; ++i) { + VERIFY_IS_EQUAL(ref3(i), input(i)); + VERIFY_IS_EQUAL(ref4(i), input(i)); + } + + for (int i = 0; i < 6; ++i) { + ref3.coeffRef(i) = i; + } + for (int i = 0; i < 6; ++i) { + VERIFY_IS_EQUAL(input(i), i); + } + for (int i = 0; i < 6; ++i) { + ref4.coeffRef(i) = -i * 2; + } + for (int i = 0; i < 6; ++i) { + VERIFY_IS_EQUAL(input(i), -i*2); + } +} + + +static void test_simple_rvalue_ref() +{ + Tensor input1(6); + input1.setRandom(); + Tensor input2(6); + input2.setRandom(); + + TensorRef> ref3(input1 + input2); + TensorRef> ref4 = input1 + input2; + + VERIFY_IS_NOT_EQUAL(ref3.data(), input1.data()); + VERIFY_IS_NOT_EQUAL(ref4.data(), input1.data()); + VERIFY_IS_NOT_EQUAL(ref3.data(), input2.data()); + VERIFY_IS_NOT_EQUAL(ref4.data(), input2.data()); + + for (int i = 0; i < 6; ++i) { + VERIFY_IS_EQUAL(ref3(i), input1(i) + input2(i)); + VERIFY_IS_EQUAL(ref4(i), input1(i) + input2(i)); + } +} + + +static void test_multiple_dims() +{ + Tensor input(3,5,7); + input.setRandom(); + + TensorRef> ref(input); + VERIFY_IS_EQUAL(ref.data(), input.data()); + VERIFY_IS_EQUAL(ref.dimension(0), 3); + VERIFY_IS_EQUAL(ref.dimension(1), 5); + VERIFY_IS_EQUAL(ref.dimension(2), 7); + + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(ref(i,j,k), input(i,j,k)); + } + } + } +} + + +static void test_slice() +{ + Tensor tensor(2,3,5,7,11); + tensor.setRandom(); + + Eigen::DSizes indices(1,2,3,4,5); + Eigen::DSizes sizes(1,1,1,1,1); + TensorRef> slice = tensor.slice(indices, sizes); + VERIFY_IS_EQUAL(slice(0,0,0,0,0), tensor(1,2,3,4,5)); + + Eigen::DSizes indices2(1,1,3,4,5); + Eigen::DSizes sizes2(1,1,2,2,3); + slice = tensor.slice(indices2, sizes2); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 2; ++j) { + for (int k = 0; k < 3; ++k) { + VERIFY_IS_EQUAL(slice(0,0,i,j,k), tensor(1,1,3+i,4+j,5+k)); + } + } + } + + Eigen::DSizes indices3(0,0,0,0,0); + Eigen::DSizes sizes3(2,3,1,1,1); + slice = tensor.slice(indices3, sizes3); + VERIFY_IS_EQUAL(slice.data(), tensor.data()); +} + + +static void test_ref_of_ref() +{ + Tensor input(3,5,7); + input.setRandom(); + + TensorRef> ref(input); + TensorRef> ref_of_ref(ref); + TensorRef> ref_of_ref2; + ref_of_ref2 = ref; + + VERIFY_IS_EQUAL(ref_of_ref.data(), input.data()); + VERIFY_IS_EQUAL(ref_of_ref.dimension(0), 3); + VERIFY_IS_EQUAL(ref_of_ref.dimension(1), 5); + VERIFY_IS_EQUAL(ref_of_ref.dimension(2), 7); + + VERIFY_IS_EQUAL(ref_of_ref2.data(), input.data()); + VERIFY_IS_EQUAL(ref_of_ref2.dimension(0), 3); + VERIFY_IS_EQUAL(ref_of_ref2.dimension(1), 5); + VERIFY_IS_EQUAL(ref_of_ref2.dimension(2), 7); + + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(ref_of_ref(i,j,k), input(i,j,k)); + VERIFY_IS_EQUAL(ref_of_ref2(i,j,k), input(i,j,k)); + } + } + } +} + + +static void test_ref_in_expr() +{ + Tensor input(3,5,7); + input.setRandom(); + TensorRef> input_ref(input); + + Tensor result(3,5,7); + result.setRandom(); + TensorRef> result_ref(result); + + Tensor bias(3,5,7); + bias.setRandom(); + + result_ref = input_ref + bias; + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(result_ref(i,j,k), input(i,j,k) + bias(i,j,k)); + VERIFY_IS_NOT_EQUAL(result(i,j,k), input(i,j,k) + bias(i,j,k)); + } + } + } + + result = result_ref; + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(result(i,j,k), input(i,j,k) + bias(i,j,k)); + } + } + } +} + + +void test_cxx11_tensor_ref() +{ + CALL_SUBTEST(test_simple_lvalue_ref()); + CALL_SUBTEST(test_simple_rvalue_ref()); + CALL_SUBTEST(test_multiple_dims()); + CALL_SUBTEST(test_slice()); + CALL_SUBTEST(test_ref_of_ref()); + CALL_SUBTEST(test_ref_in_expr()); +} From 21c0a2ce0c317fa258380c92bea4e9e16e0840f2 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 29 Oct 2014 11:29:33 +0100 Subject: [PATCH 0936/1103] Move D&C SVD to official SVD module. --- Eigen/SVD | 9 ++++-- Eigen/src/Core/MatrixBase.h | 1 + Eigen/src/Core/util/ForwardDeclarations.h | 1 + .../src/BDCSVD => Eigen/src/SVD}/BDCSVD.h | 23 +++++++------- test/CMakeLists.txt | 1 + {unsupported/test => test}/bdcsvd.cpp | 30 +++++++++---------- unsupported/Eigen/BDCSVD | 26 ---------------- unsupported/Eigen/src/BDCSVD/CMakeLists.txt | 6 ---- unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt | 13 -------- unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt | 8 ----- unsupported/Eigen/src/CMakeLists.txt | 1 - unsupported/test/CMakeLists.txt | 1 - 12 files changed, 37 insertions(+), 83 deletions(-) rename {unsupported/Eigen/src/BDCSVD => Eigen/src/SVD}/BDCSVD.h (99%) rename {unsupported/test => test}/bdcsvd.cpp (84%) delete mode 100644 unsupported/Eigen/BDCSVD delete mode 100644 unsupported/Eigen/src/BDCSVD/CMakeLists.txt delete mode 100644 unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt delete mode 100644 unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt diff --git a/Eigen/SVD b/Eigen/SVD index c13472e82..dbd37b17a 100644 --- a/Eigen/SVD +++ b/Eigen/SVD @@ -12,20 +12,25 @@ * * * This module provides SVD decomposition for matrices (both real and complex). - * This decomposition is accessible via the following MatrixBase method: + * Two decomposition algorithms are provided: + * - JacobiSVD implementing two-sided Jacobi iterations is numerically very accurate, fast for small matrices, but very slow for larger ones. + * - BDCSVD implementing a recursive divide & conquer strategy on top of an upper-bidiagonalization which remains fast for large problems. + * These decompositions are accessible via the respective classes and following MatrixBase methods: * - MatrixBase::jacobiSvd() + * - MatrixBase::bdcSvd() * * \code * #include * \endcode */ +#include "src/SVD/UpperBidiagonalization.h" #include "src/SVD/SVDBase.h" #include "src/SVD/JacobiSVD.h" +#include "src/SVD/BDCSVD.h" #if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT) #include "src/SVD/JacobiSVD_MKL.h" #endif -#include "src/SVD/UpperBidiagonalization.h" #include "src/Core/util/ReenableStupidWarnings.h" diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 001513187..135d8dfc8 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -390,6 +390,7 @@ template class MatrixBase /////////// SVD module /////////// JacobiSVD jacobiSvd(unsigned int computationOptions = 0) const; + BDCSVD bdcSvd(unsigned int computationOptions = 0) const; /////////// Geometry module /////////// diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index be156a44a..1f7503dfa 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -251,6 +251,7 @@ template class HouseholderQR; template class ColPivHouseholderQR; template class FullPivHouseholderQR; template class JacobiSVD; +template class BDCSVD; template class LLT; template class LDLT; template class HouseholderSequence; diff --git a/unsupported/Eigen/src/BDCSVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h similarity index 99% rename from unsupported/Eigen/src/BDCSVD/BDCSVD.h rename to Eigen/src/SVD/BDCSVD.h index 94bc03aac..498541050 100644 --- a/unsupported/Eigen/src/BDCSVD/BDCSVD.h +++ b/Eigen/src/SVD/BDCSVD.h @@ -514,6 +514,7 @@ void BDCSVD::divide (Index firstCol, Index lastCol, Index firstRowW, // // TODO Opportunities for optimization: better root finding algo, better stopping criterion, better // handling of round-off errors, be consistent in ordering +// For instance, to solve the secular equation using FMM, see http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf template void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V) { @@ -528,7 +529,10 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec U.resize(n+1, n+1); if (m_compV) V.resize(n, n); - if (col0.hasNaN() || diag.hasNaN()) { std::cout << "\n\nHAS NAN\n\n"; return; } +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE + if (col0.hasNaN() || diag.hasNaN()) + std::cout << "\n\nHAS NAN\n\n"; +#endif // Many singular values might have been deflated, the zero ones have been moved to the end, // but others are interleaved and we must ignore them at this stage. @@ -544,7 +548,7 @@ void BDCSVD::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec ArrayXr shifts(n), mus(n), zhat(n); -#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE +#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE std::cout << "computeSVDofM using:\n"; std::cout << " z: " << col0.transpose() << "\n"; std::cout << " d: " << diag.transpose() << "\n"; @@ -1149,21 +1153,18 @@ void BDCSVD::deflation(Index firstCol, Index lastCol, Index k, Index }//end deflation - /** \svd_module - * - * \return the singular value decomposition of \c *this computed by - * BDC Algorithm - * - * \sa class BDCSVD - */ -/* +/** \svd_module + * + * \return the singular value decomposition of \c *this computed by Divide & Conquer algorithm + * + * \sa class BDCSVD + */ template BDCSVD::PlainObject> MatrixBase::bdcSvd(unsigned int computationOptions) const { return BDCSVD(*this, computationOptions); } -*/ } // end namespace Eigen diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 63ed2c7a4..f57d8ce36 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -211,6 +211,7 @@ ei_add_test(real_qz) ei_add_test(eigensolver_generalized_real) ei_add_test(jacobi) ei_add_test(jacobisvd) +ei_add_test(bdcsvd) ei_add_test(householder) ei_add_test(geo_orthomethods) ei_add_test(geo_quaternion) diff --git a/unsupported/test/bdcsvd.cpp b/test/bdcsvd.cpp similarity index 84% rename from unsupported/test/bdcsvd.cpp rename to test/bdcsvd.cpp index 97d7880bb..d58d259d6 100644 --- a/unsupported/test/bdcsvd.cpp +++ b/test/bdcsvd.cpp @@ -15,7 +15,7 @@ #define EIGEN_RUNTIME_NO_MALLOC #include "main.h" -#include +#include #include #include @@ -35,18 +35,18 @@ void bdcsvd(const MatrixType& a = MatrixType(), bool pickrandom = true) CALL_SUBTEST(( svd_test_all_computation_options >(m, false) )); } -// template -// void bdcsvd_method() -// { -// enum { Size = MatrixType::RowsAtCompileTime }; -// typedef typename MatrixType::RealScalar RealScalar; -// typedef Matrix RealVecType; -// MatrixType m = MatrixType::Identity(); -// VERIFY_IS_APPROX(m.bdcSvd().singularValues(), RealVecType::Ones()); -// VERIFY_RAISES_ASSERT(m.bdcSvd().matrixU()); -// VERIFY_RAISES_ASSERT(m.bdcSvd().matrixV()); -// VERIFY_IS_APPROX(m.bdcSvd(ComputeFullU|ComputeFullV).solve(m), m); -// } +template +void bdcsvd_method() +{ + enum { Size = MatrixType::RowsAtCompileTime }; + typedef typename MatrixType::RealScalar RealScalar; + typedef Matrix RealVecType; + MatrixType m = MatrixType::Identity(); + VERIFY_IS_APPROX(m.bdcSvd().singularValues(), RealVecType::Ones()); + VERIFY_RAISES_ASSERT(m.bdcSvd().matrixU()); + VERIFY_RAISES_ASSERT(m.bdcSvd().matrixV()); + VERIFY_IS_APPROX(m.bdcSvd(ComputeFullU|ComputeFullV).solve(m), m); +} // compare the Singular values returned with Jacobi and Bdc template @@ -97,8 +97,8 @@ void test_bdcsvd() } // test matrixbase method -// CALL_SUBTEST_1(( bdcsvd_method() )); -// CALL_SUBTEST_3(( bdcsvd_method() )); + CALL_SUBTEST_1(( bdcsvd_method() )); + CALL_SUBTEST_3(( bdcsvd_method() )); // Test problem size constructors CALL_SUBTEST_7( BDCSVD(10,10) ); diff --git a/unsupported/Eigen/BDCSVD b/unsupported/Eigen/BDCSVD deleted file mode 100644 index 44649dbd0..000000000 --- a/unsupported/Eigen/BDCSVD +++ /dev/null @@ -1,26 +0,0 @@ -#ifndef EIGEN_BDCSVD_MODULE_H -#define EIGEN_BDCSVD_MODULE_H - -#include - -#include "../../Eigen/src/Core/util/DisableStupidWarnings.h" - -/** \defgroup BDCSVD_Module BDCSVD module - * - * - * - * This module provides Divide & Conquer SVD decomposition for matrices (both real and complex). - * This decomposition is accessible via the following MatrixBase method: - * - MatrixBase::bdcSvd() - * - * \code - * #include - * \endcode - */ - -#include "src/BDCSVD/BDCSVD.h" - -#include "../../Eigen/src/Core/util/ReenableStupidWarnings.h" - -#endif // EIGEN_BDCSVD_MODULE_H -/* vim: set filetype=cpp et sw=2 ts=2 ai: */ diff --git a/unsupported/Eigen/src/BDCSVD/CMakeLists.txt b/unsupported/Eigen/src/BDCSVD/CMakeLists.txt deleted file mode 100644 index 1045512f9..000000000 --- a/unsupported/Eigen/src/BDCSVD/CMakeLists.txt +++ /dev/null @@ -1,6 +0,0 @@ -FILE(GLOB Eigen_BDCSVD_SRCS "*.h") - -INSTALL(FILES - ${Eigen_BDCSVD_SRCS} - DESTINATION ${INCLUDE_INSTALL_DIR}/unsupported/Eigen/src/BDCSVD COMPONENT Devel - ) diff --git a/unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt b/unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt deleted file mode 100644 index 9b7bf9314..000000000 --- a/unsupported/Eigen/src/BDCSVD/TODOBdcsvd.txt +++ /dev/null @@ -1,13 +0,0 @@ -TO DO LIST - -- check more carefully single precision -- check with duplicated singularvalues -- no-malloc mode - -(optional optimization) -- do all the allocations in the allocate part -- support static matrices -- return a error at compilation time when using integer matrices (int, long, std::complex, ...) -- To solve the secular equation using FMM: -http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf - diff --git a/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt b/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt deleted file mode 100644 index 29ab9cd40..000000000 --- a/unsupported/Eigen/src/BDCSVD/doneInBDCSVD.txt +++ /dev/null @@ -1,8 +0,0 @@ -This unsupported package is about a divide and conquer algorithm to compute SVD. - -The implementation follows as closely as possible the following reference paper : -http://www.cs.yale.edu/publications/techreports/tr933.pdf - -To solve the secular equation using FMM: -http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf - diff --git a/unsupported/Eigen/src/CMakeLists.txt b/unsupported/Eigen/src/CMakeLists.txt index 654a2327f..8eb2808e3 100644 --- a/unsupported/Eigen/src/CMakeLists.txt +++ b/unsupported/Eigen/src/CMakeLists.txt @@ -12,4 +12,3 @@ ADD_SUBDIRECTORY(Skyline) ADD_SUBDIRECTORY(SparseExtra) ADD_SUBDIRECTORY(KroneckerProduct) ADD_SUBDIRECTORY(Splines) -ADD_SUBDIRECTORY(BDCSVD) diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 48b61cde0..97849a25a 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -92,7 +92,6 @@ ei_add_test(splines) ei_add_test(gmres) ei_add_test(minres) ei_add_test(levenberg_marquardt) -ei_add_test(bdcsvd) ei_add_test(kronecker_product) option(EIGEN_TEST_CXX11 "Enable testing of C++11 features (e.g. Tensor module)." OFF) From acecb7b09f36d8a073ebf399d9c9f038dd6fe3c2 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 29 Oct 2014 17:46:33 +0100 Subject: [PATCH 0937/1103] Fixed include in bdcsvd.cpp --- test/bdcsvd.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/bdcsvd.cpp b/test/bdcsvd.cpp index d58d259d6..52a02b697 100644 --- a/test/bdcsvd.cpp +++ b/test/bdcsvd.cpp @@ -22,7 +22,7 @@ #define SVD_DEFAULT(M) BDCSVD #define SVD_FOR_MIN_NORM(M) BDCSVD -#include "../../test/svd_common.h" +#include "svd_common.h" // Check all variants of JacobiSVD template From 3d25b1f5b8694638659217cf67d86534afd8db28 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 29 Oct 2014 17:46:54 +0100 Subject: [PATCH 0938/1103] Split up some test cases --- test/eigensolver_selfadjoint.cpp | 11 +++++++---- test/jacobisvd.cpp | 4 ++-- 2 files changed, 9 insertions(+), 6 deletions(-) diff --git a/test/eigensolver_selfadjoint.cpp b/test/eigensolver_selfadjoint.cpp index 8ff48462e..935736328 100644 --- a/test/eigensolver_selfadjoint.cpp +++ b/test/eigensolver_selfadjoint.cpp @@ -145,11 +145,14 @@ void test_eigensolver_selfadjoint() { int s = 0; for(int i = 0; i < g_repeat; i++) { + // trivial test for 1x1 matrices: + CALL_SUBTEST_1( selfadjointeigensolver(Matrix())); + CALL_SUBTEST_1( selfadjointeigensolver(Matrix())); // very important to test 3x3 and 2x2 matrices since we provide special paths for them - CALL_SUBTEST_1( selfadjointeigensolver(Matrix2f()) ); - CALL_SUBTEST_1( selfadjointeigensolver(Matrix2d()) ); - CALL_SUBTEST_1( selfadjointeigensolver(Matrix3f()) ); - CALL_SUBTEST_1( selfadjointeigensolver(Matrix3d()) ); + CALL_SUBTEST_12( selfadjointeigensolver(Matrix2f()) ); + CALL_SUBTEST_12( selfadjointeigensolver(Matrix2d()) ); + CALL_SUBTEST_13( selfadjointeigensolver(Matrix3f()) ); + CALL_SUBTEST_13( selfadjointeigensolver(Matrix3d()) ); CALL_SUBTEST_2( selfadjointeigensolver(Matrix4d()) ); s = internal::random(1,EIGEN_TEST_MAX_SIZE/4); CALL_SUBTEST_3( selfadjointeigensolver(MatrixXf(s,s)) ); diff --git a/test/jacobisvd.cpp b/test/jacobisvd.cpp index bfcadce95..f9de6b708 100644 --- a/test/jacobisvd.cpp +++ b/test/jacobisvd.cpp @@ -77,8 +77,8 @@ void test_jacobisvd() CALL_SUBTEST_7(( jacobisvd_verify_assert(MatrixXf(10,12)) )); CALL_SUBTEST_8(( jacobisvd_verify_assert(MatrixXcd(7,5)) )); - svd_all_trivial_2x2(jacobisvd); - svd_all_trivial_2x2(jacobisvd); + CALL_SUBTEST_11(svd_all_trivial_2x2(jacobisvd)); + CALL_SUBTEST_12(svd_all_trivial_2x2(jacobisvd)); for(int i = 0; i < g_repeat; i++) { CALL_SUBTEST_3(( jacobisvd() )); From d2fc597d5b05aaa4531ff6daede9a47c739eb70e Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Wed, 29 Oct 2014 17:51:14 +0100 Subject: [PATCH 0939/1103] Removed deprecated header (unsupported/Eigen/BDCSVD is included in Eigen/SVD now) --- lapack/svd.cpp | 1 - 1 file changed, 1 deletion(-) diff --git a/lapack/svd.cpp b/lapack/svd.cpp index ecac3bab1..df77a371c 100644 --- a/lapack/svd.cpp +++ b/lapack/svd.cpp @@ -9,7 +9,6 @@ #include "lapack_common.h" #include -#include // computes the singular values/vectors a general M-by-N matrix A using divide-and-conquer EIGEN_LAPACK_FUNC(gesdd,(char *jobz, int *m, int* n, Scalar* a, int *lda, RealScalar *s, Scalar *u, int *ldu, Scalar *vt, int *ldvt, Scalar* /*work*/, int* lwork, From e5f134006b431d7d307c8ce7706c5aa1df09895c Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 30 Oct 2014 19:59:09 +0100 Subject: [PATCH 0940/1103] EIGEN_UNUSED_VARIABLE works better than casting to void. Make this also usable from CUDA code --- Eigen/src/Core/util/Macros.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index c09854951..61d7be752 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -251,7 +251,7 @@ #endif #ifdef EIGEN_NO_DEBUG -#define EIGEN_ONLY_USED_FOR_DEBUG(x) (void)x +#define EIGEN_ONLY_USED_FOR_DEBUG(x) EIGEN_UNUSED_VARIABLE(x) #else #define EIGEN_ONLY_USED_FOR_DEBUG(x) #endif @@ -277,7 +277,7 @@ // Suppresses 'unused variable' warnings. namespace Eigen { namespace internal { - template void ignore_unused_variable(const T&) {} + template EIGEN_DEVICE_FUNC void ignore_unused_variable(const T&) {} } } #define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var); From 883168ed94ca2559231861957558a9ded13e4831 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Thu, 30 Oct 2014 20:16:16 +0100 Subject: [PATCH 0941/1103] Make select CUDA compatible (comparison operators aren't yet, so no test case yet) --- Eigen/src/Core/CoreEvaluators.h | 6 +++--- Eigen/src/Core/Select.h | 13 ++++++++----- 2 files changed, 11 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h index 836d25b3b..a0dc72c4d 100644 --- a/Eigen/src/Core/CoreEvaluators.h +++ b/Eigen/src/Core/CoreEvaluators.h @@ -859,7 +859,7 @@ struct evaluator > Flags = (unsigned int)evaluator::Flags & evaluator::Flags & HereditaryBits }; - EIGEN_DEVICE_FUNC explicit evaluator(const XprType& select) + inline EIGEN_DEVICE_FUNC explicit evaluator(const XprType& select) : m_conditionImpl(select.conditionMatrix()), m_thenImpl(select.thenMatrix()), m_elseImpl(select.elseMatrix()) @@ -868,7 +868,7 @@ struct evaluator > typedef typename XprType::Index Index; typedef typename XprType::CoeffReturnType CoeffReturnType; - EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const + inline EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const { if (m_conditionImpl.coeff(row, col)) return m_thenImpl.coeff(row, col); @@ -876,7 +876,7 @@ struct evaluator > return m_elseImpl.coeff(row, col); } - EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const + inline EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index index) const { if (m_conditionImpl.coeff(index)) return m_thenImpl.coeff(index); diff --git a/Eigen/src/Core/Select.h b/Eigen/src/Core/Select.h index 0cb85a4ad..0c09a4ff4 100644 --- a/Eigen/src/Core/Select.h +++ b/Eigen/src/Core/Select.h @@ -57,6 +57,7 @@ class Select : internal::no_assignment_operator, typedef typename internal::dense_xpr_base

 - +
Constant or Random Insertion \code -sm1.setZero(); // Set the matrix with zero elements -sm1.setConstant(val); //Replace all the nonzero values with val +sm1.setZero(); \endcode The matrix sm1 should have been created before ???Remove all non-zero coefficients
From 0feff6e987750a61f0ee14774efaef85d2fb6fac Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 14 Jan 2015 15:29:48 -0800 Subject: [PATCH 1025/1103] Expanded the functionality of index lists --- .../Eigen/CXX11/src/Tensor/TensorIndexList.h | 105 +++++++++++++- unsupported/test/cxx11_tensor_index_list.cpp | 131 ++++++++++++++++++ 2 files changed, 231 insertions(+), 5 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h index eaf0195ce..209749042 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h @@ -95,6 +95,20 @@ struct tuple_coeff { return ((i == Idx) & is_compile_time_constant >::type>::value) || tuple_coeff::value_known_statically(i, t); } + + template + static constexpr bool values_up_to_known_statically(const std::tuple& t) { + return is_compile_time_constant >::type>::value && + tuple_coeff::values_up_to_known_statically(t); + } + + template + static constexpr bool values_up_to_statically_known_to_increase(const std::tuple& t) { + return is_compile_time_constant >::type>::value && + is_compile_time_constant >::type>::value && + std::get(t) > std::get(t) && + tuple_coeff::values_up_to_statically_known_to_increase(t); + } }; template <> @@ -110,10 +124,20 @@ struct tuple_coeff<0> { update_value(std::get<0>(t), value); } template - static constexpr bool value_known_statically(const DenseIndex i, const std::tuple&) { + static constexpr bool value_known_statically(const DenseIndex i, const std::tuple& t) { // eigen_assert (i == 0); // gcc fails to compile assertions in constexpr return is_compile_time_constant >::type>::value & (i == 0); } + + template + static constexpr bool values_up_to_known_statically(const std::tuple& t) { + return is_compile_time_constant >::type>::value; + } + + template + static constexpr bool values_up_to_statically_known_to_increase(const std::tuple& t) { + return true; + } }; } // namespace internal @@ -133,6 +157,13 @@ struct IndexList : std::tuple { constexpr bool value_known_statically(const DenseIndex i) const { return internal::tuple_coeff >::value-1>::value_known_statically(i, *this); } + constexpr bool all_values_known_statically() const { + return internal::tuple_coeff >::value-1>::values_up_to_known_statically(*this); + } + + constexpr bool values_statically_known_to_increase() const { + return internal::tuple_coeff >::value-1>::values_up_to_statically_known_to_increase(*this); + } }; @@ -144,6 +175,14 @@ constexpr IndexList make_index_list(FirstType val1, Ot namespace internal { +template size_t array_prod(const IndexList& sizes) { + size_t result = 1; + for (int i = 0; i < array_size >::value; ++i) { + result *= sizes[i]; + } + return result; +} + template struct array_size > { static const size_t value = std::tuple_size >::value; }; @@ -179,6 +218,48 @@ struct index_known_statically > { } }; +template +struct all_indices_known_statically { + constexpr bool operator() () const { + return false; + } +}; + +template +struct all_indices_known_statically > { + constexpr bool operator() () const { + return IndexList().all_values_known_statically(); + } +}; + +template +struct all_indices_known_statically > { + constexpr bool operator() () const { + return IndexList().all_values_known_statically(); + } +}; + +template +struct indices_statically_known_to_increase { + constexpr bool operator() () const { + return false; + } +}; + +template +struct indices_statically_known_to_increase > { + constexpr bool operator() () const { + return IndexList().values_statically_known_to_increase(); + } +}; + +template +struct indices_statically_known_to_increase > { + constexpr bool operator() () const { + return IndexList().values_statically_known_to_increase(); + } +}; + template struct index_statically_eq { constexpr bool operator() (DenseIndex, DenseIndex) const { @@ -190,7 +271,7 @@ template struct index_statically_eq > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - (IndexList()[i] == value); + IndexList()[i] == value; } }; @@ -198,7 +279,7 @@ template struct index_statically_eq > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - (IndexList()[i] == value); + IndexList()[i] == value; } }; @@ -213,7 +294,7 @@ template struct index_statically_ne > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - (IndexList()[i] != value); + IndexList()[i] != value; } }; @@ -221,7 +302,7 @@ template struct index_statically_ne > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - (IndexList()[i] != value); + IndexList()[i] != value; } }; @@ -242,6 +323,20 @@ struct index_known_statically { } }; +template +struct all_indices_known_statically { + EIGEN_ALWAYS_INLINE EIGEN_DEVICE_FUNC bool operator() () const { + return false; + } +}; + +template +struct indices_statically_known_to_increase { + EIGEN_ALWAYS_INLINE EIGEN_DEVICE_FUNC bool operator() () const { + return false; + } +}; + template struct index_statically_eq { EIGEN_ALWAYS_INLINE EIGEN_DEVICE_FUNC bool operator() (DenseIndex, DenseIndex) const{ diff --git a/unsupported/test/cxx11_tensor_index_list.cpp b/unsupported/test/cxx11_tensor_index_list.cpp index 6a103cab1..d79a3ed45 100644 --- a/unsupported/test/cxx11_tensor_index_list.cpp +++ b/unsupported/test/cxx11_tensor_index_list.cpp @@ -44,6 +44,120 @@ static void test_static_index_list() } +static void test_type2index_list() +{ + Tensor tensor(2,3,5,7,11); + tensor.setRandom(); + tensor += tensor.constant(10.0f); + + typedef Eigen::IndexList> Dims0; + typedef Eigen::IndexList, Eigen::type2index<1>> Dims1; + typedef Eigen::IndexList, Eigen::type2index<1>, Eigen::type2index<2>> Dims2; + typedef Eigen::IndexList, Eigen::type2index<1>, Eigen::type2index<2>, Eigen::type2index<3>> Dims3; + typedef Eigen::IndexList, Eigen::type2index<1>, Eigen::type2index<2>, Eigen::type2index<3>, Eigen::type2index<4>> Dims4; + +#if 0 + EIGEN_STATIC_ASSERT((internal::indices_statically_known_to_increase()() == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::indices_statically_known_to_increase()() == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::indices_statically_known_to_increase()() == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::indices_statically_known_to_increase()() == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::indices_statically_known_to_increase()() == true), YOU_MADE_A_PROGRAMMING_MISTAKE); +#endif + + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::are_inner_most_dims::value == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + + const Dims0 reduction_axis0; + Tensor result0 = tensor.sum(reduction_axis0); + for (int m = 0; m < 11; ++m) { + for (int l = 0; l < 7; ++l) { + for (int k = 0; k < 5; ++k) { + for (int j = 0; j < 3; ++j) { + float expected = 0.0f; + for (int i = 0; i < 2; ++i) { + expected += tensor(i,j,k,l,m); + } + VERIFY_IS_APPROX(result0(j,k,l,m), expected); + } + } + } + } + + const Dims1 reduction_axis1; + Tensor result1 = tensor.sum(reduction_axis1); + for (int m = 0; m < 11; ++m) { + for (int l = 0; l < 7; ++l) { + for (int k = 0; k < 5; ++k) { + float expected = 0.0f; + for (int j = 0; j < 3; ++j) { + for (int i = 0; i < 2; ++i) { + expected += tensor(i,j,k,l,m); + } + } + VERIFY_IS_APPROX(result1(k,l,m), expected); + } + } + } + + const Dims2 reduction_axis2; + Tensor result2 = tensor.sum(reduction_axis2); + for (int m = 0; m < 11; ++m) { + for (int l = 0; l < 7; ++l) { + float expected = 0.0f; + for (int k = 0; k < 5; ++k) { + for (int j = 0; j < 3; ++j) { + for (int i = 0; i < 2; ++i) { + expected += tensor(i,j,k,l,m); + } + } + } + VERIFY_IS_APPROX(result2(l,m), expected); + } + } + + const Dims3 reduction_axis3; + Tensor result3 = tensor.sum(reduction_axis3); + for (int m = 0; m < 11; ++m) { + float expected = 0.0f; + for (int l = 0; l < 7; ++l) { + for (int k = 0; k < 5; ++k) { + for (int j = 0; j < 3; ++j) { + for (int i = 0; i < 2; ++i) { + expected += tensor(i,j,k,l,m); + } + } + } + } + VERIFY_IS_APPROX(result3(m), expected); + } + + const Dims4 reduction_axis4; + Tensor result4 = tensor.sum(reduction_axis4); + float expected = 0.0f; + for (int m = 0; m < 11; ++m) { + for (int l = 0; l < 7; ++l) { + for (int k = 0; k < 5; ++k) { + for (int j = 0; j < 3; ++j) { + for (int i = 0; i < 2; ++i) { + expected += tensor(i,j,k,l,m); + } + } + } + } + } + VERIFY_IS_APPROX(result4(0), expected); +} + + static void test_dynamic_index_list() { Tensor tensor(2,3,5,7); @@ -105,10 +219,25 @@ static void test_mixed_index_list() EIGEN_STATIC_ASSERT((internal::index_known_statically()(2) == true), YOU_MADE_A_PROGRAMMING_MISTAKE); EIGEN_STATIC_ASSERT((internal::index_statically_eq()(0, 0) == true), YOU_MADE_A_PROGRAMMING_MISTAKE); EIGEN_STATIC_ASSERT((internal::index_statically_eq()(2, 2) == true), YOU_MADE_A_PROGRAMMING_MISTAKE); +#if 0 + EIGEN_STATIC_ASSERT((internal::all_indices_known_statically()() == false), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::indices_statically_known_to_increase()() == false), YOU_MADE_A_PROGRAMMING_MISTAKE); +#endif + typedef IndexList, type2index<1>, type2index<2>, type2index<3>> ReductionList; + ReductionList reduction_list; + EIGEN_STATIC_ASSERT((internal::index_statically_eq()(0, 0) == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::index_statically_eq()(1, 1) == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::index_statically_eq()(2, 2) == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::index_statically_eq()(3, 3) == true), YOU_MADE_A_PROGRAMMING_MISTAKE); +#if 0 + EIGEN_STATIC_ASSERT((internal::all_indices_known_statically()() == true), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((internal::indices_statically_known_to_increase()() == true), YOU_MADE_A_PROGRAMMING_MISTAKE); +#endif Tensor result1 = tensor.sum(reduction_axis); Tensor result2 = tensor.sum(reduction_indices); + Tensor result3 = tensor.sum(reduction_list); float expected = 0.0f; for (int i = 0; i < 2; ++i) { @@ -122,12 +251,14 @@ static void test_mixed_index_list() } VERIFY_IS_APPROX(result1(0), expected); VERIFY_IS_APPROX(result2(0), expected); + VERIFY_IS_APPROX(result3(0), expected); } void test_cxx11_tensor_index_list() { CALL_SUBTEST(test_static_index_list()); + CALL_SUBTEST(test_type2index_list()); CALL_SUBTEST(test_dynamic_index_list()); CALL_SUBTEST(test_mixed_index_list()); } From 4cdf3fe427b4fdc271733d0404a66e2d5613cb16 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 14 Jan 2015 15:30:47 -0800 Subject: [PATCH 1026/1103] Misc fixes --- .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 136 ++++++------------ 1 file changed, 44 insertions(+), 92 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 6d9e09318..6c9a67c58 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -40,6 +40,10 @@ template struct IndexPair { // Boilerplate code namespace internal { +template struct dget { + static const std::size_t value = get::value; +}; + template struct fixed_size_tensor_index_linearization_helper @@ -49,7 +53,7 @@ struct fixed_size_tensor_index_linearization_helper const Dimensions& dimensions) { return array_get(indices) + - get::value * + dget::value * fixed_size_tensor_index_linearization_helper::run(indices, dimensions); } }; @@ -75,6 +79,10 @@ struct Sizes : internal::numeric_list { typedef internal::numeric_list Base; static const std::size_t total_size = internal::arg_prod(Indices...); + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t rank() const { + return Base::count; + } + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t TotalSize() { return internal::arg_prod(Indices...); } @@ -85,6 +93,7 @@ struct Sizes : internal::numeric_list { // todo: add assertion } #ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template Sizes(DenseIndex... indices) { } explicit Sizes(std::initializer_list /*l*/) { // todo: add assertion } @@ -121,11 +130,15 @@ struct non_zero_size<0> { typedef internal::null_type type; }; -template struct Sizes : typename internal::make_type_list::type, typename non_zero_size::type, typename non_zero_size::type, typename non_zero_size::type, typename non_zero_size::type >::type { +template struct Sizes { typedef typename internal::make_type_list::type, typename non_zero_size::type, typename non_zero_size::type, typename non_zero_size::type, typename non_zero_size::type >::type Base; static const size_t count = Base::count; static const std::size_t total_size = internal::arg_prod::value; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t rank() const { + return count; + } + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() { return internal::arg_prod::value; } @@ -160,11 +173,11 @@ template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfColMajor(const array& indices) const { - return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this)); + return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfRowMajor(const array& indices) const { - return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this)); + return internal::fixed_size_tensor_index_linearization_helper::run(indices, *static_cast(this); } }; @@ -208,6 +221,10 @@ struct DSizes : array { typedef array Base; static const std::size_t count = NumDims; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t rank() const { + return NumDims; + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() const { return internal::array_prod(*static_cast(this)); } @@ -219,31 +236,44 @@ struct DSizes : array { } EIGEN_DEVICE_FUNC explicit DSizes(const array& a) : Base(a) { } +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE explicit DSizes(DenseIndex firstDimension, IndexTypes... otherDimensions) { + EIGEN_STATIC_ASSERT(sizeof...(otherDimensions) + 1 == NumDims, YOU_MADE_A_PROGRAMMING_MISTAKE) + (*this) = array{{firstDimension, otherDimensions...}}; + } +#else EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0) { + eigen_assert(NumDims == 1); (*this)[0] = i0; } EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1) { + eigen_assert(NumDims == 2); (*this)[0] = i0; (*this)[1] = i1; } EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2) { + eigen_assert(NumDims == 3); (*this)[0] = i0; (*this)[1] = i1; (*this)[2] = i2; } EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3) { + eigen_assert(NumDims == 4); (*this)[0] = i0; (*this)[1] = i1; (*this)[2] = i2; (*this)[3] = i3; } EIGEN_DEVICE_FUNC explicit DSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3, const DenseIndex i4) { + eigen_assert(NumDims == 5); (*this)[0] = i0; (*this)[1] = i1; (*this)[2] = i2; (*this)[3] = i3; (*this)[4] = i4; } +#endif DSizes& operator = (const array& other) { *static_cast(this) = other; @@ -287,84 +317,6 @@ struct tensor_vsize_index_linearization_helper }; } // end namespace internal -template -struct VSizes : std::vector { - typedef std::vector Base; - - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t TotalSize() const { - return internal::array_prod(*static_cast(this)); - } - - EIGEN_DEVICE_FUNC VSizes() { } - EIGEN_DEVICE_FUNC explicit VSizes(const std::vector& a) : Base(a) { } - - template - EIGEN_DEVICE_FUNC explicit VSizes(const array& a) { - this->resize(NumDims); - for (int i = 0; i < NumDims; ++i) { - (*this)[i] = a[i]; - } - } - - EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0) { - this->resize(1); - (*this)[0] = i0; - } - EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1) { - this->resize(2); - (*this)[0] = i0; - (*this)[1] = i1; - } - EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2) { - this->resize(3); - (*this)[0] = i0; - (*this)[1] = i1; - (*this)[2] = i2; - } - EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3) { - this->resize(4); - (*this)[0] = i0; - (*this)[1] = i1; - (*this)[2] = i2; - (*this)[3] = i3; - } - EIGEN_DEVICE_FUNC explicit VSizes(const DenseIndex i0, const DenseIndex i1, const DenseIndex i2, const DenseIndex i3, const DenseIndex i4) { - this->resize(5); - (*this)[0] = i0; - (*this)[1] = i1; - (*this)[2] = i2; - (*this)[3] = i3; - (*this)[4] = i4; - } - - VSizes& operator = (const std::vector& other) { - *static_cast(this) = other; - return *this; - } - - // A constexpr would be so much better here - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfColMajor(const array& indices) const { - return internal::tensor_vsize_index_linearization_helper::run(indices, *static_cast(this)); - } - template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t IndexOfRowMajor(const array& indices) const { - return internal::tensor_vsize_index_linearization_helper::run(indices, *static_cast(this)); - } -}; - - -// Boilerplate -namespace internal { -template -EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex array_prod(const VSizes& sizes) { - DenseIndex total_size = 1; - for (int i = 0; i < sizes.size(); ++i) { - total_size *= sizes[i]; - } - return total_size; -} -} namespace internal { @@ -381,8 +333,8 @@ static const size_t value = Sizes::count; template struct array_size > { static const size_t value = Sizes::count; }; - template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_get(const Sizes) { - return get::Base>::value; +template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_get(const Sizes& a) { + return get >::value; } #else template struct array_size > { @@ -412,17 +364,17 @@ struct sizes_match_up_to_dim { } }; -template -bool dimensions_match(Dims1& dims1, Dims2& dims2) { - if (array_size::value != array_size::value) { - return false; - } - return sizes_match_up_to_dim::value-1>::run(dims1, dims2); -} - } // end namespace internal +template +bool dimensions_match(Dims1& dims1, Dims2& dims2) { + if (internal::array_size::value != internal::array_size::value) { + return false; + } + return internal::sizes_match_up_to_dim::value-1>::run(dims1, dims2); +} + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_DIMENSIONS_H From 703c526355c929cc6c422b7599ecfed57642e988 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 14 Jan 2015 15:31:52 -0800 Subject: [PATCH 1027/1103] Misc improvements --- .../Eigen/CXX11/src/Core/util/CXX11Meta.h | 20 +++-- .../CXX11/src/Core/util/CXX11Workarounds.h | 12 ++- .../CXX11/src/Core/util/EmulateCXX11Meta.h | 81 +++++++++++++++---- 3 files changed, 86 insertions(+), 27 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h index 1e6b97ce4..36d91e780 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -42,14 +42,14 @@ struct numeric_list { constexpr static std::size_t count = sizeof.. * typename gen_numeric_list_repeated::type numeric_list */ -template struct gen_numeric_list : gen_numeric_list {}; -template struct gen_numeric_list { typedef numeric_list type; }; +template struct gen_numeric_list : gen_numeric_list {}; +template struct gen_numeric_list { typedef numeric_list type; }; -template struct gen_numeric_list_reversed : gen_numeric_list_reversed {}; -template struct gen_numeric_list_reversed { typedef numeric_list type; }; +template struct gen_numeric_list_reversed : gen_numeric_list_reversed {}; +template struct gen_numeric_list_reversed { typedef numeric_list type; }; -template struct gen_numeric_list_swapped_pair : gen_numeric_list_swapped_pair {}; -template struct gen_numeric_list_swapped_pair { typedef numeric_list type; }; +template struct gen_numeric_list_swapped_pair : gen_numeric_list_swapped_pair {}; +template struct gen_numeric_list_swapped_pair { typedef numeric_list type; }; template struct gen_numeric_list_repeated : gen_numeric_list_repeated {}; template struct gen_numeric_list_repeated { typedef numeric_list type; }; @@ -370,6 +370,14 @@ constexpr inline auto array_prod(std::array arr) -> decltype(array_reduce< return array_reduce(arr); } +template +EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE t array_prod(const std::vector& a) { + eigen_assert(a.size() > 0); + t prod = 1; + for (size_t i = 0; i < a.size(); ++i) { prod *= a[i]; } + return prod; +} + /* zip an array */ template diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h index e30eb6ad8..a590cf4e1 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Workarounds.h @@ -48,15 +48,13 @@ namespace internal { * - libstdc++ from version 4.7 onwards has it nevertheless, * so use that * - libstdc++ older versions: use _M_instance directly - * - libc++ from version 3.4 onwards has it IF compiled with - * -std=c++1y - * - libc++ older versions or -std=c++11: use __elems_ directly + * - libc++ all versions so far: use __elems_ directly * - all other libs: use std::get to be portable, but * this may not be constexpr */ #if defined(__GLIBCXX__) && __GLIBCXX__ < 20120322 #define STD_GET_ARR_HACK a._M_instance[I] -#elif defined(_LIBCPP_VERSION) && (!defined(_LIBCPP_STD_VER) || _LIBCPP_STD_VER <= 11) +#elif defined(_LIBCPP_VERSION) #define STD_GET_ARR_HACK a.__elems_[I] #else #define STD_GET_ARR_HACK std::template get(a) @@ -70,14 +68,14 @@ template constexpr inline T& array_get(std::vector template constexpr inline T&& array_get(std::vector&& a) { return a[I]; } template constexpr inline T const& array_get(std::vector const& a) { return a[I]; } - #undef STD_GET_ARR_HACK template struct array_size; -template struct array_size > { +template struct array_size > { static const size_t value = N; }; -template struct array_size > { +template struct array_size; +template struct array_size > { static const size_t value = N; }; diff --git a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h index e45d0a3b1..494f95690 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/EmulateCXX11Meta.h @@ -29,7 +29,7 @@ template class array { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE array() { } - EIGEN_DEVICE_FUNC + explicit EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE array(const T& v) { EIGEN_STATIC_ASSERT(n==1, YOU_MADE_A_PROGRAMMING_MISTAKE) values[0] = v; @@ -106,6 +106,7 @@ template class array { #ifdef EIGEN_HAS_VARIADIC_TEMPLATES array(std::initializer_list l) { + eigen_assert(l.size() == n); std::copy(l.begin(), l.end(), values); } #endif @@ -211,6 +212,29 @@ template struct gen_numeric_list_repeated { template struct get; +template +struct get +{ + get() { eigen_assert(false && "index overflow"); } + typedef void type; + static const char value = '\0'; +}; + +template +struct get > +{ + get() { eigen_assert(false && "index overflow"); } + typedef void type; + static const char value = '\0'; +}; + +template +struct get<0, type_list > +{ + typedef typename Head::type type; + static const type value = Head::value; +}; + template struct get<0, type_list > { @@ -221,10 +245,11 @@ struct get<0, type_list > template struct get > { - typedef typename get::type type; + typedef typename Tail::HeadType::type type; static const type value = get::value; }; + template struct arg_prod { static const typename NList::HeadType::type value = get<0, NList>::value * arg_prod::value; }; @@ -354,23 +379,51 @@ struct greater_equal_zero_op { template -inline bool array_apply_and_reduce(const array& a) { - EIGEN_STATIC_ASSERT(N >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE) - bool result = Reducer::run(Op::run(a[0]), Op::run(a[1])); - for (size_t i = 2; i < N; ++i) { - result = Reducer::run(result, Op::run(a[i])); +struct ArrayApplyAndReduce { + static inline bool run(const array& a) { + EIGEN_STATIC_ASSERT(N >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE); + bool result = Reducer::run(Op::run(a[0]), Op::run(a[1])); + for (size_t i = 2; i < N; ++i) { + result = Reducer::run(result, Op::run(a[i])); + } + return result; } - return result; +}; + +template +struct ArrayApplyAndReduce { + static inline bool run(const array& a) { + return Op::run(a[0]); + } +}; + +template +inline bool array_apply_and_reduce(const array& a) { + return ArrayApplyAndReduce::run(a); } template -inline bool array_zip_and_reduce(const array& a, const array& b) { - EIGEN_STATIC_ASSERT(N >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE) - bool result = Reducer::run(Op::run(a[0], b[0]), Op::run(a[1], b[1])); - for (size_t i = 2; i < N; ++i) { - result = Reducer::run(result, Op::run(a[i], b[i])); +struct ArrayZipAndReduce { + static inline bool run(const array& a, const array& b) { + EIGEN_STATIC_ASSERT(N >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE); + bool result = Reducer::run(Op::run(a[0], b[0]), Op::run(a[1], b[1])); + for (size_t i = 2; i < N; ++i) { + result = Reducer::run(result, Op::run(a[i], b[i])); + } + return result; } - return result; +}; + +template +struct ArrayZipAndReduce { + static inline bool run(const array& a, const array& b) { + return Op::run(a[0], b[0]); + } +}; + +template +inline bool array_zip_and_reduce(const array& a, const array& b) { + return ArrayZipAndReduce::run(a, b); } } // end namespace internal From 8a382aa119274efd2eb73b822ae7cd2afa128cc5 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 14 Jan 2015 15:33:11 -0800 Subject: [PATCH 1028/1103] Improved the resizing of tensors --- unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h | 12 +++--------- 1 file changed, 3 insertions(+), 9 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index aaec39756..dfe85602a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -114,16 +114,12 @@ class TensorStorage& dimensions() const {return m_dimensions;} - void conservativeResize(DenseIndex size, const array& nbDimensions) - { - m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, internal::array_prod(m_dimensions)); - m_dimensions = nbDimensions; - } void resize(DenseIndex size, const array& nbDimensions) { - if(size != internal::array_prod(m_dimensions)) + const DenseIndex currentSz = internal::array_prod(m_dimensions); + if(size != currentSz) { - internal::conditional_aligned_delete_auto(m_data, internal::array_prod(m_dimensions)); + internal::conditional_aligned_delete_auto(m_data, currentSz); if (size) m_data = internal::conditional_aligned_new_auto(size); else @@ -139,8 +135,6 @@ class TensorStorage Date: Wed, 14 Jan 2015 15:34:50 -0800 Subject: [PATCH 1029/1103] Ensured that each thread has it's own copy of the TensorEvaluator: this avoid race conditions when the evaluator calls a non thread safe functor, eg when generating random numbers. --- .../Eigen/CXX11/src/Tensor/TensorExecutor.h | 29 ++++++++++--------- 1 file changed, 15 insertions(+), 14 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index f27f643c1..d93fdd907 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -77,17 +77,17 @@ class TensorExecutor #ifdef EIGEN_USE_THREADS template struct EvalRange { - static void run(Evaluator* evaluator, const Index first, const Index last) { + static void run(Evaluator evaluator, const Index first, const Index last) { eigen_assert(last > first); for (Index i = first; i < last; ++i) { - evaluator->evalScalar(i); + evaluator.evalScalar(i); } } }; template struct EvalRange { - static void run(Evaluator* evaluator, const Index first, const Index last) { + static void run(Evaluator evaluator, const Index first, const Index last) { eigen_assert(last > first); Index i = first; @@ -96,12 +96,12 @@ struct EvalRange { eigen_assert(first % PacketSize == 0); Index lastPacket = last - (last % PacketSize); for (; i < lastPacket; i += PacketSize) { - evaluator->evalPacket(i); + evaluator.evalPacket(i); } } for (; i < last; ++i) { - evaluator->evalScalar(i); + evaluator.evalScalar(i); } } }; @@ -130,16 +130,17 @@ class TensorExecutor std::vector results; results.reserve(numblocks); for (int i = 0; i < numblocks; ++i) { - results.push_back(device.enqueue(&EvalRange::run, &evaluator, i*blocksize, (i+1)*blocksize)); - } - - for (int i = 0; i < numblocks; ++i) { - results[i].get(); + results.push_back(device.enqueue(&EvalRange::run, evaluator, i*blocksize, (i+1)*blocksize)); } if (numblocks * blocksize < size) { - EvalRange::run(&evaluator, numblocks * blocksize, size); + EvalRange::run(evaluator, numblocks * blocksize, size); } + + for (int i = 0; i < numblocks; ++i) { + get_when_ready(&results[i]); + } + } evaluator.cleanup(); } @@ -168,7 +169,8 @@ __launch_bounds__(1024) const Index PacketSize = unpacket_traits::size; const Index vectorized_step_size = step_size * PacketSize; const Index vectorized_size = (size / PacketSize) * PacketSize; - for (Index i = first_index * PacketSize; i < vectorized_size; i += vectorized_step_size) { + for (Index i = first_index * PacketSize; i < vectorized_size; + i += vectorized_step_size) { eval.evalPacket(i); } for (Index i = vectorized_size + first_index; i < size; i += step_size) { @@ -192,8 +194,7 @@ class TensorExecutor const int block_size = maxCudaThreadsPerBlock(); const Index size = array_prod(evaluator.dimensions()); - EigenMetaKernel, Index><<>>(evaluator, size); - assert(cudaGetLastError() == cudaSuccess); + LAUNCH_CUDA_KERNEL((EigenMetaKernel, Index>), num_blocks, block_size, 0, device, evaluator, size); } evaluator.cleanup(); } From f697df723798779bc29d9f7299bb5398767d5db0 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 14 Jan 2015 15:38:48 -0800 Subject: [PATCH 1030/1103] Improved support for RowMajor tensors Misc fixes and API cleanups. --- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 12 +- .../Eigen/CXX11/src/Tensor/TensorBase.h | 315 ++++++++++++++---- .../CXX11/src/Tensor/TensorBroadcasting.h | 166 ++++++++- .../Eigen/CXX11/src/Tensor/TensorChipping.h | 208 +++++++++--- .../CXX11/src/Tensor/TensorConcatenation.h | 75 ++++- .../src/Tensor/TensorContractionThreadPool.h | 6 +- .../CXX11/src/Tensor/TensorConvolution.h | 50 +-- .../Eigen/CXX11/src/Tensor/TensorEvalTo.h | 33 +- .../Eigen/CXX11/src/Tensor/TensorForcedEval.h | 24 +- .../src/Tensor/TensorForwardDeclarations.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorImagePatch.h | 142 ++++++-- .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 227 +++++++++---- .../Eigen/CXX11/src/Tensor/TensorPadding.h | 171 ++++++++-- .../Eigen/CXX11/src/Tensor/TensorPatch.h | 46 ++- .../Eigen/CXX11/src/Tensor/TensorShuffling.h | 54 ++- .../Eigen/CXX11/src/Tensor/TensorStriding.h | 175 ++++++++-- .../Eigen/CXX11/src/Tensor/TensorTraits.h | 53 +++ 17 files changed, 1405 insertions(+), 356 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index e973c00d3..93938bd1b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -33,6 +33,8 @@ struct traits > typedef typename RhsXprType::Nested RhsNested; typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; + static const std::size_t NumDimensions = internal::traits::NumDimensions; + static const int Layout = internal::traits::Layout; enum { Flags = 0, @@ -94,12 +96,18 @@ struct TensorEvaluator, Device> enum { IsAligned = TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, }; EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) : m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device) - { } + { + EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); + // The dimensions of the lhs and the rhs tensors should be equal to prevent + // overflows and ensure the result is fully initialized. + eigen_assert(dimensions_match(m_leftImpl.dimensions(), m_leftImpl.dimensions())); + } typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -114,7 +122,7 @@ struct TensorEvaluator, Device> } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { - eigen_assert(internal::dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions())); + eigen_assert(dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions())); m_leftImpl.evalSubExprsIfNeeded(NULL); // If the lhs provides raw access to its storage area (i.e. if m_leftImpl.data() returns a non // null value), attempt to evaluate the rhs expression in place. Returns true iff in place diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index f451a3c99..8860f622b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -25,77 +25,118 @@ template class TensorBase { public: - typedef typename internal::traits::Scalar Scalar; - typedef typename internal::traits::Index Index; - typedef Scalar CoeffReturnType; - typedef typename internal::packet_traits::type PacketReturnType; + typedef internal::traits DerivedTraits; + typedef typename DerivedTraits::Scalar Scalar; + typedef typename DerivedTraits::Index Index; + typedef typename internal::remove_const::type CoeffReturnType; + typedef typename internal::packet_traits::type PacketReturnType; + static const int NumDimensions = DerivedTraits::NumDimensions; - // Dimensions - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Index dimension(std::size_t n) const { return derived().dimensions()[n]; } - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE Index size() const { return internal::array_prod(derived().dimensions()); } + // Generic nullary operation support. + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseNullaryOp + nullaryExpr(const CustomNullaryOp& func) const { + return TensorCwiseNullaryOp(derived(), func); + } - // Nullary operators + // Coefficient-wise nullary operators EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> constant(const Scalar& value) const { - return TensorCwiseNullaryOp, const Derived> - (derived(), internal::scalar_constant_op(value)); + return nullaryExpr(internal::scalar_constant_op(value)); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseNullaryOp, const Derived> random() const { - return TensorCwiseNullaryOp, const Derived>(derived()); + return nullaryExpr(internal::UniformRandomGenerator()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseNullaryOp random() const { - return TensorCwiseNullaryOp(derived()); + return nullaryExpr(RandomGenerator()); + } + + // Generic unary operation support. + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp + unaryExpr(const CustomUnaryOp& func) const { + return TensorCwiseUnaryOp(derived(), func); } // Coefficient-wise unary operators EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - operator-() const { return derived(); } + operator-() const { + return unaryExpr(internal::scalar_opposite_op()); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - sqrt() const { return derived(); } + sqrt() const { + return unaryExpr(internal::scalar_sqrt_op()); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - square() const { return derived(); } + square() const { + return unaryExpr(internal::scalar_square_op()); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - inverse() const { return derived(); } + inverse() const { + return unaryExpr(internal::scalar_inverse_op()); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - exp() const { return derived(); } + exp() const { + return unaryExpr(internal::scalar_exp_op()); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - log() const { return derived(); } + log() const { + return unaryExpr(internal::scalar_log_op()); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - abs() const { return derived(); } + abs() const { + return unaryExpr(internal::scalar_abs_op()); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> pow(Scalar exponent) const { - return TensorCwiseUnaryOp, const Derived> - (derived(), internal::scalar_pow_op(exponent)); + return unaryExpr(internal::scalar_pow_op(exponent)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + operator+ (Scalar rhs) const { + return unaryExpr(internal::scalar_add_op(rhs)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + operator- (Scalar rhs) const { + EIGEN_STATIC_ASSERT((std::numeric_limits::is_signed || internal::is_same >::value), YOU_MADE_A_PROGRAMMING_MISTAKE); + return unaryExpr(internal::scalar_sub_op(rhs)); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> - operator * (Scalar scale) const { - return TensorCwiseUnaryOp, const Derived> - (derived(), internal::scalar_multiple_op(scale)); + operator* (Scalar rhs) const { + return unaryExpr(internal::scalar_multiple_op(rhs)); + } + + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + operator/ (Scalar rhs) const { + // EIGEN_STATIC_ASSERT(!std::numeric_limits::is_integer, YOU_MADE_A_PROGRAMMING_MISTAKE); + return unaryExpr(internal::scalar_quotient1_op(rhs)); } EIGEN_DEVICE_FUNC @@ -110,86 +151,106 @@ class TensorBase return cwiseMin(constant(threshold)); } - template EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE const TensorCwiseUnaryOp - unaryExpr(const CustomUnaryOp& func) const { - return TensorCwiseUnaryOp(derived(), func); - } - template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> cast() const { - return derived(); + return unaryExpr(internal::scalar_cast_op()); + } + + // Generic binary operation support. + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseBinaryOp + binaryExpr(const OtherDerived& other, const CustomBinaryOp& func) const { + return TensorCwiseBinaryOp(derived(), other, func); } // Coefficient-wise binary operators. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator+(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), internal::scalar_sum_op()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator-(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), internal::scalar_difference_op()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator*(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), internal::scalar_product_op()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator/(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), internal::scalar_quotient_op()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> cwiseMax(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), internal::scalar_max_op()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> cwiseMin(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), internal::scalar_min_op()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp + operator&&(const OtherDerived& other) const { + return binaryExpr(other.derived(), internal::scalar_boolean_and_op()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorCwiseBinaryOp + operator||(const OtherDerived& other) const { + return binaryExpr(other.derived(), internal::scalar_boolean_or_op()); } // Comparisons and tests. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator<(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), std::less()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator<=(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), std::less_equal()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator>(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), std::greater()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator>=(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), std::greater_equal()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator==(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), std::equal_to()); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseBinaryOp, const Derived, const OtherDerived> operator!=(const OtherDerived& other) const { - return TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return binaryExpr(other.derived(), std::not_equal_to()); + } + + // Coefficient-wise ternary operators. + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorSelectOp + select(const ThenDerived& thenTensor, const ElseDerived& elseTensor) const { + return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); } // Contractions. @@ -208,29 +269,72 @@ class TensorBase return TensorConvolutionOp(derived(), kernel.derived(), dims); } - // Coefficient-wise ternary operators. - template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - const TensorSelectOp - select(const ThenDerived& thenTensor, const ElseDerived& elseTensor) const { - return TensorSelectOp(derived(), thenTensor.derived(), elseTensor.derived()); - } - // Reductions. template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - const TensorReductionOp, const Dims, const Derived> + const TensorReductionOp, const Dims, const Derived> sum(const Dims& dims) const { - return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::SumReducer()); + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::SumReducer()); } + + const TensorReductionOp, const array, const Derived> + sum() const { + array in_dims; + for (int i = 0; i < NumDimensions; ++i) in_dims[i] = i; + return TensorReductionOp, const array, const Derived>(derived(), in_dims, internal::SumReducer()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - const TensorReductionOp, const Dims, const Derived> + const TensorReductionOp, const Dims, const Derived> + mean(const Dims& dims) const { + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::MeanReducer()); + } + + const TensorReductionOp, const array, const Derived> + mean() const { + array in_dims; + for (int i = 0; i < NumDimensions; ++i) in_dims[i] = i; + return TensorReductionOp, const array, const Derived>(derived(), in_dims, internal::MeanReducer()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReductionOp, const Dims, const Derived> + prod(const Dims& dims) const { + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::ProdReducer()); + } + + const TensorReductionOp, const array, const Derived> + prod() const { + array in_dims; + for (int i = 0; i < NumDimensions; ++i) in_dims[i] = i; + return TensorReductionOp, const array, const Derived>(derived(), in_dims, internal::ProdReducer()); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReductionOp, const Dims, const Derived> maximum(const Dims& dims) const { - return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::MaxReducer()); + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::MaxReducer()); } + + const TensorReductionOp, const array, const Derived> + maximum() const { + array in_dims; + for (int i = 0; i < NumDimensions; ++i) in_dims[i] = i; + return TensorReductionOp, const array, const Derived>(derived(), in_dims, internal::MaxReducer()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE - const TensorReductionOp, const Dims, const Derived> + const TensorReductionOp, const Dims, const Derived> minimum(const Dims& dims) const { - return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::MinReducer()); + return TensorReductionOp, const Dims, const Derived>(derived(), dims, internal::MinReducer()); } + + const TensorReductionOp, const array, const Derived> + minimum() const { + array in_dims; + for (int i = 0; i < NumDimensions; ++i) in_dims[i] = i; + return TensorReductionOp, const array, const Derived>(derived(), in_dims, internal::MinReducer()); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorReductionOp reduce(const Dims& dims, const Reducer& reducer) const { @@ -258,17 +362,44 @@ class TensorBase template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorImagePatchOp extract_image_patches() const { - return TensorImagePatchOp(derived(), Rows, Cols, 1, 1); + return TensorImagePatchOp(derived(), Rows, Cols, 1, 1, PADDING_SAME); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorImagePatchOp + extract_image_patches(const PaddingType padding_type) const { + return TensorImagePatchOp(derived(), Rows, Cols, 1, 1, padding_type); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorImagePatchOp + extract_image_patches(const Index stride, const PaddingType padding_type) const { + return TensorImagePatchOp(derived(), Rows, Cols, stride, stride, padding_type); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorImagePatchOp extract_image_patches(const Index patch_rows, const Index patch_cols, const Index row_stride = 1, const Index col_stride = 1) const { - return TensorImagePatchOp(derived(), patch_rows, patch_cols, row_stride, col_stride); + return TensorImagePatchOp(derived(), patch_rows, patch_cols, row_stride, col_stride, + PADDING_SAME); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorImagePatchOp + extract_image_patches(const Index patch_rows, const Index patch_cols, + const Index row_stride, const Index col_stride, + const PaddingType padding_type) const { + return TensorImagePatchOp(derived(), patch_rows, patch_cols, row_stride, col_stride, + padding_type); } // Morphing operators. + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorLayoutSwapOp + swap_layout() const { + return TensorLayoutSwapOp(derived()); + } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorReshapingOp reshape(const NewDimensions& newDimensions) const { @@ -279,10 +410,20 @@ class TensorBase slice(const StartIndices& startIndices, const Sizes& sizes) const { return TensorSlicingOp(derived(), startIndices, sizes); } - template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorChippingOp chip(const Index offset) const { - return TensorChippingOp(derived(), offset); + return TensorChippingOp(derived(), offset, DimId); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorChippingOp + chip(const Index offset, const Index dim) const { + return TensorChippingOp(derived(), offset, dim); + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorReverseOp + reverse(const ReverseDimensions& rev) const { + return TensorReverseOp(derived(), rev); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorPaddingOp @@ -308,21 +449,24 @@ class TensorBase protected: template friend class Tensor; + template friend class TensorVarDim; template friend class TensorBase; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Derived& derived() const { return *static_cast(this); } }; - template class TensorBase : public TensorBase { public: - typedef typename internal::traits::Scalar Scalar; - typedef typename internal::traits::Index Index; + typedef internal::traits DerivedTraits; + typedef typename DerivedTraits::Scalar Scalar; + typedef typename DerivedTraits::Index Index; typedef Scalar CoeffReturnType; typedef typename internal::packet_traits::type PacketReturnType; + static const int NumDimensions = DerivedTraits::NumDimensions; template friend class Tensor; + template friend class TensorVarDim; template friend class TensorBase; EIGEN_DEVICE_FUNC @@ -337,24 +481,43 @@ class TensorBase : public TensorBaserandom(); } + template EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setRandom() { + return derived() = this->template random(); + } + +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE Derived& setValues( + const typename internal::Initializer::InitList& vals) { + TensorEvaluator eval(derived(), DefaultDevice()); + internal::initialize_tensor(eval, vals); + return derived(); + } +#endif // EIGEN_HAS_VARIADIC_TEMPLATES template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator+=(const OtherDerived& other) { - return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return derived() = derived() + other.derived(); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator-=(const OtherDerived& other) { - return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return derived() = derived() - other.derived(); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator*=(const OtherDerived& other) { - return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return derived() = derived() * other.derived(); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator/=(const OtherDerived& other) { - return derived() = TensorCwiseBinaryOp, const Derived, const OtherDerived>(derived(), other.derived()); + return derived() = derived() / other.derived(); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorLayoutSwapOp + swap_layout() const { + return TensorLayoutSwapOp(derived()); + } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReshapingOp reshape(const NewDimensions& newDimensions) const { @@ -365,16 +528,26 @@ class TensorBase : public TensorBase(derived(), startIndices, sizes); } - template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp chip(const Index offset) const { - return TensorChippingOp(derived(), offset); + return TensorChippingOp(derived(), offset, DimId); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorChippingOp + chip(const Index offset, const Index dim) const { + return TensorChippingOp(derived(), offset, dim); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp shuffle(const Shuffle& shuffle) const { return TensorShufflingOp(derived(), shuffle); } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorStridingOp + stride(const Strides& strides) const { + return TensorStridingOp(derived(), strides); + } // Select the device on which to evaluate the expression. template diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h index 8cb41aec8..ef134adf2 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h @@ -30,6 +30,8 @@ struct traits > : public traits::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions; + static const int Layout = XprTraits::Layout; }; template @@ -91,6 +93,7 @@ struct TensorEvaluator, Device> enum { IsAligned = false, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -103,11 +106,20 @@ struct TensorEvaluator, Device> m_dimensions[i] = input_dims[i] * broadcast[i]; } - m_inputStrides[0] = 1; - m_outputStrides[0] = 1; - for (int i = 1; i < NumDims; ++i) { - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; - m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + if (Layout == ColMajor) { + m_inputStrides[0] = 1; + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } + } else { + m_inputStrides[NumDims-1] = 1; + m_outputStrides[NumDims-1] = 1; + for (int i = NumDims-2; i >= 0; --i) { + m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1]; + m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1]; + } } } @@ -125,16 +137,30 @@ struct TensorEvaluator, Device> m_impl.cleanup(); } + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffReturnType coeff(Index index) const + { + if (Layout == ColMajor) { + return coeffColMajor(index); + } else { + return coeffRowMajor(index); + } + } + // TODO: attempt to speed this up. The integer divisions and modulo are slow - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffColMajor(Index index) const { Index inputIndex = 0; for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; - if (internal::index_statically_eq()(i, 1)) { - eigen_assert(idx % m_impl.dimensions()[i] == 0); + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx < m_impl.dimensions()[i]); + inputIndex += idx * m_inputStrides[i]; } else { - inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx % m_impl.dimensions()[i] == 0); + } else { + inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + } } index -= idx * m_outputStrides[i]; } @@ -142,15 +168,59 @@ struct TensorEvaluator, Device> eigen_assert(index < m_impl.dimensions()[0]); inputIndex += index; } else { - inputIndex += (index % m_impl.dimensions()[0]); + if (internal::index_statically_eq()(0, 1)) { + eigen_assert(index % m_impl.dimensions()[0] == 0); + } else { + inputIndex += (index % m_impl.dimensions()[0]); + } } return m_impl.coeff(inputIndex); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffRowMajor(Index index) const + { + Index inputIndex = 0; + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx = index / m_outputStrides[i]; + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx < m_impl.dimensions()[i]); + inputIndex += idx * m_inputStrides[i]; + } else { + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx % m_impl.dimensions()[i] == 0); + } else { + inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + } + } + index -= idx * m_outputStrides[i]; + } + if (internal::index_statically_eq()(NumDims-1, 1)) { + eigen_assert(index < m_impl.dimensions()[NumDims-1]); + inputIndex += index; + } else { + if (internal::index_statically_eq()(NumDims-1, 1)) { + eigen_assert(index % m_impl.dimensions()[NumDims-1] == 0); + } else { + inputIndex += (index % m_impl.dimensions()[NumDims-1]); + } + } + return m_impl.coeff(inputIndex); + } + + template + EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnType packet(Index index) const + { + if (Layout == ColMajor) { + return packetColMajor(index); + } else { + return packetRowMajor(index); + } + } + // Ignore the LoadMode and always use unaligned loads since we can't guarantee // the alignment at compile time. template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetColMajor(Index index) const { const int packetSize = internal::unpacket_traits::size; EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) @@ -161,10 +231,15 @@ struct TensorEvaluator, Device> Index inputIndex = 0; for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; - if (internal::index_statically_eq()(i, 1)) { - eigen_assert(idx % m_impl.dimensions()[i] == 0); + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx < m_impl.dimensions()[i]); + inputIndex += idx * m_inputStrides[i]; } else { - inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx % m_impl.dimensions()[i] == 0); + } else { + inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + } } index -= idx * m_outputStrides[i]; } @@ -173,7 +248,12 @@ struct TensorEvaluator, Device> eigen_assert(index < m_impl.dimensions()[0]); innermostLoc = index; } else { - innermostLoc = index % m_impl.dimensions()[0]; + if (internal::index_statically_eq()(0, 1)) { + eigen_assert(innermostLoc % m_impl.dimensions()[0] == 0); + innermostLoc = 0; + } else { + innermostLoc = index % m_impl.dimensions()[0]; + } } inputIndex += innermostLoc; @@ -185,13 +265,67 @@ struct TensorEvaluator, Device> EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; values[0] = m_impl.coeff(inputIndex); for (int i = 1; i < packetSize; ++i) { - values[i] = coeff(originalIndex+i); + values[i] = coeffColMajor(originalIndex+i); } PacketReturnType rslt = internal::pload(values); return rslt; } } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetRowMajor(Index index) const + { + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); + + const Index originalIndex = index; + + Index inputIndex = 0; + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx = index / m_outputStrides[i]; + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx < m_impl.dimensions()[i]); + inputIndex += idx * m_inputStrides[i]; + } else { + if (internal::index_statically_eq()(i, 1)) { + eigen_assert(idx % m_impl.dimensions()[i] == 0); + } else { + inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i]; + } + } + index -= idx * m_outputStrides[i]; + } + Index innermostLoc; + if (internal::index_statically_eq()(NumDims-1, 1)) { + eigen_assert(index < m_impl.dimensions()[NumDims-1]); + innermostLoc = index; + } else { + if (internal::index_statically_eq()(NumDims-1, 1)) { + eigen_assert(innermostLoc % m_impl.dimensions()[NumDims-1] == 0); + innermostLoc = 0; + } else { + innermostLoc = index % m_impl.dimensions()[NumDims-1]; + } + } + inputIndex += innermostLoc; + + // Todo: this could be extended to the second dimension if we're not + // broadcasting alongside the first dimension, and so on. + if (innermostLoc + packetSize <= m_impl.dimensions()[NumDims-1]) { + return m_impl.template packet(inputIndex); + } else { + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + values[0] = m_impl.coeff(inputIndex); + for (int i = 1; i < packetSize; ++i) { + values[i] = coeffRowMajor(originalIndex+i); + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + } + + Scalar* data() const { return NULL; } protected: diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h index b862a8fd3..bc336e488 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h @@ -21,34 +21,61 @@ namespace Eigen { */ namespace internal { -template +template struct traits > : public traits { typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions - 1; + static const int Layout = XprTraits::Layout; }; -template +template struct eval, Eigen::Dense> { typedef const TensorChippingOp& type; }; -template +template struct nested, 1, typename eval >::type> { typedef TensorChippingOp type; }; +template +struct DimensionId +{ + DimensionId(DenseIndex dim) { + eigen_assert(dim == DimId); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim() const { + return DimId; + } +}; +template <> +struct DimensionId +{ + DimensionId(DenseIndex dim) : actual_dim(dim) { + eigen_assert(dim >= 0); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim() const { + return actual_dim; + } + private: + const DenseIndex actual_dim; +}; + + } // end namespace internal -template +template class TensorChippingOp : public TensorBase > { public: @@ -61,34 +88,39 @@ class TensorChippingOp : public TensorBase > typedef typename Eigen::internal::traits::StorageKind StorageKind; typedef typename Eigen::internal::traits::Index Index; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp(const XprType& expr, const Index offset) - : m_xpr(expr), m_offset(offset) {} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp(const XprType& expr, const Index offset, const Index dim) + : m_xpr(expr), m_offset(offset), m_dim(dim) { + } - EIGEN_DEVICE_FUNC - const Index offset() const { return m_offset; } + EIGEN_DEVICE_FUNC + const Index offset() const { return m_offset; } + EIGEN_DEVICE_FUNC + const Index dim() const { return m_dim.actualDim(); } - EIGEN_DEVICE_FUNC - const typename internal::remove_all::type& - expression() const { return m_xpr; } + EIGEN_DEVICE_FUNC + const typename internal::remove_all::type& + expression() const { return m_xpr; } - template - EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE TensorChippingOp& operator = (const OtherDerived& other) - { - typedef TensorAssignOp Assign; - Assign assign(*this, other); - internal::TensorExecutor::run(assign, DefaultDevice()); - return *this; - } + template + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorChippingOp& operator = (const OtherDerived& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + static const bool Vectorize = TensorEvaluator::PacketAccess; + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } protected: typename XprType::Nested m_xpr; const Index m_offset; + const internal::DimensionId m_dim; }; // Eval as rvalue -template +template struct TensorEvaluator, Device> { typedef TensorChippingOp XprType; @@ -96,41 +128,50 @@ struct TensorEvaluator, Device> static const int NumDims = NumInputDims-1; typedef typename XprType::Index Index; typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; enum { // Alignment can't be guaranteed at compile time since it depends on the // slice offsets. IsAligned = false, - PacketAccess = false, // not yet implemented + PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) - : m_impl(op.expression(), device), m_device(device) + : m_impl(op.expression(), device), m_dim(op.dim()), m_device(device) { // We could also support the case where NumInputDims==1 if needed. EIGEN_STATIC_ASSERT(NumInputDims >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE); - EIGEN_STATIC_ASSERT(NumInputDims > DimId, YOU_MADE_A_PROGRAMMING_MISTAKE); + eigen_assert(NumInputDims > m_dim.actualDim()); const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); int j = 0; for (int i = 0; i < NumInputDims; ++i) { - if (i != DimId) { + if (i != m_dim.actualDim()) { m_dimensions[j] = input_dims[i]; ++j; } } - m_stride = 1; - m_inputStride = 1; - for (int i = 0; i < DimId; ++i) { - m_stride *= input_dims[i]; - m_inputStride *= input_dims[i]; - } - m_inputStride *= input_dims[DimId]; - m_inputOffset = m_stride * op.offset(); + m_stride = 1; + m_inputStride = 1; + if (Layout == ColMajor) { + for (int i = 0; i < m_dim.actualDim(); ++i) { + m_stride *= input_dims[i]; + m_inputStride *= input_dims[i]; + } + } else { + for (int i = NumInputDims-1; i > m_dim.actualDim(); --i) { + m_stride *= input_dims[i]; + m_inputStride *= input_dims[i]; + } + } + m_inputStride *= input_dims[m_dim.actualDim()]; + m_inputOffset = m_stride * op.offset(); } - typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; @@ -150,16 +191,52 @@ struct TensorEvaluator, Device> return m_impl.coeff(srcCoeff(index)); } - /* to be done template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); - }*/ + if ((Layout == ColMajor && m_dim.actualDim() == 0) || + (Layout == RowMajor && m_dim.actualDim() == NumInputDims-1)) { + // m_stride is equal to 1, so let's avoid the integer division. + eigen_assert(m_stride == 1); + Index inputIndex = index * m_inputStride + m_inputOffset; + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + for (int i = 0; i < packetSize; ++i) { + values[i] = m_impl.coeff(inputIndex); + inputIndex += m_inputStride; + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } else if ((Layout == ColMajor && m_dim.actualDim() == NumInputDims - 1) || + (Layout == RowMajor && m_dim.actualDim() == 0)) { + // m_stride is aways greater than index, so let's avoid the integer division. + eigen_assert(m_stride > index); + return m_impl.template packet(index + m_inputOffset); + } else { + const Index idx = index / m_stride; + const Index rem = index - idx * m_stride; + if (rem + packetSize <= m_stride) { + Index inputIndex = idx * m_inputStride + m_inputOffset + rem; + return m_impl.template packet(inputIndex); + } else { + // Cross the stride boundary. Fallback to slow path. + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + for (int i = 0; i < packetSize; ++i) { + values[i] = coeff(index); + ++index; + } + PacketReturnType rslt = internal::pload(values); + return rslt; + } + } + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { Scalar* result = m_impl.data(); - if (DimId == NumDims && result) { + if (m_dim.actualDim() == NumDims && result) { return result + m_inputOffset; } else { return NULL; @@ -170,11 +247,13 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const { Index inputIndex; - if (DimId == 0) { + if ((Layout == ColMajor && m_dim.actualDim() == 0) || + (Layout == RowMajor && m_dim.actualDim() == NumInputDims-1)) { // m_stride is equal to 1, so let's avoid the integer division. eigen_assert(m_stride == 1); inputIndex = index * m_inputStride + m_inputOffset; - } else if (DimId == NumInputDims-1) { + } else if ((Layout == ColMajor && m_dim.actualDim() == NumInputDims-1) || + (Layout == RowMajor && m_dim.actualDim() == 0)) { // m_stride is aways greater than index, so let's avoid the integer division. eigen_assert(m_stride > index); inputIndex = index + m_inputOffset; @@ -192,12 +271,13 @@ struct TensorEvaluator, Device> Index m_inputOffset; Index m_inputStride; TensorEvaluator m_impl; + const internal::DimensionId m_dim; const Device& m_device; }; // Eval as lvalue -template +template struct TensorEvaluator, Device> : public TensorEvaluator, Device> { @@ -207,17 +287,17 @@ struct TensorEvaluator, Device> static const int NumDims = NumInputDims-1; typedef typename XprType::Index Index; typedef DSizes Dimensions; + typedef typename XprType::Scalar Scalar; enum { IsAligned = false, - PacketAccess = false, + PacketAccess = TensorEvaluator::PacketAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : Base(op, device) { } - typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketReturnType PacketReturnType; @@ -226,11 +306,45 @@ struct TensorEvaluator, Device> return this->m_impl.coeffRef(this->srcCoeff(index)); } - /* to be done template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writePacket(Index index, const PacketReturnType& x) { - } */ + static const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + + if ((this->Layout == ColMajor && this->m_dim.actualDim() == 0) || + (this->Layout == RowMajor && this->m_dim.actualDim() == NumInputDims-1)) { + // m_stride is equal to 1, so let's avoid the integer division. + eigen_assert(this->m_stride == 1); + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + internal::pstore(values, x); + Index inputIndex = index * this->m_inputStride + this->m_inputOffset; + for (int i = 0; i < packetSize; ++i) { + this->m_impl.coeffRef(inputIndex) = values[i]; + inputIndex += this->m_inputStride; + } + } else if ((this->Layout == ColMajor && this->m_dim.actualDim() == NumInputDims-1) || + (this->Layout == RowMajor && this->m_dim.actualDim() == 0)) { + // m_stride is aways greater than index, so let's avoid the integer division. + eigen_assert(this->m_stride > index); + this->m_impl.template writePacket(index + this->m_inputOffset, x); + } else { + const Index idx = index / this->m_stride; + const Index rem = index - idx * this->m_stride; + if (rem + packetSize <= this->m_stride) { + const Index inputIndex = idx * this->m_inputStride + this->m_inputOffset + rem; + this->m_impl.template writePacket(inputIndex, x); + } else { + // Cross stride boundary. Fallback to slow path. + EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; + internal::pstore(values, x); + for (int i = 0; i < packetSize; ++i) { + this->coeffRef(index) = values[i]; + ++index; + } + } + } + } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h index 74485b15b..fb4e7fb11 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h @@ -35,6 +35,8 @@ struct traits > typedef typename RhsXprType::Nested RhsNested; typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; + static const int NumDimensions = traits::NumDimensions; + static const int Layout = traits::Layout; enum { Flags = 0 }; }; @@ -103,11 +105,13 @@ struct TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device), m_axis(op.axis()) { + EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == TensorEvaluator::Layout || NumDims == 1), YOU_MADE_A_PROGRAMMING_MISTAKE); EIGEN_STATIC_ASSERT(NumDims == RightNumDims, YOU_MADE_A_PROGRAMMING_MISTAKE) eigen_assert(0 <= m_axis && m_axis < NumDims); const Dimensions& lhs_dims = m_leftImpl.dimensions(); @@ -127,13 +131,26 @@ struct TensorEvaluator= 0; --i) { + m_leftStrides[i] = m_leftStrides[i+1] * lhs_dims[i+1]; + m_rightStrides[i] = m_rightStrides[i+1] * rhs_dims[i+1]; + m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1]; + } } } @@ -159,25 +176,49 @@ struct TensorEvaluator subs; - for (int i = NumDims - 1; i > 0; --i) { - subs[i] = index / m_outputStrides[i]; - index -= subs[i] * m_outputStrides[i]; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + subs[i] = index / m_outputStrides[i]; + index -= subs[i] * m_outputStrides[i]; + } + subs[0] = index; + } else { + for (int i = 0; i < NumDims - 1; ++i) { + subs[i] = index / m_outputStrides[i]; + index -= subs[i] * m_outputStrides[i]; + } + subs[NumDims - 1] = index; } - subs[0] = index; const Dimensions& left_dims = m_leftImpl.dimensions(); if (subs[m_axis] < left_dims[m_axis]) { - Index left_index = subs[0]; - for (int i = 1; i < NumDims; ++i) { - left_index += (subs[i] % left_dims[i]) * m_leftStrides[i]; + Index left_index; + if (Layout == ColMajor) { + left_index = subs[0]; + for (int i = 1; i < NumDims; ++i) { + left_index += (subs[i] % left_dims[i]) * m_leftStrides[i]; + } + } else { + left_index = subs[NumDims - 1]; + for (int i = NumDims - 2; i >= 0; --i) { + left_index += (subs[i] % left_dims[i]) * m_leftStrides[i]; + } } return m_leftImpl.coeff(left_index); } else { subs[m_axis] -= left_dims[m_axis]; const Dimensions& right_dims = m_rightImpl.dimensions(); - Index right_index = subs[0]; - for (int i = 1; i < NumDims; ++i) { - right_index += (subs[i] % right_dims[i]) * m_rightStrides[i]; + Index right_index; + if (Layout == ColMajor) { + right_index = subs[0]; + for (int i = 1; i < NumDims; ++i) { + right_index += (subs[i] % right_dims[i]) * m_rightStrides[i]; + } + } else { + right_index = subs[NumDims - 1]; + for (int i = NumDims - 2; i >= 0; --i) { + right_index += (subs[i] % right_dims[i]) * m_rightStrides[i]; + } } return m_rightImpl.coeff(right_index); } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h index 5851e5adc..e358e6a3a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h @@ -93,10 +93,10 @@ struct TensorEvaluator right_dim_mapper_t; typedef array contract_t; - typedef array::size> left_nocontract_t; - typedef array::size> right_nocontract_t; + typedef array::size> left_nocontract_t; + typedef array::size> right_nocontract_t; - static const int NumDims = max_n_1::size; + static const int NumDims = internal::max_n_1::size; typedef DSizes Dimensions; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index 50cb10a33..aecef3313 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -144,9 +144,9 @@ template struct traits > { // Type promotion to handle the case where the types of the lhs and the rhs are different. - typedef typename internal::promote_storage_type::ret Scalar; - typedef typename internal::packet_traits::type Packet; + typedef typename promote_storage_type::ret Scalar; + typedef typename packet_traits::type Packet; typedef typename promote_storage_type::StorageKind, typename traits::StorageKind>::ret StorageKind; typedef typename promote_index_type::Index, @@ -155,6 +155,8 @@ struct traits > typedef typename KernelXprType::Nested RhsNested; typedef typename remove_reference::type _LhsNested; typedef typename remove_reference::type _RhsNested; + static const int NumDimensions = traits::NumDimensions; + static const int Layout = traits::Layout; enum { Flags = 0, @@ -227,11 +229,17 @@ struct TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, PacketAccess = TensorEvaluator::PacketAccess & TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_inputImpl(op.inputExpression(), device), m_kernelImpl(op.kernelExpression(), device), m_kernelArg(op.kernelExpression()), m_kernel(NULL), m_local_kernel(false), m_device(device) { + EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); + // Only column major tensors are supported for now. + EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); @@ -389,10 +397,6 @@ struct TensorEvaluator m_inputStride; array m_outputStride; @@ -421,7 +425,7 @@ struct GetKernelSize { } }; template <> -struct GetKernelSize { +struct GetKernelSize { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int operator() (const int kernelSize) const { return kernelSize; } @@ -610,11 +614,17 @@ struct TensorEvaluator::IsAligned & TensorEvaluator::IsAligned, PacketAccess = false, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const GpuDevice& device) : m_inputImpl(op.inputExpression(), device), m_kernelArg(op.kernelExpression()), m_kernelImpl(op.kernelExpression(), device), m_indices(op.indices()), m_buf(NULL), m_kernel(NULL), m_local_kernel(false), m_device(device) { + EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); + // Only column major tensors are supported for now. + EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); @@ -740,19 +750,17 @@ struct TensorEvaluator indexMapper(m_inputImpl.dimensions(), kernel_dims, indices); switch(kernel_size) { case 4: { - EigenConvolutionKernel1D, Index, InputDims, 4> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel1D, Index, InputDims, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 4, data); break; } case 7: { - EigenConvolutionKernel1D, Index, InputDims, 7> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel1D, Index, InputDims, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, 7, data); break; } default: { - EigenConvolutionKernel1D, Index, InputDims, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel1D, Index, InputDims, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, kernel_size, data); } } - cudaError_t error = cudaGetLastError(); - assert(error == cudaSuccess); break; } @@ -797,11 +805,11 @@ struct TensorEvaluator, Index, InputDims, 4, 7> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D, Index, InputDims, 4, 7>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, 7, data); break; } default: { - EigenConvolutionKernel2D, Index, InputDims, 4, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D, Index, InputDims, 4, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 4, kernel_size_y, data); break; } } @@ -810,23 +818,21 @@ struct TensorEvaluator, Index, InputDims, 7, 4> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D, Index, InputDims, 7, 4>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, 4, data); break; } default: { - EigenConvolutionKernel2D, Index, InputDims, 7, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D, Index, InputDims, 7, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, 7, kernel_size_y, data); break; } } break; } default: { - EigenConvolutionKernel2D, Index, InputDims, Eigen::Dynamic, Eigen::Dynamic> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel2D, Index, InputDims, Dynamic, Dynamic>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, kernel_size_x, kernel_size_y, data); break; } } - cudaError_t error = cudaGetLastError(); - assert(error == cudaSuccess); break; } @@ -858,9 +864,7 @@ struct TensorEvaluator kernel_dims(m_kernelImpl.dimensions()[0], m_kernelImpl.dimensions()[1], m_kernelImpl.dimensions()[2]); internal::IndexMapper indexMapper(m_inputImpl.dimensions(), kernel_dims, indices); - EigenConvolutionKernel3D, Index, InputDims> <<>>(m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data); - cudaError_t error = cudaGetLastError(); - assert(error == cudaSuccess); + LAUNCH_CUDA_KERNEL((EigenConvolutionKernel3D, Index, InputDims>), num_blocks, block_size, shared_mem, m_device, m_inputImpl, indexMapper, m_kernel, numP, numX, maxX, numY, maxY, numZ, maxZ, kernel_size_x, kernel_size_y, kernel_size_z, data); break; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h index ce9d73578..93ebbe277 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h @@ -25,11 +25,14 @@ struct traits > { // Type promotion to handle the case where the types of the lhs and the rhs are different. typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions; + static const int Layout = XprTraits::Layout; enum { Flags = 0, @@ -60,24 +63,24 @@ class TensorEvalToOp : public TensorBase > typedef typename Eigen::internal::traits::Scalar Scalar; typedef typename Eigen::internal::traits::Packet Packet; typedef typename Eigen::NumTraits::Real RealScalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename internal::remove_const::type CoeffReturnType; + typedef typename internal::remove_const::type PacketReturnType; typedef typename Eigen::internal::nested::type Nested; typedef typename Eigen::internal::traits::StorageKind StorageKind; typedef typename Eigen::internal::traits::Index Index; - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvalToOp(Scalar* buffer, const XprType& expr) + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvalToOp(CoeffReturnType* buffer, const XprType& expr) : m_xpr(expr), m_buffer(buffer) {} EIGEN_DEVICE_FUNC const typename internal::remove_all::type& expression() const { return m_xpr; } - EIGEN_DEVICE_FUNC Scalar* buffer() const { return m_buffer; } + EIGEN_DEVICE_FUNC CoeffReturnType* buffer() const { return m_buffer; } protected: typename XprType::Nested m_xpr; - Scalar* m_buffer; + CoeffReturnType* m_buffer; }; @@ -93,6 +96,8 @@ struct TensorEvaluator, Device> enum { IsAligned = true, PacketAccess = true, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -103,12 +108,12 @@ struct TensorEvaluator, Device> } typedef typename XprType::Index Index; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename internal::remove_const::type CoeffReturnType; + typedef typename internal::remove_const::type PacketReturnType; EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) { m_impl.evalSubExprsIfNeeded(NULL); return true; } @@ -117,7 +122,7 @@ struct TensorEvaluator, Device> m_buffer[i] = m_impl.coeff(i); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void evalPacket(Index i) { - internal::pstoret(m_buffer + i, m_impl.template packet::IsAligned ? Aligned : Unaligned>(i)); + internal::pstoret(m_buffer + i, m_impl.template packet::IsAligned ? Aligned : Unaligned>(i)); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { @@ -135,12 +140,12 @@ struct TensorEvaluator, Device> return internal::ploadt(m_buffer + index); } - Scalar* data() const { return NULL; } + CoeffReturnType* data() const { return NULL; } private: TensorEvaluator m_impl; const Device& m_device; - Scalar* m_buffer; + CoeffReturnType* m_buffer; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h index cb14cc7f7..a9501336e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h @@ -25,11 +25,14 @@ struct traits > { // Type promotion to handle the case where the types of the lhs and the rhs are different. typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; typedef typename traits::StorageKind StorageKind; typedef typename traits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions; + static const int Layout = XprTraits::Layout; enum { Flags = 0, @@ -59,8 +62,8 @@ class TensorForcedEvalOp : public TensorBase > typedef typename Eigen::internal::traits::Scalar Scalar; typedef typename Eigen::internal::traits::Packet Packet; typedef typename Eigen::NumTraits::Real RealScalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename internal::remove_const::type CoeffReturnType; + typedef typename internal::remove_const::type PacketReturnType; typedef typename Eigen::internal::nested::type Nested; typedef typename Eigen::internal::traits::StorageKind StorageKind; typedef typename Eigen::internal::traits::Index Index; @@ -88,6 +91,7 @@ struct TensorEvaluator, Device> enum { IsAligned = true, PacketAccess = (internal::packet_traits::size > 1), + Layout = TensorEvaluator::Layout, }; EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) @@ -100,10 +104,16 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); } - EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) { + EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) { m_impl.evalSubExprsIfNeeded(NULL); - m_buffer = (Scalar*)m_device.allocate(m_impl.dimensions().TotalSize() * sizeof(Scalar)); - + const Index numValues = m_impl.dimensions().TotalSize(); + m_buffer = (CoeffReturnType*)m_device.allocate(numValues * sizeof(CoeffReturnType)); + // Should initialize the memory in case we're dealing with non POD types. + if (!internal::is_arithmetic::value) { + for (Index i = 0; i < numValues; ++i) { + new(m_buffer+i) CoeffReturnType(); + } + } typedef TensorEvalToOp EvalTo; EvalTo evalToTmp(m_buffer, m_op); internal::TensorExecutor::PacketAccess>::run(evalToTmp, m_device); @@ -132,7 +142,7 @@ struct TensorEvaluator, Device> TensorEvaluator m_impl; const ArgType m_op; const Device& m_device; - Scalar* m_buffer; + CoeffReturnType* m_buffer; }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index 85599ccfd..7bec2b10a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h @@ -29,9 +29,11 @@ template cla template class TensorPatchOp; template class TensorImagePatchOp; template class TensorBroadcastingOp; -template class TensorChippingOp; +template class TensorChippingOp; template class TensorReshapingOp; +template class TensorLayoutSwapOp; template class TensorSlicingOp; +template class TensorReverseOp; template class TensorPaddingOp; template class TensorShufflingOp; template class TensorStridingOp; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h index 0dfb6913b..585ebc778 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h @@ -37,6 +37,8 @@ struct traits > : public traits typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions + 1; + static const int Layout = XprTraits::Layout; }; template @@ -53,8 +55,6 @@ struct nested, 1, typename eval class TensorImagePatchOp : public TensorBase, ReadOnlyAccessors> { @@ -69,9 +69,11 @@ class TensorImagePatchOp : public TensorBase::Index Index; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols, - DenseIndex row_strides, DenseIndex col_strides) + DenseIndex row_strides, DenseIndex col_strides, + PaddingType padding_type) : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols), - m_row_strides(row_strides), m_col_strides(col_strides){} + m_row_strides(row_strides), m_col_strides(col_strides), + m_padding_type(padding_type) {} EIGEN_DEVICE_FUNC DenseIndex patch_rows() const { return m_patch_rows; } @@ -81,6 +83,8 @@ class TensorImagePatchOp : public TensorBase::type& @@ -92,6 +96,7 @@ class TensorImagePatchOp : public TensorBase, Device> enum { IsAligned = false, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = NumDims == 5, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_impl(op.expression(), device) { + // Only column major tensors are supported for now. + EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(NumDims >= 4, YOU_MADE_A_PROGRAMMING_MISTAKE); const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); + + // Caches a few variables. + m_inputRows = input_dims[1]; + m_inputCols = input_dims[2]; + + m_row_strides = op.row_strides(); + m_col_strides = op.col_strides(); + + // We only support same strides for both dimensions and square patches. + eigen_assert(m_row_strides == m_col_strides); + + switch (op.padding_type()) { + case PADDING_VALID: + m_outputRows = ceil((m_inputRows - op.patch_rows() + 1.f) / static_cast(m_row_strides)); + m_outputCols = ceil((m_inputCols - op.patch_cols() + 1.f) / static_cast(m_col_strides)); + // Calculate the padding + m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + op.patch_rows() - m_inputRows) / 2; + m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + op.patch_cols() - m_inputCols) / 2; + break; + case PADDING_SAME: + m_outputRows = ceil(m_inputRows / static_cast(m_row_strides)); + m_outputCols = ceil(m_inputCols / static_cast(m_col_strides)); + // Calculate the padding + m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + op.patch_rows() - m_inputRows) / 2; + m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + op.patch_cols() - m_inputCols) / 2; + break; + default: + eigen_assert(false && "unexpected padding"); + } + + // Dimensions for result of extraction. + // 0: depth + // 1: patch_rows + // 2: patch_cols + // 3: number of patches + // 4 and beyond: anything else (such as batch). m_dimensions[0] = input_dims[0]; m_dimensions[1] = op.patch_rows(); m_dimensions[2] = op.patch_cols(); - m_dimensions[3] = ceilf(static_cast(input_dims[1]) / op.row_strides()) * - ceilf(static_cast(input_dims[2]) / op.col_strides()); + m_dimensions[3] = m_outputRows * m_outputCols; for (int i = 4; i < NumDims; ++i) { m_dimensions[i] = input_dims[i-1]; } + // Strides for moving the patch in various dimensions. m_colStride = m_dimensions[1]; m_patchStride = m_colStride * m_dimensions[2] * m_dimensions[0]; m_otherStride = m_patchStride * m_dimensions[3]; - m_inputRows = input_dims[1]; - m_inputCols = input_dims[2]; - - m_rowInputStride = input_dims[0] * op.row_strides(); - m_colInputStride = input_dims[0] * input_dims[1] * op.col_strides(); + // Strides for navigating through the input tensor. + m_rowInputStride = input_dims[0]; + m_colInputStride = input_dims[0] * input_dims[1]; m_patchInputStride = input_dims[0] * input_dims[1] * input_dims[2]; - m_rowPaddingTop = op.patch_rows() / 2; - m_colPaddingLeft = op.patch_cols() / 2; - + // Fast representations of different variables. m_fastOtherStride = internal::TensorIntDivisor(m_otherStride); m_fastPatchStride = internal::TensorIntDivisor(m_patchStride); m_fastColStride = internal::TensorIntDivisor(m_colStride); - m_fastInputRows = internal::TensorIntDivisor(m_inputRows); + // Number of patches in the width dimension. + m_fastOutputRows = internal::TensorIntDivisor(m_outputRows); m_fastDimZero = internal::TensorIntDivisor(m_dimensions[0]); } @@ -162,26 +205,29 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { - // Find the location of the first element of the patch. + // Patch index corresponding to the passed in index. const Index patchIndex = index / m_fastPatchStride; // Find the offset of the element wrt the location of the first element. const Index patchOffset = (index - patchIndex * m_patchStride) / m_fastDimZero; + // Other ways to index this element. const Index otherIndex = (NumDims == 4) ? 0 : index / m_fastOtherStride; const Index patch2DIndex = (NumDims == 4) ? patchIndex : (index - otherIndex * m_otherStride) / m_fastPatchStride; - const Index colIndex = patch2DIndex / m_fastInputRows; + const Index colIndex = patch2DIndex / m_fastOutputRows; const Index colOffset = patchOffset / m_fastColStride; - const Index inputCol = colIndex + colOffset - m_colPaddingLeft; + // Calculate col index in the input original tensor. + const Index inputCol = colIndex * m_col_strides + colOffset - m_colPaddingLeft; if (inputCol < 0 || inputCol >= m_inputCols) { return Scalar(0); } - const Index rowIndex = patch2DIndex - colIndex * m_inputRows; // m_rowStride is always 1 + const Index rowIndex = patch2DIndex - colIndex * m_outputRows; const Index rowOffset = patchOffset - colOffset * m_colStride; - const Index inputRow = rowIndex + rowOffset - m_rowPaddingTop; + // Calculate row index in the original input tensor. + const Index inputRow = rowIndex * m_row_strides + rowOffset - m_rowPaddingTop; if (inputRow < 0 || inputRow >= m_inputRows) { return Scalar(0); } @@ -214,20 +260,24 @@ struct TensorEvaluator, Device> const Index patch2DIndex = (NumDims == 4) ? patchIndex : (indices[0] - otherIndex * m_otherStride) / m_fastPatchStride; eigen_assert(patch2DIndex == (indices[1] - otherIndex * m_otherStride) / m_fastPatchStride); - const Index colIndex = patch2DIndex / m_fastInputRows; + const Index colIndex = patch2DIndex / m_fastOutputRows; const Index colOffsets[2] = {patchOffsets[0] / m_fastColStride, patchOffsets[1] / m_fastColStride}; - const Index inputCols[2] = {colIndex + colOffsets[0] - m_colPaddingLeft, colIndex + colOffsets[1] - m_colPaddingLeft}; + // Calculate col indices in the original input tensor. + const Index inputCols[2] = {colIndex * m_col_strides + colOffsets[0] - + m_colPaddingLeft, colIndex * m_col_strides + colOffsets[1] - m_colPaddingLeft}; if (inputCols[1] < 0 || inputCols[0] >= m_inputCols) { // all zeros return internal::pset1(Scalar(0)); } if (inputCols[0] == inputCols[1]) { - const Index rowIndex = patch2DIndex - colIndex * m_inputRows; + const Index rowIndex = patch2DIndex - colIndex * m_outputRows; const Index rowOffsets[2] = {patchOffsets[0] - colOffsets[0]*m_colStride, patchOffsets[1] - colOffsets[1]*m_colStride}; eigen_assert(rowOffsets[0] <= rowOffsets[1]); - const Index inputRows[2] = {rowIndex + rowOffsets[0] - m_rowPaddingTop, rowIndex + rowOffsets[1] - m_rowPaddingTop}; + // Calculate col indices in the original input tensor. + const Index inputRows[2] = {rowIndex * m_row_strides + rowOffsets[0] - + m_rowPaddingTop, rowIndex * m_row_strides + rowOffsets[1] - m_rowPaddingTop}; if (inputRows[1] < 0 || inputRows[0] >= m_inputRows) { // all zeros @@ -247,6 +297,43 @@ struct TensorEvaluator, Device> Scalar* data() const { return NULL; } + const TensorEvaluator& impl() const { return m_impl; } + + Index rowPaddingTop() const { return m_rowPaddingTop; } + Index colPaddingLeft() const { return m_colPaddingLeft; } + Index outputRows() const { return m_outputRows; } + Index outputCols() const { return m_outputCols; } + Index userRowStride() const { return m_row_strides; } + Index userColStride() const { return m_col_strides; } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array& coords) const + { + // Location of the first element of the patch. + // 0: d, 1: patch_rows, 2: patch_cols, 3: number of patches, 4: number of batches + const Index patchIndex = coords[3]; + + array inputCoords; + inputCoords[0] = coords[0]; // depth + inputCoords[1] = patchIndex / m_inputCols + coords[1] - m_rowPaddingTop; + inputCoords[2] = patchIndex - patchIndex / m_inputCols * m_inputCols + coords[2] - m_colPaddingLeft; + inputCoords[3] = coords[4]; // batch + // If the computed coordinates are outside the original image perimeter, return 0. + if (inputCoords[1] < 0 || inputCoords[1] >= m_inputRows || + inputCoords[2] < 0 || inputCoords[2] >= m_inputCols) { + return Scalar(0); + } + if (TensorEvaluator::CoordAccess) { + return m_impl.coeff(inputCoords); + } else { + Index inputIndex = + inputCoords[3] * m_patchInputStride + + inputCoords[2] * m_colInputStride + + inputCoords[1] * m_rowInputStride + + inputCoords[0]; + return m_impl.coeff(inputIndex); + } + } + protected: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const { @@ -264,6 +351,8 @@ struct TensorEvaluator, Device> Index m_otherStride; Index m_patchStride; Index m_colStride; + Index m_row_strides; + Index m_col_strides; internal::TensorIntDivisor m_fastOtherStride; internal::TensorIntDivisor m_fastPatchStride; internal::TensorIntDivisor m_fastColStride; @@ -275,10 +364,13 @@ struct TensorEvaluator, Device> Index m_inputRows; Index m_inputCols; + Index m_outputRows; + Index m_outputCols; + Index m_rowPaddingTop; Index m_colPaddingLeft; - internal::TensorIntDivisor m_fastInputRows; + internal::TensorIntDivisor m_fastOutputRows; internal::TensorIntDivisor m_fastDimZero; TensorEvaluator m_impl; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 33849ed3e..23b595ac3 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -24,11 +24,14 @@ template struct traits > : public traits { typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = array_size::value; + static const int Layout = XprTraits::Layout; }; template @@ -54,8 +57,8 @@ class TensorReshapingOp : public TensorBase::Scalar Scalar; typedef typename Eigen::internal::traits::Packet Packet; typedef typename Eigen::NumTraits::Real RealScalar; - typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename internal::remove_const::type CoeffReturnType; + typedef typename internal::remove_const::type PacketReturnType; typedef typename Eigen::internal::nested::type Nested; typedef typename Eigen::internal::traits::StorageKind StorageKind; typedef typename Eigen::internal::traits::Index Index; @@ -96,11 +99,17 @@ struct TensorEvaluator, Device> enum { IsAligned = TensorEvaluator::IsAligned, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_impl(op.expression(), device), m_dimensions(op.dimensions()) - { } + { + // The total size of the reshaped tensor must be equal to the total size + // of the input tensor. + eigen_assert(internal::array_prod(m_impl.dimensions()) == internal::array_prod(op.dimensions())); + } typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; @@ -109,7 +118,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) { return m_impl.evalSubExprsIfNeeded(data); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() { @@ -127,7 +136,9 @@ struct TensorEvaluator, Device> return m_impl.template packet(index); } - Scalar* data() const { return m_impl.data(); } + CoeffReturnType* data() const { return m_impl.data(); } + + const TensorEvaluator& impl() const { return m_impl; } protected: TensorEvaluator m_impl; @@ -148,6 +159,8 @@ template enum { IsAligned = TensorEvaluator::IsAligned, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -183,11 +196,14 @@ template struct traits > : public traits { typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = array_size::value; + static const int Layout = XprTraits::Layout; }; template @@ -260,6 +276,8 @@ struct TensorEvaluator, Devi // slice offsets and sizes. IsAligned = /*TensorEvaluator::IsAligned*/false, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = TensorEvaluator::CoordAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -270,22 +288,30 @@ struct TensorEvaluator, Devi } const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; - } else { - m_inputStrides[0] = 1; - } - } - const Sizes& output_dims = op.sizes(); - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { + if (Layout == ColMajor) { + m_inputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + } + + m_outputStrides[0] = 1; + m_fastOutputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1]; m_fastOutputStrides[i] = internal::TensorIntDivisor(m_outputStrides[i]); - } else { - m_outputStrides[0] = 1; - m_fastOutputStrides[0] = 1; + } + } else { + m_inputStrides[NumDims-1] = 1; + for (int i = NumDims - 2; i >= 0; --i) { + m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1]; + } + + m_outputStrides[NumDims-1] = 1; + m_fastOutputStrides[NumDims-1] = 1; + for (int i = NumDims - 2; i >= 0; --i) { + m_outputStrides[i] = m_outputStrides[i+1] * output_dims[i+1]; + m_fastOutputStrides[i] = internal::TensorIntDivisor(m_outputStrides[i]); } } } @@ -299,14 +325,23 @@ struct TensorEvaluator, Devi EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) { m_impl.evalSubExprsIfNeeded(NULL); if (internal::is_arithmetic::value && data && m_impl.data()) { Index contiguous_values = 1; - for (int i = 0; i < NumDims; ++i) { - contiguous_values *= dimensions()[i]; - if (dimensions()[i] != m_impl.dimensions()[i]) { - break; + if (Layout == ColMajor) { + for (int i = 0; i < NumDims; ++i) { + contiguous_values *= dimensions()[i]; + if (dimensions()[i] != m_impl.dimensions()[i]) { + break; + } + } + } else { + for (int i = NumDims-1; i >= 0; --i) { + contiguous_values *= dimensions()[i]; + if (dimensions()[i] != m_impl.dimensions()[i]) { + break; + } } } // Use memcpy if it's going to be faster than using the regular evaluation. @@ -340,16 +375,29 @@ struct TensorEvaluator, Devi Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / m_fastOutputStrides[i]; - const Index idx1 = indices[1] / m_fastOutputStrides[i]; - inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; - inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; - indices[0] -= idx0 * m_outputStrides[i]; - indices[1] -= idx1 * m_outputStrides[i]; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / m_fastOutputStrides[i]; + const Index idx1 = indices[1] / m_fastOutputStrides[i]; + inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; + inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += (indices[0] + m_offsets[0]); + inputIndices[1] += (indices[1] + m_offsets[0]); + } else { + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx0 = indices[0] / m_fastOutputStrides[i]; + const Index idx1 = indices[1] / m_fastOutputStrides[i]; + inputIndices[0] += (idx0 + m_offsets[i]) * m_inputStrides[i]; + inputIndices[1] += (idx1 + m_offsets[i]) * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += (indices[0] + m_offsets[NumDims-1]); + inputIndices[1] += (indices[1] + m_offsets[NumDims-1]); } - inputIndices[0] += (indices[0] + m_offsets[0]); - inputIndices[1] += (indices[1] + m_offsets[0]); if (inputIndices[1] - inputIndices[0] == packetSize - 1) { PacketReturnType rslt = m_impl.template packet(inputIndices[0]); return rslt; @@ -366,20 +414,44 @@ struct TensorEvaluator, Devi } } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array& coords) + { + array inputCoords; + for (int i = 0; i < NumDims; ++i) { + inputCoords = coords[i] + this->m_offsets[i]; + } + return m_impl.coeff(inputCoords); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType* data() const { Scalar* result = m_impl.data(); if (result) { Index offset = 0; - for (int i = 0; i < NumDims; ++i) { - if (m_dimensions[i] != m_impl.dimensions()[i]) { - offset += m_offsets[i] * m_inputStrides[i]; - for (int j = i+1; j < NumDims; ++j) { - if (m_dimensions[j] > 1) { - return NULL; + if (Layout == ColMajor) { + for (int i = 0; i < NumDims; ++i) { + if (m_dimensions[i] != m_impl.dimensions()[i]) { + offset += m_offsets[i] * m_inputStrides[i]; + for (int j = i+1; j < NumDims; ++j) { + if (m_dimensions[j] > 1) { + return NULL; + } + offset += m_offsets[j] * m_inputStrides[j]; } - offset += m_offsets[j] * m_inputStrides[j]; + break; + } + } + } else { + for (int i = NumDims - 1; i >= 0; --i) { + if (m_dimensions[i] != m_impl.dimensions()[i]) { + offset += m_offsets[i] * m_inputStrides[i]; + for (int j = i-1; j >= 0; --j) { + if (m_dimensions[j] > 1) { + return NULL; + } + offset += m_offsets[j] * m_inputStrides[j]; + } + break; } - break; } } return result + offset; @@ -391,12 +463,21 @@ struct TensorEvaluator, Devi EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const { Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_fastOutputStrides[i]; - inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; - index -= idx * m_outputStrides[i]; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_fastOutputStrides[i]; + inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + inputIndex += (index + m_offsets[0]); + } else { + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx = index / m_fastOutputStrides[i]; + inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + inputIndex += (index + m_offsets[NumDims-1]); } - inputIndex += (index + m_offsets[0]); return inputIndex; } @@ -422,6 +503,8 @@ struct TensorEvaluator, Device> enum { IsAligned = /*TensorEvaluator::IsAligned*/false, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = TensorEvaluator::CoordAccess, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -445,16 +528,29 @@ struct TensorEvaluator, Device> const int packetSize = internal::unpacket_traits::size; Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / this->m_fastOutputStrides[i]; - const Index idx1 = indices[1] / this->m_fastOutputStrides[i]; - inputIndices[0] += (idx0 + this->m_offsets[i]) * this->m_inputStrides[i]; - inputIndices[1] += (idx1 + this->m_offsets[i]) * this->m_inputStrides[i]; - indices[0] -= idx0 * this->m_outputStrides[i]; - indices[1] -= idx1 * this->m_outputStrides[i]; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / this->m_fastOutputStrides[i]; + const Index idx1 = indices[1] / this->m_fastOutputStrides[i]; + inputIndices[0] += (idx0 + this->m_offsets[i]) * this->m_inputStrides[i]; + inputIndices[1] += (idx1 + this->m_offsets[i]) * this->m_inputStrides[i]; + indices[0] -= idx0 * this->m_outputStrides[i]; + indices[1] -= idx1 * this->m_outputStrides[i]; + } + inputIndices[0] += (indices[0] + this->m_offsets[0]); + inputIndices[1] += (indices[1] + this->m_offsets[0]); + } else { + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx0 = indices[0] / this->m_fastOutputStrides[i]; + const Index idx1 = indices[1] / this->m_fastOutputStrides[i]; + inputIndices[0] += (idx0 + this->m_offsets[i]) * this->m_inputStrides[i]; + inputIndices[1] += (idx1 + this->m_offsets[i]) * this->m_inputStrides[i]; + indices[0] -= idx0 * this->m_outputStrides[i]; + indices[1] -= idx1 * this->m_outputStrides[i]; + } + inputIndices[0] += (indices[0] + this->m_offsets[NumDims-1]); + inputIndices[1] += (indices[1] + this->m_offsets[NumDims-1]); } - inputIndices[0] += (indices[0] + this->m_offsets[0]); - inputIndices[1] += (indices[1] + this->m_offsets[0]); if (inputIndices[1] - inputIndices[0] == packetSize - 1) { this->m_impl.template writePacket(inputIndices[0], x); } @@ -468,6 +564,15 @@ struct TensorEvaluator, Device> } } } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(const array& coords) + { + array inputCoords; + for (int i = 0; i < NumDims; ++i) { + inputCoords = coords[i] + this->m_offsets[i]; + } + return this->m_impl.coeffRef(inputCoords); + } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h index d6347b054..9b14e01f4 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h @@ -24,11 +24,14 @@ template struct traits > : public traits { typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions; + static const int Layout = XprTraits::Layout; }; template @@ -88,6 +91,8 @@ struct TensorEvaluator, Device enum { IsAligned = false, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = true, }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -99,13 +104,23 @@ struct TensorEvaluator, Device m_dimensions[i] += m_padding[i].first + m_padding[i].second; } const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - m_inputStrides[0] = 1; - m_outputStrides[0] = 1; - for (int i = 1; i < NumDims; ++i) { - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; - m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + if (Layout == ColMajor) { + m_inputStrides[0] = 1; + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + } + m_outputStrides[NumDims] = m_outputStrides[NumDims-1] * m_dimensions[NumDims-1]; + } else { + m_inputStrides[NumDims - 1] = 1; + m_outputStrides[NumDims] = 1; + for (int i = NumDims - 2; i >= 0; --i) { + m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1]; + m_outputStrides[i+1] = m_outputStrides[i+2] * m_dimensions[i+1]; + } + m_outputStrides[0] = m_outputStrides[1] * m_dimensions[0]; } - m_outputStrides[NumDims] = m_outputStrides[NumDims-1] * m_dimensions[NumDims-1]; } typedef typename XprType::Scalar Scalar; @@ -126,23 +141,84 @@ struct TensorEvaluator, Device { eigen_assert(index < dimensions().TotalSize()); Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_outputStrides[i]; - if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { + return Scalar(0); + } + inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + if (index < m_padding[0].first || index >= m_dimensions[0] - m_padding[0].second) { return Scalar(0); } - inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; - index -= idx * m_outputStrides[i]; + inputIndex += (index - m_padding[0].first); + } else { + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx = index / m_outputStrides[i+1]; + if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { + return Scalar(0); + } + inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; + index -= idx * m_outputStrides[i+1]; + } + if (index < m_padding[NumDims-1].first || + index >= m_dimensions[NumDims-1] - m_padding[NumDims-1].second) { + return Scalar(0); + } + inputIndex += (index - m_padding[NumDims-1].first); } - if (index < m_padding[0].first || index >= m_dimensions[0] - m_padding[0].second) { - return Scalar(0); - } - inputIndex += (index - m_padding[0].first); return m_impl.coeff(inputIndex); } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + if (Layout == ColMajor) { + return packetColMajor(index); + } + return packetRowMajor(index); + } + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array& coords) const + { + Index inputIndex; + if (Layout == ColMajor) { + const Index idx = coords[0]; + if (idx < m_padding[0].first || idx >= m_dimensions[0] - m_padding[0].second) { + return Scalar(0); + } + inputIndex = idx - m_padding[0].first; + for (int i = 1; i < NumDims; ++i) { + const Index idx = coords[i]; + if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { + return Scalar(0); + } + inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; + } + } else { + const Index idx = coords[NumDims-1]; + if (idx < m_padding[NumDims-1].first || idx >= m_dimensions[NumDims-1] - m_padding[NumDims-1].second) { + return Scalar(0); + } + inputIndex = idx - m_padding[NumDims-1].first; + for (int i = NumDims - 2; i >= 0; --i) { + const Index idx = coords[i]; + if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { + return Scalar(0); + } + inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; + } + } + return m_impl.coeff(inputIndex); + } + + Scalar* data() const { return NULL; } + + protected: + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetColMajor(Index index) const { const int packetSize = internal::unpacket_traits::size; EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) @@ -200,9 +276,64 @@ struct TensorEvaluator, Device return packetWithPossibleZero(initialIndex); } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetRowMajor(Index index) const + { + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < dimensions().TotalSize()); - protected: + const Index initialIndex = index; + Index inputIndex = 0; + + for (int i = 0; i < NumDims - 1; ++i) { + const Index first = index; + const Index last = index + packetSize - 1; + const Index lastPaddedLeft = m_padding[i].first * m_outputStrides[i+1]; + const Index firstPaddedRight = (m_dimensions[i] - m_padding[i].second) * m_outputStrides[i+1]; + const Index lastPaddedRight = m_outputStrides[i]; + + if (last < lastPaddedLeft) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= firstPaddedRight && last < lastPaddedRight) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= lastPaddedLeft && last < firstPaddedRight) { + // all the coefficient are between the 2 padding zones. + const Index idx = index / m_outputStrides[i+1]; + inputIndex += (idx - m_padding[i].first) * m_inputStrides[i]; + index -= idx * m_outputStrides[i+1]; + } + else { + // Every other case + return packetWithPossibleZero(initialIndex); + } + } + + const Index last = index + packetSize - 1; + const Index first = index; + const Index lastPaddedLeft = m_padding[NumDims-1].first; + const Index firstPaddedRight = (m_dimensions[NumDims-1] - m_padding[NumDims-1].second); + const Index lastPaddedRight = m_outputStrides[NumDims-1]; + + if (last < lastPaddedLeft) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= firstPaddedRight && last < lastPaddedRight) { + // all the coefficient are in the padding zone. + return internal::pset1(Scalar(0)); + } + else if (first >= lastPaddedLeft && last < firstPaddedRight) { + // all the coefficient are between the 2 padding zones. + inputIndex += (index - m_padding[NumDims-1].first); + return m_impl.template packet(inputIndex); + } + // Every other case + return packetWithPossibleZero(initialIndex); + } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h index e2fe32d67..1c03d202f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h @@ -24,11 +24,14 @@ template struct traits > : public traits { typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions + 1; + static const int Layout = XprTraits::Layout; }; template @@ -89,11 +92,16 @@ struct TensorEvaluator, Device> enum { IsAligned = false, PacketAccess = TensorEvaluator::PacketAccess, - }; + Layout = TensorEvaluator::Layout, + CoordAccess = true, + }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_impl(op.expression(), device) { + // Only column major tensors are supported for now. + EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + Index num_patches = 1; const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); const PatchDim& patch_dims = op.patch_dims(); @@ -195,6 +203,35 @@ struct TensorEvaluator, Device> } } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array& coords) const + { + // Location of the first element of the patch. + const Index patchIndex = coords[NumDims - 1]; + + if (TensorEvaluator::CoordAccess) { + array inputCoords; + for (int i = NumDims - 2; i > 0; --i) { + const Index patchIdx = patchIndex / m_patchStrides[i]; + patchIndex -= patchIdx * m_patchStrides[i]; + const Index offsetIdx = coords[i]; + inputCoords[i] = coords[i] + patchIdx; + } + inputCoords[0] = (patchIndex + coords[0]); + return m_impl.coeff(inputCoords); + } + else { + Index inputIndex = 0; + for (int i = NumDims - 2; i > 0; --i) { + const Index patchIdx = patchIndex / m_patchStrides[i]; + patchIndex -= patchIdx * m_patchStrides[i]; + const Index offsetIdx = coords[i]; + inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i]; + } + inputIndex += (patchIndex + coords[0]); + return m_impl.coeff(inputIndex); + } + } + Scalar* data() const { return NULL; } protected: @@ -206,7 +243,6 @@ struct TensorEvaluator, Device> TensorEvaluator m_impl; }; - } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_PATCH_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h index 831a9f005..ab5fc6a69 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h @@ -24,11 +24,14 @@ template struct traits > : public traits { typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions; + static const int Layout = XprTraits::Layout; }; template @@ -99,6 +102,8 @@ struct TensorEvaluator, Device> enum { IsAligned = false, PacketAccess = (internal::packet_traits::size > 1), + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -112,15 +117,22 @@ struct TensorEvaluator, Device> array inputStrides; - for (int i = 0; i < NumDims; ++i) { - if (i > 0) { - inputStrides[i] = inputStrides[i-1] * input_dims[i-1]; - m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; - } else { - inputStrides[0] = 1; - m_outputStrides[0] = 1; + if (Layout == ColMajor) { + inputStrides[0] = 1; + m_outputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + inputStrides[i] = inputStrides[i - 1] * input_dims[i - 1]; + m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1]; + } + } else { + inputStrides[NumDims - 1] = 1; + m_outputStrides[NumDims - 1] = 1; + for (int i = NumDims - 2; i >= 0; --i) { + inputStrides[i] = inputStrides[i + 1] * input_dims[i + 1]; + m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1]; } } + for (int i = 0; i < NumDims; ++i) { m_inputStrides[i] = inputStrides[shuffle[i]]; } @@ -162,15 +174,23 @@ struct TensorEvaluator, Device> Scalar* data() const { return NULL; } protected: - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const - { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const { Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_outputStrides[i]; - inputIndex += idx * m_inputStrides[i]; - index -= idx * m_outputStrides[i]; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += idx * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + return inputIndex + index * m_inputStrides[0]; + } else { + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += idx * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + return inputIndex + index * m_inputStrides[NumDims - 1]; } - return inputIndex + index * m_inputStrides[0]; } Dimensions m_dimensions; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h index ecfdb762c..2fbdfadfe 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h @@ -24,11 +24,14 @@ template struct traits > : public traits { typedef typename XprType::Scalar Scalar; - typedef typename internal::packet_traits::type Packet; - typedef typename traits::StorageKind StorageKind; - typedef typename traits::Index Index; + typedef traits XprTraits; + typedef typename packet_traits::type Packet; + typedef typename XprTraits::StorageKind StorageKind; + typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; typedef typename remove_reference::type _Nested; + static const int NumDimensions = XprTraits::NumDimensions; + static const int Layout = XprTraits::Layout; }; template @@ -98,6 +101,8 @@ struct TensorEvaluator, Device> enum { IsAligned = /*TensorEvaluator::IsAligned*/false, PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -109,14 +114,25 @@ struct TensorEvaluator, Device> } const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - m_outputStrides[0] = 1; - m_inputStrides[0] = 1; - for (int i = 1; i < NumDims; ++i) { - m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; - m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; - m_inputStrides[i-1] *= op.strides()[i-1]; + if (Layout == ColMajor) { + m_outputStrides[0] = 1; + m_inputStrides[0] = 1; + for (int i = 1; i < NumDims; ++i) { + m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1]; + m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; + m_inputStrides[i-1] *= op.strides()[i-1]; + } + m_inputStrides[NumDims-1] *= op.strides()[NumDims-1]; + } else { // RowMajor + m_outputStrides[NumDims-1] = 1; + m_inputStrides[NumDims-1] = 1; + for (int i = NumDims - 2; i >= 0; --i) { + m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1]; + m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1]; + m_inputStrides[i+1] *= op.strides()[i+1]; + } + m_inputStrides[0] *= op.strides()[0]; } - m_inputStrides[NumDims-1] *= op.strides()[NumDims-1]; } typedef typename XprType::Scalar Scalar; @@ -135,14 +151,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const { - Index inputIndex = 0; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx = index / m_outputStrides[i]; - inputIndex += idx * m_inputStrides[i]; - index -= idx * m_outputStrides[i]; - } - inputIndex += index * m_inputStrides[0]; - return m_impl.coeff(inputIndex); + return m_impl.coeff(srcCoeff(index)); } template @@ -154,16 +163,29 @@ struct TensorEvaluator, Device> Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; - for (int i = NumDims - 1; i > 0; --i) { - const Index idx0 = indices[0] / m_outputStrides[i]; - const Index idx1 = indices[1] / m_outputStrides[i]; - inputIndices[0] += idx0 * m_inputStrides[i]; - inputIndices[1] += idx1 * m_inputStrides[i]; - indices[0] -= idx0 * m_outputStrides[i]; - indices[1] -= idx1 * m_outputStrides[i]; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / m_outputStrides[i]; + const Index idx1 = indices[1] / m_outputStrides[i]; + inputIndices[0] += idx0 * m_inputStrides[i]; + inputIndices[1] += idx1 * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += indices[0] * m_inputStrides[0]; + inputIndices[1] += indices[1] * m_inputStrides[0]; + } else { // RowMajor + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx0 = indices[0] / m_outputStrides[i]; + const Index idx1 = indices[1] / m_outputStrides[i]; + inputIndices[0] += idx0 * m_inputStrides[i]; + inputIndices[1] += idx1 * m_inputStrides[i]; + indices[0] -= idx0 * m_outputStrides[i]; + indices[1] -= idx1 * m_outputStrides[i]; + } + inputIndices[0] += indices[0] * m_inputStrides[NumDims-1]; + inputIndices[1] += indices[1] * m_inputStrides[NumDims-1]; } - inputIndices[0] += indices[0] * m_inputStrides[0]; - inputIndices[1] += indices[1] * m_inputStrides[0]; if (inputIndices[1] - inputIndices[0] == packetSize - 1) { PacketReturnType rslt = m_impl.template packet(inputIndices[0]); return rslt; @@ -183,6 +205,27 @@ struct TensorEvaluator, Device> Scalar* data() const { return NULL; } protected: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const + { + Index inputIndex = 0; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += idx * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + inputIndex += index * m_inputStrides[0]; + } else { // RowMajor + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx = index / m_outputStrides[i]; + inputIndex += idx * m_inputStrides[i]; + index -= idx * m_outputStrides[i]; + } + inputIndex += index * m_inputStrides[NumDims-1]; + } + return inputIndex; + } + Dimensions m_dimensions; array m_outputStrides; array m_inputStrides; @@ -190,6 +233,84 @@ struct TensorEvaluator, Device> }; +// Eval as lvalue +template +struct TensorEvaluator, Device> + : public TensorEvaluator, Device> +{ + typedef TensorStridingOp XprType; + typedef TensorEvaluator Base; + // typedef typename XprType::Index Index; + static const int NumDims = internal::array_size::Dimensions>::value; + // typedef DSizes Dimensions; + + enum { + IsAligned = /*TensorEvaluator::IsAligned*/false, + PacketAccess = TensorEvaluator::PacketAccess, + Layout = TensorEvaluator::Layout, + CoordAccess = false, // to be implemented + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : Base(op, device) { } + + typedef typename XprType::Index Index; + typedef typename XprType::Scalar Scalar; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index) + { + return this->m_impl.coeffRef(this->srcCoeff(index)); + } + + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketReturnType& x) + { + const int packetSize = internal::unpacket_traits::size; + EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) + eigen_assert(index+packetSize-1 < this->dimensions().TotalSize()); + + Index inputIndices[] = {0, 0}; + Index indices[] = {index, index + packetSize - 1}; + if (Layout == ColMajor) { + for (int i = NumDims - 1; i > 0; --i) { + const Index idx0 = indices[0] / this->m_outputStrides[i]; + const Index idx1 = indices[1] / this->m_outputStrides[i]; + inputIndices[0] += idx0 * this->m_inputStrides[i]; + inputIndices[1] += idx1 * this->m_inputStrides[i]; + indices[0] -= idx0 * this->m_outputStrides[i]; + indices[1] -= idx1 * this->m_outputStrides[i]; + } + inputIndices[0] += indices[0] * this->m_inputStrides[0]; + inputIndices[1] += indices[1] * this->m_inputStrides[0]; + } else { // RowMajor + for (int i = 0; i < NumDims - 1; ++i) { + const Index idx0 = indices[0] / this->m_outputStrides[i]; + const Index idx1 = indices[1] / this->m_outputStrides[i]; + inputIndices[0] += idx0 * this->m_inputStrides[i]; + inputIndices[1] += idx1 * this->m_inputStrides[i]; + indices[0] -= idx0 * this->m_outputStrides[i]; + indices[1] -= idx1 * this->m_outputStrides[i]; + } + inputIndices[0] += indices[0] * this->m_inputStrides[NumDims-1]; + inputIndices[1] += indices[1] * this->m_inputStrides[NumDims-1]; + } + if (inputIndices[1] - inputIndices[0] == packetSize - 1) { + this->m_impl.template writePacket(inputIndices[0], x); + } + else { + EIGEN_ALIGN_DEFAULT Scalar values[packetSize]; + internal::pstore(values, x); + this->m_impl.coeffRef(inputIndices[0]) = values[0]; + this->m_impl.coeffRef(inputIndices[1]) = values[packetSize-1]; + for (int i = 1; i < packetSize-1; ++i) { + this->coeffRef(index+i) = values[i]; + } + } + } +}; + + } // end namespace Eigen #endif // EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h index 5c0f78489..022d20360 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h @@ -50,6 +50,8 @@ struct traits > typedef Scalar_ Scalar; typedef Dense StorageKind; typedef DenseIndex Index; + static const int NumDimensions = NumIndices_; + static const int Layout = Options_ & RowMajor ? RowMajor : ColMajor; enum { Options = Options_, Flags = compute_tensor_flags::ret | LvalueBit, @@ -63,6 +65,8 @@ struct traits > typedef Scalar_ Scalar; typedef Dense StorageKind; typedef DenseIndex Index; + static const int NumDimensions = array_size::value; + static const int Layout = Options_ & RowMajor ? RowMajor : ColMajor; enum { Options = Options_, Flags = compute_tensor_flags::ret | LvalueBit, @@ -78,6 +82,8 @@ struct traits > typedef typename BaseTraits::Scalar Scalar; typedef typename BaseTraits::StorageKind StorageKind; typedef typename BaseTraits::Index Index; + static const int NumDimensions = BaseTraits::NumDimensions; + static const int Layout = BaseTraits::Layout; enum { Options = Options_, Flags = ((BaseTraits::Flags | LvalueBit) & ~AlignedBit) | (Options&Aligned ? AlignedBit : 0), @@ -92,6 +98,8 @@ struct traits > typedef typename BaseTraits::Scalar Scalar; typedef typename BaseTraits::StorageKind StorageKind; typedef typename BaseTraits::Index Index; + static const int NumDimensions = BaseTraits::NumDimensions; + static const int Layout = BaseTraits::Layout; enum { Options = BaseTraits::Options, Flags = ((BaseTraits::Flags | LvalueBit) & ~AlignedBit) | (Options&Aligned ? AlignedBit : 0), @@ -198,6 +206,51 @@ struct nested, 1, typename eval Date: Wed, 14 Jan 2015 15:43:38 -0800 Subject: [PATCH 1031/1103] Updated the list of include files --- unsupported/Eigen/CXX11/Tensor | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index aa26e5283..34107ae71 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor @@ -30,13 +30,20 @@ #include #include +#if __cplusplus > 199711 +#include +#endif + #ifdef EIGEN_USE_THREADS #include #endif -#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) +#ifdef EIGEN_USE_GPU +#include +#if defined(__CUDACC__) #include #endif +#endif #include "Eigen/Core" @@ -44,6 +51,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorInitializer.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorTraits.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h" @@ -55,15 +63,17 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h" -#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h" +#include "unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h" @@ -77,7 +87,6 @@ #include "unsupported/Eigen/CXX11/src/Tensor/Tensor.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorFixedSize.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorMap.h" - #include "unsupported/Eigen/CXX11/src/Tensor/TensorRef.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorIO.h" From b5124e7cfda27ed99dcfcec8cb1b674efa1ef4a3 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 14 Jan 2015 15:46:04 -0800 Subject: [PATCH 1032/1103] Created many additional tests --- unsupported/test/CMakeLists.txt | 13 +- unsupported/test/cxx11_tensor_assign.cpp | 73 +++ .../test/cxx11_tensor_broadcasting.cpp | 86 +++- unsupported/test/cxx11_tensor_chipping.cpp | 183 +++++-- .../test/cxx11_tensor_concatenation.cpp | 34 +- .../test/cxx11_tensor_contract_cuda.cpp | 121 +++++ unsupported/test/cxx11_tensor_contraction.cpp | 221 +++++--- unsupported/test/cxx11_tensor_cuda.cpp | 474 ++++++++++++++++++ unsupported/test/cxx11_tensor_device.cpp | 116 ++--- unsupported/test/cxx11_tensor_dimension.cpp | 9 +- unsupported/test/cxx11_tensor_expr.cpp | 40 ++ unsupported/test/cxx11_tensor_forced_eval.cpp | 27 + unsupported/test/cxx11_tensor_image_patch.cpp | 206 +++++++- unsupported/test/cxx11_tensor_map.cpp | 7 +- unsupported/test/cxx11_tensor_morphing.cpp | 145 ++++-- unsupported/test/cxx11_tensor_of_strings.cpp | 54 +- unsupported/test/cxx11_tensor_padding.cpp | 23 +- unsupported/test/cxx11_tensor_patch.cpp | 17 + unsupported/test/cxx11_tensor_reduction.cpp | 287 +++++++++-- unsupported/test/cxx11_tensor_shuffling.cpp | 28 +- unsupported/test/cxx11_tensor_simple.cpp | 3 + unsupported/test/cxx11_tensor_striding.cpp | 38 +- unsupported/test/cxx11_tensor_thread_pool.cpp | 72 +-- 23 files changed, 1909 insertions(+), 368 deletions(-) create mode 100644 unsupported/test/cxx11_tensor_contract_cuda.cpp create mode 100644 unsupported/test/cxx11_tensor_cuda.cpp diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index 89c651804..9f44e47f9 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt @@ -99,7 +99,7 @@ if(EIGEN_TEST_CXX11) # older compiler that don't support cxx11. ei_add_test(cxx11_meta "-std=c++0x") ei_add_test(cxx11_tensor_simple "-std=c++0x") - ei_add_test(cxx11_tensor_symmetry "-std=c++0x") +# ei_add_test(cxx11_tensor_symmetry "-std=c++0x") ei_add_test(cxx11_tensor_assign "-std=c++0x") ei_add_test(cxx11_tensor_dimension "-std=c++0x") ei_add_test(cxx11_tensor_index_list "-std=c++0x") @@ -126,8 +126,17 @@ if(EIGEN_TEST_CXX11) ei_add_test(cxx11_tensor_reduction "-std=c++0x") ei_add_test(cxx11_tensor_shuffling "-std=c++0x") ei_add_test(cxx11_tensor_striding "-std=c++0x") -# ei_add_test(cxx11_tensor_device "-std=c++0x") ei_add_test(cxx11_tensor_thread_pool "-std=c++0x") ei_add_test(cxx11_tensor_ref "-std=c++0x") + ei_add_test(cxx11_tensor_random "-std=c++0x") + ei_add_test(cxx11_tensor_casts "-std=c++0x") + ei_add_test(cxx11_tensor_reverse "-std=c++0x") + ei_add_test(cxx11_tensor_layout_swap "-std=c++0x") ei_add_test(cxx11_tensor_io "-std=c++0x") + + # These tests needs nvcc +# ei_add_test(cxx11_tensor_device "-std=c++0x") +# ei_add_test(cxx11_tensor_cuda "-std=c++0x") +# ei_add_test(cxx11_tensor_contract_cuda "-std=c++0x") + endif() diff --git a/unsupported/test/cxx11_tensor_assign.cpp b/unsupported/test/cxx11_tensor_assign.cpp index 0ac3f9bf9..d16aaf847 100644 --- a/unsupported/test/cxx11_tensor_assign.cpp +++ b/unsupported/test/cxx11_tensor_assign.cpp @@ -285,6 +285,78 @@ static void test_compound_assign() } } +static void test_std_initializers_tensor() { +#ifdef EIGEN_HAS_VARIADIC_TEMPLATES + Tensor a(3); + a.setValues({0, 1, 2}); + VERIFY_IS_EQUAL(a(0), 0); + VERIFY_IS_EQUAL(a(1), 1); + VERIFY_IS_EQUAL(a(2), 2); + + // It fills the top-left slice. + a.setValues({10, 20}); + VERIFY_IS_EQUAL(a(0), 10); + VERIFY_IS_EQUAL(a(1), 20); + VERIFY_IS_EQUAL(a(2), 2); + + // Chaining. + Tensor a2(3); + a2 = a.setValues({100, 200, 300}); + VERIFY_IS_EQUAL(a(0), 100); + VERIFY_IS_EQUAL(a(1), 200); + VERIFY_IS_EQUAL(a(2), 300); + VERIFY_IS_EQUAL(a2(0), 100); + VERIFY_IS_EQUAL(a2(1), 200); + VERIFY_IS_EQUAL(a2(2), 300); + + Tensor b(2, 3); + b.setValues({{0, 1, 2}, {3, 4, 5}}); + VERIFY_IS_EQUAL(b(0, 0), 0); + VERIFY_IS_EQUAL(b(0, 1), 1); + VERIFY_IS_EQUAL(b(0, 2), 2); + VERIFY_IS_EQUAL(b(1, 0), 3); + VERIFY_IS_EQUAL(b(1, 1), 4); + VERIFY_IS_EQUAL(b(1, 2), 5); + + // It fills the top-left slice. + b.setValues({{10, 20}, {30}}); + VERIFY_IS_EQUAL(b(0, 0), 10); + VERIFY_IS_EQUAL(b(0, 1), 20); + VERIFY_IS_EQUAL(b(0, 2), 2); + VERIFY_IS_EQUAL(b(1, 0), 30); + VERIFY_IS_EQUAL(b(1, 1), 4); + VERIFY_IS_EQUAL(b(1, 2), 5); + + Eigen::Tensor c(3, 2, 4); + c.setValues({{{0, 1, 2, 3}, {4, 5, 6, 7}}, + {{10, 11, 12, 13}, {14, 15, 16, 17}}, + {{20, 21, 22, 23}, {24, 25, 26, 27}}}); + VERIFY_IS_EQUAL(c(0, 0, 0), 0); + VERIFY_IS_EQUAL(c(0, 0, 1), 1); + VERIFY_IS_EQUAL(c(0, 0, 2), 2); + VERIFY_IS_EQUAL(c(0, 0, 3), 3); + VERIFY_IS_EQUAL(c(0, 1, 0), 4); + VERIFY_IS_EQUAL(c(0, 1, 1), 5); + VERIFY_IS_EQUAL(c(0, 1, 2), 6); + VERIFY_IS_EQUAL(c(0, 1, 3), 7); + VERIFY_IS_EQUAL(c(1, 0, 0), 10); + VERIFY_IS_EQUAL(c(1, 0, 1), 11); + VERIFY_IS_EQUAL(c(1, 0, 2), 12); + VERIFY_IS_EQUAL(c(1, 0, 3), 13); + VERIFY_IS_EQUAL(c(1, 1, 0), 14); + VERIFY_IS_EQUAL(c(1, 1, 1), 15); + VERIFY_IS_EQUAL(c(1, 1, 2), 16); + VERIFY_IS_EQUAL(c(1, 1, 3), 17); + VERIFY_IS_EQUAL(c(2, 0, 0), 20); + VERIFY_IS_EQUAL(c(2, 0, 1), 21); + VERIFY_IS_EQUAL(c(2, 0, 2), 22); + VERIFY_IS_EQUAL(c(2, 0, 3), 23); + VERIFY_IS_EQUAL(c(2, 1, 0), 24); + VERIFY_IS_EQUAL(c(2, 1, 1), 25); + VERIFY_IS_EQUAL(c(2, 1, 2), 26); + VERIFY_IS_EQUAL(c(2, 1, 3), 27); +#endif // EIGEN_HAS_VARIADIC_TEMPLATES +} void test_cxx11_tensor_assign() { @@ -294,4 +366,5 @@ void test_cxx11_tensor_assign() CALL_SUBTEST(test_same_type()); CALL_SUBTEST(test_auto_resize()); CALL_SUBTEST(test_compound_assign()); + CALL_SUBTEST(test_std_initializers_tensor()); } diff --git a/unsupported/test/cxx11_tensor_broadcasting.cpp b/unsupported/test/cxx11_tensor_broadcasting.cpp index 9663912a4..f0792bdcf 100644 --- a/unsupported/test/cxx11_tensor_broadcasting.cpp +++ b/unsupported/test/cxx11_tensor_broadcasting.cpp @@ -13,9 +13,10 @@ using Eigen::Tensor; +template static void test_simple_broadcasting() { - Tensor tensor(2,3,5,7); + Tensor tensor(2,3,5,7); tensor.setRandom(); array broadcasts; broadcasts[0] = 1; @@ -23,7 +24,7 @@ static void test_simple_broadcasting() broadcasts[2] = 1; broadcasts[3] = 1; - Tensor no_broadcast; + Tensor no_broadcast; no_broadcast = tensor.broadcast(broadcasts); VERIFY_IS_EQUAL(no_broadcast.dimension(0), 2); @@ -45,7 +46,7 @@ static void test_simple_broadcasting() broadcasts[1] = 3; broadcasts[2] = 1; broadcasts[3] = 4; - Tensor broadcast; + Tensor broadcast; broadcast = tensor.broadcast(broadcasts); VERIFY_IS_EQUAL(broadcast.dimension(0), 4); @@ -65,16 +66,17 @@ static void test_simple_broadcasting() } +template static void test_vectorized_broadcasting() { - Tensor tensor(8,3,5); + Tensor tensor(8,3,5); tensor.setRandom(); array broadcasts; broadcasts[0] = 2; broadcasts[1] = 3; broadcasts[2] = 4; - Tensor broadcast; + Tensor broadcast; broadcast = tensor.broadcast(broadcasts); VERIFY_IS_EQUAL(broadcast.dimension(0), 16); @@ -107,8 +109,78 @@ static void test_vectorized_broadcasting() } +template +static void test_static_broadcasting() +{ + Tensor tensor(8,3,5); + tensor.setRandom(); + Eigen::IndexList, Eigen::type2index<3>, Eigen::type2index<4>> broadcasts; + + Tensor broadcast; + broadcast = tensor.broadcast(broadcasts); + + VERIFY_IS_EQUAL(broadcast.dimension(0), 16); + VERIFY_IS_EQUAL(broadcast.dimension(1), 9); + VERIFY_IS_EQUAL(broadcast.dimension(2), 20); + + for (int i = 0; i < 16; ++i) { + for (int j = 0; j < 9; ++j) { + for (int k = 0; k < 20; ++k) { + VERIFY_IS_EQUAL(tensor(i%8,j%3,k%5), broadcast(i,j,k)); + } + } + } + + tensor.resize(11,3,5); + tensor.setRandom(); + broadcast = tensor.broadcast(broadcasts); + + VERIFY_IS_EQUAL(broadcast.dimension(0), 22); + VERIFY_IS_EQUAL(broadcast.dimension(1), 9); + VERIFY_IS_EQUAL(broadcast.dimension(2), 20); + + for (int i = 0; i < 22; ++i) { + for (int j = 0; j < 9; ++j) { + for (int k = 0; k < 20; ++k) { + VERIFY_IS_EQUAL(tensor(i%11,j%3,k%5), broadcast(i,j,k)); + } + } + } +} + + +template +static void test_fixed_size_broadcasting() +{ + // Need to add a [] operator to the Size class for this to work +#if 0 + Tensor t1(10); + t1.setRandom(); + TensorFixedSize, DataLayout> t2; + t2 = t2.constant(20.0f); + + Tensor t3 = t1 + t2.broadcast(Eigen::array{{10}}); + for (int i = 0; i < 10; ++i) { + VERIFY_IS_APPROX(t3(i), t1(i) + t2(0)); + } + + TensorMap, DataLayout> > t4(t2.data(), {{1}}); + Tensor t5 = t1 + t4.broadcast(Eigen::array{{10}}); + for (int i = 0; i < 10; ++i) { + VERIFY_IS_APPROX(t5(i), t1(i) + t2(0)); + } +#endif +} + + void test_cxx11_tensor_broadcasting() { - CALL_SUBTEST(test_simple_broadcasting()); - CALL_SUBTEST(test_vectorized_broadcasting()); + CALL_SUBTEST(test_simple_broadcasting()); + CALL_SUBTEST(test_simple_broadcasting()); + CALL_SUBTEST(test_vectorized_broadcasting()); + CALL_SUBTEST(test_vectorized_broadcasting()); + CALL_SUBTEST(test_static_broadcasting()); + CALL_SUBTEST(test_static_broadcasting()); + CALL_SUBTEST(test_fixed_size_broadcasting()); + CALL_SUBTEST(test_fixed_size_broadcasting()); } diff --git a/unsupported/test/cxx11_tensor_chipping.cpp b/unsupported/test/cxx11_tensor_chipping.cpp index 0027b2888..0de7bbac6 100644 --- a/unsupported/test/cxx11_tensor_chipping.cpp +++ b/unsupported/test/cxx11_tensor_chipping.cpp @@ -13,18 +13,20 @@ using Eigen::Tensor; - +template static void test_simple_chip() { - Tensor tensor(2,3,5,7,11); + Tensor tensor(2,3,5,7,11); tensor.setRandom(); - Tensor chip1; - chip1 = tensor.chip<0>(1); + Tensor chip1; + chip1 = tensor.template chip<0>(1); + VERIFY_IS_EQUAL(chip1.dimension(0), 3); VERIFY_IS_EQUAL(chip1.dimension(1), 5); VERIFY_IS_EQUAL(chip1.dimension(2), 7); VERIFY_IS_EQUAL(chip1.dimension(3), 11); + for (int i = 0; i < 3; ++i) { for (int j = 0; j < 5; ++j) { for (int k = 0; k < 7; ++k) { @@ -35,7 +37,7 @@ static void test_simple_chip() } } - Tensor chip2 = tensor.chip<1>(1); + Tensor chip2 = tensor.template chip<1>(1); VERIFY_IS_EQUAL(chip2.dimension(0), 2); VERIFY_IS_EQUAL(chip2.dimension(1), 5); VERIFY_IS_EQUAL(chip2.dimension(2), 7); @@ -50,7 +52,7 @@ static void test_simple_chip() } } - Tensor chip3 = tensor.chip<2>(2); + Tensor chip3 = tensor.template chip<2>(2); VERIFY_IS_EQUAL(chip3.dimension(0), 2); VERIFY_IS_EQUAL(chip3.dimension(1), 3); VERIFY_IS_EQUAL(chip3.dimension(2), 7); @@ -65,7 +67,7 @@ static void test_simple_chip() } } - Tensor chip4(tensor.chip<3>(5)); + Tensor chip4(tensor.template chip<3>(5)); VERIFY_IS_EQUAL(chip4.dimension(0), 2); VERIFY_IS_EQUAL(chip4.dimension(1), 3); VERIFY_IS_EQUAL(chip4.dimension(2), 5); @@ -80,7 +82,7 @@ static void test_simple_chip() } } - Tensor chip5(tensor.chip<4>(7)); + Tensor chip5(tensor.template chip<4>(7)); VERIFY_IS_EQUAL(chip5.dimension(0), 2); VERIFY_IS_EQUAL(chip5.dimension(1), 3); VERIFY_IS_EQUAL(chip5.dimension(2), 5); @@ -96,14 +98,97 @@ static void test_simple_chip() } } +template +static void test_dynamic_chip() +{ + Tensor tensor(2,3,5,7,11); + tensor.setRandom(); + Tensor chip1; + chip1 = tensor.chip(1, 0); + VERIFY_IS_EQUAL(chip1.dimension(0), 3); + VERIFY_IS_EQUAL(chip1.dimension(1), 5); + VERIFY_IS_EQUAL(chip1.dimension(2), 7); + VERIFY_IS_EQUAL(chip1.dimension(3), 11); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip1(i,j,k,l), tensor(1,i,j,k,l)); + } + } + } + } + + Tensor chip2 = tensor.chip(1, 1); + VERIFY_IS_EQUAL(chip2.dimension(0), 2); + VERIFY_IS_EQUAL(chip2.dimension(1), 5); + VERIFY_IS_EQUAL(chip2.dimension(2), 7); + VERIFY_IS_EQUAL(chip2.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip2(i,j,k,l), tensor(i,1,j,k,l)); + } + } + } + } + + Tensor chip3 = tensor.chip(2, 2); + VERIFY_IS_EQUAL(chip3.dimension(0), 2); + VERIFY_IS_EQUAL(chip3.dimension(1), 3); + VERIFY_IS_EQUAL(chip3.dimension(2), 7); + VERIFY_IS_EQUAL(chip3.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip3(i,j,k,l), tensor(i,j,2,k,l)); + } + } + } + } + + Tensor chip4(tensor.chip(5, 3)); + VERIFY_IS_EQUAL(chip4.dimension(0), 2); + VERIFY_IS_EQUAL(chip4.dimension(1), 3); + VERIFY_IS_EQUAL(chip4.dimension(2), 5); + VERIFY_IS_EQUAL(chip4.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(chip4(i,j,k,l), tensor(i,j,k,5,l)); + } + } + } + } + + Tensor chip5(tensor.chip(7, 4)); + VERIFY_IS_EQUAL(chip5.dimension(0), 2); + VERIFY_IS_EQUAL(chip5.dimension(1), 3); + VERIFY_IS_EQUAL(chip5.dimension(2), 5); + VERIFY_IS_EQUAL(chip5.dimension(3), 7); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(chip5(i,j,k,l), tensor(i,j,k,l,7)); + } + } + } + } +} + +template static void test_chip_in_expr() { - Tensor input1(2,3,5,7,11); + Tensor input1(2,3,5,7,11); input1.setRandom(); - Tensor input2(3,5,7,11); + Tensor input2(3,5,7,11); input2.setRandom(); - Tensor result = input1.chip<0>(0) + input2; + Tensor result = input1.template chip<0>(0) + input2; for (int i = 0; i < 3; ++i) { for (int j = 0; j < 5; ++j) { for (int k = 0; k < 7; ++k) { @@ -115,9 +200,9 @@ static void test_chip_in_expr() { } } - Tensor input3(3,7,11); + Tensor input3(3,7,11); input3.setRandom(); - Tensor result2 = input1.chip<0>(0).chip<1>(2) + input3; + Tensor result2 = input1.template chip<0>(0).template chip<1>(2) + input3; for (int i = 0; i < 3; ++i) { for (int j = 0; j < 7; ++j) { for (int k = 0; k < 11; ++k) { @@ -128,16 +213,16 @@ static void test_chip_in_expr() { } } - +template static void test_chip_as_lvalue() { - Tensor input1(2,3,5,7,11); + Tensor input1(2,3,5,7,11); input1.setRandom(); - Tensor input2(3,5,7,11); + Tensor input2(3,5,7,11); input2.setRandom(); - Tensor tensor = input1; - tensor.chip<0>(1) = input2; + Tensor tensor = input1; + tensor.template chip<0>(1) = input2; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 5; ++k) { @@ -154,10 +239,10 @@ static void test_chip_as_lvalue() } } - Tensor input3(2,5,7,11); + Tensor input3(2,5,7,11); input3.setRandom(); tensor = input1; - tensor.chip<1>(1) = input3; + tensor.template chip<1>(1) = input3; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 5; ++k) { @@ -174,10 +259,10 @@ static void test_chip_as_lvalue() } } - Tensor input4(2,3,7,11); + Tensor input4(2,3,7,11); input4.setRandom(); tensor = input1; - tensor.chip<2>(3) = input4; + tensor.template chip<2>(3) = input4; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 5; ++k) { @@ -194,10 +279,10 @@ static void test_chip_as_lvalue() } } - Tensor input5(2,3,5,11); + Tensor input5(2,3,5,11); input5.setRandom(); tensor = input1; - tensor.chip<3>(4) = input5; + tensor.template chip<3>(4) = input5; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 5; ++k) { @@ -214,10 +299,10 @@ static void test_chip_as_lvalue() } } - Tensor input6(2,3,5,7); + Tensor input6(2,3,5,7); input6.setRandom(); tensor = input1; - tensor.chip<4>(5) = input6; + tensor.template chip<4>(5) = input6; for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 5; ++k) { @@ -235,47 +320,57 @@ static void test_chip_as_lvalue() } } - +template static void test_chip_raw_data() { - Tensor tensor(2,3,5,7,11); + Tensor tensor(2,3,5,7,11); tensor.setRandom(); - typedef TensorEvaluator(3)), DefaultDevice> Evaluator4; - auto chip = Evaluator4(tensor.chip<4>(3), DefaultDevice()); + typedef TensorEvaluator(3)), DefaultDevice> Evaluator4; + auto chip = Evaluator4(tensor.template chip<4>(3), DefaultDevice()); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { for (int k = 0; k < 5; ++k) { for (int l = 0; l < 7; ++l) { - int chip_index = i + 2 * (j + 3 * (k + 5 * l)); + int chip_index; + if (DataLayout == ColMajor) { + chip_index = i + 2 * (j + 3 * (k + 5 * l)); + } else { + chip_index = 11 * (l + 7 * (k + 5 * (j + 3 * i))); + } VERIFY_IS_EQUAL(chip.data()[chip_index], tensor(i,j,k,l,3)); } } } } - typedef TensorEvaluator(0)), DefaultDevice> Evaluator0; - auto chip0 = Evaluator0(tensor.chip<0>(0), DefaultDevice()); + typedef TensorEvaluator(0)), DefaultDevice> Evaluator0; + auto chip0 = Evaluator0(tensor.template chip<0>(0), DefaultDevice()); VERIFY_IS_EQUAL(chip0.data(), static_cast(0)); - typedef TensorEvaluator(0)), DefaultDevice> Evaluator1; - auto chip1 = Evaluator1(tensor.chip<1>(0), DefaultDevice()); + typedef TensorEvaluator(0)), DefaultDevice> Evaluator1; + auto chip1 = Evaluator1(tensor.template chip<1>(0), DefaultDevice()); VERIFY_IS_EQUAL(chip1.data(), static_cast(0)); - typedef TensorEvaluator(0)), DefaultDevice> Evaluator2; - auto chip2 = Evaluator2(tensor.chip<2>(0), DefaultDevice()); + typedef TensorEvaluator(0)), DefaultDevice> Evaluator2; + auto chip2 = Evaluator2(tensor.template chip<2>(0), DefaultDevice()); VERIFY_IS_EQUAL(chip2.data(), static_cast(0)); - typedef TensorEvaluator(0)), DefaultDevice> Evaluator3; - auto chip3 = Evaluator3(tensor.chip<3>(0), DefaultDevice()); + typedef TensorEvaluator(0)), DefaultDevice> Evaluator3; + auto chip3 = Evaluator3(tensor.template chip<3>(0), DefaultDevice()); VERIFY_IS_EQUAL(chip3.data(), static_cast(0)); } - void test_cxx11_tensor_chipping() { - CALL_SUBTEST(test_simple_chip()); - CALL_SUBTEST(test_chip_in_expr()); - CALL_SUBTEST(test_chip_as_lvalue()); - CALL_SUBTEST(test_chip_raw_data()); + CALL_SUBTEST(test_simple_chip()); + CALL_SUBTEST(test_simple_chip()); + CALL_SUBTEST(test_dynamic_chip()); + CALL_SUBTEST(test_dynamic_chip()); + CALL_SUBTEST(test_chip_in_expr()); + CALL_SUBTEST(test_chip_in_expr()); + CALL_SUBTEST(test_chip_as_lvalue()); + CALL_SUBTEST(test_chip_as_lvalue()); + CALL_SUBTEST(test_chip_raw_data()); + CALL_SUBTEST(test_chip_raw_data()); } diff --git a/unsupported/test/cxx11_tensor_concatenation.cpp b/unsupported/test/cxx11_tensor_concatenation.cpp index 8fd4f5f80..9fdf33c16 100644 --- a/unsupported/test/cxx11_tensor_concatenation.cpp +++ b/unsupported/test/cxx11_tensor_concatenation.cpp @@ -13,15 +13,16 @@ using Eigen::Tensor; +template static void test_dimension_failures() { - Tensor left(2, 3, 1); - Tensor right(3, 3, 1); + Tensor left(2, 3, 1); + Tensor right(3, 3, 1); left.setRandom(); right.setRandom(); // Okay; other dimensions are equal. - Tensor concatenation = left.concatenate(right, 0); + Tensor concatenation = left.concatenate(right, 0); // Dimension mismatches. VERIFY_RAISES_ASSERT(concatenation = left.concatenate(right, 1)); @@ -32,33 +33,35 @@ static void test_dimension_failures() VERIFY_RAISES_ASSERT(concatenation = left.concatenate(right, -1)); } +template static void test_static_dimension_failure() { - Tensor left(2, 3); - Tensor right(2, 3, 1); + Tensor left(2, 3); + Tensor right(2, 3, 1); #ifdef CXX11_TENSOR_CONCATENATION_STATIC_DIMENSION_FAILURE // Technically compatible, but we static assert that the inputs have same // NumDims. - Tensor concatenation = left.concatenate(right, 0); + Tensor concatenation = left.concatenate(right, 0); #endif // This can be worked around in this case. - Tensor concatenation = left + Tensor concatenation = left .reshape(Tensor::Dimensions{{2, 3, 1}}) .concatenate(right, 0); - Tensor alternative = left + Tensor alternative = left .concatenate(right.reshape(Tensor::Dimensions{{2, 3}}), 0); } +template static void test_simple_concatenation() { - Tensor left(2, 3, 1); - Tensor right(2, 3, 1); + Tensor left(2, 3, 1); + Tensor right(2, 3, 1); left.setRandom(); right.setRandom(); - Tensor concatenation = left.concatenate(right, 0); + Tensor concatenation = left.concatenate(right, 0); VERIFY_IS_EQUAL(concatenation.dimension(0), 4); VERIFY_IS_EQUAL(concatenation.dimension(1), 3); VERIFY_IS_EQUAL(concatenation.dimension(2), 1); @@ -103,8 +106,11 @@ static void test_simple_concatenation() void test_cxx11_tensor_concatenation() { - CALL_SUBTEST(test_dimension_failures()); - CALL_SUBTEST(test_static_dimension_failure()); - CALL_SUBTEST(test_simple_concatenation()); + CALL_SUBTEST(test_dimension_failures()); + CALL_SUBTEST(test_dimension_failures()); + CALL_SUBTEST(test_static_dimension_failure()); + CALL_SUBTEST(test_static_dimension_failure()); + CALL_SUBTEST(test_simple_concatenation()); + CALL_SUBTEST(test_simple_concatenation()); // CALL_SUBTEST(test_vectorized_concatenation()); } diff --git a/unsupported/test/cxx11_tensor_contract_cuda.cpp b/unsupported/test/cxx11_tensor_contract_cuda.cpp new file mode 100644 index 000000000..9599607c6 --- /dev/null +++ b/unsupported/test/cxx11_tensor_contract_cuda.cpp @@ -0,0 +1,121 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// Copyright (C) 2014 Navdeep Jaitly +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#define EIGEN_TEST_NO_LONGDOUBLE +#define EIGEN_TEST_NO_COMPLEX +#define EIGEN_TEST_FUNC cxx11_tensor_cuda +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int +#define EIGEN_USE_GPU + + +#include "main.h" +#include + +using Eigen::Tensor; +typedef Tensor::DimensionPair DimPair; + +template +static void test_cuda_contraction(int m_size, int k_size, int n_size) +{ + cout<<"Calling with ("< t_left(Eigen::array(m_size, k_size)); + Tensor t_right(Eigen::array(k_size, n_size)); + Tensor t_result(Eigen::array(m_size, n_size)); + Tensor t_result_gpu(Eigen::array(m_size, n_size)); + Eigen::array dims(DimPair(1, 0)); + + t_left.setRandom(); + t_right.setRandom(); + + std::size_t t_left_bytes = t_left.size() * sizeof(float); + std::size_t t_right_bytes = t_right.size() * sizeof(float); + std::size_t t_result_bytes = t_result.size() * sizeof(float); + + float* d_t_left; + float* d_t_right; + float* d_t_result; + + cudaMalloc((void**)(&d_t_left), t_left_bytes); + cudaMalloc((void**)(&d_t_right), t_right_bytes); + cudaMalloc((void**)(&d_t_result), t_result_bytes); + + cudaMemcpy(d_t_left, t_left.data(), t_left_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_t_right, t_right.data(), t_right_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > + gpu_t_left(d_t_left, Eigen::array(m_size, k_size)); + Eigen::TensorMap > + gpu_t_right(d_t_right, Eigen::array(k_size, n_size)); + Eigen::TensorMap > + gpu_t_result(d_t_result, Eigen::array(m_size, n_size)); + + + gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims); + t_result = t_left.contract(t_right, dims); + + cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost); + for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { + if (fabs(t_result.data()[i] - t_result_gpu.data()[i]) >= 1e-4) { + cout << "mismatch detected at index " << i << ": " << t_result.data()[i] + << " vs " << t_result_gpu.data()[i] << endl; + assert(false); + } + } + + cudaFree((void*)d_t_left); + cudaFree((void*)d_t_right); + cudaFree((void*)d_t_result); +} + + +void test_cxx11_tensor_cuda() +{ + cout<<"Calling contraction tests"<(128, 128, 128)); + CALL_SUBTEST(test_cuda_contraction(128, 128, 128)); + for (int k = 32; k < 256; k++) { + CALL_SUBTEST(test_cuda_contraction(128, k, 128)); + CALL_SUBTEST(test_cuda_contraction(128, k, 128)); + } + for (int k = 32; k < 256; k++) { + CALL_SUBTEST(test_cuda_contraction(128, 128, k)); + CALL_SUBTEST(test_cuda_contraction(128, 128, k)); + } + for (int k = 32; k < 256; k++) { + CALL_SUBTEST(test_cuda_contraction(k, 128, 128)); + CALL_SUBTEST(test_cuda_contraction(k, 128, 128)); + } + + int m_sizes[] = {31, 39, 63, 64, 65, + 127, 129, 255, 257, 511, + 512, 513, 1023, 1024, 1025 }; + int n_sizes[] = {31, 39, 63, 64, 65, + 127, 129, 255, 257, 511, + 512, 513, 1023, 1024, 1025 }; + + int k_sizes[] = { 31, 39, 63, 64, 65, + 95, 96, 127, 129, 255, + 257, 511, 512, 513, 1023, + 1024, 1025}; + + for (int i = 0; i <15; i++) + for (int j = 0; j < 15; j++) + for (int k = 0; k < 17; k++) { + CALL_SUBTEST(test_cuda_contraction(m_sizes[i], n_sizes[j], k_sizes[k])); + CALL_SUBTEST(test_cuda_contraction(m_sizes[i], n_sizes[j], k_sizes[k])); + } +} diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp index 17bd335f7..6124818fd 100644 --- a/unsupported/test/cxx11_tensor_contraction.cpp +++ b/unsupported/test/cxx11_tensor_contraction.cpp @@ -16,18 +16,18 @@ using Eigen::Tensor; typedef Tensor::DimensionPair DimPair; - +template static void test_evals() { - Tensor mat1(2, 3); - Tensor mat2(2, 3); - Tensor mat3(3, 2); + Tensor mat1(2, 3); + Tensor mat2(2, 3); + Tensor mat3(3, 2); mat1.setRandom(); mat2.setRandom(); mat3.setRandom(); - Tensor mat4(3,3); + Tensor mat4(3,3); mat4.setZero(); Eigen::array dims3({{DimPair(0, 0)}}); typedef TensorEvaluator Evaluator; @@ -47,7 +47,7 @@ static void test_evals() VERIFY_IS_APPROX(mat4(2,1), mat1(0,2)*mat2(0,1) + mat1(1,2)*mat2(1,1)); VERIFY_IS_APPROX(mat4(2,2), mat1(0,2)*mat2(0,2) + mat1(1,2)*mat2(1,2)); - Tensor mat5(2,2); + Tensor mat5(2,2); mat5.setZero(); Eigen::array dims4({{DimPair(1, 1)}}); typedef TensorEvaluator Evaluator2; @@ -62,7 +62,7 @@ static void test_evals() VERIFY_IS_APPROX(mat5(1,0), mat1(1,0)*mat2(0,0) + mat1(1,1)*mat2(0,1) + mat1(1,2)*mat2(0,2)); VERIFY_IS_APPROX(mat5(1,1), mat1(1,0)*mat2(1,0) + mat1(1,1)*mat2(1,1) + mat1(1,2)*mat2(1,2)); - Tensor mat6(2,2); + Tensor mat6(2,2); mat6.setZero(); Eigen::array dims6({{DimPair(1, 0)}}); typedef TensorEvaluator Evaluator3; @@ -78,16 +78,16 @@ static void test_evals() VERIFY_IS_APPROX(mat6(1,1), mat1(1,0)*mat3(0,1) + mat1(1,1)*mat3(1,1) + mat1(1,2)*mat3(2,1)); } - +template static void test_scalar() { - Tensor vec1({6}); - Tensor vec2({6}); + Tensor vec1({6}); + Tensor vec2({6}); vec1.setRandom(); vec2.setRandom(); - Tensor scalar(1); + Tensor scalar(1); scalar.setZero(); Eigen::array dims({{DimPair(0, 0)}}); typedef TensorEvaluator Evaluator; @@ -102,16 +102,16 @@ static void test_scalar() VERIFY_IS_APPROX(scalar(0), expected); } - +template static void test_multidims() { - Tensor mat1(2, 2, 2); - Tensor mat2(2, 2, 2, 2); + Tensor mat1(2, 2, 2); + Tensor mat2(2, 2, 2, 2); mat1.setRandom(); mat2.setRandom(); - Tensor mat3(2, 2, 2); + Tensor mat3(2, 2, 2); mat3.setZero(); Eigen::array dims({{DimPair(1, 2), DimPair(2, 3)}}); typedef TensorEvaluator Evaluator; @@ -140,15 +140,15 @@ static void test_multidims() mat1(1,0,1)*mat2(1,1,0,1) + mat1(1,1,1)*mat2(1,1,1,1)); } - +template static void test_holes() { - Tensor t1(2, 5, 7, 3); - Tensor t2(2, 7, 11, 13, 3); + Tensor t1(2, 5, 7, 3); + Tensor t2(2, 7, 11, 13, 3); t1.setRandom(); t2.setRandom(); Eigen::array dims({{DimPair(0, 0), DimPair(3, 4)}}); - Tensor result = t1.contract(t2, dims); + Tensor result = t1.contract(t2, dims); VERIFY_IS_EQUAL(result.dimension(0), 5); VERIFY_IS_EQUAL(result.dimension(1), 7); VERIFY_IS_EQUAL(result.dimension(2), 7); @@ -174,16 +174,16 @@ static void test_holes() { } } - +template static void test_full_redux() { - Tensor t1(2, 2); - Tensor t2(2, 2, 2); + Tensor t1(2, 2); + Tensor t2(2, 2, 2); t1.setRandom(); t2.setRandom(); Eigen::array dims({{DimPair(0, 0), DimPair(1, 1)}}); - Tensor result = t1.contract(t2, dims); + Tensor result = t1.contract(t2, dims); VERIFY_IS_EQUAL(result.dimension(0), 2); VERIFY_IS_APPROX(result(0), t1(0, 0) * t2(0, 0, 0) + t1(1, 0) * t2(1, 0, 0) + t1(0, 1) * t2(0, 1, 0) + t1(1, 1) * t2(1, 1, 0)); @@ -200,13 +200,13 @@ static void test_full_redux() + t1(0, 1) * t2(1, 0, 1) + t1(1, 1) * t2(1, 1, 1)); } - +template static void test_contraction_of_contraction() { - Tensor t1(2, 2); - Tensor t2(2, 2); - Tensor t3(2, 2); - Tensor t4(2, 2); + Tensor t1(2, 2); + Tensor t2(2, 2); + Tensor t3(2, 2); + Tensor t4(2, 2); t1.setRandom(); t2.setRandom(); t3.setRandom(); @@ -216,30 +216,32 @@ static void test_contraction_of_contraction() auto contract1 = t1.contract(t2, dims); auto diff = t3 - contract1; auto contract2 = t1.contract(t4, dims); - Tensor result = contract2.contract(diff, dims); + Tensor result = contract2.contract(diff, dims); + VERIFY_IS_EQUAL(result.dimension(0), 2); VERIFY_IS_EQUAL(result.dimension(1), 2); - Eigen::Map m1(t1.data(), 2, 2); - Eigen::Map m2(t2.data(), 2, 2); - Eigen::Map m3(t3.data(), 2, 2); - Eigen::Map m4(t4.data(), 2, 2); - Eigen::MatrixXf expected = (m1 * m4) * (m3 - m1 * m2); + Eigen::Map> + m1(t1.data(), 2, 2), m2(t2.data(), 2, 2), m3(t3.data(), 2, 2), + m4(t4.data(), 2, 2); + Eigen::Matrix + expected = (m1 * m4) * (m3 - m1 * m2); + VERIFY_IS_APPROX(result(0, 0), expected(0, 0)); VERIFY_IS_APPROX(result(0, 1), expected(0, 1)); VERIFY_IS_APPROX(result(1, 0), expected(1, 0)); VERIFY_IS_APPROX(result(1, 1), expected(1, 1)); } - +template static void test_expr() { - Tensor mat1(2, 3); - Tensor mat2(3, 2); + Tensor mat1(2, 3); + Tensor mat2(3, 2); mat1.setRandom(); mat2.setRandom(); - Tensor mat3(2,2); + Tensor mat3(2,2); Eigen::array dims({{DimPair(1, 0)}}); mat3 = mat1.contract(mat2, dims); @@ -250,16 +252,16 @@ static void test_expr() VERIFY_IS_APPROX(mat3(1,1), mat1(1,0)*mat2(0,1) + mat1(1,1)*mat2(1,1) + mat1(1,2)*mat2(2,1)); } - +template static void test_out_of_order_contraction() { - Tensor mat1(2, 2, 2); - Tensor mat2(2, 2, 2); + Tensor mat1(2, 2, 2); + Tensor mat2(2, 2, 2); mat1.setRandom(); mat2.setRandom(); - Tensor mat3(2, 2); + Tensor mat3(2, 2); Eigen::array dims({{DimPair(2, 0), DimPair(0, 2)}}); mat3 = mat1.contract(mat2, dims); @@ -295,18 +297,18 @@ static void test_out_of_order_contraction() } - +template static void test_consistency() { // this does something like testing (A*B)^T = (B^T * A^T) - Tensor mat1(4, 3, 5); - Tensor mat2(3, 2, 1, 5, 4); + Tensor mat1(4, 3, 5); + Tensor mat2(3, 2, 1, 5, 4); mat1.setRandom(); mat2.setRandom(); - Tensor mat3(5, 2, 1, 5); - Tensor mat4(2, 1, 5, 5); + Tensor mat3(5, 2, 1, 5); + Tensor mat4(2, 1, 5, 5); // contract on dimensions of size 4 and 3 Eigen::array dims1({{DimPair(0, 4), DimPair(1, 0)}}); @@ -316,27 +318,40 @@ static void test_consistency() mat4 = mat2.contract(mat1, dims2); // check that these are equal except for ordering of dimensions - for (size_t i = 0; i < 5; i++) { - for (size_t j = 0; j < 10; j++) { - VERIFY_IS_APPROX(mat3.data()[i + 5 * j], mat4.data()[j + 10 * i]); + if (DataLayout == ColMajor) { + for (size_t i = 0; i < 5; i++) { + for (size_t j = 0; j < 10; j++) { + VERIFY_IS_APPROX(mat3.data()[i + 5 * j], mat4.data()[j + 10 * i]); + } + } + } else { + // Row major + for (size_t i = 0; i < 5; i++) { + for (size_t j = 0; j < 10; j++) { + VERIFY_IS_APPROX(mat3.data()[10 * i + j], mat4.data()[i + 5 * j]); + } } } } - +template static void test_large_contraction() { - Tensor t_left(30, 50, 8, 31); - Tensor t_right(8, 31, 7, 20, 10); - Tensor t_result(30, 50, 7, 20, 10); + Tensor t_left(30, 50, 8, 31); + Tensor t_right(8, 31, 7, 20, 10); + Tensor t_result(30, 50, 7, 20, 10); t_left.setRandom(); t_right.setRandom(); - typedef Map MapXf; + // Add a little offset so that the results won't be close to zero. + t_left += t_left.constant(1.0f); + t_right += t_right.constant(1.0f); + + typedef Map> MapXf; MapXf m_left(t_left.data(), 1500, 248); MapXf m_right(t_right.data(), 248, 1400); - MatrixXf m_result(1500, 1400); + Eigen::Matrix m_result(1500, 1400); // this contraction should be equivalent to a single matrix multiplication Eigen::array dims({{DimPair(2, 0), DimPair(3, 1)}}); @@ -351,20 +366,20 @@ static void test_large_contraction() } } - +template static void test_matrix_vector() { - Tensor t_left(30, 50); - Tensor t_right(50); - Tensor t_result(30); + Tensor t_left(30, 50); + Tensor t_right(50); + Tensor t_result(30); t_left.setRandom(); t_right.setRandom(); - typedef Map> MapXf; + typedef Map> MapXf; MapXf m_left(t_left.data(), 30, 50); MapXf m_right(t_right.data(), 50, 1); - Eigen::Matrix m_result(30, 1); + Eigen::Matrix m_result(30, 1); // this contraction should be equivalent to a single matrix multiplication Eigen::array dims{{DimPair(1, 0)}}; @@ -379,18 +394,19 @@ static void test_matrix_vector() } +template static void test_tensor_vector() { - Tensor t_left(7, 13, 17); - Tensor t_right(1, 7); - typedef typename Tensor::DimensionPair DimensionPair; + Tensor t_left(7, 13, 17); + Tensor t_right(1, 7); + typedef typename Tensor::DimensionPair DimensionPair; Eigen::array dim_pair01{{{0, 1}}}; - Tensor t_result = t_left.contract(t_right, dim_pair01); + Tensor t_result = t_left.contract(t_right, dim_pair01); - typedef Map> MapXf; + typedef Map> MapXf; MapXf m_left(t_left.data(), 7, 13*17); MapXf m_right(t_right.data(), 1, 7); - Eigen::Matrix m_result = m_left.transpose() * m_right.transpose(); + Eigen::Matrix m_result = m_left.transpose() * m_right.transpose(); for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { VERIFY_IS_APPROX(t_result(i), m_result(i, 0)); @@ -398,18 +414,63 @@ static void test_tensor_vector() } +template +static void test_small_blocking_factors() +{ + Tensor t_left(30, 5, 3, 31); + Tensor t_right(3, 31, 7, 20, 1); + t_left.setRandom(); + t_right.setRandom(); + + // Add a little offset so that the results won't be close to zero. + t_left += t_left.constant(1.0f); + t_right += t_right.constant(1.0f); + + // Force the cache sizes, which results in smaller blocking factors. + Eigen::setCpuCacheSizes(896, 1920, 2944); + + // this contraction should be equivalent to a single matrix multiplication + Eigen::array dims({{DimPair(2, 0), DimPair(3, 1)}}); + Tensor t_result; + t_result = t_left.contract(t_right, dims); + + // compute result using a simple eigen matrix product + Map> m_left(t_left.data(), 150, 93); + Map> m_right(t_right.data(), 93, 140); + Eigen::Matrix m_result = m_left * m_right; + + for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { + VERIFY_IS_APPROX(t_result.data()[i], m_result.data()[i]); + } +} + + void test_cxx11_tensor_contraction() { - CALL_SUBTEST(test_evals()); - CALL_SUBTEST(test_scalar()); - CALL_SUBTEST(test_multidims()); - CALL_SUBTEST(test_holes()); - CALL_SUBTEST(test_full_redux()); - CALL_SUBTEST(test_contraction_of_contraction()); - CALL_SUBTEST(test_expr()); - CALL_SUBTEST(test_out_of_order_contraction()); - CALL_SUBTEST(test_consistency()); - CALL_SUBTEST(test_large_contraction()); - CALL_SUBTEST(test_matrix_vector()); - CALL_SUBTEST(test_tensor_vector()); + CALL_SUBTEST(test_evals()); + CALL_SUBTEST(test_evals()); + CALL_SUBTEST(test_scalar()); + CALL_SUBTEST(test_scalar()); + CALL_SUBTEST(test_multidims()); + CALL_SUBTEST(test_multidims()); + CALL_SUBTEST(test_holes()); + CALL_SUBTEST(test_holes()); + CALL_SUBTEST(test_full_redux()); + CALL_SUBTEST(test_full_redux()); + CALL_SUBTEST(test_contraction_of_contraction()); + CALL_SUBTEST(test_contraction_of_contraction()); + CALL_SUBTEST(test_expr()); + CALL_SUBTEST(test_expr()); + CALL_SUBTEST(test_out_of_order_contraction()); + CALL_SUBTEST(test_out_of_order_contraction()); + CALL_SUBTEST(test_consistency()); + CALL_SUBTEST(test_consistency()); + CALL_SUBTEST(test_large_contraction()); + CALL_SUBTEST(test_large_contraction()); + CALL_SUBTEST(test_matrix_vector()); + CALL_SUBTEST(test_matrix_vector()); + CALL_SUBTEST(test_tensor_vector()); + CALL_SUBTEST(test_tensor_vector()); + CALL_SUBTEST(test_small_blocking_factors()); + CALL_SUBTEST(test_small_blocking_factors()); } diff --git a/unsupported/test/cxx11_tensor_cuda.cpp b/unsupported/test/cxx11_tensor_cuda.cpp new file mode 100644 index 000000000..059d23de1 --- /dev/null +++ b/unsupported/test/cxx11_tensor_cuda.cpp @@ -0,0 +1,474 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Benoit Steiner +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +// TODO(mdevin): Free the cuda memory. + +#define EIGEN_TEST_NO_LONGDOUBLE +#define EIGEN_TEST_NO_COMPLEX +#define EIGEN_TEST_FUNC cxx11_tensor_cuda +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int +#define EIGEN_USE_GPU + + +#include "main.h" +#include + +using Eigen::Tensor; + +void test_cuda_elementwise_small() { + Tensor in1(Eigen::array(2)); + Tensor in2(Eigen::array(2)); + Tensor out(Eigen::array(2)); + in1.setRandom(); + in2.setRandom(); + + std::size_t in1_bytes = in1.size() * sizeof(float); + std::size_t in2_bytes = in2.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(float); + + float* d_in1; + float* d_in2; + float* d_out; + cudaMalloc((void**)(&d_in1), in1_bytes); + cudaMalloc((void**)(&d_in2), in2_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap, Eigen::Aligned> gpu_in1( + d_in1, Eigen::array(2)); + Eigen::TensorMap, Eigen::Aligned> gpu_in2( + d_in2, Eigen::array(2)); + Eigen::TensorMap, Eigen::Aligned> gpu_out( + d_out, Eigen::array(2)); + + gpu_out.device(gpu_device) = gpu_in1 + gpu_in2; + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, + gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 2; ++i) { + VERIFY_IS_APPROX( + out(Eigen::array(i)), + in1(Eigen::array(i)) + in2(Eigen::array(i))); + } +} + +void test_cuda_elementwise() +{ + Tensor in1(Eigen::array(72,53,97)); + Tensor in2(Eigen::array(72,53,97)); + Tensor in3(Eigen::array(72,53,97)); + Tensor out(Eigen::array(72,53,97)); + in1.setRandom(); + in2.setRandom(); + in3.setRandom(); + + std::size_t in1_bytes = in1.size() * sizeof(float); + std::size_t in2_bytes = in2.size() * sizeof(float); + std::size_t in3_bytes = in3.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(float); + + float* d_in1; + float* d_in2; + float* d_in3; + float* d_out; + cudaMalloc((void**)(&d_in1), in1_bytes); + cudaMalloc((void**)(&d_in2), in2_bytes); + cudaMalloc((void**)(&d_in3), in3_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in3, in3.data(), in3_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > gpu_in1(d_in1, Eigen::array(72,53,97)); + Eigen::TensorMap > gpu_in2(d_in2, Eigen::array(72,53,97)); + Eigen::TensorMap > gpu_in3(d_in3, Eigen::array(72,53,97)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(72,53,97)); + + gpu_out.device(gpu_device) = gpu_in1 + gpu_in2 * gpu_in3; + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 72; ++i) { + for (int j = 0; j < 53; ++j) { + for (int k = 0; k < 97; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * in3(Eigen::array(i,j,k))); + } + } + } +} + + +void test_cuda_reduction() +{ + Tensor in1(Eigen::array(72,53,97,113)); + Tensor out(Eigen::array(72,97)); + in1.setRandom(); + + std::size_t in1_bytes = in1.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(float); + + float* d_in1; + float* d_out; + cudaMalloc((void**)(&d_in1), in1_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > gpu_in1(d_in1, Eigen::array(72,53,97,113)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(72,97)); + + array reduction_axis; + reduction_axis[0] = 1; + reduction_axis[1] = 3; + + gpu_out.device(gpu_device) = gpu_in1.maximum(reduction_axis); + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 72; ++i) { + for (int j = 0; j < 97; ++j) { + float expected = 0; + for (int k = 0; k < 53; ++k) { + for (int l = 0; l < 113; ++l) { + expected = + std::max(expected, in1(Eigen::array(i, k, j, l))); + } + } + VERIFY_IS_APPROX(out(Eigen::array(i,j)), expected); + } + } +} + +template +static void test_cuda_contraction() +{ + // with these dimensions, the output has 300 * 140 elements, which is + // more than 30 * 1024, which is the number of threads in blocks on + // a 15 SM GK110 GPU + Tensor t_left(Eigen::array(6, 50, 3, 31)); + Tensor t_right(Eigen::array(3, 31, 7, 20, 1)); + Tensor t_result(Eigen::array(6, 50, 7, 20, 1)); + + t_left.setRandom(); + t_right.setRandom(); + + std::size_t t_left_bytes = t_left.size() * sizeof(float); + std::size_t t_right_bytes = t_right.size() * sizeof(float); + std::size_t t_result_bytes = t_result.size() * sizeof(float); + + float* d_t_left; + float* d_t_right; + float* d_t_result; + + cudaMalloc((void**)(&d_t_left), t_left_bytes); + cudaMalloc((void**)(&d_t_right), t_right_bytes); + cudaMalloc((void**)(&d_t_result), t_result_bytes); + + cudaMemcpy(d_t_left, t_left.data(), t_left_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_t_right, t_right.data(), t_right_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > + gpu_t_left(d_t_left, Eigen::array(6, 50, 3, 31)); + Eigen::TensorMap > + gpu_t_right(d_t_right, Eigen::array(3, 31, 7, 20, 1)); + Eigen::TensorMap > + gpu_t_result(d_t_result, Eigen::array(6, 50, 7, 20, 1)); + + typedef Eigen::Map > MapXf; + MapXf m_left(t_left.data(), 300, 93); + MapXf m_right(t_right.data(), 93, 140); + Eigen::Matrix m_result(300, 140); + + typedef Tensor::DimensionPair DimPair; + Eigen::array dims; + dims[0] = DimPair(2, 0); + dims[1] = DimPair(3, 1); + + m_result = m_left * m_right; + gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims); + + cudaMemcpy(t_result.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost); + + for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { + if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { + cout << "mismatch detected at index " << i << ": " << t_result.data()[i] << " vs " << m_result.data()[i] << endl; + assert(false); + } + } +} + +static void test_cuda_convolution_1d() +{ + Tensor input(Eigen::array(74,37,11,137)); + Tensor kernel(Eigen::array(4)); + Tensor out(Eigen::array(74,34,11,137)); + input = input.constant(10.0f) + input.random(); + kernel = kernel.constant(7.0f) + kernel.random(); + + std::size_t input_bytes = input.size() * sizeof(float); + std::size_t kernel_bytes = kernel.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(float); + + float* d_input; + float* d_kernel; + float* d_out; + cudaMalloc((void**)(&d_input), input_bytes); + cudaMalloc((void**)(&d_kernel), kernel_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > gpu_input(d_input, Eigen::array(74,37,11,137)); + Eigen::TensorMap > gpu_kernel(d_kernel, Eigen::array(4)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(74,34,11,137)); + + Eigen::array dims(1); + gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 74; ++i) { + for (int j = 0; j < 34; ++j) { + for (int k = 0; k < 11; ++k) { + for (int l = 0; l < 137; ++l) { + const float result = out(Eigen::array(i,j,k,l)); + const float expected = input(Eigen::array(i,j+0,k,l)) * kernel(Eigen::array(0)) + + input(Eigen::array(i,j+1,k,l)) * kernel(Eigen::array(1)) + + input(Eigen::array(i,j+2,k,l)) * kernel(Eigen::array(2)) + + input(Eigen::array(i,j+3,k,l)) * kernel(Eigen::array(3)); + VERIFY_IS_APPROX(result, expected); + } + } + } + } +} + + +static void test_cuda_convolution_2d() +{ + Tensor input(Eigen::array(74,37,11,137)); + Tensor kernel(Eigen::array(3,4)); + Tensor out(Eigen::array(74,35,8,137)); + input = input.constant(10.0f) + input.random(); + kernel = kernel.constant(7.0f) + kernel.random(); + + std::size_t input_bytes = input.size() * sizeof(float); + std::size_t kernel_bytes = kernel.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(float); + + float* d_input; + float* d_kernel; + float* d_out; + cudaMalloc((void**)(&d_input), input_bytes); + cudaMalloc((void**)(&d_kernel), kernel_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > gpu_input(d_input, Eigen::array(74,37,11,137)); + Eigen::TensorMap > gpu_kernel(d_kernel, Eigen::array(3,4)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(74,35,8,137)); + + Eigen::array dims(1,2); + gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 74; ++i) { + for (int j = 0; j < 35; ++j) { + for (int k = 0; k < 8; ++k) { + for (int l = 0; l < 137; ++l) { + const float result = out(Eigen::array(i,j,k,l)); + const float expected = input(Eigen::array(i,j+0,k+0,l)) * kernel(Eigen::array(0,0)) + + input(Eigen::array(i,j+1,k+0,l)) * kernel(Eigen::array(1,0)) + + input(Eigen::array(i,j+2,k+0,l)) * kernel(Eigen::array(2,0)) + + input(Eigen::array(i,j+0,k+1,l)) * kernel(Eigen::array(0,1)) + + input(Eigen::array(i,j+1,k+1,l)) * kernel(Eigen::array(1,1)) + + input(Eigen::array(i,j+2,k+1,l)) * kernel(Eigen::array(2,1)) + + input(Eigen::array(i,j+0,k+2,l)) * kernel(Eigen::array(0,2)) + + input(Eigen::array(i,j+1,k+2,l)) * kernel(Eigen::array(1,2)) + + input(Eigen::array(i,j+2,k+2,l)) * kernel(Eigen::array(2,2)) + + input(Eigen::array(i,j+0,k+3,l)) * kernel(Eigen::array(0,3)) + + input(Eigen::array(i,j+1,k+3,l)) * kernel(Eigen::array(1,3)) + + input(Eigen::array(i,j+2,k+3,l)) * kernel(Eigen::array(2,3)); + VERIFY_IS_APPROX(result, expected); + } + } + } + } +} + + +static void test_cuda_convolution_3d() +{ + Tensor input(Eigen::array(74,37,11,137,17)); + Tensor kernel(Eigen::array(3,4,2)); + Tensor out(Eigen::array(74,35,8,136,17)); + input = input.constant(10.0f) + input.random(); + kernel = kernel.constant(7.0f) + kernel.random(); + + std::size_t input_bytes = input.size() * sizeof(float); + std::size_t kernel_bytes = kernel.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(float); + + float* d_input; + float* d_kernel; + float* d_out; + cudaMalloc((void**)(&d_input), input_bytes); + cudaMalloc((void**)(&d_kernel), kernel_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_input, input.data(), input_bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_kernel, kernel.data(), kernel_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > gpu_input(d_input, Eigen::array(74,37,11,137,17)); + Eigen::TensorMap > gpu_kernel(d_kernel, Eigen::array(3,4,2)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(74,35,8,136,17)); + + Eigen::array dims(1,2,3); + gpu_out.device(gpu_device) = gpu_input.convolve(gpu_kernel, dims); + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 74; ++i) { + for (int j = 0; j < 35; ++j) { + for (int k = 0; k < 8; ++k) { + for (int l = 0; l < 136; ++l) { + for (int m = 0; m < 17; ++m) { + const float result = out(Eigen::array(i,j,k,l,m)); + const float expected = input(Eigen::array(i,j+0,k+0,l+0,m)) * kernel(Eigen::array(0,0,0)) + + input(Eigen::array(i,j+1,k+0,l+0,m)) * kernel(Eigen::array(1,0,0)) + + input(Eigen::array(i,j+2,k+0,l+0,m)) * kernel(Eigen::array(2,0,0)) + + input(Eigen::array(i,j+0,k+1,l+0,m)) * kernel(Eigen::array(0,1,0)) + + input(Eigen::array(i,j+1,k+1,l+0,m)) * kernel(Eigen::array(1,1,0)) + + input(Eigen::array(i,j+2,k+1,l+0,m)) * kernel(Eigen::array(2,1,0)) + + input(Eigen::array(i,j+0,k+2,l+0,m)) * kernel(Eigen::array(0,2,0)) + + input(Eigen::array(i,j+1,k+2,l+0,m)) * kernel(Eigen::array(1,2,0)) + + input(Eigen::array(i,j+2,k+2,l+0,m)) * kernel(Eigen::array(2,2,0)) + + input(Eigen::array(i,j+0,k+3,l+0,m)) * kernel(Eigen::array(0,3,0)) + + input(Eigen::array(i,j+1,k+3,l+0,m)) * kernel(Eigen::array(1,3,0)) + + input(Eigen::array(i,j+2,k+3,l+0,m)) * kernel(Eigen::array(2,3,0)) + + input(Eigen::array(i,j+0,k+0,l+1,m)) * kernel(Eigen::array(0,0,1)) + + input(Eigen::array(i,j+1,k+0,l+1,m)) * kernel(Eigen::array(1,0,1)) + + input(Eigen::array(i,j+2,k+0,l+1,m)) * kernel(Eigen::array(2,0,1)) + + input(Eigen::array(i,j+0,k+1,l+1,m)) * kernel(Eigen::array(0,1,1)) + + input(Eigen::array(i,j+1,k+1,l+1,m)) * kernel(Eigen::array(1,1,1)) + + input(Eigen::array(i,j+2,k+1,l+1,m)) * kernel(Eigen::array(2,1,1)) + + input(Eigen::array(i,j+0,k+2,l+1,m)) * kernel(Eigen::array(0,2,1)) + + input(Eigen::array(i,j+1,k+2,l+1,m)) * kernel(Eigen::array(1,2,1)) + + input(Eigen::array(i,j+2,k+2,l+1,m)) * kernel(Eigen::array(2,2,1)) + + input(Eigen::array(i,j+0,k+3,l+1,m)) * kernel(Eigen::array(0,3,1)) + + input(Eigen::array(i,j+1,k+3,l+1,m)) * kernel(Eigen::array(1,3,1)) + + input(Eigen::array(i,j+2,k+3,l+1,m)) * kernel(Eigen::array(2,3,1)); + VERIFY_IS_APPROX(result, expected); + } + } + } + } + } +} + +static float* CudaCopyFloat(float* data, int size) { + const int nbytes = size * sizeof(float); + float* result = NULL; + if (cudaMalloc((void**)(&result), nbytes) != cudaSuccess) { + return NULL; + } else { + if (data != NULL) { + cudaMemcpy(result, data, nbytes, cudaMemcpyHostToDevice); + } + return result; + } +} + +static void test_cuda_constant_broadcast() +{ + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Tensor t1(10); + for (int i = 0; i < 10; ++i) { + t1(i) = 10.0f * i; + } + float* t1_cuda = CudaCopyFloat(t1.data(), t1.size()); + Eigen::TensorMap > t1_gpu(t1_cuda, 10); + + Tensor t2(1); + t2 = t2.constant(20.0f); + float* t2_cuda = CudaCopyFloat(t2.data(), t2.size()); + Eigen::TensorMap > > t2_gpu(t2_cuda, 1); + + float* t3_cuda = CudaCopyFloat(NULL, 10); + Eigen::TensorMap > t3_gpu(t3_cuda, 10); + + t3_gpu.device(gpu_device) = + t1_gpu + t2_gpu.broadcast(Eigen::array(10)); + + Eigen::Tensor t3(10); + cudaMemcpy(t3.data(), t3_gpu.data(), 10 * sizeof(float), + cudaMemcpyDeviceToHost); + + for (int i = 0; i < 10; ++i) { + VERIFY_IS_APPROX(t3(i), t1(i) + t2(0)); + } +} + +void test_cxx11_tensor_cuda() +{ + CALL_SUBTEST(test_cuda_elementwise_small()); + CALL_SUBTEST(test_cuda_elementwise()); + CALL_SUBTEST(test_cuda_reduction()); + CALL_SUBTEST(test_cuda_contraction()); + CALL_SUBTEST(test_cuda_contraction()); + CALL_SUBTEST(test_cuda_convolution_1d()); + CALL_SUBTEST(test_cuda_convolution_2d()); + CALL_SUBTEST(test_cuda_convolution_3d()); + CALL_SUBTEST(test_cuda_constant_broadcast()); +} diff --git a/unsupported/test/cxx11_tensor_device.cpp b/unsupported/test/cxx11_tensor_device.cpp index 26465ee11..f2d7e4ce6 100644 --- a/unsupported/test/cxx11_tensor_device.cpp +++ b/unsupported/test/cxx11_tensor_device.cpp @@ -22,23 +22,23 @@ using Eigen::RowMajor; // Context for evaluation on cpu struct CPUContext { - CPUContext(const Eigen::Tensor& in1, Eigen::Tensor& in2, Eigen::Tensor& out) : in1_(in1), in2_(in2), out_(out), kernel_1d_(2), kernel_2d_(Eigen::array(2,2)), kernel_3d_(Eigen::array(2,2,2)) { + CPUContext(const Eigen::Tensor& in1, Eigen::Tensor& in2, Eigen::Tensor& out) : in1_(in1), in2_(in2), out_(out), kernel_1d_(2), kernel_2d_(2,2), kernel_3d_(2,2,2) { kernel_1d_(0) = 3.14f; kernel_1d_(1) = 2.7f; - kernel_2d_(Eigen::array(0,0)) = 3.14f; - kernel_2d_(Eigen::array(1,0)) = 2.7f; - kernel_2d_(Eigen::array(0,1)) = 0.2f; - kernel_2d_(Eigen::array(1,1)) = 7.0f; + kernel_2d_(0,0) = 3.14f; + kernel_2d_(1,0) = 2.7f; + kernel_2d_(0,1) = 0.2f; + kernel_2d_(1,1) = 7.0f; - kernel_3d_(Eigen::array(0,0,0)) = 3.14f; - kernel_3d_(Eigen::array(0,1,0)) = 2.7f; - kernel_3d_(Eigen::array(0,0,1)) = 0.2f; - kernel_3d_(Eigen::array(0,1,1)) = 7.0f; - kernel_3d_(Eigen::array(1,0,0)) = -1.0f; - kernel_3d_(Eigen::array(1,1,0)) = -0.3f; - kernel_3d_(Eigen::array(1,0,1)) = -0.7f; - kernel_3d_(Eigen::array(1,1,1)) = -0.5f; + kernel_3d_(0,0,0) = 3.14f; + kernel_3d_(0,1,0) = 2.7f; + kernel_3d_(0,0,1) = 0.2f; + kernel_3d_(0,1,1) = 7.0f; + kernel_3d_(1,0,0) = -1.0f; + kernel_3d_(1,1,0) = -0.3f; + kernel_3d_(1,0,1) = -0.7f; + kernel_3d_(1,1,1) = -0.5f; } const Eigen::DefaultDevice& device() const { return cpu_device_; } @@ -93,8 +93,8 @@ struct GPUContext { const Eigen::TensorMap >& in2() const { return in2_; } Eigen::TensorMap >& out() { return out_; } Eigen::TensorMap > kernel1d() const { return Eigen::TensorMap >(kernel_1d_, 2); } - Eigen::TensorMap > kernel2d() const { return Eigen::TensorMap >(kernel_2d_, Eigen::array(2, 2)); } - Eigen::TensorMap > kernel3d() const { return Eigen::TensorMap >(kernel_3d_, Eigen::array(2, 2, 2)); } + Eigen::TensorMap > kernel2d() const { return Eigen::TensorMap >(kernel_2d_, 2, 2); } + Eigen::TensorMap > kernel3d() const { return Eigen::TensorMap >(kernel_3d_, 2, 2, 2); } private: const Eigen::TensorMap >& in1_; @@ -150,8 +150,8 @@ static void test_contraction(Context* context) template static void test_1d_convolution(Context* context) { - Eigen::DSizes indices(Eigen::array(0,0,0)); - Eigen::DSizes sizes(Eigen::array(40,49,70)); + Eigen::DSizes indices(0,0,0); + Eigen::DSizes sizes(40,49,70); Eigen::array dims(1); context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel1d(), dims); @@ -160,8 +160,8 @@ static void test_1d_convolution(Context* context) template static void test_2d_convolution(Context* context) { - Eigen::DSizes indices(Eigen::array(0,0,0)); - Eigen::DSizes sizes(Eigen::array(40,49,69)); + Eigen::DSizes indices(0,0,0); + Eigen::DSizes sizes(40,49,69); Eigen::array dims(1,2); context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel2d(), dims); @@ -170,8 +170,8 @@ static void test_2d_convolution(Context* context) template static void test_3d_convolution(Context* context) { - Eigen::DSizes indices(Eigen::array(0,0,0)); - Eigen::DSizes sizes(Eigen::array(39,49,69)); + Eigen::DSizes indices(0,0,0); + Eigen::DSizes sizes(39,49,69); Eigen::array dims(0,1,2); context->out().slice(indices, sizes).device(context->device()) = context->in1().convolve(context->kernel3d(), dims); @@ -179,9 +179,9 @@ static void test_3d_convolution(Context* context) static void test_cpu() { - Eigen::Tensor in1(Eigen::array(40,50,70)); - Eigen::Tensor in2(Eigen::array(40,50,70)); - Eigen::Tensor out(Eigen::array(40,50,70)); + Eigen::Tensor in1(40,50,70); + Eigen::Tensor in2(40,50,70); + Eigen::Tensor out(40,50,70); in1 = in1.random() + in1.constant(10.0f); in2 = in2.random() + in2.constant(10.0f); @@ -191,7 +191,7 @@ static void test_cpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 50; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); + VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f); } } } @@ -200,7 +200,7 @@ static void test_cpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 50; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k))) * 3.14f + 2.718f); + VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f); } } } @@ -209,7 +209,7 @@ static void test_cpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 50; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); + VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f); } } } @@ -217,11 +217,11 @@ static void test_cpu() { test_contraction(&context); for (int i = 0; i < 40; ++i) { for (int j = 0; j < 40; ++j) { - const float result = out(Eigen::array(i,j,0)); + const float result = out(i,j,0); float expected = 0; for (int k = 0; k < 50; ++k) { for (int l = 0; l < 70; ++l) { - expected += in1(Eigen::array(i, k, l)) * in2(Eigen::array(j, k, l)); + expected += in1(i, k, l) * in2(j, k, l); } } VERIFY_IS_APPROX(expected, result); @@ -232,7 +232,7 @@ static void test_cpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 49; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f)); + VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f)); } } } @@ -241,9 +241,9 @@ static void test_cpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 49; ++j) { for (int k = 0; k < 69; ++k) { - const float result = out(Eigen::array(i,j,k)); - const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f) + - (in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f); + const float result = out(i,j,k); + const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f) + + (in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f); if (fabs(expected) < 1e-4 && fabs(result) < 1e-4) { continue; } @@ -256,11 +256,11 @@ static void test_cpu() { for (int i = 0; i < 39; ++i) { for (int j = 0; j < 49; ++j) { for (int k = 0; k < 69; ++k) { - const float result = out(Eigen::array(i,j,k)); - const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f + - in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f) + - (in1(Eigen::array(i+1,j,k)) * -1.0f + in1(Eigen::array(i+1,j+1,k)) * -0.3f + - in1(Eigen::array(i+1,j,k+1)) * -0.7f + in1(Eigen::array(i+1,j+1,k+1)) * -0.5f); + const float result = out(i,j,k); + const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f + + in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f) + + (in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f + + in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f); if (fabs(expected) < 1e-4 && fabs(result) < 1e-4) { continue; } @@ -271,9 +271,9 @@ static void test_cpu() { } static void test_gpu() { - Eigen::Tensor in1(Eigen::array(40,50,70)); - Eigen::Tensor in2(Eigen::array(40,50,70)); - Eigen::Tensor out(Eigen::array(40,50,70)); + Eigen::Tensor in1(40,50,70); + Eigen::Tensor in2(40,50,70); + Eigen::Tensor out(40,50,70); in1 = in1.random() + in1.constant(10.0f); in2 = in2.random() + in2.constant(10.0f); @@ -291,9 +291,9 @@ static void test_gpu() { cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice); cudaMemcpy(d_in2, in2.data(), in2_bytes, cudaMemcpyHostToDevice); - Eigen::TensorMap > gpu_in1(d_in1, Eigen::array(40,50,70)); - Eigen::TensorMap > gpu_in2(d_in2, Eigen::array(40,50,70)); - Eigen::TensorMap > gpu_out(d_out, Eigen::array(40,50,70)); + Eigen::TensorMap > gpu_in1(d_in1, 40,50,70); + Eigen::TensorMap > gpu_in2(d_in2, 40,50,70); + Eigen::TensorMap > gpu_out(d_out, 40,50,70); GPUContext context(gpu_in1, gpu_in2, gpu_out); test_contextual_eval(&context); @@ -301,7 +301,7 @@ static void test_gpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 50; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); + VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f); } } } @@ -311,7 +311,7 @@ static void test_gpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 50; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k))) * 3.14f + 2.718f); + VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) + in2(i,j,k)) * 3.14f + 2.718f); } } } @@ -321,7 +321,7 @@ static void test_gpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 50; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), in1(Eigen::array(i,j,k)) + in2(Eigen::array(i,j,k)) * 3.14f + 2.718f); + VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f + 2.718f); } } } @@ -330,11 +330,11 @@ static void test_gpu() { assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess); for (int i = 0; i < 40; ++i) { for (int j = 0; j < 40; ++j) { - const float result = out(Eigen::array(i,j,0)); + const float result = out(i,j,0); float expected = 0; for (int k = 0; k < 50; ++k) { for (int l = 0; l < 70; ++l) { - expected += in1(Eigen::array(i, k, l)) * in2(Eigen::array(j, k, l)); + expected += in1(i, k, l) * in2(j, k, l); } } VERIFY_IS_APPROX(expected, result); @@ -347,7 +347,7 @@ static void test_gpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 49; ++j) { for (int k = 0; k < 70; ++k) { - VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f)); + VERIFY_IS_APPROX(out(i,j,k), (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f)); } } } @@ -358,9 +358,9 @@ static void test_gpu() { for (int i = 0; i < 40; ++i) { for (int j = 0; j < 49; ++j) { for (int k = 0; k < 69; ++k) { - const float result = out(Eigen::array(i,j,k)); - const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f + - in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f); + const float result = out(i,j,k); + const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f + + in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f); VERIFY_IS_APPROX(expected, result); } } @@ -372,11 +372,11 @@ static void test_gpu() { for (int i = 0; i < 39; ++i) { for (int j = 0; j < 49; ++j) { for (int k = 0; k < 69; ++k) { - const float result = out(Eigen::array(i,j,k)); - const float expected = (in1(Eigen::array(i,j,k)) * 3.14f + in1(Eigen::array(i,j+1,k)) * 2.7f + - in1(Eigen::array(i,j,k+1)) * 0.2f + in1(Eigen::array(i,j+1,k+1)) * 7.0f + - in1(Eigen::array(i+1,j,k)) * -1.0f + in1(Eigen::array(i+1,j+1,k)) * -0.3f + - in1(Eigen::array(i+1,j,k+1)) * -0.7f + in1(Eigen::array(i+1,j+1,k+1)) * -0.5f); + const float result = out(i,j,k); + const float expected = (in1(i,j,k) * 3.14f + in1(i,j+1,k) * 2.7f + + in1(i,j,k+1) * 0.2f + in1(i,j+1,k+1) * 7.0f + + in1(i+1,j,k) * -1.0f + in1(i+1,j+1,k) * -0.3f + + in1(i+1,j,k+1) * -0.7f + in1(i+1,j+1,k+1) * -0.5f); VERIFY_IS_APPROX(expected, result); } } diff --git a/unsupported/test/cxx11_tensor_dimension.cpp b/unsupported/test/cxx11_tensor_dimension.cpp index c806b623f..0cc4e86f7 100644 --- a/unsupported/test/cxx11_tensor_dimension.cpp +++ b/unsupported/test/cxx11_tensor_dimension.cpp @@ -16,12 +16,15 @@ using Eigen::Tensor; static void test_dynamic_size() { - Eigen::DSizes dimensions(Eigen::array{{2,3,7}}); + Eigen::DSizes dimensions(2,3,7); VERIFY_IS_EQUAL((int)Eigen::internal::array_get<0>(dimensions), 2); VERIFY_IS_EQUAL((int)Eigen::internal::array_get<1>(dimensions), 3); VERIFY_IS_EQUAL((int)Eigen::internal::array_get<2>(dimensions), 7); VERIFY_IS_EQUAL(dimensions.TotalSize(), (size_t)2*3*7); + VERIFY_IS_EQUAL((int)dimensions[0], 2); + VERIFY_IS_EQUAL((int)dimensions[1], 3); + VERIFY_IS_EQUAL((int)dimensions[2], 7); } static void test_fixed_size() @@ -37,9 +40,9 @@ static void test_fixed_size() static void test_match() { - Eigen::DSizes dyn(Eigen::array{{2,3,7}}); + Eigen::DSizes dyn(2,3,7); Eigen::Sizes<2,3,7> stat; - VERIFY_IS_EQUAL(Eigen::internal::dimensions_match(dyn, stat), true); + VERIFY_IS_EQUAL(Eigen::dimensions_match(dyn, stat), true); } diff --git a/unsupported/test/cxx11_tensor_expr.cpp b/unsupported/test/cxx11_tensor_expr.cpp index e85fcbfa9..792fdeade 100644 --- a/unsupported/test/cxx11_tensor_expr.cpp +++ b/unsupported/test/cxx11_tensor_expr.cpp @@ -125,6 +125,12 @@ static void test_3d() mat7 = mat1.cwiseMax(mat5 * 2.0f).exp(); Tensor mat8(2,3,7); mat8 = (-mat2).exp() * 3.14f; + Tensor mat9(2,3,7); + mat9 = mat2 + 3.14f; + Tensor mat10(2,3,7); + mat10 = mat2 - 3.14f; + Tensor mat11(2,3,7); + mat11 = mat2 / 3.14f; val = 1.0; for (int i = 0; i < 2; ++i) { @@ -136,6 +142,9 @@ static void test_3d() VERIFY_IS_APPROX(mat6(i,j,k), sqrtf(val) * 3.14f); VERIFY_IS_APPROX(mat7(i,j,k), expf((std::max)(val, mat5(i,j,k) * 2.0f))); VERIFY_IS_APPROX(mat8(i,j,k), expf(-val) * 3.14f); + VERIFY_IS_APPROX(mat9(i,j,k), val + 3.14f); + VERIFY_IS_APPROX(mat10(i,j,k), val - 3.14f); + VERIFY_IS_APPROX(mat11(i,j,k), val / 3.14f); val += 1.0; } } @@ -172,6 +181,36 @@ static void test_constants() } } +static void test_boolean() +{ + Tensor vec(6); + std::copy_n(std::begin({0, 1, 2, 3, 4, 5}), 6, vec.data()); + + // Test ||. + Tensor bool1 = vec < vec.constant(1) || vec > vec.constant(4); + VERIFY_IS_EQUAL(bool1[0], true); + VERIFY_IS_EQUAL(bool1[1], false); + VERIFY_IS_EQUAL(bool1[2], false); + VERIFY_IS_EQUAL(bool1[3], false); + VERIFY_IS_EQUAL(bool1[4], false); + VERIFY_IS_EQUAL(bool1[5], true); + + // Test &&, including cast of operand vec. + Tensor bool2 = vec.cast() && vec < vec.constant(4); + VERIFY_IS_EQUAL(bool2[0], false); + VERIFY_IS_EQUAL(bool2[1], true); + VERIFY_IS_EQUAL(bool2[2], true); + VERIFY_IS_EQUAL(bool2[3], true); + VERIFY_IS_EQUAL(bool2[4], false); + VERIFY_IS_EQUAL(bool2[5], false); + + // Compilation tests: + // Test Tensor against results of cast or comparison; verifies that + // CoeffReturnType is set to match Op return type of bool for Unary and Binary + // Ops. + Tensor bool3 = vec.cast() && bool2; + bool3 = vec < vec.constant(4) && bool2; +} static void test_functors() { @@ -258,6 +297,7 @@ void test_cxx11_tensor_expr() CALL_SUBTEST(test_2d()); CALL_SUBTEST(test_3d()); CALL_SUBTEST(test_constants()); + CALL_SUBTEST(test_boolean()); CALL_SUBTEST(test_functors()); CALL_SUBTEST(test_type_casting()); CALL_SUBTEST(test_select()); diff --git a/unsupported/test/cxx11_tensor_forced_eval.cpp b/unsupported/test/cxx11_tensor_forced_eval.cpp index 529584a7b..ad9de867d 100644 --- a/unsupported/test/cxx11_tensor_forced_eval.cpp +++ b/unsupported/test/cxx11_tensor_forced_eval.cpp @@ -45,7 +45,34 @@ static void test_simple() } +static void test_const() +{ + MatrixXf input(3,3); + input.setRandom(); + MatrixXf output = input; + output.rowwise() -= input.colwise().maxCoeff(); + + Eigen::array depth_dim; + depth_dim[0] = 0; + Tensor::Dimensions dims2d; + dims2d[0] = 1; + dims2d[1] = 3; + Eigen::array bcast; + bcast[0] = 3; + bcast[1] = 1; + const TensorMap> input_tensor(input.data(), 3, 3); + Tensor output_tensor= (input_tensor - input_tensor.maximum(depth_dim).eval().reshape(dims2d).broadcast(bcast)); + + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_APPROX(output(i, j), output_tensor(i, j)); + } + } +} + + void test_cxx11_tensor_forced_eval() { CALL_SUBTEST(test_simple()); + CALL_SUBTEST(test_const()); } diff --git a/unsupported/test/cxx11_tensor_image_patch.cpp b/unsupported/test/cxx11_tensor_image_patch.cpp index 55d35eac0..26854f5a4 100644 --- a/unsupported/test/cxx11_tensor_image_patch.cpp +++ b/unsupported/test/cxx11_tensor_image_patch.cpp @@ -28,6 +28,9 @@ static void test_simple_patch() VERIFY_IS_EQUAL(single_pixel_patch.dimension(4), 7); for (int i = 0; i < tensor.size(); ++i) { + if (tensor.data()[i] != single_pixel_patch.data()[i]) { + std::cout << "Mismatch detected at index " << i << " : " << tensor.data()[i] << " vs " << single_pixel_patch.data()[i] << std::endl; + } VERIFY_IS_EQUAL(single_pixel_patch.data()[i], tensor.data()[i]); } @@ -51,6 +54,9 @@ static void test_simple_patch() if (r-1+i >= 0 && c-2+j >= 0 && r-1+i < 3 && c-2+j < 5) { expected = tensor(d, r-1+i, c-2+j, b); } + if (entire_image_patch(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } VERIFY_IS_EQUAL(entire_image_patch(d, r, c, patchId, b), expected); } } @@ -68,6 +74,11 @@ static void test_simple_patch() VERIFY_IS_EQUAL(twod_patch.dimension(3), 3*5); VERIFY_IS_EQUAL(twod_patch.dimension(4), 7); + // Based on the calculation described in TensorTraits.h, padding happens to be 0. + int row_padding = 0; + int col_padding = 0; + int stride = 1; + for (int i = 0; i < 3; ++i) { for (int j = 0; j < 5; ++j) { int patchId = i+3*j; @@ -76,8 +87,13 @@ static void test_simple_patch() for (int d = 0; d < 2; ++d) { for (int b = 0; b < 7; ++b) { float expected = 0.0f; - if (r-1+i >= 0 && c-1+j >= 0 && r-1+i < 3 && c-1+j < 5) { - expected = tensor(d, r-1+i, c-1+j, b); + int row_offset = r*stride + i - row_padding; + int col_offset = c*stride + j - col_padding; + if (row_offset >= 0 && col_offset >= 0 && row_offset < tensor.dimension(1) && col_offset < tensor.dimension(2)) { + expected = tensor(d, row_offset, col_offset, b); + } + if (twod_patch(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; } VERIFY_IS_EQUAL(twod_patch(d, r, c, patchId, b), expected); } @@ -88,6 +104,156 @@ static void test_simple_patch() } } +// Verifies VALID padding (no padding) with incrementing values. +static void test_patch_padding_valid() +{ + int input_depth = 3; + int input_rows = 3; + int input_cols = 3; + int input_batches = 1; + int ksize = 2; // Corresponds to the Rows and Cols for tensor.extract_image_patches<>. + int stride = 2; // Only same stride is supported. + Tensor tensor(input_depth, input_rows, input_cols, input_batches); + // Initializes tensor with incrementing numbers. + for (int i = 0; i < tensor.size(); ++i) { + tensor.data()[i] = i + 1; + } + Tensor result = tensor.extract_image_patches(ksize, ksize, stride, stride, PADDING_VALID); + + VERIFY_IS_EQUAL(result.dimension(0), input_depth); // depth + VERIFY_IS_EQUAL(result.dimension(1), ksize); // kernel rows + VERIFY_IS_EQUAL(result.dimension(2), ksize); // kernel cols + VERIFY_IS_EQUAL(result.dimension(3), 1); // number of patches + VERIFY_IS_EQUAL(result.dimension(4), input_batches); // number of batches + + // No padding is carried out. + int row_padding = 0; + int col_padding = 0; + + for (int i = 0; (i+stride+ksize-1) < input_rows; i += stride) { // input rows + for (int j = 0; (j+stride+ksize-1) < input_cols; j += stride) { // input cols + int patchId = i+input_rows*j; + for (int r = 0; r < ksize; ++r) { // patch rows + for (int c = 0; c < ksize; ++c) { // patch cols + for (int d = 0; d < input_depth; ++d) { // depth + for (int b = 0; b < input_batches; ++b) { // batch + float expected = 0.0f; + int row_offset = r + i - row_padding; + int col_offset = c + j - col_padding; + if (row_offset >= 0 && col_offset >= 0 && row_offset < input_rows && col_offset < input_cols) { + expected = tensor(d, row_offset, col_offset, b); + } + if (result(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } + VERIFY_IS_EQUAL(result(d, r, c, patchId, b), expected); + } + } + } + } + } + } +} + +// Verifies VALID padding (no padding) with the same value. +static void test_patch_padding_valid_same_value() +{ + int input_depth = 1; + int input_rows = 5; + int input_cols = 5; + int input_batches = 2; + int ksize = 3; // Corresponds to the Rows and Cols for tensor.extract_image_patches<>. + int stride = 2; // Only same stride is supported. + Tensor tensor(input_depth, input_rows, input_cols, input_batches); + tensor = tensor.constant(11.0f); + Tensor result = tensor.extract_image_patches(ksize, ksize, stride, stride, PADDING_VALID); + + VERIFY_IS_EQUAL(result.dimension(0), input_depth); // depth + VERIFY_IS_EQUAL(result.dimension(1), ksize); // kernel rows + VERIFY_IS_EQUAL(result.dimension(2), ksize); // kernel cols + VERIFY_IS_EQUAL(result.dimension(3), 4); // number of patches + VERIFY_IS_EQUAL(result.dimension(4), input_batches); // number of batches + + // No padding is carried out. + int row_padding = 0; + int col_padding = 0; + + for (int i = 0; (i+stride+ksize-1) <= input_rows; i += stride) { // input rows + for (int j = 0; (j+stride+ksize-1) <= input_cols; j += stride) { // input cols + int patchId = i+input_rows*j; + for (int r = 0; r < ksize; ++r) { // patch rows + for (int c = 0; c < ksize; ++c) { // patch cols + for (int d = 0; d < input_depth; ++d) { // depth + for (int b = 0; b < input_batches; ++b) { // batch + float expected = 0.0f; + int row_offset = r + i - row_padding; + int col_offset = c + j - col_padding; + if (row_offset >= 0 && col_offset >= 0 && row_offset < input_rows && col_offset < input_cols) { + expected = tensor(d, row_offset, col_offset, b); + } + if (result(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } + VERIFY_IS_EQUAL(result(d, r, c, patchId, b), expected); + } + } + } + } + } + } +} + +// Verifies SAME padding. +static void test_patch_padding_same() +{ + int input_depth = 3; + int input_rows = 4; + int input_cols = 2; + int input_batches = 1; + int ksize = 2; // Corresponds to the Rows and Cols for tensor.extract_image_patches<>. + int stride = 2; // Only same stride is supported. + Tensor tensor(input_depth, input_rows, input_cols, input_batches); + // Initializes tensor with incrementing numbers. + for (int i = 0; i < tensor.size(); ++i) { + tensor.data()[i] = i + 1; + } + Tensor result = tensor.extract_image_patches(ksize, ksize, stride, stride, PADDING_SAME); + + VERIFY_IS_EQUAL(result.dimension(0), input_depth); // depth + VERIFY_IS_EQUAL(result.dimension(1), ksize); // kernel rows + VERIFY_IS_EQUAL(result.dimension(2), ksize); // kernel cols + VERIFY_IS_EQUAL(result.dimension(3), 2); // number of patches + VERIFY_IS_EQUAL(result.dimension(4), input_batches); // number of batches + + // Based on the calculation described in TensorTraits.h, padding happens to be + // 0. + int row_padding = 0; + int col_padding = 0; + + for (int i = 0; (i+stride+ksize-1) <= input_rows; i += stride) { // input rows + for (int j = 0; (j+stride+ksize-1) <= input_cols; j += stride) { // input cols + int patchId = i+input_rows*j; + for (int r = 0; r < ksize; ++r) { // patch rows + for (int c = 0; c < ksize; ++c) { // patch cols + for (int d = 0; d < input_depth; ++d) { // depth + for (int b = 0; b < input_batches; ++b) { // batch + float expected = 0.0f; + int row_offset = r*stride + i - row_padding; + int col_offset = c*stride + j - col_padding; + if (row_offset >= 0 && col_offset >= 0 && row_offset < input_rows && col_offset < input_cols) { + expected = tensor(d, row_offset, col_offset, b); + } + if (result(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } + VERIFY_IS_EQUAL(result(d, r, c, patchId, b), expected); + } + } + } + } + } + } +} static void test_patch_no_extra_dim() { @@ -103,6 +269,9 @@ static void test_patch_no_extra_dim() VERIFY_IS_EQUAL(single_pixel_patch.dimension(3), 3*5); for (int i = 0; i < tensor.size(); ++i) { + if (tensor.data()[i] != single_pixel_patch.data()[i]) { + std::cout << "Mismatch detected at index " << i << " : " << tensor.data()[i] << " vs " << single_pixel_patch.data()[i] << std::endl; + } VERIFY_IS_EQUAL(single_pixel_patch.data()[i], tensor.data()[i]); } @@ -124,6 +293,9 @@ static void test_patch_no_extra_dim() if (r-1+i >= 0 && c-2+j >= 0 && r-1+i < 3 && c-2+j < 5) { expected = tensor(d, r-1+i, c-2+j); } + if (entire_image_patch(d, r, c, patchId) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << std::endl; + } VERIFY_IS_EQUAL(entire_image_patch(d, r, c, patchId), expected); } } @@ -139,6 +311,11 @@ static void test_patch_no_extra_dim() VERIFY_IS_EQUAL(twod_patch.dimension(2), 2); VERIFY_IS_EQUAL(twod_patch.dimension(3), 3*5); + // Based on the calculation described in TensorTraits.h, padding happens to be 0. + int row_padding = 0; + int col_padding = 0; + int stride = 1; + for (int i = 0; i < 3; ++i) { for (int j = 0; j < 5; ++j) { int patchId = i+3*j; @@ -146,8 +323,13 @@ static void test_patch_no_extra_dim() for (int c = 0; c < 2; ++c) { for (int d = 0; d < 2; ++d) { float expected = 0.0f; - if (r-1+i >= 0 && c-1+j >= 0 && r-1+i < 3 && c-1+j < 5) { - expected = tensor(d, r-1+i, c-1+j); + int row_offset = r*stride + i - row_padding; + int col_offset = c*stride + j - col_padding; + if (row_offset >= 0 && col_offset >= 0 && row_offset < tensor.dimension(1) && col_offset < tensor.dimension(2)) { + expected = tensor(d, row_offset, col_offset); + } + if (twod_patch(d, r, c, patchId) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << std::endl; } VERIFY_IS_EQUAL(twod_patch(d, r, c, patchId), expected); } @@ -181,6 +363,9 @@ static void test_imagenet_patches() if (r-5+i >= 0 && c-5+j >= 0 && r-5+i < 128 && c-5+j < 128) { expected = l_in(d, r-5+i, c-5+j, b); } + if (l_out(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected); } } @@ -208,6 +393,9 @@ static void test_imagenet_patches() if (r-4+i >= 0 && c-4+j >= 0 && r-4+i < 64 && c-4+j < 64) { expected = l_in(d, r-4+i, c-4+j, b); } + if (l_out(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected); } } @@ -235,6 +423,9 @@ static void test_imagenet_patches() if (r-3+i >= 0 && c-3+j >= 0 && r-3+i < 16 && c-3+j < 16) { expected = l_in(d, r-3+i, c-3+j, b); } + if (l_out(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected); } } @@ -262,6 +453,9 @@ static void test_imagenet_patches() if (r-1+i >= 0 && c-1+j >= 0 && r-1+i < 13 && c-1+j < 13) { expected = l_in(d, r-1+i, c-1+j, b); } + if (l_out(d, r, c, patchId, b) != expected) { + std::cout << "Mismatch detected at index i=" << i << " j=" << j << " r=" << r << " c=" << c << " d=" << d << " b=" << b << std::endl; + } VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected); } } @@ -271,10 +465,12 @@ static void test_imagenet_patches() } } - void test_cxx11_tensor_image_patch() { CALL_SUBTEST(test_simple_patch()); CALL_SUBTEST(test_patch_no_extra_dim()); + CALL_SUBTEST(test_patch_padding_valid()); + CALL_SUBTEST(test_patch_padding_valid_same_value()); + CALL_SUBTEST(test_patch_padding_same()); CALL_SUBTEST(test_imagenet_patches()); } diff --git a/unsupported/test/cxx11_tensor_map.cpp b/unsupported/test/cxx11_tensor_map.cpp index 478c20306..9cf2eb150 100644 --- a/unsupported/test/cxx11_tensor_map.cpp +++ b/unsupported/test/cxx11_tensor_map.cpp @@ -29,6 +29,7 @@ static void test_1d() vec1(4) = 23; vec2(4) = 4; vec1(5) = 42; vec2(5) = 5; + VERIFY_IS_EQUAL(vec1.rank(), 1); VERIFY_IS_EQUAL(vec1.size(), 6); VERIFY_IS_EQUAL(vec1.dimension(0), 6); @@ -69,10 +70,12 @@ static void test_2d() TensorMap> mat3(mat1.data(), 2, 3); TensorMap> mat4(mat2.data(), 2, 3); + VERIFY_IS_EQUAL(mat3.rank(), 2); VERIFY_IS_EQUAL(mat3.size(), 6); VERIFY_IS_EQUAL(mat3.dimension(0), 2); VERIFY_IS_EQUAL(mat3.dimension(1), 3); + VERIFY_IS_EQUAL(mat4.rank(), 2); VERIFY_IS_EQUAL(mat4.size(), 6); VERIFY_IS_EQUAL(mat4.dimension(0), 2); VERIFY_IS_EQUAL(mat4.dimension(1), 3); @@ -109,13 +112,15 @@ static void test_3d() } TensorMap> mat3(mat1.data(), 2, 3, 7); - TensorMap> mat4(mat2.data(), 2, 3, 7); + TensorMap> mat4(mat2.data(), array{{2, 3, 7}}); + VERIFY_IS_EQUAL(mat3.rank(), 3); VERIFY_IS_EQUAL(mat3.size(), 2*3*7); VERIFY_IS_EQUAL(mat3.dimension(0), 2); VERIFY_IS_EQUAL(mat3.dimension(1), 3); VERIFY_IS_EQUAL(mat3.dimension(2), 7); + VERIFY_IS_EQUAL(mat4.rank(), 3); VERIFY_IS_EQUAL(mat4.size(), 2*3*7); VERIFY_IS_EQUAL(mat4.dimension(0), 2); VERIFY_IS_EQUAL(mat4.dimension(1), 3); diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp index 78b0dade0..b4b0a55b6 100644 --- a/unsupported/test/cxx11_tensor_morphing.cpp +++ b/unsupported/test/cxx11_tensor_morphing.cpp @@ -89,19 +89,19 @@ static void test_reshape_as_lvalue() } } - +template static void test_simple_slice() { - Tensor tensor(2,3,5,7,11); + Tensor tensor(2,3,5,7,11); tensor.setRandom(); - Tensor slice1(1,1,1,1,1); + Tensor slice1(1,1,1,1,1); Eigen::DSizes indices(1,2,3,4,5); Eigen::DSizes sizes(1,1,1,1,1); slice1 = tensor.slice(indices, sizes); VERIFY_IS_EQUAL(slice1(0,0,0,0,0), tensor(1,2,3,4,5)); - Tensor slice2(1,1,2,2,3); + Tensor slice2(1,1,2,2,3); Eigen::DSizes indices2(1,1,3,4,5); Eigen::DSizes sizes2(1,1,2,2,3); slice2 = tensor.slice(indices2, sizes2); @@ -114,7 +114,7 @@ static void test_simple_slice() } } - +// TODO(andydavis) Add RowMajor support when TensorContract supports RowMajor. static void test_slice_in_expr() { MatrixXf m1(7,7); MatrixXf m2(3,3); @@ -141,21 +141,28 @@ static void test_slice_in_expr() { VERIFY_IS_APPROX(res(i,j), m3(i,j)); } } + + // Take an arbitrary slice of an arbitrarily sized tensor. + TensorMap> tensor4(m1.data(), 7, 7); + Tensor tensor6 = tensor4.reshape(DSizes(7*7)).exp().slice(DSizes(0), DSizes(35)); + for (int i = 0; i < 35; ++i) { + VERIFY_IS_APPROX(tensor6(i), expf(tensor4.data()[i])); + } } - +template static void test_slice_as_lvalue() { - Tensor tensor1(2,2,7); + Tensor tensor1(2,2,7); tensor1.setRandom(); - Tensor tensor2(2,2,7); + Tensor tensor2(2,2,7); tensor2.setRandom(); - Tensor tensor3(4,3,5); + Tensor tensor3(4,3,5); tensor3.setRandom(); - Tensor tensor4(4,3,2); + Tensor tensor4(4,3,2); tensor4.setRandom(); - Tensor result(4,5,7); + Tensor result(4,5,7); Eigen::DSizes sizes12(2,2,7); Eigen::DSizes first_slice(0,0,0); result.slice(first_slice, sizes12) = tensor1; @@ -190,10 +197,10 @@ static void test_slice_as_lvalue() } } - +template static void test_slice_raw_data() { - Tensor tensor(3,5,7,11); + Tensor tensor(3,5,7,11); tensor.setRandom(); Eigen::DSizes offsets(1,2,3,4); @@ -203,40 +210,78 @@ static void test_slice_raw_data() VERIFY_IS_EQUAL(slice1.dimensions().TotalSize(), 1ul); VERIFY_IS_EQUAL(slice1.data()[0], tensor(1,2,3,4)); - extents = Eigen::DSizes(2,1,1,1); - auto slice2 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); - VERIFY_IS_EQUAL(slice2.dimensions().TotalSize(), 2ul); - VERIFY_IS_EQUAL(slice2.data()[0], tensor(1,2,3,4)); - VERIFY_IS_EQUAL(slice2.data()[1], tensor(2,2,3,4)); + if (DataLayout == ColMajor) { + extents = Eigen::DSizes(2,1,1,1); + auto slice2 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice2.dimensions().TotalSize(), 2ul); + VERIFY_IS_EQUAL(slice2.data()[0], tensor(1,2,3,4)); + VERIFY_IS_EQUAL(slice2.data()[1], tensor(2,2,3,4)); + } else { + extents = Eigen::DSizes(1,1,1,2); + auto slice2 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice2.dimensions().TotalSize(), 2ul); + VERIFY_IS_EQUAL(slice2.data()[0], tensor(1,2,3,4)); + VERIFY_IS_EQUAL(slice2.data()[1], tensor(1,2,3,5)); + } extents = Eigen::DSizes(1,2,1,1); auto slice3 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); VERIFY_IS_EQUAL(slice3.dimensions().TotalSize(), 2ul); VERIFY_IS_EQUAL(slice3.data(), static_cast(0)); - offsets = Eigen::DSizes(0,2,3,4); - extents = Eigen::DSizes(3,2,1,1); - auto slice4 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); - VERIFY_IS_EQUAL(slice4.dimensions().TotalSize(), 6ul); - for (int i = 0; i < 3; ++i) { - for (int j = 0; j < 2; ++j) { - VERIFY_IS_EQUAL(slice4.data()[i+3*j], tensor(i,2+j,3,4)); + if (DataLayout == ColMajor) { + offsets = Eigen::DSizes(0,2,3,4); + extents = Eigen::DSizes(3,2,1,1); + auto slice4 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice4.dimensions().TotalSize(), 6ul); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 2; ++j) { + VERIFY_IS_EQUAL(slice4.data()[i+3*j], tensor(i,2+j,3,4)); + } + } + } else { + offsets = Eigen::DSizes(1,2,3,0); + extents = Eigen::DSizes(1,1,2,11); + auto slice4 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice4.dimensions().TotalSize(), 22ul); + for (int l = 0; l < 11; ++l) { + for (int k = 0; k < 2; ++k) { + VERIFY_IS_EQUAL(slice4.data()[l+11*k], tensor(1,2,3+k,l)); + } } } - offsets = Eigen::DSizes(0,0,0,4); - extents = Eigen::DSizes(3,5,7,2); - auto slice5 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); - VERIFY_IS_EQUAL(slice5.dimensions().TotalSize(), 210ul); - for (int i = 0; i < 3; ++i) { - for (int j = 0; j < 5; ++j) { - for (int k = 0; k < 7; ++k) { - for (int l = 0; l < 2; ++l) { - int slice_index = i + 3 * (j + 5 * (k + 7 * l)); - VERIFY_IS_EQUAL(slice5.data()[slice_index], tensor(i,j,k,l+4)); + if (DataLayout == ColMajor) { + offsets = Eigen::DSizes(0,0,0,4); + extents = Eigen::DSizes(3,5,7,2); + auto slice5 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice5.dimensions().TotalSize(), 210ul); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 2; ++l) { + int slice_index = i + 3 * (j + 5 * (k + 7 * l)); + VERIFY_IS_EQUAL(slice5.data()[slice_index], tensor(i,j,k,l+4)); + } } } } + } else { + offsets = Eigen::DSizes(1,0,0,0); + extents = Eigen::DSizes(2,5,7,11); + auto slice5 = SliceEvaluator(tensor.slice(offsets, extents), DefaultDevice()); + VERIFY_IS_EQUAL(slice5.dimensions().TotalSize(), 770ul); + for (int l = 0; l < 11; ++l) { + for (int k = 0; k < 7; ++k) { + for (int j = 0; j < 5; ++j) { + for (int i = 0; i < 2; ++i) { + int slice_index = l + 11 * (k + 7 * (j + 5 * i)); + VERIFY_IS_EQUAL(slice5.data()[slice_index], tensor(i+1,j,k,l)); + } + } + } + } + } offsets = Eigen::DSizes(0,0,0,0); @@ -247,14 +292,38 @@ static void test_slice_raw_data() } +static void test_composition() +{ + Eigen::Tensor matrix(7, 11); + matrix.setRandom(); + + const DSizes newDims{{1, 1, 11}}; + Eigen::Tensor tensor = + matrix.slice(DSizes(2, 0), DSizes(1, 11)).reshape(newDims); + + VERIFY_IS_EQUAL(tensor.dimensions().TotalSize(), 11ul); + VERIFY_IS_EQUAL(tensor.dimension(0), 1); + VERIFY_IS_EQUAL(tensor.dimension(1), 1); + VERIFY_IS_EQUAL(tensor.dimension(2), 11); + for (int i = 0; i < 11; ++i) { + VERIFY_IS_EQUAL(tensor(0,0,i), matrix(2,i)); + } +} + + void test_cxx11_tensor_morphing() { CALL_SUBTEST(test_simple_reshape()); CALL_SUBTEST(test_reshape_in_expr()); CALL_SUBTEST(test_reshape_as_lvalue()); - CALL_SUBTEST(test_simple_slice()); + CALL_SUBTEST(test_simple_slice()); + CALL_SUBTEST(test_simple_slice()); CALL_SUBTEST(test_slice_in_expr()); - CALL_SUBTEST(test_slice_as_lvalue()); - CALL_SUBTEST(test_slice_raw_data()); + CALL_SUBTEST(test_slice_as_lvalue()); + CALL_SUBTEST(test_slice_as_lvalue()); + CALL_SUBTEST(test_slice_raw_data()); + CALL_SUBTEST(test_slice_raw_data()); + + CALL_SUBTEST(test_composition()); } diff --git a/unsupported/test/cxx11_tensor_of_strings.cpp b/unsupported/test/cxx11_tensor_of_strings.cpp index 0ffa341c4..8d05d154e 100644 --- a/unsupported/test/cxx11_tensor_of_strings.cpp +++ b/unsupported/test/cxx11_tensor_of_strings.cpp @@ -8,19 +8,18 @@ // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. #include "main.h" -#include + #include -using std::string; using Eigen::Tensor; using Eigen::TensorMap; static void test_assign() { - string data1[6]; - TensorMap> mat1(data1, 2, 3); - string data2[6]; - const TensorMap> mat2(data2, 2, 3); + std::string data1[6]; + TensorMap> mat1(data1, 2, 3); + std::string data2[6]; + const TensorMap> mat2(data2, 2, 3); for (int i = 0; i < 6; ++i) { std::ostringstream s1; @@ -31,16 +30,16 @@ static void test_assign() data2[i] = s2.str(); } - Tensor rslt1; + Tensor rslt1; rslt1 = mat1; - Tensor rslt2; + Tensor rslt2; rslt2 = mat2; - Tensor rslt3 = mat1; - Tensor rslt4 = mat2; + Tensor rslt3 = mat1; + Tensor rslt4 = mat2; - Tensor rslt5(mat1); - Tensor rslt6(mat2); + Tensor rslt5(mat1); + Tensor rslt6(mat2); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { @@ -57,8 +56,8 @@ static void test_assign() static void test_concat() { - Tensor t1(2, 3); - Tensor t2(2, 3); + Tensor t1(2, 3); + Tensor t2(2, 3); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { @@ -71,7 +70,7 @@ static void test_concat() } } - Tensor result = t1.concatenate(t2, 1); + Tensor result = t1.concatenate(t2, 1); VERIFY_IS_EQUAL(result.dimension(0), 2); VERIFY_IS_EQUAL(result.dimension(1), 6); @@ -86,7 +85,7 @@ static void test_concat() static void test_slices() { - Tensor data(2, 6); + Tensor data(2, 6); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { std::ostringstream s1; @@ -99,8 +98,8 @@ static void test_slices() const Eigen::DSizes first_half{{0, 0}}; const Eigen::DSizes second_half{{0, 3}}; - Tensor t1 = data.slice(first_half, half_size); - Tensor t2 = data.slice(second_half, half_size); + Tensor t1 = data.slice(first_half, half_size); + Tensor t2 = data.slice(second_half, half_size); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 3; ++j) { @@ -113,8 +112,8 @@ static void test_slices() static void test_additions() { - Tensor data1(3); - Tensor data2(3); + Tensor data1(3); + Tensor data2(3); for (int i = 0; i < 3; ++i) { data1(i) = "abc"; std::ostringstream s1; @@ -122,16 +121,26 @@ static void test_additions() data2(i) = s1.str(); } - Tensor sum = data1 + data2; + Tensor sum = data1 + data2; for (int i = 0; i < 3; ++i) { std::ostringstream concat; concat << "abc" << i; - string expected = concat.str(); + std::string expected = concat.str(); VERIFY_IS_EQUAL(sum(i), expected); } } +static void test_initialization() +{ + Tensor a(2, 3); + a.setConstant(std::string("foo")); + for (int i = 0; i < 2*3; ++i) { + VERIFY_IS_EQUAL(a(i), std::string("foo")); + } +} + + void test_cxx11_tensor_of_strings() { // Beware: none of this is likely to ever work on a GPU. @@ -139,4 +148,5 @@ void test_cxx11_tensor_of_strings() CALL_SUBTEST(test_concat()); CALL_SUBTEST(test_slices()); CALL_SUBTEST(test_additions()); + CALL_SUBTEST(test_initialization()); } diff --git a/unsupported/test/cxx11_tensor_padding.cpp b/unsupported/test/cxx11_tensor_padding.cpp index 6f74216dd..ffa19896e 100644 --- a/unsupported/test/cxx11_tensor_padding.cpp +++ b/unsupported/test/cxx11_tensor_padding.cpp @@ -13,9 +13,10 @@ using Eigen::Tensor; +template static void test_simple_padding() { - Tensor tensor(2,3,5,7); + Tensor tensor(2,3,5,7); tensor.setRandom(); array, 4> paddings; @@ -24,7 +25,7 @@ static void test_simple_padding() paddings[2] = std::make_pair(3, 4); paddings[3] = std::make_pair(0, 0); - Tensor padded; + Tensor padded; padded = tensor.pad(paddings); VERIFY_IS_EQUAL(padded.dimension(0), 2+0); @@ -47,9 +48,10 @@ static void test_simple_padding() } } +template static void test_padded_expr() { - Tensor tensor(2,3,5,7); + Tensor tensor(2,3,5,7); tensor.setRandom(); array, 4> paddings; @@ -62,17 +64,19 @@ static void test_padded_expr() reshape_dims[0] = 12; reshape_dims[1] = 84; - Tensor result; + Tensor result; result = tensor.pad(paddings).reshape(reshape_dims); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 6; ++j) { for (int k = 0; k < 12; ++k) { for (int l = 0; l < 7; ++l) { + const float result_value = DataLayout == ColMajor ? + result(i+2*j,k+12*l) : result(j+6*i,l+7*k); if (j >= 2 && j < 5 && k >= 3 && k < 8) { - VERIFY_IS_EQUAL(result(i+2*j,k+12*l), tensor(i,j-2,k-3,l)); + VERIFY_IS_EQUAL(result_value, tensor(i,j-2,k-3,l)); } else { - VERIFY_IS_EQUAL(result(i+2*j,k+12*l), 0.0f); + VERIFY_IS_EQUAL(result_value, 0.0f); } } } @@ -80,9 +84,10 @@ static void test_padded_expr() } } - void test_cxx11_tensor_padding() { - CALL_SUBTEST(test_simple_padding()); - CALL_SUBTEST(test_padded_expr()); + CALL_SUBTEST(test_simple_padding()); + CALL_SUBTEST(test_simple_padding()); + CALL_SUBTEST(test_padded_expr()); + CALL_SUBTEST(test_padded_expr()); } diff --git a/unsupported/test/cxx11_tensor_patch.cpp b/unsupported/test/cxx11_tensor_patch.cpp index e2ba5bfd8..0ee7b46d4 100644 --- a/unsupported/test/cxx11_tensor_patch.cpp +++ b/unsupported/test/cxx11_tensor_patch.cpp @@ -36,6 +36,23 @@ static void test_simple_patch() VERIFY_IS_EQUAL(tensor.data()[i], no_patch.data()[i]); } + patch_dims[0] = 2; + patch_dims[1] = 3; + patch_dims[2] = 5; + patch_dims[3] = 7; + Tensor single_patch; + single_patch = tensor.extract_patches(patch_dims); + + VERIFY_IS_EQUAL(single_patch.dimension(0), 2); + VERIFY_IS_EQUAL(single_patch.dimension(1), 3); + VERIFY_IS_EQUAL(single_patch.dimension(2), 5); + VERIFY_IS_EQUAL(single_patch.dimension(3), 7); + VERIFY_IS_EQUAL(single_patch.dimension(4), 1); + + for (int i = 0; i < tensor.size(); ++i) { + VERIFY_IS_EQUAL(tensor.data()[i], single_patch.data()[i]); + } + patch_dims[0] = 1; patch_dims[1] = 2; patch_dims[2] = 2; diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp index da9885166..99e19eba4 100644 --- a/unsupported/test/cxx11_tensor_reduction.cpp +++ b/unsupported/test/cxx11_tensor_reduction.cpp @@ -13,15 +13,15 @@ using Eigen::Tensor; -static void test_simple_reductions() -{ - Tensor tensor(2,3,5,7); +template +static void test_simple_reductions() { + Tensor tensor(2, 3, 5, 7); tensor.setRandom(); array reduction_axis; reduction_axis[0] = 1; reduction_axis[1] = 3; - Tensor result = tensor.sum(reduction_axis); + Tensor result = tensor.sum(reduction_axis); VERIFY_IS_EQUAL(result.dimension(0), 2); VERIFY_IS_EQUAL(result.dimension(1), 5); for (int i = 0; i < 2; ++i) { @@ -36,6 +36,53 @@ static void test_simple_reductions() } } + { + Tensor sum1 = tensor.sum(); + VERIFY_IS_EQUAL(sum1.dimension(0), 1); + + array reduction_axis; + reduction_axis[0] = 0; + reduction_axis[1] = 1; + reduction_axis[2] = 2; + reduction_axis[3] = 3; + Tensor sum2 = tensor.sum(reduction_axis); + VERIFY_IS_EQUAL(sum2.dimension(0), 1); + + VERIFY_IS_APPROX(sum1(0), sum2(0)); + } + + reduction_axis[0] = 0; + reduction_axis[1] = 2; + result = tensor.prod(reduction_axis); + VERIFY_IS_EQUAL(result.dimension(0), 3); + VERIFY_IS_EQUAL(result.dimension(1), 7); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 7; ++j) { + float prod = 1.0f; + for (int k = 0; k < 2; ++k) { + for (int l = 0; l < 5; ++l) { + prod *= tensor(k, i, l, j); + } + } + VERIFY_IS_APPROX(result(i, j), prod); + } + } + + { + Tensor prod1 = tensor.prod(); + VERIFY_IS_EQUAL(prod1.dimension(0), 1); + + array reduction_axis; + reduction_axis[0] = 0; + reduction_axis[1] = 1; + reduction_axis[2] = 2; + reduction_axis[3] = 3; + Tensor prod2 = tensor.prod(reduction_axis); + VERIFY_IS_EQUAL(prod2.dimension(0), 1); + + VERIFY_IS_APPROX(prod1(0), prod2(0)); + } + reduction_axis[0] = 0; reduction_axis[1] = 2; result = tensor.maximum(reduction_axis); @@ -53,6 +100,21 @@ static void test_simple_reductions() } } + { + Tensor max1 = tensor.maximum(); + VERIFY_IS_EQUAL(max1.dimension(0), 1); + + array reduction_axis; + reduction_axis[0] = 0; + reduction_axis[1] = 1; + reduction_axis[2] = 2; + reduction_axis[3] = 3; + Tensor max2 = tensor.maximum(reduction_axis); + VERIFY_IS_EQUAL(max2.dimension(0), 1); + + VERIFY_IS_APPROX(max1(0), max2(0)); + } + reduction_axis[0] = 0; reduction_axis[1] = 1; result = tensor.minimum(reduction_axis); @@ -63,24 +125,72 @@ static void test_simple_reductions() float min_val = (std::numeric_limits::max)(); for (int k = 0; k < 2; ++k) { for (int l = 0; l < 3; ++l) { - min_val = (std::min)(min_val, tensor(k, l, i, j)); + min_val = (std::min)(min_val, tensor(k, l, i, j)); } } VERIFY_IS_APPROX(result(i, j), min_val); } } + + { + Tensor min1 = tensor.minimum(); + VERIFY_IS_EQUAL(min1.dimension(0), 1); + + array reduction_axis; + reduction_axis[0] = 0; + reduction_axis[1] = 1; + reduction_axis[2] = 2; + reduction_axis[3] = 3; + Tensor min2 = tensor.minimum(reduction_axis); + VERIFY_IS_EQUAL(min2.dimension(0), 1); + + VERIFY_IS_APPROX(min1(0), min2(0)); + } + + reduction_axis[0] = 0; + reduction_axis[1] = 1; + result = tensor.mean(reduction_axis); + VERIFY_IS_EQUAL(result.dimension(0), 5); + VERIFY_IS_EQUAL(result.dimension(1), 7); + for (int i = 0; i < 5; ++i) { + for (int j = 0; j < 7; ++j) { + float sum = 0.0f; + int count = 0; + for (int k = 0; k < 2; ++k) { + for (int l = 0; l < 3; ++l) { + sum += tensor(k, l, i, j); + ++count; + } + } + VERIFY_IS_APPROX(result(i, j), sum / count); + } + } + + { + Tensor mean1 = tensor.mean(); + VERIFY_IS_EQUAL(mean1.dimension(0), 1); + + array reduction_axis; + reduction_axis[0] = 0; + reduction_axis[1] = 1; + reduction_axis[2] = 2; + reduction_axis[3] = 3; + Tensor mean2 = tensor.mean(reduction_axis); + VERIFY_IS_EQUAL(mean2.dimension(0), 1); + + VERIFY_IS_APPROX(mean1(0), mean2(0)); + } } - -static void test_full_reductions() -{ - Tensor tensor(2,3); +template +static void test_full_reductions() { + Tensor tensor(2, 3); tensor.setRandom(); array reduction_axis; reduction_axis[0] = 0; reduction_axis[1] = 1; - Tensor result = tensor.sum(reduction_axis); + Tensor result = tensor.sum(reduction_axis); VERIFY_IS_EQUAL(result.dimension(0), 1); float sum = 0.0f; @@ -103,30 +213,26 @@ static void test_full_reductions() VERIFY_IS_APPROX(result(0), sqrtf(sum)); } - struct UserReducer { - UserReducer(float offset) : offset_(offset), sum_(0.0f) {} - void reduce(const float val) { - sum_ += val * val; - } - float finalize() const { - return 1.0f / (sum_ + offset_); - } + static const bool PacketAccess = false; + UserReducer(float offset) : offset_(offset) {} + void reduce(const float val, float* accum) { *accum += val * val; } + float initialize() const { return 0; } + float finalize(const float accum) const { return 1.0f / (accum + offset_); } private: - float offset_; - float sum_; + const float offset_; }; -static void test_user_defined_reductions() -{ - Tensor tensor(5,7); +template +static void test_user_defined_reductions() { + Tensor tensor(5, 7); tensor.setRandom(); array reduction_axis; reduction_axis[0] = 1; UserReducer reducer(10.0f); - Tensor result = tensor.reduce(reduction_axis, reducer); + Tensor result = tensor.reduce(reduction_axis, reducer); VERIFY_IS_EQUAL(result.dimension(0), 5); for (int i = 0; i < 5; ++i) { float expected = 10.0f; @@ -138,22 +244,24 @@ static void test_user_defined_reductions() } } - -static void test_tensor_maps() -{ - int inputs[2*3*5*7]; - TensorMap > tensor_map(inputs, 2,3,5,7); - TensorMap > tensor_map_const(inputs, 2,3,5,7); - const TensorMap > tensor_map_const_const(inputs, 2,3,5,7); +template +static void test_tensor_maps() { + int inputs[2 * 3 * 5 * 7]; + TensorMap > tensor_map(inputs, 2, 3, 5, 7); + TensorMap > tensor_map_const(inputs, 2, 3, 5, + 7); + const TensorMap > tensor_map_const_const( + inputs, 2, 3, 5, 7); tensor_map.setRandom(); array reduction_axis; reduction_axis[0] = 1; reduction_axis[1] = 3; - Tensor result = tensor_map.sum(reduction_axis); - Tensor result2 = tensor_map_const.sum(reduction_axis); - Tensor result3 = tensor_map_const_const.sum(reduction_axis); + Tensor result = tensor_map.sum(reduction_axis); + Tensor result2 = tensor_map_const.sum(reduction_axis); + Tensor result3 = + tensor_map_const_const.sum(reduction_axis); for (int i = 0; i < 2; ++i) { for (int j = 0; j < 5; ++j) { @@ -170,11 +278,110 @@ static void test_tensor_maps() } } +template +static void test_static_dims() { + Tensor in(72, 53, 97, 113); + Tensor out(72, 97); + in.setRandom(); -void test_cxx11_tensor_reduction() -{ - CALL_SUBTEST(test_simple_reductions()); - CALL_SUBTEST(test_full_reductions()); - CALL_SUBTEST(test_user_defined_reductions()); - CALL_SUBTEST(test_tensor_maps()); +#if __cplusplus <= 199711L + array reduction_axis; + reduction_axis[0] = 1; + reduction_axis[1] = 3; +#else + Eigen::IndexList, Eigen::type2index<3> > reduction_axis; +#endif + + out = in.maximum(reduction_axis); + + for (int i = 0; i < 72; ++i) { + for (int j = 0; j < 97; ++j) { + float expected = -1e10f; + for (int k = 0; k < 53; ++k) { + for (int l = 0; l < 113; ++l) { + expected = (std::max)(expected, in(i, k, j, l)); + } + } + VERIFY_IS_APPROX(out(i, j), expected); + } + } +} + +template +static void test_innermost_last_dims() { + Tensor in(72, 53, 97, 113); + Tensor out(97, 113); + in.setRandom(); + +// Reduce on the innermost dimensions. +#if __cplusplus <= 199711L + array reduction_axis; + reduction_axis[0] = 0; + reduction_axis[1] = 1; +#else + // This triggers the use of packets for ColMajor. + Eigen::IndexList, Eigen::type2index<1> > reduction_axis; +#endif + + out = in.maximum(reduction_axis); + + for (int i = 0; i < 97; ++i) { + for (int j = 0; j < 113; ++j) { + float expected = -1e10f; + for (int k = 0; k < 53; ++k) { + for (int l = 0; l < 72; ++l) { + expected = (std::max)(expected, in(l, k, i, j)); + } + } + VERIFY_IS_APPROX(out(i, j), expected); + } + } +} + +template +static void test_innermost_first_dims() { + Tensor in(72, 53, 97, 113); + Tensor out(72, 53); + in.setRandom(); + +// Reduce on the innermost dimensions. +#if __cplusplus <= 199711L + array reduction_axis; + reduction_axis[0] = 2; + reduction_axis[1] = 3; +#else + // This triggers the use of packets for RowMajor. + Eigen::IndexList, Eigen::type2index<3>> reduction_axis; +#endif + + out = in.maximum(reduction_axis); + + for (int i = 0; i < 72; ++i) { + for (int j = 0; j < 53; ++j) { + float expected = -1e10f; + for (int k = 0; k < 97; ++k) { + for (int l = 0; l < 113; ++l) { + expected = (std::max)(expected, in(i, j, k, l)); + } + } + VERIFY_IS_APPROX(out(i, j), expected); + } + } +} + +void test_cxx11_tensor_reduction() { + CALL_SUBTEST(test_simple_reductions()); + CALL_SUBTEST(test_simple_reductions()); + CALL_SUBTEST(test_full_reductions()); + CALL_SUBTEST(test_full_reductions()); + CALL_SUBTEST(test_user_defined_reductions()); + CALL_SUBTEST(test_user_defined_reductions()); + CALL_SUBTEST(test_tensor_maps()); + CALL_SUBTEST(test_tensor_maps()); + CALL_SUBTEST(test_static_dims()); + CALL_SUBTEST(test_static_dims()); + CALL_SUBTEST(test_innermost_last_dims()); + CALL_SUBTEST(test_innermost_last_dims()); + CALL_SUBTEST(test_innermost_first_dims()); + CALL_SUBTEST(test_innermost_first_dims()); } diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp index 39c623499..ec623e1f9 100644 --- a/unsupported/test/cxx11_tensor_shuffling.cpp +++ b/unsupported/test/cxx11_tensor_shuffling.cpp @@ -14,9 +14,10 @@ using Eigen::Tensor; using Eigen::array; +template static void test_simple_shuffling() { - Tensor tensor(2,3,5,7); + Tensor tensor(2,3,5,7); tensor.setRandom(); array shuffles; shuffles[0] = 0; @@ -24,7 +25,7 @@ static void test_simple_shuffling() shuffles[2] = 2; shuffles[3] = 3; - Tensor no_shuffle; + Tensor no_shuffle; no_shuffle = tensor.shuffle(shuffles); VERIFY_IS_EQUAL(no_shuffle.dimension(0), 2); @@ -46,7 +47,7 @@ static void test_simple_shuffling() shuffles[1] = 3; shuffles[2] = 1; shuffles[3] = 0; - Tensor shuffle; + Tensor shuffle; shuffle = tensor.shuffle(shuffles); VERIFY_IS_EQUAL(shuffle.dimension(0), 5); @@ -66,9 +67,10 @@ static void test_simple_shuffling() } +template static void test_expr_shuffling() { - Tensor tensor(2,3,5,7); + Tensor tensor(2,3,5,7); tensor.setRandom(); array shuffles; @@ -76,10 +78,10 @@ static void test_expr_shuffling() shuffles[1] = 3; shuffles[2] = 1; shuffles[3] = 0; - Tensor expected; + Tensor expected; expected = tensor.shuffle(shuffles); - Tensor result(5,7,3,2); + Tensor result(5,7,3,2); array src_slice_dim{{2,3,1,7}}; array src_slice_start{{0,0,0,0}}; @@ -128,16 +130,17 @@ static void test_expr_shuffling() } +template static void test_shuffling_as_value() { - Tensor tensor(2,3,5,7); + Tensor tensor(2,3,5,7); tensor.setRandom(); array shuffles; shuffles[2] = 0; shuffles[3] = 1; shuffles[1] = 2; shuffles[0] = 3; - Tensor shuffle(5,7,3,2); + Tensor shuffle(5,7,3,2); shuffle.shuffle(shuffles) = tensor; VERIFY_IS_EQUAL(shuffle.dimension(0), 5); @@ -158,7 +161,10 @@ static void test_shuffling_as_value() void test_cxx11_tensor_shuffling() { - CALL_SUBTEST(test_simple_shuffling()); - CALL_SUBTEST(test_expr_shuffling()); - CALL_SUBTEST(test_shuffling_as_value()); + CALL_SUBTEST(test_simple_shuffling()); + CALL_SUBTEST(test_simple_shuffling()); + CALL_SUBTEST(test_expr_shuffling()); + CALL_SUBTEST(test_expr_shuffling()); + CALL_SUBTEST(test_shuffling_as_value()); + CALL_SUBTEST(test_shuffling_as_value()); } diff --git a/unsupported/test/cxx11_tensor_simple.cpp b/unsupported/test/cxx11_tensor_simple.cpp index a70591c82..23855fca0 100644 --- a/unsupported/test/cxx11_tensor_simple.cpp +++ b/unsupported/test/cxx11_tensor_simple.cpp @@ -32,6 +32,7 @@ static void test_1d() vec1(5) = 42; vec2(5) = 5; vec3(5) = 0; vec4.setZero(); + VERIFY_IS_EQUAL((vec1.rank()), 1); VERIFY_IS_EQUAL((vec1.size()), 6); VERIFY_IS_EQUAL((vec1.dimensions()[0]), 6); @@ -99,10 +100,12 @@ static void test_2d() mat2(1,1) = 4; mat2(1,2) = 5; + VERIFY_IS_EQUAL((mat1.rank()), 2); VERIFY_IS_EQUAL((mat1.size()), 6); VERIFY_IS_EQUAL((mat1.dimensions()[0]), 2); VERIFY_IS_EQUAL((mat1.dimensions()[1]), 3); + VERIFY_IS_EQUAL((mat2.rank()), 2); VERIFY_IS_EQUAL((mat2.size()), 6); VERIFY_IS_EQUAL((mat2.dimensions()[0]), 2); VERIFY_IS_EQUAL((mat2.dimensions()[1]), 3); diff --git a/unsupported/test/cxx11_tensor_striding.cpp b/unsupported/test/cxx11_tensor_striding.cpp index 502569d1d..1feb39dca 100644 --- a/unsupported/test/cxx11_tensor_striding.cpp +++ b/unsupported/test/cxx11_tensor_striding.cpp @@ -13,9 +13,10 @@ using Eigen::Tensor; +template static void test_simple_striding() { - Tensor tensor(2,3,5,7); + Tensor tensor(2,3,5,7); tensor.setRandom(); array strides; strides[0] = 1; @@ -23,7 +24,7 @@ static void test_simple_striding() strides[2] = 1; strides[3] = 1; - Tensor no_stride; + Tensor no_stride; no_stride = tensor.stride(strides); VERIFY_IS_EQUAL(no_stride.dimension(0), 2); @@ -45,7 +46,7 @@ static void test_simple_striding() strides[1] = 4; strides[2] = 2; strides[3] = 3; - Tensor stride; + Tensor stride; stride = tensor.stride(strides); VERIFY_IS_EQUAL(stride.dimension(0), 1); @@ -65,7 +66,36 @@ static void test_simple_striding() } +template +static void test_striding_as_lvalue() +{ + Tensor tensor(2,3,5,7); + tensor.setRandom(); + array strides; + strides[0] = 2; + strides[1] = 4; + strides[2] = 2; + strides[3] = 3; + + Tensor result(3, 12, 10, 21); + result.stride(strides) = tensor; + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(tensor(i,j,k,l), result(2*i,4*j,2*k,3*l)); + } + } + } + } +} + + void test_cxx11_tensor_striding() { - CALL_SUBTEST(test_simple_striding()); + CALL_SUBTEST(test_simple_striding()); + CALL_SUBTEST(test_simple_striding()); + CALL_SUBTEST(test_striding_as_lvalue()); + CALL_SUBTEST(test_striding_as_lvalue()); } diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp index f0de61f8b..e25912279 100644 --- a/unsupported/test/cxx11_tensor_thread_pool.cpp +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -9,10 +9,10 @@ #define EIGEN_USE_THREADS -#include -#include "main.h" -#include +#include "main.h" +#include +#include using Eigen::Tensor; @@ -60,12 +60,12 @@ static void test_multithread_compound_assignment() } } - +template static void test_multithread_contraction() { - Tensor t_left(30, 50, 37, 31); - Tensor t_right(37, 31, 70, 2, 10); - Tensor t_result(30, 50, 70, 2, 10); + Tensor t_left(30, 50, 37, 31); + Tensor t_right(37, 31, 70, 2, 10); + Tensor t_result(30, 50, 70, 2, 10); t_left.setRandom(); t_right.setRandom(); @@ -74,11 +74,10 @@ static void test_multithread_contraction() typedef Tensor::DimensionPair DimPair; Eigen::array dims({{DimPair(2, 0), DimPair(3, 1)}}); - - typedef Map MapXf; + typedef Map> MapXf; MapXf m_left(t_left.data(), 1500, 1147); MapXf m_right(t_right.data(), 1147, 1400); - MatrixXf m_result(1500, 1400); + Matrix m_result(1500, 1400); Eigen::ThreadPoolDevice thread_pool_device(4); @@ -95,12 +94,12 @@ static void test_multithread_contraction() } } - +template static void test_contraction_corner_cases() { - Tensor t_left(32, 500); - Tensor t_right(32, 28*28); - Tensor t_result(500, 28*28); + Tensor t_left(32, 500); + Tensor t_right(32, 28*28); + Tensor t_result(500, 28*28); t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f; t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f; @@ -110,10 +109,10 @@ static void test_contraction_corner_cases() typedef Tensor::DimensionPair DimPair; Eigen::array dims{{DimPair(0, 0)}}; - typedef Map MapXf; + typedef Map> MapXf; MapXf m_left(t_left.data(), 32, 500); MapXf m_right(t_right.data(), 32, 28*28); - MatrixXf m_result(500, 28*28); + Matrix m_result(500, 28*28); Eigen::ThreadPoolDevice thread_pool_device(12); @@ -181,18 +180,18 @@ static void test_contraction_corner_cases() } } - +template static void test_multithread_contraction_agrees_with_singlethread() { int contract_size = internal::random(1, 5000); - Tensor left(internal::random(1, 80), - contract_size, - internal::random(1, 100)); + Tensor left(internal::random(1, 80), + contract_size, + internal::random(1, 100)); - Tensor right(internal::random(1, 25), - internal::random(1, 37), - contract_size, - internal::random(1, 51)); + Tensor right(internal::random(1, 25), + internal::random(1, 37), + contract_size, + internal::random(1, 51)); left.setRandom(); right.setRandom(); @@ -206,13 +205,13 @@ static void test_multithread_contraction_agrees_with_singlethread() { Eigen::ThreadPoolDevice thread_pool_device(internal::random(2, 11)); - Tensor st_result; + Tensor st_result; st_result = left.contract(right, dims); - Tensor tp_result(st_result.dimensions()); + Tensor tp_result(st_result.dimensions()); tp_result.device(thread_pool_device) = left.contract(right, dims); - VERIFY(internal::dimensions_match(st_result.dimensions(), tp_result.dimensions())); + VERIFY(dimensions_match(st_result.dimensions(), tp_result.dimensions())); for (ptrdiff_t i = 0; i < st_result.size(); i++) { // if both of the values are very small, then do nothing (because the test will fail // due to numerical precision issues when values are small) @@ -241,17 +240,30 @@ static void test_memcpy() { } +static void test_multithread_random() +{ + Eigen::ThreadPoolDevice device(2); + Tensor t(1 << 20); + t.device(device) = t.random>(); +} + + void test_cxx11_tensor_thread_pool() { CALL_SUBTEST(test_multithread_elementwise()); CALL_SUBTEST(test_multithread_compound_assignment()); - CALL_SUBTEST(test_multithread_contraction()); + CALL_SUBTEST(test_multithread_contraction()); + CALL_SUBTEST(test_multithread_contraction()); - CALL_SUBTEST(test_multithread_contraction_agrees_with_singlethread()); + CALL_SUBTEST(test_multithread_contraction_agrees_with_singlethread()); + CALL_SUBTEST(test_multithread_contraction_agrees_with_singlethread()); // Exercise various cases that have been problematic in the past. - CALL_SUBTEST(test_contraction_corner_cases()); + CALL_SUBTEST(test_contraction_corner_cases()); + CALL_SUBTEST(test_contraction_corner_cases()); CALL_SUBTEST(test_memcpy()); + + CALL_SUBTEST(test_multithread_random()); } From 641e824c56db8fffb2f6091d18f913e040c1ea95 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 15 Jan 2015 11:11:48 -0800 Subject: [PATCH 1033/1103] Added cube() operation --- unsupported/Eigen/CXX11/src/Tensor/TensorBase.h | 6 ++++++ unsupported/test/cxx11_tensor_expr.cpp | 10 ++++++++++ 2 files changed, 16 insertions(+) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index 8860f622b..e08ac6aa1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h @@ -83,6 +83,12 @@ class TensorBase return unaryExpr(internal::scalar_square_op()); } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> + cube() const { + return unaryExpr(internal::scalar_cube_op()); + } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorCwiseUnaryOp, const Derived> inverse() const { diff --git a/unsupported/test/cxx11_tensor_expr.cpp b/unsupported/test/cxx11_tensor_expr.cpp index 792fdeade..695565e9b 100644 --- a/unsupported/test/cxx11_tensor_expr.cpp +++ b/unsupported/test/cxx11_tensor_expr.cpp @@ -32,6 +32,9 @@ static void test_1d() float data4[6]; TensorMap> vec4(data4, 6); vec4 = vec2.square(); + float data5[6]; + TensorMap> vec5(data5, 6); + vec5 = vec2.cube(); VERIFY_IS_APPROX(vec3(0), sqrtf(4.0)); VERIFY_IS_APPROX(vec3(1), sqrtf(8.0)); @@ -47,6 +50,13 @@ static void test_1d() VERIFY_IS_APPROX(vec4(4), 4.0f * 4.0f); VERIFY_IS_APPROX(vec4(5), 5.0f * 5.0f); + VERIFY_IS_APPROX(vec5(0), 0.0f); + VERIFY_IS_APPROX(vec5(1), 1.0f); + VERIFY_IS_APPROX(vec5(2), 2.0f * 2.0f * 2.0f); + VERIFY_IS_APPROX(vec5(3), 3.0f * 3.0f * 3.0f); + VERIFY_IS_APPROX(vec5(4), 4.0f * 4.0f * 4.0f); + VERIFY_IS_APPROX(vec5(5), 5.0f * 5.0f * 5.0f); + vec3 = vec1 + vec2; VERIFY_IS_APPROX(vec3(0), 4.0f + 0.0f); VERIFY_IS_APPROX(vec3(1), 8.0f + 1.0f); From 14f537c296710173c76379d8efec59bfb1d78eb7 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 16 Jan 2015 09:09:23 -0800 Subject: [PATCH 1034/1103] gcc doesn't consider that template TensorStridingOp& operator = (const OtherDerived& other) provides a valid assignment operator for the striding operation, and therefore refuses to compile code like: result.stride(foo) = source.stride(bar); Added the explicit TensorStridingOp& operator = (const TensorStridingOp& other) as a workaround to get the code to compile, and did the same in all the operations that can be used as lvalues. --- .../Eigen/CXX11/src/Tensor/TensorChipping.h | 10 +++++++++ .../Eigen/CXX11/src/Tensor/TensorMorphing.h | 19 +++++++++++++++++ .../Eigen/CXX11/src/Tensor/TensorShuffling.h | 9 ++++++++ .../Eigen/CXX11/src/Tensor/TensorStriding.h | 9 ++++++++ unsupported/test/cxx11_tensor_chipping.cpp | 21 +++++++++++++++++++ unsupported/test/cxx11_tensor_morphing.cpp | 13 ++++++++++++ unsupported/test/cxx11_tensor_shuffling.cpp | 17 +++++++++++++++ unsupported/test/cxx11_tensor_striding.cpp | 18 ++++++++++++++++ 8 files changed, 116 insertions(+) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h index bc336e488..503803d23 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h @@ -101,6 +101,16 @@ class TensorChippingOp : public TensorBase > const typename internal::remove_all::type& expression() const { return m_xpr; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorChippingOp& operator = (const TensorChippingOp& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + static const bool Vectorize = TensorEvaluator::PacketAccess; + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp& operator = (const OtherDerived& other) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 23b595ac3..87a4b0758 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -73,6 +73,15 @@ class TensorReshapingOp : public TensorBase::type& expression() const { return m_xpr; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorReshapingOp& operator = (const TensorReshapingOp& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReshapingOp& operator = (const OtherDerived& other) @@ -257,6 +266,16 @@ class TensorSlicingOp : public TensorBase Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + + protected: typename XprType::Nested m_xpr; const StartIndices m_indices; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h index ab5fc6a69..620a63ae7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h @@ -73,6 +73,15 @@ class TensorShufflingOp : public TensorBase const typename internal::remove_all::type& expression() const { return m_xpr; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorShufflingOp& operator = (const TensorShufflingOp& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp& operator = (const OtherDerived& other) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h index 2fbdfadfe..5aa2c8d3b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h @@ -73,6 +73,15 @@ class TensorStridingOp : public TensorBase > const typename internal::remove_all::type& expression() const { return m_xpr; } + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorStridingOp& operator = (const TensorStridingOp& other) + { + typedef TensorAssignOp Assign; + Assign assign(*this, other); + internal::TensorExecutor::run(assign, DefaultDevice()); + return *this; + } + template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorStridingOp& operator = (const OtherDerived& other) diff --git a/unsupported/test/cxx11_tensor_chipping.cpp b/unsupported/test/cxx11_tensor_chipping.cpp index 0de7bbac6..d83417872 100644 --- a/unsupported/test/cxx11_tensor_chipping.cpp +++ b/unsupported/test/cxx11_tensor_chipping.cpp @@ -318,8 +318,29 @@ static void test_chip_as_lvalue() } } } + + Tensor input7(2,3,5,7,11); + input7.setRandom(); + tensor = input1; + tensor.chip(0, 0) = input7.chip(0, 0); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (i != 0) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input7(i,j,k,l,m)); + } + } + } + } + } + } } + template static void test_chip_raw_data() { diff --git a/unsupported/test/cxx11_tensor_morphing.cpp b/unsupported/test/cxx11_tensor_morphing.cpp index b4b0a55b6..7fd7a283a 100644 --- a/unsupported/test/cxx11_tensor_morphing.cpp +++ b/unsupported/test/cxx11_tensor_morphing.cpp @@ -161,6 +161,8 @@ static void test_slice_as_lvalue() tensor3.setRandom(); Tensor tensor4(4,3,2); tensor4.setRandom(); + Tensor tensor5(10,13,12); + tensor5.setRandom(); Tensor result(4,5,7); Eigen::DSizes sizes12(2,2,7); @@ -195,6 +197,17 @@ static void test_slice_as_lvalue() } } } + + Eigen::DSizes sizes5(4,5,7); + Eigen::DSizes fifth_slice(0,0,0); + result.slice(fifth_slice, sizes5) = tensor5.slice(fifth_slice, sizes5); + for (int i = 0; i < 4; ++i) { + for (int j = 2; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + VERIFY_IS_EQUAL(result(i,j,k), tensor5(i,j,k)); + } + } + } } template diff --git a/unsupported/test/cxx11_tensor_shuffling.cpp b/unsupported/test/cxx11_tensor_shuffling.cpp index ec623e1f9..2f7fd9e50 100644 --- a/unsupported/test/cxx11_tensor_shuffling.cpp +++ b/unsupported/test/cxx11_tensor_shuffling.cpp @@ -157,6 +157,23 @@ static void test_shuffling_as_value() } } } + + array no_shuffle; + no_shuffle[0] = 0; + no_shuffle[1] = 1; + no_shuffle[2] = 2; + no_shuffle[3] = 3; + Tensor shuffle2(5,7,3,2); + shuffle2.shuffle(shuffles) = tensor.shuffle(no_shuffle); + for (int i = 0; i < 5; ++i) { + for (int j = 0; j < 7; ++j) { + for (int k = 0; k < 3; ++k) { + for (int l = 0; l < 2; ++l) { + VERIFY_IS_EQUAL(shuffle2(i,j,k,l), shuffle(i,j,k,l)); + } + } + } + } } void test_cxx11_tensor_shuffling() diff --git a/unsupported/test/cxx11_tensor_striding.cpp b/unsupported/test/cxx11_tensor_striding.cpp index 1feb39dca..935b908cc 100644 --- a/unsupported/test/cxx11_tensor_striding.cpp +++ b/unsupported/test/cxx11_tensor_striding.cpp @@ -89,6 +89,24 @@ static void test_striding_as_lvalue() } } } + + array no_strides; + no_strides[0] = 1; + no_strides[1] = 1; + no_strides[2] = 1; + no_strides[3] = 1; + Tensor result2(3, 12, 10, 21); + result2.stride(strides) = tensor.stride(no_strides); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(tensor(i,j,k,l), result2(2*i,4*j,2*k,3*l)); + } + } + } + } } From b9d314ae19b1c857adfe42b7eaecf2695428f0ed Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 17 Jan 2015 21:55:33 +0100 Subject: [PATCH 1035/1103] bug #329: fix typo --- Eigen/src/Eigenvalues/RealQZ.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Eigenvalues/RealQZ.h b/Eigen/src/Eigenvalues/RealQZ.h index ae10ff91e..128ef9028 100644 --- a/Eigen/src/Eigenvalues/RealQZ.h +++ b/Eigen/src/Eigenvalues/RealQZ.h @@ -313,7 +313,7 @@ namespace Eigen { using std::abs; using std::sqrt; const Index dim=m_S.cols(); - if (abs(m_S.coeff(i+1,i)==Scalar(0))) + if (abs(m_S.coeff(i+1,i))==Scalar(0)) return; Index z = findSmallDiagEntry(i,i+1); if (z==i-1) From e1f1091fde660581d64b54ff1019bc494dbbca89 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 24 Jan 2015 10:32:49 +0100 Subject: [PATCH 1036/1103] Add support for dense ?= diagonal --- Eigen/src/Core/DiagonalMatrix.h | 6 ++++++ test/diagonalmatrices.cpp | 7 +++++++ 2 files changed, 13 insertions(+) diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index e3dc71336..49b9b7925 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -326,6 +326,12 @@ struct Assignment dst.setZero(); dst.diagonal() = src.diagonal(); } + + static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op &/*func*/) + { dst.diagonal() += src.diagonal(); } + + static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op &/*func*/) + { dst.diagonal() -= src.diagonal(); } }; } // namespace internal diff --git a/test/diagonalmatrices.cpp b/test/diagonalmatrices.cpp index 149f1db2f..0227ba577 100644 --- a/test/diagonalmatrices.cpp +++ b/test/diagonalmatrices.cpp @@ -84,6 +84,13 @@ template void diagonalmatrices(const MatrixType& m) VERIFY_IS_APPROX(m1 * (rdm1 * s1), (m1 * rdm1) * s1); VERIFY_IS_APPROX(m1 * (s1 * rdm1), (m1 * rdm1) * s1); + + // Diagonal to dense + sq_m1.setRandom(); + sq_m2 = sq_m1; + VERIFY_IS_APPROX( (sq_m1 += (s1*v1).asDiagonal()), sq_m2 += (s1*v1).asDiagonal().toDenseMatrix() ); + VERIFY_IS_APPROX( (sq_m1 -= (s1*v1).asDiagonal()), sq_m2 -= (s1*v1).asDiagonal().toDenseMatrix() ); + VERIFY_IS_APPROX( (sq_m1 = (s1*v1).asDiagonal()), (s1*v1).asDiagonal().toDenseMatrix() ); } void test_diagonalmatrices() From c6eb84aabcf102aaa3ba1c288e890984f4b49277 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 26 Jan 2015 17:09:01 +0100 Subject: [PATCH 1037/1103] Enable vectorization of transposeInPlace for PacketSize x PacketSize matrices --- Eigen/src/Core/Transpose.h | 27 ++++++++++++++++++++++++--- test/adjoint.cpp | 22 ++++++++++++++++++++++ 2 files changed, 46 insertions(+), 3 deletions(-) diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index a3b95256f..3bab6092c 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -217,18 +217,39 @@ MatrixBase::adjoint() const namespace internal { template + bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic, + bool MatchPacketSize = + (int(MatrixType::RowsAtCompileTime) == int(internal::packet_traits::size)) + && (internal::evaluator::Flags&PacketAccessBit) > struct inplace_transpose_selector; template -struct inplace_transpose_selector { // square matrix +struct inplace_transpose_selector { // square matrix static void run(MatrixType& m) { m.matrix().template triangularView().swap(m.matrix().transpose()); } }; +// TODO: vectorized path is currently limited to LargestPacketSize x LargestPacketSize cases only. template -struct inplace_transpose_selector { // non square matrix +struct inplace_transpose_selector { // PacketSize x PacketSize + static void run(MatrixType& m) { + typedef typename MatrixType::Scalar Scalar; + typedef typename internal::packet_traits::type Packet; + typedef typename MatrixType::Index Index; + const Index PacketSize = internal::packet_traits::size; + const Index Alignment = internal::evaluator::Flags&AlignedBit ? Aligned : Unaligned; + PacketBlock A; + for (Index i=0; i(i,0); + internal::ptranspose(A); + for (Index i=0; i(m.rowIndexByOuterInner(i,0), m.colIndexByOuterInner(i,0), A.packet[i]); + } +}; + +template +struct inplace_transpose_selector { // non square matrix static void run(MatrixType& m) { if (m.rows()==m.cols()) m.matrix().template triangularView().swap(m.matrix().transpose()); diff --git a/test/adjoint.cpp b/test/adjoint.cpp index ea36f7841..3b2a53c91 100644 --- a/test/adjoint.cpp +++ b/test/adjoint.cpp @@ -64,6 +64,7 @@ template void adjoint(const MatrixType& m) typedef typename NumTraits::Real RealScalar; typedef Matrix VectorType; typedef Matrix SquareMatrixType; + const Index PacketSize = internal::packet_traits::size; Index rows = m.rows(); Index cols = m.cols(); @@ -108,6 +109,17 @@ template void adjoint(const MatrixType& m) VERIFY_IS_APPROX(m3,m1.transpose()); m3.transposeInPlace(); VERIFY_IS_APPROX(m3,m1); + + if(PacketSize(0,m3.rows()-PacketSize); + Index j = internal::random(0,m3.cols()-PacketSize); + m3.template block(i,j).transposeInPlace(); + VERIFY_IS_APPROX( (m3.template block(i,j)), (m1.template block(i,j).transpose()) ); + m3.template block(i,j).transposeInPlace(); + VERIFY_IS_APPROX(m3,m1); + } // check inplace adjoint m3 = m1; @@ -129,9 +141,19 @@ void test_adjoint() CALL_SUBTEST_1( adjoint(Matrix()) ); CALL_SUBTEST_2( adjoint(Matrix3d()) ); CALL_SUBTEST_3( adjoint(Matrix4f()) ); + CALL_SUBTEST_4( adjoint(MatrixXcf(internal::random(1,EIGEN_TEST_MAX_SIZE/2), internal::random(1,EIGEN_TEST_MAX_SIZE/2))) ); CALL_SUBTEST_5( adjoint(MatrixXi(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); CALL_SUBTEST_6( adjoint(MatrixXf(internal::random(1,EIGEN_TEST_MAX_SIZE), internal::random(1,EIGEN_TEST_MAX_SIZE))) ); + + // Complement for 128 bits vectorization: + CALL_SUBTEST_8( adjoint(Matrix2d()) ); + CALL_SUBTEST_9( adjoint(Matrix()) ); + + // 256 bits vectorization: + CALL_SUBTEST_10( adjoint(Matrix()) ); + CALL_SUBTEST_11( adjoint(Matrix()) ); + CALL_SUBTEST_12( adjoint(Matrix()) ); } // test a large static matrix only once CALL_SUBTEST_7( adjoint(Matrix()) ); From 46fc881e4ae23ef577ee20dcd61a5a74cba8b874 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Mon, 26 Jan 2015 17:46:40 -0800 Subject: [PATCH 1038/1103] Added a few benchmarks for the tensor code --- bench/tensors/tensor_benchmarks.h | 305 +++++++++++++++++++++++++ bench/tensors/tensor_benchmarks_cpu.cc | 156 +++++++++++++ bench/tensors/tensor_benchmarks_gpu.cc | 75 ++++++ 3 files changed, 536 insertions(+) create mode 100644 bench/tensors/tensor_benchmarks.h create mode 100644 bench/tensors/tensor_benchmarks_cpu.cc create mode 100644 bench/tensors/tensor_benchmarks_gpu.cc diff --git a/bench/tensors/tensor_benchmarks.h b/bench/tensors/tensor_benchmarks.h new file mode 100644 index 000000000..525b9acda --- /dev/null +++ b/bench/tensors/tensor_benchmarks.h @@ -0,0 +1,305 @@ +#ifndef THIRD_PARTY_EIGEN3_TENSOR_BENCHMARKS_H_ +#define THIRD_PARTY_EIGEN3_TENSOR_BENCHMARKS_H_ + +typedef int TensorIndex; +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int + +#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" +#include "testing/base/public/benchmark.h" + +using Eigen::Tensor; +using Eigen::TensorMap; + + +// TODO(bsteiner): also templatize on the input type since we have users +// for int8 as well as floats. +template class BenchmarkSuite { + public: + BenchmarkSuite(const Device& device, size_t m, size_t k, size_t n) + : m_(m), k_(k), n_(n), device_(device) { + initialize(); + } + + BenchmarkSuite(const Device& device, size_t m) + : m_(m), k_(m), n_(m), device_(device) { + initialize(); + } + + ~BenchmarkSuite() { + device_.deallocate(a_); + device_.deallocate(b_); + device_.deallocate(c_); + } + + void memcpy(int num_iters) { + eigen_assert(m_ == k_ && k_ == n_); + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + device_.memcpy(c_, a_, m_ * m_ * sizeof(float)); + } + // Record the number of values copied per second + finalizeBenchmark(m_ * m_ * num_iters); + } + + void random(int num_iters) { + eigen_assert(m_ == k_ && k_ == n_); + const Eigen::array sizes(m_, m_); + TensorMap, Eigen::Aligned> C(c_, sizes); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = C.random(); + } + // Record the number of random numbers generated per second + finalizeBenchmark(m_ * m_ * num_iters); + } + + void slicing(int num_iters) { + eigen_assert(m_ == k_ && k_ == n_); + const Eigen::array sizes(m_, m_); + const TensorMap, Eigen::Aligned> A(a_, sizes); + const TensorMap, Eigen::Aligned> B(b_, sizes); + TensorMap, Eigen::Aligned> C(c_, sizes); + + const Eigen::DSizes quarter_sizes(Eigen::array(m_/2, m_/2)); + const Eigen::DSizes first_quadrant(Eigen::array(0, 0)); + const Eigen::DSizes second_quadrant(Eigen::array(0, m_/2)); + const Eigen::DSizes third_quadrant(Eigen::array(m_/2, 0)); + const Eigen::DSizes fourth_quadrant(Eigen::array(m_/2, m_/2)); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.slice(first_quadrant, quarter_sizes).device(device_) = + A.slice(first_quadrant, quarter_sizes); + C.slice(second_quadrant, quarter_sizes).device(device_) = + B.slice(second_quadrant, quarter_sizes); + C.slice(third_quadrant, quarter_sizes).device(device_) = + A.slice(third_quadrant, quarter_sizes); + C.slice(fourth_quadrant, quarter_sizes).device(device_) = + B.slice(fourth_quadrant, quarter_sizes); + } + // Record the number of values copied from the rhs slice to the lhs slice + // each second + finalizeBenchmark(m_ * m_ * num_iters); + } + + void shuffling(int num_iters) { + eigen_assert(m_ == n_); + const Eigen::array size_a(m_, k_); + const TensorMap, Eigen::Aligned> A(a_, size_a); + const Eigen::array size_b(k_, m_); + TensorMap, Eigen::Aligned> B(b_, size_b); + + const Eigen::array shuffle(1, 0); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + B.device(device_) = A.shuffle(shuffle); + } + // Record the number of values shuffled from A and copied to B each second + finalizeBenchmark(m_ * k_ * num_iters); + } + + void padding(int num_iters) { + eigen_assert(m_ == k_); + const Eigen::array size_a(m_, k_-3); + const TensorMap, Eigen::Aligned> A(a_, size_a); + const Eigen::array size_b(k_, m_); + TensorMap, Eigen::Aligned> B(b_, size_b); + + Eigen::array, 2> paddings; + paddings[0] = Eigen::IndexPair(0, 0); + paddings[1] = Eigen::IndexPair(2, 1); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + B.device(device_) = A.pad(paddings); + } + // Record the number of values copied from the padded tensor A each second + finalizeBenchmark(m_ * k_ * num_iters); + } + + void striding(int num_iters) { + eigen_assert(m_ == k_); + const Eigen::array size_a(m_, k_); + const TensorMap, Eigen::Aligned> A(a_, size_a); + const Eigen::array size_b(m_, k_ / 2); + TensorMap, Eigen::Aligned> B(b_, size_b); + + const Eigen::array strides(1, 2); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + B.device(device_) = A.stride(strides); + } + // Record the number of values copied from the padded tensor A each second + finalizeBenchmark(m_ * k_ * num_iters); + } + + void broadcasting(int num_iters) { + const Eigen::array size_a(m_, 1); + const TensorMap, Eigen::Aligned> A(a_, size_a); + const Eigen::array size_c(m_, n_); + TensorMap, Eigen::Aligned> C(c_, size_c); + +#if defined(__CUDACC__) + // nvcc doesn't support cxx11 + const Eigen::array broadcast(1, n_); +#else + // Take advantage of cxx11 to give the compiler information it can use to + // optimize the code. + Eigen::IndexList, int> broadcast; + broadcast.set(1, n_); +#endif + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = A.broadcast(broadcast); + } + // Record the number of values broadcasted from A and copied to C each second + finalizeBenchmark(m_ * n_ * num_iters); + } + + void coeffWiseOp(int num_iters) { + eigen_assert(m_ == k_ && k_ == n_); + const Eigen::array sizes(m_, m_); + const TensorMap, Eigen::Aligned> A(a_, sizes); + const TensorMap, Eigen::Aligned> B(b_, sizes); + TensorMap, Eigen::Aligned> C(c_, sizes); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = A * A.constant(3.14) + B * B.constant(2.7); + } + // Record the number of FLOP executed per second (2 multiplications and + // 1 addition per value) + finalizeBenchmark(3 * m_ * m_ * num_iters); + } + + void algebraicFunc(int num_iters) { + eigen_assert(m_ == k_ && k_ == n_); + const Eigen::array sizes(m_, m_); + const TensorMap, Eigen::Aligned> A(a_, sizes); + const TensorMap, Eigen::Aligned> B(b_, sizes); + TensorMap, Eigen::Aligned> C(c_, sizes); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = A.rsqrt() + B.sqrt() * B.square(); + } + // Record the number of FLOP executed per second (assuming one operation + // per value) + finalizeBenchmark(m_ * m_ * num_iters); + } + + void transcendentalFunc(int num_iters) { + eigen_assert(m_ == k_ && k_ == n_); + const Eigen::array sizes(m_, m_); + const TensorMap, Eigen::Aligned> A(a_, sizes); + const TensorMap, Eigen::Aligned> B(b_, sizes); + TensorMap, Eigen::Aligned> C(c_, sizes); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = A.exp() + B.log(); + } + // Record the number of FLOP executed per second (assuming one operation + // per value) + finalizeBenchmark(m_ * m_ * num_iters); + } + + // Simple reduction + void reduction(int num_iters) { + const Eigen::array input_size(k_, n_); + const TensorMap, Eigen::Aligned> B(b_, input_size); + const Eigen::array output_size(n_); + TensorMap, Eigen::Aligned> C(c_, output_size); + + const Eigen::array sum_along_dim(0); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = B.sum(sum_along_dim); + } + // Record the number of FLOP executed per second (assuming one operation + // per value) + finalizeBenchmark(m_ * m_ * num_iters); + } + + // do a contraction which is equivalent to a matrix multiplication + void contraction(int num_iters) { + const Eigen::array sizeA(m_, k_); + const Eigen::array sizeB(k_, n_); + const Eigen::array sizeC(m_, n_); + + const TensorMap, Eigen::Aligned> A(a_, sizeA); + const TensorMap, Eigen::Aligned> B(b_, sizeB); + TensorMap, Eigen::Aligned> C(c_, sizeC); + + typedef typename Tensor::DimensionPair DimPair; + const Eigen::array dims(DimPair(1, 0)); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = A.contract(B, dims); + } + // Record the number of FLOP executed per second (size_ multiplications and + // additions for each value in the resulting tensor) + finalizeBenchmark(static_cast(2) * m_ * n_ * k_ * num_iters); + } + + void convolution(int num_iters, int kernel_x, int kernel_y) { + const Eigen::array input_sizes(m_, n_); + TensorMap, Eigen::Aligned> A(a_, input_sizes); + const Eigen::array kernel_sizes(kernel_x, kernel_y); + TensorMap, Eigen::Aligned> B(b_, kernel_sizes); + const Eigen::array result_sizes( + m_ - kernel_x + 1, n_ - kernel_y + 1); + TensorMap, Eigen::Aligned> C(c_, result_sizes); + Eigen::array::Index, 2> dims(0, 1); + + StartBenchmarkTiming(); + for (int iter = 0; iter < num_iters; ++iter) { + C.device(device_) = A.convolve(B, dims); + } + // Record the number of FLOP executed per second (kernel_size + // multiplications and additions for each value in the resulting tensor) + finalizeBenchmark( + (m_ - kernel_x + 1) * (n_ - kernel_y + 1) * kernel_x * kernel_y * 2 * num_iters); + } + + private: + void initialize() { + a_ = (float *) device_.allocate(m_ * k_ * sizeof(float)); + b_ = (float *) device_.allocate(k_ * n_ * sizeof(float)); + c_ = (float *) device_.allocate(m_ * n_ * sizeof(float)); + + // Initialize the content of the memory pools to prevent asan from + // complaining. + device_.memset(a_, 12, m_ * k_ * sizeof(float)); + device_.memset(b_, 23, k_ * n_ * sizeof(float)); + device_.memset(c_, 31, m_ * n_ * sizeof(float)); + + BenchmarkUseRealTime(); + } + + inline void finalizeBenchmark(int64 num_items) { +#if defined(EIGEN_USE_GPU) && defined(__CUDACC__) + if (Eigen::internal::is_same::value) { + device_.synchronize(); + } +#endif + StopBenchmarkTiming(); + SetBenchmarkItemsProcessed(num_items); + } + + + size_t m_; + size_t k_; + size_t n_; + float* a_; + float* b_; + float* c_; + Device device_; +}; +#endif // THIRD_PARTY_EIGEN3_TENSOR_BENCHMARKS_H_ diff --git a/bench/tensors/tensor_benchmarks_cpu.cc b/bench/tensors/tensor_benchmarks_cpu.cc new file mode 100644 index 000000000..68653ba15 --- /dev/null +++ b/bench/tensors/tensor_benchmarks_cpu.cc @@ -0,0 +1,156 @@ +#define EIGEN_USE_THREADS + +#include "base/sysinfo.h" +#include "strings/strcat.h" +#include "third_party/eigen3/tensor_benchmarks.h" +#include "thread/threadpool.h" + +#ifdef __ANDROID__ +#define CREATE_THREAD_POOL(threads) \ +Eigen::ThreadPoolDevice device(threads); +#else +#define CREATE_THREAD_POOL(threads) \ +ThreadPool tp(threads); \ +tp.StartWorkers(); \ +Eigen::ThreadPoolDevice device(&tp, threads); +#endif + +// Simple functions +#define BM_FuncCPU(FUNC, THREADS) \ + static void BM_##FUNC##_##THREADS##T(int iters, int N) { \ + StopBenchmarkTiming(); \ + CREATE_THREAD_POOL(THREADS); \ + BenchmarkSuite suite(device, N); \ + suite.FUNC(iters); \ + SetBenchmarkLabel(StrCat("using ", THREADS, " threads")); \ + } \ + BENCHMARK_RANGE(BM_##FUNC##_##THREADS##T, 10, 5000); + +BM_FuncCPU(memcpy, 4); +BM_FuncCPU(memcpy, 8); +BM_FuncCPU(memcpy, 12); + +BM_FuncCPU(random, 4); +BM_FuncCPU(random, 8); +BM_FuncCPU(random, 12); + +BM_FuncCPU(slicing, 4); +BM_FuncCPU(slicing, 8); +BM_FuncCPU(slicing, 12); + +BM_FuncCPU(shuffling, 4); +BM_FuncCPU(shuffling, 8); +BM_FuncCPU(shuffling, 12); + +BM_FuncCPU(padding, 4); +BM_FuncCPU(padding, 8); +BM_FuncCPU(padding, 12); + +BM_FuncCPU(striding, 4); +BM_FuncCPU(striding, 8); +BM_FuncCPU(striding, 12); + +BM_FuncCPU(broadcasting, 4); +BM_FuncCPU(broadcasting, 8); +BM_FuncCPU(broadcasting, 12); + +BM_FuncCPU(coeffWiseOp, 4); +BM_FuncCPU(coeffWiseOp, 8); +BM_FuncCPU(coeffWiseOp, 12); + +BM_FuncCPU(algebraicFunc, 4); +BM_FuncCPU(algebraicFunc, 8); +BM_FuncCPU(algebraicFunc, 12); + +BM_FuncCPU(transcendentalFunc, 4); +BM_FuncCPU(transcendentalFunc, 8); +BM_FuncCPU(transcendentalFunc, 12); + +BM_FuncCPU(reduction, 4); +BM_FuncCPU(reduction, 8); +BM_FuncCPU(reduction, 12); + + +// Contractions +#define BM_FuncWithInputDimsCPU(FUNC, D1, D2, D3, THREADS) \ + static void BM_##FUNC##_##D1##x##D2##x##D3##_##THREADS##T(int iters, int N) {\ + StopBenchmarkTiming(); \ + if (THREADS == 1) { \ + Eigen::DefaultDevice device; \ + BenchmarkSuite suite(device, D1, D2, D3); \ + suite.FUNC(iters); \ + } else { \ + CREATE_THREAD_POOL(THREADS); \ + BenchmarkSuite suite(device, D1, D2, D3); \ + suite.FUNC(iters); \ + } \ + SetBenchmarkLabel(StrCat("using ", THREADS, " threads")); \ + } \ + BENCHMARK_RANGE(BM_##FUNC##_##D1##x##D2##x##D3##_##THREADS##T, 10, 5000); + + +BM_FuncWithInputDimsCPU(contraction, N, N, N, 1); +BM_FuncWithInputDimsCPU(contraction, N, N, N, 4); +BM_FuncWithInputDimsCPU(contraction, N, N, N, 8); +BM_FuncWithInputDimsCPU(contraction, N, N, N, 12); +BM_FuncWithInputDimsCPU(contraction, N, N, N, 16); + +BM_FuncWithInputDimsCPU(contraction, 64, N, N, 1); +BM_FuncWithInputDimsCPU(contraction, 64, N, N, 4); +BM_FuncWithInputDimsCPU(contraction, 64, N, N, 8); +BM_FuncWithInputDimsCPU(contraction, 64, N, N, 12); +BM_FuncWithInputDimsCPU(contraction, 64, N, N, 16); + +BM_FuncWithInputDimsCPU(contraction, N, 64, N, 1); +BM_FuncWithInputDimsCPU(contraction, N, 64, N, 4); +BM_FuncWithInputDimsCPU(contraction, N, 64, N, 8); +BM_FuncWithInputDimsCPU(contraction, N, 64, N, 12); +BM_FuncWithInputDimsCPU(contraction, N, 64, N, 16); + +BM_FuncWithInputDimsCPU(contraction, 1, N, N, 1); +BM_FuncWithInputDimsCPU(contraction, 1, N, N, 4); +BM_FuncWithInputDimsCPU(contraction, 1, N, N, 8); +BM_FuncWithInputDimsCPU(contraction, 1, N, N, 12); +BM_FuncWithInputDimsCPU(contraction, 1, N, N, 16); + +BM_FuncWithInputDimsCPU(contraction, N, N, 1, 1); +BM_FuncWithInputDimsCPU(contraction, N, N, 1, 4); +BM_FuncWithInputDimsCPU(contraction, N, N, 1, 8); +BM_FuncWithInputDimsCPU(contraction, N, N, 1, 12); +BM_FuncWithInputDimsCPU(contraction, N, N, 1, 16); + + +// Convolutions +#define BM_FuncWithKernelDimsCPU(FUNC, DIM1, DIM2, THREADS) \ + static void BM_##FUNC##_##DIM1##x##DIM2##_##THREADS##T(int iters, int N) { \ + StopBenchmarkTiming(); \ + CREATE_THREAD_POOL(THREADS); \ + BenchmarkSuite suite(device, N); \ + suite.FUNC(iters, DIM1, DIM2); \ + SetBenchmarkLabel(StrCat("using ", THREADS, " threads")); \ + } \ + BENCHMARK_RANGE(BM_##FUNC##_##DIM1##x##DIM2##_##THREADS##T, 128, 5000); + +BM_FuncWithKernelDimsCPU(convolution, 7, 1, 4); +BM_FuncWithKernelDimsCPU(convolution, 7, 1, 8); +BM_FuncWithKernelDimsCPU(convolution, 7, 1, 12); + +BM_FuncWithKernelDimsCPU(convolution, 1, 7, 4); +BM_FuncWithKernelDimsCPU(convolution, 1, 7, 8); +BM_FuncWithKernelDimsCPU(convolution, 1, 7, 12); + +BM_FuncWithKernelDimsCPU(convolution, 7, 4, 4); +BM_FuncWithKernelDimsCPU(convolution, 7, 4, 8); +BM_FuncWithKernelDimsCPU(convolution, 7, 4, 12); + +BM_FuncWithKernelDimsCPU(convolution, 4, 7, 4); +BM_FuncWithKernelDimsCPU(convolution, 4, 7, 8); +BM_FuncWithKernelDimsCPU(convolution, 4, 7, 12); + +BM_FuncWithKernelDimsCPU(convolution, 7, 64, 4); +BM_FuncWithKernelDimsCPU(convolution, 7, 64, 8); +BM_FuncWithKernelDimsCPU(convolution, 7, 64, 12); + +BM_FuncWithKernelDimsCPU(convolution, 64, 7, 4); +BM_FuncWithKernelDimsCPU(convolution, 64, 7, 8); +BM_FuncWithKernelDimsCPU(convolution, 64, 7, 12); diff --git a/bench/tensors/tensor_benchmarks_gpu.cc b/bench/tensors/tensor_benchmarks_gpu.cc new file mode 100644 index 000000000..adea754ad --- /dev/null +++ b/bench/tensors/tensor_benchmarks_gpu.cc @@ -0,0 +1,75 @@ +#define EIGEN_USE_GPU + +#include +#include +#include +#include "strings/strcat.h" +#include "third_party/eigen3/tensor_benchmarks.h" + + + +// Simple functions +#define BM_FuncGPU(FUNC) \ + static void BM_##FUNC(int iters, int N) { \ + StopBenchmarkTiming(); \ + cudaStream_t stream; \ + cudaStreamCreate(&stream); \ + Eigen::GpuDevice device(&stream); \ + BenchmarkSuite suite(device, N); \ + cudaDeviceSynchronize(); \ + suite.FUNC(iters); \ + cudaStreamDestroy(stream); \ + } \ + BENCHMARK_RANGE(BM_##FUNC, 10, 5000); + +BM_FuncGPU(memcpy); +BM_FuncGPU(random); +BM_FuncGPU(slicing); +BM_FuncGPU(shuffling); +BM_FuncGPU(padding); +BM_FuncGPU(striding); +BM_FuncGPU(broadcasting); +BM_FuncGPU(coeffWiseOp); +BM_FuncGPU(reduction); + + +// Contractions +#define BM_FuncWithInputDimsGPU(FUNC, D1, D2, D3) \ + static void BM_##FUNC##_##D1##x##D2##x##D3(int iters, int N) { \ + StopBenchmarkTiming(); \ + cudaStream_t stream; \ + cudaStreamCreate(&stream); \ + Eigen::GpuDevice device(&stream); \ + BenchmarkSuite suite(device, D1, D2, D3); \ + cudaDeviceSynchronize(); \ + suite.FUNC(iters); \ + cudaStreamDestroy(stream); \ + } \ + BENCHMARK_RANGE(BM_##FUNC##_##D1##x##D2##x##D3, 10, 5000); + + +BM_FuncWithInputDimsGPU(contraction, N, N, N); +BM_FuncWithInputDimsGPU(contraction, 64, N, N); +BM_FuncWithInputDimsGPU(contraction, N, 64, N); + + +// Convolutions +#define BM_FuncWithKernelDimsGPU(FUNC, DIM1, DIM2) \ + static void BM_##FUNC##_##DIM1##x##DIM2(int iters, int N) { \ + StopBenchmarkTiming(); \ + cudaStream_t stream; \ + cudaStreamCreate(&stream); \ + Eigen::GpuDevice device(&stream); \ + BenchmarkSuite suite(device, N); \ + cudaDeviceSynchronize(); \ + suite.FUNC(iters, DIM1, DIM2); \ + cudaStreamDestroy(stream); \ + } \ + BENCHMARK_RANGE(BM_##FUNC##_##DIM1##x##DIM2, 128, 5000); + +BM_FuncWithKernelDimsGPU(convolution, 7, 1); +BM_FuncWithKernelDimsGPU(convolution, 1, 7); +BM_FuncWithKernelDimsGPU(convolution, 7, 4); +BM_FuncWithKernelDimsGPU(convolution, 4, 7); +BM_FuncWithKernelDimsGPU(convolution, 7, 64); +BM_FuncWithKernelDimsGPU(convolution, 64, 7); From 9dfdbd7e568bd3aa9a4610986dcfc679b9ea425d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 27 Jan 2015 14:15:31 -0800 Subject: [PATCH 1039/1103] mproved the performance of tensor reductions that preserve the inner most dimension(s). --- .../Eigen/CXX11/src/Tensor/TensorReduction.h | 64 ++++++++++++++++++- 1 file changed, 63 insertions(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index eebcc4850..c6a8ecb5d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -48,6 +48,11 @@ template struct are_inner_most_dims { static const bool value = false; }; +template +struct preserve_inner_most_dims { + static const bool value = false; +}; + #if __cplusplus > 199711L template struct are_inner_most_dims{ @@ -61,6 +66,16 @@ struct are_inner_most_dims{ index_statically_eq()(0, NumTensorDims - array_size::value) && index_statically_eq()(array_size::value - 1, NumTensorDims - 1); }; +template +struct preserve_inner_most_dims{ + static const bool value = indices_statically_known_to_increase()() && + index_statically_gt()(0, 0); +}; +template +struct preserve_inner_most_dims{ + static const bool value = indices_statically_known_to_increase()() && + index_statically_lt()(array_size::value - 1, NumTensorDims - 1); +}; #endif @@ -108,7 +123,35 @@ struct InnerMostDimReducer { for (typename Self::Index j = VectorizedSize; j < numValuesToReduce; ++j) { reducer.reduce(self.m_impl.coeff(firstIndex + j), &accum); } - return reducer.finalizePacket(accum, p); + return reducer.finalizeBoth(accum, p); + } +}; + +template +struct InnerMostDimPreserver { + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::PacketReturnType* accum) { + eigen_assert(false && "should never be called"); + } +}; + +template +struct InnerMostDimPreserver { + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::PacketReturnType* accum) { + EIGEN_STATIC_ASSERT(DimIndex > 0, YOU_MADE_A_PROGRAMMING_MISTAKE); + for (int j = 0; j < self.m_reducedDims[DimIndex]; ++j) { + const typename Self::Index input = firstIndex + j * self.m_reducedStrides[DimIndex]; + InnerMostDimPreserver::reduce(self, input, reducer, accum); + } + } +}; + +template +struct InnerMostDimPreserver<0, Self, Op, true> { + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::PacketReturnType* accum) { + for (int j = 0; j < self.m_reducedDims[0]; ++j) { + const typename Self::Index input = firstIndex + j * self.m_reducedStrides[0]; + reducer.reducePacket(self.m_impl.template packet(input), accum); + } } }; @@ -168,11 +211,14 @@ struct TensorEvaluator, Device> }; static const bool ReducingInnerMostDims = internal::are_inner_most_dims::value; + static const bool PreservingInnerMostDims = internal::preserve_inner_most_dims::value; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : m_impl(op.expression(), device), m_reducer(op.reducer()) { EIGEN_STATIC_ASSERT(NumInputDims >= NumReducedDims, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((!ReducingInnerMostDims | !PreservingInnerMostDims | (NumReducedDims == NumInputDims)), + YOU_MADE_A_PROGRAMMING_MISTAKE); // Bitmap indicating if an input dimension is reduced or not. array reduced; @@ -291,6 +337,20 @@ struct TensorEvaluator, Device> values[i] = internal::InnerMostDimReducer::reduce(*this, firstIndex + i * num_values_to_reduce, num_values_to_reduce, reducer); } + } else if (PreservingInnerMostDims) { + const Index firstIndex = firstInput(index); + const int innermost_dim = (Layout == ColMajor) ? 0 : NumOutputDims - 1; + // TBD: extend this the the n innermost dimensions that we preserve. + if (((firstIndex % m_dimensions[innermost_dim]) + packetSize - 1) < m_dimensions[innermost_dim]) { + Op reducer(m_reducer); + typename Self::PacketReturnType accum = reducer.template initializePacket(); + internal::InnerMostDimPreserver::reduce(*this, firstIndex, reducer, &accum); + return reducer.finalizePacket(accum); + } else { + for (int i = 0; i < packetSize; ++i) { + values[i] = coeff(index + i); + } + } } else { for (int i = 0; i < packetSize; ++i) { values[i] = coeff(index + i); @@ -305,6 +365,7 @@ struct TensorEvaluator, Device> private: template friend struct internal::GenericDimReducer; template friend struct internal::InnerMostDimReducer; + template friend struct internal::InnerMostDimPreserver; // Returns the Index in the input tensor of the first value that needs to be // used to compute the reduction at output index "index". @@ -316,6 +377,7 @@ struct TensorEvaluator, Device> return index * m_preservedStrides[NumOutputDims - 1]; } } + // TBD: optimize the case where we preserve the innermost dimensions. Index startInput = 0; if (Layout == ColMajor) { for (int i = NumOutputDims - 1; i > 0; --i) { From a727a2c4edfa85303ad1ad406e65415187fbb770 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 28 Jan 2015 16:07:51 +0100 Subject: [PATCH 1040/1103] bug #933: RealSchur, do not consider the input matrix norm to check negligible sub-diag entries. This also makes this test consistent with the complex and self-adjoint cases. --- Eigen/src/Eigenvalues/RealSchur.h | 12 ++++-------- 1 file changed, 4 insertions(+), 8 deletions(-) diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h index 10f5fb174..51e61ba38 100644 --- a/Eigen/src/Eigenvalues/RealSchur.h +++ b/Eigen/src/Eigenvalues/RealSchur.h @@ -234,7 +234,7 @@ template class RealSchur typedef Matrix Vector3s; Scalar computeNormOfT(); - Index findSmallSubdiagEntry(Index iu, const Scalar& norm); + Index findSmallSubdiagEntry(Index iu); void splitOffTwoRows(Index iu, bool computeU, const Scalar& exshift); void computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo); void initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector); @@ -286,7 +286,7 @@ RealSchur& RealSchur::computeFromHessenberg(const HessMa { while (iu >= 0) { - Index il = findSmallSubdiagEntry(iu, norm); + Index il = findSmallSubdiagEntry(iu); // Check for convergence if (il == iu) // One root found @@ -343,16 +343,14 @@ inline typename MatrixType::Scalar RealSchur::computeNormOfT() /** \internal Look for single small sub-diagonal element and returns its index */ template -inline typename MatrixType::Index RealSchur::findSmallSubdiagEntry(Index iu, const Scalar& norm) +inline typename MatrixType::Index RealSchur::findSmallSubdiagEntry(Index iu) { using std::abs; Index res = iu; while (res > 0) { Scalar s = abs(m_matT.coeff(res-1,res-1)) + abs(m_matT.coeff(res,res)); - if (s == 0.0) - s = norm; - if (abs(m_matT.coeff(res,res-1)) < NumTraits::epsilon() * s) + if (abs(m_matT.coeff(res,res-1)) <= NumTraits::epsilon() * s) break; res--; } @@ -457,9 +455,7 @@ inline void RealSchur::initFrancisQRStep(Index il, Index iu, const V const Scalar lhs = m_matT.coeff(im,im-1) * (abs(v.coeff(1)) + abs(v.coeff(2))); const Scalar rhs = v.coeff(0) * (abs(m_matT.coeff(im-1,im-1)) + abs(Tmm) + abs(m_matT.coeff(im+1,im+1))); if (abs(lhs) < NumTraits::epsilon() * rhs) - { break; - } } } From 5a6ea4edf61b5626a781070c6342fc16606b490a Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Wed, 28 Jan 2015 10:02:47 -0800 Subject: [PATCH 1041/1103] Added more tests to cover tensor reductions --- .../Eigen/CXX11/src/Tensor/TensorFunctors.h | 43 +++++++++---- .../Eigen/CXX11/src/Tensor/TensorIndexList.h | 62 ++++++++++++++++++- unsupported/test/cxx11_tensor_reduction.cpp | 37 ++++++++++- 3 files changed, 128 insertions(+), 14 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h index 7b8d34321..38586d067 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h @@ -37,7 +37,11 @@ template struct SumReducer return accum; } template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizePacket(const T saccum, const Packet& vaccum) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const { + return vaccum; + } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const { return saccum + predux(vaccum); } }; @@ -45,16 +49,16 @@ template struct SumReducer template struct MeanReducer { static const bool PacketAccess = true; - MeanReducer() : count_(0) { } + MeanReducer() : scalarCount_(0), packetCount_(0) { } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) { (*accum) += t; - count_++; + scalarCount_++; } template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) { (*accum) = padd(*accum, p); - count_ += packet_traits::size; + packetCount_++; } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T initialize() const { @@ -65,15 +69,20 @@ template struct MeanReducer return pset1(0); } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalize(const T accum) const { - return accum / count_; + return accum / scalarCount_; } template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizePacket(const T saccum, const Packet& vaccum) const { - return (saccum + predux(vaccum)) / count_; + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const { + return pdiv(vaccum, pset1(packetCount_)); + } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const { + return (saccum + predux(vaccum)) / (scalarCount_ + packetCount_ * packet_traits::size); } protected: - int count_; + int scalarCount_; + int packetCount_; }; template struct MaxReducer @@ -99,7 +108,11 @@ template struct MaxReducer return accum; } template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizePacket(const T saccum, const Packet& vaccum) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const { + return vaccum; + } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const { return (std::max)(saccum, predux_max(vaccum)); } }; @@ -127,7 +140,11 @@ template struct MinReducer return accum; } template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizePacket(const T saccum, const Packet& vaccum) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const { + return vaccum; + } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const { return (std::min)(saccum, predux_min(vaccum)); } }; @@ -156,7 +173,11 @@ template struct ProdReducer return accum; } template - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizePacket(const T saccum, const Packet& vaccum) const { + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet finalizePacket(const Packet& vaccum) const { + return vaccum; + } + template + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const { return saccum * predux_mul(vaccum); } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h index 209749042..7ff47673d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h @@ -181,7 +181,7 @@ template size_t array_prod(const Ind result *= sizes[i]; } return result; -} +}; template struct array_size > { static const size_t value = std::tuple_size >::value; @@ -307,6 +307,52 @@ struct index_statically_ne > { }; +template +struct index_statically_gt { + constexpr bool operator() (DenseIndex, DenseIndex) const { + return false; + } +}; + +template +struct index_statically_gt > { + constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { + return IndexList().value_known_statically(i) & + IndexList()[i] > value; + } +}; + +template +struct index_statically_gt > { + constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { + return IndexList().value_known_statically(i) & + IndexList()[i] > value; + } +}; + +template +struct index_statically_lt { + constexpr bool operator() (DenseIndex, DenseIndex) const { + return false; + } +}; + +template +struct index_statically_lt > { + constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { + return IndexList().value_known_statically(i) & + IndexList()[i] < value; + } +}; + +template +struct index_statically_lt > { + constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { + return IndexList().value_known_statically(i) & + IndexList()[i] < value; + } +}; + } // end namespace internal } // end namespace Eigen @@ -351,6 +397,20 @@ struct index_statically_ne { } }; +template +struct index_statically_gt { + EIGEN_ALWAYS_INLINE EIGEN_DEVICE_FUNC bool operator() (DenseIndex, DenseIndex) const{ + return false; + } +}; + +template +struct index_statically_lt { + EIGEN_ALWAYS_INLINE EIGEN_DEVICE_FUNC bool operator() (DenseIndex, DenseIndex) const{ + return false; + } +}; + } // end namespace internal } // end namespace Eigen diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp index 99e19eba4..5c3184833 100644 --- a/unsupported/test/cxx11_tensor_reduction.cpp +++ b/unsupported/test/cxx11_tensor_reduction.cpp @@ -369,6 +369,37 @@ static void test_innermost_first_dims() { } } +template +static void test_reduce_middle_dims() { + Tensor in(72, 53, 97, 113); + Tensor out(72, 53); + in.setRandom(); + +// Reduce on the innermost dimensions. +#if __cplusplus <= 199711L + array reduction_axis; + reduction_axis[0] = 1; + reduction_axis[1] = 2; +#else + // This triggers the use of packets for RowMajor. + Eigen::IndexList, Eigen::type2index<2>> reduction_axis; +#endif + + out = in.maximum(reduction_axis); + + for (int i = 0; i < 72; ++i) { + for (int j = 0; j < 113; ++j) { + float expected = -1e10f; + for (int k = 0; k < 53; ++k) { + for (int l = 0; l < 97; ++l) { + expected = (std::max)(expected, in(i, k, l, j)); + } + } + VERIFY_IS_APPROX(out(i, j), expected); + } + } +} + void test_cxx11_tensor_reduction() { CALL_SUBTEST(test_simple_reductions()); CALL_SUBTEST(test_simple_reductions()); @@ -380,8 +411,10 @@ void test_cxx11_tensor_reduction() { CALL_SUBTEST(test_tensor_maps()); CALL_SUBTEST(test_static_dims()); CALL_SUBTEST(test_static_dims()); - CALL_SUBTEST(test_innermost_last_dims()); CALL_SUBTEST(test_innermost_last_dims()); - CALL_SUBTEST(test_innermost_first_dims()); + CALL_SUBTEST(test_innermost_last_dims()); CALL_SUBTEST(test_innermost_first_dims()); + CALL_SUBTEST(test_innermost_first_dims()); + CALL_SUBTEST(test_reduce_middle_dims()); + CALL_SUBTEST(test_reduce_middle_dims()); } From e896c0ade7c77a18acb1b3ef01f22ef698c1a2a2 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Thu, 29 Jan 2015 10:29:47 -0800 Subject: [PATCH 1042/1103] Marked the contraction operation as read only, since its result can't be assigned. --- unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index a02a273e7..af843654c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -423,7 +423,7 @@ struct traits -class TensorContractionOp : public TensorBase > +class TensorContractionOp : public TensorBase, ReadOnlyAccessors> { public: typedef typename Eigen::internal::traits::Scalar Scalar; From 9d82f7e30d53086d090ae13d69dffef771eb6263 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 30 Jan 2015 17:24:40 +0100 Subject: [PATCH 1043/1103] Supernodes was disabled. --- Eigen/src/SparseLU/SparseLU.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index d72d7f150..79b78da99 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -309,7 +309,7 @@ class SparseLU : public SparseSolverBase >, // Functions void initperfvalues() { - m_perfv.panel_size = 1; + m_perfv.panel_size = 16; m_perfv.relax = 1; m_perfv.maxsuper = 128; m_perfv.rowblk = 16; From f1092d2f736bbe541b3d8b5cca893f668f9c6b5f Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 30 Jan 2015 19:04:04 +0100 Subject: [PATCH 1044/1103] bug #941: fix accuracy issue in ColPivHouseholderQR, do not stop decomposition on a small pivot --- Eigen/src/QR/ColPivHouseholderQR.h | 24 ++++++++---------------- 1 file changed, 8 insertions(+), 16 deletions(-) diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index de77e8411..370cb69e3 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -476,20 +476,10 @@ ColPivHouseholderQR& ColPivHouseholderQR::compute(const // we store that back into our table: it can't hurt to correct our table. m_colSqNorms.coeffRef(biggest_col_index) = biggest_col_sq_norm; - // if the current biggest column is smaller than epsilon times the initial biggest column, - // terminate to avoid generating nan/inf values. - // Note that here, if we test instead for "biggest == 0", we get a failure every 1000 (or so) - // repetitions of the unit test, with the result of solve() filled with large values of the order - // of 1/(size*epsilon). - if(biggest_col_sq_norm < threshold_helper * RealScalar(rows-k)) - { + // Track the number of meaningful pivots but do not stop the decomposition to make + // sure that the initial matrix is properly reproduced. See bug 941. + if(m_nonzero_pivots==size && biggest_col_sq_norm < threshold_helper * RealScalar(rows-k)) m_nonzero_pivots = k; - m_hCoeffs.tail(size-k).setZero(); - m_qr.bottomRightCorner(rows-k,cols-k) - .template triangularView() - .setZero(); - break; - } // apply the transposition to the columns m_colsTranspositions.coeffRef(k) = biggest_col_index; @@ -518,7 +508,7 @@ ColPivHouseholderQR& ColPivHouseholderQR::compute(const } m_colsPermutation.setIdentity(PermIndexType(cols)); - for(PermIndexType k = 0; k < m_nonzero_pivots; ++k) + for(PermIndexType k = 0; k < size/*m_nonzero_pivots*/; ++k) m_colsPermutation.applyTranspositionOnTheRight(k, PermIndexType(m_colsTranspositions.coeff(k))); m_det_pq = (number_of_transpositions%2) ? -1 : 1; @@ -574,13 +564,15 @@ struct Assignment >, interna } // end namespace internal -/** \returns the matrix Q as a sequence of householder transformations */ +/** \returns the matrix Q as a sequence of householder transformations. + * You can extract the meaningful part only by using: + * \code qr.householderQ().setLength(qr.nonzeroPivots()) */ template typename ColPivHouseholderQR::HouseholderSequenceType ColPivHouseholderQR ::householderQ() const { eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized."); - return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate()).setLength(m_nonzero_pivots); + return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate()); } #ifndef __CUDACC__ From 759bd92a85393617a56405ec0372e87416cfaebb Mon Sep 17 00:00:00 2001 From: Benoit Jacob Date: Fri, 30 Jan 2015 17:27:56 -0500 Subject: [PATCH 1045/1103] bug #935: Add asm comments in GEBP kernels to work around a bug in both GCC and Clang on ARM/NEON, whereby they spill registers, severely harming performance. The reason why the asm comments make a difference is that they prevent the compiler from reordering code across these boundaries, which has the effect of extending the lifetime of local variables and increasing register pressure on this register-tight code. --- .../Core/products/GeneralBlockPanelKernel.h | 160 +++++++++++------- Eigen/src/Core/util/Macros.h | 8 +- 2 files changed, 108 insertions(+), 60 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 7b2ed6728..1b39642fb 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -760,31 +760,36 @@ void gebp_kernel for(Index k=0; k blB += pk*4*RhsProgress; blA += pk*3*Traits::LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 3pX4"); } // process remaining peeled loop for(Index k=peeled_kc; k for(Index k=0; k blB += pk*RhsProgress; blA += pk*3*Traits::LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 3pX1"); } // process remaining peeled loop @@ -963,21 +977,26 @@ void gebp_kernel for(Index k=0; k blB += pk*4*RhsProgress; blA += pk*(2*Traits::LhsProgress); + + EIGEN_ASM_COMMENT("end gebp micro kernel 2pX4"); } // process remaining peeled loop for(Index k=peeled_kc; k for(Index k=0; k blB += pk*RhsProgress; blA += pk*2*Traits::LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 2pX1"); } // process remaining peeled loop @@ -1137,16 +1165,21 @@ void gebp_kernel for(Index k=0; k blB += pk*4*RhsProgress; blA += pk*1*LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 1pX4"); } // process remaining peeled loop for(Index k=peeled_kc; k for(Index k=0; k blB += pk*RhsProgress; blA += pk*1*Traits::LhsProgress; + + EIGEN_ASM_COMMENT("end gebp micro kernel 2pX1"); } // process remaining peeled loop diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 687ba41dd..13f8fdd4e 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -160,6 +160,12 @@ #define EIGEN_ARCH_ARM64 0 #endif +#if EIGEN_ARCH_ARM || EIGEN_ARCH_ARM64 + #define EIGEN_ARCH_ARM_OR_ARM64 1 +#else + #define EIGEN_ARCH_ARM_OR_ARM64 0 +#endif + /// \internal EIGEN_ARCH_MIPS set to 1 if the architecture is MIPS #if defined(__mips__) || defined(__mips) #define EIGEN_ARCH_MIPS 1 @@ -526,7 +532,7 @@ namespace Eigen { #define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var); #if !defined(EIGEN_ASM_COMMENT) - #if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64 + #if EIGEN_COMP_GNUC && (EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64) #define EIGEN_ASM_COMMENT(X) __asm__("#" X) #else #define EIGEN_ASM_COMMENT(X) From 9f99f61e69a70e0a209d5f93e78a9257685cd70d Mon Sep 17 00:00:00 2001 From: Benoit Jacob Date: Fri, 30 Jan 2015 17:43:56 -0500 Subject: [PATCH 1046/1103] bug #936, patch 1/3: some cleanup and renaming for consistency. --- Eigen/src/Core/arch/AltiVec/PacketMath.h | 4 ++-- Eigen/src/Core/arch/NEON/PacketMath.h | 4 ++-- .../Core/products/GeneralBlockPanelKernel.h | 18 +++++++++--------- 3 files changed, 13 insertions(+), 13 deletions(-) diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index fa02f57a1..27df5a025 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -22,8 +22,8 @@ namespace internal { #define EIGEN_HAS_FUSED_MADD 1 #endif -#ifndef EIGEN_HAS_FUSE_CJMADD -#define EIGEN_HAS_FUSE_CJMADD 1 +#ifndef EIGEN_HAS_FUSED_CJMADD +#define EIGEN_HAS_FUSED_CJMADD #endif // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16 diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index f83f8db0e..5a6eb8c1d 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -24,8 +24,8 @@ namespace internal { #define EIGEN_HAS_FUSED_MADD 1 #endif -#ifndef EIGEN_HAS_FUSE_CJMADD -#define EIGEN_HAS_FUSE_CJMADD 1 +#ifndef EIGEN_HAS_FUSED_CJMADD +#define EIGEN_HAS_FUSED_CJMADD #endif // FIXME NEON has 16 quad registers, but since the current register allocator diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 1b39642fb..ae2fd9006 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -120,8 +120,8 @@ inline void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) computeProductBlockingSizes(k, m, n); } -#ifdef EIGEN_HAS_FUSE_CJMADD - #define MADD(CJ,A,B,C,T) C = CJ.pmadd(A,B,C); +#ifdef EIGEN_HAS_FUSED_CJMADD + #define CJMADD(CJ,A,B,C,T) C = CJ.pmadd(A,B,C); #else // FIXME (a bit overkill maybe ?) @@ -146,8 +146,8 @@ inline void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) gebp_madd_selector::run(cj,a,b,c,t); } - #define MADD(CJ,A,B,C,T) gebp_madd(CJ,A,B,C,T); -// #define MADD(CJ,A,B,C,T) T = B; T = CJ.pmul(A,T); C = padd(C,T); + #define CJMADD(CJ,A,B,C,T) gebp_madd(CJ,A,B,C,T); +// #define CJMADD(CJ,A,B,C,T) T = B; T = CJ.pmul(A,T); C = padd(C,T); #endif /* Vectorization logic @@ -1402,13 +1402,13 @@ void gebp_kernel B_0 = blB[0]; B_1 = blB[1]; - MADD(cj,A0,B_0,C0, B_0); - MADD(cj,A0,B_1,C1, B_1); + CJMADD(cj,A0,B_0,C0, B_0); + CJMADD(cj,A0,B_1,C1, B_1); B_0 = blB[2]; B_1 = blB[3]; - MADD(cj,A0,B_0,C2, B_0); - MADD(cj,A0,B_1,C3, B_1); + CJMADD(cj,A0,B_0,C2, B_0); + CJMADD(cj,A0,B_1,C3, B_1); blB += 4; } @@ -1434,7 +1434,7 @@ void gebp_kernel { LhsScalar A0 = blA[k]; RhsScalar B_0 = blB[k]; - MADD(cj, A0, B_0, C0, B_0); + CJMADD(cj, A0, B_0, C0, B_0); } res[(j2+0)*resStride + i] += alpha*C0; } From 340b8afb14bb06788570ba22ba4ccba674402f09 Mon Sep 17 00:00:00 2001 From: Benoit Jacob Date: Sat, 31 Jan 2015 14:15:57 -0500 Subject: [PATCH 1047/1103] bug #936, patch 1.5/3: rename _FUSED_ macros to _SINGLE_INSTRUCTION_, because this is what they are about. "Fused" means "no intermediate rounding between the mul and the add, only one rounding at the end". Instead, what we are concerned about here is whether a temporary register is needed, i.e. whether the MUL and ADD are separate instructions. Concretely, on ARM NEON, a single-instruction mul-add is always available: VMLA. But a true fused mul-add is only available on VFPv4: VFMA. --- Eigen/src/Core/arch/AVX/PacketMath.h | 4 ++-- Eigen/src/Core/arch/AltiVec/PacketMath.h | 8 ++++---- Eigen/src/Core/arch/NEON/PacketMath.h | 8 ++++---- Eigen/src/Core/arch/SSE/PacketMath.h | 4 ++-- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 12 ++++++------ 5 files changed, 18 insertions(+), 18 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index e2376bd1f..1d8c674a6 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -23,8 +23,8 @@ namespace internal { #endif #ifdef EIGEN_VECTORIZE_FMA -#ifndef EIGEN_HAS_FUSED_MADD -#define EIGEN_HAS_FUSED_MADD 1 +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 #endif #endif diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 27df5a025..578b303a0 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -18,12 +18,12 @@ namespace internal { #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4 #endif -#ifndef EIGEN_HAS_FUSED_MADD -#define EIGEN_HAS_FUSED_MADD 1 +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 #endif -#ifndef EIGEN_HAS_FUSED_CJMADD -#define EIGEN_HAS_FUSED_CJMADD +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD #endif // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16 diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 5a6eb8c1d..9cfb9c358 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -20,12 +20,12 @@ namespace internal { #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 #endif -#ifndef EIGEN_HAS_FUSED_MADD -#define EIGEN_HAS_FUSED_MADD 1 +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 #endif -#ifndef EIGEN_HAS_FUSED_CJMADD -#define EIGEN_HAS_FUSED_CJMADD +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD #endif // FIXME NEON has 16 quad registers, but since the current register allocator diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 28427c308..202aaa72f 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -23,8 +23,8 @@ namespace internal { #endif #ifdef EIGEN_VECTORIZE_FMA -#ifndef EIGEN_HAS_FUSED_MADD -#define EIGEN_HAS_FUSED_MADD 1 +#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 #endif #endif diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index ae2fd9006..b5f06d831 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -120,7 +120,7 @@ inline void computeProductBlockingSizes(SizeType& k, SizeType& m, SizeType& n) computeProductBlockingSizes(k, m, n); } -#ifdef EIGEN_HAS_FUSED_CJMADD +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD #define CJMADD(CJ,A,B,C,T) C = CJ.pmadd(A,B,C); #else @@ -182,7 +182,7 @@ public: nr = 4, // register block size along the M direction (currently, this one cannot be modified) -#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) +#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -248,7 +248,7 @@ public: // let gcc allocate the register in which to store the result of the pmul // (in the case where there is no FMA) gcc fails to figure out how to avoid // spilling register. -#ifdef EIGEN_HAS_FUSED_MADD +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD EIGEN_UNUSED_VARIABLE(tmp); c = pmadd(a,b,c); #else @@ -290,7 +290,7 @@ public: NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, nr = 4, -#if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) +#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) // we assume 16 registers mr = 3*LhsPacketSize, #else @@ -353,7 +353,7 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { -#ifdef EIGEN_HAS_FUSED_MADD +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD EIGEN_UNUSED_VARIABLE(tmp); c.v = pmadd(a.v,b,c.v); #else @@ -637,7 +637,7 @@ public: EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const { -#ifdef EIGEN_HAS_FUSED_MADD +#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD EIGEN_UNUSED_VARIABLE(tmp); c.v = pmadd(a,b.v,c.v); #else From 0f216136980503c3792a90e382b4d6bbdbb870c0 Mon Sep 17 00:00:00 2001 From: Benoit Jacob Date: Fri, 30 Jan 2015 17:44:26 -0500 Subject: [PATCH 1048/1103] bug #936, patch 2/3: Remove EIGEN_VECTORIZE_FMA, was redundant with EIGEN_HAS_SINGLE_INSTRUCTION_MADD --- Eigen/Core | 4 +--- Eigen/src/Core/arch/AVX/PacketMath.h | 4 ++-- Eigen/src/Core/arch/AltiVec/PacketMath.h | 2 +- Eigen/src/Core/arch/NEON/PacketMath.h | 2 +- Eigen/src/Core/arch/SSE/PacketMath.h | 2 +- 5 files changed, 6 insertions(+), 8 deletions(-) diff --git a/Eigen/Core b/Eigen/Core index dcb20bfd0..b5af63623 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -125,9 +125,7 @@ #define EIGEN_VECTORIZE_SSE4_1 #define EIGEN_VECTORIZE_SSE4_2 #endif - #ifdef __FMA__ - #define EIGEN_VECTORIZE_FMA - #endif + // include files // This extern "C" works around a MINGW-w64 compilation issue diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 1d8c674a6..485bac10b 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -22,9 +22,9 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) #endif -#ifdef EIGEN_VECTORIZE_FMA +#ifdef __FMA__ #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD -#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD #endif #endif diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 578b303a0..6b68fc7a5 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -19,7 +19,7 @@ namespace internal { #endif #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD -#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD #endif #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 9cfb9c358..71255ac85 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -21,7 +21,7 @@ namespace internal { #endif #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD -#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 +#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD #endif #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 202aaa72f..3f6fb0254 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -22,7 +22,7 @@ namespace internal { #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) #endif -#ifdef EIGEN_VECTORIZE_FMA +#ifdef __FMA__ #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1 #endif From 5ef95fabee5e9a9357c082cd32ae3b4affb2eff6 Mon Sep 17 00:00:00 2001 From: Benoit Jacob Date: Fri, 30 Jan 2015 17:45:03 -0500 Subject: [PATCH 1049/1103] bug #936, patch 3/3: Properly detect FMA support on ARM (requires VFPv4) and use it instead of MLA when available, because it's both more accurate, and faster. --- Eigen/src/Core/arch/NEON/PacketMath.h | 19 +++++++++++++++++-- 1 file changed, 17 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 71255ac85..9afd86bec 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -177,8 +177,19 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv(const Packet4i& /*a*/, co return pset1(0); } -// for some weird raisons, it has to be overloaded for packet of integers +#ifdef __ARM_FEATURE_FMA +// See bug 936. +// FMA is available on VFPv4 i.e. when compiling with -mfpu=neon-vfpv4. +// FMA is a true fused multiply-add i.e. only 1 rounding at the end, no intermediate rounding. +// MLA is not fused i.e. does 2 roundings. +// In addition to giving better accuracy, FMA also gives better performance here on a Krait (Nexus 4): +// MLA: 10 GFlop/s ; FMA: 12 GFlops/s. +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vfmaq_f32(c,a,b); } +#else template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vmlaq_f32(c,a,b); } +#endif + +// No FMA instruction for int, so use MLA unconditionally. template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); } template<> EIGEN_STRONG_INLINE Packet4f pmin(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); } @@ -551,8 +562,12 @@ template<> EIGEN_STRONG_INLINE Packet2d pmul(const Packet2d& a, const template<> EIGEN_STRONG_INLINE Packet2d pdiv(const Packet2d& a, const Packet2d& b) { return vdivq_f64(a,b); } -// for some weird raisons, it has to be overloaded for packet of integers +#ifdef __ARM_FEATURE_FMA +// See bug 936. See above comment about FMA for float. +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vfmaq_f64(c,a,b); } +#else template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vmlaq_f64(c,a,b); } +#endif template<> EIGEN_STRONG_INLINE Packet2d pmin(const Packet2d& a, const Packet2d& b) { return vminq_f64(a,b); } From 590f4b0aa3583c98fe9a0682e26c24ebfaffeaa6 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 30 Jan 2015 19:46:30 -0800 Subject: [PATCH 1050/1103] Silenced some compilation warnings --- .../Eigen/CXX11/src/Tensor/TensorIndexList.h | 22 +++++++++---------- .../CXX11/src/Tensor/TensorInitializer.h | 12 ---------- .../Eigen/CXX11/src/Tensor/TensorReduction.h | 2 +- 3 files changed, 12 insertions(+), 24 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h index 7ff47673d..c94ed977e 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h @@ -124,18 +124,18 @@ struct tuple_coeff<0> { update_value(std::get<0>(t), value); } template - static constexpr bool value_known_statically(const DenseIndex i, const std::tuple& t) { + static constexpr bool value_known_statically(const DenseIndex i, const std::tuple&) { // eigen_assert (i == 0); // gcc fails to compile assertions in constexpr return is_compile_time_constant >::type>::value & (i == 0); } template - static constexpr bool values_up_to_known_statically(const std::tuple& t) { + static constexpr bool values_up_to_known_statically(const std::tuple&) { return is_compile_time_constant >::type>::value; } template - static constexpr bool values_up_to_statically_known_to_increase(const std::tuple& t) { + static constexpr bool values_up_to_statically_known_to_increase(const std::tuple&) { return true; } }; @@ -271,7 +271,7 @@ template struct index_statically_eq > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] == value; + (IndexList()[i] == value); } }; @@ -279,7 +279,7 @@ template struct index_statically_eq > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] == value; + (IndexList()[i] == value); } }; @@ -294,7 +294,7 @@ template struct index_statically_ne > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] != value; + (IndexList()[i] != value); } }; @@ -302,7 +302,7 @@ template struct index_statically_ne > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] != value; + (IndexList()[i] != value); } }; @@ -318,7 +318,7 @@ template struct index_statically_gt > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] > value; + (IndexList()[i] > value); } }; @@ -326,7 +326,7 @@ template struct index_statically_gt > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] > value; + (IndexList()[i] > value); } }; @@ -341,7 +341,7 @@ template struct index_statically_lt > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] < value; + (IndexList()[i] < value); } }; @@ -349,7 +349,7 @@ template struct index_statically_lt > { constexpr bool operator() (const DenseIndex i, const DenseIndex value) const { return IndexList().value_known_statically(i) & - IndexList()[i] < value; + (IndexList()[i] < value); } }; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorInitializer.h b/unsupported/Eigen/CXX11/src/Tensor/TensorInitializer.h index 6afef0fbb..4303e3536 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorInitializer.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorInitializer.h @@ -55,18 +55,6 @@ struct Initializer { } }; -template -struct Initializer { - typedef std::initializer_list::Scalar> InitList; - - static void run(TensorEvaluator& tensor, - Eigen::array::Index, traits::NumDimensions>* indices, - const InitList& vals) { - // Static initialization not implemented for VarDims tensors. - eigen_assert(false); - } -}; - template void initialize_tensor(TensorEvaluator& tensor, const typename Initializer::NumDimensions>::InitList& vals) { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index c6a8ecb5d..83ba1df71 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -129,7 +129,7 @@ struct InnerMostDimReducer { template struct InnerMostDimPreserver { - static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self& self, typename Self::Index firstIndex, Op& reducer, typename Self::PacketReturnType* accum) { + static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const Self&, typename Self::Index, Op&, typename Self::PacketReturnType*) { eigen_assert(false && "should never be called"); } }; From f64045a060ae22c6445b78ecea3783cef7c1ca3b Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 30 Jan 2015 19:52:01 -0800 Subject: [PATCH 1051/1103] Silenced a few more compilation warnings --- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index e125ca799..0e8a4b8d6 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -369,7 +369,7 @@ class Tensor : public TensorBase > void resize(const DSizes& dimensions) { array dims; - for (int i = 0; i < NumIndices; ++i) { + for (std::size_t i = 0; i < NumIndices; ++i) { dims[i] = dimensions[i]; } resize(dims); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index 6c9a67c58..d81197e6d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -93,7 +93,7 @@ struct Sizes : internal::numeric_list { // todo: add assertion } #ifdef EIGEN_HAS_VARIADIC_TEMPLATES - template Sizes(DenseIndex... indices) { } + template Sizes(DenseIndex...) { } explicit Sizes(std::initializer_list /*l*/) { // todo: add assertion } @@ -333,7 +333,7 @@ static const size_t value = Sizes::count; template struct array_size > { static const size_t value = Sizes::count; }; -template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_get(const Sizes& a) { +template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t array_get(const Sizes&) { return get >::value; } #else From ebdf6a2dbbc66d0f6fb045a3c5b0023bc2e89851 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 3 Feb 2015 22:32:34 +0100 Subject: [PATCH 1052/1103] SPQR: fix default threshold value --- Eigen/src/SPQRSupport/SuiteSparseQRSupport.h | 51 +++++++++++++++----- 1 file changed, 40 insertions(+), 11 deletions(-) diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h index 44f6a1acb..54a1b21b8 100644 --- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h @@ -68,13 +68,13 @@ class SPQR : public SparseSolverBase > typedef Map > PermutationType; public: SPQR() - : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) + : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()), m_useDefaultThreshold(true) { cholmod_l_start(&m_cc); } explicit SPQR(const _MatrixType& matrix) - : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()) + : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits::epsilon()), m_useDefaultThreshold(true) { cholmod_l_start(&m_cc); compute(matrix); @@ -99,10 +99,25 @@ class SPQR : public SparseSolverBase > if(m_isInitialized) SPQR_free(); MatrixType mat(matrix); + + /* Compute the default threshold as in MatLab, see: + * Tim Davis, "Algorithm 915, SuiteSparseQR: Multifrontal Multithreaded Rank-Revealing + * Sparse QR Factorization, ACM Trans. on Math. Soft. 38(1), 2011, Page 8:3 + */ + RealScalar pivotThreshold = m_tolerance; + if(m_useDefaultThreshold) + { + RealScalar max2Norm = 0.0; + for (int j = 0; j < mat.cols(); j++) max2Norm = numext::maxi(max2Norm, mat.col(j).norm()); + if(max2Norm==RealScalar(0)) + max2Norm = RealScalar(1); + pivotThreshold = 20 * (mat.rows() + mat.cols()) * max2Norm * NumTraits::epsilon(); + } + cholmod_sparse A; A = viewAsCholmod(mat); Index col = matrix.cols(); - m_rank = SuiteSparseQR(m_ordering, m_tolerance, col, &A, + m_rank = SuiteSparseQR(m_ordering, pivotThreshold, col, &A, &m_cR, &m_E, &m_H, &m_HPinv, &m_HTau, &m_cc); if (!m_cR) @@ -118,7 +133,7 @@ class SPQR : public SparseSolverBase > /** * Get the number of rows of the input matrix and the Q matrix */ - inline Index rows() const {return m_H->nrow; } + inline Index rows() const {return m_cR->nrow; } /** * Get the number of columns of the input matrix. @@ -130,16 +145,25 @@ class SPQR : public SparseSolverBase > { eigen_assert(m_isInitialized && " The QR factorization should be computed first, call compute()"); eigen_assert(b.cols()==1 && "This method is for vectors only"); - + //Compute Q^T * b - typename Dest::PlainObject y; + typename Dest::PlainObject y, y2; y = matrixQ().transpose() * b; - // Solves with the triangular matrix R + + // Solves with the triangular matrix R Index rk = this->rank(); - y.topRows(rk) = this->matrixR().topLeftCorner(rk, rk).template triangularView().solve(y.topRows(rk)); - y.bottomRows(cols()-rk).setZero(); + y2 = y; + y.resize((std::max)(cols(),Index(y.rows())),y.cols()); + y.topRows(rk) = this->matrixR().topLeftCorner(rk, rk).template triangularView().solve(y2.topRows(rk)); + // Apply the column permutation - dest.topRows(cols()) = colsPermutation() * y.topRows(cols()); + // colsPermutation() performs a copy of the permutation, + // so let's apply it manually: + for(Index i = 0; i < rk; ++i) dest.row(m_E[i]) = y.row(i); + for(Index i = rk; i < cols(); ++i) dest.row(m_E[i]).setZero(); + +// y.bottomRows(y.rows()-rk).setZero(); +// dest = colsPermutation() * y.topRows(cols()); m_info = Success; } @@ -178,7 +202,11 @@ class SPQR : public SparseSolverBase > /// Set the fill-reducing ordering method to be used void setSPQROrdering(int ord) { m_ordering = ord;} /// Set the tolerance tol to treat columns with 2-norm < =tol as zero - void setPivotThreshold(const RealScalar& tol) { m_tolerance = tol; } + void setPivotThreshold(const RealScalar& tol) + { + m_useDefaultThreshold = false; + m_tolerance = tol; + } /** \returns a pointer to the SPQR workspace */ cholmod_common *cholmodCommon() const { return &m_cc; } @@ -210,6 +238,7 @@ class SPQR : public SparseSolverBase > mutable cholmod_dense *m_HTau; // The Householder coefficients mutable Index m_rank; // The rank of the matrix mutable cholmod_common m_cc; // Workspace and parameters + bool m_useDefaultThreshold; // Use default threshold template friend struct SPQR_QProduct; }; From b1eca55328436f9778254c6bbc8852910a0d3d2a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 3 Feb 2015 23:46:05 +0100 Subject: [PATCH 1053/1103] Use Ref<> to ensure that both x and b in Ax=b are compatible with Umfpack/SuperLU expectations --- Eigen/src/SuperLUSupport/SuperLUSupport.h | 23 +++++++++++++++++++---- Eigen/src/UmfPackSupport/UmfPackSupport.h | 13 ++++++++++++- 2 files changed, 31 insertions(+), 5 deletions(-) diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h index ef73587a7..6de5b3dc5 100644 --- a/Eigen/src/SuperLUSupport/SuperLUSupport.h +++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h @@ -627,8 +627,12 @@ void SuperLU::_solve_impl(const MatrixBase &b, MatrixBase m_sluFerr.resize(rhsCols); m_sluBerr.resize(rhsCols); - m_sluB = SluMatrix::Map(b.const_cast_derived()); - m_sluX = SluMatrix::Map(x.derived()); + + Ref > b_ref(b); + Ref > x_ref(x); + + m_sluB = SluMatrix::Map(b_ref.const_cast_derived()); + m_sluX = SluMatrix::Map(x_ref.const_cast_derived()); typename Rhs::PlainObject b_cpy; if(m_sluEqued!='N') @@ -651,6 +655,10 @@ void SuperLU::_solve_impl(const MatrixBase &b, MatrixBase &m_sluFerr[0], &m_sluBerr[0], &m_sluStat, &info, Scalar()); StatFree(&m_sluStat); + + if(&x.coeffRef(0) != x_ref.data()) + x = x_ref; + m_info = info==0 ? Success : NumericalIssue; } @@ -938,8 +946,12 @@ void SuperILU::_solve_impl(const MatrixBase &b, MatrixBase > b_ref(b); + Ref > x_ref(x); + + m_sluB = SluMatrix::Map(b_ref.const_cast_derived()); + m_sluX = SluMatrix::Map(x_ref.const_cast_derived()); typename Rhs::PlainObject b_cpy; if(m_sluEqued!='N') @@ -962,6 +974,9 @@ void SuperILU::_solve_impl(const MatrixBase &b, MatrixBase::_solve_impl(const MatrixBase &b, MatrixBas eigen_assert(b.derived().data() != x.derived().data() && " Umfpack does not support inplace solve"); int errorCode; + Scalar* x_ptr = 0; + Matrix x_tmp; + if(x.innerStride()!=1) + { + x_tmp.resize(x.rows()); + x_ptr = x_tmp.data(); + } for (int j=0; j Date: Wed, 4 Feb 2015 18:37:51 +0000 Subject: [PATCH 1054/1103] Using numext::pow instead of std::pow in poly_eval function. --- unsupported/Eigen/src/Polynomials/PolynomialUtils.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/unsupported/Eigen/src/Polynomials/PolynomialUtils.h b/unsupported/Eigen/src/Polynomials/PolynomialUtils.h index 2bb8bc84a..40ba65b7e 100644 --- a/unsupported/Eigen/src/Polynomials/PolynomialUtils.h +++ b/unsupported/Eigen/src/Polynomials/PolynomialUtils.h @@ -56,7 +56,7 @@ T poly_eval( const Polynomials& poly, const T& x ) for( DenseIndex i=1; i Date: Fri, 6 Feb 2015 02:51:59 -0800 Subject: [PATCH 1055/1103] Added the EIGEN_HAS_CONSTEXPR define Gate the tensor index list code based on the value of EIGEN_HAS_CONSTEXPR --- Eigen/src/Core/util/Macros.h | 6 ++++++ unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 2 +- unsupported/test/cxx11_tensor_index_list.cpp | 4 ++++ unsupported/test/cxx11_tensor_reduction.cpp | 8 ++++---- 5 files changed, 16 insertions(+), 6 deletions(-) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 001907a0b..40a28d4d6 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -133,6 +133,12 @@ #define EIGEN_HAS_VARIADIC_TEMPLATES 1 #endif +// Does the compiler support const expressions? +#if (defined(__plusplus) && __cplusplus >= 201402L) || \ + EIGEN_GNUC_AT_LEAST(4,9) +#define EIGEN_HAS_CONSTEXPR 1 +#endif + /** Allows to disable some optimizations which might affect the accuracy of the result. * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them. * They currently include: diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h index c94ed977e..eed0a9f05 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIndexList.h @@ -10,7 +10,7 @@ #ifndef EIGEN_CXX11_TENSOR_TENSOR_INDEX_LIST_H #define EIGEN_CXX11_TENSOR_TENSOR_INDEX_LIST_H -#if __cplusplus > 199711L +#ifdef EIGEN_HAS_CONSTEXPR namespace Eigen { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index 83ba1df71..21416afe0 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -53,7 +53,7 @@ struct preserve_inner_most_dims { static const bool value = false; }; -#if __cplusplus > 199711L +#ifdef EIGEN_HAS_CONSTEXPR template struct are_inner_most_dims{ static const bool value = indices_statically_known_to_increase()() && diff --git a/unsupported/test/cxx11_tensor_index_list.cpp b/unsupported/test/cxx11_tensor_index_list.cpp index d79a3ed45..c4d4f244f 100644 --- a/unsupported/test/cxx11_tensor_index_list.cpp +++ b/unsupported/test/cxx11_tensor_index_list.cpp @@ -11,6 +11,7 @@ #include +#ifdef EIGEN_HAS_CONSTEXPR static void test_static_index_list() { @@ -254,11 +255,14 @@ static void test_mixed_index_list() VERIFY_IS_APPROX(result3(0), expected); } +#endif void test_cxx11_tensor_index_list() { +#ifdef EIGEN_HAS_CONSTEXPR CALL_SUBTEST(test_static_index_list()); CALL_SUBTEST(test_type2index_list()); CALL_SUBTEST(test_dynamic_index_list()); CALL_SUBTEST(test_mixed_index_list()); +#endif } diff --git a/unsupported/test/cxx11_tensor_reduction.cpp b/unsupported/test/cxx11_tensor_reduction.cpp index 5c3184833..0269853a9 100644 --- a/unsupported/test/cxx11_tensor_reduction.cpp +++ b/unsupported/test/cxx11_tensor_reduction.cpp @@ -284,7 +284,7 @@ static void test_static_dims() { Tensor out(72, 97); in.setRandom(); -#if __cplusplus <= 199711L +#ifndef EIGEN_HAS_CONSTEXPR array reduction_axis; reduction_axis[0] = 1; reduction_axis[1] = 3; @@ -314,7 +314,7 @@ static void test_innermost_last_dims() { in.setRandom(); // Reduce on the innermost dimensions. -#if __cplusplus <= 199711L +#ifndef EIGEN_HAS_CONSTEXPR array reduction_axis; reduction_axis[0] = 0; reduction_axis[1] = 1; @@ -345,7 +345,7 @@ static void test_innermost_first_dims() { in.setRandom(); // Reduce on the innermost dimensions. -#if __cplusplus <= 199711L +#ifndef EIGEN_HAS_CONSTEXPR array reduction_axis; reduction_axis[0] = 2; reduction_axis[1] = 3; @@ -376,7 +376,7 @@ static void test_reduce_middle_dims() { in.setRandom(); // Reduce on the innermost dimensions. -#if __cplusplus <= 199711L +#ifndef EIGEN_HAS_CONSTEXPR array reduction_axis; reduction_axis[0] = 1; reduction_axis[1] = 2; From 2559fa9b0f20ea138cfb019d441ad1757221568d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Fri, 6 Feb 2015 02:55:18 -0800 Subject: [PATCH 1056/1103] Fixed compilation error in the tensor broadcasting test --- unsupported/test/cxx11_tensor_broadcasting.cpp | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/unsupported/test/cxx11_tensor_broadcasting.cpp b/unsupported/test/cxx11_tensor_broadcasting.cpp index f0792bdcf..2ddf47234 100644 --- a/unsupported/test/cxx11_tensor_broadcasting.cpp +++ b/unsupported/test/cxx11_tensor_broadcasting.cpp @@ -114,7 +114,15 @@ static void test_static_broadcasting() { Tensor tensor(8,3,5); tensor.setRandom(); + +#ifdef EIGEN_HAS_CONSTEXPR Eigen::IndexList, Eigen::type2index<3>, Eigen::type2index<4>> broadcasts; +#else + Eigen::array broadcasts; + broadcasts[0] = 2; + broadcasts[1] = 3; + broadcasts[2] = 4; +#endif Tensor broadcast; broadcast = tensor.broadcast(broadcasts); From 668518aed69c3d20efb480acd5944a79df7e5410 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 6 Feb 2015 14:25:41 +0100 Subject: [PATCH 1057/1103] Fix non initialized entries and comparison of very small numbers --- unsupported/test/cxx11_tensor_contraction.cpp | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/unsupported/test/cxx11_tensor_contraction.cpp b/unsupported/test/cxx11_tensor_contraction.cpp index 6124818fd..2bcae90b8 100644 --- a/unsupported/test/cxx11_tensor_contraction.cpp +++ b/unsupported/test/cxx11_tensor_contraction.cpp @@ -389,7 +389,7 @@ static void test_matrix_vector() m_result = m_left * m_right; for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { - VERIFY_IS_APPROX(t_result(i), m_result(i, 0)); + VERIFY(internal::isApprox(t_result(i), m_result(i, 0), 1)); } } @@ -399,6 +399,10 @@ static void test_tensor_vector() { Tensor t_left(7, 13, 17); Tensor t_right(1, 7); + + t_left.setRandom(); + t_right.setRandom(); + typedef typename Tensor::DimensionPair DimensionPair; Eigen::array dim_pair01{{{0, 1}}}; Tensor t_result = t_left.contract(t_right, dim_pair01); @@ -409,7 +413,7 @@ static void test_tensor_vector() Eigen::Matrix m_result = m_left.transpose() * m_right.transpose(); for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { - VERIFY_IS_APPROX(t_result(i), m_result(i, 0)); + VERIFY(internal::isApprox(t_result(i), m_result(i, 0), 1)); } } From c03c73c9b7032f984bcd6c52d9ca3a430ce19c69 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 6 Feb 2015 14:26:12 +0100 Subject: [PATCH 1058/1103] Fix clang compilation --- unsupported/test/cxx11_tensor_thread_pool.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp index e25912279..f49523683 100644 --- a/unsupported/test/cxx11_tensor_thread_pool.cpp +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -15,6 +15,7 @@ #include using Eigen::Tensor; +using std::isnan; static void test_multithread_elementwise() { From 74e460b9950503ef5a306337a136e1d37795deae Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 6 Feb 2015 14:26:24 +0100 Subject: [PATCH 1059/1103] Fix symmetric product --- Eigen/src/Core/products/SelfadjointMatrixMatrix.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index 21f8175d2..860e233b9 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -374,7 +374,7 @@ EIGEN_DONT_INLINE void product_selfadjoint_matrix Date: Fri, 6 Feb 2015 06:00:59 -0800 Subject: [PATCH 1060/1103] Fixed the cxx11_meta test --- unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h index 36d91e780..3a08628be 100644 --- a/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h +++ b/unsupported/Eigen/CXX11/src/Core/util/CXX11Meta.h @@ -42,14 +42,14 @@ struct numeric_list { constexpr static std::size_t count = sizeof.. * typename gen_numeric_list_repeated::type numeric_list */ -template struct gen_numeric_list : gen_numeric_list {}; -template struct gen_numeric_list { typedef numeric_list type; }; +template struct gen_numeric_list : gen_numeric_list {}; +template struct gen_numeric_list { typedef numeric_list type; }; -template struct gen_numeric_list_reversed : gen_numeric_list_reversed {}; -template struct gen_numeric_list_reversed { typedef numeric_list type; }; +template struct gen_numeric_list_reversed : gen_numeric_list_reversed {}; +template struct gen_numeric_list_reversed { typedef numeric_list type; }; -template struct gen_numeric_list_swapped_pair : gen_numeric_list_swapped_pair {}; -template struct gen_numeric_list_swapped_pair { typedef numeric_list type; }; +template struct gen_numeric_list_swapped_pair : gen_numeric_list_swapped_pair {}; +template struct gen_numeric_list_swapped_pair { typedef numeric_list type; }; template struct gen_numeric_list_repeated : gen_numeric_list_repeated {}; template struct gen_numeric_list_repeated { typedef numeric_list type; }; @@ -112,7 +112,7 @@ template struct get<0, type_lis template struct get> { static_assert((n - n) < 0, "meta-template get: The element to extract from a list must be smaller than the size of the list."); }; template struct get> : get> {}; -template struct get<0, numeric_list> { constexpr static T value = a; }; +template struct get<0, numeric_list> { constexpr static int value = a; }; template struct get> { static_assert((n - n) < 0, "meta-template get: The element to extract from a list must be smaller than the size of the list."); }; /* always get type, regardless of dummy; good for parameter pack expansion */ From 7838fda82cfc49b40e8d3a615bb05711815b50e1 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 7 Feb 2015 22:00:46 +0100 Subject: [PATCH 1061/1103] Add a SparseCompressedBase class providing (un)compressed accessors (like data()/*Stride() for dense matrices), and a CompressedAccessBit flag (similar to DirectAccessBit for dense matrices). --- Eigen/SparseCore | 1 + Eigen/src/Core/util/Constants.h | 13 ++ Eigen/src/SparseCore/SparseCompressedBase.h | 199 ++++++++++++++++++++ Eigen/src/SparseCore/SparseMatrix.h | 131 ++----------- Eigen/src/SparseCore/SparseUtil.h | 6 +- 5 files changed, 231 insertions(+), 119 deletions(-) create mode 100644 Eigen/src/SparseCore/SparseCompressedBase.h diff --git a/Eigen/SparseCore b/Eigen/SparseCore index d5c0f6271..e2071519b 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -31,6 +31,7 @@ #include "src/SparseCore/SparseAssign.h" #include "src/SparseCore/CompressedStorage.h" #include "src/SparseCore/AmbiVector.h" +#include "src/SparseCore/SparseCompressedBase.h" #include "src/SparseCore/SparseMatrix.h" #include "src/SparseCore/MappedSparseMatrix.h" #include "src/SparseCore/SparseVector.h" diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 9b40093f0..d1855b50b 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -163,6 +163,19 @@ const unsigned int NestByRefBit = 0x100; * \sa \ref RowMajorBit, \ref TopicStorageOrders */ const unsigned int NoPreferredStorageOrderBit = 0x200; +/** \ingroup flags + * + * Means that the underlying coefficients can be accessed through pointers to the sparse (un)compressed storage format, + * that is, the expression provides: + * \code + inline const Scalar* valuePtr() const; + inline const Index* innerIndexPtr() const; + inline const Index* outerIndexPtr() const; + inline const Index* innerNonZeroPtr() const; + \endcode + */ +const unsigned int CompressedAccessBit = 0x400; + // list of flags that are inherited by default const unsigned int HereditaryBits = RowMajorBit diff --git a/Eigen/src/SparseCore/SparseCompressedBase.h b/Eigen/src/SparseCore/SparseCompressedBase.h new file mode 100644 index 000000000..00658181e --- /dev/null +++ b/Eigen/src/SparseCore/SparseCompressedBase.h @@ -0,0 +1,199 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SPARSE_COMPRESSED_BASE_H +#define EIGEN_SPARSE_COMPRESSED_BASE_H + +namespace Eigen { + +template class SparseCompressedBase; + +namespace internal { + +template +struct traits > : traits +{}; + +} // end namespace internal + +template +class SparseCompressedBase + : public SparseMatrixBase +{ + public: + typedef SparseMatrixBase Base; + _EIGEN_SPARSE_PUBLIC_INTERFACE(SparseCompressedBase) + using Base::operator=; + using Base::IsRowMajor; + + class InnerIterator; + class ReverseInnerIterator; + + /** \returns a const pointer to the array of values. + * This function is aimed at interoperability with other libraries. + * \sa innerIndexPtr(), outerIndexPtr() */ + inline const Scalar* valuePtr() const { return derived().valuePtr(); } + /** \returns a non-const pointer to the array of values. + * This function is aimed at interoperability with other libraries. + * \sa innerIndexPtr(), outerIndexPtr() */ + inline Scalar* valuePtr() { return derived().valuePtr(); } + + /** \returns a const pointer to the array of inner indices. + * This function is aimed at interoperability with other libraries. + * \sa valuePtr(), outerIndexPtr() */ + inline const Index* innerIndexPtr() const { return derived().innerIndexPtr(); } + /** \returns a non-const pointer to the array of inner indices. + * This function is aimed at interoperability with other libraries. + * \sa valuePtr(), outerIndexPtr() */ + inline Index* innerIndexPtr() { return derived().innerIndexPtr(); } + + /** \returns a const pointer to the array of the starting positions of the inner vectors. + * This function is aimed at interoperability with other libraries. + * \sa valuePtr(), innerIndexPtr() */ + inline const Index* outerIndexPtr() const { return derived().outerIndexPtr(); } + /** \returns a non-const pointer to the array of the starting positions of the inner vectors. + * This function is aimed at interoperability with other libraries. + * \sa valuePtr(), innerIndexPtr() */ + inline Index* outerIndexPtr() { return derived().outerIndexPtr(); } + + /** \returns a const pointer to the array of the number of non zeros of the inner vectors. + * This function is aimed at interoperability with other libraries. + * \warning it returns the null pointer 0 in compressed mode */ + inline const Index* innerNonZeroPtr() const { return derived().innerNonZeroPtr(); } + /** \returns a non-const pointer to the array of the number of non zeros of the inner vectors. + * This function is aimed at interoperability with other libraries. + * \warning it returns the null pointer 0 in compressed mode */ + inline Index* innerNonZeroPtr() { return derived().innerNonZeroPtr(); } + + /** \returns whether \c *this is in compressed form. */ + inline bool isCompressed() const { return innerNonZeroPtr()==0; } + +}; + +template +class SparseCompressedBase::InnerIterator +{ + public: + InnerIterator(const SparseCompressedBase& mat, Index outer) + : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(outer), m_id(mat.outerIndexPtr()[outer]) + { + if(mat.isCompressed()) + m_end = mat.outerIndexPtr()[outer+1]; + else + m_end = m_id + mat.innerNonZeroPtr()[outer]; + } + + inline InnerIterator& operator++() { m_id++; return *this; } + + inline const Scalar& value() const { return m_values[m_id]; } + inline Scalar& valueRef() { return const_cast(m_values[m_id]); } + + inline Index index() const { return m_indices[m_id]; } + inline Index outer() const { return m_outer; } + inline Index row() const { return IsRowMajor ? m_outer : index(); } + inline Index col() const { return IsRowMajor ? index() : m_outer; } + + inline operator bool() const { return (m_id < m_end); } + + protected: + const Scalar* m_values; + const Index* m_indices; + const Index m_outer; + Index m_id; + Index m_end; + private: + // If you get here, then you're not using the right InnerIterator type, e.g.: + // SparseMatrix A; + // SparseMatrix::InnerIterator it(A,0); + template InnerIterator(const SparseMatrixBase&,Index outer); +}; + +template +class SparseCompressedBase::ReverseInnerIterator +{ + public: + ReverseInnerIterator(const SparseCompressedBase& mat, Index outer) + : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(outer), m_start(mat.outerIndexPtr()[outer]) + { + if(mat.isCompressed()) + m_id = mat.outerIndexPtr()[outer+1]; + else + m_id = m_start + mat.innerNonZeroPtr()[outer]; + } + + inline ReverseInnerIterator& operator--() { --m_id; return *this; } + + inline const Scalar& value() const { return m_values[m_id-1]; } + inline Scalar& valueRef() { return const_cast(m_values[m_id-1]); } + + inline Index index() const { return m_indices[m_id-1]; } + inline Index outer() const { return m_outer; } + inline Index row() const { return IsRowMajor ? m_outer : index(); } + inline Index col() const { return IsRowMajor ? index() : m_outer; } + + inline operator bool() const { return (m_id > m_start); } + + protected: + const Scalar* m_values; + const Index* m_indices; + const Index m_outer; + Index m_id; + const Index m_start; +}; + +namespace internal { + +template +struct evaluator > + : evaluator_base +{ + typedef typename Derived::Scalar Scalar; + typedef typename Derived::Index Index; + typedef typename Derived::InnerIterator InnerIterator; + typedef typename Derived::ReverseInnerIterator ReverseInnerIterator; + + enum { + CoeffReadCost = NumTraits::ReadCost, + Flags = Derived::Flags + }; + + evaluator() : m_matrix(0) {} + explicit evaluator(const Derived &mat) : m_matrix(&mat) {} + + operator Derived&() { return m_matrix->const_cast_derived(); } + operator const Derived&() const { return *m_matrix; } + + typedef typename DenseCoeffsBase::CoeffReturnType CoeffReturnType; + Scalar coeff(Index row, Index col) const + { return m_matrix->coeff(row,col); } + + Scalar& coeffRef(Index row, Index col) + { + eigen_internal_assert(row>=0 && rowrows() && col>=0 && colcols()); + + const Index outer = Derived::IsRowMajor ? row : col; + const Index inner = Derived::IsRowMajor ? col : row; + + Index start = m_matrix->outerIndexPtr()[outer]; + Index end = m_matrix->isCompressed() ? m_matrix->outerIndexPtr()[outer+1] : m_matrix->outerIndexPtr()[outer] + m_matrix->innerNonZeroPtr()[outer]; + eigen_assert(end>start && "you are using a non finalized sparse matrix or written coefficient does not exist"); + const Index p = std::lower_bound(m_matrix->innerIndexPtr()+start, m_matrix->innerIndexPtr()+end,inner) + - m_matrix->innerIndexPtr(); + eigen_assert((pinnerIndexPtr()[p]==inner) && "written coefficient does not exist"); + return m_matrix->const_cast_derived().valuePtr()[p]; + } + + const Derived *m_matrix; +}; + +} + +} // end namespace Eigen + +#endif // EIGEN_SPARSE_COMPRESSED_BASE_H diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 93677c786..74b4c6a9d 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -51,7 +51,7 @@ struct traits > ColsAtCompileTime = Dynamic, MaxRowsAtCompileTime = Dynamic, MaxColsAtCompileTime = Dynamic, - Flags = _Options | NestByRefBit | LvalueBit, + Flags = _Options | NestByRefBit | LvalueBit | CompressedAccessBit, SupportedAccessPatterns = InnerRandomAccessPattern }; }; @@ -90,16 +90,20 @@ struct traits, DiagIndex> template class SparseMatrix - : public SparseMatrixBase > + : public SparseCompressedBase > { public: - EIGEN_SPARSE_PUBLIC_INTERFACE(SparseMatrix) + typedef SparseCompressedBase Base; + using Base::isCompressed; + _EIGEN_SPARSE_PUBLIC_INTERFACE(SparseMatrix) EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(SparseMatrix, +=) EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(SparseMatrix, -=) typedef MappedSparseMatrix Map; typedef Diagonal DiagonalReturnType; typedef Diagonal ConstDiagonalReturnType; + typedef typename Base::InnerIterator InnerIterator; + typedef typename Base::ReverseInnerIterator ReverseInnerIterator; using Base::IsRowMajor; @@ -123,9 +127,6 @@ class SparseMatrix public: - /** \returns whether \c *this is in compressed form. */ - inline bool isCompressed() const { return m_innerNonZeros==0; } - /** \returns the number of rows of the matrix */ inline Index rows() const { return IsRowMajor ? m_outerSize : m_innerSize; } /** \returns the number of columns of the matrix */ @@ -241,9 +242,6 @@ class SparseMatrix public: - class InnerIterator; - class ReverseInnerIterator; - /** Removes all non zeros but keep allocated memory */ inline void setZero() { @@ -875,76 +873,6 @@ private: }; }; -template -class SparseMatrix::InnerIterator -{ - public: - InnerIterator(const SparseMatrix& mat, Index outer) - : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(outer), m_id(mat.m_outerIndex[outer]) - { - if(mat.isCompressed()) - m_end = mat.m_outerIndex[outer+1]; - else - m_end = m_id + mat.m_innerNonZeros[outer]; - } - - inline InnerIterator& operator++() { m_id++; return *this; } - - inline const Scalar& value() const { return m_values[m_id]; } - inline Scalar& valueRef() { return const_cast(m_values[m_id]); } - - inline Index index() const { return m_indices[m_id]; } - inline Index outer() const { return m_outer; } - inline Index row() const { return IsRowMajor ? m_outer : index(); } - inline Index col() const { return IsRowMajor ? index() : m_outer; } - - inline operator bool() const { return (m_id < m_end); } - - protected: - const Scalar* m_values; - const Index* m_indices; - const Index m_outer; - Index m_id; - Index m_end; - private: - // If you get here, then you're not using the right InnerIterator type, e.g.: - // SparseMatrix A; - // SparseMatrix::InnerIterator it(A,0); - template InnerIterator(const SparseMatrixBase&,Index outer); -}; - -template -class SparseMatrix::ReverseInnerIterator -{ - public: - ReverseInnerIterator(const SparseMatrix& mat, Index outer) - : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(outer), m_start(mat.m_outerIndex[outer]) - { - if(mat.isCompressed()) - m_id = mat.m_outerIndex[outer+1]; - else - m_id = m_start + mat.m_innerNonZeros[outer]; - } - - inline ReverseInnerIterator& operator--() { --m_id; return *this; } - - inline const Scalar& value() const { return m_values[m_id-1]; } - inline Scalar& valueRef() { return const_cast(m_values[m_id-1]); } - - inline Index index() const { return m_indices[m_id-1]; } - inline Index outer() const { return m_outer; } - inline Index row() const { return IsRowMajor ? m_outer : index(); } - inline Index col() const { return IsRowMajor ? index() : m_outer; } - - inline operator bool() const { return (m_id > m_start); } - - protected: - const Scalar* m_values; - const Index* m_indices; - const Index m_outer; - Index m_id; - const Index m_start; -}; namespace internal { @@ -1075,6 +1003,10 @@ EIGEN_DONT_INLINE SparseMatrix& SparseMatrix::value), YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY) + #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN + EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN + #endif + const bool needToTranspose = (Flags & RowMajorBit) != (internal::evaluator::Flags & RowMajorBit); if (needToTranspose) { @@ -1277,45 +1209,12 @@ namespace internal { template struct evaluator > - : evaluator_base > + : evaluator > > { - typedef _Scalar Scalar; - typedef _Index Index; + typedef evaluator > > Base; typedef SparseMatrix<_Scalar,_Options,_Index> SparseMatrixType; - typedef typename SparseMatrixType::InnerIterator InnerIterator; - typedef typename SparseMatrixType::ReverseInnerIterator ReverseInnerIterator; - - enum { - CoeffReadCost = NumTraits<_Scalar>::ReadCost, - Flags = SparseMatrixType::Flags - }; - - evaluator() : m_matrix(0) {} - explicit evaluator(const SparseMatrixType &mat) : m_matrix(&mat) {} - - operator SparseMatrixType&() { return m_matrix->const_cast_derived(); } - operator const SparseMatrixType&() const { return *m_matrix; } - - typedef typename DenseCoeffsBase::CoeffReturnType CoeffReturnType; - Scalar coeff(Index row, Index col) const - { return m_matrix->coeff(row,col); } - - Scalar& coeffRef(Index row, Index col) - { - eigen_internal_assert(row>=0 && rowrows() && col>=0 && colcols()); - - const Index outer = SparseMatrixType::IsRowMajor ? row : col; - const Index inner = SparseMatrixType::IsRowMajor ? col : row; - - Index start = m_matrix->outerIndexPtr()[outer]; - Index end = m_matrix->isCompressed() ? m_matrix->outerIndexPtr()[outer+1] : m_matrix->outerIndexPtr()[outer] + m_matrix->innerNonZeroPtr()[outer]; - eigen_assert(end>start && "you are using a non finalized sparse matrix or written coefficient does not exist"); - const Index p = m_matrix->data().searchLowerIndex(start,end-1,inner); - eigen_assert((pdata().index(p)==inner) && "written coefficient does not exist"); - return m_matrix->const_cast_derived().data().value(p); - } - - const SparseMatrixType *m_matrix; + evaluator() : Base() {} + explicit evaluator(const SparseMatrixType &mat) : Base(mat) {} }; } diff --git a/Eigen/src/SparseCore/SparseUtil.h b/Eigen/src/SparseCore/SparseUtil.h index 8de227b88..ba4803646 100644 --- a/Eigen/src/SparseCore/SparseUtil.h +++ b/Eigen/src/SparseCore/SparseUtil.h @@ -43,8 +43,7 @@ EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(Derived, -=) \ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, *=) \ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) -#define _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, BaseClass) \ - typedef BaseClass Base; \ +#define _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) \ typedef typename Eigen::internal::traits::Scalar Scalar; \ typedef typename Eigen::NumTraits::Real RealScalar; \ typedef typename Eigen::internal::nested::type Nested; \ @@ -59,7 +58,8 @@ EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, /=) using Base::const_cast_derived; #define EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) \ - _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived, Eigen::SparseMatrixBase) + typedef Eigen::SparseMatrixBase Base; \ + _EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) const int CoherentAccessPattern = 0x1; const int InnerRandomAccessPattern = 0x2 | CoherentAccessPattern; From 08081f829397de11651d8d779b9eed916ccc88d7 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 7 Feb 2015 22:02:14 +0100 Subject: [PATCH 1062/1103] Make SparseTranspose inherit SparseCompressBase when possible --- Eigen/src/SparseCore/SparseTranspose.h | 34 +++++++++++++++++++++++--- 1 file changed, 31 insertions(+), 3 deletions(-) diff --git a/Eigen/src/SparseCore/SparseTranspose.h b/Eigen/src/SparseCore/SparseTranspose.h index c3d2d1a16..37ce7b0d5 100644 --- a/Eigen/src/SparseCore/SparseTranspose.h +++ b/Eigen/src/SparseCore/SparseTranspose.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008-2014 Gael Guennebaud +// Copyright (C) 2008-2015 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -12,13 +12,41 @@ namespace Eigen { +namespace internal { + template + class SparseTransposeImpl + : public SparseMatrixBase > + {}; + + template + class SparseTransposeImpl + : public SparseCompressedBase > + { + typedef SparseCompressedBase > Base; + public: + using Base::derived; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + + inline const Scalar* valuePtr() const { return derived().nestedExpression().valuePtr(); } + inline const Index* innerIndexPtr() const { return derived().nestedExpression().innerIndexPtr(); } + inline const Index* outerIndexPtr() const { return derived().nestedExpression().outerIndexPtr(); } + inline const Index* innerNonZeroPtr() const { return derived().nestedExpression().innerNonZeroPtr(); } + + inline Scalar* valuePtr() { return derived().nestedExpression().valuePtr(); } + inline Index* innerIndexPtr() { return derived().nestedExpression().innerIndexPtr(); } + inline Index* outerIndexPtr() { return derived().nestedExpression().outerIndexPtr(); } + inline Index* innerNonZeroPtr() { return derived().nestedExpression().innerNonZeroPtr(); } + }; +} + // Implement nonZeros() for transpose. I'm not sure that's the best approach for that. // Perhaps it should be implemented in Transpose<> itself. template class TransposeImpl - : public SparseMatrixBase > + : public internal::SparseTransposeImpl { protected: - typedef SparseMatrixBase > Base; + typedef internal::SparseTransposeImpl Base; public: inline typename MatrixType::Index nonZeros() const { return Base::derived().nestedExpression().nonZeros(); } }; From f3be317614b6b6709737d6ef7bea0ffe96c285d4 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 7 Feb 2015 22:03:25 +0100 Subject: [PATCH 1063/1103] Add a Map specialization. --- Eigen/SparseCore | 1 + Eigen/src/SparseCore/SparseMap.h | 211 +++++++++++++++++++++++++++++++ 2 files changed, 212 insertions(+) create mode 100644 Eigen/src/SparseCore/SparseMap.h diff --git a/Eigen/SparseCore b/Eigen/SparseCore index e2071519b..91cdfd3f0 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -33,6 +33,7 @@ #include "src/SparseCore/AmbiVector.h" #include "src/SparseCore/SparseCompressedBase.h" #include "src/SparseCore/SparseMatrix.h" +#include "src/SparseCore/SparseMap.h" #include "src/SparseCore/MappedSparseMatrix.h" #include "src/SparseCore/SparseVector.h" #include "src/SparseCore/SparseCwiseUnaryOp.h" diff --git a/Eigen/src/SparseCore/SparseMap.h b/Eigen/src/SparseCore/SparseMap.h new file mode 100644 index 000000000..91e8f7480 --- /dev/null +++ b/Eigen/src/SparseCore/SparseMap.h @@ -0,0 +1,211 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SPARSE_MAP_H +#define EIGEN_SPARSE_MAP_H + +namespace Eigen { + +template::has_write_access ? WriteAccessors : ReadOnlyAccessors +> class SparseMapBase; + +template +class SparseMapBase + : public SparseCompressedBase +{ + public: + typedef SparseCompressedBase Base; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + enum { IsRowMajor = Base::IsRowMajor }; + + protected: + + typedef typename internal::conditional< + bool(internal::is_lvalue::value), + Scalar *, const Scalar *>::type ScalarPointer; + typedef typename internal::conditional< + bool(internal::is_lvalue::value), + Index *, const Index *>::type IndexPointer; + + Index m_outerSize; + Index m_innerSize; + Index m_nnz; + IndexPointer m_outerIndex; + IndexPointer m_innerIndices; + ScalarPointer m_values; + IndexPointer m_innerNonZeros; + + public: + + inline Index rows() const { return IsRowMajor ? m_outerSize : m_innerSize; } + inline Index cols() const { return IsRowMajor ? m_innerSize : m_outerSize; } + inline Index innerSize() const { return m_innerSize; } + inline Index outerSize() const { return m_outerSize; } + + bool isCompressed() const { return m_innerNonZeros==0; } + + //---------------------------------------- + // direct access interface + inline const Scalar* valuePtr() const { return m_values; } + inline const Index* innerIndexPtr() const { return m_innerIndices; } + inline const Index* outerIndexPtr() const { return m_outerIndex; } + inline const Index* innerNonZeroPtr() const { return m_innerNonZeros; } + //---------------------------------------- + + inline Scalar coeff(Index row, Index col) const + { + const Index outer = IsRowMajor ? row : col; + const Index inner = IsRowMajor ? col : row; + + Index start = m_outerIndex[outer]; + Index end = isCompressed() ? m_outerIndex[outer+1] : start + m_innerNonZeros[outer]; + if (start==end) + return Scalar(0); + else if (end>0 && inner==m_innerIndices[end-1]) + return m_values[end-1]; + // ^^ optimization: let's first check if it is the last coefficient + // (very common in high level algorithms) + + const Index* r = std::lower_bound(&m_innerIndices[start],&m_innerIndices[end-1],inner); + const Index id = r-&m_innerIndices[0]; + return ((*r==inner) && (id +class SparseMapBase + : public SparseMapBase +{ + typedef MapBase ReadOnlyMapBase; + + public: + typedef SparseMapBase Base; + typedef typename Base::Scalar Scalar; + typedef typename Base::Index Index; + enum { IsRowMajor = Base::IsRowMajor }; + + public: + + //---------------------------------------- + // direct access interface + using Base::valuePtr; + using Base::innerIndexPtr; + using Base::outerIndexPtr; + using Base::innerNonZeroPtr; + inline Scalar* valuePtr() { return Base::m_values; } + inline Index* innerIndexPtr() { return Base::m_innerIndices; } + inline Index* outerIndexPtr() { return Base::m_outerIndex; } + inline Index* innerNonZeroPtr() { return Base::m_innerNonZeros; } + //---------------------------------------- + + inline Scalar& coeffRef(Index row, Index col) + { + const Index outer = IsRowMajor ? row : col; + const Index inner = IsRowMajor ? col : row; + + Index start = Base::m_outerIndex[outer]; + Index end = Base::isCompressed() ? Base::m_outerIndex[outer+1] : start + Base::m_innerNonZeros[outer]; + eigen_assert(end>=start && "you probably called coeffRef on a non finalized matrix"); + eigen_assert(end>start && "coeffRef cannot be called on a zero coefficient"); + Index* r = std::lower_bound(&Base::m_innerIndices[start],&Base::m_innerIndices[end],inner); + const Index id = r - &Base::m_innerIndices[0]; + eigen_assert((*r==inner) && (id(Base::m_values)[id]; + } + + inline SparseMapBase(Index rows, Index cols, Index nnz, Index* outerIndexPtr, Index* innerIndexPtr, + Scalar* valuePtr, Index* innerNonZerosPtr = 0) + : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZerosPtr) + {} + + /** Empty destructor */ + inline ~SparseMapBase() {} +}; + +template +class Map, Options, StrideType> + : public SparseMapBase, Options, StrideType> > +{ + public: + typedef SparseMapBase > Base; + _EIGEN_SPARSE_PUBLIC_INTERFACE(Map) + enum { IsRowMajor = Base::IsRowMajor }; + + public: + + inline Map(Index rows, Index cols, Index nnz, Index* outerIndexPtr, + Index* innerIndexPtr, Scalar* valuePtr, Index* innerNonZerosPtr = 0) + : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZerosPtr) + {} + + /** Empty destructor */ + inline ~Map() {} +}; + +template +class Map, Options, StrideType> + : public SparseMapBase, Options, StrideType> > +{ + public: + typedef SparseMapBase > Base; + _EIGEN_SPARSE_PUBLIC_INTERFACE(Map) + enum { IsRowMajor = Base::IsRowMajor }; + + public: + + inline Map(Index rows, Index cols, Index nnz, const Index* outerIndexPtr, + const Index* innerIndexPtr, const Scalar* valuePtr, const Index* innerNonZerosPtr = 0) + : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZerosPtr) + {} + + /** Empty destructor */ + inline ~Map() {} +}; + +namespace internal { + +template +struct evaluator, Options, StrideType> > + : evaluator, Options, StrideType> > > +{ + typedef evaluator, Options, StrideType> > > Base; + typedef Map, Options, StrideType> XprType; + evaluator() : Base() {} + explicit evaluator(const XprType &mat) : Base(mat) {} +}; + +template +struct evaluator, Options, StrideType> > + : evaluator, Options, StrideType> > > +{ + typedef evaluator, Options, StrideType> > > Base; + typedef Map, Options, StrideType> XprType; + evaluator() : Base() {} + explicit evaluator(const XprType &mat) : Base(mat) {} +}; + +} + +} // end namespace Eigen + +#endif // EIGEN_SPARSE_MAP_H From f2ff8c091e02b4aab8c7568807c09e43f51d1156 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Sat, 7 Feb 2015 22:04:18 +0100 Subject: [PATCH 1064/1103] Add a Ref specialization. --- Eigen/SparseCore | 1 + Eigen/src/SparseCore/SparseRef.h | 200 +++++++++++++++++++++++++++++++ test/CMakeLists.txt | 1 + test/sparse_ref.cpp | 96 +++++++++++++++ 4 files changed, 298 insertions(+) create mode 100644 Eigen/src/SparseCore/SparseRef.h create mode 100644 test/sparse_ref.cpp diff --git a/Eigen/SparseCore b/Eigen/SparseCore index 91cdfd3f0..48ed967b8 100644 --- a/Eigen/SparseCore +++ b/Eigen/SparseCore @@ -36,6 +36,7 @@ #include "src/SparseCore/SparseMap.h" #include "src/SparseCore/MappedSparseMatrix.h" #include "src/SparseCore/SparseVector.h" +#include "src/SparseCore/SparseRef.h" #include "src/SparseCore/SparseCwiseUnaryOp.h" #include "src/SparseCore/SparseCwiseBinaryOp.h" #include "src/SparseCore/SparseTranspose.h" diff --git a/Eigen/src/SparseCore/SparseRef.h b/Eigen/src/SparseCore/SparseRef.h new file mode 100644 index 000000000..cfea6ce8a --- /dev/null +++ b/Eigen/src/SparseCore/SparseRef.h @@ -0,0 +1,200 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2015 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_SPARSE_REF_H +#define EIGEN_SPARSE_REF_H + +namespace Eigen { + +namespace internal { + +template class SparseRefBase; + +template +struct traits, _Options, _StrideType> > + : public traits > +{ + typedef SparseMatrix PlainObjectType; + enum { + Options = _Options, + Flags = traits >::Flags | CompressedAccessBit | NestByRefBit + }; + + template struct match { + enum { + StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || Derived::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)), + MatchAtCompileTime = (Derived::Flags&CompressedAccessBit) && StorageOrderMatch + }; + typedef typename internal::conditional::type type; + }; + +}; + +template +struct traits, _Options, _StrideType> > + : public traits, _Options, _StrideType> > +{ + enum { + Flags = (traits >::Flags | CompressedAccessBit | NestByRefBit) & ~LvalueBit + }; +}; + +template +struct traits > : public traits {}; + +template class SparseRefBase +// : public MappedSparseMatrix + : public SparseMapBase +{ +// typedef typename internal::traits::PlainObjectType PlainObjectType; + +public: + + typedef SparseMapBase Base; + _EIGEN_SPARSE_PUBLIC_INTERFACE(SparseRefBase) + + SparseRefBase() + : Base(RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime, 0, 0, 0, 0, 0) + {} + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(SparseRefBase) + +protected: + + + template + void construct(Expression& expr) + { + ::new (static_cast(this)) Base(expr.rows(), expr.cols(), expr.nonZeros(), expr.outerIndexPtr(), expr.innerIndexPtr(), expr.valuePtr(), expr.innerNonZeroPtr()); + } +}; + +} // namespace internal + +template +class Ref, Options, StrideType > + : public internal::SparseRefBase, Options, StrideType > > +{ + typedef SparseMatrix PlainObjectType; + typedef internal::traits Traits; + template + inline Ref(const SparseMatrix& expr); + template + inline Ref(const MappedSparseMatrix& expr); + public: + + typedef internal::SparseRefBase Base; + _EIGEN_SPARSE_PUBLIC_INTERFACE(Ref) + + + #ifndef EIGEN_PARSED_BY_DOXYGEN + template + inline Ref(SparseMatrix& expr) + { + EIGEN_STATIC_ASSERT(bool(Traits::template match >::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH); + Base::construct(expr.derived()); + } + + template + inline Ref(MappedSparseMatrix& expr) + { + EIGEN_STATIC_ASSERT(bool(Traits::template match >::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH); + Base::construct(expr.derived()); + } + + template + inline Ref(const SparseCompressedBase& expr) + #else + template + inline Ref(SparseCompressedBase& expr) + #endif + { + EIGEN_STATIC_ASSERT(bool(internal::is_lvalue::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY); + EIGEN_STATIC_ASSERT(bool(Traits::template match::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH); + Base::construct(expr.const_cast_derived()); + } + + EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Ref) + +}; + +// this is the const ref version +template +class Ref, Options, StrideType> + : public internal::SparseRefBase, Options, StrideType> > +{ + typedef SparseMatrix TPlainObjectType; + typedef internal::traits Traits; + public: + + typedef internal::SparseRefBase Base; + _EIGEN_SPARSE_PUBLIC_INTERFACE(Ref) + + template + inline Ref(const SparseMatrixBase& expr) + { + construct(expr.derived(), typename Traits::template match::type()); + } + + inline Ref(const Ref& other) : Base(other) { + // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy + } + + template + inline Ref(const RefBase& other) { + construct(other.derived(), typename Traits::template match::type()); + } + + protected: + + template + void construct(const Expression& expr,internal::true_type) + { + Base::construct(expr); + } + + template + void construct(const Expression& expr, internal::false_type) + { + m_object = expr; + Base::construct(m_object); + } + + protected: + TPlainObjectType m_object; +}; + + +namespace internal { + +template +struct evaluator, Options, StrideType> > + : evaluator, Options, StrideType> > > +{ + typedef evaluator, Options, StrideType> > > Base; + typedef Ref, Options, StrideType> XprType; + evaluator() : Base() {} + explicit evaluator(const XprType &mat) : Base(mat) {} +}; + +template +struct evaluator, Options, StrideType> > + : evaluator, Options, StrideType> > > +{ + typedef evaluator, Options, StrideType> > > Base; + typedef Ref, Options, StrideType> XprType; + evaluator() : Base() {} + explicit evaluator(const XprType &mat) : Base(mat) {} +}; + +} + +} // end namespace Eigen + +#endif // EIGEN_SPARSE_REF_H diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index f57d8ce36..168749634 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -228,6 +228,7 @@ ei_add_test(stddeque) ei_add_test(sparse_basic) ei_add_test(sparse_vector) ei_add_test(sparse_product) +ei_add_test(sparse_ref) ei_add_test(sparse_solvers) ei_add_test(sparse_permutations) ei_add_test(simplicial_cholesky) diff --git a/test/sparse_ref.cpp b/test/sparse_ref.cpp new file mode 100644 index 000000000..b261cecf3 --- /dev/null +++ b/test/sparse_ref.cpp @@ -0,0 +1,96 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 20015 Gael Guennebaud +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +// This unit test cannot be easily written to work with EIGEN_DEFAULT_TO_ROW_MAJOR +#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR +#undef EIGEN_DEFAULT_TO_ROW_MAJOR +#endif + +static long int nb_temporaries; + +inline void on_temporary_creation() { + // here's a great place to set a breakpoint when debugging failures in this test! + nb_temporaries++; +} + +#define EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN { on_temporary_creation(); } + +#include "main.h" +#include + +#define VERIFY_EVALUATION_COUNT(XPR,N) {\ + nb_temporaries = 0; \ + XPR; \ + if(nb_temporaries!=N) std::cerr << "nb_temporaries == " << nb_temporaries << "\n"; \ + VERIFY( (#XPR) && nb_temporaries==N ); \ + } + +template void check_const_correctness(const PlainObjectType&) +{ + // verify that ref-to-const don't have LvalueBit + typedef typename internal::add_const::type ConstPlainObjectType; + VERIFY( !(internal::traits >::Flags & LvalueBit) ); + VERIFY( !(internal::traits >::Flags & LvalueBit) ); + VERIFY( !(Ref::Flags & LvalueBit) ); + VERIFY( !(Ref::Flags & LvalueBit) ); +} + +template +EIGEN_DONT_INLINE void call_ref_1(Ref > a, const B &b) { VERIFY_IS_EQUAL(a.toDense(),b.toDense()); } + +template +EIGEN_DONT_INLINE void call_ref_2(const Ref >& a, const B &b) { VERIFY_IS_EQUAL(a.toDense(),b.toDense()); } + +void call_ref() +{ +// SparseVector > ca = VectorXcf::Random(10).sparseView(); +// SparseVector a = VectorXf::Random(10).sparseView(); + SparseMatrix A = MatrixXf::Random(10,10).sparseView(); + SparseMatrix B = MatrixXf::Random(10,10).sparseView(); + const SparseMatrix& Ac(A); + Block > Ab(A,0,1, 3,3); + const Block > Abc(A,0,1,3,3); + + + VERIFY_EVALUATION_COUNT( call_ref_1(A, A), 0); +// VERIFY_EVALUATION_COUNT( call_ref_1(Ac, Ac), 0); // does not compile on purpose + VERIFY_EVALUATION_COUNT( call_ref_2(A, A), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(A.transpose(), A.transpose()), 1); + VERIFY_EVALUATION_COUNT( call_ref_2(Ac,Ac), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(A+A,2*Ac), 1); + VERIFY_EVALUATION_COUNT( call_ref_2(B, B), 1); + VERIFY_EVALUATION_COUNT( call_ref_2(B.transpose(), B.transpose()), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(A*A, A*A), 1); + + Ref > Ar(A); + VERIFY_EVALUATION_COUNT( call_ref_1(Ar, Ar), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(Ar, Ar), 0); + + Ref > Br(B); + VERIFY_EVALUATION_COUNT( call_ref_1(Br.transpose(), Br.transpose()), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(Br, Br), 1); + VERIFY_EVALUATION_COUNT( call_ref_2(Br.transpose(), Br.transpose()), 0); + + Ref > Arc(A); +// VERIFY_EVALUATION_COUNT( call_ref_1(Arc, Arc), 0); // does not compile on purpose + VERIFY_EVALUATION_COUNT( call_ref_2(Arc, Arc), 0); + + VERIFY_EVALUATION_COUNT( call_ref_2(A.middleCols(1,3), A.middleCols(1,3)), 1); // should be 0 + + VERIFY_EVALUATION_COUNT( call_ref_2(A.block(1,1,3,3), A.block(1,1,3,3)), 1); // should be 0 (allocate starts/nnz only) +} + +void test_sparse_ref() +{ + for(int i = 0; i < g_repeat; i++) { + CALL_SUBTEST_1( check_const_correctness(SparseMatrix()) ); + CALL_SUBTEST_1( check_const_correctness(SparseMatrix()) ); + CALL_SUBTEST_2( call_ref() ); + } +} From 3af29caae87b00ed8f002605573d227ad1b629a4 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 9 Feb 2015 10:23:45 +0100 Subject: [PATCH 1065/1103] Cleaning and add more unit tests for Ref and Map --- Eigen/src/SparseCore/MappedSparseMatrix.h | 172 ++-------------------- Eigen/src/SparseCore/SparseMap.h | 34 ++++- Eigen/src/SparseCore/SparseRef.h | 12 +- test/sparse_basic.cpp | 20 +++ test/sparse_ref.cpp | 5 +- 5 files changed, 72 insertions(+), 171 deletions(-) diff --git a/Eigen/src/SparseCore/MappedSparseMatrix.h b/Eigen/src/SparseCore/MappedSparseMatrix.h index 2852c669a..533479fd0 100644 --- a/Eigen/src/SparseCore/MappedSparseMatrix.h +++ b/Eigen/src/SparseCore/MappedSparseMatrix.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2008 Gael Guennebaud +// Copyright (C) 2008-2014 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -10,9 +10,10 @@ #ifndef EIGEN_MAPPED_SPARSEMATRIX_H #define EIGEN_MAPPED_SPARSEMATRIX_H -namespace Eigen { +namespace Eigen { -/** \class MappedSparseMatrix +/** \deprecated Use Map > + * \class MappedSparseMatrix * * \brief Sparse matrix * @@ -25,179 +26,38 @@ namespace internal { template struct traits > : traits > {}; -} +} // end namespace internal template class MappedSparseMatrix - : public SparseMatrixBase > + : public Map > { - public: - EIGEN_SPARSE_PUBLIC_INTERFACE(MappedSparseMatrix) - enum { IsRowMajor = Base::IsRowMajor }; - - protected: - - Index m_outerSize; - Index m_innerSize; - Index m_nnz; - Index* m_outerIndex; - Index* m_innerIndices; - Scalar* m_values; + typedef Map > Base; public: - - inline Index rows() const { return IsRowMajor ? m_outerSize : m_innerSize; } - inline Index cols() const { return IsRowMajor ? m_innerSize : m_outerSize; } - inline Index innerSize() const { return m_innerSize; } - inline Index outerSize() const { return m_outerSize; } - bool isCompressed() const { return true; } + typedef typename Base::Index Index; + typedef typename Base::Scalar Scalar; - //---------------------------------------- - // direct access interface - inline const Scalar* valuePtr() const { return m_values; } - inline Scalar* valuePtr() { return m_values; } - - inline const Index* innerIndexPtr() const { return m_innerIndices; } - inline Index* innerIndexPtr() { return m_innerIndices; } - - inline const Index* outerIndexPtr() const { return m_outerIndex; } - inline Index* outerIndexPtr() { return m_outerIndex; } - //---------------------------------------- - - inline Scalar coeff(Index row, Index col) const - { - const Index outer = IsRowMajor ? row : col; - const Index inner = IsRowMajor ? col : row; - - Index start = m_outerIndex[outer]; - Index end = m_outerIndex[outer+1]; - if (start==end) - return Scalar(0); - else if (end>0 && inner==m_innerIndices[end-1]) - return m_values[end-1]; - // ^^ optimization: let's first check if it is the last coefficient - // (very common in high level algorithms) - - const Index* r = std::lower_bound(&m_innerIndices[start],&m_innerIndices[end-1],inner); - const Index id = r-&m_innerIndices[0]; - return ((*r==inner) && (id=start && "you probably called coeffRef on a non finalized matrix"); - eigen_assert(end>start && "coeffRef cannot be called on a zero coefficient"); - Index* r = std::lower_bound(&m_innerIndices[start],&m_innerIndices[end],inner); - const Index id = r-&m_innerIndices[0]; - eigen_assert((*r==inner) && (id -class MappedSparseMatrix::InnerIterator -{ - public: - InnerIterator(const MappedSparseMatrix& mat, Index outer) - : m_matrix(mat), - m_outer(outer), - m_id(mat.outerIndexPtr()[outer]), - m_start(m_id), - m_end(mat.outerIndexPtr()[outer+1]) - {} - - inline InnerIterator& operator++() { m_id++; return *this; } - - inline Scalar value() const { return m_matrix.valuePtr()[m_id]; } - inline Scalar& valueRef() { return const_cast(m_matrix.valuePtr()[m_id]); } - - inline Index index() const { return m_matrix.innerIndexPtr()[m_id]; } - inline Index row() const { return IsRowMajor ? m_outer : index(); } - inline Index col() const { return IsRowMajor ? index() : m_outer; } - - inline operator bool() const { return (m_id < m_end) && (m_id>=m_start); } - - protected: - const MappedSparseMatrix& m_matrix; - const Index m_outer; - Index m_id; - const Index m_start; - const Index m_end; -}; - -template -class MappedSparseMatrix::ReverseInnerIterator -{ - public: - ReverseInnerIterator(const MappedSparseMatrix& mat, Index outer) - : m_matrix(mat), - m_outer(outer), - m_id(mat.outerIndexPtr()[outer+1]), - m_start(mat.outerIndexPtr()[outer]), - m_end(m_id) - {} - - inline ReverseInnerIterator& operator--() { m_id--; return *this; } - - inline Scalar value() const { return m_matrix.valuePtr()[m_id-1]; } - inline Scalar& valueRef() { return const_cast(m_matrix.valuePtr()[m_id-1]); } - - inline Index index() const { return m_matrix.innerIndexPtr()[m_id-1]; } - inline Index row() const { return IsRowMajor ? m_outer : index(); } - inline Index col() const { return IsRowMajor ? index() : m_outer; } - - inline operator bool() const { return (m_id <= m_end) && (m_id>m_start); } - - protected: - const MappedSparseMatrix& m_matrix; - const Index m_outer; - Index m_id; - const Index m_start; - const Index m_end; -}; - namespace internal { template struct evaluator > - : evaluator_base > + : evaluator > > { - typedef MappedSparseMatrix<_Scalar,_Options,_Index> MappedSparseMatrixType; - typedef typename MappedSparseMatrixType::InnerIterator InnerIterator; - typedef typename MappedSparseMatrixType::ReverseInnerIterator ReverseInnerIterator; + typedef MappedSparseMatrix<_Scalar,_Options,_Index> XprType; + typedef evaluator > Base; - enum { - CoeffReadCost = NumTraits<_Scalar>::ReadCost, - Flags = MappedSparseMatrixType::Flags - }; - - evaluator() : m_matrix(0) {} - explicit evaluator(const MappedSparseMatrixType &mat) : m_matrix(&mat) {} - - operator MappedSparseMatrixType&() { return m_matrix->const_cast_derived(); } - operator const MappedSparseMatrixType&() const { return *m_matrix; } - - const MappedSparseMatrixType *m_matrix; + evaluator() : Base() {} + explicit evaluator(const XprType &mat) : Base(mat) {} }; } diff --git a/Eigen/src/SparseCore/SparseMap.h b/Eigen/src/SparseCore/SparseMap.h index 91e8f7480..72dedb1ec 100644 --- a/Eigen/src/SparseCore/SparseMap.h +++ b/Eigen/src/SparseCore/SparseMap.h @@ -12,6 +12,32 @@ namespace Eigen { +namespace internal { + +template +struct traits, Options, StrideType> > + : public traits > +{ + typedef SparseMatrix PlainObjectType; + typedef traits TraitsBase; + enum { + Flags = TraitsBase::Flags & (~NestByRefBit) + }; +}; + +template +struct traits, Options, StrideType> > + : public traits > +{ + typedef SparseMatrix PlainObjectType; + typedef traits TraitsBase; + enum { + Flags = TraitsBase::Flags & (~ (NestByRefBit | LvalueBit)) + }; +}; + +} // end namespace internal + template::has_write_access ? WriteAccessors : ReadOnlyAccessors > class SparseMapBase; @@ -25,7 +51,7 @@ class SparseMapBase typedef typename Base::Scalar Scalar; typedef typename Base::Index Index; enum { IsRowMajor = Base::IsRowMajor }; - + using Base::operator=; protected: typedef typename internal::conditional< @@ -103,6 +129,8 @@ class SparseMapBase typedef typename Base::Scalar Scalar; typedef typename Base::Index Index; enum { IsRowMajor = Base::IsRowMajor }; + + using Base::operator=; public: @@ -147,7 +175,7 @@ class Map, Options, StrideType> : public SparseMapBase, Options, StrideType> > { public: - typedef SparseMapBase > Base; + typedef SparseMapBase Base; _EIGEN_SPARSE_PUBLIC_INTERFACE(Map) enum { IsRowMajor = Base::IsRowMajor }; @@ -167,7 +195,7 @@ class Map, Options, StrideType : public SparseMapBase, Options, StrideType> > { public: - typedef SparseMapBase > Base; + typedef SparseMapBase Base; _EIGEN_SPARSE_PUBLIC_INTERFACE(Map) enum { IsRowMajor = Base::IsRowMajor }; diff --git a/Eigen/src/SparseCore/SparseRef.h b/Eigen/src/SparseCore/SparseRef.h index cfea6ce8a..2ca039323 100644 --- a/Eigen/src/SparseCore/SparseRef.h +++ b/Eigen/src/SparseCore/SparseRef.h @@ -23,7 +23,7 @@ struct traits, _Options, _Stride typedef SparseMatrix PlainObjectType; enum { Options = _Options, - Flags = traits >::Flags | CompressedAccessBit | NestByRefBit + Flags = traits >::Flags | CompressedAccessBit | NestByRefBit }; template struct match { @@ -41,7 +41,7 @@ struct traits, _Options, _ : public traits, _Options, _StrideType> > { enum { - Flags = (traits >::Flags | CompressedAccessBit | NestByRefBit) & ~LvalueBit + Flags = (traits >::Flags | CompressedAccessBit | NestByRefBit) & ~LvalueBit }; }; @@ -49,11 +49,8 @@ template struct traits > : public traits {}; template class SparseRefBase -// : public MappedSparseMatrix : public SparseMapBase { -// typedef typename internal::traits::PlainObjectType PlainObjectType; - public: typedef SparseMapBase Base; @@ -63,8 +60,6 @@ public: : Base(RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime, 0, 0, 0, 0, 0) {} - EIGEN_INHERIT_ASSIGNMENT_OPERATORS(SparseRefBase) - protected: @@ -119,9 +114,6 @@ class Ref, Options, StrideType > EIGEN_STATIC_ASSERT(bool(Traits::template match::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH); Base::construct(expr.const_cast_derived()); } - - EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Ref) - }; // this is the const ref version diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index 097959c84..a19de296a 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -408,6 +408,26 @@ template void sparse_basic(const SparseMatrixType& re m.setFromTriplets(triplets.begin(), triplets.end()); VERIFY_IS_APPROX(m, refMat); } + + // test Map + { + DenseMatrix refMat2(rows, cols), refMat3(rows, cols); + SparseMatrixType m2(rows, cols), m3(rows, cols); + initSparse(density, refMat2, m2); + initSparse(density, refMat3, m3); + { + Map mapMat2(m2.rows(), m2.cols(), m2.nonZeros(), m2.outerIndexPtr(), m2.innerIndexPtr(), m2.valuePtr(), m2.innerNonZeroPtr()); + Map mapMat3(m3.rows(), m3.cols(), m3.nonZeros(), m3.outerIndexPtr(), m3.innerIndexPtr(), m3.valuePtr(), m3.innerNonZeroPtr()); + VERIFY_IS_APPROX(mapMat2+mapMat3, refMat2+refMat3); + VERIFY_IS_APPROX(mapMat2+mapMat3, refMat2+refMat3); + } + { + MappedSparseMatrix mapMat2(m2.rows(), m2.cols(), m2.nonZeros(), m2.outerIndexPtr(), m2.innerIndexPtr(), m2.valuePtr(), m2.innerNonZeroPtr()); + MappedSparseMatrix mapMat3(m3.rows(), m3.cols(), m3.nonZeros(), m3.outerIndexPtr(), m3.innerIndexPtr(), m3.valuePtr(), m3.innerNonZeroPtr()); + VERIFY_IS_APPROX(mapMat2+mapMat3, refMat2+refMat3); + VERIFY_IS_APPROX(mapMat2+mapMat3, refMat2+refMat3); + } + } // test triangularView { diff --git a/test/sparse_ref.cpp b/test/sparse_ref.cpp index b261cecf3..27700f827 100644 --- a/test/sparse_ref.cpp +++ b/test/sparse_ref.cpp @@ -69,8 +69,9 @@ void call_ref() VERIFY_EVALUATION_COUNT( call_ref_2(A*A, A*A), 1); Ref > Ar(A); - VERIFY_EVALUATION_COUNT( call_ref_1(Ar, Ar), 0); - VERIFY_EVALUATION_COUNT( call_ref_2(Ar, Ar), 0); + VERIFY_IS_APPROX(Ar+Ar, A+A); + VERIFY_EVALUATION_COUNT( call_ref_1(Ar, A), 0); + VERIFY_EVALUATION_COUNT( call_ref_2(Ar, A), 0); Ref > Br(B); VERIFY_EVALUATION_COUNT( call_ref_1(Br.transpose(), Br.transpose()), 0); From 554aa9b31de06cf1d4464b1fe8e5a956091641ba Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 9 Feb 2015 10:24:07 +0100 Subject: [PATCH 1066/1103] Add failtests for Ref --- failtest/CMakeLists.txt | 6 ++++++ failtest/sparse_ref_1.cpp | 18 ++++++++++++++++++ failtest/sparse_ref_2.cpp | 15 +++++++++++++++ failtest/sparse_ref_3.cpp | 15 +++++++++++++++ failtest/sparse_ref_4.cpp | 15 +++++++++++++++ failtest/sparse_ref_5.cpp | 16 ++++++++++++++++ 6 files changed, 85 insertions(+) create mode 100644 failtest/sparse_ref_1.cpp create mode 100644 failtest/sparse_ref_2.cpp create mode 100644 failtest/sparse_ref_3.cpp create mode 100644 failtest/sparse_ref_4.cpp create mode 100644 failtest/sparse_ref_5.cpp diff --git a/failtest/CMakeLists.txt b/failtest/CMakeLists.txt index d2fea7bdc..c8795a344 100644 --- a/failtest/CMakeLists.txt +++ b/failtest/CMakeLists.txt @@ -41,6 +41,12 @@ ei_add_failtest("ref_5") ei_add_failtest("swap_1") ei_add_failtest("swap_2") +ei_add_failtest("sparse_ref_1") +ei_add_failtest("sparse_ref_2") +ei_add_failtest("sparse_ref_3") +ei_add_failtest("sparse_ref_4") +ei_add_failtest("sparse_ref_5") + if (EIGEN_FAILTEST_FAILURE_COUNT) message(FATAL_ERROR "${EIGEN_FAILTEST_FAILURE_COUNT} out of ${EIGEN_FAILTEST_COUNT} failtests FAILED. " diff --git a/failtest/sparse_ref_1.cpp b/failtest/sparse_ref_1.cpp new file mode 100644 index 000000000..d78d1f9b1 --- /dev/null +++ b/failtest/sparse_ref_1.cpp @@ -0,0 +1,18 @@ +#include "../Eigen/Sparse" + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +#define CV_QUALIFIER const +#else +#define CV_QUALIFIER +#endif + +using namespace Eigen; + +void call_ref(Ref > a) { } + +int main() +{ + SparseMatrix a(10,10); + CV_QUALIFIER SparseMatrix& ac(a); + call_ref(ac); +} diff --git a/failtest/sparse_ref_2.cpp b/failtest/sparse_ref_2.cpp new file mode 100644 index 000000000..46c9440c2 --- /dev/null +++ b/failtest/sparse_ref_2.cpp @@ -0,0 +1,15 @@ +#include "../Eigen/Sparse" + +using namespace Eigen; + +void call_ref(Ref > a) { } + +int main() +{ + SparseMatrix A(10,10); +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD + call_ref(A.row(3)); +#else + call_ref(A.col(3)); +#endif +} diff --git a/failtest/sparse_ref_3.cpp b/failtest/sparse_ref_3.cpp new file mode 100644 index 000000000..a9949b552 --- /dev/null +++ b/failtest/sparse_ref_3.cpp @@ -0,0 +1,15 @@ +#include "../Eigen/Sparse" + +using namespace Eigen; + +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD +void call_ref(Ref > a) { } +#else +void call_ref(const Ref > &a) { } +#endif + +int main() +{ + SparseMatrix a(10,10); + call_ref(a+a); +} diff --git a/failtest/sparse_ref_4.cpp b/failtest/sparse_ref_4.cpp new file mode 100644 index 000000000..57bb6a1fc --- /dev/null +++ b/failtest/sparse_ref_4.cpp @@ -0,0 +1,15 @@ +#include "../Eigen/Sparse" + +using namespace Eigen; + +void call_ref(Ref > a) {} + +int main() +{ + SparseMatrix A(10,10); +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD + call_ref(A.transpose()); +#else + call_ref(A); +#endif +} diff --git a/failtest/sparse_ref_5.cpp b/failtest/sparse_ref_5.cpp new file mode 100644 index 000000000..4478f6f2f --- /dev/null +++ b/failtest/sparse_ref_5.cpp @@ -0,0 +1,16 @@ +#include "../Eigen/Sparse" + +using namespace Eigen; + +void call_ref(Ref > a) { } + +int main() +{ + SparseMatrix a(10,10); + SparseMatrixBase > &ac(a); +#ifdef EIGEN_SHOULD_FAIL_TO_BUILD + call_ref(ac); +#else + call_ref(ac.derived()); +#endif +} From d4ec48575e94ec469ef9fbf005a0a6e67571f528 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 9 Feb 2015 11:14:36 +0100 Subject: [PATCH 1067/1103] Make Block inherit SparseCompressedBase in the case of an inner-panels and fix valuePtr() innerIndexPtr() --- Eigen/src/Core/Block.h | 2 +- Eigen/src/SparseCore/SparseBlock.h | 17 ++++++++++++----- test/sparse_ref.cpp | 8 +++++--- 3 files changed, 18 insertions(+), 9 deletions(-) diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h index 9cf9d5432..5f6307206 100644 --- a/Eigen/src/Core/Block.h +++ b/Eigen/src/Core/Block.h @@ -87,7 +87,7 @@ struct traits > : traits::value ? LvalueBit : 0, FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0, - Flags = (traits::Flags & DirectAccessBit) | FlagsLvalueBit | FlagsRowMajorBit + Flags = (traits::Flags & (DirectAccessBit | (InnerPanel?CompressedAccessBit:0))) | FlagsLvalueBit | FlagsRowMajorBit // FIXME DirectAccessBit should not be handled by expressions }; }; diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 9e4da2057..6b8ade5aa 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -74,7 +74,7 @@ namespace internal { template class sparse_matrix_block_impl - : public SparseMatrixBase > + : public SparseCompressedBase > { typedef typename internal::remove_all::type _MatrixTypeNested; typedef Block BlockType; @@ -172,19 +172,24 @@ public: } inline const Scalar* valuePtr() const - { return m_matrix.valuePtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + { return m_matrix.valuePtr(); } inline Scalar* valuePtr() - { return m_matrix.const_cast_derived().valuePtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + { return m_matrix.const_cast_derived().valuePtr(); } inline const Index* innerIndexPtr() const - { return m_matrix.innerIndexPtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + { return m_matrix.innerIndexPtr(); } inline Index* innerIndexPtr() - { return m_matrix.const_cast_derived().innerIndexPtr() + m_matrix.outerIndexPtr()[m_outerStart]; } + { return m_matrix.const_cast_derived().innerIndexPtr(); } inline const Index* outerIndexPtr() const { return m_matrix.outerIndexPtr() + m_outerStart; } inline Index* outerIndexPtr() { return m_matrix.const_cast_derived().outerIndexPtr() + m_outerStart; } + + inline const Index* innerNonZeroPtr() const + { return isCompressed() ? 0 : m_matrix.innerNonZeroPtr(); } + inline Index* innerNonZeroPtr() + { return isCompressed() ? 0 : m_matrix.const_cast_derived().innerNonZeroPtr(); } Index nonZeros() const { @@ -196,6 +201,8 @@ public: else return Map >(m_matrix.innerNonZeroPtr()+m_outerStart, m_outerSize.value()).sum(); } + + bool isCompressed() const { return m_matrix.innerNonZeroPtr()==0; } const Scalar& lastCoeff() const { diff --git a/test/sparse_ref.cpp b/test/sparse_ref.cpp index 27700f827..e7380ba21 100644 --- a/test/sparse_ref.cpp +++ b/test/sparse_ref.cpp @@ -51,8 +51,8 @@ void call_ref() { // SparseVector > ca = VectorXcf::Random(10).sparseView(); // SparseVector a = VectorXf::Random(10).sparseView(); - SparseMatrix A = MatrixXf::Random(10,10).sparseView(); - SparseMatrix B = MatrixXf::Random(10,10).sparseView(); + SparseMatrix A = MatrixXf::Random(10,10).sparseView(0.5,1); + SparseMatrix B = MatrixXf::Random(10,10).sparseView(0.5,1); const SparseMatrix& Ac(A); Block > Ab(A,0,1, 3,3); const Block > Abc(A,0,1,3,3); @@ -82,7 +82,9 @@ void call_ref() // VERIFY_EVALUATION_COUNT( call_ref_1(Arc, Arc), 0); // does not compile on purpose VERIFY_EVALUATION_COUNT( call_ref_2(Arc, Arc), 0); - VERIFY_EVALUATION_COUNT( call_ref_2(A.middleCols(1,3), A.middleCols(1,3)), 1); // should be 0 + VERIFY_EVALUATION_COUNT( call_ref_2(A.middleCols(1,3), A.middleCols(1,3)), 0); + + VERIFY_EVALUATION_COUNT( call_ref_2(A.col(2), A.col(2)), 0); VERIFY_EVALUATION_COUNT( call_ref_2(A.block(1,1,3,3), A.block(1,1,3,3)), 1); // should be 0 (allocate starts/nnz only) } From 87629cd6395433966977e868ec091c3fc754956c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 9 Feb 2015 11:41:25 +0100 Subject: [PATCH 1068/1103] bug #897: makes iterative sparse solvers use a Ref instead of a SparseMatrix pointer. This fixes usage of iterative solvers with a Map. --- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 2 +- .../ConjugateGradient.h | 2 +- .../IterativeSolverBase.h | 53 +++++++++++++------ test/sparse_solver.h | 5 +- 4 files changed, 41 insertions(+), 21 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index 224fe913f..a50680133 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -193,7 +193,7 @@ public: m_error = Base::m_tolerance; typename Dest::ColXpr xj(x,j); - if(!internal::bicgstab(*mp_matrix, b.col(j), xj, Base::m_preconditioner, m_iterations, m_error)) + if(!internal::bicgstab(mp_matrix, b.col(j), xj, Base::m_preconditioner, m_iterations, m_error)) failed = true; } m_info = failed ? NumericalIssue diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index b5ef6d60f..3e024bda1 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -206,7 +206,7 @@ public: m_error = Base::m_tolerance; typename Dest::ColXpr xj(x,j); - internal::conjugate_gradient(mp_matrix->template selfadjointView(), b.col(j), xj, + internal::conjugate_gradient(mp_matrix.template selfadjointView(), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error); } diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h index f33c868bb..6f48075b4 100644 --- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h @@ -37,7 +37,7 @@ public: /** Default constructor. */ IterativeSolverBase() - : mp_matrix(0) + : m_dummy(0,0), mp_matrix(m_dummy) { init(); } @@ -52,10 +52,11 @@ public: * this class becomes invalid. Call compute() to update it with the new * matrix A, or modify a copy of A. */ - explicit IterativeSolverBase(const MatrixType& A) + template + explicit IterativeSolverBase(const SparseMatrixBase& A) { init(); - compute(A); + compute(A.derived()); } ~IterativeSolverBase() {} @@ -65,9 +66,11 @@ public: * Currently, this function mostly calls analyzePattern on the preconditioner. In the future * we might, for instance, implement column reordering for faster matrix vector products. */ - Derived& analyzePattern(const MatrixType& A) + template + Derived& analyzePattern(const SparseMatrixBase& A) { - m_preconditioner.analyzePattern(A); + grab(A); + m_preconditioner.analyzePattern(mp_matrix); m_isInitialized = true; m_analysisIsOk = true; m_info = Success; @@ -83,11 +86,12 @@ public: * this class becomes invalid. Call compute() to update it with the new * matrix A, or modify a copy of A. */ - Derived& factorize(const MatrixType& A) + template + Derived& factorize(const SparseMatrixBase& A) { eigen_assert(m_analysisIsOk && "You must first call analyzePattern()"); - mp_matrix = &A; - m_preconditioner.factorize(A); + grab(A); + m_preconditioner.factorize(mp_matrix); m_factorizationIsOk = true; m_info = Success; return derived(); @@ -103,10 +107,11 @@ public: * this class becomes invalid. Call compute() to update it with the new * matrix A, or modify a copy of A. */ - Derived& compute(const MatrixType& A) + template + Derived& compute(const SparseMatrixBase& A) { - mp_matrix = &A; - m_preconditioner.compute(A); + grab(A); + m_preconditioner.compute(mp_matrix); m_isInitialized = true; m_analysisIsOk = true; m_factorizationIsOk = true; @@ -115,9 +120,9 @@ public: } /** \internal */ - Index rows() const { return mp_matrix ? mp_matrix->rows() : 0; } + Index rows() const { return mp_matrix.rows(); } /** \internal */ - Index cols() const { return mp_matrix ? mp_matrix->cols() : 0; } + Index cols() const { return mp_matrix.cols(); } /** \returns the tolerance threshold used by the stopping criteria */ RealScalar tolerance() const { return m_tolerance; } @@ -135,13 +140,18 @@ public: /** \returns a read-only reference to the preconditioner. */ const Preconditioner& preconditioner() const { return m_preconditioner; } - /** \returns the max number of iterations */ + /** \returns the max number of iterations. + * It is either the value setted by setMaxIterations or, by default, + * twice the number of columns of the matrix. + */ int maxIterations() const { - return (mp_matrix && m_maxIterations<0) ? mp_matrix->cols() : m_maxIterations; + return (m_maxIterations<0) ? 2*mp_matrix.cols() : m_maxIterations; } - /** Sets the max number of iterations */ + /** Sets the max number of iterations. + * Default is twice the number of columns of the matrix. + */ Derived& setMaxIterations(int maxIters) { m_maxIterations = maxIters; @@ -210,7 +220,16 @@ protected: m_maxIterations = -1; m_tolerance = NumTraits::epsilon(); } - const MatrixType* mp_matrix; + + template + void grab(const SparseMatrixBase &A) + { + mp_matrix.~Ref(); + ::new (&mp_matrix) Ref(A); + } + + MatrixType m_dummy; + Ref mp_matrix; Preconditioner m_preconditioner; int m_maxIterations; diff --git a/test/sparse_solver.h b/test/sparse_solver.h index ee350d561..887ff02a9 100644 --- a/test/sparse_solver.h +++ b/test/sparse_solver.h @@ -32,7 +32,7 @@ void check_sparse_solving(Solver& solver, const typename Solver::MatrixType& A, x = solver.solve(b); if (solver.info() != Success) { - std::cerr << "sparse solver testing: solving failed\n"; + std::cerr << "sparse solver testing: solving failed (" << typeid(Solver).name() << ")\n"; return; } VERIFY(oldb.isApprox(b) && "sparse solver testing: the rhs should not be modified!"); @@ -75,7 +75,8 @@ void check_sparse_solving(Solver& solver, const typename Solver::MatrixType& A, xm = solver.solve(bm); if (solver.info() != Success) { - std::cerr << "sparse solver testing: solving failed\n"; + std::cerr << "sparse solver testing: solving with a Map failed\n"; + exit(0); return; } VERIFY(oldb.isApprox(bm) && "sparse solver testing: the rhs should not be modified!"); From d10d6a40dda3fb5ac9f401b8e6d9cede3f3ca34a Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 10 Feb 2015 13:02:59 +0100 Subject: [PATCH 1069/1103] bug #897: Update unsupported iterative solvers based on IterativeSolverBased. --- unsupported/Eigen/src/IterativeSolvers/DGMRES.h | 2 +- unsupported/Eigen/src/IterativeSolvers/GMRES.h | 2 +- unsupported/Eigen/src/IterativeSolvers/MINRES.h | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h index 0e1b7d977..52eb65a2f 100644 --- a/unsupported/Eigen/src/IterativeSolvers/DGMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/DGMRES.h @@ -150,7 +150,7 @@ class DGMRES : public IterativeSolverBase > m_error = Base::m_tolerance; typename Dest::ColXpr xj(x,j); - dgmres(*mp_matrix, b.col(j), xj, Base::m_preconditioner); + dgmres(mp_matrix, b.col(j), xj, Base::m_preconditioner); } m_info = failed ? NumericalIssue : m_error <= Base::m_tolerance ? Success diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index cd15ce0bf..39610c074 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -327,7 +327,7 @@ public: m_error = Base::m_tolerance; typename Dest::ColXpr xj(x,j); - if(!internal::gmres(*mp_matrix, b.col(j), xj, Base::m_preconditioner, m_iterations, m_restart, m_error)) + if(!internal::gmres(mp_matrix, b.col(j), xj, Base::m_preconditioner, m_iterations, m_restart, m_error)) failed = true; } m_info = failed ? NumericalIssue diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index aaf42c78a..a34902001 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -259,7 +259,7 @@ namespace Eigen { m_error = Base::m_tolerance; typename Dest::ColXpr xj(x,j); - internal::minres(mp_matrix->template selfadjointView(), b.col(j), xj, + internal::minres(mp_matrix.template selfadjointView(), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error); } From c6e8caf0900ae303e9e7399bed00af705015ff17 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 10 Feb 2015 18:57:41 +0100 Subject: [PATCH 1070/1103] Allows Lower|Upper as a template argument of CG and MINRES: in this case the full matrix will be considered. --- Eigen/src/IterativeLinearSolvers/ConjugateGradient.h | 11 +++++++---- test/conjugate_gradient.cpp | 6 ++++-- test/sparse_solver.h | 5 ++++- unsupported/Eigen/src/IterativeSolvers/MINRES.h | 7 ++++++- unsupported/test/minres.cpp | 2 ++ 5 files changed, 23 insertions(+), 8 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index 3e024bda1..4857dd9e9 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -113,8 +113,8 @@ struct traits > * The matrix A must be selfadjoint. The matrix A and the vectors x and b can be either dense or sparse. * * \tparam _MatrixType the type of the matrix A, can be a dense or a sparse matrix. - * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower - * or Upper. Default is Lower. + * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower, + * Upper, or Lower|Upper in which the full matrix entries will be considered. Default is Lower. * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner * * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations() @@ -197,6 +197,10 @@ public: template void _solve_with_guess_impl(const Rhs& b, Dest& x) const { + typedef typename internal::conditional&, + SparseSelfAdjointView, UpLo> + >::type MatrixWrapperType; m_iterations = Base::maxIterations(); m_error = Base::m_tolerance; @@ -206,8 +210,7 @@ public: m_error = Base::m_tolerance; typename Dest::ColXpr xj(x,j); - internal::conjugate_gradient(mp_matrix.template selfadjointView(), b.col(j), xj, - Base::m_preconditioner, m_iterations, m_error); + internal::conjugate_gradient(MatrixWrapperType(mp_matrix), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error); } m_isInitialized = true; diff --git a/test/conjugate_gradient.cpp b/test/conjugate_gradient.cpp index 869051b31..019cc4d64 100644 --- a/test/conjugate_gradient.cpp +++ b/test/conjugate_gradient.cpp @@ -12,13 +12,15 @@ template void test_conjugate_gradient_T() { - ConjugateGradient, Lower> cg_colmajor_lower_diag; - ConjugateGradient, Upper> cg_colmajor_upper_diag; + ConjugateGradient, Lower > cg_colmajor_lower_diag; + ConjugateGradient, Upper > cg_colmajor_upper_diag; + ConjugateGradient, Lower|Upper> cg_colmajor_loup_diag; ConjugateGradient, Lower, IdentityPreconditioner> cg_colmajor_lower_I; ConjugateGradient, Upper, IdentityPreconditioner> cg_colmajor_upper_I; CALL_SUBTEST( check_sparse_spd_solving(cg_colmajor_lower_diag) ); CALL_SUBTEST( check_sparse_spd_solving(cg_colmajor_upper_diag) ); + CALL_SUBTEST( check_sparse_spd_solving(cg_colmajor_loup_diag) ); CALL_SUBTEST( check_sparse_spd_solving(cg_colmajor_lower_I) ); CALL_SUBTEST( check_sparse_spd_solving(cg_colmajor_upper_I) ); } diff --git a/test/sparse_solver.h b/test/sparse_solver.h index 887ff02a9..6cf99e0b0 100644 --- a/test/sparse_solver.h +++ b/test/sparse_solver.h @@ -195,7 +195,10 @@ int generate_sparse_spd_problem(Solver& , typename Solver::MatrixType& A, typena dA = dM * dM.adjoint(); halfA.resize(size,size); - halfA.template selfadjointView().rankUpdate(M); + if(Solver::UpLo==(Lower|Upper)) + halfA = A; + else + halfA.template selfadjointView().rankUpdate(M); return size; } diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index a34902001..65cffc255 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -250,6 +250,11 @@ namespace Eigen { template void _solve_with_guess_impl(const Rhs& b, Dest& x) const { + typedef typename internal::conditional&, + SparseSelfAdjointView, UpLo> + >::type MatrixWrapperType; + m_iterations = Base::maxIterations(); m_error = Base::m_tolerance; @@ -259,7 +264,7 @@ namespace Eigen { m_error = Base::m_tolerance; typename Dest::ColXpr xj(x,j); - internal::minres(mp_matrix.template selfadjointView(), b.col(j), xj, + internal::minres(MatrixWrapperType(mp_matrix), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error); } diff --git a/unsupported/test/minres.cpp b/unsupported/test/minres.cpp index 81b762c37..f6e526bbd 100644 --- a/unsupported/test/minres.cpp +++ b/unsupported/test/minres.cpp @@ -21,6 +21,7 @@ template void test_minres_T() // Diagonal preconditioner MINRES, Lower, DiagonalPreconditioner > minres_colmajor_lower_diag; MINRES, Upper, DiagonalPreconditioner > minres_colmajor_upper_diag; + MINRES, Upper, DiagonalPreconditioner > minres_colmajor_uplo_diag; // call tests for SPD matrix CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_lower_I) ); @@ -28,6 +29,7 @@ template void test_minres_T() CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_lower_diag) ); CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_upper_diag) ); + CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_uplo_diag) ); // TO DO: symmetric semi-definite matrix // TO DO: symmetric indefinite matrix From deecff97edfb6f75e7613e1db97a1e3e5504e971 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 10 Feb 2015 19:22:05 +0100 Subject: [PATCH 1071/1103] typo --- blas/common.h | 3 +-- blas/level3_impl.h | 2 +- blas/xerbla.cpp | 4 ++-- unsupported/Eigen/src/IterativeSolvers/MINRES.h | 4 ++-- unsupported/test/minres.cpp | 2 +- 5 files changed, 7 insertions(+), 8 deletions(-) diff --git a/blas/common.h b/blas/common.h index c39cc63a8..5ecb153e2 100644 --- a/blas/common.h +++ b/blas/common.h @@ -1,7 +1,7 @@ // This file is part of Eigen, a lightweight C++ template library // for linear algebra. // -// Copyright (C) 2009-2010 Gael Guennebaud +// Copyright (C) 2009-2015 Gael Guennebaud // // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed @@ -13,7 +13,6 @@ #include #include -#include #include #ifndef SCALAR diff --git a/blas/level3_impl.h b/blas/level3_impl.h index a05872666..32313e814 100644 --- a/blas/level3_impl.h +++ b/blas/level3_impl.h @@ -8,7 +8,7 @@ // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. #include "common.h" - +#include int EIGEN_BLAS_FUNC(gemm)(char *opa, char *opb, int *m, int *n, int *k, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *pb, int *ldb, RealScalar *pbeta, RealScalar *pc, int *ldc) { // std::cerr << "in gemm " << *opa << " " << *opb << " " << *m << " " << *n << " " << *k << " " << *lda << " " << *ldb << " " << *ldc << " " << *palpha << " " << *pbeta << "\n"; diff --git a/blas/xerbla.cpp b/blas/xerbla.cpp index 0422f79b7..8775b88cd 100644 --- a/blas/xerbla.cpp +++ b/blas/xerbla.cpp @@ -1,5 +1,5 @@ -#include +#include #if (defined __GNUC__) && (!defined __MINGW32__) #define EIGEN_WEAK_LINKING __attribute__ ((weak)) @@ -14,7 +14,7 @@ extern "C" EIGEN_WEAK_LINKING int xerbla_(const char * msg, int *info, int) { - std::cerr << "Eigen BLAS ERROR #" << *info << ": " << msg << "\n"; + printf("Eigen BLAS ERROR #%i: %s\n", *info, msg ); return 0; } diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index 65cffc255..93a83e5b7 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -165,8 +165,8 @@ namespace Eigen { * The vectors x and b can be either dense or sparse. * * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix. - * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower - * or Upper. Default is Lower. + * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower, + * Upper, or Lower|Upper in which the full matrix entries will be considered. Default is Lower. * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner * * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations() diff --git a/unsupported/test/minres.cpp b/unsupported/test/minres.cpp index f6e526bbd..8b300b78a 100644 --- a/unsupported/test/minres.cpp +++ b/unsupported/test/minres.cpp @@ -21,7 +21,7 @@ template void test_minres_T() // Diagonal preconditioner MINRES, Lower, DiagonalPreconditioner > minres_colmajor_lower_diag; MINRES, Upper, DiagonalPreconditioner > minres_colmajor_upper_diag; - MINRES, Upper, DiagonalPreconditioner > minres_colmajor_uplo_diag; + MINRES, Lower|Upper, DiagonalPreconditioner > minres_colmajor_uplo_diag; // call tests for SPD matrix CALL_SUBTEST( check_sparse_spd_solving(minres_colmajor_lower_I) ); From c3f3580b8f7e4af89f4c7cdbe036ac1cac750128 Mon Sep 17 00:00:00 2001 From: Jan Blechta Date: Tue, 10 Feb 2015 14:24:39 +0100 Subject: [PATCH 1072/1103] Fix bug #733: step by step solving is not a good example for solveWithGuess --- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 6 +----- .../IterativeLinearSolvers/ConjugateGradient.h | 15 +-------------- unsupported/Eigen/src/IterativeSolvers/GMRES.h | 17 ++--------------- unsupported/Eigen/src/IterativeSolvers/MINRES.h | 15 +-------------- 4 files changed, 5 insertions(+), 48 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index a50680133..31a43cb56 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -139,11 +139,7 @@ struct traits > * \include BiCGSTAB_simple.cpp * * By default the iterations start with x=0 as an initial guess of the solution. - * One can control the start using the solveWithGuess() method. Here is a step by - * step execution example starting with a random guess and printing the evolution - * of the estimated error: - * \include BiCGSTAB_step_by_step.cpp - * Note that such a step by step execution is slightly slower. + * One can control the start using the solveWithGuess() method. * * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner */ diff --git a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h index 4857dd9e9..1e819fc9f 100644 --- a/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h +++ b/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h @@ -137,20 +137,7 @@ struct traits > * \endcode * * By default the iterations start with x=0 as an initial guess of the solution. - * One can control the start using the solveWithGuess() method. Here is a step by - * step execution example starting with a random guess and printing the evolution - * of the estimated error: - * * \code - * x = VectorXd::Random(n); - * cg.setMaxIterations(1); - * int i = 0; - * do { - * x = cg.solveWithGuess(b,x); - * std::cout << i << " : " << cg.error() << std::endl; - * ++i; - * } while (cg.info()!=Success && i<100); - * \endcode - * Note that such a step by step excution is slightly slower. + * One can control the start using the solveWithGuess() method. * * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner */ diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index 39610c074..30f82b52e 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -250,21 +250,8 @@ struct traits > * \endcode * * By default the iterations start with x=0 as an initial guess of the solution. - * One can control the start using the solveWithGuess() method. Here is a step by - * step execution example starting with a random guess and printing the evolution - * of the estimated error: - * * \code - * x = VectorXd::Random(n); - * solver.setMaxIterations(1); - * int i = 0; - * do { - * x = solver.solveWithGuess(b,x); - * std::cout << i << " : " << solver.error() << std::endl; - * ++i; - * } while (solver.info()!=Success && i<100); - * \endcode - * Note that such a step by step excution is slightly slower. - * + * One can control the start using the solveWithGuess() method. + * * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner */ template< typename _MatrixType, typename _Preconditioner> diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index 93a83e5b7..c4d969a72 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -189,20 +189,7 @@ namespace Eigen { * \endcode * * By default the iterations start with x=0 as an initial guess of the solution. - * One can control the start using the solveWithGuess() method. Here is a step by - * step execution example starting with a random guess and printing the evolution - * of the estimated error: - * * \code - * x = VectorXd::Random(n); - * mr.setMaxIterations(1); - * int i = 0; - * do { - * x = mr.solveWithGuess(b,x); - * std::cout << i << " : " << mr.error() << std::endl; - * ++i; - * } while (mr.info()!=Success && i<100); - * \endcode - * Note that such a step by step excution is slightly slower. + * One can control the start using the solveWithGuess() method. * * \sa class ConjugateGradient, BiCGSTAB, SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner */ From 91fe3a30043874e51225c8f25964687320c9b601 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 10:29:28 -0800 Subject: [PATCH 1073/1103] Removed a debug printf statement. --- unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h | 3 --- 1 file changed, 3 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h index a96d705a4..7e448f7c0 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h @@ -89,9 +89,6 @@ class TensorLayoutSwapOp : public TensorBase, WriteA EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorLayoutSwapOp& operator = (const OtherDerived& other) { - -std::cout << "In assignment operator " << std::endl; - typedef TensorAssignOp Assign; Assign assign(*this, other); internal::TensorExecutor::run(assign, DefaultDevice()); From 4716c2c6666eb7018dac2e2ed050ead45c8933e1 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 12:06:19 -0800 Subject: [PATCH 1074/1103] Fixed compilation error --- unsupported/test/cxx11_tensor_thread_pool.cpp | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/unsupported/test/cxx11_tensor_thread_pool.cpp b/unsupported/test/cxx11_tensor_thread_pool.cpp index f49523683..6fe65c7f9 100644 --- a/unsupported/test/cxx11_tensor_thread_pool.cpp +++ b/unsupported/test/cxx11_tensor_thread_pool.cpp @@ -15,7 +15,7 @@ #include using Eigen::Tensor; -using std::isnan; + static void test_multithread_elementwise() { @@ -122,7 +122,7 @@ static void test_contraction_corner_cases() m_result = m_left.transpose() * m_right; for (ptrdiff_t i = 0; i < t_result.size(); i++) { - assert(!isnan(t_result.data()[i])); + assert(!std::isnan(t_result.data()[i])); if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { std::cout << "mismatch detected at index " << i << " : " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; assert(false); @@ -137,7 +137,7 @@ static void test_contraction_corner_cases() new(&m_left) MapXf(t_left.data(), 32, 1); m_result = m_left.transpose() * m_right; for (ptrdiff_t i = 0; i < t_result.size(); i++) { - assert(!isnan(t_result.data()[i])); + assert(!std::isnan(t_result.data()[i])); if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; assert(false); @@ -155,7 +155,7 @@ static void test_contraction_corner_cases() new(&m_right) MapXf(t_right.data(), 32, 4); m_result = m_left.transpose() * m_right; for (ptrdiff_t i = 0; i < t_result.size(); i++) { - assert(!isnan(t_result.data()[i])); + assert(!std::isnan(t_result.data()[i])); if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; assert(false); @@ -173,7 +173,7 @@ static void test_contraction_corner_cases() new(&m_right) MapXf(t_right.data(), 32, 4); m_result = m_left.transpose() * m_right; for (ptrdiff_t i = 0; i < t_result.size(); i++) { - assert(!isnan(t_result.data()[i])); + assert(!std::isnan(t_result.data()[i])); if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) { std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl; assert(false); From 410895a7e4276fa2e1f78dbb953c7045818a86ae Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 12:13:19 -0800 Subject: [PATCH 1075/1103] Silenced several compilation warnings --- .../Eigen/CXX11/src/Tensor/TensorAssign.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorChipping.h | 16 ++++++++-------- .../Eigen/CXX11/src/Tensor/TensorContraction.h | 18 +++++++++--------- .../src/Tensor/TensorContractionThreadPool.h | 4 ++-- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 4 ++-- .../Eigen/CXX11/src/Tensor/TensorLayoutSwap.h | 4 ++-- 6 files changed, 24 insertions(+), 24 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h index 93938bd1b..a4f73b2a1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h @@ -103,7 +103,7 @@ struct TensorEvaluator, Device> m_leftImpl(op.lhsExpression(), device), m_rightImpl(op.rhsExpression(), device) { - EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == static_cast(TensorEvaluator::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE); // The dimensions of the lhs and the rhs tensors should be equal to prevent // overflows and ensure the result is fully initialized. eigen_assert(dimensions_match(m_leftImpl.dimensions(), m_leftImpl.dimensions())); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h index 503803d23..698bcfe18 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h @@ -257,13 +257,13 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const { Index inputIndex; - if ((Layout == ColMajor && m_dim.actualDim() == 0) || - (Layout == RowMajor && m_dim.actualDim() == NumInputDims-1)) { + if ((static_cast(Layout) == static_cast(ColMajor) && m_dim.actualDim() == 0) || + (static_cast(Layout) == static_cast(RowMajor) && m_dim.actualDim() == NumInputDims-1)) { // m_stride is equal to 1, so let's avoid the integer division. eigen_assert(m_stride == 1); inputIndex = index * m_inputStride + m_inputOffset; - } else if ((Layout == ColMajor && m_dim.actualDim() == NumInputDims-1) || - (Layout == RowMajor && m_dim.actualDim() == 0)) { + } else if ((static_cast(Layout) == static_cast(ColMajor) && m_dim.actualDim() == NumInputDims-1) || + (static_cast(Layout) == static_cast(RowMajor) && m_dim.actualDim() == 0)) { // m_stride is aways greater than index, so let's avoid the integer division. eigen_assert(m_stride > index); inputIndex = index + m_inputOffset; @@ -322,8 +322,8 @@ struct TensorEvaluator, Device> static const int packetSize = internal::unpacket_traits::size; EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) - if ((this->Layout == ColMajor && this->m_dim.actualDim() == 0) || - (this->Layout == RowMajor && this->m_dim.actualDim() == NumInputDims-1)) { + if ((static_cast(this->Layout) == static_cast(ColMajor) && this->m_dim.actualDim() == 0) || + (static_cast(this->Layout) == static_cast(RowMajor) && this->m_dim.actualDim() == NumInputDims-1)) { // m_stride is equal to 1, so let's avoid the integer division. eigen_assert(this->m_stride == 1); EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; @@ -333,8 +333,8 @@ struct TensorEvaluator, Device> this->m_impl.coeffRef(inputIndex) = values[i]; inputIndex += this->m_inputStride; } - } else if ((this->Layout == ColMajor && this->m_dim.actualDim() == NumInputDims-1) || - (this->Layout == RowMajor && this->m_dim.actualDim() == 0)) { + } else if ((static_cast(this->Layout) == static_cast(ColMajor) && this->m_dim.actualDim() == NumInputDims-1) || + (static_cast(this->Layout) == static_cast(RowMajor) && this->m_dim.actualDim() == 0)) { // m_stride is aways greater than index, so let's avoid the integer division. eigen_assert(this->m_stride > index); this->m_impl.template writePacket(index + this->m_inputOffset, x); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index af843654c..e750c21e7 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -499,9 +499,9 @@ struct TensorContractionEvaluatorBase // If we want to compute A * B = C, where A is LHS and B is RHS, the code // will pretend B is LHS and A is RHS. typedef typename internal::conditional< - Layout == ColMajor, LeftArgType, RightArgType>::type EvalLeftArgType; + static_cast(Layout) == static_cast(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType; typedef typename internal::conditional< - Layout == ColMajor, RightArgType, LeftArgType>::type EvalRightArgType; + static_cast(Layout) == static_cast(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType; static const int LDims = internal::array_size::Dimensions>::value; @@ -520,14 +520,14 @@ struct TensorContractionEvaluatorBase EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionEvaluatorBase(const XprType& op, const Device& device) - : m_leftImpl(choose(Cond(), + : m_leftImpl(choose(Cond(Layout) == static_cast(ColMajor)>(), op.lhsExpression(), op.rhsExpression()), device), - m_rightImpl(choose(Cond(), + m_rightImpl(choose(Cond(Layout) == static_cast(ColMajor)>(), op.rhsExpression(), op.lhsExpression()), device), m_device(device), m_result(NULL) { - EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == - TensorEvaluator::Layout), + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == + static_cast(TensorEvaluator::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE); eigen_assert((internal::array_size::value > 0) && "Must contract on some indices"); @@ -681,7 +681,7 @@ struct TensorContractionEvaluatorBase } // If the layout is RowMajor, we need to reverse the m_dimensions - if (Layout == RowMajor) { + if (static_cast(Layout) == static_cast(RowMajor)) { for (int i = 0, j = NumDims - 1; i < j; i++, j--) { std::swap(m_dimensions[i], m_dimensions[j]); } @@ -855,9 +855,9 @@ struct TensorEvaluator::type EvalLeftArgType; + static_cast(Layout) == static_cast(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType; typedef typename internal::conditional< - Layout == ColMajor, RightArgType, LeftArgType>::type EvalRightArgType; + static_cast(Layout) == static_cast(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType; static const int LDims = internal::array_size::Dimensions>::value; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h index e358e6a3a..8b87f1045 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h @@ -79,9 +79,9 @@ struct TensorEvaluator::type EvalLeftArgType; + static_cast(Layout) == static_cast(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType; typedef typename internal::conditional< - Layout == ColMajor, RightArgType, LeftArgType>::type EvalRightArgType; + static_cast(Layout) == static_cast(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType; static const int LDims = internal::array_size::Dimensions>::value; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index 97f225f0a..5e167d4aa 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -94,14 +94,14 @@ struct TensorEvaluator EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(const array& coords) { eigen_assert(m_data); - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { return m_data[m_dims.IndexOfColMajor(coords)]; } else { return m_data[m_dims.IndexOfRowMajor(coords)]; } } - Scalar* data() const { return m_data; } + EIGEN_DEVICE_FUNC Scalar* data() const { return m_data; } protected: Scalar* m_data; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h index 7e448f7c0..c00810594 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h @@ -112,7 +112,7 @@ struct TensorEvaluator, Device> enum { IsAligned = TensorEvaluator::IsAligned, PacketAccess = TensorEvaluator::PacketAccess, - Layout = (TensorEvaluator::Layout == ColMajor) ? RowMajor : ColMajor, + Layout = (static_cast(TensorEvaluator::Layout) == static_cast(ColMajor)) ? RowMajor : ColMajor, CoordAccess = false, // to be implemented }; @@ -169,7 +169,7 @@ template enum { IsAligned = TensorEvaluator::IsAligned, PacketAccess = TensorEvaluator::PacketAccess, - Layout = (TensorEvaluator::Layout == ColMajor) ? RowMajor : ColMajor, + Layout = (static_cast(TensorEvaluator::Layout) == static_cast(ColMajor)) ? RowMajor : ColMajor, CoordAccess = false, // to be implemented }; From 114e863f086077fc949baf5dfe1f4102222c938e Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 12:20:24 -0800 Subject: [PATCH 1076/1103] Silcenced a few compilation warnings --- .../Eigen/CXX11/src/Tensor/TensorBroadcasting.h | 8 ++++---- .../Eigen/CXX11/src/Tensor/TensorContraction.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorEvaluator.h | 12 ++++++------ .../Eigen/CXX11/src/Tensor/TensorReduction.h | 14 +++++++------- 4 files changed, 18 insertions(+), 18 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h index ef134adf2..5790e19d6 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h @@ -106,7 +106,7 @@ struct TensorEvaluator, Device> m_dimensions[i] = input_dims[i] * broadcast[i]; } - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { m_inputStrides[0] = 1; m_outputStrides[0] = 1; for (int i = 1; i < NumDims; ++i) { @@ -139,7 +139,7 @@ struct TensorEvaluator, Device> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffReturnType coeff(Index index) const { - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { return coeffColMajor(index); } else { return coeffRowMajor(index); @@ -210,7 +210,7 @@ struct TensorEvaluator, Device> template EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnType packet(Index index) const { - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { return packetColMajor(index); } else { return packetRowMajor(index); @@ -326,7 +326,7 @@ struct TensorEvaluator, Device> } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } protected: Dimensions m_dimensions; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h index e750c21e7..f7254a24d 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h @@ -536,7 +536,7 @@ struct TensorContractionEvaluatorBase DSizes eval_left_dims; DSizes eval_right_dims; array, ContractDims> eval_op_indices; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { // For ColMajor, we keep using the existing dimensions for (int i = 0; i < LDims; i++) { eval_left_dims[i] = m_leftImpl.dimensions()[i]; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index 5e167d4aa..488d32cb4 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -167,7 +167,7 @@ struct TensorEvaluator #endif } - const Scalar* data() const { return m_data; } + EIGEN_DEVICE_FUNC const Scalar* data() const { return m_data; } protected: const Scalar* m_data; @@ -218,7 +218,7 @@ struct TensorEvaluator, Device> return m_functor.packetOp(index); } - CoeffReturnType* data() const { return NULL; } + EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; } private: const NullaryOp m_functor; @@ -273,7 +273,7 @@ struct TensorEvaluator, Device> return m_functor.packetOp(m_argImpl.template packet(index)); } - CoeffReturnType* data() const { return NULL; } + EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; } private: const UnaryOp m_functor; @@ -301,7 +301,7 @@ struct TensorEvaluator::Layout == TensorEvaluator::Layout || internal::traits::NumDimensions == 1), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == static_cast(TensorEvaluator::Layout) || internal::traits::NumDimensions == 1), YOU_MADE_A_PROGRAMMING_MISTAKE); eigen_assert(dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions())); } @@ -337,7 +337,7 @@ struct TensorEvaluator(index), m_rightImpl.template packet(index)); } - CoeffReturnType* data() const { return NULL; } + EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; } private: const BinaryOp m_functor; @@ -413,7 +413,7 @@ struct TensorEvaluator m_elseImpl.template packet(index)); } - CoeffReturnType* data() const { return NULL; } + EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; } private: TensorEvaluator m_condImpl; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index 21416afe0..7643d4cdc 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -245,7 +245,7 @@ struct TensorEvaluator, Device> } // Precompute output strides. - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { m_outputStrides[0] = 1; for (int i = 1; i < NumOutputDims; ++i) { m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1]; @@ -259,7 +259,7 @@ struct TensorEvaluator, Device> // Precompute input strides. array input_strides; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { input_strides[0] = 1; for (int i = 1; i < NumInputDims; ++i) { input_strides[i] = input_strides[i-1] * input_dims[i-1]; @@ -309,7 +309,7 @@ struct TensorEvaluator, Device> Op reducer(m_reducer); if (ReducingInnerMostDims) { const Index num_values_to_reduce = - (Layout == ColMajor) ? m_preservedStrides[0] : m_preservedStrides[NumOutputDims - 1]; + (static_cast(Layout) == static_cast(ColMajor)) ? m_preservedStrides[0] : m_preservedStrides[NumOutputDims - 1]; return internal::InnerMostDimReducer::reduce(*this, firstInput(index), num_values_to_reduce, reducer); } else { @@ -330,7 +330,7 @@ struct TensorEvaluator, Device> EIGEN_ALIGN_DEFAULT typename internal::remove_const::type values[packetSize]; if (ReducingInnerMostDims) { const Index num_values_to_reduce = - (Layout == ColMajor) ? m_preservedStrides[0] : m_preservedStrides[NumOutputDims - 1]; + (static_cast(Layout) == static_cast(ColMajor)) ? m_preservedStrides[0] : m_preservedStrides[NumOutputDims - 1]; const Index firstIndex = firstInput(index); for (Index i = 0; i < packetSize; ++i) { Op reducer(m_reducer); @@ -339,7 +339,7 @@ struct TensorEvaluator, Device> } } else if (PreservingInnerMostDims) { const Index firstIndex = firstInput(index); - const int innermost_dim = (Layout == ColMajor) ? 0 : NumOutputDims - 1; + const int innermost_dim = (static_cast(Layout) == static_cast(ColMajor)) ? 0 : NumOutputDims - 1; // TBD: extend this the the n innermost dimensions that we preserve. if (((firstIndex % m_dimensions[innermost_dim]) + packetSize - 1) < m_dimensions[innermost_dim]) { Op reducer(m_reducer); @@ -371,7 +371,7 @@ struct TensorEvaluator, Device> // used to compute the reduction at output index "index". EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index firstInput(Index index) const { if (ReducingInnerMostDims) { - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { return index * m_preservedStrides[0]; } else { return index * m_preservedStrides[NumOutputDims - 1]; @@ -379,7 +379,7 @@ struct TensorEvaluator, Device> } // TBD: optimize the case where we preserve the innermost dimensions. Index startInput = 0; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumOutputDims - 1; i > 0; --i) { // This is index_i in the output tensor. const Index idx = index / m_outputStrides[i]; From 057cfd2f02f06650db0634aca6abfbd09da36897 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 12:25:02 -0800 Subject: [PATCH 1077/1103] Silenced more compilation warnings --- unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h | 8 ++++---- unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h | 12 ++++++------ unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h | 4 ++-- 3 files changed, 12 insertions(+), 12 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h index 488d32cb4..d084880de 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h @@ -85,7 +85,7 @@ struct TensorEvaluator EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array& coords) const { eigen_assert(m_data); - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { return m_data[m_dims.IndexOfColMajor(coords)]; } else { return m_data[m_dims.IndexOfRowMajor(coords)]; @@ -158,7 +158,7 @@ struct TensorEvaluator EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array& coords) const { eigen_assert(m_data); - const Index index = (Layout == ColMajor) ? m_dims.IndexOfColMajor(coords) + const Index index = (static_cast(Layout) == static_cast(ColMajor)) ? m_dims.IndexOfColMajor(coords) : m_dims.IndexOfRowMajor(coords); #ifdef __CUDA_ARCH__ return __ldg(m_data+index); @@ -366,8 +366,8 @@ struct TensorEvaluator m_thenImpl(op.thenExpression(), device), m_elseImpl(op.elseExpression(), device) { - EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); - EIGEN_STATIC_ASSERT((TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == static_cast(TensorEvaluator::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == static_cast(TensorEvaluator::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE); eigen_assert(dimensions_match(m_condImpl.dimensions(), m_thenImpl.dimensions())); eigen_assert(dimensions_match(m_thenImpl.dimensions(), m_elseImpl.dimensions())); } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 87a4b0758..1191b2411 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -308,7 +308,7 @@ struct TensorEvaluator, Devi const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); const Sizes& output_dims = op.sizes(); - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { m_inputStrides[0] = 1; for (int i = 1; i < NumDims; ++i) { m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1]; @@ -348,7 +348,7 @@ struct TensorEvaluator, Devi m_impl.evalSubExprsIfNeeded(NULL); if (internal::is_arithmetic::value && data && m_impl.data()) { Index contiguous_values = 1; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = 0; i < NumDims; ++i) { contiguous_values *= dimensions()[i]; if (dimensions()[i] != m_impl.dimensions()[i]) { @@ -394,7 +394,7 @@ struct TensorEvaluator, Devi Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx0 = indices[0] / m_fastOutputStrides[i]; const Index idx1 = indices[1] / m_fastOutputStrides[i]; @@ -446,7 +446,7 @@ struct TensorEvaluator, Devi Scalar* result = m_impl.data(); if (result) { Index offset = 0; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = 0; i < NumDims; ++i) { if (m_dimensions[i] != m_impl.dimensions()[i]) { offset += m_offsets[i] * m_inputStrides[i]; @@ -482,7 +482,7 @@ struct TensorEvaluator, Devi EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const { Index inputIndex = 0; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_fastOutputStrides[i]; inputIndex += (idx + m_offsets[i]) * m_inputStrides[i]; @@ -547,7 +547,7 @@ struct TensorEvaluator, Device> const int packetSize = internal::unpacket_traits::size; Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx0 = indices[0] / this->m_fastOutputStrides[i]; const Index idx1 = indices[1] / this->m_fastOutputStrides[i]; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h index 439cf3230..82969b4c0 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h @@ -106,7 +106,7 @@ struct TensorEvaluator, Device { // Compute strides m_dimensions = m_impl.dimensions(); - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { m_strides[0] = 1; for (int i = 1; i < NumDims; ++i) { m_strides[i] = m_strides[i-1] * m_dimensions[i-1]; @@ -138,7 +138,7 @@ struct TensorEvaluator, Device { eigen_assert(index < dimensions().TotalSize()); Index inputIndex = 0; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { Index idx = index / m_strides[i]; index -= idx * m_strides[i]; From c21e45fbc5b82a2f99113e8b6ab0005ca01a7428 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 12:36:26 -0800 Subject: [PATCH 1078/1103] Fixed a few more compilation warnings --- .../Eigen/CXX11/src/Tensor/TensorConcatenation.h | 10 +++++----- unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h | 6 +++--- unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h | 10 +++++----- 3 files changed, 13 insertions(+), 13 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h index fb4e7fb11..57a14a037 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h @@ -111,7 +111,7 @@ struct TensorEvaluator::Layout == TensorEvaluator::Layout || NumDims == 1), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == static_cast(TensorEvaluator::Layout) || NumDims == 1), YOU_MADE_A_PROGRAMMING_MISTAKE); EIGEN_STATIC_ASSERT(NumDims == RightNumDims, YOU_MADE_A_PROGRAMMING_MISTAKE) eigen_assert(0 <= m_axis && m_axis < NumDims); const Dimensions& lhs_dims = m_leftImpl.dimensions(); @@ -131,7 +131,7 @@ struct TensorEvaluator(Layout) == static_cast(ColMajor)) { m_leftStrides[0] = 1; m_rightStrides[0] = 1; m_outputStrides[0] = 1; @@ -176,7 +176,7 @@ struct TensorEvaluator subs; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { subs[i] = index / m_outputStrides[i]; index -= subs[i] * m_outputStrides[i]; @@ -193,7 +193,7 @@ struct TensorEvaluator(Layout) == static_cast(ColMajor)) { left_index = subs[0]; for (int i = 1; i < NumDims; ++i) { left_index += (subs[i] % left_dims[i]) * m_leftStrides[i]; @@ -209,7 +209,7 @@ struct TensorEvaluator(Layout) == static_cast(ColMajor)) { right_index = subs[0]; for (int i = 1; i < NumDims; ++i) { right_index += (subs[i] % right_dims[i]) * m_rightStrides[i]; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h index 620a63ae7..1012ecd69 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h @@ -126,7 +126,7 @@ struct TensorEvaluator, Device> array inputStrides; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { inputStrides[0] = 1; m_outputStrides[0] = 1; for (int i = 1; i < NumDims; ++i) { @@ -180,12 +180,12 @@ struct TensorEvaluator, Device> return rslt; } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } protected: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const { Index inputIndex = 0; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; inputIndex += idx * m_inputStrides[i]; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h index 5aa2c8d3b..00cb8e373 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h @@ -123,7 +123,7 @@ struct TensorEvaluator, Device> } const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { m_outputStrides[0] = 1; m_inputStrides[0] = 1; for (int i = 1; i < NumDims; ++i) { @@ -172,7 +172,7 @@ struct TensorEvaluator, Device> Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx0 = indices[0] / m_outputStrides[i]; const Index idx1 = indices[1] / m_outputStrides[i]; @@ -211,13 +211,13 @@ struct TensorEvaluator, Device> } } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } protected: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const { Index inputIndex = 0; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; inputIndex += idx * m_inputStrides[i]; @@ -281,7 +281,7 @@ struct TensorEvaluator, Device> Index inputIndices[] = {0, 0}; Index indices[] = {index, index + packetSize - 1}; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx0 = indices[0] / this->m_outputStrides[i]; const Index idx1 = indices[1] / this->m_outputStrides[i]; From 780b2422e2b3fd2b50121a6e5642c94b030fbf5b Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 12:43:55 -0800 Subject: [PATCH 1079/1103] Silenced the last batch of compilation warnings triggered by gcc 4.8 --- unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h | 10 +++++----- unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h | 10 +++++----- unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h | 4 ++-- unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h | 10 +++++----- unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h | 4 ++-- 5 files changed, 19 insertions(+), 19 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h index 698bcfe18..dc9586cbc 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h @@ -167,7 +167,7 @@ struct TensorEvaluator, Device> m_stride = 1; m_inputStride = 1; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = 0; i < m_dim.actualDim(); ++i) { m_stride *= input_dims[i]; m_inputStride *= input_dims[i]; @@ -208,8 +208,8 @@ struct TensorEvaluator, Device> EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE) eigen_assert(index+packetSize-1 < dimensions().TotalSize()); - if ((Layout == ColMajor && m_dim.actualDim() == 0) || - (Layout == RowMajor && m_dim.actualDim() == NumInputDims-1)) { + if ((static_cast(Layout) == static_cast(ColMajor) && m_dim.actualDim() == 0) || + (static_cast(Layout) == static_cast(RowMajor) && m_dim.actualDim() == NumInputDims-1)) { // m_stride is equal to 1, so let's avoid the integer division. eigen_assert(m_stride == 1); Index inputIndex = index * m_inputStride + m_inputOffset; @@ -220,8 +220,8 @@ struct TensorEvaluator, Device> } PacketReturnType rslt = internal::pload(values); return rslt; - } else if ((Layout == ColMajor && m_dim.actualDim() == NumInputDims - 1) || - (Layout == RowMajor && m_dim.actualDim() == 0)) { + } else if ((static_cast(Layout) == static_cast(ColMajor) && m_dim.actualDim() == NumInputDims - 1) || + (static_cast(Layout) == static_cast(RowMajor) && m_dim.actualDim() == 0)) { // m_stride is aways greater than index, so let's avoid the integer division. eigen_assert(m_stride > index); return m_impl.template packet(index + m_inputOffset); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h index aecef3313..591fd2464 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h @@ -236,9 +236,9 @@ struct TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == static_cast(TensorEvaluator::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE); // Only column major tensors are supported for now. - EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(Layout) == static_cast(ColMajor)), YOU_MADE_A_PROGRAMMING_MISTAKE); const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); @@ -339,7 +339,7 @@ struct TensorEvaluator::Layout == TensorEvaluator::Layout), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(TensorEvaluator::Layout) == static_cast(TensorEvaluator::Layout)), YOU_MADE_A_PROGRAMMING_MISTAKE); // Only column major tensors are supported for now. - EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(Layout) == static_cast(ColMajor)), YOU_MADE_A_PROGRAMMING_MISTAKE); const typename TensorEvaluator::Dimensions& input_dims = m_inputImpl.dimensions(); const typename TensorEvaluator::Dimensions& kernel_dims = m_kernelImpl.dimensions(); diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h index 585ebc778..bf0e7edfb 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h @@ -121,7 +121,7 @@ struct TensorEvaluator, Device> : m_impl(op.expression(), device) { // Only column major tensors are supported for now. - EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(Layout) == static_cast(ColMajor)), YOU_MADE_A_PROGRAMMING_MISTAKE); EIGEN_STATIC_ASSERT(NumDims >= 4, YOU_MADE_A_PROGRAMMING_MISTAKE); @@ -295,7 +295,7 @@ struct TensorEvaluator, Device> return packetWithPossibleZero(index); } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } const TensorEvaluator& impl() const { return m_impl; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h index 9b14e01f4..2a7dd45c0 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h @@ -104,7 +104,7 @@ struct TensorEvaluator, Device m_dimensions[i] += m_padding[i].first + m_padding[i].second; } const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { m_inputStrides[0] = 1; m_outputStrides[0] = 1; for (int i = 1; i < NumDims; ++i) { @@ -141,7 +141,7 @@ struct TensorEvaluator, Device { eigen_assert(index < dimensions().TotalSize()); Index inputIndex = 0; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { for (int i = NumDims - 1; i > 0; --i) { const Index idx = index / m_outputStrides[i]; if (idx < m_padding[i].first || idx >= m_dimensions[i] - m_padding[i].second) { @@ -175,7 +175,7 @@ struct TensorEvaluator, Device template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const { - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { return packetColMajor(index); } return packetRowMajor(index); @@ -184,7 +184,7 @@ struct TensorEvaluator, Device EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array& coords) const { Index inputIndex; - if (Layout == ColMajor) { + if (static_cast(Layout) == static_cast(ColMajor)) { const Index idx = coords[0]; if (idx < m_padding[0].first || idx >= m_dimensions[0] - m_padding[0].second) { return Scalar(0); @@ -214,7 +214,7 @@ struct TensorEvaluator, Device return m_impl.coeff(inputIndex); } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } protected: diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h index 1c03d202f..8a42ab6b1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h @@ -100,7 +100,7 @@ struct TensorEvaluator, Device> : m_impl(op.expression(), device) { // Only column major tensors are supported for now. - EIGEN_STATIC_ASSERT((Layout == ColMajor), YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT((static_cast(Layout) == static_cast(ColMajor)), YOU_MADE_A_PROGRAMMING_MISTAKE); Index num_patches = 1; const typename TensorEvaluator::Dimensions& input_dims = m_impl.dimensions(); @@ -232,7 +232,7 @@ struct TensorEvaluator, Device> } } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } protected: Dimensions m_dimensions; From fefec723aa44703c1b7884b2ccfa73877a58f500 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 13:16:22 -0800 Subject: [PATCH 1080/1103] Fixed compilation error triggered when trying to vectorize a non vectorizable cuda kernel. --- .../Eigen/CXX11/src/Tensor/TensorExecutor.h | 84 +++++++++++++------ 1 file changed, 60 insertions(+), 24 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index d93fdd907..05ac9bd2f 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -22,8 +22,13 @@ namespace Eigen { */ namespace internal { +template +struct IsVectorizable { + static const bool value = TensorEvaluator::PacketAccess; +}; + // Default strategy: the expression is evaluated with a single cpu thread. -template::PacketAccess> +template::value> class TensorExecutor { public: @@ -153,34 +158,45 @@ class TensorExecutor template __global__ void __launch_bounds__(1024) - EigenMetaKernel(Evaluator eval, Index size) { +EigenMetaKernel_NonVectorizable(Evaluator eval, Index size) { const Index first_index = blockIdx.x * blockDim.x + threadIdx.x; const Index step_size = blockDim.x * gridDim.x; - if (!Evaluator::PacketAccess || !Evaluator::IsAligned) { - // Use the scalar path - for (Index i = first_index; i < size; i += step_size) { - eval.evalScalar(i); - } - } - else { - // Use the vector path - const Index PacketSize = unpacket_traits::size; - const Index vectorized_step_size = step_size * PacketSize; - const Index vectorized_size = (size / PacketSize) * PacketSize; - for (Index i = first_index * PacketSize; i < vectorized_size; - i += vectorized_step_size) { - eval.evalPacket(i); - } - for (Index i = vectorized_size + first_index; i < size; i += step_size) { - eval.evalScalar(i); - } + // Use the scalar path + for (Index i = first_index; i < size; i += step_size) { + eval.evalScalar(i); } } -template -class TensorExecutor +template +__global__ void +__launch_bounds__(1024) +EigenMetaKernel_Vectorizable(Evaluator eval, Index size) { + + const Index first_index = blockIdx.x * blockDim.x + threadIdx.x; + const Index step_size = blockDim.x * gridDim.x; + + // Use the vector path + const Index PacketSize = unpacket_traits::size; + const Index vectorized_step_size = step_size * PacketSize; + const Index vectorized_size = (size / PacketSize) * PacketSize; + for (Index i = first_index * PacketSize; i < vectorized_size; + i += vectorized_step_size) { + eval.evalPacket(i); + } + for (Index i = vectorized_size + first_index; i < size; i += step_size) { + eval.evalScalar(i); + } +} + +template +struct IsVectorizable { + static const bool value = TensorEvaluator::PacketAccess && TensorEvaluator::IsAligned; +}; + +template +class TensorExecutor { public: typedef typename Expression::Index Index; @@ -192,13 +208,33 @@ class TensorExecutor { const int num_blocks = getNumCudaMultiProcessors() * maxCudaThreadsPerMultiProcessor() / maxCudaThreadsPerBlock(); const int block_size = maxCudaThreadsPerBlock(); - const Index size = array_prod(evaluator.dimensions()); - LAUNCH_CUDA_KERNEL((EigenMetaKernel, Index>), num_blocks, block_size, 0, device, evaluator, size); + LAUNCH_CUDA_KERNEL((EigenMetaKernel_NonVectorizable, Index>), num_blocks, block_size, 0, device, evaluator, size); } evaluator.cleanup(); } }; + +template +class TensorExecutor +{ + public: + typedef typename Expression::Index Index; + static inline void run(const Expression& expr, const GpuDevice& device) + { + TensorEvaluator evaluator(expr, device); + const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL); + if (needs_assign) + { + const int num_blocks = getNumCudaMultiProcessors() * maxCudaThreadsPerMultiProcessor() / maxCudaThreadsPerBlock(); + const int block_size = maxCudaThreadsPerBlock(); + const Index size = array_prod(evaluator.dimensions()); + LAUNCH_CUDA_KERNEL((EigenMetaKernel_Vectorizable, Index>), num_blocks, block_size, 0, device, evaluator, size); + } + evaluator.cleanup(); + } +}; + #endif } // end namespace internal From d771295554b4d0238f136588ae526a4c23be9c6c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 10 Feb 2015 22:59:27 +0100 Subject: [PATCH 1081/1103] remove useless include --- blas/level3_impl.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/blas/level3_impl.h b/blas/level3_impl.h index 32313e814..a05872666 100644 --- a/blas/level3_impl.h +++ b/blas/level3_impl.h @@ -8,7 +8,7 @@ // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. #include "common.h" -#include + int EIGEN_BLAS_FUNC(gemm)(char *opa, char *opb, int *m, int *n, int *k, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *pb, int *ldb, RealScalar *pbeta, RealScalar *pc, int *ldc) { // std::cerr << "in gemm " << *opa << " " << *opb << " " << *m << " " << *n << " " << *k << " " << *lda << " " << *ldb << " " << *ldc << " " << *palpha << " " << *pbeta << "\n"; From cc5d7ff5238da45ef7416ec94f18227486ed9643 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 14:02:38 -0800 Subject: [PATCH 1082/1103] Added vectorized implementation of the exponential function for ARM/NEON --- Eigen/Core | 1 + Eigen/src/Core/arch/NEON/MathFunctions.h | 91 ++++++++++++++++++++++++ Eigen/src/Core/arch/NEON/PacketMath.h | 2 +- 3 files changed, 93 insertions(+), 1 deletion(-) create mode 100644 Eigen/src/Core/arch/NEON/MathFunctions.h diff --git a/Eigen/Core b/Eigen/Core index b5af63623..416e901ae 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -300,6 +300,7 @@ using std::ptrdiff_t; #include "src/Core/arch/AltiVec/Complex.h" #elif defined EIGEN_VECTORIZE_NEON #include "src/Core/arch/NEON/PacketMath.h" + #include "src/Core/arch/NEON/MathFunctions.h" #include "src/Core/arch/NEON/Complex.h" #endif diff --git a/Eigen/src/Core/arch/NEON/MathFunctions.h b/Eigen/src/Core/arch/NEON/MathFunctions.h new file mode 100644 index 000000000..6bb05bb92 --- /dev/null +++ b/Eigen/src/Core/arch/NEON/MathFunctions.h @@ -0,0 +1,91 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +/* The sin, cos, exp, and log functions of this file come from + * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ + */ + +#ifndef EIGEN_MATH_FUNCTIONS_NEON_H +#define EIGEN_MATH_FUNCTIONS_NEON_H + +namespace Eigen { + +namespace internal { + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f pexp(const Packet4f& _x) +{ + Packet4f x = _x; + Packet4f tmp, fx; + + _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f); + _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f); + _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f); + _EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f); + _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f); + _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f); + + x = vminq_f32(x, p4f_exp_hi); + x = vmaxq_f32(x, p4f_exp_lo); + + /* express exp(x) as exp(g + n*log(2)) */ + fx = vmlaq_f32(p4f_half, x, p4f_cephes_LOG2EF); + + /* perform a floorf */ + tmp = vcvtq_f32_s32(vcvtq_s32_f32(fx)); + + /* if greater, substract 1 */ + Packet4ui mask = vcgtq_f32(tmp, fx); + mask = vandq_u32(mask, vreinterpretq_u32_f32(p4f_1)); + + fx = vsubq_f32(tmp, vreinterpretq_f32_u32(mask)); + + tmp = vmulq_f32(fx, p4f_cephes_exp_C1); + Packet4f z = vmulq_f32(fx, p4f_cephes_exp_C2); + x = vsubq_f32(x, tmp); + x = vsubq_f32(x, z); + + Packet4f y = vmulq_f32(p4f_cephes_exp_p0, x); + z = vmulq_f32(x, x); + y = vaddq_f32(y, p4f_cephes_exp_p1); + y = vmulq_f32(y, x); + y = vaddq_f32(y, p4f_cephes_exp_p2); + y = vmulq_f32(y, x); + y = vaddq_f32(y, p4f_cephes_exp_p3); + y = vmulq_f32(y, x); + y = vaddq_f32(y, p4f_cephes_exp_p4); + y = vmulq_f32(y, x); + y = vaddq_f32(y, p4f_cephes_exp_p5); + + y = vmulq_f32(y, z); + y = vaddq_f32(y, x); + y = vaddq_f32(y, p4f_1); + + /* build 2^n */ + int32x4_t mm; + mm = vcvtq_s32_f32(fx); + mm = vaddq_s32(mm, p4i_0x7f); + mm = vshlq_n_s32(mm, 23); + Packet4f pow2n = vreinterpretq_f32_s32(mm); + + y = vmulq_f32(y, pow2n); + return y; +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_NEON_H diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 9afd86bec..559682cf7 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -88,7 +88,7 @@ template<> struct packet_traits : default_packet_traits HasSin = 0, HasCos = 0, HasLog = 0, - HasExp = 0, + HasExp = 1, HasSqrt = 0 }; }; From fe25f3b8e3ee86d808b5f8f75f789cee8dffd581 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Tue, 10 Feb 2015 23:11:35 +0100 Subject: [PATCH 1083/1103] FMA has been wrongly disabled --- Eigen/src/Core/arch/AVX/PacketMath.h | 2 +- Eigen/src/Core/arch/SSE/PacketMath.h | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 485bac10b..fc2e78b2d 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -133,7 +133,7 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv(const Packet8i& /*a*/, co return pset1(0); } -#ifdef EIGEN_VECTORIZE_FMA +#ifdef __FMA__ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { #if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers, diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 3f6fb0254..63a859699 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -227,7 +227,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv(const Packet4i& /*a*/, co // for some weird raisons, it has to be overloaded for packet of integers template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); } -#ifdef EIGEN_VECTORIZE_FMA +#ifdef __FMA__ template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); } template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); } #endif From f669f5656ab550010c5dd92ce2da7d3fab07babd Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 14:29:47 -0800 Subject: [PATCH 1084/1103] Marked a few functions as EIGEN_DEVICE_FUNC to enable the use of tensors in cuda kernels. --- Eigen/src/Core/util/Memory.h | 48 +++++++++---------- unsupported/Eigen/CXX11/src/Tensor/Tensor.h | 4 +- .../Eigen/CXX11/src/Tensor/TensorDimensions.h | 2 +- .../Eigen/CXX11/src/Tensor/TensorStorage.h | 4 +- 4 files changed, 29 insertions(+), 29 deletions(-) diff --git a/Eigen/src/Core/util/Memory.h b/Eigen/src/Core/util/Memory.h index bacf236fb..16f8cc1b0 100644 --- a/Eigen/src/Core/util/Memory.h +++ b/Eigen/src/Core/util/Memory.h @@ -143,8 +143,8 @@ inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = *** Implementation of generic aligned realloc (when no realloc can be used)*** *****************************************************************************/ -void* aligned_malloc(std::size_t size); -void aligned_free(void *ptr); +EIGEN_DEVICE_FUNC void* aligned_malloc(std::size_t size); +EIGEN_DEVICE_FUNC void aligned_free(void *ptr); /** \internal * \brief Reallocates aligned memory. @@ -185,33 +185,33 @@ inline void* generic_aligned_realloc(void* ptr, size_t size, size_t old_size) *****************************************************************************/ #ifdef EIGEN_NO_MALLOC -inline void check_that_malloc_is_allowed() +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() { eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)"); } #elif defined EIGEN_RUNTIME_NO_MALLOC -inline bool is_malloc_allowed_impl(bool update, bool new_value = false) +EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false) { static bool value = true; if (update == 1) value = new_value; return value; } -inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); } -inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); } -inline void check_that_malloc_is_allowed() +EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); } +EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); } +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() { eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)"); } #else -inline void check_that_malloc_is_allowed() +EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed() {} #endif /** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements. * On allocation error, the returned pointer is null, and std::bad_alloc is thrown. */ -inline void* aligned_malloc(size_t size) +EIGEN_DEVICE_FUNC inline void* aligned_malloc(size_t size) { check_that_malloc_is_allowed(); @@ -237,7 +237,7 @@ inline void* aligned_malloc(size_t size) } /** \internal Frees memory allocated with aligned_malloc. */ -inline void aligned_free(void *ptr) +EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr) { #if !EIGEN_ALIGN std::free(ptr); @@ -298,12 +298,12 @@ inline void* aligned_realloc(void *ptr, size_t new_size, size_t old_size) /** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned. * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown. */ -template inline void* conditional_aligned_malloc(size_t size) +template EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(size_t size) { return aligned_malloc(size); } -template<> inline void* conditional_aligned_malloc(size_t size) +template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(size_t size) { check_that_malloc_is_allowed(); @@ -314,12 +314,12 @@ template<> inline void* conditional_aligned_malloc(size_t size) } /** \internal Frees memory allocated with conditional_aligned_malloc */ -template inline void conditional_aligned_free(void *ptr) +template EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr) { aligned_free(ptr); } -template<> inline void conditional_aligned_free(void *ptr) +template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr) { std::free(ptr); } @@ -341,7 +341,7 @@ template<> inline void* conditional_aligned_realloc(void* ptr, size_t new /** \internal Destructs the elements of an array. * The \a size parameters tells on how many objects to call the destructor of T. */ -template inline void destruct_elements_of_array(T *ptr, size_t size) +template EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, size_t size) { // always destruct an array starting from the end. if(ptr) @@ -351,7 +351,7 @@ template inline void destruct_elements_of_array(T *ptr, size_t size) /** \internal Constructs the elements of an array. * The \a size parameter tells on how many objects to call the constructor of T. */ -template inline T* construct_elements_of_array(T *ptr, size_t size) +template EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, size_t size) { size_t i; EIGEN_TRY @@ -371,7 +371,7 @@ template inline T* construct_elements_of_array(T *ptr, size_t size) *****************************************************************************/ template -EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size) +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size) { if(size > size_t(-1) / sizeof(T)) throw_std_bad_alloc(); @@ -381,7 +381,7 @@ EIGEN_ALWAYS_INLINE void check_size_for_overflow(size_t size) * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown. * The default constructor of T is called. */ -template inline T* aligned_new(size_t size) +template EIGEN_DEVICE_FUNC inline T* aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(aligned_malloc(sizeof(T)*size)); @@ -396,7 +396,7 @@ template inline T* aligned_new(size_t size) } } -template inline T* conditional_aligned_new(size_t size) +template EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(size_t size) { check_size_for_overflow(size); T *result = reinterpret_cast(conditional_aligned_malloc(sizeof(T)*size)); @@ -414,7 +414,7 @@ template inline T* conditional_aligned_new(size_t size) /** \internal Deletes objects constructed with aligned_new * The \a size parameters tells on how many objects to call the destructor of T. */ -template inline void aligned_delete(T *ptr, size_t size) +template EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, size_t size) { destruct_elements_of_array(ptr, size); aligned_free(ptr); @@ -423,13 +423,13 @@ template inline void aligned_delete(T *ptr, size_t size) /** \internal Deletes objects constructed with conditional_aligned_new * The \a size parameters tells on how many objects to call the destructor of T. */ -template inline void conditional_aligned_delete(T *ptr, size_t size) +template EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, size_t size) { destruct_elements_of_array(ptr, size); conditional_aligned_free(ptr); } -template inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size) +template EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, size_t new_size, size_t old_size) { check_size_for_overflow(new_size); check_size_for_overflow(old_size); @@ -452,7 +452,7 @@ template inline T* conditional_aligned_realloc_new(T* pt } -template inline T* conditional_aligned_new_auto(size_t size) +template EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(size_t size) { if(size==0) return 0; // short-cut. Also fixes Bug 884 @@ -495,7 +495,7 @@ template inline T* conditional_aligned_realloc_new_auto( return result; } -template inline void conditional_aligned_delete_auto(T *ptr, size_t size) +template EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, size_t size) { if(NumTraits::RequireInitialization) destruct_elements_of_array(ptr, size); diff --git a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h index 0e8a4b8d6..037219f23 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/Tensor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/Tensor.h @@ -350,7 +350,7 @@ class Tensor : public TensorBase > } #endif - void resize(const array& dimensions) + EIGEN_DEVICE_FUNC void resize(const array& dimensions) { std::size_t i; Index size = Index(1); @@ -367,7 +367,7 @@ class Tensor : public TensorBase > #endif } - void resize(const DSizes& dimensions) { + EIGEN_DEVICE_FUNC void resize(const DSizes& dimensions) { array dims; for (std::size_t i = 0; i < NumIndices; ++i) { dims[i] = dimensions[i]; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h index d81197e6d..2ad52b2f9 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDimensions.h @@ -275,7 +275,7 @@ struct DSizes : array { } #endif - DSizes& operator = (const array& other) { + EIGEN_DEVICE_FUNC DSizes& operator = (const array& other) { *static_cast(this) = other; return *this; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h index dfe85602a..1b227e8c2 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h @@ -112,9 +112,9 @@ class TensorStorage& dimensions() const {return m_dimensions;} + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const DSizes& dimensions() const {return m_dimensions;} - void resize(DenseIndex size, const array& nbDimensions) + EIGEN_DEVICE_FUNC void resize(DenseIndex size, const array& nbDimensions) { const DenseIndex currentSz = internal::array_prod(m_dimensions); if(size != currentSz) From 6620aaa4b3ad3ae9f38b7b6213e874021579bcd7 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 14:34:42 -0800 Subject: [PATCH 1085/1103] Silenced a few compilation warnings generated by nvcc --- unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h | 2 +- unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h | 2 +- 7 files changed, 7 insertions(+), 7 deletions(-) diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h index 57a14a037..a1dec76d1 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h @@ -240,7 +240,7 @@ struct TensorEvaluator, Device> return internal::ploadt(m_buffer + index); } - CoeffReturnType* data() const { return NULL; } + EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return NULL; } private: TensorEvaluator m_impl; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h index a9501336e..41a36cb75 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h @@ -136,7 +136,7 @@ struct TensorEvaluator, Device> return internal::ploadt(m_buffer + index); } - Scalar* data() const { return m_buffer; } + EIGEN_DEVICE_FUNC Scalar* data() const { return m_buffer; } private: TensorEvaluator m_impl; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h index c00810594..c119b30e2 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h @@ -148,7 +148,7 @@ struct TensorEvaluator, Device> return m_impl.template packet(index); } - CoeffReturnType* data() const { return m_impl.data(); } + EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_impl.data(); } const TensorEvaluator& impl() const { return m_impl; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h index 1191b2411..a93f48ccb 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h @@ -145,7 +145,7 @@ struct TensorEvaluator, Device> return m_impl.template packet(index); } - CoeffReturnType* data() const { return m_impl.data(); } + EIGEN_DEVICE_FUNC CoeffReturnType* data() const { return m_impl.data(); } const TensorEvaluator& impl() const { return m_impl; } diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h index 7643d4cdc..de5747905 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h @@ -360,7 +360,7 @@ struct TensorEvaluator, Device> return rslt; } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } private: template friend struct internal::GenericDimReducer; diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h index 82969b4c0..ad21e966b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReverse.h @@ -190,7 +190,7 @@ struct TensorEvaluator, Device return rslt; } - Scalar* data() const { return NULL; } + EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } protected: Dimensions m_dimensions; From 4470c9997559522e9b81810948d9783b58444ae4 Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 10 Feb 2015 14:40:18 -0800 Subject: [PATCH 1086/1103] Added a test to validate tensor casting on cuda devices --- unsupported/test/cxx11_tensor_cuda.cpp | 40 ++++++++++++++++++++++++++ 1 file changed, 40 insertions(+) diff --git a/unsupported/test/cxx11_tensor_cuda.cpp b/unsupported/test/cxx11_tensor_cuda.cpp index 059d23de1..8c1ca1bf8 100644 --- a/unsupported/test/cxx11_tensor_cuda.cpp +++ b/unsupported/test/cxx11_tensor_cuda.cpp @@ -460,6 +460,45 @@ static void test_cuda_constant_broadcast() } } + +void test_cuda_cast() +{ + Tensor in(Eigen::array(72,53,97)); + Tensor out(Eigen::array(72,53,97)); + in.setRandom(); + + std::size_t in_bytes = in.size() * sizeof(double); + std::size_t out_bytes = out.size() * sizeof(float); + + double* d_in; + float* d_out; + cudaMalloc((void**)(&d_in), in_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_in, in.data(), in_bytes, cudaMemcpyHostToDevice); + + cudaStream_t stream; + assert(cudaStreamCreate(&stream) == cudaSuccess); + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap > gpu_in(d_in, Eigen::array(72,53,97)); + Eigen::TensorMap > gpu_out(d_out, Eigen::array(72,53,97)); + + gpu_out.device(gpu_device) = gpu_in.template cast(); + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 72; ++i) { + for (int j = 0; j < 53; ++j) { + for (int k = 0; k < 97; ++k) { + VERIFY_IS_APPROX(out(Eigen::array(i,j,k)), static_cast(in(Eigen::array(i,j,k)))); + } + } + } +} + + void test_cxx11_tensor_cuda() { CALL_SUBTEST(test_cuda_elementwise_small()); @@ -471,4 +510,5 @@ void test_cxx11_tensor_cuda() CALL_SUBTEST(test_cuda_convolution_2d()); CALL_SUBTEST(test_cuda_convolution_3d()); CALL_SUBTEST(test_cuda_constant_broadcast()); + CALL_SUBTEST(test_cuda_cast()); } From 409547a0c83604b6dea70b8523674ac19e2af958 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Thu, 12 Feb 2015 21:04:31 +0100 Subject: [PATCH 1087/1103] update EIGEN_FAST_MATH documentation --- Eigen/src/Core/util/Macros.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 13f8fdd4e..27167d3dc 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -385,7 +385,7 @@ /** Allows to disable some optimizations which might affect the accuracy of the result. * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them. * They currently include: - * - single precision Cwise::sin() and Cwise::cos() when SSE vectorization is enabled. + * - single precision ArrayBase::sin() and ArrayBase::cos() when SSE vectorization is enabled. */ #ifndef EIGEN_FAST_MATH #define EIGEN_FAST_MATH 1 From fc202bab398ed9b78ef8452f8e4ef35e233965f6 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Fri, 13 Feb 2015 18:57:41 +0100 Subject: [PATCH 1088/1103] Index refactoring: StorageIndex must be used for storage only (and locally when it make sense). In all other cases use the global Index type. --- Eigen/src/Core/BooleanRedux.h | 2 +- Eigen/src/Core/CoreIterators.h | 15 +-- Eigen/src/Core/CwiseBinaryOp.h | 4 +- Eigen/src/Core/DenseBase.h | 15 +-- Eigen/src/Core/Diagonal.h | 24 ++-- Eigen/src/Core/DiagonalMatrix.h | 1 - Eigen/src/Core/EigenBase.h | 19 ++- Eigen/src/Core/Inverse.h | 2 +- Eigen/src/Core/MatrixBase.h | 3 +- Eigen/src/Core/PermutationMatrix.h | 64 +++++----- Eigen/src/Core/Product.h | 5 +- Eigen/src/Core/ProductEvaluators.h | 10 -- Eigen/src/Core/ReturnByValue.h | 4 +- Eigen/src/Core/SelfAdjointView.h | 2 - Eigen/src/Core/Solve.h | 4 +- Eigen/src/Core/Swap.h | 1 - Eigen/src/Core/Transpose.h | 4 +- Eigen/src/Core/TriangularMatrix.h | 12 +- Eigen/src/Geometry/Transform.h | 7 +- Eigen/src/Householder/HouseholderSequence.h | 3 - .../IterativeLinearSolvers/IncompleteLUT.h | 74 ++++++------ Eigen/src/LU/FullPivLU.h | 6 +- Eigen/src/LU/PartialPivLU.h | 2 +- Eigen/src/QR/ColPivHouseholderQR.h | 2 +- Eigen/src/QR/FullPivHouseholderQR.h | 4 +- Eigen/src/QR/HouseholderQR.h | 2 +- Eigen/src/SPQRSupport/SuiteSparseQRSupport.h | 6 +- Eigen/src/SVD/SVDBase.h | 1 - Eigen/src/SparseCholesky/SimplicialCholesky.h | 8 +- .../SparseCholesky/SimplicialCholesky_impl.h | 20 ++-- Eigen/src/SparseCore/AmbiVector.h | 4 +- Eigen/src/SparseCore/CompressedStorage.h | 53 +++++---- Eigen/src/SparseCore/SparseBlock.h | 112 +++++++++--------- Eigen/src/SparseCore/SparseColEtree.h | 8 +- Eigen/src/SparseCore/SparseCwiseBinaryOp.h | 32 +++-- Eigen/src/SparseCore/SparseDenseProduct.h | 11 +- Eigen/src/SparseCore/SparseDiagonalProduct.h | 16 ++- Eigen/src/SparseCore/SparseMatrix.h | 24 ++-- Eigen/src/SparseCore/SparseMatrixBase.h | 12 +- Eigen/src/SparseCore/SparseSelfAdjointView.h | 8 +- .../SparseSparseProductWithPruning.h | 6 +- Eigen/src/SparseCore/SparseTranspose.h | 11 +- Eigen/src/SparseCore/SparseTriangularView.h | 19 ++- Eigen/src/SparseCore/SparseVector.h | 40 +++---- Eigen/src/SparseCore/SparseView.h | 20 ++-- Eigen/src/SparseLU/SparseLU.h | 26 ++-- Eigen/src/SparseLU/SparseLUImpl.h | 8 +- Eigen/src/SparseLU/SparseLU_Memory.h | 15 ++- Eigen/src/SparseLU/SparseLU_Structs.h | 3 +- .../src/SparseLU/SparseLU_SupernodalMatrix.h | 18 +-- Eigen/src/SparseLU/SparseLU_Utils.h | 8 +- Eigen/src/SparseLU/SparseLU_column_bmod.h | 5 +- Eigen/src/SparseLU/SparseLU_column_dfs.h | 16 +-- Eigen/src/SparseLU/SparseLU_copy_to_ucol.h | 5 +- Eigen/src/SparseLU/SparseLU_gemm_kernel.h | 2 +- .../src/SparseLU/SparseLU_heap_relax_snode.h | 6 +- Eigen/src/SparseLU/SparseLU_kernel_bmod.h | 11 +- Eigen/src/SparseLU/SparseLU_panel_bmod.h | 4 +- Eigen/src/SparseLU/SparseLU_panel_dfs.h | 14 +-- Eigen/src/SparseLU/SparseLU_pivotL.h | 6 +- Eigen/src/SparseLU/SparseLU_pruneL.h | 5 +- Eigen/src/SparseLU/SparseLU_relax_snode.h | 4 +- Eigen/src/SparseQR/SparseQR.h | 26 ++-- Eigen/src/SuperLUSupport/SuperLUSupport.h | 6 +- Eigen/src/UmfPackSupport/UmfPackSupport.h | 4 +- test/nullary.cpp | 3 - test/sparse_basic.cpp | 4 +- test/sparse_product.cpp | 6 +- .../Eigen/src/IterativeSolvers/GMRES.h | 1 - .../Eigen/src/IterativeSolvers/MINRES.h | 1 - .../KroneckerProduct/KroneckerTensorProduct.h | 5 +- .../Eigen/src/LevenbergMarquardt/LMcovar.h | 1 - .../Eigen/src/LevenbergMarquardt/LMpar.h | 2 +- .../Eigen/src/LevenbergMarquardt/LMqrsolv.h | 2 - .../LevenbergMarquardt/LevenbergMarquardt.h | 3 +- .../Eigen/src/SparseExtra/BlockSparseMatrix.h | 98 +++++++-------- .../src/SparseExtra/DynamicSparseMatrix.h | 50 ++++---- .../Eigen/src/SparseExtra/RandomSetter.h | 8 +- 78 files changed, 514 insertions(+), 564 deletions(-) diff --git a/Eigen/src/Core/BooleanRedux.h b/Eigen/src/Core/BooleanRedux.h index dac1887e0..31fbb9214 100644 --- a/Eigen/src/Core/BooleanRedux.h +++ b/Eigen/src/Core/BooleanRedux.h @@ -130,7 +130,7 @@ inline bool DenseBase::any() const * \sa all(), any() */ template -inline typename DenseBase::Index DenseBase::count() const +inline Eigen::Index DenseBase::count() const { return derived().template cast().template cast().sum(); } diff --git a/Eigen/src/Core/CoreIterators.h b/Eigen/src/Core/CoreIterators.h index 141eaa2eb..c76bdf68e 100644 --- a/Eigen/src/Core/CoreIterators.h +++ b/Eigen/src/Core/CoreIterators.h @@ -36,7 +36,6 @@ protected: typedef internal::inner_iterator_selector::Kind> IteratorType; typedef typename internal::evaluator::type EvaluatorType; typedef typename internal::traits::Scalar Scalar; - typedef typename internal::traits::StorageIndex StorageIndex; public: /** Construct an iterator over the \a outerId -th row or column of \a xpr */ InnerIterator(const XprType &xpr, const Index &outerId) @@ -50,11 +49,11 @@ public: */ EIGEN_STRONG_INLINE InnerIterator& operator++() { m_iter.operator++(); return *this; } /// \returns the column or row index of the current coefficient. - EIGEN_STRONG_INLINE StorageIndex index() const { return m_iter.index(); } + EIGEN_STRONG_INLINE Index index() const { return m_iter.index(); } /// \returns the row index of the current coefficient. - EIGEN_STRONG_INLINE StorageIndex row() const { return m_iter.row(); } + EIGEN_STRONG_INLINE Index row() const { return m_iter.row(); } /// \returns the column index of the current coefficient. - EIGEN_STRONG_INLINE StorageIndex col() const { return m_iter.col(); } + EIGEN_STRONG_INLINE Index col() const { return m_iter.col(); } /// \returns \c true if the iterator \c *this still references a valid coefficient. EIGEN_STRONG_INLINE operator bool() const { return m_iter; } @@ -77,7 +76,6 @@ class inner_iterator_selector protected: typedef typename evaluator::type EvaluatorType; typedef typename traits::Scalar Scalar; - typedef typename traits::StorageIndex StorageIndex; enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit }; public: @@ -93,9 +91,9 @@ public: EIGEN_STRONG_INLINE inner_iterator_selector& operator++() { m_inner++; return *this; } - EIGEN_STRONG_INLINE StorageIndex index() const { return m_inner; } - inline StorageIndex row() const { return IsRowMajor ? m_outer : index(); } - inline StorageIndex col() const { return IsRowMajor ? index() : m_outer; } + EIGEN_STRONG_INLINE Index index() const { return m_inner; } + inline Index row() const { return IsRowMajor ? m_outer : index(); } + inline Index col() const { return IsRowMajor ? index() : m_outer; } EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; } @@ -115,7 +113,6 @@ class inner_iterator_selector protected: typedef typename evaluator::InnerIterator Base; typedef typename evaluator::type EvaluatorType; - typedef typename traits::StorageIndex StorageIndex; public: EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &/*innerSize*/) diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h index 4d4626279..0c9d9fbf2 100644 --- a/Eigen/src/Core/CwiseBinaryOp.h +++ b/Eigen/src/Core/CwiseBinaryOp.h @@ -111,7 +111,7 @@ class CwiseBinaryOp : } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE StorageIndex rows() const { + EIGEN_STRONG_INLINE Index rows() const { // return the fixed size type if available to enable compile time optimizations if (internal::traits::type>::RowsAtCompileTime==Dynamic) return m_rhs.rows(); @@ -119,7 +119,7 @@ class CwiseBinaryOp : return m_lhs.rows(); } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE StorageIndex cols() const { + EIGEN_STRONG_INLINE Index cols() const { // return the fixed size type if available to enable compile time optimizations if (internal::traits::type>::ColsAtCompileTime==Dynamic) return m_rhs.cols(); diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h index 628fc61fa..322daad8f 100644 --- a/Eigen/src/Core/DenseBase.h +++ b/Eigen/src/Core/DenseBase.h @@ -58,16 +58,10 @@ template class DenseBase typedef typename internal::traits::StorageKind StorageKind; - /** \brief The interface type of indices - * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE. - * \sa \ref TopicPreprocessorDirectives, StorageIndex. - */ - typedef Eigen::Index Index; - /** - * \brief The type used to store indices - * \details This typedef is relevant for types that store multiple indices such as - * PermutationMatrix or Transpositions, otherwise it defaults to Eigen::Index + * \brief The type used to store indices + * \details This typedef is relevant for types that store multiple indices such as + * PermutationMatrix or Transpositions, otherwise it defaults to Eigen::Index * \sa \ref TopicPreprocessorDirectives, Eigen::Index, SparseMatrixBase. */ typedef typename internal::traits::StorageIndex StorageIndex; @@ -76,7 +70,8 @@ template class DenseBase typedef typename internal::packet_traits::type PacketScalar; typedef typename NumTraits::Real RealScalar; - typedef DenseCoeffsBase Base; + typedef internal::special_scalar_op_base::Scalar, + typename NumTraits::Scalar>::Real> Base; using Base::derived; using Base::const_cast_derived; using Base::rows; diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h index 18f061179..2446a18d4 100644 --- a/Eigen/src/Core/Diagonal.h +++ b/Eigen/src/Core/Diagonal.h @@ -70,28 +70,28 @@ template class Diagonal EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal) EIGEN_DEVICE_FUNC - explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(internal::convert_index(a_index)) {} + explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {} EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal) EIGEN_DEVICE_FUNC - inline StorageIndex rows() const + inline Index rows() const { - return m_index.value()<0 ? numext::mini(m_matrix.cols(),m_matrix.rows()+m_index.value()) - : numext::mini(m_matrix.rows(),m_matrix.cols()-m_index.value()); + return m_index.value()<0 ? numext::mini(m_matrix.cols(),m_matrix.rows()+m_index.value()) + : numext::mini(m_matrix.rows(),m_matrix.cols()-m_index.value()); } EIGEN_DEVICE_FUNC - inline StorageIndex cols() const { return 1; } + inline Index cols() const { return 1; } EIGEN_DEVICE_FUNC - inline StorageIndex innerStride() const + inline Index innerStride() const { return m_matrix.outerStride() + 1; } EIGEN_DEVICE_FUNC - inline StorageIndex outerStride() const + inline Index outerStride() const { return 0; } @@ -153,23 +153,23 @@ template class Diagonal } EIGEN_DEVICE_FUNC - inline StorageIndex index() const + inline Index index() const { return m_index.value(); } protected: typename MatrixType::Nested m_matrix; - const internal::variable_if_dynamicindex m_index; + const internal::variable_if_dynamicindex m_index; private: // some compilers may fail to optimize std::max etc in case of compile-time constants... EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE StorageIndex absDiagIndex() const { return m_index.value()>0 ? m_index.value() : -m_index.value(); } + EIGEN_STRONG_INLINE Index absDiagIndex() const { return m_index.value()>0 ? m_index.value() : -m_index.value(); } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE StorageIndex rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); } + EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); } EIGEN_DEVICE_FUNC - EIGEN_STRONG_INLINE StorageIndex colOffset() const { return m_index.value()>0 ? m_index.value() : 0; } + EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; } // trigger a compile time error is someone try to call packet template typename MatrixType::PacketReturnType packet(Index) const; template typename MatrixType::PacketReturnType packet(Index,Index) const; diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h index 56beaf3bc..5a9e3abd4 100644 --- a/Eigen/src/Core/DiagonalMatrix.h +++ b/Eigen/src/Core/DiagonalMatrix.h @@ -108,7 +108,6 @@ struct traits > { typedef Matrix<_Scalar,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1> DiagonalVectorType; typedef DiagonalShape StorageKind; -// typedef DenseIndex Index; enum { Flags = LvalueBit | NoPreferredStorageOrderBit }; diff --git a/Eigen/src/Core/EigenBase.h b/Eigen/src/Core/EigenBase.h index c98ca467a..79dabda37 100644 --- a/Eigen/src/Core/EigenBase.h +++ b/Eigen/src/Core/EigenBase.h @@ -13,7 +13,9 @@ namespace Eigen { -/** Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T). +/** \class EigenBase + * + * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T). * * In other words, an EigenBase object is an object that can be copied into a MatrixBase. * @@ -26,9 +28,16 @@ namespace Eigen { template struct EigenBase { // typedef typename internal::plain_matrix_type::type PlainObject; + + /** \brief The interface type of indices + * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE. + * \deprecated Since Eigen 3.3, its usage is deprecated. Use Eigen::Index instead. + * \sa StorageIndex, \ref TopicPreprocessorDirectives. + */ + typedef Eigen::Index Index; + // FIXME is it needed? typedef typename internal::traits::StorageKind StorageKind; - typedef typename internal::traits::StorageIndex StorageIndex; /** \returns a reference to the derived object */ EIGEN_DEVICE_FUNC @@ -46,14 +55,14 @@ template struct EigenBase /** \returns the number of rows. \sa cols(), RowsAtCompileTime */ EIGEN_DEVICE_FUNC - inline StorageIndex rows() const { return derived().rows(); } + inline Index rows() const { return derived().rows(); } /** \returns the number of columns. \sa rows(), ColsAtCompileTime*/ EIGEN_DEVICE_FUNC - inline StorageIndex cols() const { return derived().cols(); } + inline Index cols() const { return derived().cols(); } /** \returns the number of coefficients, which is rows()*cols(). * \sa rows(), cols(), SizeAtCompileTime. */ EIGEN_DEVICE_FUNC - inline StorageIndex size() const { return rows() * cols(); } + inline Index size() const { return rows() * cols(); } /** \internal Don't use it, but do the equivalent: \code dst = *this; \endcode */ template diff --git a/Eigen/src/Core/Inverse.h b/Eigen/src/Core/Inverse.h index f3b0dff87..f3fa82a01 100644 --- a/Eigen/src/Core/Inverse.h +++ b/Eigen/src/Core/Inverse.h @@ -45,7 +45,7 @@ template class Inverse : public InverseImpl::StorageKind> { public: - typedef typename XprType::Index Index; + typedef typename XprType::StorageIndex StorageIndex; typedef typename XprType::PlainObject PlainObject; typedef typename internal::nested::type XprTypeNested; typedef typename internal::remove_all::type XprTypeNestedCleaned; diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h index 5c00d6a63..ed28b4d07 100644 --- a/Eigen/src/Core/MatrixBase.h +++ b/Eigen/src/Core/MatrixBase.h @@ -52,8 +52,7 @@ template class MatrixBase #ifndef EIGEN_PARSED_BY_DOXYGEN typedef MatrixBase StorageBaseType; typedef typename internal::traits::StorageKind StorageKind; - typedef Eigen::Index Index; - typedef Index StorageIndex; + typedef typename internal::traits::StorageIndex StorageIndex; typedef typename internal::traits::Scalar Scalar; typedef typename internal::packet_traits::type PacketScalar; typedef typename NumTraits::Real RealScalar; diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index 886d59a2c..1da27c06c 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -66,11 +66,10 @@ class PermutationBase : public EigenBase MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime, MaxColsAtCompileTime = Traits::MaxColsAtCompileTime }; - typedef typename Traits::StorageIndexType StorageIndexType; typedef typename Traits::StorageIndex StorageIndex; - typedef Matrix + typedef Matrix DenseMatrixType; - typedef PermutationMatrix + typedef PermutationMatrix PlainPermutationType; using Base::derived; typedef Transpose TransposeReturnType; @@ -148,7 +147,7 @@ class PermutationBase : public EigenBase /** Sets *this to be the identity permutation matrix */ void setIdentity() { - for(StorageIndexType i = 0; i < size(); ++i) + for(Index i = 0; i < size(); ++i) indices().coeffRef(i) = i; } @@ -174,8 +173,8 @@ class PermutationBase : public EigenBase eigen_assert(i>=0 && j>=0 && i * * \param SizeAtCompileTime the number of rows/cols, or Dynamic * \param MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it. - * \param StorageIndexType the integer type of the indices + * \param StorageIndex the integer type of the indices * * This class represents a permutation matrix, internally stored as a vector of integers. * @@ -271,19 +270,18 @@ class PermutationBase : public EigenBase */ namespace internal { -template -struct traits > - : traits > +template +struct traits > + : traits > { typedef PermutationStorage StorageKind; - typedef Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; - typedef typename IndicesType::StorageIndex StorageIndex; - typedef _StorageIndexType StorageIndexType; + typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; + typedef _StorageIndex StorageIndex; }; } -template -class PermutationMatrix : public PermutationBase > +template +class PermutationMatrix : public PermutationBase > { typedef PermutationBase Base; typedef internal::traits Traits; @@ -293,7 +291,6 @@ class PermutationMatrix : public PermutationBase::highest()); + eigen_internal_assert(size <= NumTraits::highest()); } /** Copy constructor. */ @@ -376,9 +373,9 @@ class PermutationMatrix : public PermutationBase >& other) : m_indices(other.nestedPermutation().size()) { - eigen_internal_assert(m_indices.size() <= NumTraits::highest()); - StorageIndexType end = StorageIndexType(m_indices.size()); - for (StorageIndexType i=0; i::highest()); + StorageIndex end = StorageIndex(m_indices.size()); + for (StorageIndex i=0; i @@ -396,20 +393,19 @@ class PermutationMatrix : public PermutationBase -struct traits,_PacketAccess> > - : traits > +template +struct traits,_PacketAccess> > + : traits > { typedef PermutationStorage StorageKind; - typedef Map, _PacketAccess> IndicesType; - typedef typename IndicesType::StorageIndex StorageIndex; - typedef _StorageIndexType StorageIndexType; + typedef Map, _PacketAccess> IndicesType; + typedef _StorageIndex StorageIndex; }; } -template -class Map,_PacketAccess> - : public PermutationBase,_PacketAccess> > +template +class Map,_PacketAccess> + : public PermutationBase,_PacketAccess> > { typedef PermutationBase Base; typedef internal::traits Traits; @@ -417,15 +413,14 @@ class Map > { typedef PermutationStorage StorageKind; typedef typename _IndicesType::Scalar Scalar; - typedef typename _IndicesType::Scalar StorageIndexType; - typedef typename _IndicesType::StorageIndex StorageIndex; + typedef typename _IndicesType::Scalar StorageIndex; typedef _IndicesType IndicesType; enum { RowsAtCompileTime = _IndicesType::SizeAtCompileTime, diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h index 8ff13fbba..74b895792 100644 --- a/Eigen/src/Core/Product.h +++ b/Eigen/src/Core/Product.h @@ -120,8 +120,8 @@ class Product : public ProductImpl<_Lhs,_Rhs,Option, && "if you wanted a coeff-wise or a dot product use the respective explicit functions"); } - EIGEN_DEVICE_FUNC inline StorageIndex rows() const { return m_lhs.rows(); } - EIGEN_DEVICE_FUNC inline StorageIndex cols() const { return m_rhs.cols(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); } EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; } EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; } @@ -149,7 +149,6 @@ class dense_product_base public: using Base::derived; typedef typename Base::Scalar Scalar; - typedef typename Base::StorageIndex StorageIndex; operator const Scalar() const { diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h index 7f9d135f7..d84e7776b 100644 --- a/Eigen/src/Core/ProductEvaluators.h +++ b/Eigen/src/Core/ProductEvaluators.h @@ -373,7 +373,6 @@ struct product_evaluator, ProductTag, DenseShape, : evaluator_base > { typedef Product XprType; - typedef typename XprType::StorageIndex StorageIndex; typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketScalar PacketScalar; @@ -525,7 +524,6 @@ struct product_evaluator, LazyCoeffBasedProduc template struct etor_product_packet_impl { - typedef typename Lhs::StorageIndex StorageIndex; static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res) { etor_product_packet_impl::run(row, col, lhs, rhs, innerDim, res); @@ -536,7 +534,6 @@ struct etor_product_packet_impl struct etor_product_packet_impl { - typedef typename Lhs::StorageIndex StorageIndex; static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res) { etor_product_packet_impl::run(row, col, lhs, rhs, innerDim, res); @@ -547,7 +544,6 @@ struct etor_product_packet_impl struct etor_product_packet_impl { - typedef typename Lhs::StorageIndex StorageIndex; static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res) { res = pmul(pset1(lhs.coeff(row, 0)),rhs.template packet(0, col)); @@ -557,7 +553,6 @@ struct etor_product_packet_impl template struct etor_product_packet_impl { - typedef typename Lhs::StorageIndex StorageIndex; static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res) { res = pmul(lhs.template packet(row, 0), pset1(rhs.coeff(0, col))); @@ -567,7 +562,6 @@ struct etor_product_packet_impl template struct etor_product_packet_impl { - typedef typename Lhs::StorageIndex StorageIndex; static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res) { eigen_assert(innerDim>0 && "you are using a non initialized matrix"); @@ -580,7 +574,6 @@ struct etor_product_packet_impl template struct etor_product_packet_impl { - typedef typename Lhs::StorageIndex StorageIndex; static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res) { eigen_assert(innerDim>0 && "you are using a non initialized matrix"); @@ -669,7 +662,6 @@ template { - typedef typename MatrixType::StorageIndex StorageIndex; typedef typename scalar_product_traits::ReturnType Scalar; typedef typename internal::packet_traits::type PacketScalar; public: @@ -734,7 +726,6 @@ struct product_evaluator, ProductTag, DiagonalSha using Base::coeff; using Base::packet_impl; typedef typename Base::Scalar Scalar; - typedef typename Base::StorageIndex StorageIndex; typedef typename Base::PacketScalar PacketScalar; typedef Product XprType; @@ -782,7 +773,6 @@ struct product_evaluator, ProductTag, DenseShape, using Base::coeff; using Base::packet_impl; typedef typename Base::Scalar Scalar; - typedef typename Base::StorageIndex StorageIndex; typedef typename Base::PacketScalar PacketScalar; typedef Product XprType; diff --git a/Eigen/src/Core/ReturnByValue.h b/Eigen/src/Core/ReturnByValue.h index d2b80d872..af01a5567 100644 --- a/Eigen/src/Core/ReturnByValue.h +++ b/Eigen/src/Core/ReturnByValue.h @@ -61,8 +61,8 @@ template class ReturnByValue EIGEN_DEVICE_FUNC inline void evalTo(Dest& dst) const { static_cast(this)->evalTo(dst); } - EIGEN_DEVICE_FUNC inline StorageIndex rows() const { return static_cast(this)->rows(); } - EIGEN_DEVICE_FUNC inline StorageIndex cols() const { return static_cast(this)->cols(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return static_cast(this)->rows(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return static_cast(this)->cols(); } #ifndef EIGEN_PARSED_BY_DOXYGEN #define Unusable YOU_ARE_TRYING_TO_ACCESS_A_SINGLE_COEFFICIENT_IN_A_SPECIAL_EXPRESSION_WHERE_THAT_IS_NOT_ALLOWED_BECAUSE_THAT_WOULD_BE_INEFFICIENT diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h index 2d5760066..a05746ad2 100644 --- a/Eigen/src/Core/SelfAdjointView.h +++ b/Eigen/src/Core/SelfAdjointView.h @@ -58,7 +58,6 @@ template class SelfAdjointView /** \brief The type of coefficients in this matrix */ typedef typename internal::traits::Scalar Scalar; - typedef typename MatrixType::StorageIndex StorageIndex; enum { @@ -224,7 +223,6 @@ public: typedef typename Base::DstEvaluatorType DstEvaluatorType; typedef typename Base::SrcEvaluatorType SrcEvaluatorType; typedef typename Base::Scalar Scalar; - typedef typename Base::StorageIndex StorageIndex; typedef typename Base::AssignmentTraits AssignmentTraits; diff --git a/Eigen/src/Core/Solve.h b/Eigen/src/Core/Solve.h index 5a3a4235e..47446b49e 100644 --- a/Eigen/src/Core/Solve.h +++ b/Eigen/src/Core/Solve.h @@ -70,8 +70,8 @@ public: : m_dec(dec), m_rhs(rhs) {} - EIGEN_DEVICE_FUNC StorageIndex rows() const { return m_dec.cols(); } - EIGEN_DEVICE_FUNC StorageIndex cols() const { return m_rhs.cols(); } + EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); } + EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); } EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; } EIGEN_DEVICE_FUNC const RhsType& rhs() const { return m_rhs; } diff --git a/Eigen/src/Core/Swap.h b/Eigen/src/Core/Swap.h index 3d4d8b802..dcb42821f 100644 --- a/Eigen/src/Core/Swap.h +++ b/Eigen/src/Core/Swap.h @@ -28,7 +28,6 @@ protected: public: typedef typename Base::Scalar Scalar; - typedef typename Base::StorageIndex StorageIndex; typedef typename Base::DstXprType DstXprType; typedef swap_assign_op Functor; diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 5d60ba149..7e41769a3 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -64,8 +64,8 @@ template class Transpose EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose) - EIGEN_DEVICE_FUNC inline StorageIndex rows() const { return m_matrix.cols(); } - EIGEN_DEVICE_FUNC inline StorageIndex cols() const { return m_matrix.rows(); } + EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.cols(); } + EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.rows(); } /** \returns the nested expression */ EIGEN_DEVICE_FUNC diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h index d8135be27..fd53ae4cb 100644 --- a/Eigen/src/Core/TriangularMatrix.h +++ b/Eigen/src/Core/TriangularMatrix.h @@ -54,9 +54,9 @@ template class TriangularBase : public EigenBase inline TriangularBase() { eigen_assert(!((Mode&UnitDiag) && (Mode&ZeroDiag))); } EIGEN_DEVICE_FUNC - inline StorageIndex rows() const { return derived().rows(); } + inline Index rows() const { return derived().rows(); } EIGEN_DEVICE_FUNC - inline StorageIndex cols() const { return derived().cols(); } + inline Index cols() const { return derived().cols(); } EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().outerStride(); } EIGEN_DEVICE_FUNC @@ -199,7 +199,6 @@ template class TriangularView public: typedef typename internal::traits::StorageKind StorageKind; - typedef typename internal::traits::StorageIndex StorageIndex; typedef typename internal::traits::MatrixTypeNestedCleaned NestedExpression; enum { @@ -222,9 +221,9 @@ template class TriangularView { return Base::operator=(other); } EIGEN_DEVICE_FUNC - inline StorageIndex rows() const { return m_matrix.rows(); } + inline Index rows() const { return m_matrix.rows(); } EIGEN_DEVICE_FUNC - inline StorageIndex cols() const { return m_matrix.cols(); } + inline Index cols() const { return m_matrix.cols(); } EIGEN_DEVICE_FUNC const NestedExpression& nestedExpression() const { return m_matrix; } @@ -325,7 +324,6 @@ template class TriangularViewImpl<_Mat using Base::derived; typedef typename internal::traits::StorageKind StorageKind; - typedef typename internal::traits::StorageIndex StorageIndex; enum { Mode = _Mode, @@ -688,7 +686,6 @@ public: typedef typename Base::DstEvaluatorType DstEvaluatorType; typedef typename Base::SrcEvaluatorType SrcEvaluatorType; typedef typename Base::Scalar Scalar; - typedef typename Base::StorageIndex StorageIndex; typedef typename Base::AssignmentTraits AssignmentTraits; @@ -831,7 +828,6 @@ struct triangular_assignment_loop template struct triangular_assignment_loop { - typedef typename Kernel::StorageIndex StorageIndex; typedef typename Kernel::Scalar Scalar; EIGEN_DEVICE_FUNC static inline void run(Kernel &kernel) diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 54842b5ff..34d337c90 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -66,7 +66,7 @@ template struct traits > { typedef _Scalar Scalar; - typedef DenseIndex StorageIndex; + typedef Eigen::Index StorageIndex; typedef Dense StorageKind; enum { Dim1 = _Dim==Dynamic ? _Dim : _Dim + 1, @@ -204,8 +204,9 @@ public: }; /** the scalar type of the coefficients */ typedef _Scalar Scalar; - typedef DenseIndex StorageIndex; - typedef DenseIndex Index; + typedef Eigen::Index StorageIndex; + /** \deprecated */ + typedef Eigen::Index Index; /** type of the matrix used to represent the transformation */ typedef typename internal::make_proper_matrix_type::type MatrixType; /** constified MatrixType */ diff --git a/Eigen/src/Householder/HouseholderSequence.h b/Eigen/src/Householder/HouseholderSequence.h index bf2bb59ab..74cd0a472 100644 --- a/Eigen/src/Householder/HouseholderSequence.h +++ b/Eigen/src/Householder/HouseholderSequence.h @@ -87,7 +87,6 @@ struct hseq_side_dependent_impl { typedef Block EssentialVectorType; typedef HouseholderSequence HouseholderSequenceType; - typedef typename VectorsType::StorageIndex StorageIndex; static inline const EssentialVectorType essentialVector(const HouseholderSequenceType& h, Index k) { Index start = k+1+h.m_shift; @@ -100,7 +99,6 @@ struct hseq_side_dependent_impl { typedef Transpose > EssentialVectorType; typedef HouseholderSequence HouseholderSequenceType; - typedef typename VectorsType::StorageIndex StorageIndex; static inline const EssentialVectorType essentialVector(const HouseholderSequenceType& h, Index k) { Index start = k+1+h.m_shift; @@ -131,7 +129,6 @@ template class HouseholderS MaxColsAtCompileTime = internal::traits::MaxColsAtCompileTime }; typedef typename internal::traits::Scalar Scalar; - typedef typename VectorsType::StorageIndex StorageIndex; typedef HouseholderSequence< typename internal::conditional::IsComplex, diff --git a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h index c413e9e1a..6d63d45e4 100644 --- a/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h +++ b/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h @@ -25,14 +25,14 @@ namespace internal { * \param ind The array of index for the elements in @p row * \param ncut The number of largest elements to keep **/ -template +template Index QuickSplit(VectorV &row, VectorI &ind, Index ncut) { typedef typename VectorV::RealScalar RealScalar; using std::swap; using std::abs; Index mid; - Index n = convert_index(row.size()); /* length of the vector */ + Index n = row.size(); /* length of the vector */ Index first, last ; ncut--; /* to fit the zero-based indices */ @@ -124,9 +124,9 @@ class IncompleteLUT : public SparseSolverBase > compute(mat); } - StorageIndex rows() const { return m_lu.rows(); } + Index rows() const { return m_lu.rows(); } - StorageIndex cols() const { return m_lu.cols(); } + Index cols() const { return m_lu.cols(); } /** \brief Reports whether previous computation was successful. * @@ -239,9 +239,10 @@ void IncompleteLUT::factorize(const _MatrixType& amat) using std::sqrt; using std::swap; using std::abs; + using internal::convert_index; eigen_assert((amat.rows() == amat.cols()) && "The factorization should be done on a square matrix"); - StorageIndex n = amat.cols(); // Size of the matrix + Index n = amat.cols(); // Size of the matrix m_lu.resize(n,n); // Declare Working vectors and variables Vector u(n) ; // real values of the row -- maximum size is n -- @@ -259,36 +260,36 @@ void IncompleteLUT::factorize(const _MatrixType& amat) u.fill(0); // number of largest elements to keep in each row: - StorageIndex fill_in = static_cast (amat.nonZeros()*m_fillfactor)/n+1; + Index fill_in = (amat.nonZeros()*m_fillfactor)/n + 1; if (fill_in > n) fill_in = n; // number of largest nonzero elements to keep in the L and the U part of the current row: - StorageIndex nnzL = fill_in/2; - StorageIndex nnzU = nnzL; + Index nnzL = fill_in/2; + Index nnzU = nnzL; m_lu.reserve(n * (nnzL + nnzU + 1)); // global loop over the rows of the sparse matrix - for (StorageIndex ii = 0; ii < n; ii++) + for (Index ii = 0; ii < n; ii++) { // 1 - copy the lower and the upper part of the row i of mat in the working vector u - StorageIndex sizeu = 1; // number of nonzero elements in the upper part of the current row - StorageIndex sizel = 0; // number of nonzero elements in the lower part of the current row - ju(ii) = ii; + Index sizeu = 1; // number of nonzero elements in the upper part of the current row + Index sizel = 0; // number of nonzero elements in the lower part of the current row + ju(ii) = convert_index(ii); u(ii) = 0; - jr(ii) = ii; + jr(ii) = convert_index(ii); RealScalar rownorm = 0; typename FactorType::InnerIterator j_it(mat, ii); // Iterate through the current row ii for (; j_it; ++j_it) { - StorageIndex k = j_it.index(); + Index k = j_it.index(); if (k < ii) { // copy the lower part - ju(sizel) = k; + ju(sizel) = convert_index(k); u(sizel) = j_it.value(); - jr(k) = sizel; + jr(k) = convert_index(sizel); ++sizel; } else if (k == ii) @@ -298,10 +299,10 @@ void IncompleteLUT::factorize(const _MatrixType& amat) else { // copy the upper part - StorageIndex jpos = ii + sizeu; - ju(jpos) = k; + Index jpos = ii + sizeu; + ju(jpos) = convert_index(k); u(jpos) = j_it.value(); - jr(k) = jpos; + jr(k) = convert_index(jpos); ++sizeu; } rownorm += numext::abs2(j_it.value()); @@ -317,21 +318,22 @@ void IncompleteLUT::factorize(const _MatrixType& amat) rownorm = sqrt(rownorm); // 3 - eliminate the previous nonzero rows - StorageIndex jj = 0; - StorageIndex len = 0; + Index jj = 0; + Index len = 0; while (jj < sizel) { // In order to eliminate in the correct order, // we must select first the smallest column index among ju(jj:sizel) - StorageIndex k; - StorageIndex minrow = ju.segment(jj,sizel-jj).minCoeff(&k); // k is relative to the segment + Index k; + Index minrow = ju.segment(jj,sizel-jj).minCoeff(&k); // k is relative to the segment k += jj; if (minrow != ju(jj)) { // swap the two locations - StorageIndex j = ju(jj); + Index j = ju(jj); swap(ju(jj), ju(k)); - jr(minrow) = jj; jr(j) = k; + jr(minrow) = convert_index(jj); + jr(j) = convert_index(k); swap(u(jj), u(k)); } // Reset this location @@ -355,11 +357,11 @@ void IncompleteLUT::factorize(const _MatrixType& amat) for (; ki_it; ++ki_it) { Scalar prod = fact * ki_it.value(); - StorageIndex j = ki_it.index(); - StorageIndex jpos = jr(j); + Index j = ki_it.index(); + Index jpos = jr(j); if (jpos == -1) // fill-in element { - StorageIndex newpos; + Index newpos; if (j >= ii) // dealing with the upper part { newpos = ii + sizeu; @@ -372,23 +374,23 @@ void IncompleteLUT::factorize(const _MatrixType& amat) sizel++; eigen_internal_assert(sizel<=ii); } - ju(newpos) = j; + ju(newpos) = convert_index(j); u(newpos) = -prod; - jr(j) = newpos; + jr(j) = convert_index(newpos); } else u(jpos) -= prod; } // store the pivot element - u(len) = fact; - ju(len) = minrow; + u(len) = fact; + ju(len) = convert_index(minrow); ++len; jj++; } // end of the elimination on the row ii // reset the upper part of the pointer jr to zero - for(StorageIndex k = 0; k ::factorize(const _MatrixType& amat) // store the largest m_fill elements of the L part m_lu.startVec(ii); - for(StorageIndex k = 0; k < len; k++) + for(Index k = 0; k < len; k++) m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k); // store the diagonal element @@ -413,7 +415,7 @@ void IncompleteLUT::factorize(const _MatrixType& amat) // sort the U-part of the row // apply the dropping rule first len = 0; - for(StorageIndex k = 1; k < sizeu; k++) + for(Index k = 1; k < sizeu; k++) { if(abs(u(ii+k)) > m_droptol * rownorm ) { @@ -429,7 +431,7 @@ void IncompleteLUT::factorize(const _MatrixType& amat) internal::QuickSplit(uu, juu, len); // store the largest elements of the U part - for(StorageIndex k = ii + 1; k < ii + len; k++) + for(Index k = ii + 1; k < ii + len; k++) m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k); } diff --git a/Eigen/src/LU/FullPivLU.h b/Eigen/src/LU/FullPivLU.h index eb4520004..49c8b183d 100644 --- a/Eigen/src/LU/FullPivLU.h +++ b/Eigen/src/LU/FullPivLU.h @@ -66,10 +66,10 @@ template class FullPivLU typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; typedef typename internal::traits::StorageKind StorageKind; - typedef typename MatrixType::Index Index; + // FIXME should be int typedef typename MatrixType::StorageIndex StorageIndex; - typedef typename internal::plain_row_type::type IntRowVectorType; - typedef typename internal::plain_col_type::type IntColVectorType; + typedef typename internal::plain_row_type::type IntRowVectorType; + typedef typename internal::plain_col_type::type IntColVectorType; typedef PermutationMatrix PermutationQType; typedef PermutationMatrix PermutationPType; typedef typename MatrixType::PlainObject PlainObject; diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index 7e2c8b471..43c2a716e 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -72,7 +72,7 @@ template class PartialPivLU typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; typedef typename internal::traits::StorageKind StorageKind; - typedef typename MatrixType::Index Index; + // FIXME should be int typedef typename MatrixType::StorageIndex StorageIndex; typedef PermutationMatrix PermutationType; typedef Transpositions TranspositionType; diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index 0dee139ce..c500529da 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -57,7 +57,7 @@ template class ColPivHouseholderQR }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; + // FIXME should be int typedef typename MatrixType::StorageIndex StorageIndex; typedef Matrix MatrixQType; typedef typename internal::plain_diag_type::type HCoeffsType; diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index 90ab25b2b..a7a0d9138 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -66,11 +66,11 @@ template class FullPivHouseholderQR }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; + // FIXME should be int typedef typename MatrixType::StorageIndex StorageIndex; typedef internal::FullPivHouseholderQRMatrixQReturnType MatrixQReturnType; typedef typename internal::plain_diag_type::type HCoeffsType; - typedef Matrix IntDiagSizeVectorType; typedef PermutationMatrix PermutationType; diff --git a/Eigen/src/QR/HouseholderQR.h b/Eigen/src/QR/HouseholderQR.h index 5156a6e4b..8f25cf728 100644 --- a/Eigen/src/QR/HouseholderQR.h +++ b/Eigen/src/QR/HouseholderQR.h @@ -53,7 +53,7 @@ template class HouseholderQR }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; + // FIXME should be int typedef typename MatrixType::StorageIndex StorageIndex; typedef Matrix MatrixQType; typedef typename internal::plain_diag_type::type HCoeffsType; diff --git a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h index 58efa6024..4ad22f8b4 100644 --- a/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h +++ b/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h @@ -236,7 +236,7 @@ class SPQR : public SparseSolverBase > mutable cholmod_sparse *m_H; //The householder vectors mutable StorageIndex *m_HPinv; // The row permutation of H mutable cholmod_dense *m_HTau; // The Householder coefficients - mutable StorageIndex m_rank; // The rank of the matrix + mutable Index m_rank; // The rank of the matrix mutable cholmod_common m_cc; // Workspace and parameters bool m_useDefaultThreshold; // Use default threshold template friend struct SPQR_QProduct; @@ -250,8 +250,8 @@ struct SPQR_QProduct : ReturnByValue > //Define the constructor to get reference to argument types SPQR_QProduct(const SPQRType& spqr, const Derived& other, bool transpose) : m_spqr(spqr),m_other(other),m_transpose(transpose) {} - inline StorageIndex rows() const { return m_transpose ? m_spqr.rows() : m_spqr.cols(); } - inline StorageIndex cols() const { return m_other.cols(); } + inline Index rows() const { return m_transpose ? m_spqr.rows() : m_spqr.cols(); } + inline Index cols() const { return m_other.cols(); } // Assign to a vector template void evalTo(ResType& res) const diff --git a/Eigen/src/SVD/SVDBase.h b/Eigen/src/SVD/SVDBase.h index 0bc2ede28..95d378da9 100644 --- a/Eigen/src/SVD/SVDBase.h +++ b/Eigen/src/SVD/SVDBase.h @@ -52,7 +52,6 @@ public: typedef typename internal::traits::MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; typedef typename MatrixType::StorageIndex StorageIndex; enum { RowsAtCompileTime = MatrixType::RowsAtCompileTime, diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index c47077e50..2580151de 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -92,8 +92,8 @@ class SimplicialCholeskyBase : public SparseSolverBase Derived& derived() { return *static_cast(this); } const Derived& derived() const { return *static_cast(this); } - inline StorageIndex cols() const { return m_matrix.cols(); } - inline StorageIndex rows() const { return m_matrix.rows(); } + inline Index cols() const { return m_matrix.cols(); } + inline Index rows() const { return m_matrix.rows(); } /** \brief Reports whether previous computation was successful. * @@ -108,12 +108,12 @@ class SimplicialCholeskyBase : public SparseSolverBase /** \returns the permutation P * \sa permutationPinv() */ - const PermutationMatrix& permutationP() const + const PermutationMatrix& permutationP() const { return m_P; } /** \returns the inverse P^-1 of the permutation P * \sa permutationP() */ - const PermutationMatrix& permutationPinv() const + const PermutationMatrix& permutationPinv() const { return m_Pinv; } /** Sets the shift parameters that will be used to adjust the diagonal coefficients during the numerical factorization. diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h index a93189df2..9e2e878e0 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h @@ -57,7 +57,7 @@ void SimplicialCholeskyBase::analyzePattern_preordered(const CholMatrix ei_declare_aligned_stack_constructed_variable(Index, tags, size, 0); - for(StorageIndex k = 0; k < size; ++k) + for(Index k = 0; k < size; ++k) { /* L(k,:) pattern: all nodes reachable in etree from nz in A(0:k-1,k) */ m_parent[k] = -1; /* parent of k is not yet known */ @@ -104,7 +104,7 @@ void SimplicialCholeskyBase::factorize_preordered(const CholMatrixType& eigen_assert(m_analysisIsOk && "You must first call analyzePattern()"); eigen_assert(ap.rows()==ap.cols()); - const StorageIndex size = ap.rows(); + const Index size = ap.rows(); eigen_assert(m_parent.size()==size); eigen_assert(m_nonZerosPerCol.size()==size); @@ -119,20 +119,20 @@ void SimplicialCholeskyBase::factorize_preordered(const CholMatrixType& bool ok = true; m_diag.resize(DoLDLT ? size : 0); - for(StorageIndex k = 0; k < size; ++k) + for(Index k = 0; k < size; ++k) { // compute nonzero pattern of kth row of L, in topological order y[k] = 0.0; // Y(0:k) is now all zero - StorageIndex top = size; // stack for pattern is empty + Index top = size; // stack for pattern is empty tags[k] = k; // mark node k as visited m_nonZerosPerCol[k] = 0; // count of nonzeros in column k of L for(typename CholMatrixType::InnerIterator it(ap,k); it; ++it) { - StorageIndex i = it.index(); + Index i = it.index(); if(i <= k) { y[i] += numext::conj(it.value()); /* scatter A(i,k) into Y (sum duplicates) */ - StorageIndex len; + Index len; for(len = 0; tags[i] != k; i = m_parent[i]) { pattern[len++] = i; /* L(k,i) is nonzero */ @@ -149,7 +149,7 @@ void SimplicialCholeskyBase::factorize_preordered(const CholMatrixType& y[k] = 0.0; for(; top < size; ++top) { - StorageIndex i = pattern[top]; /* pattern[top:n-1] is pattern of L(:,k) */ + Index i = pattern[top]; /* pattern[top:n-1] is pattern of L(:,k) */ Scalar yi = y[i]; /* get and clear Y(i) */ y[i] = 0.0; @@ -160,8 +160,8 @@ void SimplicialCholeskyBase::factorize_preordered(const CholMatrixType& else yi = l_ki = yi / Lx[Lp[i]]; - StorageIndex p2 = Lp[i] + m_nonZerosPerCol[i]; - StorageIndex p; + Index p2 = Lp[i] + m_nonZerosPerCol[i]; + Index p; for(p = Lp[i] + (DoLDLT ? 0 : 1); p < p2; ++p) y[Li[p]] -= numext::conj(Lx[p]) * yi; d -= numext::real(l_ki * numext::conj(yi)); @@ -180,7 +180,7 @@ void SimplicialCholeskyBase::factorize_preordered(const CholMatrixType& } else { - StorageIndex p = Lp[k] + m_nonZerosPerCol[k]++; + Index p = Lp[k] + m_nonZerosPerCol[k]++; Li[p] = k ; /* store L(k,k) = sqrt (d) in column k */ if(d <= RealScalar(0)) { ok = false; /* failure, matrix is not positive definite */ diff --git a/Eigen/src/SparseCore/AmbiVector.h b/Eigen/src/SparseCore/AmbiVector.h index 8bde5d58e..1233e164e 100644 --- a/Eigen/src/SparseCore/AmbiVector.h +++ b/Eigen/src/SparseCore/AmbiVector.h @@ -36,7 +36,7 @@ class AmbiVector void init(double estimatedDensity); void init(int mode); - StorageIndex nonZeros() const; + Index nonZeros() const; /** Specifies a sub-vector to work on */ void setBounds(Index start, Index end) { m_start = convert_index(start); m_end = convert_index(end); } @@ -126,7 +126,7 @@ class AmbiVector /** \returns the number of non zeros in the current sub vector */ template -_StorageIndex AmbiVector<_Scalar,_StorageIndex>::nonZeros() const +Index AmbiVector<_Scalar,_StorageIndex>::nonZeros() const { if (m_mode==IsSparse) return m_llSize; diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index 2d4f2bcf8..bba8a104b 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -36,7 +36,7 @@ class CompressedStorage : m_values(0), m_indices(0), m_size(0), m_allocatedSize(0) {} - explicit CompressedStorage(size_t size) + explicit CompressedStorage(Index size) : m_values(0), m_indices(0), m_size(0), m_allocatedSize(0) { resize(size); @@ -70,9 +70,9 @@ class CompressedStorage delete[] m_indices; } - void reserve(size_t size) + void reserve(Index size) { - size_t newAllocatedSize = m_size + size; + Index newAllocatedSize = m_size + size; if (newAllocatedSize > m_allocatedSize) reallocate(newAllocatedSize); } @@ -83,13 +83,14 @@ class CompressedStorage reallocate(m_size); } - void resize(size_t size, double reserveSizeFactor = 0) + void resize(Index size, double reserveSizeFactor = 0) { if (m_allocatedSize(i); } - inline size_t size() const { return m_size; } - inline size_t allocatedSize() const { return m_allocatedSize; } + inline Index size() const { return m_size; } + inline Index allocatedSize() const { return m_allocatedSize; } inline void clear() { m_size = 0; } - inline Scalar& value(size_t i) { return m_values[i]; } - inline const Scalar& value(size_t i) const { return m_values[i]; } + inline Scalar& value(Index i) { return m_values[i]; } + inline const Scalar& value(Index i) const { return m_values[i]; } - inline StorageIndex& index(size_t i) { return m_indices[i]; } - inline const StorageIndex& index(size_t i) const { return m_indices[i]; } + inline StorageIndex& index(Index i) { return m_indices[i]; } + inline const StorageIndex& index(Index i) const { return m_indices[i]; } /** \returns the largest \c k such that for all \c j in [0,k) index[\c j]\<\a key */ - inline StorageIndex searchLowerIndex(Index key) const + inline StorageIndex searchLowerIndex(StorageIndex key) const { return searchLowerIndex(0, m_size, key); } /** \returns the largest \c k in [start,end) such that for all \c j in [start,k) index[\c j]\<\a key */ - inline StorageIndex searchLowerIndex(size_t start, size_t end, Index key) const + inline Index searchLowerIndex(Index start, Index end, StorageIndex key) const { while(end>start) { - size_t mid = (end+start)>>1; + Index mid = (end+start)>>1; if (m_indices[mid](start); + return start; } /** \returns the stored value at index \a key @@ -138,12 +139,12 @@ class CompressedStorage return m_values[m_size-1]; // ^^ optimization: let's first check if it is the last coefficient // (very common in high level algorithms) - const size_t id = searchLowerIndex(0,m_size-1,key); + const Index id = searchLowerIndex(0,m_size-1,key); return ((id=end) return defaultValue; @@ -151,7 +152,7 @@ class CompressedStorage return m_values[end-1]; // ^^ optimization: let's first check if it is the last coefficient // (very common in high level algorithms) - const size_t id = searchLowerIndex(start,end-1,key); + const Index id = searchLowerIndex(start,end-1,key); return ((id=m_size || m_indices[id]!=key) { if (m_allocatedSize::dummy_precision()) { - size_t k = 0; - size_t n = size(); - for (size_t i=0; i newValues(size); internal::scoped_array newIndices(size); - size_t copySize = (std::min)(size, m_size); + Index copySize = (std::min)(size, m_size); internal::smart_copy(m_values, m_values+copySize, newValues.ptr()); internal::smart_copy(m_indices, m_indices+copySize, newIndices.ptr()); std::swap(m_values,newValues.ptr()); @@ -228,8 +229,8 @@ class CompressedStorage protected: Scalar* m_values; StorageIndex* m_indices; - size_t m_size; - size_t m_allocatedSize; + Index m_size; + Index m_allocatedSize; }; diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index dfdf4314a..5256bf950 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -34,14 +34,14 @@ public: : m_matrix(xpr), m_outerStart(convert_index(IsRowMajor ? startRow : startCol)), m_outerSize(convert_index(IsRowMajor ? blockRows : blockCols)) {} - EIGEN_STRONG_INLINE StorageIndex rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } - EIGEN_STRONG_INLINE StorageIndex cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } + EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } + EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } - StorageIndex nonZeros() const + Index nonZeros() const { typedef typename internal::evaluator::type EvaluatorType; EvaluatorType matEval(m_matrix); - StorageIndex nnz = 0; + Index nnz = 0; Index end = m_outerStart + m_outerSize.value(); for(Index j=m_outerStart; j m_outerSize; + Index m_outerStart; + const internal::variable_if_dynamic m_outerSize; public: EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) @@ -106,11 +106,11 @@ public: SparseMatrix tmp(other); // 2 - let's check whether there is enough allocated memory - StorageIndex nnz = tmp.nonZeros(); - StorageIndex start = m_outerStart==0 ? 0 : matrix.outerIndexPtr()[m_outerStart]; // starting position of the current block - StorageIndex end = m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]; // ending position of the current block - StorageIndex block_size = end - start; // available room in the current block - StorageIndex tail_size = m_matrix.outerIndexPtr()[m_matrix.outerSize()] - end; + Index nnz = tmp.nonZeros(); + Index start = m_outerStart==0 ? 0 : matrix.outerIndexPtr()[m_outerStart]; // starting position of the current block + Index end = m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]; // ending position of the current block + Index block_size = end - start; // available room in the current block + Index tail_size = m_matrix.outerIndexPtr()[m_matrix.outerSize()] - end; Index free_size = m_matrix.isCompressed() ? Index(matrix.data().allocatedSize()) + block_size @@ -192,7 +192,7 @@ public: inline StorageIndex* innerNonZeroPtr() { return isCompressed() ? 0 : m_matrix.const_cast_derived().innerNonZeroPtr(); } - StorageIndex nonZeros() const + Index nonZeros() const { if(m_matrix.isCompressed()) return ( (m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]) @@ -215,20 +215,20 @@ public: return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart]+m_matrix.innerNonZeroPtr()[m_outerStart]-1]; } - EIGEN_STRONG_INLINE StorageIndex rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } - EIGEN_STRONG_INLINE StorageIndex cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } + EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } + EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } - StorageIndex startRow() const { return IsRowMajor ? m_outerStart : 0; } - StorageIndex startCol() const { return IsRowMajor ? 0 : m_outerStart; } - StorageIndex blockRows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } - StorageIndex blockCols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } + Index startRow() const { return IsRowMajor ? m_outerStart : 0; } + Index startCol() const { return IsRowMajor ? 0 : m_outerStart; } + Index blockRows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); } + Index blockCols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); } protected: typename SparseMatrixType::Nested m_matrix; - StorageIndex m_outerStart; - const internal::variable_if_dynamic m_outerSize; + Index m_outerStart; + const internal::variable_if_dynamic m_outerSize; }; @@ -353,8 +353,8 @@ public: : m_matrix(xpr), m_startRow(convert_index(startRow)), m_startCol(convert_index(startCol)), m_blockRows(convert_index(blockRows)), m_blockCols(convert_index(blockCols)) {} - inline StorageIndex rows() const { return m_blockRows.value(); } - inline StorageIndex cols() const { return m_blockCols.value(); } + inline Index rows() const { return m_blockRows.value(); } + inline Index cols() const { return m_blockCols.value(); } inline Scalar& coeffRef(Index row, Index col) { @@ -382,10 +382,10 @@ public: } inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; } - StorageIndex startRow() const { return m_startRow.value(); } - StorageIndex startCol() const { return m_startCol.value(); } - StorageIndex blockRows() const { return m_blockRows.value(); } - StorageIndex blockCols() const { return m_blockCols.value(); } + Index startRow() const { return m_startRow.value(); } + Index startCol() const { return m_startCol.value(); } + Index blockRows() const { return m_blockRows.value(); } + Index blockCols() const { return m_blockCols.value(); } protected: friend class internal::GenericSparseBlockInnerIteratorImpl; @@ -394,10 +394,10 @@ public: EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl) typename XprType::Nested m_matrix; - const internal::variable_if_dynamic m_startRow; - const internal::variable_if_dynamic m_startCol; - const internal::variable_if_dynamic m_blockRows; - const internal::variable_if_dynamic m_blockCols; + const internal::variable_if_dynamic m_startRow; + const internal::variable_if_dynamic m_startCol; + const internal::variable_if_dynamic m_blockRows; + const internal::variable_if_dynamic m_blockCols; }; @@ -425,10 +425,10 @@ namespace internal { Base::operator++(); } - inline StorageIndex index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } - inline StorageIndex outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } - inline StorageIndex row() const { return Base::row() - m_block.m_startRow.value(); } - inline StorageIndex col() const { return Base::col() - m_block.m_startCol.value(); } + inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } + inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); } + inline Index row() const { return Base::row() - m_block.m_startRow.value(); } + inline Index col() const { return Base::col() - m_block.m_startCol.value(); } inline operator bool() const { return Base::operator bool() && Base::index() < m_end; } }; @@ -445,10 +445,10 @@ namespace internal { typedef typename BlockType::StorageIndex StorageIndex; typedef typename BlockType::Scalar Scalar; const BlockType& m_block; - StorageIndex m_outerPos; - StorageIndex m_innerIndex; + Index m_outerPos; + Index m_innerIndex; Scalar m_value; - StorageIndex m_end; + Index m_end; public: explicit EIGEN_STRONG_INLINE GenericSparseBlockInnerIteratorImpl(const BlockType& block, Index outer = 0) @@ -464,10 +464,10 @@ namespace internal { ++(*this); } - inline StorageIndex index() const { return m_outerPos - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } - inline StorageIndex outer() const { return 0; } - inline StorageIndex row() const { return IsRowMajor ? 0 : index(); } - inline StorageIndex col() const { return IsRowMajor ? index() : 0; } + inline Index index() const { return m_outerPos - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); } + inline Index outer() const { return 0; } + inline Index row() const { return IsRowMajor ? 0 : index(); } + inline Index col() const { return IsRowMajor ? index() : 0; } inline Scalar value() const { return m_value; } @@ -546,10 +546,10 @@ public: EvalIterator::operator++(); } - inline StorageIndex index() const { return EvalIterator::index() - (IsRowMajor ? m_block.startCol() : m_block.startRow()); } - inline StorageIndex outer() const { return EvalIterator::outer() - (IsRowMajor ? m_block.startRow() : m_block.startCol()); } - inline StorageIndex row() const { return EvalIterator::row() - m_block.startRow(); } - inline StorageIndex col() const { return EvalIterator::col() - m_block.startCol(); } + inline Index index() const { return EvalIterator::index() - (IsRowMajor ? m_block.startCol() : m_block.startRow()); } + inline Index outer() const { return EvalIterator::outer() - (IsRowMajor ? m_block.startRow() : m_block.startCol()); } + inline Index row() const { return EvalIterator::row() - m_block.startRow(); } + inline Index col() const { return EvalIterator::col() - m_block.startCol(); } inline operator bool() const { return EvalIterator::operator bool() && EvalIterator::index() < m_end; } }; @@ -558,10 +558,10 @@ template class unary_evaluator, IteratorBased>::OuterVectorInnerIterator { const unary_evaluator& m_eval; - StorageIndex m_outerPos; - StorageIndex m_innerIndex; + Index m_outerPos; + Index m_innerIndex; Scalar m_value; - StorageIndex m_end; + Index m_end; public: EIGEN_STRONG_INLINE OuterVectorInnerIterator(const unary_evaluator& aEval, Index outer) @@ -576,10 +576,10 @@ public: ++(*this); } - inline StorageIndex index() const { return m_outerPos - (IsRowMajor ? m_eval.m_block.startCol() : m_eval.m_block.startRow()); } - inline StorageIndex outer() const { return 0; } - inline StorageIndex row() const { return IsRowMajor ? 0 : index(); } - inline StorageIndex col() const { return IsRowMajor ? index() : 0; } + inline Index index() const { return m_outerPos - (IsRowMajor ? m_eval.m_block.startCol() : m_eval.m_block.startRow()); } + inline Index outer() const { return 0; } + inline Index row() const { return IsRowMajor ? 0 : index(); } + inline Index col() const { return IsRowMajor ? index() : 0; } inline Scalar value() const { return m_value; } diff --git a/Eigen/src/SparseCore/SparseColEtree.h b/Eigen/src/SparseCore/SparseColEtree.h index 88c799068..28fb2d175 100644 --- a/Eigen/src/SparseCore/SparseColEtree.h +++ b/Eigen/src/SparseCore/SparseColEtree.h @@ -60,7 +60,7 @@ Index etree_find (Index i, IndexVector& pp) template int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowElt, typename MatrixType::StorageIndex *perm=0) { - typedef typename MatrixType::StorageIndex Index; + typedef typename MatrixType::StorageIndex StorageIndex; Index nc = mat.cols(); // Number of columns Index m = mat.rows(); Index diagSize = (std::min)(nc,m); @@ -70,7 +70,7 @@ int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowEl pp.setZero(); // Initialize disjoint sets parent.resize(mat.cols()); //Compute first nonzero column in each row - Index row,col; + StorageIndex row,col; firstRowElt.resize(m); firstRowElt.setConstant(nc); firstRowElt.segment(0, diagSize).setLinSpaced(diagSize, 0, diagSize-1); @@ -127,7 +127,7 @@ int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowEl * Depth-first search from vertex n. No recursion. * This routine was contributed by Cédric Doucet, CEDRAT Group, Meylan, France. */ -template +template void nr_etdfs (Index n, IndexVector& parent, IndexVector& first_kid, IndexVector& next_kid, IndexVector& post, Index postnum) { Index current = n, first, next; @@ -174,7 +174,7 @@ void nr_etdfs (Index n, IndexVector& parent, IndexVector& first_kid, IndexVector * \param parent Input tree * \param post postordered tree */ -template +template void treePostorder(Index n, IndexVector& parent, IndexVector& post) { IndexVector first_kid, next_kid; // Linked list of children diff --git a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h index afb09ad91..3b4e9df59 100644 --- a/Eigen/src/SparseCore/SparseCwiseBinaryOp.h +++ b/Eigen/src/SparseCore/SparseCwiseBinaryOp.h @@ -56,7 +56,6 @@ public: class InnerIterator { typedef typename traits::Scalar Scalar; - typedef typename XprType::StorageIndex StorageIndex; public: @@ -97,9 +96,9 @@ public: EIGEN_STRONG_INLINE Scalar value() const { return m_value; } - EIGEN_STRONG_INLINE StorageIndex index() const { return m_id; } - EIGEN_STRONG_INLINE StorageIndex row() const { return Lhs::IsRowMajor ? m_lhsIter.row() : index(); } - EIGEN_STRONG_INLINE StorageIndex col() const { return Lhs::IsRowMajor ? index() : m_lhsIter.col(); } + EIGEN_STRONG_INLINE Index index() const { return m_id; } + EIGEN_STRONG_INLINE Index row() const { return Lhs::IsRowMajor ? m_lhsIter.row() : index(); } + EIGEN_STRONG_INLINE Index col() const { return Lhs::IsRowMajor ? index() : m_lhsIter.col(); } EIGEN_STRONG_INLINE operator bool() const { return m_id>=0; } @@ -108,7 +107,7 @@ public: RhsIterator m_rhsIter; const BinaryOp& m_functor; Scalar m_value; - StorageIndex m_id; + Index m_id; }; @@ -145,7 +144,6 @@ public: class InnerIterator { typedef typename traits::Scalar Scalar; - typedef typename XprType::StorageIndex StorageIndex; public: @@ -177,9 +175,9 @@ public: EIGEN_STRONG_INLINE Scalar value() const { return m_functor(m_lhsIter.value(), m_rhsIter.value()); } - EIGEN_STRONG_INLINE StorageIndex index() const { return m_lhsIter.index(); } - EIGEN_STRONG_INLINE StorageIndex row() const { return m_lhsIter.row(); } - EIGEN_STRONG_INLINE StorageIndex col() const { return m_lhsIter.col(); } + EIGEN_STRONG_INLINE Index index() const { return m_lhsIter.index(); } + EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); } + EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); } EIGEN_STRONG_INLINE operator bool() const { return (m_lhsIter && m_rhsIter); } @@ -223,7 +221,6 @@ public: class InnerIterator { typedef typename traits::Scalar Scalar; - typedef typename XprType::StorageIndex StorageIndex; enum { IsRowMajor = (int(Rhs::Flags)&RowMajorBit)==RowMajorBit }; public: @@ -241,9 +238,9 @@ public: EIGEN_STRONG_INLINE Scalar value() const { return m_functor(m_lhsEval.coeff(IsRowMajor?m_outer:m_rhsIter.index(),IsRowMajor?m_rhsIter.index():m_outer), m_rhsIter.value()); } - EIGEN_STRONG_INLINE StorageIndex index() const { return m_rhsIter.index(); } - EIGEN_STRONG_INLINE StorageIndex row() const { return m_rhsIter.row(); } - EIGEN_STRONG_INLINE StorageIndex col() const { return m_rhsIter.col(); } + EIGEN_STRONG_INLINE Index index() const { return m_rhsIter.index(); } + EIGEN_STRONG_INLINE Index row() const { return m_rhsIter.row(); } + EIGEN_STRONG_INLINE Index col() const { return m_rhsIter.col(); } EIGEN_STRONG_INLINE operator bool() const { return m_rhsIter; } @@ -288,7 +285,6 @@ public: class InnerIterator { typedef typename traits::Scalar Scalar; - typedef typename XprType::StorageIndex StorageIndex; enum { IsRowMajor = (int(Lhs::Flags)&RowMajorBit)==RowMajorBit }; public: @@ -307,9 +303,9 @@ public: { return m_functor(m_lhsIter.value(), m_rhsEval.coeff(IsRowMajor?m_outer:m_lhsIter.index(),IsRowMajor?m_lhsIter.index():m_outer)); } - EIGEN_STRONG_INLINE StorageIndex index() const { return m_lhsIter.index(); } - EIGEN_STRONG_INLINE StorageIndex row() const { return m_lhsIter.row(); } - EIGEN_STRONG_INLINE StorageIndex col() const { return m_lhsIter.col(); } + EIGEN_STRONG_INLINE Index index() const { return m_lhsIter.index(); } + EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); } + EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); } EIGEN_STRONG_INLINE operator bool() const { return m_lhsIter; } @@ -317,7 +313,7 @@ public: LhsIterator m_lhsIter; const RhsEvaluator &m_rhsEval; const BinaryOp& m_functor; - const StorageIndex m_outer; + const Index m_outer; }; diff --git a/Eigen/src/SparseCore/SparseDenseProduct.h b/Eigen/src/SparseCore/SparseDenseProduct.h index f6e6fab29..edb9d5998 100644 --- a/Eigen/src/SparseCore/SparseDenseProduct.h +++ b/Eigen/src/SparseCore/SparseDenseProduct.h @@ -29,7 +29,6 @@ struct sparse_time_dense_product_impl::type Lhs; typedef typename internal::remove_all::type Rhs; typedef typename internal::remove_all::type Res; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::InnerIterator LhsInnerIterator; static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) { @@ -62,7 +61,6 @@ struct sparse_time_dense_product_impl::type Lhs; typedef typename internal::remove_all::type Rhs; typedef typename internal::remove_all::type Res; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::InnerIterator LhsInnerIterator; static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha) { @@ -86,7 +84,6 @@ struct sparse_time_dense_product_impl::type Lhs; typedef typename internal::remove_all::type Rhs; typedef typename internal::remove_all::type Res; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::InnerIterator LhsInnerIterator; static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) { @@ -106,7 +103,6 @@ struct sparse_time_dense_product_impl::type Lhs; typedef typename internal::remove_all::type Rhs; typedef typename internal::remove_all::type Res; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::InnerIterator LhsInnerIterator; static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha) { @@ -193,7 +189,6 @@ protected: typedef typename evaluator::type RhsEval; typedef typename evaluator::InnerIterator LhsIterator; typedef typename ProdXprType::Scalar Scalar; - typedef typename ProdXprType::StorageIndex StorageIndex; public: enum { @@ -211,9 +206,9 @@ public: m_factor(get(xprEval.m_rhsXprImpl, outer, typename internal::traits::StorageKind() )) {} - EIGEN_STRONG_INLINE StorageIndex outer() const { return m_outer; } - EIGEN_STRONG_INLINE StorageIndex row() const { return NeedToTranspose ? m_outer : LhsIterator::index(); } - EIGEN_STRONG_INLINE StorageIndex col() const { return NeedToTranspose ? LhsIterator::index() : m_outer; } + EIGEN_STRONG_INLINE Index outer() const { return m_outer; } + EIGEN_STRONG_INLINE Index row() const { return NeedToTranspose ? m_outer : LhsIterator::index(); } + EIGEN_STRONG_INLINE Index col() const { return NeedToTranspose ? LhsIterator::index() : m_outer; } EIGEN_STRONG_INLINE Scalar value() const { return LhsIterator::value() * m_factor; } EIGEN_STRONG_INLINE operator bool() const { return LhsIterator::operator bool() && (!m_empty); } diff --git a/Eigen/src/SparseCore/SparseDiagonalProduct.h b/Eigen/src/SparseCore/SparseDiagonalProduct.h index 19a79edad..b7598c885 100644 --- a/Eigen/src/SparseCore/SparseDiagonalProduct.h +++ b/Eigen/src/SparseCore/SparseDiagonalProduct.h @@ -66,7 +66,6 @@ struct sparse_diagonal_product_evaluator::InnerIterator SparseXprInnerIterator; typedef typename SparseXprType::Scalar Scalar; - typedef typename SparseXprType::StorageIndex StorageIndex; public: class InnerIterator : public SparseXprInnerIterator @@ -96,7 +95,6 @@ template struct sparse_diagonal_product_evaluator { typedef typename SparseXprType::Scalar Scalar; - typedef typename SparseXprType::StorageIndex StorageIndex; typedef CwiseBinaryOp, const typename SparseXprType::ConstInnerVectorReturnType, @@ -111,14 +109,14 @@ struct sparse_diagonal_product_evaluator(outer)) + m_outer(outer) {} - inline Scalar value() const { return m_cwiseIter.value(); } - inline StorageIndex index() const { return convert_index(m_cwiseIter.index()); } - inline StorageIndex outer() const { return m_outer; } - inline StorageIndex col() const { return SparseXprType::IsRowMajor ? m_cwiseIter.index() : m_outer; } - inline StorageIndex row() const { return SparseXprType::IsRowMajor ? m_outer : m_cwiseIter.index(); } + inline Scalar value() const { return m_cwiseIter.value(); } + inline Index index() const { return m_cwiseIter.index(); } + inline Index outer() const { return m_outer; } + inline Index col() const { return SparseXprType::IsRowMajor ? m_cwiseIter.index() : m_outer; } + inline Index row() const { return SparseXprType::IsRowMajor ? m_outer : m_cwiseIter.index(); } EIGEN_STRONG_INLINE InnerIterator& operator++() { ++m_cwiseIter; return *this; } @@ -127,7 +125,7 @@ struct sparse_diagonal_product_evaluator TransposedSparseMatrix; - StorageIndex m_outerSize; - StorageIndex m_innerSize; + Index m_outerSize; + Index m_innerSize; StorageIndex* m_outerIndex; StorageIndex* m_innerNonZeros; // optional, if null then the data is compressed Storage m_data; @@ -129,14 +129,14 @@ class SparseMatrix public: /** \returns the number of rows of the matrix */ - inline StorageIndex rows() const { return IsRowMajor ? m_outerSize : m_innerSize; } + inline Index rows() const { return IsRowMajor ? m_outerSize : m_innerSize; } /** \returns the number of columns of the matrix */ - inline StorageIndex cols() const { return IsRowMajor ? m_innerSize : m_outerSize; } + inline Index cols() const { return IsRowMajor ? m_innerSize : m_outerSize; } /** \returns the number of rows (resp. columns) of the matrix if the storage order column major (resp. row major) */ - inline StorageIndex innerSize() const { return m_innerSize; } + inline Index innerSize() const { return m_innerSize; } /** \returns the number of columns (resp. rows) of the matrix if the storage order column major (resp. row major) */ - inline StorageIndex outerSize() const { return m_outerSize; } + inline Index outerSize() const { return m_outerSize; } /** \returns a const pointer to the array of values. * This function is aimed at interoperability with other libraries. @@ -253,7 +253,7 @@ class SparseMatrix } /** \returns the number of non zero coefficients */ - inline StorageIndex nonZeros() const + inline Index nonZeros() const { if(m_innerNonZeros) return innerNonZeros().sum(); @@ -299,7 +299,7 @@ class SparseMatrix { if(isCompressed()) { - std::size_t totalReserveSize = 0; + Index totalReserveSize = 0; // turn the matrix into non-compressed mode m_innerNonZeros = static_cast(std::malloc(m_outerSize * sizeof(StorageIndex))); if (!m_innerNonZeros) internal::throw_std_bad_alloc(); @@ -424,7 +424,7 @@ class SparseMatrix { if(isCompressed()) { - StorageIndex size = internal::convert_index(Index(m_data.size())); + StorageIndex size = internal::convert_index(m_data.size()); Index i = m_outerSize; // find the last filled column while (i>=0 && m_outerIndex[i]==0) @@ -605,8 +605,8 @@ class SparseMatrix */ void resize(Index rows, Index cols) { - const StorageIndex outerSize = convert_index(IsRowMajor ? rows : cols); - m_innerSize = convert_index(IsRowMajor ? cols : rows); + const Index outerSize = IsRowMajor ? rows : cols; + m_innerSize = IsRowMajor ? cols : rows; m_data.clear(); if (m_outerSize != outerSize || m_outerSize==0) { @@ -1069,7 +1069,7 @@ EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& Sparse { eigen_assert(!isCompressed()); - const StorageIndex outer = convert_index(IsRowMajor ? row : col); + const Index outer = IsRowMajor ? row : col; const StorageIndex inner = convert_index(IsRowMajor ? col : row); Index room = m_outerIndex[outer+1] - m_outerIndex[outer]; diff --git a/Eigen/src/SparseCore/SparseMatrixBase.h b/Eigen/src/SparseCore/SparseMatrixBase.h index c55a6a930..9039ebcec 100644 --- a/Eigen/src/SparseCore/SparseMatrixBase.h +++ b/Eigen/src/SparseCore/SparseMatrixBase.h @@ -144,15 +144,15 @@ template class SparseMatrixBase : public EigenBase #undef EIGEN_CURRENT_STORAGE_BASE_CLASS /** \returns the number of rows. \sa cols() */ - inline StorageIndex rows() const { return derived().rows(); } + inline Index rows() const { return derived().rows(); } /** \returns the number of columns. \sa rows() */ - inline StorageIndex cols() const { return derived().cols(); } + inline Index cols() const { return derived().cols(); } /** \returns the number of coefficients, which is \a rows()*cols(). * \sa rows(), cols(). */ - inline StorageIndex size() const { return rows() * cols(); } + inline Index size() const { return rows() * cols(); } /** \returns the number of nonzero coefficients which is in practice the number * of stored coefficients. */ - inline StorageIndex nonZeros() const { return derived().nonZeros(); } + inline Index nonZeros() const { return derived().nonZeros(); } /** \returns true if either the number of rows or the number of columns is equal to 1. * In other words, this function returns * \code rows()==1 || cols()==1 \endcode @@ -160,10 +160,10 @@ template class SparseMatrixBase : public EigenBase inline bool isVector() const { return rows()==1 || cols()==1; } /** \returns the size of the storage major dimension, * i.e., the number of columns for a columns major matrix, and the number of rows otherwise */ - StorageIndex outerSize() const { return (int(Flags)&RowMajorBit) ? this->rows() : this->cols(); } + Index outerSize() const { return (int(Flags)&RowMajorBit) ? this->rows() : this->cols(); } /** \returns the size of the inner dimension according to the storage order, * i.e., the number of rows for a columns major matrix, and the number of cols otherwise */ - StorageIndex innerSize() const { return (int(Flags)&RowMajorBit) ? this->cols() : this->rows(); } + Index innerSize() const { return (int(Flags)&RowMajorBit) ? this->cols() : this->rows(); } bool isRValue() const { return m_isRValue; } Derived& markAsRValue() { m_isRValue = true; return derived(); } diff --git a/Eigen/src/SparseCore/SparseSelfAdjointView.h b/Eigen/src/SparseCore/SparseSelfAdjointView.h index e13f98144..05be8e57c 100644 --- a/Eigen/src/SparseCore/SparseSelfAdjointView.h +++ b/Eigen/src/SparseCore/SparseSelfAdjointView.h @@ -58,8 +58,8 @@ template class SparseSelfAdjointView eigen_assert(rows()==cols() && "SelfAdjointView is only for squared matrices"); } - inline StorageIndex rows() const { return m_matrix.rows(); } - inline StorageIndex cols() const { return m_matrix.cols(); } + inline Index rows() const { return m_matrix.rows(); } + inline Index cols() const { return m_matrix.cols(); } /** \internal \returns a reference to the nested matrix */ const _MatrixTypeNested& matrix() const { return m_matrix; } @@ -530,8 +530,8 @@ class SparseSymmetricPermutationProduct : m_matrix(mat), m_perm(perm) {} - inline StorageIndex rows() const { return m_matrix.rows(); } - inline StorageIndex cols() const { return m_matrix.cols(); } + inline Index rows() const { return m_matrix.rows(); } + inline Index cols() const { return m_matrix.cols(); } template void evalTo(SparseMatrix& _dest) const diff --git a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h index 1384fbbff..3db01bf2d 100644 --- a/Eigen/src/SparseCore/SparseSparseProductWithPruning.h +++ b/Eigen/src/SparseCore/SparseSparseProductWithPruning.h @@ -25,8 +25,8 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r typedef typename remove_all::type::StorageIndex StorageIndex; // make sure to call innerSize/outerSize since we fake the storage order. - StorageIndex rows = lhs.innerSize(); - StorageIndex cols = rhs.outerSize(); + Index rows = lhs.innerSize(); + Index cols = rhs.outerSize(); //Index size = lhs.outerSize(); eigen_assert(lhs.outerSize() == rhs.innerSize()); @@ -39,7 +39,7 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r // the product of a rhs column with the lhs is X+Y where X is the average number of non zero // per column of the lhs. // Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs) - StorageIndex estimated_nnz_prod = lhs.nonZeros() + rhs.nonZeros(); + Index estimated_nnz_prod = lhs.nonZeros() + rhs.nonZeros(); // mimics a resizeByInnerOuter: if(ResultType::IsRowMajor) diff --git a/Eigen/src/SparseCore/SparseTranspose.h b/Eigen/src/SparseCore/SparseTranspose.h index 84413c374..45d9c6700 100644 --- a/Eigen/src/SparseCore/SparseTranspose.h +++ b/Eigen/src/SparseCore/SparseTranspose.h @@ -48,7 +48,7 @@ template class TransposeImpl protected: typedef internal::SparseTransposeImpl Base; public: - inline typename MatrixType::StorageIndex nonZeros() const { return Base::derived().nestedExpression().nonZeros(); } + inline Index nonZeros() const { return Base::derived().nestedExpression().nonZeros(); } }; namespace internal { @@ -61,7 +61,6 @@ struct unary_evaluator, IteratorBased> typedef typename evaluator::ReverseInnerIterator EvalReverseIterator; public: typedef Transpose XprType; - typedef typename XprType::StorageIndex StorageIndex; class InnerIterator : public EvalIterator { @@ -70,8 +69,8 @@ struct unary_evaluator, IteratorBased> : EvalIterator(unaryOp.m_argImpl,outer) {} - StorageIndex row() const { return EvalIterator::col(); } - StorageIndex col() const { return EvalIterator::row(); } + Index row() const { return EvalIterator::col(); } + Index col() const { return EvalIterator::row(); } }; class ReverseInnerIterator : public EvalReverseIterator @@ -81,8 +80,8 @@ struct unary_evaluator, IteratorBased> : EvalReverseIterator(unaryOp.m_argImpl,outer) {} - StorageIndex row() const { return EvalReverseIterator::col(); } - StorageIndex col() const { return EvalReverseIterator::row(); } + Index row() const { return EvalReverseIterator::col(); } + Index col() const { return EvalReverseIterator::row(); } }; enum { diff --git a/Eigen/src/SparseCore/SparseTriangularView.h b/Eigen/src/SparseCore/SparseTriangularView.h index 15bdbacb5..b5fbcbdde 100644 --- a/Eigen/src/SparseCore/SparseTriangularView.h +++ b/Eigen/src/SparseCore/SparseTriangularView.h @@ -64,7 +64,6 @@ template class TriangularViewImpl::InnerIterator : public MatrixTypeNestedCleaned::InnerIterator { typedef typename MatrixTypeNestedCleaned::InnerIterator Base; - typedef typename TriangularViewType::StorageIndex StorageIndex; public: EIGEN_STRONG_INLINE InnerIterator(const TriangularViewImpl& view, Index outer) @@ -102,9 +101,9 @@ class TriangularViewImpl::InnerIterator : public MatrixT return *this; } - inline StorageIndex row() const { return (MatrixType::Flags&RowMajorBit ? Base::outer() : this->index()); } - inline StorageIndex col() const { return (MatrixType::Flags&RowMajorBit ? this->index() : Base::outer()); } - inline StorageIndex index() const + inline Index row() const { return (MatrixType::Flags&RowMajorBit ? Base::outer() : this->index()); } + inline Index col() const { return (MatrixType::Flags&RowMajorBit ? this->index() : Base::outer()); } + inline Index index() const { if(HasUnitDiag && m_returnOne) return Base::outer(); else return Base::index(); @@ -134,7 +133,6 @@ template class TriangularViewImpl::ReverseInnerIterator : public MatrixTypeNestedCleaned::ReverseInnerIterator { typedef typename MatrixTypeNestedCleaned::ReverseInnerIterator Base; - typedef typename TriangularViewImpl::StorageIndex StorageIndex; public: EIGEN_STRONG_INLINE ReverseInnerIterator(const TriangularViewType& view, Index outer) @@ -150,8 +148,8 @@ class TriangularViewImpl::ReverseInnerIterator : public EIGEN_STRONG_INLINE ReverseInnerIterator& operator--() { Base::operator--(); return *this; } - inline StorageIndex row() const { return Base::row(); } - inline StorageIndex col() const { return Base::col(); } + inline Index row() const { return Base::row(); } + inline Index col() const { return Base::col(); } EIGEN_STRONG_INLINE operator bool() const { @@ -175,7 +173,6 @@ struct unary_evaluator, IteratorBased> protected: typedef typename XprType::Scalar Scalar; - typedef typename XprType::StorageIndex StorageIndex; typedef typename evaluator::InnerIterator EvalIterator; enum { SkipFirst = ((Mode&Lower) && !(ArgType::Flags&RowMajorBit)) @@ -246,9 +243,9 @@ public: } } -// inline StorageIndex row() const { return (ArgType::Flags&RowMajorBit ? Base::outer() : this->index()); } -// inline StorageIndex col() const { return (ArgType::Flags&RowMajorBit ? this->index() : Base::outer()); } - inline StorageIndex index() const +// inline Index row() const { return (ArgType::Flags&RowMajorBit ? Base::outer() : this->index()); } +// inline Index col() const { return (ArgType::Flags&RowMajorBit ? this->index() : Base::outer()); } + inline Index index() const { if(HasUnitDiag && m_returnOne) return Base::outer(); else return Base::index(); diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index fd70cf2bc..b1cc4df77 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -79,10 +79,10 @@ class SparseVector Options = _Options }; - EIGEN_STRONG_INLINE StorageIndex rows() const { return IsColVector ? m_size : 1; } - EIGEN_STRONG_INLINE StorageIndex cols() const { return IsColVector ? 1 : m_size; } - EIGEN_STRONG_INLINE StorageIndex innerSize() const { return m_size; } - EIGEN_STRONG_INLINE StorageIndex outerSize() const { return 1; } + EIGEN_STRONG_INLINE Index rows() const { return IsColVector ? m_size : 1; } + EIGEN_STRONG_INLINE Index cols() const { return IsColVector ? 1 : m_size; } + EIGEN_STRONG_INLINE Index innerSize() const { return m_size; } + EIGEN_STRONG_INLINE Index outerSize() const { return 1; } EIGEN_STRONG_INLINE const Scalar* valuePtr() const { return &m_data.value(0); } EIGEN_STRONG_INLINE Scalar* valuePtr() { return &m_data.value(0); } @@ -132,7 +132,7 @@ class SparseVector inline void setZero() { m_data.clear(); } /** \returns the number of non zero coefficients */ - inline StorageIndex nonZeros() const { return static_cast(m_data.size()); } + inline Index nonZeros() const { return m_data.size(); } inline void startVec(Index outer) { @@ -213,7 +213,7 @@ class SparseVector void resize(Index newSize) { - m_size = convert_index(newSize); + m_size = newSize; m_data.clear(); } @@ -353,7 +353,7 @@ protected: } Storage m_data; - StorageIndex m_size; + Index m_size; }; template @@ -361,14 +361,14 @@ class SparseVector::InnerIterator { public: explicit InnerIterator(const SparseVector& vec, Index outer=0) - : m_data(vec.m_data), m_id(0), m_end(convert_index(m_data.size())) + : m_data(vec.m_data), m_id(0), m_end(m_data.size()) { EIGEN_UNUSED_VARIABLE(outer); eigen_assert(outer==0); } explicit InnerIterator(const internal::CompressedStorage& data) - : m_data(data), m_id(0), m_end(convert_index(m_data.size())) + : m_data(data), m_id(0), m_end(m_data.size()) {} inline InnerIterator& operator++() { m_id++; return *this; } @@ -376,16 +376,16 @@ class SparseVector::InnerIterator inline Scalar value() const { return m_data.value(m_id); } inline Scalar& valueRef() { return const_cast(m_data.value(m_id)); } - inline StorageIndex index() const { return m_data.index(m_id); } - inline StorageIndex row() const { return IsColVector ? index() : 0; } - inline StorageIndex col() const { return IsColVector ? 0 : index(); } + inline Index index() const { return m_data.index(m_id); } + inline Index row() const { return IsColVector ? index() : 0; } + inline Index col() const { return IsColVector ? 0 : index(); } inline operator bool() const { return (m_id < m_end); } protected: const internal::CompressedStorage& m_data; - StorageIndex m_id; - const StorageIndex m_end; + Index m_id; + const Index m_end; private: // If you get here, then you're not using the right InnerIterator type, e.g.: // SparseMatrix A; @@ -398,14 +398,14 @@ class SparseVector::ReverseInnerIterator { public: explicit ReverseInnerIterator(const SparseVector& vec, Index outer=0) - : m_data(vec.m_data), m_id(convert_index(m_data.size())), m_start(0) + : m_data(vec.m_data), m_id(m_data.size()), m_start(0) { EIGEN_UNUSED_VARIABLE(outer); eigen_assert(outer==0); } explicit ReverseInnerIterator(const internal::CompressedStorage& data) - : m_data(data), m_id(convert_index(m_data.size())), m_start(0) + : m_data(data), m_id(m_data.size()), m_start(0) {} inline ReverseInnerIterator& operator--() { m_id--; return *this; } @@ -413,15 +413,15 @@ class SparseVector::ReverseInnerIterator inline Scalar value() const { return m_data.value(m_id-1); } inline Scalar& valueRef() { return const_cast(m_data.value(m_id-1)); } - inline StorageIndex index() const { return m_data.index(m_id-1); } - inline StorageIndex row() const { return IsColVector ? index() : 0; } - inline StorageIndex col() const { return IsColVector ? 0 : index(); } + inline Index index() const { return m_data.index(m_id-1); } + inline Index row() const { return IsColVector ? index() : 0; } + inline Index col() const { return IsColVector ? 0 : index(); } inline operator bool() const { return (m_id > m_start); } protected: const internal::CompressedStorage& m_data; - StorageIndex m_id; + Index m_id; const Index m_start; }; diff --git a/Eigen/src/SparseCore/SparseView.h b/Eigen/src/SparseCore/SparseView.h index e26016c16..d6042d970 100644 --- a/Eigen/src/SparseCore/SparseView.h +++ b/Eigen/src/SparseCore/SparseView.h @@ -40,11 +40,11 @@ public: RealScalar m_epsilon = NumTraits::dummy_precision()) : m_matrix(mat), m_reference(m_reference), m_epsilon(m_epsilon) {} - inline StorageIndex rows() const { return m_matrix.rows(); } - inline StorageIndex cols() const { return m_matrix.cols(); } + inline Index rows() const { return m_matrix.rows(); } + inline Index cols() const { return m_matrix.cols(); } - inline StorageIndex innerSize() const { return m_matrix.innerSize(); } - inline StorageIndex outerSize() const { return m_matrix.outerSize(); } + inline Index innerSize() const { return m_matrix.innerSize(); } + inline Index outerSize() const { return m_matrix.outerSize(); } /** \returns the nested expression */ const typename internal::remove_all::type& @@ -153,17 +153,17 @@ struct unary_evaluator, IndexBased> : m_sve.m_argImpl.coeff(m_inner, m_outer); } - EIGEN_STRONG_INLINE StorageIndex index() const { return m_inner; } - inline StorageIndex row() const { return IsRowMajor ? m_outer : index(); } - inline StorageIndex col() const { return IsRowMajor ? index() : m_outer; } + EIGEN_STRONG_INLINE Index index() const { return m_inner; } + inline Index row() const { return IsRowMajor ? m_outer : index(); } + inline Index col() const { return IsRowMajor ? index() : m_outer; } EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; } protected: const unary_evaluator &m_sve; - StorageIndex m_inner; - const StorageIndex m_outer; - const StorageIndex m_end; + Index m_inner; + const Index m_outer; + const Index m_end; private: void incrementToNonZero() diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 380ba25c0..f60e4ac9d 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -122,8 +122,8 @@ class SparseLU : public SparseSolverBase >, factorize(matrix); } - inline StorageIndex rows() const { return m_mat.rows(); } - inline StorageIndex cols() const { return m_mat.cols(); } + inline Index rows() const { return m_mat.rows(); } + inline Index cols() const { return m_mat.cols(); } /** Indicate that the pattern of the input matrix is symmetric */ void isSymmetric(bool sym) { @@ -334,10 +334,10 @@ class SparseLU : public SparseSolverBase >, // SparseLU options bool m_symmetricmode; // values for performance - internal::perfvalues m_perfv; + internal::perfvalues m_perfv; RealScalar m_diagpivotthresh; // Specifies the threshold used for a diagonal entry to be an acceptable pivot - StorageIndex m_nnzL, m_nnzU; // Nonzeros in L and U factors - StorageIndex m_detPermR; // Determinant of the coefficient matrix + Index m_nnzL, m_nnzU; // Nonzeros in L and U factors + Index m_detPermR; // Determinant of the coefficient matrix private: // Disable copy constructor SparseLU (const SparseLU& ); @@ -449,7 +449,7 @@ void SparseLU::factorize(const MatrixType& matrix) eigen_assert(m_analysisIsOk && "analyzePattern() should be called first"); eigen_assert((matrix.rows() == matrix.cols()) && "Only for squared matrices"); - typedef typename IndexVector::Scalar Index; + typedef typename IndexVector::Scalar StorageIndex; m_isInitialized = true; @@ -461,11 +461,11 @@ void SparseLU::factorize(const MatrixType& matrix) { m_mat.uncompress(); //NOTE: The effect of this command is only to create the InnerNonzeros pointers. //Then, permute only the column pointers - const Index * outerIndexPtr; + const StorageIndex * outerIndexPtr; if (matrix.isCompressed()) outerIndexPtr = matrix.outerIndexPtr(); else { - Index* outerIndexPtr_t = new Index[matrix.cols()+1]; + StorageIndex* outerIndexPtr_t = new StorageIndex[matrix.cols()+1]; for(Index i = 0; i <= matrix.cols(); i++) outerIndexPtr_t[i] = m_mat.outerIndexPtr()[i]; outerIndexPtr = outerIndexPtr_t; } @@ -649,12 +649,11 @@ void SparseLU::factorize(const MatrixType& matrix) template struct SparseLUMatrixLReturnType : internal::no_assignment_operator { - typedef typename MappedSupernodalType::StorageIndex StorageIndex; typedef typename MappedSupernodalType::Scalar Scalar; explicit SparseLUMatrixLReturnType(const MappedSupernodalType& mapL) : m_mapL(mapL) { } - StorageIndex rows() { return m_mapL.rows(); } - StorageIndex cols() { return m_mapL.cols(); } + Index rows() { return m_mapL.rows(); } + Index cols() { return m_mapL.cols(); } template void solveInPlace( MatrixBase &X) const { @@ -666,13 +665,12 @@ struct SparseLUMatrixLReturnType : internal::no_assignment_operator template struct SparseLUMatrixUReturnType : internal::no_assignment_operator { - typedef typename MatrixLType::StorageIndex StorageIndex; typedef typename MatrixLType::Scalar Scalar; explicit SparseLUMatrixUReturnType(const MatrixLType& mapL, const MatrixUType& mapU) : m_mapL(mapL),m_mapU(mapU) { } - StorageIndex rows() { return m_mapL.rows(); } - StorageIndex cols() { return m_mapL.cols(); } + Index rows() { return m_mapL.rows(); } + Index cols() { return m_mapL.cols(); } template void solveInPlace(MatrixBase &X) const { diff --git a/Eigen/src/SparseLU/SparseLUImpl.h b/Eigen/src/SparseLU/SparseLUImpl.h index 14d70897d..e735fd5c8 100644 --- a/Eigen/src/SparseLU/SparseLUImpl.h +++ b/Eigen/src/SparseLU/SparseLUImpl.h @@ -16,17 +16,17 @@ namespace internal { * \class SparseLUImpl * Base class for sparseLU */ -template +template class SparseLUImpl { public: typedef Matrix ScalarVector; - typedef Matrix IndexVector; + typedef Matrix IndexVector; typedef typename ScalarVector::RealScalar RealScalar; typedef Ref > BlockScalarVector; - typedef Ref > BlockIndexVector; + typedef Ref > BlockIndexVector; typedef LU_GlobalLU_t GlobalLU_t; - typedef SparseMatrix MatrixType; + typedef SparseMatrix MatrixType; protected: template diff --git a/Eigen/src/SparseLU/SparseLU_Memory.h b/Eigen/src/SparseLU/SparseLU_Memory.h index 1ffa7d54e..1cf8bebc7 100644 --- a/Eigen/src/SparseLU/SparseLU_Memory.h +++ b/Eigen/src/SparseLU/SparseLU_Memory.h @@ -36,13 +36,12 @@ namespace internal { enum { LUNoMarker = 3 }; enum {emptyIdxLU = -1}; -template inline Index LUnumTempV(Index& m, Index& w, Index& t, Index& b) { return (std::max)(m, (t+b)*w); } -template< typename Scalar, typename Index> +template< typename Scalar> inline Index LUTempSpace(Index&m, Index& w) { return (2*w + 4 + LUNoMarker) * m * sizeof(Index) + (w + 1) * m * sizeof(Scalar); @@ -59,9 +58,9 @@ inline Index LUTempSpace(Index&m, Index& w) * \param keep_prev 1: use length and do not expand the vector; 0: compute new_len and expand * \param[in,out] num_expansions Number of times the memory has been expanded */ -template +template template -Index SparseLUImpl::expand(VectorType& vec, Index& length, Index nbElts, Index keep_prev, Index& num_expansions) +Index SparseLUImpl::expand(VectorType& vec, Index& length, Index nbElts, Index keep_prev, Index& num_expansions) { float alpha = 1.5; // Ratio of the memory increase @@ -148,8 +147,8 @@ Index SparseLUImpl::expand(VectorType& vec, Index& length, Index * \return an estimated size of the required memory if lwork = -1; otherwise, return the size of actually allocated memory when allocation failed, and 0 on success * \note Unlike SuperLU, this routine does not support successive factorization with the same pattern and the same row permutation */ -template -Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lwork, Index fillratio, Index panel_size, GlobalLU_t& glu) +template +Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lwork, Index fillratio, Index panel_size, GlobalLU_t& glu) { Index& num_expansions = glu.num_expansions; //No memory expansions so far num_expansions = 0; @@ -205,9 +204,9 @@ Index SparseLUImpl::memInit(Index m, Index n, Index annz, Index lw * \param num_expansions Number of expansions * \return 0 on success, > 0 size of the memory allocated so far */ -template +template template -Index SparseLUImpl::memXpand(VectorType& vec, Index& maxlen, Index nbElts, MemType memtype, Index& num_expansions) +Index SparseLUImpl::memXpand(VectorType& vec, Index& maxlen, Index nbElts, MemType memtype, Index& num_expansions) { Index failed_size; if (memtype == USUB) diff --git a/Eigen/src/SparseLU/SparseLU_Structs.h b/Eigen/src/SparseLU/SparseLU_Structs.h index 24d6bf179..cf5ec449b 100644 --- a/Eigen/src/SparseLU/SparseLU_Structs.h +++ b/Eigen/src/SparseLU/SparseLU_Structs.h @@ -75,7 +75,7 @@ typedef enum {LUSUP, UCOL, LSUB, USUB, LLVL, ULVL} MemType; template struct LU_GlobalLU_t { - typedef typename IndexVector::Scalar Index; + typedef typename IndexVector::Scalar StorageIndex; IndexVector xsup; //First supernode column ... xsup(s) points to the beginning of the s-th supernode IndexVector supno; // Supernode number corresponding to this column (column to supernode mapping) ScalarVector lusup; // nonzero values of L ordered by columns @@ -93,7 +93,6 @@ struct LU_GlobalLU_t { }; // Values to set for performance -template struct perfvalues { Index panel_size; // a panel consists of at most consecutive columns Index relax; // To control degree of relaxing supernodes. If the number of nodes (columns) diff --git a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h index 098763765..f7ffc2d9c 100644 --- a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h +++ b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h @@ -42,7 +42,7 @@ class MappedSuperNodalMatrix { } - MappedSuperNodalMatrix(StorageIndex m, StorageIndex n, ScalarVector& nzval, IndexVector& nzval_colptr, IndexVector& rowind, + MappedSuperNodalMatrix(Index m, Index n, ScalarVector& nzval, IndexVector& nzval_colptr, IndexVector& rowind, IndexVector& rowind_colptr, IndexVector& col_to_sup, IndexVector& sup_to_col ) { setInfos(m, n, nzval, nzval_colptr, rowind, rowind_colptr, col_to_sup, sup_to_col); @@ -58,7 +58,7 @@ class MappedSuperNodalMatrix * FIXME This class will be modified such that it can be use in the course * of the factorization. */ - void setInfos(StorageIndex m, StorageIndex n, ScalarVector& nzval, IndexVector& nzval_colptr, IndexVector& rowind, + void setInfos(Index m, Index n, ScalarVector& nzval, IndexVector& nzval_colptr, IndexVector& rowind, IndexVector& rowind_colptr, IndexVector& col_to_sup, IndexVector& sup_to_col ) { m_row = m; @@ -75,12 +75,12 @@ class MappedSuperNodalMatrix /** * Number of rows */ - StorageIndex rows() { return m_row; } + Index rows() { return m_row; } /** * Number of columns */ - StorageIndex cols() { return m_col; } + Index cols() { return m_col; } /** * Return the array of nonzero values packed by column @@ -148,7 +148,7 @@ class MappedSuperNodalMatrix /** * Return the number of supernodes */ - StorageIndex nsuper() const + Index nsuper() const { return m_nsuper; } @@ -161,9 +161,9 @@ class MappedSuperNodalMatrix protected: - StorageIndex m_row; // Number of rows - StorageIndex m_col; // Number of columns - StorageIndex m_nsuper; // Number of supernodes + Index m_row; // Number of rows + Index m_col; // Number of columns + Index m_nsuper; // Number of supernodes Scalar* m_nzval; //array of nonzero values packed by column StorageIndex* m_nzval_colptr; //nzval_colptr[j] Stores the location in nzval[] which starts column j StorageIndex* m_rowind; // Array of compressed row indices of rectangular supernodes @@ -184,7 +184,7 @@ class MappedSuperNodalMatrix::InnerIterator public: InnerIterator(const MappedSuperNodalMatrix& mat, Eigen::Index outer) : m_matrix(mat), - m_outer(convert_index(outer)), + m_outer(outer), m_supno(mat.colToSup()[outer]), m_idval(mat.colIndexPtr()[outer]), m_startidval(m_idval), diff --git a/Eigen/src/SparseLU/SparseLU_Utils.h b/Eigen/src/SparseLU/SparseLU_Utils.h index 15352ac33..b48157d9f 100644 --- a/Eigen/src/SparseLU/SparseLU_Utils.h +++ b/Eigen/src/SparseLU/SparseLU_Utils.h @@ -17,8 +17,8 @@ namespace internal { /** * \brief Count Nonzero elements in the factors */ -template -void SparseLUImpl::countnz(const Index n, Index& nnzL, Index& nnzU, GlobalLU_t& glu) +template +void SparseLUImpl::countnz(const Index n, Index& nnzL, Index& nnzU, GlobalLU_t& glu) { nnzL = 0; nnzU = (glu.xusub)(n); @@ -48,8 +48,8 @@ void SparseLUImpl::countnz(const Index n, Index& nnzL, Index& nnzU * and applies permutation to the remaining subscripts * */ -template -void SparseLUImpl::fixupL(const Index n, const IndexVector& perm_r, GlobalLU_t& glu) +template +void SparseLUImpl::fixupL(const Index n, const IndexVector& perm_r, GlobalLU_t& glu) { Index fsupc, i, j, k, jstart; diff --git a/Eigen/src/SparseLU/SparseLU_column_bmod.h b/Eigen/src/SparseLU/SparseLU_column_bmod.h index f24bd87d3..bda01dcb3 100644 --- a/Eigen/src/SparseLU/SparseLU_column_bmod.h +++ b/Eigen/src/SparseLU/SparseLU_column_bmod.h @@ -49,8 +49,9 @@ namespace internal { * > 0 - number of bytes allocated when run out of space * */ -template -Index SparseLUImpl::column_bmod(const Index jcol, const Index nseg, BlockScalarVector dense, ScalarVector& tempv, BlockIndexVector segrep, BlockIndexVector repfnz, Index fpanelc, GlobalLU_t& glu) +template +Index SparseLUImpl::column_bmod(const Index jcol, const Index nseg, BlockScalarVector dense, ScalarVector& tempv, + BlockIndexVector segrep, BlockIndexVector repfnz, Index fpanelc, GlobalLU_t& glu) { Index jsupno, k, ksub, krep, ksupno; Index lptr, nrow, isub, irow, nextlu, new_next, ufirst; diff --git a/Eigen/src/SparseLU/SparseLU_column_dfs.h b/Eigen/src/SparseLU/SparseLU_column_dfs.h index 4c04b0e44..17c9e6adb 100644 --- a/Eigen/src/SparseLU/SparseLU_column_dfs.h +++ b/Eigen/src/SparseLU/SparseLU_column_dfs.h @@ -30,7 +30,7 @@ #ifndef SPARSELU_COLUMN_DFS_H #define SPARSELU_COLUMN_DFS_H -template class SparseLUImpl; +template class SparseLUImpl; namespace Eigen { namespace internal { @@ -39,8 +39,8 @@ template struct column_dfs_traits : no_assignment_operator { typedef typename ScalarVector::Scalar Scalar; - typedef typename IndexVector::Scalar Index; - column_dfs_traits(Index jcol, Index& jsuper, typename SparseLUImpl::GlobalLU_t& glu, SparseLUImpl& luImpl) + typedef typename IndexVector::Scalar StorageIndex; + column_dfs_traits(Index jcol, Index& jsuper, typename SparseLUImpl::GlobalLU_t& glu, SparseLUImpl& luImpl) : m_jcol(jcol), m_jsuper_ref(jsuper), m_glu(glu), m_luImpl(luImpl) {} bool update_segrep(Index /*krep*/, Index /*jj*/) @@ -57,8 +57,8 @@ struct column_dfs_traits : no_assignment_operator Index m_jcol; Index& m_jsuper_ref; - typename SparseLUImpl::GlobalLU_t& m_glu; - SparseLUImpl& m_luImpl; + typename SparseLUImpl::GlobalLU_t& m_glu; + SparseLUImpl& m_luImpl; }; @@ -89,8 +89,10 @@ struct column_dfs_traits : no_assignment_operator * > 0 number of bytes allocated when run out of space * */ -template -Index SparseLUImpl::column_dfs(const Index m, const Index jcol, IndexVector& perm_r, Index maxsuper, Index& nseg, BlockIndexVector lsub_col, IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune, IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu) +template +Index SparseLUImpl::column_dfs(const Index m, const Index jcol, IndexVector& perm_r, Index maxsuper, Index& nseg, + BlockIndexVector lsub_col, IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune, + IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu) { Index jsuper = glu.supno(jcol); diff --git a/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h b/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h index 170610d9f..bf237951d 100644 --- a/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h +++ b/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h @@ -46,8 +46,9 @@ namespace internal { * > 0 - number of bytes allocated when run out of space * */ -template -Index SparseLUImpl::copy_to_ucol(const Index jcol, const Index nseg, IndexVector& segrep, BlockIndexVector repfnz ,IndexVector& perm_r, BlockScalarVector dense, GlobalLU_t& glu) +template +Index SparseLUImpl::copy_to_ucol(const Index jcol, const Index nseg, IndexVector& segrep, + BlockIndexVector repfnz ,IndexVector& perm_r, BlockScalarVector dense, GlobalLU_t& glu) { Index ksub, krep, ksupno; diff --git a/Eigen/src/SparseLU/SparseLU_gemm_kernel.h b/Eigen/src/SparseLU/SparseLU_gemm_kernel.h index 9e4e3e72b..7420b4d17 100644 --- a/Eigen/src/SparseLU/SparseLU_gemm_kernel.h +++ b/Eigen/src/SparseLU/SparseLU_gemm_kernel.h @@ -21,7 +21,7 @@ namespace internal { * - lda and ldc must be multiples of the respective packet size * - C must have the same alignment as A */ -template +template EIGEN_DONT_INLINE void sparselu_gemm(Index m, Index n, Index d, const Scalar* A, Index lda, const Scalar* B, Index ldb, Scalar* C, Index ldc) { diff --git a/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h b/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h index 7a4e4305a..4092f842f 100644 --- a/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h +++ b/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h @@ -42,8 +42,8 @@ namespace internal { * \param descendants Number of descendants of each node in the etree * \param relax_end last column in a supernode */ -template -void SparseLUImpl::heap_relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end) +template +void SparseLUImpl::heap_relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end) { // The etree may not be postordered, but its heap ordered @@ -75,7 +75,7 @@ void SparseLUImpl::heap_relax_snode (const Index n, IndexVector& e } // Identify the relaxed supernodes by postorder traversal of the etree Index snode_start; // beginning of a snode - Index k; + StorageIndex k; Index nsuper_et_post = 0; // Number of relaxed snodes in postordered etree Index nsuper_et = 0; // Number of relaxed snodes in the original etree Index l; diff --git a/Eigen/src/SparseLU/SparseLU_kernel_bmod.h b/Eigen/src/SparseLU/SparseLU_kernel_bmod.h index cad149ded..9513f8369 100644 --- a/Eigen/src/SparseLU/SparseLU_kernel_bmod.h +++ b/Eigen/src/SparseLU/SparseLU_kernel_bmod.h @@ -30,13 +30,13 @@ namespace internal { */ template struct LU_kernel_bmod { - template + template static EIGEN_DONT_INLINE void run(const Index segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros); }; template -template +template EIGEN_DONT_INLINE void LU_kernel_bmod::run(const Index segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros) { @@ -91,21 +91,22 @@ EIGEN_DONT_INLINE void LU_kernel_bmod::run(const Index seg template <> struct LU_kernel_bmod<1> { - template + template static EIGEN_DONT_INLINE void run(const Index /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros); }; -template +template EIGEN_DONT_INLINE void LU_kernel_bmod<1>::run(const Index /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr, const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros) { typedef typename ScalarVector::Scalar Scalar; + typedef typename IndexVector::Scalar StorageIndex; Scalar f = dense(lsub(lptr + no_zeros)); luptr += lda * no_zeros + no_zeros + 1; const Scalar* a(lusup.data() + luptr); - const /*typename IndexVector::Scalar*/Index* irow(lsub.data()+lptr + no_zeros + 1); + const StorageIndex* irow(lsub.data()+lptr + no_zeros + 1); Index i = 0; for (; i+1 < nrow; i+=2) { diff --git a/Eigen/src/SparseLU/SparseLU_panel_bmod.h b/Eigen/src/SparseLU/SparseLU_panel_bmod.h index da0e0fc3c..bd3cf87b9 100644 --- a/Eigen/src/SparseLU/SparseLU_panel_bmod.h +++ b/Eigen/src/SparseLU/SparseLU_panel_bmod.h @@ -52,8 +52,8 @@ namespace internal { * * */ -template -void SparseLUImpl::panel_bmod(const Index m, const Index w, const Index jcol, +template +void SparseLUImpl::panel_bmod(const Index m, const Index w, const Index jcol, const Index nseg, ScalarVector& dense, ScalarVector& tempv, IndexVector& segrep, IndexVector& repfnz, GlobalLU_t& glu) { diff --git a/Eigen/src/SparseLU/SparseLU_panel_dfs.h b/Eigen/src/SparseLU/SparseLU_panel_dfs.h index dc0054efd..f4a908ee5 100644 --- a/Eigen/src/SparseLU/SparseLU_panel_dfs.h +++ b/Eigen/src/SparseLU/SparseLU_panel_dfs.h @@ -37,8 +37,8 @@ namespace internal { template struct panel_dfs_traits { - typedef typename IndexVector::Scalar Index; - panel_dfs_traits(Index jcol, Index* marker) + typedef typename IndexVector::Scalar StorageIndex; + panel_dfs_traits(Index jcol, StorageIndex* marker) : m_jcol(jcol), m_marker(marker) {} bool update_segrep(Index krep, Index jj) @@ -53,13 +53,13 @@ struct panel_dfs_traits void mem_expand(IndexVector& /*glu.lsub*/, Index /*nextl*/, Index /*chmark*/) {} enum { ExpandMem = false }; Index m_jcol; - Index* m_marker; + StorageIndex* m_marker; }; -template +template template -void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& perm_r, +void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& perm_r, Index& nseg, IndexVector& panel_lsub, IndexVector& segrep, Ref repfnz_col, IndexVector& xprune, Ref marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu, @@ -215,8 +215,8 @@ void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& perm_r, * */ -template -void SparseLUImpl::panel_dfs(const Index m, const Index w, const Index jcol, MatrixType& A, IndexVector& perm_r, Index& nseg, ScalarVector& dense, IndexVector& panel_lsub, IndexVector& segrep, IndexVector& repfnz, IndexVector& xprune, IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu) +template +void SparseLUImpl::panel_dfs(const Index m, const Index w, const Index jcol, MatrixType& A, IndexVector& perm_r, Index& nseg, ScalarVector& dense, IndexVector& panel_lsub, IndexVector& segrep, IndexVector& repfnz, IndexVector& xprune, IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu) { Index nextl_col; // Next available position in panel_lsub[*,jj] diff --git a/Eigen/src/SparseLU/SparseLU_pivotL.h b/Eigen/src/SparseLU/SparseLU_pivotL.h index 457789c78..01f5ba4e9 100644 --- a/Eigen/src/SparseLU/SparseLU_pivotL.h +++ b/Eigen/src/SparseLU/SparseLU_pivotL.h @@ -56,8 +56,8 @@ namespace internal { * \return 0 if success, i > 0 if U(i,i) is exactly zero * */ -template -Index SparseLUImpl::pivotL(const Index jcol, const RealScalar& diagpivotthresh, IndexVector& perm_r, IndexVector& iperm_c, Index& pivrow, GlobalLU_t& glu) +template +Index SparseLUImpl::pivotL(const Index jcol, const RealScalar& diagpivotthresh, IndexVector& perm_r, IndexVector& iperm_c, Index& pivrow, GlobalLU_t& glu) { Index fsupc = (glu.xsup)((glu.supno)(jcol)); // First column in the supernode containing the column jcol @@ -67,7 +67,7 @@ Index SparseLUImpl::pivotL(const Index jcol, const RealScalar& dia Index lda = glu.xlusup(fsupc+1) - glu.xlusup(fsupc); // leading dimension Scalar* lu_sup_ptr = &(glu.lusup.data()[glu.xlusup(fsupc)]); // Start of the current supernode Scalar* lu_col_ptr = &(glu.lusup.data()[glu.xlusup(jcol)]); // Start of jcol in the supernode - Index* lsub_ptr = &(glu.lsub.data()[lptr]); // Start of row indices of the supernode + StorageIndex* lsub_ptr = &(glu.lsub.data()[lptr]); // Start of row indices of the supernode // Determine the largest abs numerical value for partial pivoting Index diagind = iperm_c(jcol); // diagonal index diff --git a/Eigen/src/SparseLU/SparseLU_pruneL.h b/Eigen/src/SparseLU/SparseLU_pruneL.h index 66460d168..13133fcc2 100644 --- a/Eigen/src/SparseLU/SparseLU_pruneL.h +++ b/Eigen/src/SparseLU/SparseLU_pruneL.h @@ -49,8 +49,9 @@ namespace internal { * \param glu Global LU data * */ -template -void SparseLUImpl::pruneL(const Index jcol, const IndexVector& perm_r, const Index pivrow, const Index nseg, const IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune, GlobalLU_t& glu) +template +void SparseLUImpl::pruneL(const Index jcol, const IndexVector& perm_r, const Index pivrow, const Index nseg, + const IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune, GlobalLU_t& glu) { // For each supernode-rep irep in U(*,j] Index jsupno = glu.supno(jcol); diff --git a/Eigen/src/SparseLU/SparseLU_relax_snode.h b/Eigen/src/SparseLU/SparseLU_relax_snode.h index 58ec32e27..21c182d56 100644 --- a/Eigen/src/SparseLU/SparseLU_relax_snode.h +++ b/Eigen/src/SparseLU/SparseLU_relax_snode.h @@ -43,8 +43,8 @@ namespace internal { * \param descendants Number of descendants of each node in the etree * \param relax_end last column in a supernode */ -template -void SparseLUImpl::relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end) +template +void SparseLUImpl::relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end) { // compute the number of descendants of each node in the etree diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 58bfc1cb4..920b884e5 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -109,11 +109,11 @@ class SparseQR : public SparseSolverBase > /** \returns the number of rows of the represented matrix. */ - inline StorageIndex rows() const { return m_pmat.rows(); } + inline Index rows() const { return m_pmat.rows(); } /** \returns the number of columns of the represented matrix. */ - inline StorageIndex cols() const { return m_pmat.cols();} + inline Index cols() const { return m_pmat.cols();} /** \returns a const reference to the \b sparse upper triangular matrix R of the QR factorization. */ @@ -123,7 +123,7 @@ class SparseQR : public SparseSolverBase > * * \sa setPivotThreshold() */ - StorageIndex rank() const + Index rank() const { eigen_assert(m_isInitialized && "The factorization should be called first, use compute()"); return m_nonzeropivots; @@ -260,7 +260,7 @@ class SparseQR : public SparseSolverBase > PermutationType m_outputPerm_c; // The final column permutation RealScalar m_threshold; // Threshold to determine null Householder reflections bool m_useDefaultThreshold; // Use default threshold - StorageIndex m_nonzeropivots; // Number of non zero pivots found + Index m_nonzeropivots; // Number of non zero pivots found IndexVector m_etree; // Column elimination tree IndexVector m_firstRowElt; // First element in each row bool m_isQSorted; // whether Q is sorted or not @@ -289,9 +289,9 @@ void SparseQR::analyzePattern(const MatrixType& mat) // Compute the column fill reducing ordering OrderingType ord; ord(matCpy, m_perm_c); - StorageIndex n = mat.cols(); - StorageIndex m = mat.rows(); - StorageIndex diagSize = (std::min)(m,n); + Index n = mat.cols(); + Index m = mat.rows(); + Index diagSize = (std::min)(m,n); if (!m_perm_c.size()) { @@ -327,9 +327,9 @@ void SparseQR::factorize(const MatrixType& mat) using std::abs; eigen_assert(m_analysisIsok && "analyzePattern() should be called before this step"); - Index m = mat.rows(); - Index n = mat.cols(); - Index diagSize = (std::min)(m,n); + StorageIndex m = mat.rows(); + StorageIndex n = mat.cols(); + StorageIndex diagSize = (std::min)(m,n); IndexVector mark((std::max)(m,n)); mark.setConstant(-1); // Record the visited nodes IndexVector Ridx(n), Qidx(m); // Store temporarily the row indexes for the current column of R and Q Index nzcolR, nzcolQ; // Number of nonzero for the current column of R and Q @@ -578,7 +578,6 @@ struct SparseQR_QProduct : ReturnByValue struct SparseQRMatrixQReturnType : public EigenBase > { - typedef typename SparseQRType::StorageIndex StorageIndex; typedef typename SparseQRType::Scalar Scalar; typedef Matrix DenseMatrix; explicit SparseQRMatrixQReturnType(const SparseQRType& qr) : m_qr(qr) {} @@ -647,8 +645,8 @@ struct SparseQRMatrixQReturnType : public EigenBase(m_qr); } - inline StorageIndex rows() const { return m_qr.rows(); } - inline StorageIndex cols() const { return (std::min)(m_qr.rows(),m_qr.cols()); } + inline Index rows() const { return m_qr.rows(); } + inline Index cols() const { return (std::min)(m_qr.rows(),m_qr.cols()); } // To use for operations with the transpose of Q SparseQRMatrixQTransposeReturnType transpose() const { diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h index 3a9b5fd74..8779eb74c 100644 --- a/Eigen/src/SuperLUSupport/SuperLUSupport.h +++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h @@ -313,8 +313,8 @@ class SuperLUBase : public SparseSolverBase clearFactors(); } - inline StorageIndex rows() const { return m_matrix.rows(); } - inline StorageIndex cols() const { return m_matrix.cols(); } + inline Index rows() const { return m_matrix.rows(); } + inline Index cols() const { return m_matrix.cols(); } /** \returns a reference to the Super LU option object to configure the Super LU algorithms. */ inline superlu_options_t& options() { return m_sluOptions; } @@ -616,7 +616,7 @@ void SuperLU::_solve_impl(const MatrixBase &b, MatrixBase { eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or analyzePattern()/factorize()"); - const StorageIndex size = m_matrix.rows(); + const Index size = m_matrix.rows(); const Index rhsCols = b.cols(); eigen_assert(size==b.rows()); diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h index 47e8b6304..dcbd4ab71 100644 --- a/Eigen/src/UmfPackSupport/UmfPackSupport.h +++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h @@ -164,8 +164,8 @@ class UmfPackLU : public SparseSolverBase > if(m_numeric) umfpack_free_numeric(&m_numeric,Scalar()); } - inline StorageIndex rows() const { return m_copyMatrix.rows(); } - inline StorageIndex cols() const { return m_copyMatrix.cols(); } + inline Index rows() const { return m_copyMatrix.rows(); } + inline Index cols() const { return m_copyMatrix.cols(); } /** \brief Reports whether previous computation was successful. * diff --git a/test/nullary.cpp b/test/nullary.cpp index 8344855df..2c148e205 100644 --- a/test/nullary.cpp +++ b/test/nullary.cpp @@ -12,7 +12,6 @@ template bool equalsIdentity(const MatrixType& A) { - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; Scalar zero = static_cast(0); @@ -35,7 +34,6 @@ bool equalsIdentity(const MatrixType& A) template void testVectorType(const VectorType& base) { - typedef typename VectorType::Index Index; typedef typename VectorType::Scalar Scalar; const Index size = base.size(); @@ -104,7 +102,6 @@ void testVectorType(const VectorType& base) template void testMatrixType(const MatrixType& m) { - typedef typename MatrixType::Index Index; const Index rows = m.rows(); const Index cols = m.cols(); diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index 8fd759c93..b06956974 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -16,8 +16,8 @@ template void sparse_basic(const SparseMatrixType& re typedef typename SparseMatrixType::StorageIndex StorageIndex; typedef Matrix Vector2; - const StorageIndex rows = ref.rows(); - const StorageIndex cols = ref.cols(); + const Index rows = ref.rows(); + const Index cols = ref.cols(); const Index inner = ref.innerSize(); const Index outer = ref.outerSize(); diff --git a/test/sparse_product.cpp b/test/sparse_product.cpp index b3f653d0e..480a660fc 100644 --- a/test/sparse_product.cpp +++ b/test/sparse_product.cpp @@ -11,7 +11,7 @@ template void sparse_product() { - typedef typename SparseMatrixType::StorageIndex Index; + typedef typename SparseMatrixType::StorageIndex StorageIndex; Index n = 100; const Index rows = internal::random(1,n); const Index cols = internal::random(1,n); @@ -23,8 +23,8 @@ template void sparse_product() typedef Matrix DenseMatrix; typedef Matrix DenseVector; typedef Matrix RowDenseVector; - typedef SparseVector ColSpVector; - typedef SparseVector RowSpVector; + typedef SparseVector ColSpVector; + typedef SparseVector RowSpVector; Scalar s1 = internal::random(); Scalar s2 = internal::random(); diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index 6e847e110..873f2bf2a 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -271,7 +271,6 @@ public: using Base::_solve_impl; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::StorageIndex StorageIndex; typedef typename MatrixType::RealScalar RealScalar; typedef _Preconditioner Preconditioner; diff --git a/unsupported/Eigen/src/IterativeSolvers/MINRES.h b/unsupported/Eigen/src/IterativeSolvers/MINRES.h index 2845b9cfd..ea8b73d38 100644 --- a/unsupported/Eigen/src/IterativeSolvers/MINRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/MINRES.h @@ -207,7 +207,6 @@ namespace Eigen { using Base::_solve_impl; typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::StorageIndex StorageIndex; typedef typename MatrixType::RealScalar RealScalar; typedef _Preconditioner Preconditioner; diff --git a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h index d0b51970d..4406437cc 100644 --- a/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h +++ b/unsupported/Eigen/src/KroneckerProduct/KroneckerTensorProduct.h @@ -31,7 +31,6 @@ class KroneckerProductBase : public ReturnByValue protected: typedef typename Traits::Lhs Lhs; typedef typename Traits::Rhs Rhs; - typedef typename Traits::StorageIndex StorageIndex; public: /*! \brief Constructor. */ @@ -39,8 +38,8 @@ class KroneckerProductBase : public ReturnByValue : m_A(A), m_B(B) {} - inline StorageIndex rows() const { return m_A.rows() * m_B.rows(); } - inline StorageIndex cols() const { return m_A.cols() * m_B.cols(); } + inline Index rows() const { return m_A.rows() * m_B.rows(); } + inline Index cols() const { return m_A.cols() * m_B.cols(); } /*! * This overrides ReturnByValue::coeff because this function is diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LMcovar.h b/unsupported/Eigen/src/LevenbergMarquardt/LMcovar.h index 32d3ad518..b75bea25f 100644 --- a/unsupported/Eigen/src/LevenbergMarquardt/LMcovar.h +++ b/unsupported/Eigen/src/LevenbergMarquardt/LMcovar.h @@ -23,7 +23,6 @@ void covar( Scalar tol = std::sqrt(NumTraits::epsilon()) ) { using std::abs; - typedef DenseIndex Index; /* Local variables */ Index i, j, k, l, ii, jj; bool sing; diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LMpar.h b/unsupported/Eigen/src/LevenbergMarquardt/LMpar.h index 9532042d9..731862341 100644 --- a/unsupported/Eigen/src/LevenbergMarquardt/LMpar.h +++ b/unsupported/Eigen/src/LevenbergMarquardt/LMpar.h @@ -30,7 +30,7 @@ namespace internal { using std::abs; typedef typename QRSolver::MatrixType MatrixType; typedef typename QRSolver::Scalar Scalar; - typedef typename QRSolver::Index Index; + typedef typename QRSolver::StorageIndex StorageIndex; /* Local variables */ Index j; diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h b/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h index db3a0ef2c..ae9d793b1 100644 --- a/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h +++ b/unsupported/Eigen/src/LevenbergMarquardt/LMqrsolv.h @@ -28,8 +28,6 @@ void lmqrsolv( Matrix &x, Matrix &sdiag) { - typedef typename Matrix::Index Index; - /* Local variables */ Index i, j, k; Scalar temp; diff --git a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h index 7cebe4e06..9eca33d04 100644 --- a/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h +++ b/unsupported/Eigen/src/LevenbergMarquardt/LevenbergMarquardt.h @@ -115,8 +115,7 @@ class LevenbergMarquardt : internal::no_assignment_operator typedef typename FunctorType::JacobianType JacobianType; typedef typename JacobianType::Scalar Scalar; typedef typename JacobianType::RealScalar RealScalar; - typedef typename JacobianType::Index Index; - typedef typename QRSolver::Index PermIndex; + typedef typename QRSolver::StorageIndex PermIndex; typedef Matrix FVectorType; typedef PermutationMatrix PermutationType; public: diff --git a/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h b/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h index d92fd0ef1..8a7e0e57f 100644 --- a/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h +++ b/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h @@ -535,7 +535,7 @@ class BlockSparseMatrix : public SparseMatrixBase coeffRef(StorageIndex brow, StorageIndex bcol) + Ref coeffRef(Index brow, Index bcol) { eigen_assert(brow < blockRows() && "BLOCK ROW INDEX OUT OF BOUNDS"); eigen_assert(bcol < blockCols() && "BLOCK nzblocksFlagCOLUMN OUT OF BOUNDS"); @@ -829,7 +829,7 @@ class BlockSparseMatrix : public SparseMatrixBase coeff(StorageIndex brow, StorageIndex bcol) const + Map coeff(Index brow, Index bcol) const { eigen_assert(brow < blockRows() && "BLOCK ROW INDEX OUT OF BOUNDS"); eigen_assert(bcol < blockCols() && "BLOCK COLUMN OUT OF BOUNDS"); @@ -857,9 +857,9 @@ class BlockSparseMatrix : public SparseMatrixBase(m_values);} // inline Scalar *valuePtr(){ return m_values; } @@ -873,7 +873,7 @@ class BlockSparseMatrix : public SparseMatrixBase in the array of values */ - StorageIndex blockPtr(Index id) const + Index blockPtr(Index id) const { if(m_blockSize == Dynamic) return m_blockPtr[id]; else return id * m_blockSize * m_blockSize; @@ -955,17 +955,17 @@ class BlockSparseMatrix : public SparseMatrixBase insert(Index brow, Index bcol); - StorageIndex m_innerBSize; // Number of block rows - StorageIndex m_outerBSize; // Number of block columns + Index m_innerBSize; // Number of block rows + Index m_outerBSize; // Number of block columns StorageIndex *m_innerOffset; // Starting index of each inner block (size m_innerBSize+1) StorageIndex *m_outerOffset; // Starting index of each outer block (size m_outerBSize+1) - StorageIndex m_nonzerosblocks; // Total nonzeros blocks (lower than m_innerBSize x m_outerBSize) - StorageIndex m_nonzeros; // Total nonzeros elements + Index m_nonzerosblocks; // Total nonzeros blocks (lower than m_innerBSize x m_outerBSize) + Index m_nonzeros; // Total nonzeros elements Scalar *m_values; //Values stored block column after block column (size m_nonzeros) StorageIndex *m_blockPtr; // Pointer to the beginning of each block in m_values, size m_nonzeroblocks ... null for fixed-size blocks StorageIndex *m_indices; //Inner block indices, size m_nonzerosblocks ... OK StorageIndex *m_outerIndex; // Starting pointer of each block column in m_indices (size m_outerBSize)... OK - StorageIndex m_blockSize; // Size of a block for fixed-size blocks, otherwise -1 + Index m_blockSize; // Size of a block for fixed-size blocks, otherwise -1 }; template @@ -977,7 +977,7 @@ class BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _StorageIndex>:: Flags = _Options }; - BlockInnerIterator(const BlockSparseMatrix& mat, const StorageIndex outer) + BlockInnerIterator(const BlockSparseMatrix& mat, const Index outer) : m_mat(mat),m_outer(outer), m_id(mat.m_outerIndex[outer]), m_end(mat.m_outerIndex[outer+1]) @@ -997,23 +997,23 @@ class BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _StorageIndex>:: rows(),cols()); } // Block inner index - inline StorageIndex index() const {return m_mat.m_indices[m_id]; } - inline StorageIndex outer() const { return m_outer; } + inline Index index() const {return m_mat.m_indices[m_id]; } + inline Index outer() const { return m_outer; } // block row index - inline StorageIndex row() const {return index(); } + inline Index row() const {return index(); } // block column index - inline StorageIndex col() const {return outer(); } + inline Index col() const {return outer(); } // FIXME Number of rows in the current block - inline StorageIndex rows() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_innerOffset[index()+1] - m_mat.m_innerOffset[index()]) : m_mat.m_blockSize; } + inline Index rows() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_innerOffset[index()+1] - m_mat.m_innerOffset[index()]) : m_mat.m_blockSize; } // Number of columns in the current block ... - inline StorageIndex cols() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_outerOffset[m_outer+1]-m_mat.m_outerOffset[m_outer]) : m_mat.m_blockSize;} + inline Index cols() const { return (m_mat.m_blockSize==Dynamic) ? (m_mat.m_outerOffset[m_outer+1]-m_mat.m_outerOffset[m_outer]) : m_mat.m_blockSize;} inline operator bool() const { return (m_id < m_end); } protected: const BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, StorageIndex>& m_mat; - const StorageIndex m_outer; - StorageIndex m_id; - StorageIndex m_end; + const Index m_outer; + Index m_id; + Index m_end; }; template @@ -1055,23 +1055,23 @@ class BlockSparseMatrix<_Scalar, _BlockAtCompileTime, _Options, _StorageIndex>:: { return itb.valueRef().coeff(m_id - m_start, m_offset); } - inline StorageIndex index() const { return m_id; } - inline StorageIndex outer() const {return m_outer; } - inline StorageIndex col() const {return outer(); } - inline StorageIndex row() const { return index();} + inline Index index() const { return m_id; } + inline Index outer() const {return m_outer; } + inline Index col() const {return outer(); } + inline Index row() const { return index();} inline operator bool() const { return itb; } protected: const BlockSparseMatrix& m_mat; - const StorageIndex m_outer; - const StorageIndex m_outerB; + const Index m_outer; + const Index m_outerB; BlockInnerIterator itb; // Iterator through the blocks - const StorageIndex m_offset; // Position of this column in the block - StorageIndex m_start; // starting inner index of this block - StorageIndex m_id; // current inner index in the block - StorageIndex m_end; // starting inner index of the next block + const Index m_offset; // Position of this column in the block + Index m_start; // starting inner index of this block + Index m_id; // current inner index in the block + Index m_end; // starting inner index of the next block }; } // end namespace Eigen diff --git a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h index bedb1dec5..e1284c782 100644 --- a/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h +++ b/unsupported/Eigen/src/SparseExtra/DynamicSparseMatrix.h @@ -33,11 +33,11 @@ namespace Eigen { */ namespace internal { -template -struct traits > +template +struct traits > { typedef _Scalar Scalar; - typedef _Index StorageIndex; + typedef _StorageIndex StorageIndex; typedef Sparse StorageKind; typedef MatrixXpr XprKind; enum { @@ -52,9 +52,9 @@ struct traits > }; } -template +template class DynamicSparseMatrix - : public SparseMatrixBase > + : public SparseMatrixBase > { public: EIGEN_SPARSE_PUBLIC_INTERFACE(DynamicSparseMatrix) @@ -72,16 +72,16 @@ template typedef DynamicSparseMatrix TransposedSparseMatrix; - StorageIndex m_innerSize; + Index m_innerSize; std::vector > m_data; public: - inline StorageIndex rows() const { return IsRowMajor ? outerSize() : m_innerSize; } - inline StorageIndex cols() const { return IsRowMajor ? m_innerSize : outerSize(); } - inline StorageIndex innerSize() const { return m_innerSize; } - inline StorageIndex outerSize() const { return convert_index(m_data.size()); } - inline StorageIndex innerNonZeros(Index j) const { return m_data[j].size(); } + inline Index rows() const { return IsRowMajor ? outerSize() : m_innerSize; } + inline Index cols() const { return IsRowMajor ? m_innerSize : outerSize(); } + inline Index innerSize() const { return m_innerSize; } + inline Index outerSize() const { return convert_index(m_data.size()); } + inline Index innerNonZeros(Index j) const { return m_data[j].size(); } std::vector >& _data() { return m_data; } const std::vector >& _data() const { return m_data; } @@ -117,11 +117,11 @@ template } /** \returns the number of non zero coefficients */ - StorageIndex nonZeros() const + Index nonZeros() const { - StorageIndex res = 0; + Index res = 0; for (Index j=0; j # endif }; -template -class DynamicSparseMatrix::InnerIterator : public SparseVector::InnerIterator +template +class DynamicSparseMatrix::InnerIterator : public SparseVector::InnerIterator { - typedef typename SparseVector::InnerIterator Base; + typedef typename SparseVector::InnerIterator Base; public: InnerIterator(const DynamicSparseMatrix& mat, Index outer) : Base(mat.m_data[outer]), m_outer(outer) @@ -337,10 +337,10 @@ class DynamicSparseMatrix::InnerIterator : public Sparse const Index m_outer; }; -template -class DynamicSparseMatrix::ReverseInnerIterator : public SparseVector::ReverseInnerIterator +template +class DynamicSparseMatrix::ReverseInnerIterator : public SparseVector::ReverseInnerIterator { - typedef typename SparseVector::ReverseInnerIterator Base; + typedef typename SparseVector::ReverseInnerIterator Base; public: ReverseInnerIterator(const DynamicSparseMatrix& mat, Index outer) : Base(mat.m_data[outer]), m_outer(outer) @@ -356,13 +356,13 @@ class DynamicSparseMatrix::ReverseInnerIterator : public namespace internal { -template -struct evaluator > - : evaluator_base > +template +struct evaluator > + : evaluator_base > { typedef _Scalar Scalar; - typedef _Index Index; - typedef DynamicSparseMatrix<_Scalar,_Options,_Index> SparseMatrixType; + typedef _StorageIndex Index; + typedef DynamicSparseMatrix<_Scalar,_Options,_StorageIndex> SparseMatrixType; typedef typename SparseMatrixType::InnerIterator InnerIterator; typedef typename SparseMatrixType::ReverseInnerIterator ReverseInnerIterator; diff --git a/unsupported/Eigen/src/SparseExtra/RandomSetter.h b/unsupported/Eigen/src/SparseExtra/RandomSetter.h index 807ba9d94..0b71ed3ad 100644 --- a/unsupported/Eigen/src/SparseExtra/RandomSetter.h +++ b/unsupported/Eigen/src/SparseExtra/RandomSetter.h @@ -292,10 +292,10 @@ class RandomSetter /** \returns a reference to the coefficient at given coordinates \a row, \a col */ Scalar& operator() (Index row, Index col) { - const StorageIndex outer = internal::convert_index(SetterRowMajor ? row : col); - const StorageIndex inner = internal::convert_index(SetterRowMajor ? col : row); - const StorageIndex outerMajor = outer >> OuterPacketBits; // index of the packet/map - const StorageIndex outerMinor = outer & OuterPacketMask; // index of the inner vector in the packet + const Index outer = SetterRowMajor ? row : col; + const Index inner = SetterRowMajor ? col : row; + const Index outerMajor = outer >> OuterPacketBits; // index of the packet/map + const Index outerMinor = outer & OuterPacketMask; // index of the inner vector in the packet const KeyType key = (KeyType(outerMinor)< Date: Fri, 13 Feb 2015 14:52:36 -0500 Subject: [PATCH 1089/1103] bug #953 - Fix prefetches in 3px4 product kernel This gives a 10% speedup on nexus 4 and on nexus 5. --- Eigen/src/Core/products/GeneralBlockPanelKernel.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h index 11e5f591d..15bf04d1f 100644 --- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h +++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h @@ -810,6 +810,7 @@ void gebp_kernel Date: Fri, 13 Feb 2015 16:07:08 -0800 Subject: [PATCH 1090/1103] Optimized version of the sin(), exp(), log() and sqrt() function for AVX --- Eigen/Core | 1 + Eigen/src/Core/arch/AVX/MathFunctions.h | 317 ++++++++++++++++++++++++ Eigen/src/Core/arch/AVX/PacketMath.h | 19 +- 3 files changed, 331 insertions(+), 6 deletions(-) create mode 100644 Eigen/src/Core/arch/AVX/MathFunctions.h diff --git a/Eigen/Core b/Eigen/Core index 416e901ae..68748a3f6 100644 --- a/Eigen/Core +++ b/Eigen/Core @@ -290,6 +290,7 @@ using std::ptrdiff_t; #include "src/Core/arch/SSE/PacketMath.h" #include "src/Core/arch/SSE/Complex.h" #include "src/Core/arch/AVX/PacketMath.h" + #include "src/Core/arch/AVX/MathFunctions.h" #include "src/Core/arch/AVX/Complex.h" #elif defined EIGEN_VECTORIZE_SSE #include "src/Core/arch/SSE/PacketMath.h" diff --git a/Eigen/src/Core/arch/AVX/MathFunctions.h b/Eigen/src/Core/arch/AVX/MathFunctions.h new file mode 100644 index 000000000..2810a7a0b --- /dev/null +++ b/Eigen/src/Core/arch/AVX/MathFunctions.h @@ -0,0 +1,317 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com) +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_MATH_FUNCTIONS_AVX_H +#define EIGEN_MATH_FUNCTIONS_AVX_H + +// For some reason, this function didn't make it into the avxintirn.h +// used by the compiler, so we'll just wrap it. +#define _mm256_setr_m128(lo, hi) \ + _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1) + +/* The sin, cos, exp, and log functions of this file are loosely derived from + * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/ + */ + +namespace Eigen { + +namespace internal { + +// Sine function +// Computes sin(x) by wrapping x to the interval [-Pi/4,3*Pi/4] and +// evaluating interpolants in [-Pi/4,Pi/4] or [Pi/4,3*Pi/4]. The interpolants +// are (anti-)symmetric and thus have only odd/even coefficients +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +psin(const Packet8f& _x) { + Packet8f x = _x; + + // Some useful values. + _EIGEN_DECLARE_CONST_Packet8i(one, 1); + _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f); + _EIGEN_DECLARE_CONST_Packet8f(two, 2.0f); + _EIGEN_DECLARE_CONST_Packet8f(one_over_four, 0.25f); + _EIGEN_DECLARE_CONST_Packet8f(one_over_pi, 3.183098861837907e-01f); + _EIGEN_DECLARE_CONST_Packet8f(neg_pi_first, -3.140625000000000e+00); + _EIGEN_DECLARE_CONST_Packet8f(neg_pi_second, -9.670257568359375e-04); + _EIGEN_DECLARE_CONST_Packet8f(neg_pi_third, -6.278329571784980e-07); + _EIGEN_DECLARE_CONST_Packet8f(four_over_pi, 1.273239544735163e+00); + + // Map x from [-Pi/4,3*Pi/4] to z in [-1,3] and subtract the shifted period. + Packet8f z = pmul(x, p8f_one_over_pi); + Packet8f shift = _mm256_floor_ps(padd(z, p8f_one_over_four)); + x = pmadd(shift, p8f_neg_pi_first, x); + x = pmadd(shift, p8f_neg_pi_second, x); + x = pmadd(shift, p8f_neg_pi_third, x); + z = pmul(x, p8f_four_over_pi); + + // Make a mask for the entries that need flipping, i.e. wherever the shift + // is odd. + Packet8i shift_ints = _mm256_cvtps_epi32(shift); + Packet8i shift_isodd = + (__m256i)_mm256_and_ps((__m256)shift_ints, (__m256)p8i_one); +#ifdef EIGEN_VECTORIZE_AVX2 + Packet8i sign_flip_mask = _mm256_slli_epi32(shift_isodd, 31); +#else + __m128i lo = + _mm_slli_epi32(_mm256_extractf128_si256((__m256i)shift_isodd, 0), 31); + __m128i hi = + _mm_slli_epi32(_mm256_extractf128_si256((__m256i)shift_isodd, 1), 31); + Packet8i sign_flip_mask = _mm256_setr_m128(lo, hi); +#endif + + // Create a mask for which interpolant to use, i.e. if z > 1, then the mask + // is set to ones for that entry. + Packet8f ival_mask = _mm256_cmp_ps(z, p8f_one, _CMP_GT_OQ); + + // Evaluate the polynomial for the interval [1,3] in z. + _EIGEN_DECLARE_CONST_Packet8f(coeff_right_0, 9.999999724233232e-01f); + _EIGEN_DECLARE_CONST_Packet8f(coeff_right_2, -3.084242535619928e-01); + _EIGEN_DECLARE_CONST_Packet8f(coeff_right_4, 1.584991525700324e-02); + _EIGEN_DECLARE_CONST_Packet8f(coeff_right_6, -3.188805084631342e-04); + Packet8f z_minus_two = psub(z, p8f_two); + Packet8f z_minus_two2 = pmul(z_minus_two, z_minus_two); + Packet8f right = pmadd(p8f_coeff_right_6, z_minus_two2, p8f_coeff_right_4); + right = pmadd(right, z_minus_two2, p8f_coeff_right_2); + right = pmadd(right, z_minus_two2, p8f_coeff_right_0); + + // Evaluate the polynomial for the interval [-1,1] in z. + _EIGEN_DECLARE_CONST_Packet8f(coeff_left_1, 7.853981525427295e-01); + _EIGEN_DECLARE_CONST_Packet8f(coeff_left_3, -8.074536727092352e-02); + _EIGEN_DECLARE_CONST_Packet8f(coeff_left_5, 2.489871967827018e-03); + _EIGEN_DECLARE_CONST_Packet8f(coeff_left_7, -3.587725841214251e-05); + Packet8f z2 = pmul(z, z); + Packet8f left = pmadd(p8f_coeff_left_7, z2, p8f_coeff_left_5); + left = pmadd(left, z2, p8f_coeff_left_3); + left = pmadd(left, z2, p8f_coeff_left_1); + left = pmul(left, z); + + // Assemble the results, i.e. select the left and right polynomials. + left = _mm256_andnot_ps(ival_mask, left); + right = _mm256_and_ps(ival_mask, right); + Packet8f res = _mm256_or_ps(left, right); + + // Flip the sign on the odd intervals and return the result. + res = _mm256_xor_ps(res, (__m256)sign_flip_mask); + return res; +} + +// Natural logarithm +// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2) +// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can +// be easily approximated by a polynomial centered on m=1 for stability. +// TODO(gonnet): Further reduce the interval allowing for lower-degree +// polynomial interpolants -> ... -> profit! +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +plog(const Packet8f& _x) { + Packet8f x = _x; + _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f); + _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f); + _EIGEN_DECLARE_CONST_Packet8f(126f, 126.0f); + + _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inv_mant_mask, ~0x7f800000); + + // The smallest non denormalized float number. + _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(min_norm_pos, 0x00800000); + _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(minus_inf, 0xff800000); + + // Polynomial coefficients. + _EIGEN_DECLARE_CONST_Packet8f(cephes_SQRTHF, 0.707106781186547524f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p0, 7.0376836292E-2f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p1, -1.1514610310E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p2, 1.1676998740E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p3, -1.2420140846E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p4, +1.4249322787E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p5, -1.6668057665E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p6, +2.0000714765E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p7, -2.4999993993E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p8, +3.3333331174E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q1, -2.12194440e-4f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q2, 0.693359375f); + + Packet8f invalid_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_NGE_UQ); // not greater equal is true if x is NaN + Packet8f iszero_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_EQ_OQ); + + // Truncate input values to the minimum positive normal. + x = pmax(x, p8f_min_norm_pos); + +// Extract the shifted exponents (No bitwise shifting in regular AVX, so +// convert to SSE and do it there). +#ifdef EIGEN_VECTORIZE_AVX2 + Packet8f emm0 = _mm256_cvtepi32_ps(_mm256_srli_epi32((__m256i)x, 23)); +#else + __m128i lo = _mm_srli_epi32(_mm256_extractf128_si256((__m256i)x, 0), 23); + __m128i hi = _mm_srli_epi32(_mm256_extractf128_si256((__m256i)x, 1), 23); + Packet8f emm0 = _mm256_cvtepi32_ps(_mm256_setr_m128(lo, hi)); +#endif + Packet8f e = _mm256_sub_ps(emm0, p8f_126f); + + // Set the exponents to -1, i.e. x are in the range [0.5,1). + x = _mm256_and_ps(x, p8f_inv_mant_mask); + x = _mm256_or_ps(x, p8f_half); + + // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2)) + // and shift by -1. The values are then centered around 0, which improves + // the stability of the polynomial evaluation. + // if( x < SQRTHF ) { + // e -= 1; + // x = x + x - 1.0; + // } else { x = x - 1.0; } + Packet8f mask = _mm256_cmp_ps(x, p8f_cephes_SQRTHF, _CMP_LT_OQ); + Packet8f tmp = _mm256_and_ps(x, mask); + x = psub(x, p8f_1); + e = psub(e, _mm256_and_ps(p8f_1, mask)); + x = padd(x, tmp); + + Packet8f x2 = pmul(x, x); + Packet8f x3 = pmul(x2, x); + + // Evaluate the polynomial approximant of degree 8 in three parts, probably + // to improve instruction-level parallelism. + Packet8f y, y1, y2; + y = pmadd(p8f_cephes_log_p0, x, p8f_cephes_log_p1); + y1 = pmadd(p8f_cephes_log_p3, x, p8f_cephes_log_p4); + y2 = pmadd(p8f_cephes_log_p6, x, p8f_cephes_log_p7); + y = pmadd(y, x, p8f_cephes_log_p2); + y1 = pmadd(y1, x, p8f_cephes_log_p5); + y2 = pmadd(y2, x, p8f_cephes_log_p8); + y = pmadd(y, x3, y1); + y = pmadd(y, x3, y2); + y = pmul(y, x3); + + // Add the logarithm of the exponent back to the result of the interpolation. + y1 = pmul(e, p8f_cephes_log_q1); + tmp = pmul(x2, p8f_half); + y = padd(y, y1); + x = psub(x, tmp); + y2 = pmul(e, p8f_cephes_log_q2); + x = padd(x, y); + x = padd(x, y2); + + // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF. + return _mm256_or_ps( + _mm256_andnot_ps(iszero_mask, _mm256_or_ps(x, invalid_mask)), + _mm256_and_ps(iszero_mask, p8f_minus_inf)); +} + +// Exponential function. Works by writing "x = m*log(2) + r" where +// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then +// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1). +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +pexp(const Packet8f& _x) { + _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f); + _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f); + _EIGEN_DECLARE_CONST_Packet8f(127, 127.0f); + + _EIGEN_DECLARE_CONST_Packet8f(exp_hi, 88.3762626647950f); + _EIGEN_DECLARE_CONST_Packet8f(exp_lo, -88.3762626647949f); + + _EIGEN_DECLARE_CONST_Packet8f(cephes_LOG2EF, 1.44269504088896341f); + + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p0, 1.9875691500E-4f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p1, 1.3981999507E-3f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p2, 8.3334519073E-3f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p3, 4.1665795894E-2f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p4, 1.6666665459E-1f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p5, 5.0000001201E-1f); + + // Clamp x. + Packet8f x = pmax(pmin(_x, p8f_exp_hi), p8f_exp_lo); + + // Express exp(x) as exp(m*ln(2) + r), start by extracting + // m = floor(x/ln(2) + 0.5). + Packet8f m = _mm256_floor_ps(pmadd(x, p8f_cephes_LOG2EF, p8f_half)); + +// Get r = x - m*ln(2). If no FMA instructions are available, m*ln(2) is +// subtracted out in two parts, m*C1+m*C2 = m*ln(2), to avoid accumulating +// truncation errors. Note that we don't use the "pmadd" function here to +// ensure that a precision-preserving FMA instruction is used. +#ifdef EIGEN_VECTORIZE_FMA + _EIGEN_DECLARE_CONST_Packet8f(nln2, -0.6931471805599453f); + Packet8f r = _mm256_fmadd_ps(m, p8f_nln2, x); +#else + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C1, 0.693359375f); + _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C2, -2.12194440e-4f); + Packet8f r = psub(x, pmul(m, p8f_cephes_exp_C1)); + r = psub(r, pmul(m, p8f_cephes_exp_C2)); +#endif + + Packet8f r2 = pmul(r, r); + + // TODO(gonnet): Split into odd/even polynomials and try to exploit + // instruction-level parallelism. + Packet8f y = p8f_cephes_exp_p0; + y = pmadd(y, r, p8f_cephes_exp_p1); + y = pmadd(y, r, p8f_cephes_exp_p2); + y = pmadd(y, r, p8f_cephes_exp_p3); + y = pmadd(y, r, p8f_cephes_exp_p4); + y = pmadd(y, r, p8f_cephes_exp_p5); + y = pmadd(y, r2, r); + y = padd(y, p8f_1); + + // Build emm0 = 2^m. + Packet8i emm0 = _mm256_cvttps_epi32(padd(m, p8f_127)); +#ifdef EIGEN_VECTORIZE_AVX2 + emm0 = _mm256_slli_epi32(emm0, 23); +#else + __m128i lo = _mm_slli_epi32(_mm256_extractf128_si256(emm0, 0), 23); + __m128i hi = _mm_slli_epi32(_mm256_extractf128_si256(emm0, 1), 23); + emm0 = _mm256_setr_m128(lo, hi); +#endif + + // Return 2^m * exp(r). + return pmax(pmul(y, _mm256_castsi256_ps(emm0)), _x); +} + +// Functions for sqrt. +// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step +// of Newton's method, at a cost of 1-2 bits of precision as opposed to the +// exact solution. The main advantage of this approach is not just speed, but +// also the fact that it can be inlined and pipelined with other computations, +// further reducing its effective latency. +#if EIGEN_FAST_MATH +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +psqrt(const Packet8f& _x) { + _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f); + _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f); + _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000); + + Packet8f neg_half = pmul(_x, p8f_minus_half); + + // select only the inverse sqrt of positive normal inputs (denormals are + // flushed to zero and cause infs as well). + Packet8f non_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_GE_OQ); + Packet8f x = _mm256_and_ps(non_zero_mask, _mm256_rsqrt_ps(_x)); + + // Do a single step of Newton's iteration. + x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five)); + + // Multiply the original _x by it's reciprocal square root to extract the + // square root. + return pmul(_x, x); +} +#else +template <> +EIGEN_STRONG_INLINE Packet8f psqrt(const Packet8f& x) { + return _mm256_sqrt_ps(x); +} +#endif +template <> +EIGEN_STRONG_INLINE Packet4d psqrt(const Packet4d& x) { + return _mm256_sqrt_pd(x); +} + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_MATH_FUNCTIONS_AVX_H diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index fc2e78b2d..c404676c4 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -42,6 +42,12 @@ template<> struct is_arithmetic<__m256d> { enum { value = true }; }; #define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \ const Packet4d p4d_##NAME = pset1(X) +#define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \ + const Packet8f p8f_##NAME = (__m256)pset1(X) + +#define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \ + const Packet8i p8i_##NAME = pset1(X) + template<> struct packet_traits : default_packet_traits { @@ -54,13 +60,13 @@ template<> struct packet_traits : default_packet_traits HasHalfPacket = 1, HasDiv = 1, - HasSin = 0, + HasSin = 1, HasCos = 0, - HasLog = 0, - HasExp = 0, - HasSqrt = 0 + HasLog = 1, + HasExp = 1, + HasSqrt = 1 }; - }; +}; template<> struct packet_traits : default_packet_traits { typedef Packet4d type; @@ -72,7 +78,8 @@ template<> struct packet_traits : default_packet_traits HasHalfPacket = 1, HasDiv = 1, - HasExp = 0 + HasExp = 0, + HasSqrt = 1 }; }; From bd511dde9da3a1f147b3d0d70e76298f759c4420 Mon Sep 17 00:00:00 2001 From: Christoph Hertzberg Date: Sun, 15 Feb 2015 06:08:25 +0100 Subject: [PATCH 1091/1103] bug #952: Missing \endcode made doxygen fail to build ColPivHouseholderQR --- Eigen/src/QR/ColPivHouseholderQR.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Eigen/src/QR/ColPivHouseholderQR.h b/Eigen/src/QR/ColPivHouseholderQR.h index 370cb69e3..2696c4b2f 100644 --- a/Eigen/src/QR/ColPivHouseholderQR.h +++ b/Eigen/src/QR/ColPivHouseholderQR.h @@ -566,7 +566,7 @@ struct Assignment >, interna /** \returns the matrix Q as a sequence of householder transformations. * You can extract the meaningful part only by using: - * \code qr.householderQ().setLength(qr.nonzeroPivots()) */ + * \code qr.householderQ().setLength(qr.nonzeroPivots()) \endcode*/ template typename ColPivHouseholderQR::HouseholderSequenceType ColPivHouseholderQR ::householderQ() const From aa6c516ec17fb44dff85b1abf3a1b05d58d3bc01 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 13:19:05 +0100 Subject: [PATCH 1092/1103] Fix many long to int conversion warnings: - fix usage of Index (API) versus StorageIndex (when multiple indexes are stored) - use StorageIndex(val) when the input has already been check - use internal::convert_index(val) when val is potentially unsafe (directly comes from user input) --- Eigen/src/Core/CwiseUnaryOp.h | 4 +- Eigen/src/Core/PermutationMatrix.h | 3 +- Eigen/src/IterativeLinearSolvers/BiCGSTAB.h | 12 +- .../ConjugateGradient.h | 10 +- .../IterativeSolverBase.h | 16 +- Eigen/src/OrderingMethods/Amd.h | 74 ++-- Eigen/src/OrderingMethods/Eigen_Colamd.h | 390 +++++++++--------- Eigen/src/OrderingMethods/Ordering.h | 24 +- Eigen/src/PaStiXSupport/PaStiXSupport.h | 6 +- Eigen/src/SparseCholesky/SimplicialCholesky.h | 4 +- .../SparseCholesky/SimplicialCholesky_impl.h | 22 +- Eigen/src/SparseCore/CompressedStorage.h | 8 +- Eigen/src/SparseCore/SparseBlock.h | 4 +- Eigen/src/SparseCore/SparseColEtree.h | 36 +- Eigen/src/SparseCore/SparseMatrix.h | 10 +- Eigen/src/SparseCore/SparsePermutation.h | 2 +- Eigen/src/SparseCore/SparseSelfAdjointView.h | 2 +- Eigen/src/SparseCore/SparseSolverBase.h | 12 +- Eigen/src/SparseCore/SparseVector.h | 4 +- Eigen/src/SparseLU/SparseLU.h | 4 +- Eigen/src/SparseLU/SparseLUImpl.h | 2 +- .../src/SparseLU/SparseLU_SupernodalMatrix.h | 6 +- Eigen/src/SparseLU/SparseLU_Utils.h | 2 +- Eigen/src/SparseLU/SparseLU_column_bmod.h | 2 +- Eigen/src/SparseLU/SparseLU_column_dfs.h | 22 +- Eigen/src/SparseLU/SparseLU_copy_to_ucol.h | 2 +- .../src/SparseLU/SparseLU_heap_relax_snode.h | 15 +- Eigen/src/SparseLU/SparseLU_panel_dfs.h | 32 +- Eigen/src/SparseLU/SparseLU_pivotL.h | 4 +- Eigen/src/SparseLU/SparseLU_pruneL.h | 2 +- Eigen/src/SparseLU/SparseLU_relax_snode.h | 8 +- Eigen/src/SparseQR/SparseQR.h | 10 +- Eigen/src/UmfPackSupport/UmfPackSupport.h | 5 +- test/sparse_basic.cpp | 4 +- test/spqr_support.cpp | 2 +- .../Eigen/src/IterativeSolvers/GMRES.h | 24 +- .../Eigen/src/IterativeSolvers/MINRES.h | 6 +- 37 files changed, 397 insertions(+), 398 deletions(-) diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h index 5388af216..da1d1992d 100644 --- a/Eigen/src/Core/CwiseUnaryOp.h +++ b/Eigen/src/Core/CwiseUnaryOp.h @@ -66,9 +66,9 @@ class CwiseUnaryOp : public CwiseUnaryOpImpl /** Sets *this to be the identity permutation matrix */ void setIdentity() { - for(Index i = 0; i < size(); ++i) + StorageIndex n = StorageIndex(size()); + for(StorageIndex i = 0; i < n; ++i) indices().coeffRef(i) = i; } diff --git a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h index a715c7285..e67f09184 100644 --- a/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h +++ b/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h @@ -27,7 +27,7 @@ namespace internal { */ template bool bicgstab(const MatrixType& mat, const Rhs& rhs, Dest& x, - const Preconditioner& precond, int& iters, + const Preconditioner& precond, Index& iters, typename Dest::RealScalar& tol_error) { using std::sqrt; @@ -36,9 +36,9 @@ bool bicgstab(const MatrixType& mat, const Rhs& rhs, Dest& x, typedef typename Dest::Scalar Scalar; typedef Matrix VectorType; RealScalar tol = tol_error; - int maxIters = iters; + Index maxIters = iters; - int n = mat.cols(); + Index n = mat.cols(); VectorType r = rhs - mat * x; VectorType r0 = r; @@ -61,8 +61,8 @@ bool bicgstab(const MatrixType& mat, const Rhs& rhs, Dest& x, RealScalar tol2 = tol*tol; RealScalar eps2 = NumTraits::epsilon()*NumTraits::epsilon(); - int i = 0; - int restarts = 0; + Index i = 0; + Index restarts = 0; while ( r.squaredNorm()/rhs_sqnorm > tol2 && i EIGEN_DONT_INLINE void conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x, - const Preconditioner& precond, int& iters, + const Preconditioner& precond, Index& iters, typename Dest::RealScalar& tol_error) { using std::sqrt; @@ -36,9 +36,9 @@ void conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x, typedef Matrix VectorType; RealScalar tol = tol_error; - int maxIters = iters; + Index maxIters = iters; - int n = mat.cols(); + Index n = mat.cols(); VectorType residual = rhs - mat * x; //initial residual @@ -64,7 +64,7 @@ void conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x, VectorType z(n), tmp(n); RealScalar absNew = numext::real(residual.dot(p)); // the square of the absolute value of r scaled by invM - int i = 0; + Index i = 0; while(i < maxIters) { tmp.noalias() = mat * p; // the bottleneck of the algorithm @@ -190,7 +190,7 @@ public: m_iterations = Base::maxIterations(); m_error = Base::m_tolerance; - for(int j=0; j tb(size); Eigen::Matrix tx(size); - for(int k=0; k mp_matrix; Preconditioner m_preconditioner; - int m_maxIterations; + Index m_maxIterations; RealScalar m_tolerance; mutable RealScalar m_error; - mutable int m_iterations; + mutable Index m_iterations; mutable ComputationInfo m_info; mutable bool m_analysisIsOk, m_factorizationIsOk; }; diff --git a/Eigen/src/OrderingMethods/Amd.h b/Eigen/src/OrderingMethods/Amd.h index 50022d1ca..3d2981f0c 100644 --- a/Eigen/src/OrderingMethods/Amd.h +++ b/Eigen/src/OrderingMethods/Amd.h @@ -41,10 +41,10 @@ template inline bool amd_marked(const T0* w, const T1& template inline void amd_mark(const T0* w, const T1& j) { return w[j] = amd_flip(w[j]); } /* clear w */ -template -static Index cs_wclear (Index mark, Index lemax, Index *w, Index n) +template +static StorageIndex cs_wclear (StorageIndex mark, StorageIndex lemax, StorageIndex *w, StorageIndex n) { - Index k; + StorageIndex k; if(mark < 2 || (mark + lemax < 0)) { for(k = 0; k < n; k++) @@ -56,10 +56,10 @@ static Index cs_wclear (Index mark, Index lemax, Index *w, Index n) } /* depth-first search and postorder of a tree rooted at node j */ -template -Index cs_tdfs(Index j, Index k, Index *head, const Index *next, Index *post, Index *stack) +template +StorageIndex cs_tdfs(StorageIndex j, StorageIndex k, StorageIndex *head, const StorageIndex *next, StorageIndex *post, StorageIndex *stack) { - Index i, p, top = 0; + StorageIndex i, p, top = 0; if(!head || !next || !post || !stack) return (-1); /* check inputs */ stack[0] = j; /* place j on the stack */ while (top >= 0) /* while (stack is not empty) */ @@ -87,41 +87,39 @@ Index cs_tdfs(Index j, Index k, Index *head, const Index *next, Index *post, Ind * \returns the permutation P reducing the fill-in of the input matrix \a C * The input matrix \a C must be a selfadjoint compressed column major SparseMatrix object. Both the upper and lower parts have to be stored, but the diagonal entries are optional. * On exit the values of C are destroyed */ -template -void minimum_degree_ordering(SparseMatrix& C, PermutationMatrix& perm) +template +void minimum_degree_ordering(SparseMatrix& C, PermutationMatrix& perm) { using std::sqrt; - Index d, dk, dext, lemax = 0, e, elenk, eln, i, j, k, k1, - k2, k3, jlast, ln, dense, nzmax, mindeg = 0, nvi, nvj, nvk, mark, wnvi, - ok, nel = 0, p, p1, p2, p3, p4, pj, pk, pk1, pk2, pn, q, t; + StorageIndex d, dk, dext, lemax = 0, e, elenk, eln, i, j, k, k1, + k2, k3, jlast, ln, dense, nzmax, mindeg = 0, nvi, nvj, nvk, mark, wnvi, + ok, nel = 0, p, p1, p2, p3, p4, pj, pk, pk1, pk2, pn, q, t, h; - std::size_t h; - - Index n = C.cols(); - dense = std::max (16, Index(10 * sqrt(double(n)))); /* find dense threshold */ - dense = std::min (n-2, dense); + StorageIndex n = StorageIndex(C.cols()); + dense = std::max (16, StorageIndex(10 * sqrt(double(n)))); /* find dense threshold */ + dense = (std::min)(n-2, dense); - Index cnz = C.nonZeros(); + StorageIndex cnz = StorageIndex(C.nonZeros()); perm.resize(n+1); t = cnz + cnz/5 + 2*n; /* add elbow room to C */ C.resizeNonZeros(t); // get workspace - ei_declare_aligned_stack_constructed_variable(Index,W,8*(n+1),0); - Index* len = W; - Index* nv = W + (n+1); - Index* next = W + 2*(n+1); - Index* head = W + 3*(n+1); - Index* elen = W + 4*(n+1); - Index* degree = W + 5*(n+1); - Index* w = W + 6*(n+1); - Index* hhead = W + 7*(n+1); - Index* last = perm.indices().data(); /* use P as workspace for last */ + ei_declare_aligned_stack_constructed_variable(StorageIndex,W,8*(n+1),0); + StorageIndex* len = W; + StorageIndex* nv = W + (n+1); + StorageIndex* next = W + 2*(n+1); + StorageIndex* head = W + 3*(n+1); + StorageIndex* elen = W + 4*(n+1); + StorageIndex* degree = W + 5*(n+1); + StorageIndex* w = W + 6*(n+1); + StorageIndex* hhead = W + 7*(n+1); + StorageIndex* last = perm.indices().data(); /* use P as workspace for last */ /* --- Initialize quotient graph ---------------------------------------- */ - Index* Cp = C.outerIndexPtr(); - Index* Ci = C.innerIndexPtr(); + StorageIndex* Cp = C.outerIndexPtr(); + StorageIndex* Ci = C.innerIndexPtr(); for(k = 0; k < n; k++) len[k] = Cp[k+1] - Cp[k]; len[n] = 0; @@ -138,7 +136,7 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation elen[i] = 0; // Ek of node i is empty degree[i] = len[i]; // degree of node i } - mark = internal::cs_wclear(0, 0, w, n); /* clear w */ + mark = internal::cs_wclear(0, 0, w, n); /* clear w */ elen[n] = -2; /* n is a dead element */ Cp[n] = -1; /* n is a root of assembly tree */ w[n] = 0; /* n is a dead element */ @@ -253,7 +251,7 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation elen[k] = -2; /* k is now an element */ /* --- Find set differences ----------------------------------------- */ - mark = internal::cs_wclear(mark, lemax, w, n); /* clear w if necessary */ + mark = internal::cs_wclear(mark, lemax, w, n); /* clear w if necessary */ for(pk = pk1; pk < pk2; pk++) /* scan 1: find |Le\Lk| */ { i = Ci[pk]; @@ -323,7 +321,7 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation } else { - degree[i] = std::min (degree[i], d); /* update degree(i) */ + degree[i] = std::min (degree[i], d); /* update degree(i) */ Ci[pn] = Ci[p3]; /* move first node to end */ Ci[p3] = Ci[p1]; /* move 1st el. to end of Ei */ Ci[p1] = k; /* add k as 1st element in of Ei */ @@ -331,12 +329,12 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation h %= n; /* finalize hash of i */ next[i] = hhead[h]; /* place i in hash bucket */ hhead[h] = i; - last[i] = Index(h); /* save hash of i in last[i] */ + last[i] = h; /* save hash of i in last[i] */ } } /* scan2 is done */ degree[k] = dk; /* finalize |Lk| */ - lemax = std::max(lemax, dk); - mark = internal::cs_wclear(mark+lemax, lemax, w, n); /* clear w */ + lemax = std::max(lemax, dk); + mark = internal::cs_wclear(mark+lemax, lemax, w, n); /* clear w */ /* --- Supernode detection ------------------------------------------ */ for(pk = pk1; pk < pk2; pk++) @@ -384,12 +382,12 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation if((nvi = -nv[i]) <= 0) continue;/* skip if i is dead */ nv[i] = nvi; /* restore nv[i] */ d = degree[i] + dk - nvi; /* compute external degree(i) */ - d = std::min (d, n - nel - nvi); + d = std::min (d, n - nel - nvi); if(head[d] != -1) last[head[d]] = i; next[i] = head[d]; /* put i back in degree list */ last[i] = -1; head[d] = i; - mindeg = std::min (mindeg, d); /* find new minimum degree */ + mindeg = std::min (mindeg, d); /* find new minimum degree */ degree[i] = d; Ci[p++] = i; /* place i in Lk */ } @@ -422,7 +420,7 @@ void minimum_degree_ordering(SparseMatrix& C, Permutation } for(k = 0, i = 0; i <= n; i++) /* postorder the assembly tree */ { - if(Cp[i] == -1) k = internal::cs_tdfs(i, k, head, next, perm.indices().data(), w); + if(Cp[i] == -1) k = internal::cs_tdfs(i, k, head, next, perm.indices().data(), w); } perm.indices().conservativeResize(n); diff --git a/Eigen/src/OrderingMethods/Eigen_Colamd.h b/Eigen/src/OrderingMethods/Eigen_Colamd.h index 44548f660..6238676e5 100644 --- a/Eigen/src/OrderingMethods/Eigen_Colamd.h +++ b/Eigen/src/OrderingMethods/Eigen_Colamd.h @@ -135,54 +135,54 @@ namespace internal { /* ========================================================================== */ // == Row and Column structures == -template +template struct colamd_col { - Index start ; /* index for A of first row in this column, or DEAD */ + IndexType start ; /* index for A of first row in this column, or DEAD */ /* if column is dead */ - Index length ; /* number of rows in this column */ + IndexType length ; /* number of rows in this column */ union { - Index thickness ; /* number of original columns represented by this */ + IndexType thickness ; /* number of original columns represented by this */ /* col, if the column is alive */ - Index parent ; /* parent in parent tree super-column structure, if */ + IndexType parent ; /* parent in parent tree super-column structure, if */ /* the column is dead */ } shared1 ; union { - Index score ; /* the score used to maintain heap, if col is alive */ - Index order ; /* pivot ordering of this column, if col is dead */ + IndexType score ; /* the score used to maintain heap, if col is alive */ + IndexType order ; /* pivot ordering of this column, if col is dead */ } shared2 ; union { - Index headhash ; /* head of a hash bucket, if col is at the head of */ + IndexType headhash ; /* head of a hash bucket, if col is at the head of */ /* a degree list */ - Index hash ; /* hash value, if col is not in a degree list */ - Index prev ; /* previous column in degree list, if col is in a */ + IndexType hash ; /* hash value, if col is not in a degree list */ + IndexType prev ; /* previous column in degree list, if col is in a */ /* degree list (but not at the head of a degree list) */ } shared3 ; union { - Index degree_next ; /* next column, if col is in a degree list */ - Index hash_next ; /* next column, if col is in a hash list */ + IndexType degree_next ; /* next column, if col is in a degree list */ + IndexType hash_next ; /* next column, if col is in a hash list */ } shared4 ; }; -template +template struct Colamd_Row { - Index start ; /* index for A of first col in this row */ - Index length ; /* number of principal columns in this row */ + IndexType start ; /* index for A of first col in this row */ + IndexType length ; /* number of principal columns in this row */ union { - Index degree ; /* number of principal & non-principal columns in row */ - Index p ; /* used as a row pointer in init_rows_cols () */ + IndexType degree ; /* number of principal & non-principal columns in row */ + IndexType p ; /* used as a row pointer in init_rows_cols () */ } shared1 ; union { - Index mark ; /* for computing set differences and marking dead rows*/ - Index first_column ;/* first column in row (used in garbage collection) */ + IndexType mark ; /* for computing set differences and marking dead rows*/ + IndexType first_column ;/* first column in row (used in garbage collection) */ } shared2 ; }; @@ -202,38 +202,38 @@ struct Colamd_Row This macro is not needed when using symamd. - Explicit typecast to Index added Sept. 23, 2002, COLAMD version 2.2, to avoid + Explicit typecast to IndexType added Sept. 23, 2002, COLAMD version 2.2, to avoid gcc -pedantic warning messages. */ -template -inline Index colamd_c(Index n_col) -{ return Index( ((n_col) + 1) * sizeof (colamd_col) / sizeof (Index) ) ; } +template +inline IndexType colamd_c(IndexType n_col) +{ return IndexType( ((n_col) + 1) * sizeof (colamd_col) / sizeof (IndexType) ) ; } -template -inline Index colamd_r(Index n_row) -{ return Index(((n_row) + 1) * sizeof (Colamd_Row) / sizeof (Index)); } +template +inline IndexType colamd_r(IndexType n_row) +{ return IndexType(((n_row) + 1) * sizeof (Colamd_Row) / sizeof (IndexType)); } // Prototypes of non-user callable routines -template -static Index init_rows_cols (Index n_row, Index n_col, Colamd_Row Row [], colamd_col col [], Index A [], Index p [], Index stats[COLAMD_STATS] ); +template +static IndexType init_rows_cols (IndexType n_row, IndexType n_col, Colamd_Row Row [], colamd_col col [], IndexType A [], IndexType p [], IndexType stats[COLAMD_STATS] ); -template -static void init_scoring (Index n_row, Index n_col, Colamd_Row Row [], colamd_col Col [], Index A [], Index head [], double knobs[COLAMD_KNOBS], Index *p_n_row2, Index *p_n_col2, Index *p_max_deg); +template +static void init_scoring (IndexType n_row, IndexType n_col, Colamd_Row Row [], colamd_col Col [], IndexType A [], IndexType head [], double knobs[COLAMD_KNOBS], IndexType *p_n_row2, IndexType *p_n_col2, IndexType *p_max_deg); -template -static Index find_ordering (Index n_row, Index n_col, Index Alen, Colamd_Row Row [], colamd_col Col [], Index A [], Index head [], Index n_col2, Index max_deg, Index pfree); +template +static IndexType find_ordering (IndexType n_row, IndexType n_col, IndexType Alen, Colamd_Row Row [], colamd_col Col [], IndexType A [], IndexType head [], IndexType n_col2, IndexType max_deg, IndexType pfree); -template -static void order_children (Index n_col, colamd_col Col [], Index p []); +template +static void order_children (IndexType n_col, colamd_col Col [], IndexType p []); -template -static void detect_super_cols (colamd_col Col [], Index A [], Index head [], Index row_start, Index row_length ) ; +template +static void detect_super_cols (colamd_col Col [], IndexType A [], IndexType head [], IndexType row_start, IndexType row_length ) ; -template -static Index garbage_collection (Index n_row, Index n_col, Colamd_Row Row [], colamd_col Col [], Index A [], Index *pfree) ; +template +static IndexType garbage_collection (IndexType n_row, IndexType n_col, Colamd_Row Row [], colamd_col Col [], IndexType A [], IndexType *pfree) ; -template -static inline Index clear_mark (Index n_row, Colamd_Row Row [] ) ; +template +static inline IndexType clear_mark (IndexType n_row, Colamd_Row Row [] ) ; /* === No debugging ========================================================= */ @@ -260,8 +260,8 @@ static inline Index clear_mark (Index n_row, Colamd_Row Row [] ) ; * \param n_col number of columns in A * \return recommended value of Alen for use by colamd */ -template -inline Index colamd_recommended ( Index nnz, Index n_row, Index n_col) +template +inline IndexType colamd_recommended ( IndexType nnz, IndexType n_row, IndexType n_col) { if ((nnz) < 0 || (n_row) < 0 || (n_col) < 0) return (-1); @@ -325,22 +325,22 @@ static inline void colamd_set_defaults(double knobs[COLAMD_KNOBS]) * \param knobs parameter settings for colamd * \param stats colamd output statistics and error codes */ -template -static bool colamd(Index n_row, Index n_col, Index Alen, Index *A, Index *p, double knobs[COLAMD_KNOBS], Index stats[COLAMD_STATS]) +template +static bool colamd(IndexType n_row, IndexType n_col, IndexType Alen, IndexType *A, IndexType *p, double knobs[COLAMD_KNOBS], IndexType stats[COLAMD_STATS]) { /* === Local variables ================================================== */ - Index i ; /* loop index */ - Index nnz ; /* nonzeros in A */ - Index Row_size ; /* size of Row [], in integers */ - Index Col_size ; /* size of Col [], in integers */ - Index need ; /* minimum required length of A */ - Colamd_Row *Row ; /* pointer into A of Row [0..n_row] array */ - colamd_col *Col ; /* pointer into A of Col [0..n_col] array */ - Index n_col2 ; /* number of non-dense, non-empty columns */ - Index n_row2 ; /* number of non-dense, non-empty rows */ - Index ngarbage ; /* number of garbage collections performed */ - Index max_deg ; /* maximum row degree */ + IndexType i ; /* loop index */ + IndexType nnz ; /* nonzeros in A */ + IndexType Row_size ; /* size of Row [], in integers */ + IndexType Col_size ; /* size of Col [], in integers */ + IndexType need ; /* minimum required length of A */ + Colamd_Row *Row ; /* pointer into A of Row [0..n_row] array */ + colamd_col *Col ; /* pointer into A of Col [0..n_col] array */ + IndexType n_col2 ; /* number of non-dense, non-empty columns */ + IndexType n_row2 ; /* number of non-dense, non-empty rows */ + IndexType ngarbage ; /* number of garbage collections performed */ + IndexType max_deg ; /* maximum row degree */ double default_knobs [COLAMD_KNOBS] ; /* default knobs array */ @@ -431,8 +431,8 @@ static bool colamd(Index n_row, Index n_col, Index Alen, Index *A, Index *p, dou } Alen -= Col_size + Row_size ; - Col = (colamd_col *) &A [Alen] ; - Row = (Colamd_Row *) &A [Alen + Col_size] ; + Col = (colamd_col *) &A [Alen] ; + Row = (Colamd_Row *) &A [Alen + Col_size] ; /* === Construct the row and column data structures ===================== */ @@ -485,29 +485,29 @@ static bool colamd(Index n_row, Index n_col, Index Alen, Index *A, Index *p, dou column form of the matrix. Returns false if the matrix is invalid, true otherwise. Not user-callable. */ -template -static Index init_rows_cols /* returns true if OK, or false otherwise */ +template +static IndexType init_rows_cols /* returns true if OK, or false otherwise */ ( /* === Parameters ======================================================= */ - Index n_row, /* number of rows of A */ - Index n_col, /* number of columns of A */ - Colamd_Row Row [], /* of size n_row+1 */ - colamd_col Col [], /* of size n_col+1 */ - Index A [], /* row indices of A, of size Alen */ - Index p [], /* pointers to columns in A, of size n_col+1 */ - Index stats [COLAMD_STATS] /* colamd statistics */ + IndexType n_row, /* number of rows of A */ + IndexType n_col, /* number of columns of A */ + Colamd_Row Row [], /* of size n_row+1 */ + colamd_col Col [], /* of size n_col+1 */ + IndexType A [], /* row indices of A, of size Alen */ + IndexType p [], /* pointers to columns in A, of size n_col+1 */ + IndexType stats [COLAMD_STATS] /* colamd statistics */ ) { /* === Local variables ================================================== */ - Index col ; /* a column index */ - Index row ; /* a row index */ - Index *cp ; /* a column pointer */ - Index *cp_end ; /* a pointer to the end of a column */ - Index *rp ; /* a row pointer */ - Index *rp_end ; /* a pointer to the end of a row */ - Index last_row ; /* previous row */ + IndexType col ; /* a column index */ + IndexType row ; /* a row index */ + IndexType *cp ; /* a column pointer */ + IndexType *cp_end ; /* a pointer to the end of a column */ + IndexType *rp ; /* a row pointer */ + IndexType *rp_end ; /* a pointer to the end of a row */ + IndexType last_row ; /* previous row */ /* === Initialize columns, and check column pointers ==================== */ @@ -701,40 +701,40 @@ static Index init_rows_cols /* returns true if OK, or false otherwise */ Kills dense or empty columns and rows, calculates an initial score for each column, and places all columns in the degree lists. Not user-callable. */ -template +template static void init_scoring ( /* === Parameters ======================================================= */ - Index n_row, /* number of rows of A */ - Index n_col, /* number of columns of A */ - Colamd_Row Row [], /* of size n_row+1 */ - colamd_col Col [], /* of size n_col+1 */ - Index A [], /* column form and row form of A */ - Index head [], /* of size n_col+1 */ + IndexType n_row, /* number of rows of A */ + IndexType n_col, /* number of columns of A */ + Colamd_Row Row [], /* of size n_row+1 */ + colamd_col Col [], /* of size n_col+1 */ + IndexType A [], /* column form and row form of A */ + IndexType head [], /* of size n_col+1 */ double knobs [COLAMD_KNOBS],/* parameters */ - Index *p_n_row2, /* number of non-dense, non-empty rows */ - Index *p_n_col2, /* number of non-dense, non-empty columns */ - Index *p_max_deg /* maximum row degree */ + IndexType *p_n_row2, /* number of non-dense, non-empty rows */ + IndexType *p_n_col2, /* number of non-dense, non-empty columns */ + IndexType *p_max_deg /* maximum row degree */ ) { /* === Local variables ================================================== */ - Index c ; /* a column index */ - Index r, row ; /* a row index */ - Index *cp ; /* a column pointer */ - Index deg ; /* degree of a row or column */ - Index *cp_end ; /* a pointer to the end of a column */ - Index *new_cp ; /* new column pointer */ - Index col_length ; /* length of pruned column */ - Index score ; /* current column score */ - Index n_col2 ; /* number of non-dense, non-empty columns */ - Index n_row2 ; /* number of non-dense, non-empty rows */ - Index dense_row_count ; /* remove rows with more entries than this */ - Index dense_col_count ; /* remove cols with more entries than this */ - Index min_score ; /* smallest column score */ - Index max_deg ; /* maximum row degree */ - Index next_col ; /* Used to add to degree list.*/ + IndexType c ; /* a column index */ + IndexType r, row ; /* a row index */ + IndexType *cp ; /* a column pointer */ + IndexType deg ; /* degree of a row or column */ + IndexType *cp_end ; /* a pointer to the end of a column */ + IndexType *new_cp ; /* new column pointer */ + IndexType col_length ; /* length of pruned column */ + IndexType score ; /* current column score */ + IndexType n_col2 ; /* number of non-dense, non-empty columns */ + IndexType n_row2 ; /* number of non-dense, non-empty rows */ + IndexType dense_row_count ; /* remove rows with more entries than this */ + IndexType dense_col_count ; /* remove cols with more entries than this */ + IndexType min_score ; /* smallest column score */ + IndexType max_deg ; /* maximum row degree */ + IndexType next_col ; /* Used to add to degree list.*/ /* === Extract knobs ==================================================== */ @@ -845,7 +845,7 @@ static void init_scoring score = COLAMD_MIN (score, n_col) ; } /* determine pruned column length */ - col_length = (Index) (new_cp - &A [Col [c].start]) ; + col_length = (IndexType) (new_cp - &A [Col [c].start]) ; if (col_length == 0) { /* a newly-made null column (all rows in this col are "dense" */ @@ -938,56 +938,56 @@ static void init_scoring (no supercolumns on input). Uses a minimum approximate column minimum degree ordering method. Not user-callable. */ -template -static Index find_ordering /* return the number of garbage collections */ +template +static IndexType find_ordering /* return the number of garbage collections */ ( /* === Parameters ======================================================= */ - Index n_row, /* number of rows of A */ - Index n_col, /* number of columns of A */ - Index Alen, /* size of A, 2*nnz + n_col or larger */ - Colamd_Row Row [], /* of size n_row+1 */ - colamd_col Col [], /* of size n_col+1 */ - Index A [], /* column form and row form of A */ - Index head [], /* of size n_col+1 */ - Index n_col2, /* Remaining columns to order */ - Index max_deg, /* Maximum row degree */ - Index pfree /* index of first free slot (2*nnz on entry) */ + IndexType n_row, /* number of rows of A */ + IndexType n_col, /* number of columns of A */ + IndexType Alen, /* size of A, 2*nnz + n_col or larger */ + Colamd_Row Row [], /* of size n_row+1 */ + colamd_col Col [], /* of size n_col+1 */ + IndexType A [], /* column form and row form of A */ + IndexType head [], /* of size n_col+1 */ + IndexType n_col2, /* Remaining columns to order */ + IndexType max_deg, /* Maximum row degree */ + IndexType pfree /* index of first free slot (2*nnz on entry) */ ) { /* === Local variables ================================================== */ - Index k ; /* current pivot ordering step */ - Index pivot_col ; /* current pivot column */ - Index *cp ; /* a column pointer */ - Index *rp ; /* a row pointer */ - Index pivot_row ; /* current pivot row */ - Index *new_cp ; /* modified column pointer */ - Index *new_rp ; /* modified row pointer */ - Index pivot_row_start ; /* pointer to start of pivot row */ - Index pivot_row_degree ; /* number of columns in pivot row */ - Index pivot_row_length ; /* number of supercolumns in pivot row */ - Index pivot_col_score ; /* score of pivot column */ - Index needed_memory ; /* free space needed for pivot row */ - Index *cp_end ; /* pointer to the end of a column */ - Index *rp_end ; /* pointer to the end of a row */ - Index row ; /* a row index */ - Index col ; /* a column index */ - Index max_score ; /* maximum possible score */ - Index cur_score ; /* score of current column */ + IndexType k ; /* current pivot ordering step */ + IndexType pivot_col ; /* current pivot column */ + IndexType *cp ; /* a column pointer */ + IndexType *rp ; /* a row pointer */ + IndexType pivot_row ; /* current pivot row */ + IndexType *new_cp ; /* modified column pointer */ + IndexType *new_rp ; /* modified row pointer */ + IndexType pivot_row_start ; /* pointer to start of pivot row */ + IndexType pivot_row_degree ; /* number of columns in pivot row */ + IndexType pivot_row_length ; /* number of supercolumns in pivot row */ + IndexType pivot_col_score ; /* score of pivot column */ + IndexType needed_memory ; /* free space needed for pivot row */ + IndexType *cp_end ; /* pointer to the end of a column */ + IndexType *rp_end ; /* pointer to the end of a row */ + IndexType row ; /* a row index */ + IndexType col ; /* a column index */ + IndexType max_score ; /* maximum possible score */ + IndexType cur_score ; /* score of current column */ unsigned int hash ; /* hash value for supernode detection */ - Index head_column ; /* head of hash bucket */ - Index first_col ; /* first column in hash bucket */ - Index tag_mark ; /* marker value for mark array */ - Index row_mark ; /* Row [row].shared2.mark */ - Index set_difference ; /* set difference size of row with pivot row */ - Index min_score ; /* smallest column score */ - Index col_thickness ; /* "thickness" (no. of columns in a supercol) */ - Index max_mark ; /* maximum value of tag_mark */ - Index pivot_col_thickness ; /* number of columns represented by pivot col */ - Index prev_col ; /* Used by Dlist operations. */ - Index next_col ; /* Used by Dlist operations. */ - Index ngarbage ; /* number of garbage collections performed */ + IndexType head_column ; /* head of hash bucket */ + IndexType first_col ; /* first column in hash bucket */ + IndexType tag_mark ; /* marker value for mark array */ + IndexType row_mark ; /* Row [row].shared2.mark */ + IndexType set_difference ; /* set difference size of row with pivot row */ + IndexType min_score ; /* smallest column score */ + IndexType col_thickness ; /* "thickness" (no. of columns in a supercol) */ + IndexType max_mark ; /* maximum value of tag_mark */ + IndexType pivot_col_thickness ; /* number of columns represented by pivot col */ + IndexType prev_col ; /* Used by Dlist operations. */ + IndexType next_col ; /* Used by Dlist operations. */ + IndexType ngarbage ; /* number of garbage collections performed */ /* === Initialization and clear mark ==================================== */ @@ -1277,7 +1277,7 @@ static Index find_ordering /* return the number of garbage collections */ } /* recompute the column's length */ - Col [col].length = (Index) (new_cp - &A [Col [col].start]) ; + Col [col].length = (IndexType) (new_cp - &A [Col [col].start]) ; /* === Further mass elimination ================================= */ @@ -1325,7 +1325,7 @@ static Index find_ordering /* return the number of garbage collections */ Col [col].shared4.hash_next = first_col ; /* save hash function in Col [col].shared3.hash */ - Col [col].shared3.hash = (Index) hash ; + Col [col].shared3.hash = (IndexType) hash ; COLAMD_ASSERT (COL_IS_ALIVE (col)) ; } } @@ -1420,7 +1420,7 @@ static Index find_ordering /* return the number of garbage collections */ /* update pivot row length to reflect any cols that were killed */ /* during super-col detection and mass elimination */ Row [pivot_row].start = pivot_row_start ; - Row [pivot_row].length = (Index) (new_rp - &A[pivot_row_start]) ; + Row [pivot_row].length = (IndexType) (new_rp - &A[pivot_row_start]) ; Row [pivot_row].shared1.degree = pivot_row_degree ; Row [pivot_row].shared2.mark = 0 ; /* pivot row is no longer dead */ @@ -1449,22 +1449,22 @@ static Index find_ordering /* return the number of garbage collections */ taken by this routine is O (n_col), that is, linear in the number of columns. Not user-callable. */ -template +template static inline void order_children ( /* === Parameters ======================================================= */ - Index n_col, /* number of columns of A */ - colamd_col Col [], /* of size n_col+1 */ - Index p [] /* p [0 ... n_col-1] is the column permutation*/ + IndexType n_col, /* number of columns of A */ + colamd_col Col [], /* of size n_col+1 */ + IndexType p [] /* p [0 ... n_col-1] is the column permutation*/ ) { /* === Local variables ================================================== */ - Index i ; /* loop counter for all columns */ - Index c ; /* column index */ - Index parent ; /* index of column's parent */ - Index order ; /* column's order */ + IndexType i ; /* loop counter for all columns */ + IndexType c ; /* column index */ + IndexType parent ; /* index of column's parent */ + IndexType order ; /* column's order */ /* === Order each non-principal column ================================== */ @@ -1550,33 +1550,33 @@ static inline void order_children just been computed in the approximate degree computation. Not user-callable. */ -template +template static void detect_super_cols ( /* === Parameters ======================================================= */ - colamd_col Col [], /* of size n_col+1 */ - Index A [], /* row indices of A */ - Index head [], /* head of degree lists and hash buckets */ - Index row_start, /* pointer to set of columns to check */ - Index row_length /* number of columns to check */ + colamd_col Col [], /* of size n_col+1 */ + IndexType A [], /* row indices of A */ + IndexType head [], /* head of degree lists and hash buckets */ + IndexType row_start, /* pointer to set of columns to check */ + IndexType row_length /* number of columns to check */ ) { /* === Local variables ================================================== */ - Index hash ; /* hash value for a column */ - Index *rp ; /* pointer to a row */ - Index c ; /* a column index */ - Index super_c ; /* column index of the column to absorb into */ - Index *cp1 ; /* column pointer for column super_c */ - Index *cp2 ; /* column pointer for column c */ - Index length ; /* length of column super_c */ - Index prev_c ; /* column preceding c in hash bucket */ - Index i ; /* loop counter */ - Index *rp_end ; /* pointer to the end of the row */ - Index col ; /* a column index in the row to check */ - Index head_column ; /* first column in hash bucket or degree list */ - Index first_col ; /* first column in hash bucket */ + IndexType hash ; /* hash value for a column */ + IndexType *rp ; /* pointer to a row */ + IndexType c ; /* a column index */ + IndexType super_c ; /* column index of the column to absorb into */ + IndexType *cp1 ; /* column pointer for column super_c */ + IndexType *cp2 ; /* column pointer for column c */ + IndexType length ; /* length of column super_c */ + IndexType prev_c ; /* column preceding c in hash bucket */ + IndexType i ; /* loop counter */ + IndexType *rp_end ; /* pointer to the end of the row */ + IndexType col ; /* a column index in the row to check */ + IndexType head_column ; /* first column in hash bucket or degree list */ + IndexType first_col ; /* first column in hash bucket */ /* === Consider each column in the row ================================== */ @@ -1701,27 +1701,27 @@ static void detect_super_cols itself linear in the number of nonzeros in the input matrix. Not user-callable. */ -template -static Index garbage_collection /* returns the new value of pfree */ +template +static IndexType garbage_collection /* returns the new value of pfree */ ( /* === Parameters ======================================================= */ - Index n_row, /* number of rows */ - Index n_col, /* number of columns */ - Colamd_Row Row [], /* row info */ - colamd_col Col [], /* column info */ - Index A [], /* A [0 ... Alen-1] holds the matrix */ - Index *pfree /* &A [0] ... pfree is in use */ + IndexType n_row, /* number of rows */ + IndexType n_col, /* number of columns */ + Colamd_Row Row [], /* row info */ + colamd_col Col [], /* column info */ + IndexType A [], /* A [0 ... Alen-1] holds the matrix */ + IndexType *pfree /* &A [0] ... pfree is in use */ ) { /* === Local variables ================================================== */ - Index *psrc ; /* source pointer */ - Index *pdest ; /* destination pointer */ - Index j ; /* counter */ - Index r ; /* a row index */ - Index c ; /* a column index */ - Index length ; /* length of a row or column */ + IndexType *psrc ; /* source pointer */ + IndexType *pdest ; /* destination pointer */ + IndexType j ; /* counter */ + IndexType r ; /* a row index */ + IndexType c ; /* a column index */ + IndexType length ; /* length of a row or column */ /* === Defragment the columns =========================================== */ @@ -1734,7 +1734,7 @@ static Index garbage_collection /* returns the new value of pfree */ /* move and compact the column */ COLAMD_ASSERT (pdest <= psrc) ; - Col [c].start = (Index) (pdest - &A [0]) ; + Col [c].start = (IndexType) (pdest - &A [0]) ; length = Col [c].length ; for (j = 0 ; j < length ; j++) { @@ -1744,7 +1744,7 @@ static Index garbage_collection /* returns the new value of pfree */ *pdest++ = r ; } } - Col [c].length = (Index) (pdest - &A [Col [c].start]) ; + Col [c].length = (IndexType) (pdest - &A [Col [c].start]) ; } } @@ -1791,7 +1791,7 @@ static Index garbage_collection /* returns the new value of pfree */ /* move and compact the row */ COLAMD_ASSERT (pdest <= psrc) ; - Row [r].start = (Index) (pdest - &A [0]) ; + Row [r].start = (IndexType) (pdest - &A [0]) ; length = Row [r].length ; for (j = 0 ; j < length ; j++) { @@ -1801,7 +1801,7 @@ static Index garbage_collection /* returns the new value of pfree */ *pdest++ = c ; } } - Row [r].length = (Index) (pdest - &A [Row [r].start]) ; + Row [r].length = (IndexType) (pdest - &A [Row [r].start]) ; } } @@ -1810,7 +1810,7 @@ static Index garbage_collection /* returns the new value of pfree */ /* === Return the new value of pfree ==================================== */ - return ((Index) (pdest - &A [0])) ; + return ((IndexType) (pdest - &A [0])) ; } @@ -1822,18 +1822,18 @@ static Index garbage_collection /* returns the new value of pfree */ Clears the Row [].shared2.mark array, and returns the new tag_mark. Return value is the new tag_mark. Not user-callable. */ -template -static inline Index clear_mark /* return the new value for tag_mark */ +template +static inline IndexType clear_mark /* return the new value for tag_mark */ ( /* === Parameters ======================================================= */ - Index n_row, /* number of rows in A */ - Colamd_Row Row [] /* Row [0 ... n_row-1].shared2.mark is set to zero */ + IndexType n_row, /* number of rows in A */ + Colamd_Row Row [] /* Row [0 ... n_row-1].shared2.mark is set to zero */ ) { /* === Local variables ================================================== */ - Index r ; + IndexType r ; for (r = 0 ; r < n_row ; r++) { diff --git a/Eigen/src/OrderingMethods/Ordering.h b/Eigen/src/OrderingMethods/Ordering.h index f3c31f9cb..e88e637a4 100644 --- a/Eigen/src/OrderingMethods/Ordering.h +++ b/Eigen/src/OrderingMethods/Ordering.h @@ -111,12 +111,12 @@ class NaturalOrdering * Functor computing the \em column \em approximate \em minimum \em degree ordering * The matrix should be in column-major and \b compressed format (see SparseMatrix::makeCompressed()). */ -template +template class COLAMDOrdering { public: - typedef PermutationMatrix PermutationType; - typedef Matrix IndexVector; + typedef PermutationMatrix PermutationType; + typedef Matrix IndexVector; /** Compute the permutation vector \a perm form the sparse matrix \a mat * \warning The input sparse matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()). @@ -126,26 +126,26 @@ class COLAMDOrdering { eigen_assert(mat.isCompressed() && "COLAMDOrdering requires a sparse matrix in compressed mode. Call .makeCompressed() before passing it to COLAMDOrdering"); - Index m = mat.rows(); - Index n = mat.cols(); - Index nnz = mat.nonZeros(); + StorageIndex m = StorageIndex(mat.rows()); + StorageIndex n = StorageIndex(mat.cols()); + StorageIndex nnz = StorageIndex(mat.nonZeros()); // Get the recommended value of Alen to be used by colamd - Index Alen = internal::colamd_recommended(nnz, m, n); + StorageIndex Alen = internal::colamd_recommended(nnz, m, n); // Set the default parameters double knobs [COLAMD_KNOBS]; - Index stats [COLAMD_STATS]; + StorageIndex stats [COLAMD_STATS]; internal::colamd_set_defaults(knobs); IndexVector p(n+1), A(Alen); - for(Index i=0; i <= n; i++) p(i) = mat.outerIndexPtr()[i]; - for(Index i=0; i < nnz; i++) A(i) = mat.innerIndexPtr()[i]; + for(StorageIndex i=0; i <= n; i++) p(i) = mat.outerIndexPtr()[i]; + for(StorageIndex i=0; i < nnz; i++) A(i) = mat.innerIndexPtr()[i]; // Call Colamd routine to compute the ordering - Index info = internal::colamd(m, n, Alen, A.data(), p.data(), knobs, stats); + StorageIndex info = internal::colamd(m, n, Alen, A.data(), p.data(), knobs, stats); EIGEN_UNUSED_VARIABLE(info); eigen_assert( info && "COLAMD failed " ); perm.resize(n); - for (Index i = 0; i < n; i++) perm.indices()(p(i)) = i; + for (StorageIndex i = 0; i < n; i++) perm.indices()(p(i)) = i; } }; diff --git a/Eigen/src/PaStiXSupport/PaStiXSupport.h b/Eigen/src/PaStiXSupport/PaStiXSupport.h index e20c9ba2a..4e73edf5b 100644 --- a/Eigen/src/PaStiXSupport/PaStiXSupport.h +++ b/Eigen/src/PaStiXSupport/PaStiXSupport.h @@ -308,7 +308,7 @@ void PastixBase::analyzePattern(ColSpMatrix& mat) if(m_size>0) clean(); - m_size = mat.rows(); + m_size = internal::convert_index(mat.rows()); m_perm.resize(m_size); m_invp.resize(m_size); @@ -337,7 +337,7 @@ void PastixBase::factorize(ColSpMatrix& mat) eigen_assert(m_analysisIsOk && "The analysis phase should be called before the factorization phase"); m_iparm(IPARM_START_TASK) = API_TASK_NUMFACT; m_iparm(IPARM_END_TASK) = API_TASK_NUMFACT; - m_size = mat.rows(); + m_size = internal::convert_index(mat.rows()); internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, m_size, mat.outerIndexPtr(), mat.innerIndexPtr(), mat.valuePtr(), m_perm.data(), m_invp.data(), 0, 0, m_iparm.data(), m_dparm.data()); @@ -373,7 +373,7 @@ bool PastixBase::_solve_impl(const MatrixBase &b, MatrixBase &x m_iparm[IPARM_START_TASK] = API_TASK_SOLVE; m_iparm[IPARM_END_TASK] = API_TASK_REFINE; - internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, x.rows(), 0, 0, 0, + internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, internal::convert_index(x.rows()), 0, 0, 0, m_perm.data(), m_invp.data(), &x(0, i), rhs, m_iparm.data(), m_dparm.data()); } diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky.h b/Eigen/src/SparseCholesky/SimplicialCholesky.h index 2580151de..e2d7f95f2 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky.h @@ -202,7 +202,7 @@ class SimplicialCholeskyBase : public SparseSolverBase void factorize(const MatrixType& a) { eigen_assert(a.rows()==a.cols()); - int size = a.cols(); + Index size = a.cols(); CholMatrixType tmp(size,size); ConstCholMatrixPtr pmat; @@ -226,7 +226,7 @@ class SimplicialCholeskyBase : public SparseSolverBase void analyzePattern(const MatrixType& a, bool doLDLT) { eigen_assert(a.rows()==a.cols()); - int size = a.cols(); + Index size = a.cols(); CholMatrixType tmp(size,size); ConstCholMatrixPtr pmat; ordering(a, pmat, tmp); diff --git a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h index 9e2e878e0..31e06995b 100644 --- a/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h +++ b/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h @@ -50,14 +50,14 @@ namespace Eigen { template void SimplicialCholeskyBase::analyzePattern_preordered(const CholMatrixType& ap, bool doLDLT) { - const Index size = ap.rows(); + const StorageIndex size = StorageIndex(ap.rows()); m_matrix.resize(size, size); m_parent.resize(size); m_nonZerosPerCol.resize(size); - ei_declare_aligned_stack_constructed_variable(Index, tags, size, 0); + ei_declare_aligned_stack_constructed_variable(StorageIndex, tags, size, 0); - for(Index k = 0; k < size; ++k) + for(StorageIndex k = 0; k < size; ++k) { /* L(k,:) pattern: all nodes reachable in etree from nz in A(0:k-1,k) */ m_parent[k] = -1; /* parent of k is not yet known */ @@ -65,7 +65,7 @@ void SimplicialCholeskyBase::analyzePattern_preordered(const CholMatrix m_nonZerosPerCol[k] = 0; /* count of nonzeros in column k of L */ for(typename CholMatrixType::InnerIterator it(ap,k); it; ++it) { - Index i = it.index(); + StorageIndex i = it.index(); if(i < k) { /* follow path from i to root of etree, stop at flagged node */ @@ -84,7 +84,7 @@ void SimplicialCholeskyBase::analyzePattern_preordered(const CholMatrix /* construct Lp index array from m_nonZerosPerCol column counts */ StorageIndex* Lp = m_matrix.outerIndexPtr(); Lp[0] = 0; - for(Index k = 0; k < size; ++k) + for(StorageIndex k = 0; k < size; ++k) Lp[k+1] = Lp[k] + m_nonZerosPerCol[k] + (doLDLT ? 0 : 1); m_matrix.resizeNonZeros(Lp[size]); @@ -104,10 +104,10 @@ void SimplicialCholeskyBase::factorize_preordered(const CholMatrixType& eigen_assert(m_analysisIsOk && "You must first call analyzePattern()"); eigen_assert(ap.rows()==ap.cols()); - const Index size = ap.rows(); - eigen_assert(m_parent.size()==size); - eigen_assert(m_nonZerosPerCol.size()==size); + eigen_assert(m_parent.size()==ap.rows()); + eigen_assert(m_nonZerosPerCol.size()==ap.rows()); + const StorageIndex size = StorageIndex(ap.rows()); const StorageIndex* Lp = m_matrix.outerIndexPtr(); StorageIndex* Li = m_matrix.innerIndexPtr(); Scalar* Lx = m_matrix.valuePtr(); @@ -119,16 +119,16 @@ void SimplicialCholeskyBase::factorize_preordered(const CholMatrixType& bool ok = true; m_diag.resize(DoLDLT ? size : 0); - for(Index k = 0; k < size; ++k) + for(StorageIndex k = 0; k < size; ++k) { // compute nonzero pattern of kth row of L, in topological order y[k] = 0.0; // Y(0:k) is now all zero - Index top = size; // stack for pattern is empty + StorageIndex top = size; // stack for pattern is empty tags[k] = k; // mark node k as visited m_nonZerosPerCol[k] = 0; // count of nonzeros in column k of L for(typename CholMatrixType::InnerIterator it(ap,k); it; ++it) { - Index i = it.index(); + StorageIndex i = it.index(); if(i <= k) { y[i] += numext::conj(it.value()); /* scatter A(i,k) into Y (sum duplicates) */ diff --git a/Eigen/src/SparseCore/CompressedStorage.h b/Eigen/src/SparseCore/CompressedStorage.h index bba8a104b..454462ad5 100644 --- a/Eigen/src/SparseCore/CompressedStorage.h +++ b/Eigen/src/SparseCore/CompressedStorage.h @@ -131,7 +131,7 @@ class CompressedStorage /** \returns the stored value at index \a key * If the value does not exist, then the value \a defaultValue is returned without any insertion. */ - inline Scalar at(Index key, const Scalar& defaultValue = Scalar(0)) const + inline Scalar at(StorageIndex key, const Scalar& defaultValue = Scalar(0)) const { if (m_size==0) return defaultValue; @@ -144,7 +144,7 @@ class CompressedStorage } /** Like at(), but the search is performed in the range [start,end) */ - inline Scalar atInRange(Index start, Index end, Index key, const Scalar &defaultValue = Scalar(0)) const + inline Scalar atInRange(Index start, Index end, StorageIndex key, const Scalar &defaultValue = Scalar(0)) const { if (start>=end) return defaultValue; @@ -159,7 +159,7 @@ class CompressedStorage /** \returns a reference to the value at index \a key * If the value does not exist, then the value \a defaultValue is inserted * such that the keys are sorted. */ - inline Scalar& atWithInsertion(Index key, const Scalar& defaultValue = Scalar(0)) + inline Scalar& atWithInsertion(StorageIndex key, const Scalar& defaultValue = Scalar(0)) { Index id = searchLowerIndex(0,m_size,key); if (id>=m_size || m_indices[id]!=key) @@ -189,7 +189,7 @@ class CompressedStorage internal::smart_memmove(m_indices+id, m_indices+m_size, m_indices+id+1); } m_size++; - m_indices[id] = convert_index(key); + m_indices[id] = key; m_values[id] = defaultValue; } return m_values[id]; diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index 5256bf950..b8604a219 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -149,10 +149,10 @@ public: // update innerNonZeros if(!m_matrix.isCompressed()) for(Index j=0; j int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowElt, typename MatrixType::StorageIndex *perm=0) { typedef typename MatrixType::StorageIndex StorageIndex; - Index nc = mat.cols(); // Number of columns - Index m = mat.rows(); - Index diagSize = (std::min)(nc,m); + StorageIndex nc = convert_index(mat.cols()); // Number of columns + StorageIndex m = convert_index(mat.rows()); + StorageIndex diagSize = (std::min)(nc,m); IndexVector root(nc); // root of subtree of etree root.setZero(); IndexVector pp(nc); // disjoint sets pp.setZero(); // Initialize disjoint sets parent.resize(mat.cols()); //Compute first nonzero column in each row - StorageIndex row,col; firstRowElt.resize(m); firstRowElt.setConstant(nc); firstRowElt.segment(0, diagSize).setLinSpaced(diagSize, 0, diagSize-1); bool found_diag; - for (col = 0; col < nc; col++) + for (StorageIndex col = 0; col < nc; col++) { - Index pcol = col; + StorageIndex pcol = col; if(perm) pcol = perm[col]; for (typename MatrixType::InnerIterator it(mat, pcol); it; ++it) { - row = it.row(); + Index row = it.row(); firstRowElt(row) = (std::min)(firstRowElt(row), col); } } @@ -89,8 +88,8 @@ int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowEl except use (firstRowElt[r],c) in place of an edge (r,c) of A. Thus each row clique in A'*A is replaced by a star centered at its first vertex, which has the same fill. */ - Index rset, cset, rroot; - for (col = 0; col < nc; col++) + StorageIndex rset, cset, rroot; + for (StorageIndex col = 0; col < nc; col++) { found_diag = col>=m; pp(col) = col; @@ -99,7 +98,7 @@ int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowEl parent(col) = nc; /* The diagonal element is treated here even if it does not exist in the matrix * hence the loop is executed once more */ - Index pcol = col; + StorageIndex pcol = col; if(perm) pcol = perm[col]; for (typename MatrixType::InnerIterator it(mat, pcol); it||!found_diag; ++it) { // A sequence of interleaved find and union is performed @@ -107,7 +106,7 @@ int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowEl if(it) i = it.index(); if (i == col) found_diag = true; - row = firstRowElt(i); + StorageIndex row = firstRowElt(i); if (row >= col) continue; rset = internal::etree_find(row, pp); // Find the name of the set containing row rroot = root(rset); @@ -128,9 +127,10 @@ int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowEl * This routine was contributed by Cédric Doucet, CEDRAT Group, Meylan, France. */ template -void nr_etdfs (Index n, IndexVector& parent, IndexVector& first_kid, IndexVector& next_kid, IndexVector& post, Index postnum) +void nr_etdfs (typename IndexVector::Scalar n, IndexVector& parent, IndexVector& first_kid, IndexVector& next_kid, IndexVector& post, typename IndexVector::Scalar postnum) { - Index current = n, first, next; + typedef typename IndexVector::Scalar StorageIndex; + StorageIndex current = n, first, next; while (postnum != n) { // No kid for the current node @@ -175,21 +175,21 @@ void nr_etdfs (Index n, IndexVector& parent, IndexVector& first_kid, IndexVector * \param post postordered tree */ template -void treePostorder(Index n, IndexVector& parent, IndexVector& post) +void treePostorder(typename IndexVector::Scalar n, IndexVector& parent, IndexVector& post) { + typedef typename IndexVector::Scalar StorageIndex; IndexVector first_kid, next_kid; // Linked list of children - Index postnum; + StorageIndex postnum; // Allocate storage for working arrays and results first_kid.resize(n+1); next_kid.setZero(n+1); post.setZero(n+1); // Set up structure describing children - Index v, dad; first_kid.setConstant(-1); - for (v = n-1; v >= 0; v--) + for (StorageIndex v = n-1; v >= 0; v--) { - dad = parent(v); + StorageIndex dad = parent(v); next_kid(v) = first_kid(dad); first_kid(dad) = v; } diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 3cfd7ae9b..4cf4f1826 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -188,7 +188,7 @@ class SparseMatrix const Index outer = IsRowMajor ? row : col; const Index inner = IsRowMajor ? col : row; Index end = m_innerNonZeros ? m_outerIndex[outer] + m_innerNonZeros[outer] : m_outerIndex[outer+1]; - return m_data.atInRange(m_outerIndex[outer], end, inner); + return m_data.atInRange(m_outerIndex[outer], end, StorageIndex(inner)); } /** \returns a non-const reference to the value of the matrix at position \a i, \a j @@ -211,7 +211,7 @@ class SparseMatrix eigen_assert(end>=start && "you probably called coeffRef on a non finalized matrix"); if(end<=start) return insert(row,col); - const Index p = m_data.searchLowerIndex(start,end-1,inner); + const Index p = m_data.searchLowerIndex(start,end-1,StorageIndex(inner)); if((pm_data.resize(rows()); - Eigen::Map(&this->m_data.index(0), rows()).setLinSpaced(0, rows()-1); + Eigen::Map(&this->m_data.index(0), rows()).setLinSpaced(0, StorageIndex(rows()-1)); Eigen::Map(&this->m_data.value(0), rows()).setOnes(); - Eigen::Map(this->m_outerIndex, rows()+1).setLinSpaced(0, rows()); + Eigen::Map(this->m_outerIndex, rows()+1).setLinSpaced(0, StorageIndex(rows())); } inline SparseMatrix& operator=(const SparseMatrix& other) { diff --git a/Eigen/src/SparseCore/SparsePermutation.h b/Eigen/src/SparseCore/SparsePermutation.h index 80e5c5fef..4be93c18c 100644 --- a/Eigen/src/SparseCore/SparsePermutation.h +++ b/Eigen/src/SparseCore/SparsePermutation.h @@ -61,7 +61,7 @@ struct permut_sparsematrix_product_retval for(Index j=0; j tmp(size,tmpCols); Eigen::Matrix tmpX(size,tmpCols); - for(int k=0; k(rhsCols-k, NbColsAtOnce); + Index actualCols = std::min(rhsCols-k, NbColsAtOnce); tmp.leftCols(actualCols) = rhs.middleCols(k,actualCols); tmpX.leftCols(actualCols) = dec.solve(tmp.leftCols(actualCols)); dest.middleCols(k,actualCols) = tmpX.leftCols(actualCols).sparseView(); diff --git a/Eigen/src/SparseCore/SparseVector.h b/Eigen/src/SparseCore/SparseVector.h index b1cc4df77..35bcec819 100644 --- a/Eigen/src/SparseCore/SparseVector.h +++ b/Eigen/src/SparseCore/SparseVector.h @@ -103,7 +103,7 @@ class SparseVector inline Scalar coeff(Index i) const { eigen_assert(i>=0 && i=0 && i::analyzePattern(const MatrixType& mat) if (!m_symmetricmode) { IndexVector post, iwork; // Post order etree - internal::treePostorder(m_mat.cols(), m_etree, post); + internal::treePostorder(StorageIndex(m_mat.cols()), m_etree, post); // Renumber etree in postorder @@ -479,7 +479,7 @@ void SparseLU::factorize(const MatrixType& matrix) else { //FIXME This should not be needed if the empty permutation is handled transparently m_perm_c.resize(matrix.cols()); - for(Index i = 0; i < matrix.cols(); ++i) m_perm_c.indices()(i) = i; + for(StorageIndex i = 0; i < matrix.cols(); ++i) m_perm_c.indices()(i) = i; } Index m = m_mat.rows(); diff --git a/Eigen/src/SparseLU/SparseLUImpl.h b/Eigen/src/SparseLU/SparseLUImpl.h index e735fd5c8..731d1652c 100644 --- a/Eigen/src/SparseLU/SparseLUImpl.h +++ b/Eigen/src/SparseLU/SparseLUImpl.h @@ -40,7 +40,7 @@ class SparseLUImpl Index snode_bmod (const Index jcol, const Index fsupc, ScalarVector& dense, GlobalLU_t& glu); Index pivotL(const Index jcol, const RealScalar& diagpivotthresh, IndexVector& perm_r, IndexVector& iperm_c, Index& pivrow, GlobalLU_t& glu); template - void dfs_kernel(const Index jj, IndexVector& perm_r, + void dfs_kernel(const StorageIndex jj, IndexVector& perm_r, Index& nseg, IndexVector& panel_lsub, IndexVector& segrep, Ref repfnz_col, IndexVector& xprune, Ref marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu, Index& nextl_col, Index krow, Traits& traits); diff --git a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h index f7ffc2d9c..b37b93cf1 100644 --- a/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h +++ b/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h @@ -178,11 +178,11 @@ class MappedSuperNodalMatrix * \brief InnerIterator class to iterate over nonzero values of the current column in the supernodal matrix L * */ -template -class MappedSuperNodalMatrix::InnerIterator +template +class MappedSuperNodalMatrix::InnerIterator { public: - InnerIterator(const MappedSuperNodalMatrix& mat, Eigen::Index outer) + InnerIterator(const MappedSuperNodalMatrix& mat, Index outer) : m_matrix(mat), m_outer(outer), m_supno(mat.colToSup()[outer]), diff --git a/Eigen/src/SparseLU/SparseLU_Utils.h b/Eigen/src/SparseLU/SparseLU_Utils.h index b48157d9f..9e3dab44d 100644 --- a/Eigen/src/SparseLU/SparseLU_Utils.h +++ b/Eigen/src/SparseLU/SparseLU_Utils.h @@ -53,7 +53,7 @@ void SparseLUImpl::fixupL(const Index n, const IndexVector& { Index fsupc, i, j, k, jstart; - Index nextl = 0; + StorageIndex nextl = 0; Index nsuper = (glu.supno)(n); // For each supernode diff --git a/Eigen/src/SparseLU/SparseLU_column_bmod.h b/Eigen/src/SparseLU/SparseLU_column_bmod.h index bda01dcb3..be190997d 100644 --- a/Eigen/src/SparseLU/SparseLU_column_bmod.h +++ b/Eigen/src/SparseLU/SparseLU_column_bmod.h @@ -138,7 +138,7 @@ Index SparseLUImpl::column_bmod(const Index jcol, const Ind glu.lusup.segment(nextlu,offset).setZero(); nextlu += offset; } - glu.xlusup(jcol + 1) = nextlu; // close L\U(*,jcol); + glu.xlusup(jcol + 1) = StorageIndex(nextlu); // close L\U(*,jcol); /* For more updates within the panel (also within the current supernode), * should start from the first column of the panel, or the first column diff --git a/Eigen/src/SparseLU/SparseLU_column_dfs.h b/Eigen/src/SparseLU/SparseLU_column_dfs.h index 17c9e6adb..c98b30e32 100644 --- a/Eigen/src/SparseLU/SparseLU_column_dfs.h +++ b/Eigen/src/SparseLU/SparseLU_column_dfs.h @@ -112,13 +112,13 @@ Index SparseLUImpl::column_dfs(const Index m, const Index j // krow was visited before, go to the next nonz; if (kmark == jcol) continue; - dfs_kernel(jcol, perm_r, nseg, glu.lsub, segrep, repfnz, xprune, marker2, parent, + dfs_kernel(StorageIndex(jcol), perm_r, nseg, glu.lsub, segrep, repfnz, xprune, marker2, parent, xplore, glu, nextl, krow, traits); } // for each nonzero ... - Index fsupc, jptr, jm1ptr, ito, ifrom, istop; - Index nsuper = glu.supno(jcol); - Index jcolp1 = jcol + 1; + Index fsupc; + StorageIndex nsuper = glu.supno(jcol); + StorageIndex jcolp1 = StorageIndex(jcol) + 1; Index jcolm1 = jcol - 1; // check to see if j belongs in the same supernode as j-1 @@ -129,8 +129,8 @@ Index SparseLUImpl::column_dfs(const Index m, const Index j else { fsupc = glu.xsup(nsuper); - jptr = glu.xlsub(jcol); // Not yet compressed - jm1ptr = glu.xlsub(jcolm1); + StorageIndex jptr = glu.xlsub(jcol); // Not yet compressed + StorageIndex jm1ptr = glu.xlsub(jcolm1); // Use supernodes of type T2 : see SuperLU paper if ( (nextl-jptr != jptr-jm1ptr-1) ) jsuper = emptyIdxLU; @@ -148,13 +148,13 @@ Index SparseLUImpl::column_dfs(const Index m, const Index j { // starts a new supernode if ( (fsupc < jcolm1-1) ) { // >= 3 columns in nsuper - ito = glu.xlsub(fsupc+1); + StorageIndex ito = glu.xlsub(fsupc+1); glu.xlsub(jcolm1) = ito; - istop = ito + jptr - jm1ptr; + StorageIndex istop = ito + jptr - jm1ptr; xprune(jcolm1) = istop; // intialize xprune(jcol-1) glu.xlsub(jcol) = istop; - for (ifrom = jm1ptr; ifrom < nextl; ++ifrom, ++ito) + for (StorageIndex ifrom = jm1ptr; ifrom < nextl; ++ifrom, ++ito) glu.lsub(ito) = glu.lsub(ifrom); nextl = ito; // = istop + length(jcol) } @@ -166,8 +166,8 @@ Index SparseLUImpl::column_dfs(const Index m, const Index j // Tidy up the pointers before exit glu.xsup(nsuper+1) = jcolp1; glu.supno(jcolp1) = nsuper; - xprune(jcol) = nextl; // Intialize upper bound for pruning - glu.xlsub(jcolp1) = nextl; + xprune(jcol) = StorageIndex(nextl); // Intialize upper bound for pruning + glu.xlsub(jcolp1) = StorageIndex(nextl); return 0; } diff --git a/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h b/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h index bf237951d..c32d8d8b1 100644 --- a/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h +++ b/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h @@ -56,7 +56,7 @@ Index SparseLUImpl::copy_to_ucol(const Index jcol, const In // For each nonzero supernode segment of U[*,j] in topological order Index k = nseg - 1, i; - Index nextu = glu.xusub(jcol); + StorageIndex nextu = glu.xusub(jcol); Index kfnz, isub, segsize; Index new_next,irow; Index fsupc, mem; diff --git a/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h b/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h index 4092f842f..6f75d500e 100644 --- a/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h +++ b/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h @@ -48,15 +48,14 @@ void SparseLUImpl::heap_relax_snode (const Index n, IndexVe // The etree may not be postordered, but its heap ordered IndexVector post; - internal::treePostorder(n, et, post); // Post order etree + internal::treePostorder(StorageIndex(n), et, post); // Post order etree IndexVector inv_post(n+1); - Index i; - for (i = 0; i < n+1; ++i) inv_post(post(i)) = i; // inv_post = post.inverse()??? + for (StorageIndex i = 0; i < n+1; ++i) inv_post(post(i)) = i; // inv_post = post.inverse()??? // Renumber etree in postorder IndexVector iwork(n); IndexVector et_save(n+1); - for (i = 0; i < n; ++i) + for (Index i = 0; i < n; ++i) { iwork(post(i)) = post(et(i)); } @@ -78,7 +77,7 @@ void SparseLUImpl::heap_relax_snode (const Index n, IndexVe StorageIndex k; Index nsuper_et_post = 0; // Number of relaxed snodes in postordered etree Index nsuper_et = 0; // Number of relaxed snodes in the original etree - Index l; + StorageIndex l; for (j = 0; j < n; ) { parent = et(j); @@ -90,8 +89,8 @@ void SparseLUImpl::heap_relax_snode (const Index n, IndexVe } // Found a supernode in postordered etree, j is the last column ++nsuper_et_post; - k = n; - for (i = snode_start; i <= j; ++i) + k = StorageIndex(n); + for (Index i = snode_start; i <= j; ++i) k = (std::min)(k, inv_post(i)); l = inv_post(j); if ( (l - k) == (j - snode_start) ) // Same number of columns in the snode @@ -102,7 +101,7 @@ void SparseLUImpl::heap_relax_snode (const Index n, IndexVe } else { - for (i = snode_start; i <= j; ++i) + for (Index i = snode_start; i <= j; ++i) { l = inv_post(i); if (descendants(i) == 0) diff --git a/Eigen/src/SparseLU/SparseLU_panel_dfs.h b/Eigen/src/SparseLU/SparseLU_panel_dfs.h index f4a908ee5..155df7336 100644 --- a/Eigen/src/SparseLU/SparseLU_panel_dfs.h +++ b/Eigen/src/SparseLU/SparseLU_panel_dfs.h @@ -41,7 +41,7 @@ struct panel_dfs_traits panel_dfs_traits(Index jcol, StorageIndex* marker) : m_jcol(jcol), m_marker(marker) {} - bool update_segrep(Index krep, Index jj) + bool update_segrep(Index krep, StorageIndex jj) { if(m_marker[krep] template -void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& perm_r, +void SparseLUImpl::dfs_kernel(const StorageIndex jj, IndexVector& perm_r, Index& nseg, IndexVector& panel_lsub, IndexVector& segrep, Ref repfnz_col, IndexVector& xprune, Ref marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu, @@ -67,14 +67,14 @@ void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& ) { - Index kmark = marker(krow); + StorageIndex kmark = marker(krow); // For each unmarked krow of jj marker(krow) = jj; - Index kperm = perm_r(krow); + StorageIndex kperm = perm_r(krow); if (kperm == emptyIdxLU ) { // krow is in L : place it in structure of L(*, jj) - panel_lsub(nextl_col++) = krow; // krow is indexed into A + panel_lsub(nextl_col++) = StorageIndex(krow); // krow is indexed into A traits.mem_expand(panel_lsub, nextl_col, kmark); } @@ -83,9 +83,9 @@ void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& // krow is in U : if its supernode-representative krep // has been explored, update repfnz(*) // krep = supernode representative of the current row - Index krep = glu.xsup(glu.supno(kperm)+1) - 1; + StorageIndex krep = glu.xsup(glu.supno(kperm)+1) - 1; // First nonzero element in the current column: - Index myfnz = repfnz_col(krep); + StorageIndex myfnz = repfnz_col(krep); if (myfnz != emptyIdxLU ) { @@ -96,26 +96,26 @@ void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& else { // Otherwise, perform dfs starting at krep - Index oldrep = emptyIdxLU; + StorageIndex oldrep = emptyIdxLU; parent(krep) = oldrep; repfnz_col(krep) = kperm; - Index xdfs = glu.xlsub(krep); + StorageIndex xdfs = glu.xlsub(krep); Index maxdfs = xprune(krep); - Index kpar; + StorageIndex kpar; do { // For each unmarked kchild of krep while (xdfs < maxdfs) { - Index kchild = glu.lsub(xdfs); + StorageIndex kchild = glu.lsub(xdfs); xdfs++; - Index chmark = marker(kchild); + StorageIndex chmark = marker(kchild); if (chmark != jj ) { marker(kchild) = jj; - Index chperm = perm_r(kchild); + StorageIndex chperm = perm_r(kchild); if (chperm == emptyIdxLU) { @@ -128,7 +128,7 @@ void SparseLUImpl::dfs_kernel(const Index jj, IndexVector& // case kchild is in U : // chrep = its supernode-rep. If its rep has been explored, // update its repfnz(*) - Index chrep = glu.xsup(glu.supno(chperm)+1) - 1; + StorageIndex chrep = glu.xsup(glu.supno(chperm)+1) - 1; myfnz = repfnz_col(chrep); if (myfnz != emptyIdxLU) @@ -227,7 +227,7 @@ void SparseLUImpl::panel_dfs(const Index m, const Index w, panel_dfs_traits traits(jcol, marker1.data()); // For each column in the panel - for (Index jj = jcol; jj < jcol + w; jj++) + for (StorageIndex jj = StorageIndex(jcol); jj < jcol + w; jj++) { nextl_col = (jj - jcol) * m; @@ -241,7 +241,7 @@ void SparseLUImpl::panel_dfs(const Index m, const Index w, Index krow = it.row(); dense_col(krow) = it.value(); - Index kmark = marker(krow); + StorageIndex kmark = marker(krow); if (kmark == jj) continue; // krow visited before, go to the next nonzero diff --git a/Eigen/src/SparseLU/SparseLU_pivotL.h b/Eigen/src/SparseLU/SparseLU_pivotL.h index 01f5ba4e9..562128b69 100644 --- a/Eigen/src/SparseLU/SparseLU_pivotL.h +++ b/Eigen/src/SparseLU/SparseLU_pivotL.h @@ -89,7 +89,7 @@ Index SparseLUImpl::pivotL(const Index jcol, const RealScal // Test for singularity if ( pivmax == 0.0 ) { pivrow = lsub_ptr[pivptr]; - perm_r(pivrow) = jcol; + perm_r(pivrow) = StorageIndex(jcol); return (jcol+1); } @@ -110,7 +110,7 @@ Index SparseLUImpl::pivotL(const Index jcol, const RealScal } // Record pivot row - perm_r(pivrow) = jcol; + perm_r(pivrow) = StorageIndex(jcol); // Interchange row subscripts if (pivptr != nsupc ) { diff --git a/Eigen/src/SparseLU/SparseLU_pruneL.h b/Eigen/src/SparseLU/SparseLU_pruneL.h index 13133fcc2..ad32fed5e 100644 --- a/Eigen/src/SparseLU/SparseLU_pruneL.h +++ b/Eigen/src/SparseLU/SparseLU_pruneL.h @@ -124,7 +124,7 @@ void SparseLUImpl::pruneL(const Index jcol, const IndexVect } } // end while - xprune(irep) = kmin; //Pruning + xprune(irep) = StorageIndex(kmin); //Pruning } // end if do_prune } // end pruning } // End for each U-segment diff --git a/Eigen/src/SparseLU/SparseLU_relax_snode.h b/Eigen/src/SparseLU/SparseLU_relax_snode.h index 21c182d56..c408d01b4 100644 --- a/Eigen/src/SparseLU/SparseLU_relax_snode.h +++ b/Eigen/src/SparseLU/SparseLU_relax_snode.h @@ -48,10 +48,10 @@ void SparseLUImpl::relax_snode (const Index n, IndexVector& { // compute the number of descendants of each node in the etree - Index j, parent; + Index parent; relax_end.setConstant(emptyIdxLU); descendants.setZero(); - for (j = 0; j < n; j++) + for (Index j = 0; j < n; j++) { parent = et(j); if (parent != n) // not the dummy root @@ -59,7 +59,7 @@ void SparseLUImpl::relax_snode (const Index n, IndexVector& } // Identify the relaxed supernodes by postorder traversal of the etree Index snode_start; // beginning of a snode - for (j = 0; j < n; ) + for (Index j = 0; j < n; ) { parent = et(j); snode_start = j; @@ -69,7 +69,7 @@ void SparseLUImpl::relax_snode (const Index n, IndexVector& parent = et(j); } // Found a supernode in postordered etree, j is the last column - relax_end(snode_start) = j; // Record last column + relax_end(snode_start) = StorageIndex(j); // Record last column j++; // Search for a new leaf while (descendants(j) != 0 && j < n) j++; diff --git a/Eigen/src/SparseQR/SparseQR.h b/Eigen/src/SparseQR/SparseQR.h index 920b884e5..ce4a70454 100644 --- a/Eigen/src/SparseQR/SparseQR.h +++ b/Eigen/src/SparseQR/SparseQR.h @@ -296,7 +296,7 @@ void SparseQR::analyzePattern(const MatrixType& mat) if (!m_perm_c.size()) { m_perm_c.resize(n); - m_perm_c.indices().setLinSpaced(n, 0,n-1); + m_perm_c.indices().setLinSpaced(n, 0,StorageIndex(n-1)); } // Compute the column elimination tree of the permuted matrix @@ -327,8 +327,8 @@ void SparseQR::factorize(const MatrixType& mat) using std::abs; eigen_assert(m_analysisIsok && "analyzePattern() should be called before this step"); - StorageIndex m = mat.rows(); - StorageIndex n = mat.cols(); + StorageIndex m = StorageIndex(mat.rows()); + StorageIndex n = StorageIndex(mat.cols()); StorageIndex diagSize = (std::min)(m,n); IndexVector mark((std::max)(m,n)); mark.setConstant(-1); // Record the visited nodes IndexVector Ridx(n), Qidx(m); // Store temporarily the row indexes for the current column of R and Q @@ -406,7 +406,7 @@ void SparseQR::factorize(const MatrixType& mat) for (typename QRMatrixType::InnerIterator itp(m_pmat, col); itp || !found_diag; ++itp) { StorageIndex curIdx = nonzeroCol; - if(itp) curIdx = itp.row(); + if(itp) curIdx = StorageIndex(itp.row()); if(curIdx == nonzeroCol) found_diag = true; // Get the nonzeros indexes of the current column of R @@ -467,7 +467,7 @@ void SparseQR::factorize(const MatrixType& mat) { for (typename QRMatrixType::InnerIterator itq(m_Q, curIdx); itq; ++itq) { - StorageIndex iQ = itq.row(); + StorageIndex iQ = StorageIndex(itq.row()); if (mark(iQ) != col) { Qidx(nzcolQ++) = iQ; // Add this row to the pattern of Q, diff --git a/Eigen/src/UmfPackSupport/UmfPackSupport.h b/Eigen/src/UmfPackSupport/UmfPackSupport.h index dcbd4ab71..3d30403c7 100644 --- a/Eigen/src/UmfPackSupport/UmfPackSupport.h +++ b/Eigen/src/UmfPackSupport/UmfPackSupport.h @@ -313,8 +313,9 @@ class UmfPackLU : public SparseSolverBase > void analyzePattern_impl() { int errorCode = 0; - errorCode = umfpack_symbolic(m_copyMatrix.rows(), m_copyMatrix.cols(), m_outerIndexPtr, m_innerIndexPtr, m_valuePtr, - &m_symbolic, 0, 0); + errorCode = umfpack_symbolic(internal::convert_index(m_copyMatrix.rows()), + internal::convert_index(m_copyMatrix.cols()), + m_outerIndexPtr, m_innerIndexPtr, m_valuePtr, &m_symbolic, 0, 0); m_isInitialized = true; m_info = errorCode ? InvalidInput : Success; diff --git a/test/sparse_basic.cpp b/test/sparse_basic.cpp index b06956974..8021f4db6 100644 --- a/test/sparse_basic.cpp +++ b/test/sparse_basic.cpp @@ -398,8 +398,8 @@ template void sparse_basic(const SparseMatrixType& re refMat.setZero(); for(Index i=0;i(0,rows-1); - StorageIndex c = internal::random(0,cols-1); + StorageIndex r = internal::random(0,StorageIndex(rows-1)); + StorageIndex c = internal::random(0,StorageIndex(cols-1)); Scalar v = internal::random(); triplets.push_back(TripletType(r,c,v)); refMat(r,c) += v; diff --git a/test/spqr_support.cpp b/test/spqr_support.cpp index 901c42c40..baa25a0c2 100644 --- a/test/spqr_support.cpp +++ b/test/spqr_support.cpp @@ -37,7 +37,7 @@ template void test_spqr_scalar() SPQR solver; generate_sparse_rectangular_problem(A,dA); - int m = A.rows(); + Index m = A.rows(); b = DenseVector::Random(m); solver.compute(A); if (solver.info() != Success) diff --git a/unsupported/Eigen/src/IterativeSolvers/GMRES.h b/unsupported/Eigen/src/IterativeSolvers/GMRES.h index 873f2bf2a..3e733e053 100644 --- a/unsupported/Eigen/src/IterativeSolvers/GMRES.h +++ b/unsupported/Eigen/src/IterativeSolvers/GMRES.h @@ -54,7 +54,7 @@ namespace internal { */ template bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Preconditioner & precond, - int &iters, const int &restart, typename Dest::RealScalar & tol_error) { + Index &iters, const Index &restart, typename Dest::RealScalar & tol_error) { using std::sqrt; using std::abs; @@ -65,10 +65,10 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition typedef Matrix < Scalar, Dynamic, Dynamic > FMatrixType; RealScalar tol = tol_error; - const int maxIters = iters; + const Index maxIters = iters; iters = 0; - const int m = mat.rows(); + const Index m = mat.rows(); // residual and preconditioned residual const VectorType p0 = rhs - mat*x; @@ -97,14 +97,14 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition w(0)=(Scalar) beta; H.bottomLeftCorner(m - 1, 1) = e; - for (int k = 1; k <= restart; ++k) { + for (Index k = 1; k <= restart; ++k) { ++iters; VectorType v = VectorType::Unit(m, k - 1), workspace(m); // apply Householder reflections H_{1} ... H_{k-1} to v - for (int i = k - 1; i >= 0; --i) { + for (Index i = k - 1; i >= 0; --i) { v.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data()); } @@ -113,7 +113,7 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition v=precond.solve(t); // apply Householder reflections H_{k-1} ... H_{1} to v - for (int i = 0; i < k; ++i) { + for (Index i = 0; i < k; ++i) { v.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data()); } @@ -133,7 +133,7 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition } if (k > 1) { - for (int i = 0; i < k - 1; ++i) { + for (Index i = 0; i < k - 1; ++i) { // apply old Givens rotations to v v.applyOnTheLeft(i, i + 1, G[i].adjoint()); } @@ -166,7 +166,7 @@ bool gmres(const MatrixType & mat, const Rhs & rhs, Dest & x, const Precondition // apply Householder reflection H_{k} to x_new x_new.tail(m - k + 1).applyHouseholderOnTheLeft(H.col(k - 1).tail(m - k), tau.coeffRef(k - 1), workspace.data()); - for (int i = k - 2; i >= 0; --i) { + for (Index i = k - 2; i >= 0; --i) { x_new += y(i) * VectorType::Unit(m, i); // apply Householder reflection H_{i} to x_new x_new.tail(m - i).applyHouseholderOnTheLeft(H.col(i).tail(m - i - 1), tau.coeffRef(i), workspace.data()); @@ -265,7 +265,7 @@ class GMRES : public IterativeSolverBase > using Base::m_isInitialized; private: - int m_restart; + Index m_restart; public: using Base::_solve_impl; @@ -295,19 +295,19 @@ public: /** Get the number of iterations after that a restart is performed. */ - int get_restart() { return m_restart; } + Index get_restart() { return m_restart; } /** Set the number of iterations after that a restart is performed. * \param restart number of iterations for a restarti, default is 30. */ - void set_restart(const int restart) { m_restart=restart; } + void set_restart(const Index restart) { m_restart=restart; } /** \internal */ template void _solve_with_guess_impl(const Rhs& b, Dest& x) const { bool failed = false; - for(int j=0; j EIGEN_DONT_INLINE void minres(const MatrixType& mat, const Rhs& rhs, Dest& x, - const Preconditioner& precond, int& iters, + const Preconditioner& precond, Index& iters, typename Dest::RealScalar& tol_error) { using std::sqrt; @@ -48,8 +48,8 @@ namespace Eigen { } // initialize - const int maxIters(iters); // initialize maxIters to iters - const int N(mat.cols()); // the size of the matrix + const Index maxIters(iters); // initialize maxIters to iters + const Index N(mat.cols()); // the size of the matrix const RealScalar threshold2(tol_error*tol_error*rhsNorm2); // convergence threshold (compared to residualNorm2) // Initialize preconditioned Lanczos From cc641aabb710ab002b6c641bd5c3e1deed0e634d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 14:46:51 +0100 Subject: [PATCH 1093/1103] Remove deprecated usage of expr::Index. --- Eigen/src/Cholesky/LDLT.h | 8 +-- Eigen/src/Cholesky/LLT.h | 19 +++--- Eigen/src/Cholesky/LLT_MKL.h | 10 +-- Eigen/src/Core/Assign_MKL.h | 1 - Eigen/src/Core/CwiseNullaryOp.h | 1 - Eigen/src/Core/DenseCoeffsBase.h | 6 +- Eigen/src/Core/GeneralProduct.h | 4 -- Eigen/src/Core/IO.h | 1 - Eigen/src/Core/Map.h | 2 - Eigen/src/Core/MapBase.h | 2 +- Eigen/src/Core/Redux.h | 4 -- Eigen/src/Core/SolveTriangular.h | 4 +- Eigen/src/Core/Transpose.h | 1 - Eigen/src/Core/Transpositions.h | 68 ++++++++----------- Eigen/src/Core/VectorwiseOp.h | 2 +- Eigen/src/Core/Visitor.h | 5 -- Eigen/src/Core/products/GeneralMatrixMatrix.h | 1 - .../products/GeneralMatrixMatrixTriangular.h | 3 - .../Core/products/SelfadjointMatrixMatrix.h | 1 - .../Core/products/SelfadjointMatrixVector.h | 1 - Eigen/src/Core/products/SelfadjointProduct.h | 2 - .../Core/products/TriangularMatrixMatrix.h | 1 - .../Core/products/TriangularMatrixVector.h | 2 - Eigen/src/Core/util/Macros.h | 1 - Eigen/src/Eigenvalues/ComplexEigenSolver.h | 2 +- Eigen/src/Eigenvalues/ComplexSchur.h | 2 +- Eigen/src/Eigenvalues/EigenSolver.h | 2 +- .../src/Eigenvalues/GeneralizedEigenSolver.h | 2 +- .../GeneralizedSelfAdjointEigenSolver.h | 1 - .../src/Eigenvalues/HessenbergDecomposition.h | 3 +- Eigen/src/Eigenvalues/RealQZ.h | 6 +- Eigen/src/Eigenvalues/RealSchur.h | 4 +- .../src/Eigenvalues/SelfAdjointEigenSolver.h | 7 +- Eigen/src/Eigenvalues/Tridiagonalization.h | 5 +- Eigen/src/Geometry/Homogeneous.h | 2 - Eigen/src/Geometry/OrthoMethods.h | 1 - Eigen/src/Geometry/Transform.h | 3 +- Eigen/src/Geometry/Umeyama.h | 1 - Eigen/src/Householder/BlockHouseholder.h | 3 - .../IterativeLinearSolvers/SolveWithGuess.h | 1 - Eigen/src/Jacobi/Jacobi.h | 5 +- Eigen/src/LU/PartialPivLU.h | 7 +- Eigen/src/QR/FullPivHouseholderQR.h | 1 - Eigen/src/QR/HouseholderQR.h | 5 +- Eigen/src/QR/HouseholderQR_MKL.h | 3 +- Eigen/src/SVD/BDCSVD.h | 1 - Eigen/src/SVD/JacobiSVD.h | 12 ---- Eigen/src/SVD/SVDBase.h | 1 + Eigen/src/SVD/UpperBidiagonalization.h | 9 +-- Eigen/src/SparseCore/SparseBlock.h | 2 +- Eigen/src/SparseCore/SparseView.h | 3 +- Eigen/src/SuperLUSupport/SuperLUSupport.h | 1 - Eigen/src/misc/Image.h | 2 - Eigen/src/misc/Kernel.h | 2 - 54 files changed, 80 insertions(+), 169 deletions(-) diff --git a/Eigen/src/Cholesky/LDLT.h b/Eigen/src/Cholesky/LDLT.h index e2a6ca2b2..f46f7b758 100644 --- a/Eigen/src/Cholesky/LDLT.h +++ b/Eigen/src/Cholesky/LDLT.h @@ -59,7 +59,7 @@ template class LDLT }; typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef typename MatrixType::StorageIndex StorageIndex; typedef Matrix TmpMatrixType; @@ -252,8 +252,7 @@ template<> struct ldlt_inplace using std::abs; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; - typedef typename TranspositionType::StorageIndexType IndexType; + typedef typename TranspositionType::StorageIndex IndexType; eigen_assert(mat.rows()==mat.cols()); const Index size = mat.rows(); @@ -343,7 +342,6 @@ template<> struct ldlt_inplace using numext::isfinite; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; const Index size = mat.rows(); eigen_assert(mat.cols() == size && w.size()==size); @@ -451,7 +449,7 @@ template template LDLT& LDLT::rankUpdate(const MatrixBase& w, const typename NumTraits::Real& sigma) { - typedef typename TranspositionType::StorageIndexType IndexType; + typedef typename TranspositionType::StorageIndex IndexType; const Index size = w.rows(); if (m_isInitialized) { diff --git a/Eigen/src/Cholesky/LLT.h b/Eigen/src/Cholesky/LLT.h index 5e0cf6c88..629c87161 100644 --- a/Eigen/src/Cholesky/LLT.h +++ b/Eigen/src/Cholesky/LLT.h @@ -59,7 +59,7 @@ template class LLT }; typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef typename MatrixType::StorageIndex StorageIndex; enum { @@ -184,12 +184,11 @@ namespace internal { template struct llt_inplace; template -static typename MatrixType::Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) +static Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) { using std::sqrt; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; typedef typename MatrixType::ColXpr ColXpr; typedef typename internal::remove_all::type ColXprCleaned; typedef typename ColXprCleaned::SegmentReturnType ColXprSegment; @@ -258,10 +257,9 @@ template struct llt_inplace { typedef typename NumTraits::Real RealScalar; template - static typename MatrixType::Index unblocked(MatrixType& mat) + static Index unblocked(MatrixType& mat) { using std::sqrt; - typedef typename MatrixType::Index Index; eigen_assert(mat.rows()==mat.cols()); const Index size = mat.rows(); @@ -285,9 +283,8 @@ template struct llt_inplace } template - static typename MatrixType::Index blocked(MatrixType& m) + static Index blocked(MatrixType& m) { - typedef typename MatrixType::Index Index; eigen_assert(m.rows()==m.cols()); Index size = m.rows(); if(size<32) @@ -318,7 +315,7 @@ template struct llt_inplace } template - static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma) + static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma) { return Eigen::internal::llt_rank_update_lower(mat, vec, sigma); } @@ -329,19 +326,19 @@ template struct llt_inplace typedef typename NumTraits::Real RealScalar; template - static EIGEN_STRONG_INLINE typename MatrixType::Index unblocked(MatrixType& mat) + static EIGEN_STRONG_INLINE Index unblocked(MatrixType& mat) { Transpose matt(mat); return llt_inplace::unblocked(matt); } template - static EIGEN_STRONG_INLINE typename MatrixType::Index blocked(MatrixType& mat) + static EIGEN_STRONG_INLINE Index blocked(MatrixType& mat) { Transpose matt(mat); return llt_inplace::blocked(matt); } template - static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma) + static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma) { Transpose matt(mat); return llt_inplace::rankUpdate(matt, vec.conjugate(), sigma); diff --git a/Eigen/src/Cholesky/LLT_MKL.h b/Eigen/src/Cholesky/LLT_MKL.h index 64daa445c..09bf59d43 100644 --- a/Eigen/src/Cholesky/LLT_MKL.h +++ b/Eigen/src/Cholesky/LLT_MKL.h @@ -46,7 +46,7 @@ template struct mkl_llt; template<> struct mkl_llt \ { \ template \ - static inline typename MatrixType::Index potrf(MatrixType& m, char uplo) \ + static inline Index potrf(MatrixType& m, char uplo) \ { \ lapack_int matrix_order; \ lapack_int size, lda, info, StorageOrder; \ @@ -67,23 +67,23 @@ template<> struct mkl_llt \ template<> struct llt_inplace \ { \ template \ - static typename MatrixType::Index blocked(MatrixType& m) \ + static Index blocked(MatrixType& m) \ { \ return mkl_llt::potrf(m, 'L'); \ } \ template \ - static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \ + static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \ { return Eigen::internal::llt_rank_update_lower(mat, vec, sigma); } \ }; \ template<> struct llt_inplace \ { \ template \ - static typename MatrixType::Index blocked(MatrixType& m) \ + static Index blocked(MatrixType& m) \ { \ return mkl_llt::potrf(m, 'U'); \ } \ template \ - static typename MatrixType::Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \ + static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \ { \ Transpose matt(mat); \ return llt_inplace::rankUpdate(matt, vec.conjugate(), sigma); \ diff --git a/Eigen/src/Core/Assign_MKL.h b/Eigen/src/Core/Assign_MKL.h index 97134ffd7..a7b9e9293 100644 --- a/Eigen/src/Core/Assign_MKL.h +++ b/Eigen/src/Core/Assign_MKL.h @@ -84,7 +84,6 @@ template { typedef typename Derived1::Scalar Scalar; - typedef typename Derived1::Index Index; static inline void run(Derived1& dst, const CwiseUnaryOp& src) { // in case we want to (or have to) skip VML at runtime we can call: diff --git a/Eigen/src/Core/CwiseNullaryOp.h b/Eigen/src/Core/CwiseNullaryOp.h index 05c4fedd0..009fd845d 100644 --- a/Eigen/src/Core/CwiseNullaryOp.h +++ b/Eigen/src/Core/CwiseNullaryOp.h @@ -753,7 +753,6 @@ struct setIdentity_impl template struct setIdentity_impl { - typedef typename Derived::Index Index; EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Derived& run(Derived& m) { diff --git a/Eigen/src/Core/DenseCoeffsBase.h b/Eigen/src/Core/DenseCoeffsBase.h index 569fed956..f08380bed 100644 --- a/Eigen/src/Core/DenseCoeffsBase.h +++ b/Eigen/src/Core/DenseCoeffsBase.h @@ -583,14 +583,14 @@ namespace internal { template struct first_aligned_impl { - static inline typename Derived::Index run(const Derived&) + static inline Index run(const Derived&) { return 0; } }; template struct first_aligned_impl { - static inline typename Derived::Index run(const Derived& m) + static inline Index run(const Derived& m) { return internal::first_aligned(&m.const_cast_derived().coeffRef(0,0), m.size()); } @@ -602,7 +602,7 @@ struct first_aligned_impl * documentation. */ template -static inline typename Derived::Index first_aligned(const Derived& m) +static inline Index first_aligned(const Derived& m) { return first_aligned_impl diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h index 81750722c..7027130e0 100644 --- a/Eigen/src/Core/GeneralProduct.h +++ b/Eigen/src/Core/GeneralProduct.h @@ -221,7 +221,6 @@ template<> struct gemv_dense_sense_selector template static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef typename Dest::Index Index; typedef typename Lhs::Scalar LhsScalar; typedef typename Rhs::Scalar RhsScalar; typedef typename Dest::Scalar ResScalar; @@ -298,7 +297,6 @@ template<> struct gemv_dense_sense_selector template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef typename Dest::Index Index; typedef typename Lhs::Scalar LhsScalar; typedef typename Rhs::Scalar RhsScalar; typedef typename Dest::Scalar ResScalar; @@ -352,7 +350,6 @@ template<> struct gemv_dense_sense_selector template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef typename Dest::Index Index; // TODO makes sure dest is sequentially stored in memory, otherwise use a temp const Index size = rhs.rows(); for(Index k=0; k struct gemv_dense_sense_selector template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef typename Dest::Index Index; // TODO makes sure rhs is sequentially stored in memory, otherwise use a temp const Index rows = dest.rows(); for(Index i=0; i > : public traits { typedef traits TraitsBase; - typedef typename PlainObjectType::Index Index; - typedef typename PlainObjectType::Scalar Scalar; enum { InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0 ? int(PlainObjectType::InnerStrideAtCompileTime) diff --git a/Eigen/src/Core/MapBase.h b/Eigen/src/Core/MapBase.h index 8dca9796d..acac74aa4 100644 --- a/Eigen/src/Core/MapBase.h +++ b/Eigen/src/Core/MapBase.h @@ -178,7 +178,7 @@ template class MapBase typedef typename Base::Scalar Scalar; typedef typename Base::PacketScalar PacketScalar; - typedef typename Base::Index Index; + typedef typename Base::StorageIndex StorageIndex; typedef typename Base::PointerType PointerType; using Base::derived; diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h index f6546917e..1a0a00481 100644 --- a/Eigen/src/Core/Redux.h +++ b/Eigen/src/Core/Redux.h @@ -191,7 +191,6 @@ template struct redux_impl { typedef typename Derived::Scalar Scalar; - typedef typename Derived::Index Index; EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func) { @@ -217,7 +216,6 @@ struct redux_impl { typedef typename Derived::Scalar Scalar; typedef typename packet_traits::type PacketScalar; - typedef typename Derived::Index Index; static Scalar run(const Derived &mat, const Func& func) { @@ -275,7 +273,6 @@ struct redux_impl { typedef typename Derived::Scalar Scalar; typedef typename packet_traits::type PacketScalar; - typedef typename Derived::Index Index; EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func) { @@ -342,7 +339,6 @@ public: typedef _XprType XprType; EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} - typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; typedef typename XprType::PacketScalar PacketScalar; diff --git a/Eigen/src/Core/SolveTriangular.h b/Eigen/src/Core/SolveTriangular.h index f97048bda..9bac726f7 100644 --- a/Eigen/src/Core/SolveTriangular.h +++ b/Eigen/src/Core/SolveTriangular.h @@ -68,7 +68,7 @@ struct triangular_solver_selector if(!useRhsDirectly) MappedRhs(actualRhs,rhs.size()) = rhs; - triangular_solve_vector ::run(actualLhs.cols(), actualLhs.data(), actualLhs.outerStride(), actualRhs); @@ -82,7 +82,6 @@ template struct triangular_solver_selector { typedef typename Rhs::Scalar Scalar; - typedef typename Rhs::Index Index; typedef blas_traits LhsProductTraits; typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType; @@ -232,7 +231,6 @@ template struct triangular_solv { typedef typename remove_all::type RhsNestedCleaned; typedef ReturnByValue Base; - typedef typename Base::Index Index; triangular_solve_retval(const TriangularType& tri, const Rhs& rhs) : m_triangularMatrix(tri), m_rhs(rhs) diff --git a/Eigen/src/Core/Transpose.h b/Eigen/src/Core/Transpose.h index 7e41769a3..2c967abca 100644 --- a/Eigen/src/Core/Transpose.h +++ b/Eigen/src/Core/Transpose.h @@ -232,7 +232,6 @@ struct inplace_transpose_selector { // PacketSize x Packet static void run(MatrixType& m) { typedef typename MatrixType::Scalar Scalar; typedef typename internal::packet_traits::type Packet; - typedef typename MatrixType::Index Index; const Index PacketSize = internal::packet_traits::size; const Index Alignment = internal::evaluator::Flags&AlignedBit ? Aligned : Unaligned; PacketBlock A; diff --git a/Eigen/src/Core/Transpositions.h b/Eigen/src/Core/Transpositions.h index 77e7d6f45..e9b54c2ba 100644 --- a/Eigen/src/Core/Transpositions.h +++ b/Eigen/src/Core/Transpositions.h @@ -53,8 +53,8 @@ class TranspositionsBase public: typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar StorageIndexType; - typedef typename IndicesType::Index Index; + typedef typename IndicesType::Scalar StorageIndex; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 Derived& derived() { return *static_cast(this); } const Derived& derived() const { return *static_cast(this); } @@ -82,17 +82,17 @@ class TranspositionsBase inline Index size() const { return indices().size(); } /** Direct access to the underlying index vector */ - inline const StorageIndexType& coeff(Index i) const { return indices().coeff(i); } + inline const StorageIndex& coeff(Index i) const { return indices().coeff(i); } /** Direct access to the underlying index vector */ - inline StorageIndexType& coeffRef(Index i) { return indices().coeffRef(i); } + inline StorageIndex& coeffRef(Index i) { return indices().coeffRef(i); } /** Direct access to the underlying index vector */ - inline const StorageIndexType& operator()(Index i) const { return indices()(i); } + inline const StorageIndex& operator()(Index i) const { return indices()(i); } /** Direct access to the underlying index vector */ - inline StorageIndexType& operator()(Index i) { return indices()(i); } + inline StorageIndex& operator()(Index i) { return indices()(i); } /** Direct access to the underlying index vector */ - inline const StorageIndexType& operator[](Index i) const { return indices()(i); } + inline const StorageIndex& operator[](Index i) const { return indices()(i); } /** Direct access to the underlying index vector */ - inline StorageIndexType& operator[](Index i) { return indices()(i); } + inline StorageIndex& operator[](Index i) { return indices()(i); } /** const version of indices(). */ const IndicesType& indices() const { return derived().indices(); } @@ -108,7 +108,7 @@ class TranspositionsBase /** Sets \c *this to represents an identity transformation */ void setIdentity() { - for(StorageIndexType i = 0; i < indices().size(); ++i) + for(StorageIndex i = 0; i < indices().size(); ++i) coeffRef(i) = i; } @@ -145,26 +145,23 @@ class TranspositionsBase }; namespace internal { -template -struct traits > +template +struct traits > { - typedef Matrix<_StorageIndexType, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; - typedef typename IndicesType::Index Index; - typedef _StorageIndexType StorageIndexType; + typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType; + typedef _StorageIndex StorageIndex; }; } -template -class Transpositions : public TranspositionsBase > +template +class Transpositions : public TranspositionsBase > { typedef internal::traits Traits; public: typedef TranspositionsBase Base; typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar StorageIndexType; - typedef typename IndicesType::Index Index; - + typedef typename IndicesType::Scalar StorageIndex; inline Transpositions() {} @@ -219,32 +216,30 @@ class Transpositions : public TranspositionsBase -struct traits,_PacketAccess> > +template +struct traits,_PacketAccess> > { - typedef Map, _PacketAccess> IndicesType; - typedef typename IndicesType::Index Index; - typedef _StorageIndexType StorageIndexType; + typedef Map, _PacketAccess> IndicesType; + typedef _StorageIndex StorageIndex; }; } -template -class Map,PacketAccess> - : public TranspositionsBase,PacketAccess> > +template +class Map,PacketAccess> + : public TranspositionsBase,PacketAccess> > { typedef internal::traits Traits; public: typedef TranspositionsBase Base; typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar StorageIndexType; - typedef typename IndicesType::Index Index; + typedef typename IndicesType::Scalar StorageIndex; - explicit inline Map(const StorageIndexType* indicesPtr) + explicit inline Map(const StorageIndex* indicesPtr) : m_indices(indicesPtr) {} - inline Map(const StorageIndexType* indicesPtr, Index size) + inline Map(const StorageIndex* indicesPtr, Index size) : m_indices(indicesPtr,size) {} @@ -281,8 +276,7 @@ namespace internal { template struct traits > { - typedef typename _IndicesType::Scalar StorageIndexType; - typedef typename _IndicesType::Index Index; + typedef typename _IndicesType::Scalar StorageIndex; typedef _IndicesType IndicesType; }; } @@ -296,8 +290,7 @@ class TranspositionsWrapper typedef TranspositionsBase Base; typedef typename Traits::IndicesType IndicesType; - typedef typename IndicesType::Scalar StorageIndexType; - typedef typename IndicesType::Index Index; + typedef typename IndicesType::Scalar StorageIndex; explicit inline TranspositionsWrapper(IndicesType& a_indices) : m_indices(a_indices) @@ -370,8 +363,7 @@ struct transposition_matrix_product_retval : public ReturnByValue > { typedef typename remove_all::type MatrixTypeNestedCleaned; - typedef typename TranspositionType::Index Index; - typedef typename TranspositionType::StorageIndexType StorageIndexType; + typedef typename TranspositionType::StorageIndex StorageIndex; transposition_matrix_product_retval(const TranspositionType& tr, const MatrixType& matrix) : m_transpositions(tr), m_matrix(matrix) @@ -383,7 +375,7 @@ struct transposition_matrix_product_retval template inline void evalTo(Dest& dst) const { const Index size = m_transpositions.size(); - StorageIndexType j = 0; + StorageIndex j = 0; if(!(is_same::value && extract_data(dst) == extract_data(m_matrix))) dst = m_matrix; diff --git a/Eigen/src/Core/VectorwiseOp.h b/Eigen/src/Core/VectorwiseOp.h index a626310ec..b3dc0c224 100644 --- a/Eigen/src/Core/VectorwiseOp.h +++ b/Eigen/src/Core/VectorwiseOp.h @@ -159,7 +159,7 @@ template class VectorwiseOp typedef typename ExpressionType::Scalar Scalar; typedef typename ExpressionType::RealScalar RealScalar; - typedef typename ExpressionType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef typename internal::conditional::ret, ExpressionType, ExpressionType&>::type ExpressionTypeNested; typedef typename internal::remove_all::type ExpressionTypeNestedCleaned; diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h index 02bd4eff3..6b1ecae8b 100644 --- a/Eigen/src/Core/Visitor.h +++ b/Eigen/src/Core/Visitor.h @@ -41,7 +41,6 @@ struct visitor_impl template struct visitor_impl { - typedef typename Derived::Index Index; static inline void run(const Derived& mat, Visitor& visitor) { visitor.init(mat.coeff(0,0), 0, 0); @@ -60,7 +59,6 @@ class visitor_evaluator public: explicit visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {} - typedef typename XprType::Index Index; typedef typename XprType::Scalar Scalar; typedef typename XprType::CoeffReturnType CoeffReturnType; @@ -124,7 +122,6 @@ namespace internal { template struct coeff_visitor { - typedef typename Derived::Index Index; typedef typename Derived::Scalar Scalar; Index row, col; Scalar res; @@ -144,7 +141,6 @@ struct coeff_visitor template struct min_coeff_visitor : coeff_visitor { - typedef typename Derived::Index Index; typedef typename Derived::Scalar Scalar; void operator() (const Scalar& value, Index i, Index j) { @@ -172,7 +168,6 @@ struct functor_traits > { template struct max_coeff_visitor : coeff_visitor { - typedef typename Derived::Index Index; typedef typename Derived::Scalar Scalar; void operator() (const Scalar& value, Index i, Index j) { diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index fd9443cd2..8210ea584 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -386,7 +386,6 @@ struct generic_product_impl : generic_product_impl_base > { typedef typename Product::Scalar Scalar; - typedef typename Product::Index Index; typedef typename Lhs::Scalar LhsScalar; typedef typename Rhs::Scalar RhsScalar; diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h index e55994900..a36eb2fe0 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h @@ -198,7 +198,6 @@ struct general_product_to_triangular_selector static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha) { typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; typedef typename internal::remove_all::type Lhs; typedef internal::blas_traits LhsBlasTraits; @@ -243,8 +242,6 @@ struct general_product_to_triangular_selector { static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha) { - typedef typename MatrixType::Index Index; - typedef typename internal::remove_all::type Lhs; typedef internal::blas_traits LhsBlasTraits; typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs; diff --git a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h index 4b6316d63..f84f54982 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixMatrix.h +++ b/Eigen/src/Core/products/SelfadjointMatrixMatrix.h @@ -474,7 +474,6 @@ template struct selfadjoint_product_impl { typedef typename Product::Scalar Scalar; - typedef typename Product::Index Index; typedef internal::blas_traits LhsBlasTraits; typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; diff --git a/Eigen/src/Core/products/SelfadjointMatrixVector.h b/Eigen/src/Core/products/SelfadjointMatrixVector.h index 372a44e47..5d6ef9913 100644 --- a/Eigen/src/Core/products/SelfadjointMatrixVector.h +++ b/Eigen/src/Core/products/SelfadjointMatrixVector.h @@ -174,7 +174,6 @@ template struct selfadjoint_product_impl { typedef typename Product::Scalar Scalar; - typedef typename Product::Index Index; typedef internal::blas_traits LhsBlasTraits; typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType; diff --git a/Eigen/src/Core/products/SelfadjointProduct.h b/Eigen/src/Core/products/SelfadjointProduct.h index 6ca4ae6c0..2af00058d 100644 --- a/Eigen/src/Core/products/SelfadjointProduct.h +++ b/Eigen/src/Core/products/SelfadjointProduct.h @@ -53,7 +53,6 @@ struct selfadjoint_product_selector static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha) { typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; typedef internal::blas_traits OtherBlasTraits; typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType; typedef typename internal::remove_all::type _ActualOtherType; @@ -86,7 +85,6 @@ struct selfadjoint_product_selector static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha) { typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; typedef internal::blas_traits OtherBlasTraits; typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType; typedef typename internal::remove_all::type _ActualOtherType; diff --git a/Eigen/src/Core/products/TriangularMatrixMatrix.h b/Eigen/src/Core/products/TriangularMatrixMatrix.h index 60c99dcd2..5f01eb5a8 100644 --- a/Eigen/src/Core/products/TriangularMatrixMatrix.h +++ b/Eigen/src/Core/products/TriangularMatrixMatrix.h @@ -388,7 +388,6 @@ struct triangular_product_impl { template static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha) { - typedef typename Dest::Index Index; typedef typename Dest::Scalar Scalar; typedef internal::blas_traits LhsBlasTraits; diff --git a/Eigen/src/Core/products/TriangularMatrixVector.h b/Eigen/src/Core/products/TriangularMatrixVector.h index 4d88a710b..7c014b72a 100644 --- a/Eigen/src/Core/products/TriangularMatrixVector.h +++ b/Eigen/src/Core/products/TriangularMatrixVector.h @@ -206,7 +206,6 @@ template struct trmv_selector template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef typename Dest::Index Index; typedef typename Lhs::Scalar LhsScalar; typedef typename Rhs::Scalar RhsScalar; typedef typename Dest::Scalar ResScalar; @@ -283,7 +282,6 @@ template struct trmv_selector template static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha) { - typedef typename Dest::Index Index; typedef typename Lhs::Scalar LhsScalar; typedef typename Rhs::Scalar RhsScalar; typedef typename Dest::Scalar ResScalar; diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 07923848a..e607cdd12 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -660,7 +660,6 @@ namespace Eigen { #define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \ EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \ typedef typename Base::PacketScalar PacketScalar; \ - typedef Eigen::Index Index; \ enum { MaxRowsAtCompileTime = Eigen::internal::traits::MaxRowsAtCompileTime, \ MaxColsAtCompileTime = Eigen::internal::traits::MaxColsAtCompileTime}; \ using Base::derived; \ diff --git a/Eigen/src/Eigenvalues/ComplexEigenSolver.h b/Eigen/src/Eigenvalues/ComplexEigenSolver.h index 25082546e..075a62848 100644 --- a/Eigen/src/Eigenvalues/ComplexEigenSolver.h +++ b/Eigen/src/Eigenvalues/ComplexEigenSolver.h @@ -60,7 +60,7 @@ template class ComplexEigenSolver /** \brief Scalar type for matrices of type #MatrixType. */ typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 /** \brief Complex scalar type for #MatrixType. * diff --git a/Eigen/src/Eigenvalues/ComplexSchur.h b/Eigen/src/Eigenvalues/ComplexSchur.h index a3a5a4649..993ee7e1e 100644 --- a/Eigen/src/Eigenvalues/ComplexSchur.h +++ b/Eigen/src/Eigenvalues/ComplexSchur.h @@ -63,7 +63,7 @@ template class ComplexSchur /** \brief Scalar type for matrices of type \p _MatrixType. */ typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 /** \brief Complex scalar type for \p _MatrixType. * diff --git a/Eigen/src/Eigenvalues/EigenSolver.h b/Eigen/src/Eigenvalues/EigenSolver.h index 9372021ff..a63a42341 100644 --- a/Eigen/src/Eigenvalues/EigenSolver.h +++ b/Eigen/src/Eigenvalues/EigenSolver.h @@ -79,7 +79,7 @@ template class EigenSolver /** \brief Scalar type for matrices of type #MatrixType. */ typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 /** \brief Complex scalar type for #MatrixType. * diff --git a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h index c20ea03e6..c9da6740a 100644 --- a/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h +++ b/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h @@ -72,7 +72,7 @@ template class GeneralizedEigenSolver /** \brief Scalar type for matrices of type #MatrixType. */ typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 /** \brief Complex scalar type for #MatrixType. * diff --git a/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h index 1ce1f5f58..5f6bb8289 100644 --- a/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h @@ -50,7 +50,6 @@ class GeneralizedSelfAdjointEigenSolver : public SelfAdjointEigenSolver<_MatrixT typedef SelfAdjointEigenSolver<_MatrixType> Base; public: - typedef typename Base::Index Index; typedef _MatrixType MatrixType; /** \brief Default constructor for fixed-size matrices. diff --git a/Eigen/src/Eigenvalues/HessenbergDecomposition.h b/Eigen/src/Eigenvalues/HessenbergDecomposition.h index 2615a9f23..87a5bcb69 100644 --- a/Eigen/src/Eigenvalues/HessenbergDecomposition.h +++ b/Eigen/src/Eigenvalues/HessenbergDecomposition.h @@ -71,7 +71,7 @@ template class HessenbergDecomposition /** \brief Scalar type for matrices of type #MatrixType. */ typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 /** \brief Type for vector of Householder coefficients. * @@ -337,7 +337,6 @@ namespace internal { template struct HessenbergDecompositionMatrixHReturnType : public ReturnByValue > { - typedef typename MatrixType::Index Index; public: /** \brief Constructor. * diff --git a/Eigen/src/Eigenvalues/RealQZ.h b/Eigen/src/Eigenvalues/RealQZ.h index 128ef9028..ca75f2f50 100644 --- a/Eigen/src/Eigenvalues/RealQZ.h +++ b/Eigen/src/Eigenvalues/RealQZ.h @@ -67,7 +67,7 @@ namespace Eigen { }; typedef typename MatrixType::Scalar Scalar; typedef std::complex::Real> ComplexScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef Matrix EigenvalueType; typedef Matrix ColumnVectorType; @@ -276,7 +276,7 @@ namespace Eigen { /** \internal Look for single small sub-diagonal element S(res, res-1) and return res (or 0) */ template - inline typename MatrixType::Index RealQZ::findSmallSubdiagEntry(Index iu) + inline Index RealQZ::findSmallSubdiagEntry(Index iu) { using std::abs; Index res = iu; @@ -294,7 +294,7 @@ namespace Eigen { /** \internal Look for single small diagonal element T(res, res) for res between f and l, and return res (or f-1) */ template - inline typename MatrixType::Index RealQZ::findSmallDiagEntry(Index f, Index l) + inline Index RealQZ::findSmallDiagEntry(Index f, Index l) { using std::abs; Index res = l; diff --git a/Eigen/src/Eigenvalues/RealSchur.h b/Eigen/src/Eigenvalues/RealSchur.h index 51e61ba38..60ade50a0 100644 --- a/Eigen/src/Eigenvalues/RealSchur.h +++ b/Eigen/src/Eigenvalues/RealSchur.h @@ -64,7 +64,7 @@ template class RealSchur }; typedef typename MatrixType::Scalar Scalar; typedef std::complex::Real> ComplexScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef Matrix EigenvalueType; typedef Matrix ColumnVectorType; @@ -343,7 +343,7 @@ inline typename MatrixType::Scalar RealSchur::computeNormOfT() /** \internal Look for single small sub-diagonal element and returns its index */ template -inline typename MatrixType::Index RealSchur::findSmallSubdiagEntry(Index iu) +inline Index RealSchur::findSmallSubdiagEntry(Index iu) { using std::abs; Index res = iu; diff --git a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h index 54f60b197..66d1154cf 100644 --- a/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h +++ b/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h @@ -21,7 +21,7 @@ class GeneralizedSelfAdjointEigenSolver; namespace internal { template struct direct_selfadjoint_eigenvalues; template -ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const typename MatrixType::Index maxIterations, bool computeEigenvectors, MatrixType& eivec); +ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec); } /** \eigenvalues_module \ingroup Eigenvalues_Module @@ -81,7 +81,7 @@ template class SelfAdjointEigenSolver /** \brief Scalar type for matrices of type \p _MatrixType. */ typedef typename MatrixType::Scalar Scalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 /** \brief Real scalar type for \p _MatrixType. * @@ -456,12 +456,11 @@ namespace internal { * \returns \c Success or \c NoConvergence */ template -ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const typename MatrixType::Index maxIterations, bool computeEigenvectors, MatrixType& eivec) +ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec) { using std::abs; ComputationInfo info; - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; Index n = diag.size(); diff --git a/Eigen/src/Eigenvalues/Tridiagonalization.h b/Eigen/src/Eigenvalues/Tridiagonalization.h index bedd1cb34..a6fb00b21 100644 --- a/Eigen/src/Eigenvalues/Tridiagonalization.h +++ b/Eigen/src/Eigenvalues/Tridiagonalization.h @@ -69,7 +69,7 @@ template class Tridiagonalization typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 enum { Size = MatrixType::RowsAtCompileTime, @@ -345,7 +345,6 @@ template void tridiagonalization_inplace(MatrixType& matA, CoeffVectorType& hCoeffs) { using numext::conj; - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; Index n = matA.rows(); @@ -437,7 +436,6 @@ struct tridiagonalization_inplace_selector { typedef typename Tridiagonalization::CoeffVectorType CoeffVectorType; typedef typename Tridiagonalization::HouseholderSequenceType HouseholderSequenceType; - typedef typename MatrixType::Index Index; template static void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType& subdiag, bool extractQ) { @@ -525,7 +523,6 @@ struct tridiagonalization_inplace_selector template struct TridiagonalizationMatrixTReturnType : public ReturnByValue > { - typedef typename MatrixType::Index Index; public: /** \brief Constructor. * diff --git a/Eigen/src/Geometry/Homogeneous.h b/Eigen/src/Geometry/Homogeneous.h index 7f1907542..f16451656 100644 --- a/Eigen/src/Geometry/Homogeneous.h +++ b/Eigen/src/Geometry/Homogeneous.h @@ -238,7 +238,6 @@ struct homogeneous_left_product_impl,Lhs> typedef typename traits::LhsMatrixType LhsMatrixType; typedef typename remove_all::type LhsMatrixTypeCleaned; typedef typename remove_all::type LhsMatrixTypeNested; - typedef typename MatrixType::Index Index; homogeneous_left_product_impl(const Lhs& lhs, const MatrixType& rhs) : m_lhs(take_matrix_for_product::run(lhs)), m_rhs(rhs) @@ -278,7 +277,6 @@ struct homogeneous_right_product_impl,Rhs> : public ReturnByValue,Rhs> > { typedef typename remove_all::type RhsNested; - typedef typename MatrixType::Index Index; homogeneous_right_product_impl(const MatrixType& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs) {} diff --git a/Eigen/src/Geometry/OrthoMethods.h b/Eigen/src/Geometry/OrthoMethods.h index a245c79d3..6b2e57392 100644 --- a/Eigen/src/Geometry/OrthoMethods.h +++ b/Eigen/src/Geometry/OrthoMethods.h @@ -133,7 +133,6 @@ struct unitOrthogonal_selector typedef typename plain_matrix_type::type VectorType; typedef typename traits::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename Derived::Index Index; typedef Matrix Vector2; EIGEN_DEVICE_FUNC static inline VectorType run(const Derived& src) diff --git a/Eigen/src/Geometry/Transform.h b/Eigen/src/Geometry/Transform.h index 34d337c90..8c9d7049b 100644 --- a/Eigen/src/Geometry/Transform.h +++ b/Eigen/src/Geometry/Transform.h @@ -205,8 +205,7 @@ public: /** the scalar type of the coefficients */ typedef _Scalar Scalar; typedef Eigen::Index StorageIndex; - /** \deprecated */ - typedef Eigen::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 /** type of the matrix used to represent the transformation */ typedef typename internal::make_proper_matrix_type::type MatrixType; /** constified MatrixType */ diff --git a/Eigen/src/Geometry/Umeyama.h b/Eigen/src/Geometry/Umeyama.h index 5e20662f8..8d9b7a154 100644 --- a/Eigen/src/Geometry/Umeyama.h +++ b/Eigen/src/Geometry/Umeyama.h @@ -97,7 +97,6 @@ umeyama(const MatrixBase& src, const MatrixBase& dst, boo typedef typename internal::umeyama_transform_matrix_type::type TransformationMatrixType; typedef typename internal::traits::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename Derived::Index Index; EIGEN_STATIC_ASSERT(!NumTraits::IsComplex, NUMERIC_TYPE_MUST_BE_REAL) EIGEN_STATIC_ASSERT((internal::is_same::Scalar>::value), diff --git a/Eigen/src/Householder/BlockHouseholder.h b/Eigen/src/Householder/BlockHouseholder.h index 35dbf80a1..39bf8c83d 100644 --- a/Eigen/src/Householder/BlockHouseholder.h +++ b/Eigen/src/Householder/BlockHouseholder.h @@ -21,7 +21,6 @@ namespace internal { // template // void make_block_householder_triangular_factor(TriangularFactorType& triFactor, const VectorsType& vectors, const CoeffsType& hCoeffs) // { -// typedef typename TriangularFactorType::Index Index; // typedef typename VectorsType::Scalar Scalar; // const Index nbVecs = vectors.cols(); // eigen_assert(triFactor.rows() == nbVecs && triFactor.cols() == nbVecs && vectors.rows()>=nbVecs); @@ -51,7 +50,6 @@ namespace internal { template void make_block_householder_triangular_factor(TriangularFactorType& triFactor, const VectorsType& vectors, const CoeffsType& hCoeffs) { - typedef typename TriangularFactorType::Index Index; const Index nbVecs = vectors.cols(); eigen_assert(triFactor.rows() == nbVecs && triFactor.cols() == nbVecs && vectors.rows()>=nbVecs); @@ -80,7 +78,6 @@ void make_block_householder_triangular_factor(TriangularFactorType& triFactor, c template void apply_block_householder_on_the_left(MatrixType& mat, const VectorsType& vectors, const CoeffsType& hCoeffs, bool forward) { - typedef typename MatrixType::Index Index; enum { TFactorSize = MatrixType::ColsAtCompileTime }; Index nbVecs = vectors.cols(); Matrix T(nbVecs,nbVecs); diff --git a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h index 251c6fa1a..ef7efc9cf 100644 --- a/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h +++ b/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h @@ -41,7 +41,6 @@ template class SolveWithGuess : public internal::generic_xpr_base, MatrixXpr, typename internal::traits::StorageKind>::type { public: - typedef typename RhsType::Index Index; typedef typename internal::traits::Scalar Scalar; typedef typename internal::traits::PlainObject PlainObject; typedef typename internal::generic_xpr_base, MatrixXpr, typename internal::traits::StorageKind>::type Base; diff --git a/Eigen/src/Jacobi/Jacobi.h b/Eigen/src/Jacobi/Jacobi.h index da9fb53d0..25eabe984 100644 --- a/Eigen/src/Jacobi/Jacobi.h +++ b/Eigen/src/Jacobi/Jacobi.h @@ -62,7 +62,7 @@ template class JacobiRotation JacobiRotation adjoint() const { using numext::conj; return JacobiRotation(conj(m_c), -m_s); } template - bool makeJacobi(const MatrixBase&, typename Derived::Index p, typename Derived::Index q); + bool makeJacobi(const MatrixBase&, Index p, Index q); bool makeJacobi(const RealScalar& x, const Scalar& y, const RealScalar& z); void makeGivens(const Scalar& p, const Scalar& q, Scalar* z=0); @@ -123,7 +123,7 @@ bool JacobiRotation::makeJacobi(const RealScalar& x, const Scalar& y, co */ template template -inline bool JacobiRotation::makeJacobi(const MatrixBase& m, typename Derived::Index p, typename Derived::Index q) +inline bool JacobiRotation::makeJacobi(const MatrixBase& m, Index p, Index q) { return makeJacobi(numext::real(m.coeff(p,p)), m.coeff(p,q), numext::real(m.coeff(q,q))); } @@ -300,7 +300,6 @@ namespace internal { template void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(VectorX& _x, VectorY& _y, const JacobiRotation& j) { - typedef typename VectorX::Index Index; typedef typename VectorX::Scalar Scalar; enum { PacketSize = packet_traits::size }; typedef typename packet_traits::type Packet; diff --git a/Eigen/src/LU/PartialPivLU.h b/Eigen/src/LU/PartialPivLU.h index 43c2a716e..e57b36bc5 100644 --- a/Eigen/src/LU/PartialPivLU.h +++ b/Eigen/src/LU/PartialPivLU.h @@ -262,7 +262,6 @@ struct partial_lu_impl typedef Block MatrixType; typedef Block BlockType; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; /** \internal performs the LU decomposition in-place of the matrix \a lu * using an unblocked algorithm. @@ -409,13 +408,13 @@ struct partial_lu_impl /** \internal performs the LU decomposition with partial pivoting in-place. */ template -void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, typename TranspositionType::StorageIndexType& nb_transpositions) +void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, typename TranspositionType::StorageIndex& nb_transpositions) { eigen_assert(lu.cols() == row_transpositions.size()); eigen_assert((&row_transpositions.coeffRef(1)-&row_transpositions.coeffRef(0)) == 1); partial_lu_impl - + ::blocked_lu(lu.rows(), lu.cols(), &lu.coeffRef(0,0), lu.outerStride(), &row_transpositions.coeffRef(0), nb_transpositions); } @@ -434,7 +433,7 @@ PartialPivLU& PartialPivLU::compute(const MatrixType& ma m_rowsTranspositions.resize(size); - typename TranspositionType::StorageIndexType nb_transpositions; + typename TranspositionType::StorageIndex nb_transpositions; internal::partial_lu_inplace(m_lu, m_rowsTranspositions, nb_transpositions); m_det_p = (nb_transpositions%2) ? -1 : 1; diff --git a/Eigen/src/QR/FullPivHouseholderQR.h b/Eigen/src/QR/FullPivHouseholderQR.h index a7a0d9138..7d5e58d2f 100644 --- a/Eigen/src/QR/FullPivHouseholderQR.h +++ b/Eigen/src/QR/FullPivHouseholderQR.h @@ -557,7 +557,6 @@ template struct FullPivHouseholderQRMatrixQReturnType : public ReturnByValue > { public: - typedef typename MatrixType::Index Index; typedef typename FullPivHouseholderQR::IntDiagSizeVectorType IntDiagSizeVectorType; typedef typename internal::plain_diag_type::type HCoeffsType; typedef Matrix void householder_qr_inplace_unblocked(MatrixQR& mat, HCoeffs& hCoeffs, typename MatrixQR::Scalar* tempData = 0) { - typedef typename MatrixQR::Index Index; typedef typename MatrixQR::Scalar Scalar; typedef typename MatrixQR::RealScalar RealScalar; Index rows = mat.rows(); @@ -264,11 +263,9 @@ template BlockType; diff --git a/Eigen/src/QR/HouseholderQR_MKL.h b/Eigen/src/QR/HouseholderQR_MKL.h index 8a3a7e406..84ab640a1 100644 --- a/Eigen/src/QR/HouseholderQR_MKL.h +++ b/Eigen/src/QR/HouseholderQR_MKL.h @@ -46,8 +46,7 @@ namespace internal { template \ struct householder_qr_inplace_blocked \ { \ - static void run(MatrixQR& mat, HCoeffs& hCoeffs, \ - typename MatrixQR::Index = 32, \ + static void run(MatrixQR& mat, HCoeffs& hCoeffs, Index = 32, \ typename MatrixQR::Scalar* = 0) \ { \ lapack_int m = (lapack_int) mat.rows(); \ diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h index dad59bcca..3fe17b27f 100644 --- a/Eigen/src/SVD/BDCSVD.h +++ b/Eigen/src/SVD/BDCSVD.h @@ -65,7 +65,6 @@ public: typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; enum { RowsAtCompileTime = MatrixType::RowsAtCompileTime, ColsAtCompileTime = MatrixType::ColsAtCompileTime, diff --git a/Eigen/src/SVD/JacobiSVD.h b/Eigen/src/SVD/JacobiSVD.h index 444187ae7..fcf01f518 100644 --- a/Eigen/src/SVD/JacobiSVD.h +++ b/Eigen/src/SVD/JacobiSVD.h @@ -52,7 +52,6 @@ template class qr_preconditioner_impl { public: - typedef typename MatrixType::Index Index; void allocate(const JacobiSVD&) {} bool run(JacobiSVD&, const MatrixType&) { @@ -66,7 +65,6 @@ template class qr_preconditioner_impl { public: - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; enum { @@ -107,7 +105,6 @@ template class qr_preconditioner_impl { public: - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; enum { @@ -157,8 +154,6 @@ template class qr_preconditioner_impl { public: - typedef typename MatrixType::Index Index; - void allocate(const JacobiSVD& svd) { if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) @@ -198,7 +193,6 @@ template class qr_preconditioner_impl { public: - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; enum { @@ -257,8 +251,6 @@ template class qr_preconditioner_impl { public: - typedef typename MatrixType::Index Index; - void allocate(const JacobiSVD& svd) { if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) @@ -297,7 +289,6 @@ template class qr_preconditioner_impl { public: - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; enum { @@ -359,7 +350,6 @@ template struct svd_precondition_2x2_block_to_be_real { typedef JacobiSVD SVD; - typedef typename SVD::Index Index; static void run(typename SVD::WorkMatrixType&, SVD&, Index, Index) {} }; @@ -369,7 +359,6 @@ struct svd_precondition_2x2_block_to_be_real typedef JacobiSVD SVD; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename SVD::Index Index; static void run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q) { using std::sqrt; @@ -514,7 +503,6 @@ template class JacobiSVD typedef _MatrixType MatrixType; typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef typename MatrixType::Index Index; enum { RowsAtCompileTime = MatrixType::RowsAtCompileTime, ColsAtCompileTime = MatrixType::ColsAtCompileTime, diff --git a/Eigen/src/SVD/SVDBase.h b/Eigen/src/SVD/SVDBase.h index 95d378da9..8903755e7 100644 --- a/Eigen/src/SVD/SVDBase.h +++ b/Eigen/src/SVD/SVDBase.h @@ -53,6 +53,7 @@ public: typedef typename MatrixType::Scalar Scalar; typedef typename NumTraits::Real RealScalar; typedef typename MatrixType::StorageIndex StorageIndex; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 enum { RowsAtCompileTime = MatrixType::RowsAtCompileTime, ColsAtCompileTime = MatrixType::ColsAtCompileTime, diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index eaa6bb86e..9dc470fd9 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -29,7 +29,7 @@ template class UpperBidiagonalization }; typedef typename MatrixType::Scalar Scalar; typedef typename MatrixType::RealScalar RealScalar; - typedef typename MatrixType::Index Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef Matrix RowVectorType; typedef Matrix ColVectorType; typedef BandMatrix BidiagonalType; @@ -95,7 +95,6 @@ void upperbidiagonalization_inplace_unblocked(MatrixType& mat, typename MatrixType::RealScalar *upper_diagonal, typename MatrixType::Scalar* tempData = 0) { - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; Index rows = mat.rows(); @@ -153,13 +152,12 @@ template void upperbidiagonalization_blocked_helper(MatrixType& A, typename MatrixType::RealScalar *diagonal, typename MatrixType::RealScalar *upper_diagonal, - typename MatrixType::Index bs, + Index bs, Ref::Flags & RowMajorBit> > X, Ref::Flags & RowMajorBit> > Y) { - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; enum { StorageOrder = traits::Flags & RowMajorBit }; typedef InnerStride ColInnerStride; @@ -282,10 +280,9 @@ void upperbidiagonalization_blocked_helper(MatrixType& A, */ template void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagonal, - typename MatrixType::Index maxBlockSize=32, + Index maxBlockSize=32, typename MatrixType::Scalar* /*tempData*/ = 0) { - typedef typename MatrixType::Index Index; typedef typename MatrixType::Scalar Scalar; typedef Block BlockType; diff --git a/Eigen/src/SparseCore/SparseBlock.h b/Eigen/src/SparseCore/SparseBlock.h index b8604a219..40dc1a2bd 100644 --- a/Eigen/src/SparseCore/SparseBlock.h +++ b/Eigen/src/SparseCore/SparseBlock.h @@ -158,7 +158,7 @@ public: matrix.outerIndexPtr()[m_outerStart+k] = p; p += tmp.innerVector(k).nonZeros(); } - std::ptrdiff_t offset = nnz - block_size; + StorageIndex offset = internal::convert_index(nnz - block_size); for(Index k = m_outerStart + m_outerSize.value(); k<=matrix.outerSize(); ++k) { matrix.outerIndexPtr()[k] += offset; diff --git a/Eigen/src/SparseCore/SparseView.h b/Eigen/src/SparseCore/SparseView.h index d6042d970..1c69aa458 100644 --- a/Eigen/src/SparseCore/SparseView.h +++ b/Eigen/src/SparseCore/SparseView.h @@ -78,7 +78,7 @@ struct unary_evaluator, IteratorBased> typedef typename XprType::Scalar Scalar; public: - EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& sve, typename XprType::Index outer) + EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& sve, Index outer) : EvalIterator(sve.m_argImpl,outer), m_view(sve.m_view) { incrementToNonZero(); @@ -126,7 +126,6 @@ struct unary_evaluator, IndexBased> typedef SparseView XprType; protected: enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit }; - typedef typename XprType::StorageIndex StorageIndex; typedef typename XprType::Scalar Scalar; public: diff --git a/Eigen/src/SuperLUSupport/SuperLUSupport.h b/Eigen/src/SuperLUSupport/SuperLUSupport.h index 8779eb74c..d182b59b3 100644 --- a/Eigen/src/SuperLUSupport/SuperLUSupport.h +++ b/Eigen/src/SuperLUSupport/SuperLUSupport.h @@ -808,7 +808,6 @@ class SuperILU : public SuperLUBase<_MatrixType,SuperILU<_MatrixType> > typedef _MatrixType MatrixType; typedef typename Base::Scalar Scalar; typedef typename Base::RealScalar RealScalar; - typedef typename Base::Index Index; public: using Base::_solve_impl; diff --git a/Eigen/src/misc/Image.h b/Eigen/src/misc/Image.h index 75c5f433a..b8b8a0455 100644 --- a/Eigen/src/misc/Image.h +++ b/Eigen/src/misc/Image.h @@ -38,7 +38,6 @@ template struct image_retval_base typedef _DecompositionType DecompositionType; typedef typename DecompositionType::MatrixType MatrixType; typedef ReturnByValue Base; - typedef typename Base::Index Index; image_retval_base(const DecompositionType& dec, const MatrixType& originalMatrix) : m_dec(dec), m_rank(dec.rank()), @@ -69,7 +68,6 @@ template struct image_retval_base typedef typename DecompositionType::MatrixType MatrixType; \ typedef typename MatrixType::Scalar Scalar; \ typedef typename MatrixType::RealScalar RealScalar; \ - typedef typename MatrixType::Index Index; \ typedef Eigen::internal::image_retval_base Base; \ using Base::dec; \ using Base::originalMatrix; \ diff --git a/Eigen/src/misc/Kernel.h b/Eigen/src/misc/Kernel.h index 4b03e44c1..bef5d6ff5 100644 --- a/Eigen/src/misc/Kernel.h +++ b/Eigen/src/misc/Kernel.h @@ -39,7 +39,6 @@ template struct kernel_retval_base { typedef _DecompositionType DecompositionType; typedef ReturnByValue Base; - typedef typename Base::Index Index; explicit kernel_retval_base(const DecompositionType& dec) : m_dec(dec), @@ -68,7 +67,6 @@ template struct kernel_retval_base typedef typename DecompositionType::MatrixType MatrixType; \ typedef typename MatrixType::Scalar Scalar; \ typedef typename MatrixType::RealScalar RealScalar; \ - typedef typename MatrixType::Index Index; \ typedef Eigen::internal::kernel_retval_base Base; \ using Base::dec; \ using Base::rank; \ From 45cbb0bbb1f66bbc923dd4dd1034b919f6b4a191 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 15:05:41 +0100 Subject: [PATCH 1094/1103] The usage of DenseIndex is deprecated, so let's replace DenseIndex by Index --- Eigen/src/Core/BandMatrix.h | 2 +- Eigen/src/Core/DenseStorage.h | 96 +++++++++---------- Eigen/src/Core/GenericPacketMath.h | 4 +- Eigen/src/Core/Matrix.h | 2 +- Eigen/src/Core/Stride.h | 4 +- Eigen/src/Core/arch/AVX/Complex.h | 8 +- Eigen/src/Core/arch/AVX/PacketMath.h | 8 +- Eigen/src/Core/arch/AltiVec/Complex.h | 8 +- Eigen/src/Core/arch/AltiVec/PacketMath.h | 12 +-- Eigen/src/Core/arch/NEON/Complex.h | 8 +- Eigen/src/Core/arch/NEON/PacketMath.h | 12 +-- Eigen/src/Core/arch/SSE/Complex.h | 4 +- Eigen/src/Core/arch/SSE/PacketMath.h | 12 +-- Eigen/src/Core/functors/NullaryFunctors.h | 2 +- Eigen/src/Core/products/GeneralMatrixMatrix.h | 24 ++--- Eigen/src/Geometry/AlignedBox.h | 2 +- Eigen/src/Geometry/Hyperplane.h | 2 +- Eigen/src/Geometry/ParametrizedLine.h | 2 +- Eigen/src/Geometry/Quaternion.h | 7 +- Eigen/src/PardisoSupport/PardisoSupport.h | 68 ++++++------- Eigen/src/SVD/BDCSVD.h | 2 +- 21 files changed, 143 insertions(+), 146 deletions(-) diff --git a/Eigen/src/Core/BandMatrix.h b/Eigen/src/Core/BandMatrix.h index d07ea7056..87c124fdf 100644 --- a/Eigen/src/Core/BandMatrix.h +++ b/Eigen/src/Core/BandMatrix.h @@ -179,7 +179,7 @@ struct traits > { typedef _Scalar Scalar; typedef Dense StorageKind; - typedef DenseIndex StorageIndex; + typedef Eigen::Index StorageIndex; enum { CoeffReadCost = NumTraits::ReadCost, RowsAtCompileTime = _Rows, diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 852648639..4c37fadbd 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -140,12 +140,12 @@ template class DenseSt if (this != &other) m_data = other.m_data; return *this; } - EIGEN_DEVICE_FUNC DenseStorage(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC DenseStorage(Index,Index,Index) {} EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); } - EIGEN_DEVICE_FUNC static DenseIndex rows(void) {return _Rows;} - EIGEN_DEVICE_FUNC static DenseIndex cols(void) {return _Cols;} - EIGEN_DEVICE_FUNC void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {} - EIGEN_DEVICE_FUNC void resize(DenseIndex,DenseIndex,DenseIndex) {} + EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;} + EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {} + EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {} EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; } EIGEN_DEVICE_FUNC T *data() { return m_data.array; } }; @@ -158,12 +158,12 @@ template class DenseStorage class DenseStorage class DenseStorage { internal::plain_array m_data; - DenseIndex m_rows; - DenseIndex m_cols; + Index m_rows; + Index m_cols; public: EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0), m_cols(0) {} explicit DenseStorage(internal::constructor_without_unaligned_array_assert) @@ -199,13 +199,13 @@ template class DenseStorage class DenseStorage class DenseStorage { internal::plain_array m_data; - DenseIndex m_rows; + Index m_rows; public: EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0) {} explicit DenseStorage(internal::constructor_without_unaligned_array_assert) @@ -229,12 +229,12 @@ template class DenseStorage class DenseStorage class DenseStorage { internal::plain_array m_data; - DenseIndex m_cols; + Index m_cols; public: EIGEN_DEVICE_FUNC DenseStorage() : m_cols(0) {} explicit DenseStorage(internal::constructor_without_unaligned_array_assert) @@ -258,12 +258,12 @@ template class DenseStorage class DenseStorage class DenseStorage { T *m_data; - DenseIndex m_rows; - DenseIndex m_cols; + Index m_rows; + Index m_cols; public: EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0), m_cols(0) {} explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0), m_cols(0) {} - DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols) + DenseStorage(Index size, Index nbRows, Index nbCols) : m_data(internal::conditional_aligned_new_auto(size)), m_rows(nbRows), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } DenseStorage(const DenseStorage& other) @@ -317,15 +317,15 @@ template class DenseStorage(m_data, m_rows*m_cols); } void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); } - EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return m_rows;} - EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return m_cols;} - void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols) + EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;} + EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;} + void conservativeResize(Index size, Index nbRows, Index nbCols) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, m_rows*m_cols); m_rows = nbRows; m_cols = nbCols; } - void resize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols) + void resize(Index size, Index nbRows, Index nbCols) { if(size != m_rows*m_cols) { @@ -347,11 +347,11 @@ template class DenseStorage class DenseStorage { T *m_data; - DenseIndex m_cols; + Index m_cols; public: EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_cols(0) {} explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {} - DenseStorage(DenseIndex size, DenseIndex, DenseIndex nbCols) : m_data(internal::conditional_aligned_new_auto(size)), m_cols(nbCols) + DenseStorage(Index size, Index, Index nbCols) : m_data(internal::conditional_aligned_new_auto(size)), m_cols(nbCols) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } DenseStorage(const DenseStorage& other) : m_data(internal::conditional_aligned_new_auto(_Rows*other.m_cols)) @@ -385,14 +385,14 @@ template class DenseStorage(m_data, _Rows*m_cols); } void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); } - EIGEN_DEVICE_FUNC static DenseIndex rows(void) {return _Rows;} - EIGEN_DEVICE_FUNC DenseIndex cols(void) const {return m_cols;} - void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols) + EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;} + EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;} + void conservativeResize(Index size, Index, Index nbCols) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, _Rows*m_cols); m_cols = nbCols; } - EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex, DenseIndex nbCols) + EIGEN_STRONG_INLINE void resize(Index size, Index, Index nbCols) { if(size != _Rows*m_cols) { @@ -413,11 +413,11 @@ template class DenseStorage class DenseStorage { T *m_data; - DenseIndex m_rows; + Index m_rows; public: EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0) {} explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {} - DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex) : m_data(internal::conditional_aligned_new_auto(size)), m_rows(nbRows) + DenseStorage(Index size, Index nbRows, Index) : m_data(internal::conditional_aligned_new_auto(size)), m_rows(nbRows) { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN } DenseStorage(const DenseStorage& other) : m_data(internal::conditional_aligned_new_auto(other.m_rows*_Cols)) @@ -451,14 +451,14 @@ template class DenseStorage(m_data, _Cols*m_rows); } void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); } - EIGEN_DEVICE_FUNC DenseIndex rows(void) const {return m_rows;} - EIGEN_DEVICE_FUNC static DenseIndex cols(void) {return _Cols;} - void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex) + EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;} + EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;} + void conservativeResize(Index size, Index nbRows, Index) { m_data = internal::conditional_aligned_realloc_new_auto(m_data, size, m_rows*_Cols); m_rows = nbRows; } - EIGEN_STRONG_INLINE void resize(DenseIndex size, DenseIndex nbRows, DenseIndex) + EIGEN_STRONG_INLINE void resize(Index size, Index nbRows, Index) { if(size != m_rows*_Cols) { diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 8759cd06c..74e1174ae 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -236,10 +236,10 @@ template EIGEN_DEVICE_FUNC inline void pstore( template EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from) { (*to) = from; } - template EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, DenseIndex /*stride*/) + template EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, Index /*stride*/) { return ploadu(from); } - template EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, DenseIndex /*stride*/) + template EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, Index /*stride*/) { pstore(to, from); } /** \internal tries to do cache prefetching of \a addr */ diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h index 94b1a966e..a10d1856f 100644 --- a/Eigen/src/Core/Matrix.h +++ b/Eigen/src/Core/Matrix.h @@ -107,7 +107,7 @@ struct traits > { typedef _Scalar Scalar; typedef Dense StorageKind; - typedef DenseIndex StorageIndex; + typedef Eigen::Index StorageIndex; typedef MatrixXpr XprKind; enum { RowsAtCompileTime = _Rows, diff --git a/Eigen/src/Core/Stride.h b/Eigen/src/Core/Stride.h index e46faad34..9a2f4f1eb 100644 --- a/Eigen/src/Core/Stride.h +++ b/Eigen/src/Core/Stride.h @@ -44,7 +44,7 @@ template class Stride { public: - typedef DenseIndex Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 enum { InnerStrideAtCompileTime = _InnerStrideAtCompileTime, OuterStrideAtCompileTime = _OuterStrideAtCompileTime @@ -91,7 +91,6 @@ class InnerStride : public Stride<0, Value> { typedef Stride<0, Value> Base; public: - typedef DenseIndex Index; EIGEN_DEVICE_FUNC InnerStride() : Base() {} EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} // FIXME making this explicit could break valid code }; @@ -103,7 +102,6 @@ class OuterStride : public Stride { typedef Stride Base; public: - typedef DenseIndex Index; EIGEN_DEVICE_FUNC OuterStride() : Base() {} EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} // FIXME making this explicit could break valid code }; diff --git a/Eigen/src/Core/arch/AVX/Complex.h b/Eigen/src/Core/arch/AVX/Complex.h index aa5aa1e34..003a1fc3c 100644 --- a/Eigen/src/Core/arch/AVX/Complex.h +++ b/Eigen/src/Core/arch/AVX/Complex.h @@ -92,7 +92,7 @@ template<> EIGEN_STRONG_INLINE Packet4cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); } -template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packet4cf>(const std::complex* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packet4cf>(const std::complex* from, Index stride) { return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]), std::imag(from[2*stride]), std::real(from[2*stride]), @@ -100,7 +100,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather, Packe std::imag(from[0*stride]), std::real(from[0*stride]))); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet4cf>(std::complex* to, const Packet4cf& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet4cf>(std::complex* to, const Packet4cf& from, Index stride) { __m128 low = _mm256_extractf128_ps(from.v, 0); to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)), @@ -310,13 +310,13 @@ template<> EIGEN_STRONG_INLINE Packet2cd ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } -template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather, Packet2cd>(const std::complex* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather, Packet2cd>(const std::complex* from, Index stride) { return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]), std::imag(from[0*stride]), std::real(from[0*stride]))); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cd>(std::complex* to, const Packet2cd& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cd>(std::complex* to, const Packet2cd& from, Index stride) { __m128d low = _mm256_extractf128_pd(from.v, 0); to[stride*0] = std::complex(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1))); diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index be66a502a..ff6cc6b56 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -226,17 +226,17 @@ template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet8i& // NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available // NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4); -template<> EIGEN_DEVICE_FUNC inline Packet8f pgather(const float* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet8f pgather(const float* from, Index stride) { return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride], from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline Packet4d pgather(const double* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4d pgather(const double* from, Index stride) { return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet8f& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet8f& from, Index stride) { __m128 low = _mm256_extractf128_ps(from, 0); to[stride*0] = _mm_cvtss_f32(low); @@ -250,7 +250,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, co to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)); to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet4d& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet4d& from, Index stride) { __m128d low = _mm256_extractf128_pd(from, 0); to[stride*0] = _mm_cvtsd_f64(low); diff --git a/Eigen/src/Core/arch/AltiVec/Complex.h b/Eigen/src/Core/arch/AltiVec/Complex.h index f9b93a42b..565d2ece0 100644 --- a/Eigen/src/Core/arch/AltiVec/Complex.h +++ b/Eigen/src/Core/arch/AltiVec/Complex.h @@ -67,14 +67,14 @@ template<> EIGEN_STRONG_INLINE Packet2cf pset1(const std::complex EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, Index stride) { std::complex EIGEN_ALIGN16 af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return Packet2cf(vec_ld(0, (const float*)af)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, Index stride) { std::complex EIGEN_ALIGN16 af[2]; vec_st(from.v, 0, (float*)af); @@ -285,14 +285,14 @@ template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex< template<> EIGEN_STRONG_INLINE Packet1cd pset1(const std::complex& from) { /* here we really have to use unaligned loads :( */ return ploadu(&from); } -template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, Index stride) { std::complex EIGEN_ALIGN16 af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return pload(af); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet1cd>(std::complex* to, const Packet1cd& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet1cd>(std::complex* to, const Packet1cd& from, Index stride) { std::complex EIGEN_ALIGN16 af[2]; pstore >(af, from); diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h index 6b68fc7a5..d647427ce 100755 --- a/Eigen/src/Core/arch/AltiVec/PacketMath.h +++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h @@ -252,7 +252,7 @@ pbroadcast4(const int *a, a3 = vec_splat(a3, 3); } -template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, Index stride) { float EIGEN_ALIGN16 af[4]; af[0] = from[0*stride]; @@ -261,7 +261,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const floa af[3] = from[3*stride]; return pload(af); } -template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, Index stride) { int EIGEN_ALIGN16 ai[4]; ai[0] = from[0*stride]; @@ -270,7 +270,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* f ai[3] = from[3*stride]; return pload(ai); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, Index stride) { float EIGEN_ALIGN16 af[4]; pstore(af, from); @@ -279,7 +279,7 @@ template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, co to[2*stride] = af[2]; to[3*stride] = af[3]; } -template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, Index stride) { int EIGEN_ALIGN16 ai[4]; pstore((int *)ai, from); @@ -793,14 +793,14 @@ pbroadcast4(const double *a, a2 = vec_splat_dbl(a3, 0); a3 = vec_splat_dbl(a3, 1); } -template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, Index stride) { double EIGEN_ALIGN16 af[2]; af[0] = from[0*stride]; af[1] = from[1*stride]; return pload(af); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, Index stride) { double EIGEN_ALIGN16 af[2]; pstore(af, from); diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h index 0fdcb0741..154daa7a7 100644 --- a/Eigen/src/Core/arch/NEON/Complex.h +++ b/Eigen/src/Core/arch/NEON/Complex.h @@ -112,7 +112,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup(const std::complex< template<> EIGEN_STRONG_INLINE void pstore >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } -template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, Index stride) { Packet4f res; res = vsetq_lane_f32(std::real(from[0*stride]), res, 0); @@ -122,7 +122,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packe return Packet2cf(res); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, Index stride) { to[stride*0] = std::complex(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1)); to[stride*1] = std::complex(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3)); @@ -363,7 +363,7 @@ template<> EIGEN_STRONG_INLINE void pstoreu >(std::complex< template<> EIGEN_STRONG_INLINE void prefetch >(const std::complex * addr) { EIGEN_ARM_PREFETCH((double *)addr); } -template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Packet1cd>(const std::complex* from, Index stride) { Packet2d res; res = vsetq_lane_f64(std::real(from[0*stride]), res, 0); @@ -371,7 +371,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather, Pack return Packet1cd(res); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet1cd>(std::complex* to, const Packet1cd& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet1cd>(std::complex* to, const Packet1cd& from, Index stride) { to[stride*0] = std::complex(vgetq_lane_f64(from.v, 0), vgetq_lane_f64(from.v, 1)); } diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 559682cf7..8149aed7f 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -250,7 +250,7 @@ template<> EIGEN_STRONG_INLINE void pstore(int* to, const Packet4i& f template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); } -template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, Index stride) { Packet4f res; res = vsetq_lane_f32(from[0*stride], res, 0); @@ -259,7 +259,7 @@ template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const floa res = vsetq_lane_f32(from[3*stride], res, 3); return res; } -template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, Index stride) { Packet4i res; res = vsetq_lane_s32(from[0*stride], res, 0); @@ -269,14 +269,14 @@ template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* f return res; } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, Index stride) { to[stride*0] = vgetq_lane_f32(from, 0); to[stride*1] = vgetq_lane_f32(from, 1); to[stride*2] = vgetq_lane_f32(from, 2); to[stride*3] = vgetq_lane_f32(from, 3); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, Index stride) { to[stride*0] = vgetq_lane_s32(from, 0); to[stride*1] = vgetq_lane_s32(from, 1); @@ -606,14 +606,14 @@ template<> EIGEN_STRONG_INLINE void pstore(double* to, const Packet2d& template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to, from); } -template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, Index stride) { Packet2d res; res = vsetq_lane_f64(from[0*stride], res, 0); res = vsetq_lane_f64(from[1*stride], res, 1); return res; } -template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, Index stride) { to[stride*0] = vgetq_lane_f64(from, 0); to[stride*1] = vgetq_lane_f64(from, 1); diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h index 565e448fe..acb49abf8 100644 --- a/Eigen/src/Core/arch/SSE/Complex.h +++ b/Eigen/src/Core/arch/SSE/Complex.h @@ -115,13 +115,13 @@ template<> EIGEN_STRONG_INLINE void pstore >(std::complex EIGEN_STRONG_INLINE void pstoreu >(std::complex * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); } -template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather, Packet2cf>(const std::complex* from, Index stride) { return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]), std::imag(from[0*stride]), std::real(from[0*stride]))); } -template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter, Packet2cf>(std::complex* to, const Packet2cf& from, Index stride) { to[stride*0] = std::complex(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)), _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1))); diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 898cb9ab0..86d94dffa 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -387,32 +387,32 @@ template<> EIGEN_STRONG_INLINE void pstoreu(double* to, const Packet2d& template<> EIGEN_STRONG_INLINE void pstoreu(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), Packet2d(_mm_castps_pd(from))); } template<> EIGEN_STRONG_INLINE void pstoreu(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast(to), Packet2d(_mm_castsi128_pd(from))); } -template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4f pgather(const float* from, Index stride) { return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet2d pgather(const double* from, Index stride) { return _mm_set_pd(from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline Packet4i pgather(const int* from, Index stride) { return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(float* to, const Packet4f& from, Index stride) { to[stride*0] = _mm_cvtss_f32(from); to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1)); to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2)); to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(double* to, const Packet2d& from, Index stride) { to[stride*0] = _mm_cvtsd_f64(from); to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1)); } -template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, DenseIndex stride) +template<> EIGEN_DEVICE_FUNC inline void pscatter(int* to, const Packet4i& from, Index stride) { to[stride*0] = _mm_cvtsi128_si32(from); to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1)); diff --git a/Eigen/src/Core/functors/NullaryFunctors.h b/Eigen/src/Core/functors/NullaryFunctors.h index be03fbf52..2362b3a7f 100644 --- a/Eigen/src/Core/functors/NullaryFunctors.h +++ b/Eigen/src/Core/functors/NullaryFunctors.h @@ -112,7 +112,7 @@ template struct functor_traits< linspaced_o template struct linspaced_op { typedef typename packet_traits::type Packet; - linspaced_op(const Scalar& low, const Scalar& high, DenseIndex num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1))) {} + linspaced_op(const Scalar& low, const Scalar& high, Index num_steps) : impl((num_steps==1 ? high : low), (num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1))) {} template EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (Index i) const { return impl(i); } diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h index 8210ea584..44e44b986 100644 --- a/Eigen/src/Core/products/GeneralMatrixMatrix.h +++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h @@ -257,9 +257,9 @@ class level3_blocking LhsScalar* m_blockA; RhsScalar* m_blockB; - DenseIndex m_mc; - DenseIndex m_nc; - DenseIndex m_kc; + Index m_mc; + Index m_nc; + Index m_kc; public: @@ -267,9 +267,9 @@ class level3_blocking : m_blockA(0), m_blockB(0), m_mc(0), m_nc(0), m_kc(0) {} - inline DenseIndex mc() const { return m_mc; } - inline DenseIndex nc() const { return m_nc; } - inline DenseIndex kc() const { return m_kc; } + inline Index mc() const { return m_mc; } + inline Index nc() const { return m_nc; } + inline Index kc() const { return m_kc; } inline LhsScalar* blockA() { return m_blockA; } inline RhsScalar* blockB() { return m_blockB; } @@ -299,7 +299,7 @@ class gemm_blocking_spacem_mc = ActualRows; this->m_nc = ActualCols; @@ -326,12 +326,12 @@ class gemm_blocking_space::type RhsScalar; typedef gebp_traits Traits; - DenseIndex m_sizeA; - DenseIndex m_sizeB; + Index m_sizeA; + Index m_sizeB; public: - gemm_blocking_space(DenseIndex rows, DenseIndex cols, DenseIndex depth, int num_threads, bool l3_blocking) + gemm_blocking_space(Index rows, Index cols, Index depth, int num_threads, bool l3_blocking) { this->m_mc = Transpose ? cols : rows; this->m_nc = Transpose ? rows : cols; @@ -343,8 +343,8 @@ class gemm_blocking_spacem_mc; - DenseIndex n = this->m_nc; + Index m = this->m_mc; + Index n = this->m_nc; computeProductBlockingSizes(this->m_kc, m, n, num_threads); } diff --git a/Eigen/src/Geometry/AlignedBox.h b/Eigen/src/Geometry/AlignedBox.h index d6c5c1293..b7c02e8db 100644 --- a/Eigen/src/Geometry/AlignedBox.h +++ b/Eigen/src/Geometry/AlignedBox.h @@ -32,7 +32,7 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim) enum { AmbientDimAtCompileTime = _AmbientDim }; typedef _Scalar Scalar; typedef NumTraits ScalarTraits; - typedef DenseIndex Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef typename ScalarTraits::Real RealScalar; typedef typename ScalarTraits::NonInteger NonInteger; typedef Matrix VectorType; diff --git a/Eigen/src/Geometry/Hyperplane.h b/Eigen/src/Geometry/Hyperplane.h index 00b7c4300..2d076d7f8 100644 --- a/Eigen/src/Geometry/Hyperplane.h +++ b/Eigen/src/Geometry/Hyperplane.h @@ -41,7 +41,7 @@ public: }; typedef _Scalar Scalar; typedef typename NumTraits::Real RealScalar; - typedef DenseIndex Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef Matrix VectorType; typedef Matrix::Real RealScalar; - typedef DenseIndex Index; + typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3 typedef Matrix VectorType; /** Default constructor without initialization */ diff --git a/Eigen/src/Geometry/Quaternion.h b/Eigen/src/Geometry/Quaternion.h index 508eba767..e1ad803bb 100644 --- a/Eigen/src/Geometry/Quaternion.h +++ b/Eigen/src/Geometry/Quaternion.h @@ -724,7 +724,6 @@ template struct quaternionbase_assign_impl { typedef typename Other::Scalar Scalar; - typedef DenseIndex Index; template static inline void run(QuaternionBase& q, const Other& mat) { using std::sqrt; @@ -742,13 +741,13 @@ struct quaternionbase_assign_impl } else { - DenseIndex i = 0; + Index i = 0; if (mat.coeff(1,1) > mat.coeff(0,0)) i = 1; if (mat.coeff(2,2) > mat.coeff(i,i)) i = 2; - DenseIndex j = (i+1)%3; - DenseIndex k = (j+1)%3; + Index j = (i+1)%3; + Index k = (j+1)%3; t = sqrt(mat.coeff(i,i)-mat.coeff(j,j)-mat.coeff(k,k) + Scalar(1.0)); q.coeffs().coeffRef(i) = Scalar(0.5) * t; diff --git a/Eigen/src/PardisoSupport/PardisoSupport.h b/Eigen/src/PardisoSupport/PardisoSupport.h index 7c75dcb7f..7ab2e3e6b 100644 --- a/Eigen/src/PardisoSupport/PardisoSupport.h +++ b/Eigen/src/PardisoSupport/PardisoSupport.h @@ -40,13 +40,13 @@ template class PardisoLDLT; namespace internal { - template + template struct pardiso_run_selector { - static Index run( _MKL_DSS_HANDLE_t pt, Index maxfct, Index mnum, Index type, Index phase, Index n, void *a, - Index *ia, Index *ja, Index *perm, Index nrhs, Index *iparm, Index msglvl, void *b, void *x) + static IndexType run( _MKL_DSS_HANDLE_t pt, IndexType maxfct, IndexType mnum, IndexType type, IndexType phase, IndexType n, void *a, + IndexType *ia, IndexType *ja, IndexType *perm, IndexType nrhs, IndexType *iparm, IndexType msglvl, void *b, void *x) { - Index error = 0; + IndexType error = 0; ::pardiso(pt, &maxfct, &mnum, &type, &phase, &n, a, ia, ja, perm, &nrhs, iparm, &msglvl, b, x, &error); return error; } @@ -54,11 +54,11 @@ namespace internal template<> struct pardiso_run_selector { - typedef long long int Index; - static Index run( _MKL_DSS_HANDLE_t pt, Index maxfct, Index mnum, Index type, Index phase, Index n, void *a, - Index *ia, Index *ja, Index *perm, Index nrhs, Index *iparm, Index msglvl, void *b, void *x) + typedef long long int IndexTypeType; + static IndexType run( _MKL_DSS_HANDLE_t pt, IndexType maxfct, IndexType mnum, IndexType type, IndexType phase, IndexType n, void *a, + IndexType *ia, IndexType *ja, IndexType *perm, IndexType nrhs, IndexType *iparm, IndexType msglvl, void *b, void *x) { - Index error = 0; + IndexType error = 0; ::pardiso_64(pt, &maxfct, &mnum, &type, &phase, &n, a, ia, ja, perm, &nrhs, iparm, &msglvl, b, x, &error); return error; } @@ -72,7 +72,7 @@ namespace internal typedef _MatrixType MatrixType; typedef typename _MatrixType::Scalar Scalar; typedef typename _MatrixType::RealScalar RealScalar; - typedef typename _MatrixType::Index Index; + typedef typename _MatrixType::StorageIndex StorageIndex; }; template @@ -81,7 +81,7 @@ namespace internal typedef _MatrixType MatrixType; typedef typename _MatrixType::Scalar Scalar; typedef typename _MatrixType::RealScalar RealScalar; - typedef typename _MatrixType::Index Index; + typedef typename _MatrixType::StorageIndex StorageIndex; }; template @@ -90,7 +90,7 @@ namespace internal typedef _MatrixType MatrixType; typedef typename _MatrixType::Scalar Scalar; typedef typename _MatrixType::RealScalar RealScalar; - typedef typename _MatrixType::Index Index; + typedef typename _MatrixType::StorageIndex StorageIndex; }; } @@ -111,18 +111,18 @@ class PardisoImpl : public SparseSolveBase typedef typename Traits::Scalar Scalar; typedef typename Traits::RealScalar RealScalar; typedef typename Traits::StorageIndex StorageIndex; - typedef SparseMatrix SparseMatrixType; + typedef SparseMatrix SparseMatrixType; typedef Matrix VectorType; - typedef Matrix IntRowVectorType; - typedef Matrix IntColVectorType; - typedef Array ParameterType; + typedef Matrix IntRowVectorType; + typedef Matrix IntColVectorType; + typedef Array ParameterType; enum { ScalarIsComplex = NumTraits::IsComplex }; PardisoImpl() { - eigen_assert((sizeof(Index) >= sizeof(_INTEGER_t) && sizeof(Index) <= 8) && "Non-supported index type"); + eigen_assert((sizeof(StorageIndex) >= sizeof(_INTEGER_t) && sizeof(StorageIndex) <= 8) && "Non-supported index type"); m_iparm.setZero(); m_msglvl = 0; // No output m_isInitialized = false; @@ -181,7 +181,7 @@ class PardisoImpl : public SparseSolveBase { if(m_isInitialized) // Factorization ran at least once { - internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, -1, m_size, 0, 0, 0, m_perm.data(), 0, + internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, -1, m_size, 0, 0, 0, m_perm.data(), 0, m_iparm.data(), m_msglvl, 0, 0); } } @@ -261,9 +261,9 @@ Derived& PardisoImpl::compute(const MatrixType& a) derived().getMatrix(a); Index error; - error = internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, 12, m_size, - m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), - m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL); + error = internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, 12, m_size, + m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), + m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL); manageErrorCode(error); m_analysisIsOk = true; @@ -284,9 +284,9 @@ Derived& PardisoImpl::analyzePattern(const MatrixType& a) derived().getMatrix(a); Index error; - error = internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, 11, m_size, - m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), - m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL); + error = internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, 11, m_size, + m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), + m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL); manageErrorCode(error); m_analysisIsOk = true; @@ -304,9 +304,9 @@ Derived& PardisoImpl::factorize(const MatrixType& a) derived().getMatrix(a); Index error; - error = internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, 22, m_size, - m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), - m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL); + error = internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, 22, m_size, + m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), + m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL); manageErrorCode(error); m_factorizationIsOk = true; @@ -348,10 +348,10 @@ bool PardisoImpl::_solve_impl(const MatrixBase &b, MatrixBase::run(m_pt, 1, 1, m_type, 33, m_size, - m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), - m_perm.data(), nrhs, m_iparm.data(), m_msglvl, - rhs_ptr, x.derived().data()); + error = internal::pardiso_run_selector::run(m_pt, 1, 1, m_type, 33, m_size, + m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(), + m_perm.data(), nrhs, m_iparm.data(), m_msglvl, + rhs_ptr, x.derived().data()); return error==0; } @@ -424,7 +424,7 @@ class PardisoLLT : public PardisoImpl< PardisoLLT > protected: typedef PardisoImpl< PardisoLLT > Base; typedef typename Base::Scalar Scalar; - typedef typename Base::Index Index; + typedef typename Base::StorageIndex StorageIndex; typedef typename Base::RealScalar RealScalar; using Base::pardisoInit; using Base::m_matrix; @@ -454,7 +454,7 @@ class PardisoLLT : public PardisoImpl< PardisoLLT > void getMatrix(const MatrixType& matrix) { // PARDISO supports only upper, row-major matrices - PermutationMatrix p_null; + PermutationMatrix p_null; m_matrix.resize(matrix.rows(), matrix.cols()); m_matrix.template selfadjointView() = matrix.template selfadjointView().twistedBy(p_null); } @@ -482,7 +482,7 @@ class PardisoLDLT : public PardisoImpl< PardisoLDLT > protected: typedef PardisoImpl< PardisoLDLT > Base; typedef typename Base::Scalar Scalar; - typedef typename Base::Index Index; + typedef typename Base::StorageIndex StorageIndex; typedef typename Base::RealScalar RealScalar; using Base::pardisoInit; using Base::m_matrix; @@ -510,7 +510,7 @@ class PardisoLDLT : public PardisoImpl< PardisoLDLT > void getMatrix(const MatrixType& matrix) { // PARDISO supports only upper, row-major matrices - PermutationMatrix p_null; + PermutationMatrix p_null; m_matrix.resize(matrix.rows(), matrix.cols()); m_matrix.template selfadjointView() = matrix.template selfadjointView().twistedBy(p_null); } diff --git a/Eigen/src/SVD/BDCSVD.h b/Eigen/src/SVD/BDCSVD.h index 3fe17b27f..fd7c8a4b2 100644 --- a/Eigen/src/SVD/BDCSVD.h +++ b/Eigen/src/SVD/BDCSVD.h @@ -306,7 +306,7 @@ void BDCSVD::structured_update(Block A, co { // If the matrices are large enough, let's exploit the sparse structure of A by // splitting it in half (wrt n1), and packing the non-zero columns. - DenseIndex n2 = n - n1; + Index n2 = n - n1; MatrixXr A1(n1,n), A2(n2,n), B1(n,n), B2(n,n); Index k1=0, k2=0; for(Index j=0; j Date: Mon, 16 Feb 2015 15:56:11 +0100 Subject: [PATCH 1095/1103] Doc: explain how to free allocated memory in SparseMAtrix --- Eigen/src/SparseCore/SparseMatrix.h | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/Eigen/src/SparseCore/SparseMatrix.h b/Eigen/src/SparseCore/SparseMatrix.h index 4cf4f1826..4c8965802 100644 --- a/Eigen/src/SparseCore/SparseMatrix.h +++ b/Eigen/src/SparseCore/SparseMatrix.h @@ -243,7 +243,13 @@ class SparseMatrix public: - /** Removes all non zeros but keep allocated memory */ + /** Removes all non zeros but keep allocated memory + * + * This function does not free the currently allocated memory. To release as much as memory as possible, + * call \code mat.data().squeeze(); \endcode after resizing it. + * + * \sa resize(Index,Index), data() + */ inline void setZero() { m_data.clear(); @@ -601,6 +607,10 @@ class SparseMatrix } /** Resizes the matrix to a \a rows x \a cols matrix and initializes it to zero. + * + * This function does not free the currently allocated memory. To release as much as memory as possible, + * call \code mat.data().squeeze(); \endcode after resizing it. + * * \sa resizeNonZeros(Index), reserve(), setZero() */ void resize(Index rows, Index cols) From 4dded7322715ab1940539b903014f2a115de1984 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 16:26:47 +0100 Subject: [PATCH 1096/1103] bug #914: fix compiler detection on windows (grafted from 77af14fb6213b833cc37a01e2a8c7eb762db9af8 ) --- cmake/EigenTesting.cmake | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-) diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake index 3eb4e4c07..dc69f65aa 100644 --- a/cmake/EigenTesting.cmake +++ b/cmake/EigenTesting.cmake @@ -359,22 +359,29 @@ macro(ei_get_compilerver VAR) endif() else() # on all other system we rely on ${CMAKE_CXX_COMPILER} - # supporting a "--version" flag + # supporting a "--version" or "/version" flag + + if(WIN32 AND NOT CYGWIN) + set(EIGEN_CXX_FLAG_VERSION "/version") + else() + set(EIGEN_CXX_FLAG_VERSION "--version") + endif() # check whether the head command exists find_program(HEAD_EXE head NO_CMAKE_ENVIRONMENT_PATH NO_CMAKE_PATH NO_CMAKE_SYSTEM_PATH) if(HEAD_EXE) - execute_process(COMMAND ${CMAKE_CXX_COMPILER} --version + execute_process(COMMAND ${CMAKE_CXX_COMPILER} ${EIGEN_CXX_FLAG_VERSION} COMMAND head -n 1 OUTPUT_VARIABLE eigen_cxx_compiler_version_string OUTPUT_STRIP_TRAILING_WHITESPACE) else() - execute_process(COMMAND ${CMAKE_CXX_COMPILER} --version + execute_process(COMMAND ${CMAKE_CXX_COMPILER} ${EIGEN_CXX_FLAG_VERSION} OUTPUT_VARIABLE eigen_cxx_compiler_version_string OUTPUT_STRIP_TRAILING_WHITESPACE) string(REGEX REPLACE "[\n\r].*" "" eigen_cxx_compiler_version_string ${eigen_cxx_compiler_version_string}) endif() ei_get_compilerver_from_cxx_version_string("${eigen_cxx_compiler_version_string}" CNAME CVER) set(${VAR} "${CNAME}-${CVER}") + endif() endmacro(ei_get_compilerver) From 470d26d5803caaf6e41dda5d3e864d8757e32f2d Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 16:48:21 +0100 Subject: [PATCH 1097/1103] Remove some useless typedefs --- Eigen/src/SparseCore/TriangularSolver.h | 4 ---- 1 file changed, 4 deletions(-) diff --git a/Eigen/src/SparseCore/TriangularSolver.h b/Eigen/src/SparseCore/TriangularSolver.h index ccfbdc762..fd1a55bc6 100644 --- a/Eigen/src/SparseCore/TriangularSolver.h +++ b/Eigen/src/SparseCore/TriangularSolver.h @@ -28,7 +28,6 @@ template struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::type LhsEval; typedef typename evaluator::InnerIterator LhsIterator; static void run(const Lhs& lhs, Rhs& other) @@ -66,7 +65,6 @@ template struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::type LhsEval; typedef typename evaluator::InnerIterator LhsIterator; static void run(const Lhs& lhs, Rhs& other) @@ -106,7 +104,6 @@ template struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::type LhsEval; typedef typename evaluator::InnerIterator LhsIterator; static void run(const Lhs& lhs, Rhs& other) @@ -142,7 +139,6 @@ template struct sparse_solve_triangular_selector { typedef typename Rhs::Scalar Scalar; - typedef typename Lhs::StorageIndex StorageIndex; typedef typename evaluator::type LhsEval; typedef typename evaluator::InnerIterator LhsIterator; static void run(const Lhs& lhs, Rhs& other) From 0f464d9d876d9bf6bdfe64d7fe47cd8e35ad3c2c Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 17:05:10 +0100 Subject: [PATCH 1098/1103] bug #897: fix regression in BiCGSTAB(mat) ctor (an all other iterative solvers). Add respective regression unit test. --- .../IterativeSolverBase.h | 18 ++++++++++++++---- test/sparse_solver.h | 9 +++++++++ 2 files changed, 23 insertions(+), 4 deletions(-) diff --git a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h index 38334028a..46bc0ac78 100644 --- a/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h +++ b/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h @@ -54,9 +54,10 @@ public: */ template explicit IterativeSolverBase(const SparseMatrixBase& A) + : mp_matrix(A) { init(); - compute(A.derived()); + compute(mp_matrix); } ~IterativeSolverBase() {} @@ -69,7 +70,7 @@ public: template Derived& analyzePattern(const SparseMatrixBase& A) { - grab(A); + grab(A.derived()); m_preconditioner.analyzePattern(mp_matrix); m_isInitialized = true; m_analysisIsOk = true; @@ -90,7 +91,7 @@ public: Derived& factorize(const SparseMatrixBase& A) { eigen_assert(m_analysisIsOk && "You must first call analyzePattern()"); - grab(A); + grab(A.derived()); m_preconditioner.factorize(mp_matrix); m_factorizationIsOk = true; m_info = Success; @@ -110,7 +111,7 @@ public: template Derived& compute(const SparseMatrixBase& A) { - grab(A); + grab(A.derived()); m_preconditioner.compute(mp_matrix); m_isInitialized = true; m_analysisIsOk = true; @@ -229,6 +230,15 @@ protected: ::new (&mp_matrix) Ref(A); } + void grab(const Ref &A) + { + if(&(A.derived()) != &mp_matrix) + { + mp_matrix.~Ref(); + ::new (&mp_matrix) Ref(A); + } + } + MatrixType m_dummy; Ref mp_matrix; Preconditioner m_preconditioner; diff --git a/test/sparse_solver.h b/test/sparse_solver.h index 42b365eaa..45cfdad25 100644 --- a/test/sparse_solver.h +++ b/test/sparse_solver.h @@ -83,6 +83,15 @@ void check_sparse_solving(Solver& solver, const typename Solver::MatrixType& A, VERIFY(xm.isApprox(refX,test_precision())); } + // test initialization ctor + { + Rhs x(b.rows(), b.cols()); + Solver solver2(A); + VERIFY(solver2.info() == Success); + x = solver2.solve(b); + VERIFY(x.isApprox(refX,test_precision())); + } + // test dense Block as the result and rhs: { DenseRhs x(db.rows(), db.cols()); From 69fa405096f16899ce625eb690440a473a7864bf Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 17:21:16 +0100 Subject: [PATCH 1099/1103] Update circulant custom expression example --- doc/examples/TutorialLinAlgInverseDeterminant.cpp | 2 +- doc/examples/make_circulant.cpp.evaluator | 1 - doc/examples/make_circulant.cpp.expression | 2 +- doc/examples/make_circulant.cpp.traits | 4 ++-- 4 files changed, 4 insertions(+), 5 deletions(-) diff --git a/doc/examples/TutorialLinAlgInverseDeterminant.cpp b/doc/examples/TutorialLinAlgInverseDeterminant.cpp index 43970ff05..14dde5b35 100644 --- a/doc/examples/TutorialLinAlgInverseDeterminant.cpp +++ b/doc/examples/TutorialLinAlgInverseDeterminant.cpp @@ -13,4 +13,4 @@ int main() cout << "Here is the matrix A:\n" << A << endl; cout << "The determinant of A is " << A.determinant() << endl; cout << "The inverse of A is:\n" << A.inverse() << endl; -} \ No newline at end of file +} diff --git a/doc/examples/make_circulant.cpp.evaluator b/doc/examples/make_circulant.cpp.evaluator index 98713cdc0..71c8d4027 100644 --- a/doc/examples/make_circulant.cpp.evaluator +++ b/doc/examples/make_circulant.cpp.evaluator @@ -8,7 +8,6 @@ namespace Eigen { typedef typename nested_eval::type ArgTypeNested; typedef typename remove_all::type ArgTypeNestedCleaned; typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::Index Index; enum { CoeffReadCost = evaluator::CoeffReadCost, diff --git a/doc/examples/make_circulant.cpp.expression b/doc/examples/make_circulant.cpp.expression index a68bcb730..380cd4450 100644 --- a/doc/examples/make_circulant.cpp.expression +++ b/doc/examples/make_circulant.cpp.expression @@ -11,7 +11,7 @@ public: typedef typename Eigen::internal::ref_selector::type Nested; - typedef typename Eigen::internal::traits::Index Index; + typedef Eigen::Index Index; Index rows() const { return m_arg.rows(); } Index cols() const { return m_arg.rows(); } diff --git a/doc/examples/make_circulant.cpp.traits b/doc/examples/make_circulant.cpp.traits index f91e43717..4e04535d3 100644 --- a/doc/examples/make_circulant.cpp.traits +++ b/doc/examples/make_circulant.cpp.traits @@ -1,11 +1,11 @@ namespace Eigen { namespace internal { template - struct traits > + struct traits > { typedef Eigen::Dense StorageKind; typedef Eigen::MatrixXpr XprKind; - typedef typename ArgType::Index Index; + typedef typename ArgType::StorageIndex StorageIndex; typedef typename ArgType::Scalar Scalar; enum { Flags = Eigen::ColMajor, From 1c0e8bcf09db440793abb02b3f5d68e1f4a27585 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 17:21:30 +0100 Subject: [PATCH 1100/1103] Fix unused variable warning. --- Eigen/src/SVD/UpperBidiagonalization.h | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Eigen/src/SVD/UpperBidiagonalization.h b/Eigen/src/SVD/UpperBidiagonalization.h index 9dc470fd9..deaf3dcc4 100644 --- a/Eigen/src/SVD/UpperBidiagonalization.h +++ b/Eigen/src/SVD/UpperBidiagonalization.h @@ -358,6 +358,7 @@ UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::comput { Index rows = matrix.rows(); Index cols = matrix.cols(); + EIGEN_ONLY_USED_FOR_DEBUG(cols); eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols."); @@ -379,6 +380,7 @@ UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::comput { Index rows = matrix.rows(); Index cols = matrix.cols(); + EIGEN_ONLY_USED_FOR_DEBUG(cols); eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols."); From 64b29e06b9f23a496b28fab4b9e29680bc486cc1 Mon Sep 17 00:00:00 2001 From: Martin Drozdik Date: Mon, 16 Feb 2015 18:18:46 +0900 Subject: [PATCH 1101/1103] bug #956: Fixed bug in move constructors of DenseStorage which caused "moved-from" objects to be in an invalid state. --- Eigen/src/Core/DenseStorage.h | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/Eigen/src/Core/DenseStorage.h b/Eigen/src/Core/DenseStorage.h index 4c37fadbd..b52b4dc7b 100644 --- a/Eigen/src/Core/DenseStorage.h +++ b/Eigen/src/Core/DenseStorage.h @@ -304,6 +304,8 @@ template class DenseStorage class DenseStorage class DenseStorage Date: Mon, 16 Feb 2015 19:08:25 +0100 Subject: [PATCH 1102/1103] Add PermutationMatrix::determinant method. --- Eigen/src/Core/PermutationMatrix.h | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+) diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h index de824c129..065000c12 100644 --- a/Eigen/src/Core/PermutationMatrix.h +++ b/Eigen/src/Core/PermutationMatrix.h @@ -251,6 +251,35 @@ class PermutationBase : public EigenBase template friend inline PlainPermutationType operator*(const Transpose >& other, const PermutationBase& perm) { return PlainPermutationType(internal::PermPermProduct, other.eval(), perm); } + + /** \returns the determinant of the permutation matrix, which is either 1 or -1 depending on the parity of the permutation. + * + * This function is O(\c n) procedure allocating a buffer of \c n booleans. + */ + Index determinant() const + { + Index res = 1; + Index n = size(); + Matrix mask(n); + mask.fill(false); + Index r = 0; + while(r < n) + { + // search for the next seed + while(r=n) + break; + // we got one, let's follow it until we are back to the seed + Index k0 = r++; + mask.coeffRef(k0) = true; + for(Index k=indices().coeff(k0); k!=k0; k=indices().coeff(k)) + { + mask.coeffRef(k) = true; + res = -res; + } + } + return res; + } protected: From f0b1b1df9bb893b5fb7d6673b33dad794105c811 Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Mon, 16 Feb 2015 19:09:22 +0100 Subject: [PATCH 1103/1103] Fix SparseLU::signDeterminant() method, and add a SparseLU::determinant() method. --- Eigen/src/SparseLU/SparseLU.h | 55 ++++++++++++++++++++++++++++++++--- test/sparselu.cpp | 3 ++ 2 files changed, 54 insertions(+), 4 deletions(-) diff --git a/Eigen/src/SparseLU/SparseLU.h b/Eigen/src/SparseLU/SparseLU.h index 1e448f2ab..1a21c2a08 100644 --- a/Eigen/src/SparseLU/SparseLU.h +++ b/Eigen/src/SparseLU/SparseLU.h @@ -302,7 +302,50 @@ class SparseLU : public SparseSolverBase >, Scalar signDeterminant() { eigen_assert(m_factorizationIsOk && "The matrix should be factorized first."); - return Scalar(m_detPermR); + // Initialize with the determinant of the row matrix + Index det = 1; + // Note that the diagonal blocks of U are stored in supernodes, + // which are available in the L part :) + for (Index j = 0; j < this->cols(); ++j) + { + for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it) + { + if(it.index() == j) + { + if(it.value()<0) + det = -det; + else if(it.value()==0) + return 0; + break; + } + } + } + return det * m_detPermR * m_detPermC; + } + + /** \returns The determinant of the matrix. + * + * \sa absDeterminant(), logAbsDeterminant() + */ + Scalar determinant() + { + eigen_assert(m_factorizationIsOk && "The matrix should be factorized first."); + // Initialize with the determinant of the row matrix + Scalar det = Scalar(1.); + // Note that the diagonal blocks of U are stored in supernodes, + // which are available in the L part :) + for (Index j = 0; j < this->cols(); ++j) + { + for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it) + { + if(it.index() == j) + { + det *= it.value(); + break; + } + } + } + return det * (m_detPermR * m_detPermC); } protected: @@ -337,7 +380,7 @@ class SparseLU : public SparseSolverBase >, internal::perfvalues m_perfv; RealScalar m_diagpivotthresh; // Specifies the threshold used for a diagonal entry to be an acceptable pivot Index m_nnzL, m_nnzU; // Nonzeros in L and U factors - Index m_detPermR; // Determinant of the coefficient matrix + Index m_detPermR, m_detPermC; // Determinants of the permutation matrices private: // Disable copy constructor SparseLU (const SparseLU& ); @@ -617,7 +660,8 @@ void SparseLU::factorize(const MatrixType& matrix) } // Update the determinant of the row permutation matrix - if (pivrow != jj) m_detPermR *= -1; + // FIXME: the following test is not correct, we should probably take iperm_c into account and pivrow is not directly the row pivot. + if (pivrow != jj) m_detPermR = -m_detPermR; // Prune columns (0:jj-1) using column jj Base::pruneL(jj, m_perm_r.indices(), pivrow, nseg, segrep, repfnz_k, xprune, m_glu); @@ -632,10 +676,13 @@ void SparseLU::factorize(const MatrixType& matrix) jcol += panel_size; // Move to the next panel } // end for -- end elimination + m_detPermR = m_perm_r.determinant(); + m_detPermC = m_perm_c.determinant(); + // Count the number of nonzeros in factors Base::countnz(n, m_nnzL, m_nnzU, m_glu); // Apply permutation to the L subscripts - Base::fixupL(n, m_perm_r.indices(), m_glu); + Base::fixupL(n, m_perm_r.indices(), m_glu); // Create supernode matrix L m_Lstore.setInfos(m, n, m_glu.lusup, m_glu.xlusup, m_glu.lsub, m_glu.xlsub, m_glu.supno, m_glu.xsup); diff --git a/test/sparselu.cpp b/test/sparselu.cpp index 52371cb12..37eb069a9 100644 --- a/test/sparselu.cpp +++ b/test/sparselu.cpp @@ -47,6 +47,9 @@ template void test_sparselu_T() check_sparse_square_abs_determinant(sparselu_colamd); check_sparse_square_abs_determinant(sparselu_amd); + + check_sparse_square_determinant(sparselu_colamd); + check_sparse_square_determinant(sparselu_amd); } void test_sparselu()